RNA Structure and Dynamics: From Molecular Foundations to Therapeutic Design

Abstract

This article provides a comprehensive exploration of RNA structure and dynamics, tailored for researchers and drug development professionals. It covers the foundational principles of RNA's structural hierarchy, from primary sequence to tertiary folding, and its intrinsic dynamics. The review critically examines the latest methodological advancements, including deep learning for 3D structure prediction, atomistic simulations, and integrative experimental techniques. It further addresses central challenges in the field, such as force field accuracy and data scarcity, while presenting optimization strategies. Finally, the article offers a comparative analysis of validation frameworks, benchmarking state-of-the-art computational tools against experimental data. The synthesis of these areas aims to bridge fundamental RNA biology with its rapidly expanding applications in therapeutics.

The RNA Structural Hierarchy: From Sequence to Functional Dynamics

Ribonucleic acid (RNA) is a versatile macromolecule central to cellular functions, serving not only as a genetic information carrier but also as an essential regulator and structural component influencing numerous biological processes [1]. The functionality of RNA is intrinsically linked to its form, where its biological roles—ranging from catalysis to the regulation of gene expression—are dictated by a specific, hierarchical architecture [2] [1]. This architecture is organized into three distinct, yet interdependent, structural levels: primary, secondary, and tertiary. Understanding this structural lexicon is fundamental to advancing our knowledge of cellular biology and is a critical foundation for developing RNA-based therapeutics [2]. This guide provides an in-depth technical examination of these structural levels, framed within the context of modern research on RNA structure and dynamics, and is intended for researchers, scientists, and drug development professionals.

Primary Structure: The Sequence Foundation

The primary structure of an RNA molecule is its most fundamental definition, referring to the precise linear sequence of ribonucleotides—adenosine (A), uridine (U), cytidine (C), and guanosine (G)—linked by phosphodiester bonds [3]. This sequence is the onedimensional blueprint from which all higher-order structures emerge. The acquisition of primary structure is relatively straightforward, achieved through RNA sequencing techniques [3]. It is the primary sequence that contains the information necessary for the molecule to fold into specific shapes via intramolecular interactions, primarily through complementary base pairing. While the primary structure is a simple string of characters, it encodes the potential for the complex two- and three-dimensional folds that define an RNA's functional state.

Secondary Structure: The Two-Dimensional Fold

RNA secondary structure represents the two-dimensional arrangement of the molecule, formed through hydrogen bonding between complementary bases within the same strand [3]. This level of structure is characterized by the formation of double-stranded helical regions interspersed with various single-stranded loops. The folding into secondary structure is a critical step, as unlike proteins, a significant portion of the stabilizing free energy for the RNA molecule is derived from its secondary structure [3]. Furthermore, secondary structures are often well-conserved throughout evolution, aiding in both prediction algorithms and the identification of non-coding RNAs [3].

Core Elements of Secondary Structure

The following are the most common and foundational elements that constitute RNA secondary structure:

• Helices/Stems: Double-stranded regions formed by canonical Watson-Crick base pairing (G-C, A-U) within the same molecule. These are the most stable structural elements.
• Hairpin Loops: Occur when the RNA strand forms a sharp turn upon itself, creating a stem-loop structure. These are among the most frequent secondary structure motifs.
• Internal Loops: Unpaired nucleotides on both strands of a duplex, interrupting an otherwise continuous helix.
• Bulges: Unpaired nucleotides on only one strand of a duplex, causing a local bulge in the double helix.
• Junctions (Multibranch Loops): Points where three or more helices come together. These are often critical for the formation of complex tertiary structures.

Formally, a secondary structure can be defined as a vertex-labeled graph on n vertices (nucleotides) with an adjacency matrix that fulfills specific conditions: the backbone is continuous; each base can pair with at most one other non-adjacent base; and pseudoknots are excluded, meaning if pairs (i, j) and (k, l) exist with i < k < j, then i < l < j must hold [3].

Quantitative Parameters for Secondary Structure Prediction

Computational prediction of secondary structure often relies on thermodynamic parameters derived from empirical data. The table below summarizes key parameters used in free energy minimization algorithms, such as those employed by the RNAfold and RNAstructure software packages.

Table 1: Representative Thermodynamic Parameters (37°C) for RNA Secondary Structure Prediction

Structural Element	Sequence Context	Free Energy Contribution (ΔG° in kcal/mol)	Notes
Stacked Base Pair	5'-GC/3'-CG	-3.3	Most stable stacking interaction
	5'-AU/3'-UA	-1.1	Less stable than GC pair
Hairpin Loop	3-nucleotide loop	~4.0 - 7.0	Penalty depends on loop sequence; highly destabilizing
	4-nucleotide loop	~3.0 - 5.0	Sequence-dependent stability
Internal Loop	1x1 nucleotide loop	~0.5 - 2.0	Small, symmetric loops
	2x2 nucleotide loop	~0.5 - 2.5	Can be stabilizing or destabilizing
Bulge	1-nucleotide bulge	~3.0 - 4.0	Destabilizing due to loss of stacking
	2-nucleotide bulge	~5.0 - 7.0	Increased penalty with size
GU Closing Pair	5'-GU/3'-UG	~-1.0	Non-canonical, but common and stable

Figure 1: RNA secondary structure elements and their relationship to primary and tertiary structure.

Tertiary Structure: The Three-Dimensional Architecture

RNA tertiary structure is the precise three-dimensional shape adopted by the entire nucleic acid polymer [4]. It is the final, functional form of the molecule, achieved through the packing of secondary structural elements against one another via long-range interactions. These interactions are stabilized by a variety of molecular forces, including non-canonical hydrogen bonding, coordination with metal ions (e.g., Mg²⁺), and base stacking. The tertiary structure is what enables RNAs to perform sophisticated functions like molecular recognition and catalysis [4].

Key Tertiary Structural Motifs

Several recurring three-dimensional motifs serve as molecular building blocks for RNA tertiary structures [4]:

• Coaxial Stacking: The end-to-end stacking of two or more helices, forming a quasi-continuous helix. This is a major determinant of tertiary structure stability and is frequently observed in motifs like kissing loops and pseudoknots [4]. The stability of these interactions can often be predicted by adaptations of thermodynamic rules, with prediction accuracy improving significantly when coaxial stacking is considered [4].
• Tetraloop-Receptor Interactions: A highly specific interaction where the nucleotides of a tetraloop (a common 4-nucleotide hairpin loop) dock into a receptor site, often located within an RNA duplex. This combines base-pairing and stacking interactions to stabilize the global fold [4].
• A-Minor Motif: The insertion of an adenosine base into the minor groove of a neighboring helix. This is a ubiquitous and energetically favorable interaction that densely packs helical regions [4].
• Ribose Zipper: A motif involving hydrogen bonding between the 2'-OH groups of two ribose sugars and bases across two closely packed RNA strands.
• Triple Helices (Triplexes): Structures where a third nucleotide strand interacts with the major or minor groove of an RNA double helix via Hoogsteen or reversed Hoogsteen hydrogen bonds. The A-minor motif is a specific type of minor groove triplex [4].
• G-Quadruplexes: Four-stranded structures formed by guanine-rich sequences, where four guanines associate through Hoogsteen hydrogen bonding to form a planar G-quartet. Multiple quartets can stack on top of each other, creating a highly stable structure [4].

Table 2: Experimentally Determined RNA Tertiary Structures and Key Features

RNA Molecule	Function	Key Tertiary Motifs	Primary Experimental Method
tRNA-Phe	Transfer of amino acids	Two coaxial stacks (D-/anticodon arms; acceptor/T arms), L-shape	X-ray Crystallography [4]
Group I Intron	Self-splicing ribozyme	Coaxial stacking (P4-P6 helices), tetraloop-receptors, A-minor motifs	X-ray Crystallography [4]
Group II Intron	Self-splicing ribozyme	Major groove triplex (catalytic core), five-way junction	X-ray Crystallography [4]
Ribosomal RNA	Protein synthesis scaffold	Extensive coaxial stacking (up to 70 bp), multiple A-minor motifs, kink-turns	Cryo-EM [5]
SAM-II Riboswitch	Gene regulation	Major groove triplex, pseudoknot	X-ray Crystallography [4]

Experimental Determination of RNA Structure

A suite of experimental techniques has been developed to probe RNA structure at each hierarchical level. Recent advances have significantly increased the resolution and throughput of these methods [2].

Probing Secondary and Tertiary Structure with Chemical Probes

Chemical probing is a powerful method for obtaining local structural and dynamic information at single-nucleotide resolution. The protocol outlined below is a generalized workflow for in vitro probing of large RNA molecules, which can be adapted for small RNAs or in vivo studies [6].

Protocol: RNA Structure Analysis by Chemical Modification [6]

• RNA Preparation: Synthesize or purify the target RNA. For large RNAs, in vitro transcription followed by gel purification is standard. Refold the RNA by denaturing at 95°C for 2 minutes and snap-cooling on ice, then incubating in the appropriate folding buffer (e.g., containing Mg²⁺) at 37°C for 20-60 minutes.
• Chemical Modification:
  • Reagent Selection: Choose a reagent based on the desired structural information.
    • SHAPE Reagents: 1M7, 1M6, NMIA. These acylate the 2'-hydroxyl group of flexible, unconstrained nucleotides [7].
    • DMS (Dimethyl Sulfate): Methylates the N7 of guanine (Hoogsteen edge) and the N3 of cytosine (Watson-Crick edge), primarily in single-stranded or flexible regions [6].
  • Reaction Incubation: Incubate the folded RNA with an optimal concentration of the chemical probe. Include a "no reagent" control (mock modification) and an "untreated" RNA control.
• RNA Extraction: Terminate the reaction and precipitate the RNA to remove excess reagent and salts.
• Modification Detection by Primer Extension:
  • Use a 5'-end fluorescently or radioactively labeled DNA primer complementary to a region 50-150 nucleotides downstream of the area of interest.
  • Perform reverse transcription. The polymerase will typically stop one nucleotide before a modified base, producing a truncated cDNA product.
  • Run the cDNA products on a high-resolution denaturing polyacrylamide gel alongside a sequencing ladder generated from the untreated RNA using dideoxy sequencing.
• Data Analysis and Normalization:
  • Quantify the band intensities from the gel to determine the reactivity at each nucleotide position.
  • Normalize the data. For SHAPE, reactivities are typically normalized to a scale from 0 (low reactivity/structured) to ~2 (high reactivity/flexible) [7].
  • Integrate the normalized reactivity data as soft constraints into RNA structure prediction algorithms (e.g., RNAstructure) to guide and improve the accuracy of secondary structure modeling.

Figure 2: Experimental workflow for RNA structure probing via chemical modification.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Reagents for RNA Structural Studies

Reagent / Material	Function / Application	Key Characteristics
1M7 (1-methyl-7-nitroisatoic anhydride)	SHAPE reagent for probing nucleotide flexibility [7].	Cell-permeable, highly reactive; provides single-nucleotide resolution data on local backbone dynamics.
DMS (Dimethyl Sulfate)	Chemical probe for base-pairing status (G N7, C N3) [6].	Reveals nucleotide accessibility; can be used in vivo and in vitro.
Mg²⁺ (Magnesium Ions)	Essential cation for RNA folding [4].	Stabilizes tertiary structure by shielding negative charge and forming specific inner-sphere coordination complexes.
Reverse Transcriptase	Enzyme for primer extension in chemical probing [6].	Generates cDNA fragments truncated at modification sites; processivity affects data quality.
Fluorescent DNA Primers	Detection in primer extension assays [6].	Enable high-sensitivity, multiplexed analysis (e.g., SHAPE-Seq).
Structure Prediction Software (e.g., RNAstructure)	Computational modeling of secondary structure [7].	Integrates thermodynamic parameters with experimental data (e.g., SHAPE reactivities) for improved accuracy.
7-Methylguanosine	7-Methylguanosine (m7G)
1-Stearoyl-sn-glycero-3-phosphocholine	1-Stearoyl-sn-glycero-3-phosphocholine, CAS:19420-57-6, MF:C26H54NO7P, MW:523.7 g/mol	Chemical Reagent

Computational Modeling and the Rise of AI

Computational methods are indispensable for interpreting experimental data and predicting RNA structure, especially as the complexity increases from secondary to tertiary folds.

Traditional and Modern Prediction Methods

• Thermodynamics-Based Methods (e.g., RNAfold): These algorithms predict the minimum free-energy secondary structure by calculating the stability of all possible folds using nearest-neighbor parameters [1]. Their accuracy is constrained by the completeness and precision of the underlying energy rules.
• Comparative Sequence Analysis: This method identifies covarying mutations across evolutionarily related sequences to infer base-paired regions. It is highly accurate but requires numerous homologous sequences, which are not always available [1].
• Deep Learning-Based Models: Recent advances have leveraged artificial intelligence to predict RNA structure. A leading example is ERNIE-RNA, a pre-trained language model based on a modified BERT architecture [1]. Its key innovation is an attention mechanism informed by base-pairing rules, which allows it to develop comprehensive representations of RNA architecture during pre-training. ERNIE-RNA has demonstrated a superior ability to capture RNA structural features through zero-shot prediction, outperforming conventional methods, and achieves state-of-the-art performance on various downstream tasks after fine-tuning [1].

The hierarchical lexicon of RNA structure—primary, secondary, and tertiary—provides the conceptual framework for understanding how RNA sequence dictates function. The primary sequence encodes the potential for folding, which is realized in the two-dimensional secondary structure stabilized by base pairing. This two-dimensional scaffold then folds into a complex, functional three-dimensional architecture stabilized by specific tertiary motifs. The integrated use of experimental techniques, such as chemical probing, with advanced computational models, including deep learning, is rapidly advancing our capacity to accurately define RNA structures at all levels. This comprehensive understanding is not only answering fundamental biological questions but is also paving the way for rational design of RNA-targeted therapeutics and synthetic RNA devices, making the mastery of this structural lexicon more critical than ever for researchers and drug developers.

The classical view of RNA molecules as static, well-defined structures has been fundamentally overturned. It is now recognized that many functional RNAs exist not as single conformations but as dynamic ensembles—populations of alternative structural states that interconvert. These conformational landscapes are not mere structural curiosities; they are fundamental to RNA's ability to regulate gene expression, catalyze reactions, and respond to cellular signals with exquisite timing and precision. For researchers and drug development professionals, understanding these ensembles is paramount, as they represent a new frontier for therapeutic intervention. The functional properties of crucial viral proteins, such as RNA-dependent RNA polymerases (RdRps), are demonstrably modulated by native substrates of dynamic and interconvertible conformational ensembles, many populated by essential flexible or intrinsically disordered regions (IDRs) [8]. This whitepater explores the experimental and computational frameworks illuminating RNA dynamics, providing a resource to accelerate the discovery of regulatory RNA switches and novel antiviral strategies.

Experimental Mapping of RNA Structural Ensembles

Moving from studying a single RNA structure to characterizing its full conformational repertoire requires sophisticated experimental methodologies that can capture structural heterogeneity.

Transcriptome-Wide Ensemble Deconvolution with DRACO

A significant advance in the field is the development of DRACO, an algorithm designed to deconvolve RNA structure ensembles from chemical probing data read out through Mutational Profiling (MaP) [9]. In MaP experiments, RNA in its native cellular environment is treated with a chemical probe like Dimethyl Sulfate (DMS), which covalently modifies unpaired adenines and cytosines. During reverse transcription, these modifications are recorded as cDNA mutations, which are then decoded by high-throughput sequencing. DRACO analyzes co-mutation patterns within individual sequencing reads to estimate the number of conformations in an ensemble, reconstruct their secondary structures, and determine their relative stoichiometries [9].

Experimental Protocol: DMS-MaPseq for In Vivo RNA Structural Ensembles

• Cell Culture and Probing: Grow E. coli (e.g., DH5α or TOP10 strains) to exponential phase in suitable media at 37°C. Treat living cells with DMS for a precise, optimized duration (e.g., 5-10 minutes) to achieve single-hit kinetics.
• RNA Extraction and rRNA Depletion: Lyse cells and extract total RNA. Remove ribosomal RNA to enrich for mRNA and non-coding RNAs.
• Library Preparation and Sequencing: Perform reverse transcription with a MarathonRT enzyme, which is tolerant of DMS modifications and introduces mutations at modification sites. Construct sequencing libraries and sequence on an Illumina platform to a high depth (>1 billion paired-end reads recommended for transcriptome-wide coverage).
• Data Processing and Ensemble Deconvolution:
  • Map sequencing reads to a reference transcriptome.
  • Calculate mutation rates per nucleotide from the sequencing data.
  • Run DRACO analysis on sliding windows across transcripts. The tool requires a minimum effective read depth (e.g., 2,000x) to robustly identify multiple conformations.
  • DRACO outputs the number of significant conformations, their predicted secondary structures, and their relative abundances for each RNA region.

This approach has revealed the astonishing complexity of the RNA structural landscape. A transcriptome-wide analysis in E. coli showed that approximately 16.6% of analyzed genomic regions populated two or more distinct conformations under standard growth conditions [9]. This methodology has successfully identified known regulatory switches like RNA thermometers and riboswitches, confirming its utility in discovering functional dynamics.

Key Reagents for Ensemble Mapping

Table 1: Essential Research Reagents for RNA Ensemble Mapping Experiments

Reagent / Tool	Function	Example Use Case
Dimethyl Sulfate (DMS)	Chemical probing reagent that modifies unpaired A and C nucleotides in vivo and in vitro.	Mapping single-stranded regions in living cells for DRACO analysis [9].
MarathonRT	Reverse transcriptase engineered to read through DMS modifications and record them as mutations in cDNA.	Essential for DMS-MaPseq protocol to generate mutation data for ensemble deconvolution [9].
DRACO Algorithm	Computational tool for deconvolving RNA structural ensembles from MaP sequencing data.	Identifying the number, structure, and population of conformations from co-mutation patterns [9].
5'UTR-MaP Method	Specialized protocol for transcriptome-wide mapping of 5' untranslated region structures in eukaryotes.	Uncovering RNA structural switches regulating open reading frame usage in human cells [9].

Computational Approaches for Prediction and Design

Complementing experimental methods, computational frameworks provide a powerful means to predict, model, and design RNA conformational states.

Graph-Theoretic Models with RNA-As-Graphs (RAG)

The RNA-As-Graphs (RAG) approach utilizes graph theory to represent, analyze, and organize RNA secondary structures [10] [11]. This coarse-grained method simplifies RNA 2D structures into tree graphs where unpaired loops are vertices and base-paired helices are edges. This abstraction reduces complexity and allows the application of graph theory to classify existing RNA topologies and predict novel, plausible RNA motifs [11].

The RAG-Web server provides a user-friendly interface for three key modules, forming a pipeline for RNA structure prediction and design [10]:

• RAG Sampler: Takes an RNA sequence and 2D structure as input and uses the RAGTOP tool to sample candidate 3D graph topologies, scoring them with a knowledge-based potential.
• RAG Builder: Builds 3D atomic models from candidate graphs generated by RAG Sampler. It uses the Fragment Assembly for RAG (F-RAG) tool to assemble atomic fragments from the RAG-3D database.
• RAG Designer: Designs RNA sequences that are predicted to fold into a target tree graph topology, facilitating the de novo design of RNA molecules.

A current limitation of the RAG pipeline is that it handles RNAs of up to 200 nucleotides and topologies with a maximum of 13 vertices, and it does not yet incorporate energy minimization [10].

Predicting Conformational Heterogeneity in Proteins with AlphaFold2

While not an RNA-specific tool, AlphaFold2 (AF2) has impacted the study of RNA-binding proteins, such as viral RdRps. A key insight is that AF2's low per-residue confidence scores (pLDDT) often correlate with intrinsically disordered regions (IDRs) [8]. These IDRs, which lack a stable 3D structure, make up nearly 16% of conserved RdRp domains across Riboviria and are essential for their function as dynamic conformational ensembles [8]. Thus, AF2's low-confidence predictions can serve as a proxy for identifying regions likely to undergo conformational dynamics, guiding further experimental investigation.

Quantitative Data and Analysis

The study of RNA dynamics generates critical quantitative metrics that allow for the comparison and validation of conformational states.

Table 2: Quantitative Metrics from RNA Dynamics Studies

Metric	Value / Example	Context and Significance
Genomic Regions in Ensembles	16.6% [9]	Percentage of the E. coli transcriptome found to populate two or more conformations.
IDR Content in RdRps	~16% [8]	The fraction of conserved RdRp domain composed of intrinsically disordered regions, crucial for dynamics.
Effective Sequencing Depth	>2,000x [9]	Minimum recommended read depth for robust ensemble deconvolution with DRACO.
DMS Mutation Preference	A: ~56%, C: ~34% [9]	Typical distribution of DMS-induced mutations, reflecting the reagent's specificity for unpaired A and C nucleotides.
RAG Size Limit	200 nucleotides, 13 vertices [10]	Current upper size boundary for structures processed by the RAG-Web server.

Applications in Virology and Drug Discovery

The understanding of RNA and protein conformational dynamics has direct and profound implications for virology and therapeutic development. RdRps, the central enzymes in RNA virus replication, are proven potent antiviral targets due to their high evolutionary conservation [8]. However, traditional structural biology often provides static snapshots that miss the essential dynamics of these proteins. Research shows that RdRps exist as "dynamic conformational ensembles that adapt to the functional requirements of the viral cycle" [8]. For instance, molecular dynamics studies have revealed that conserved structural motifs within the RdRp act as sequence-specific conformational switches during the nucleotide incorporation cycle [8].

Targeting these specific conformational states, rather than the static structure, offers a promising avenue for designing novel antivirals that can trap the enzyme in a non-functional state or allosterically modulate its dynamics. The quantitative and predictive frameworks provided by ensemble mapping and computational design are thus invaluable tools for rational drug discovery against RNA viruses.

Visualizing Experimental and Analytical Workflows

The following diagrams illustrate the core methodologies discussed in this whitepaper, providing a clear visual representation of the complex workflows involved in studying RNA ensembles.

DRACO Ensemble Deconvolution Workflow

Diagram Title: DRACO Ensemble Deconvolution from MaP Data

RAG Computational Pipeline

Diagram Title: RAG Structure Prediction and Design Pipeline

RNA biology is fundamentally governed by a hierarchical relationship between its structure and dynamic behavior, moving beyond the classical view of RNA as a mere information carrier. The intricate three-dimensional architectures and conformational ensembles adopted by RNA molecules are essential for their diverse functions in gene regulation, splicing, and catalysis. This whitepaper explores the intrinsic connection between RNA structure, dynamics, and function, emphasizing how recent methodological advances in experimental biophysics, computational modeling, and artificial intelligence are illuminating these relationships. Within the context of a broader thesis on RNA structure and dynamics, this review highlights how understanding these principles is accelerating the development of RNA-targeted therapeutics, enabling researchers to target previously "undruggable" pathways in various diseases, including viral infections, cancer, and genetic disorders.

The central dogma of molecular biology has undergone a significant transformation, with RNA now recognized not merely as a passive messenger but as a versatile macromolecule whose functions are critically determined by its structural dynamics [12]. RNA molecules exhibit a hierarchical organization where their primary sequences fold into specific secondary and tertiary structures that ultimately dictate their biological activities [1]. Unlike the relatively stable DNA double helix, RNA structures are notably dynamic and flexible, adopting multiple conformational states that enable them to perform diverse regulatory and catalytic functions [13].

The intrinsic dynamic flexibility and pronounced conformational heterogeneity of RNA endow it with diverse functional capabilities that are fundamental to cellular processes [13]. RNA can fold into complex three-dimensional structures, enabling it to perform a variety of functions beyond coding for proteins, with non-coding RNAs (ncRNAs) serving as crucial regulators of gene expression [12]. This structural complexity creates unique binding sites for small molecules, proteins, and other RNAs, making RNA an attractive target for therapeutic intervention [14]. Understanding the conformational ensembles of RNA is therefore fundamental for elucidating its intricate mechanisms of action, advancing RNA-targeted drug discovery, and facilitating the design of RNA-based therapeutic strategies [13].

RNA Structural Hierarchy and Functional Implications

Secondary Structure Motifs and Their Functional Roles

The secondary structure of RNA is defined by canonical base pairing, including Watson-Crick pairs (A-U and G-C) and the wobble base pair (G-U), which form through hydrogen bonding and create structural motifs that serve as the building blocks for higher-order organization [15]. These paired nucleotides form helices, while unpaired bases create various structural motifs with distinct functional implications:

• Hairpin loops: Often serve as recognition elements for proteins and other RNAs, playing critical roles in transcriptional termination and RNA interference.
• Bulges and internal loops: Introduce flexibility and create binding platforms for ligands and proteins, frequently found in ribosomal RNA and riboswitches.
• Multi-branch loops (junctions): Act as organizational hubs that direct the three-dimensional folding of complex RNA structures.
• Pseudoknots: Involve base pairing between a loop and complementary sequence outside the loop, creating complex tertiary architectures essential for ribosomal frameshifting and ribozyme activity.

Tertiary Structure and Dynamics

RNA tertiary structure emerges from the spatial arrangement of secondary structure elements through long-range interactions, including A-minor motifs, ribose zippers, and tetraloop-receptor interactions [15]. These tertiary interactions stabilize the overall RNA architecture and create specific binding pockets and catalytic sites. The functional state of RNA is not a single static structure but rather a dynamic conformational ensemble, where transitions between different states mediate biological function [13]. For example, riboswitches undergo conformational changes upon ligand binding that modulate gene expression, and the HIV-1 Trans-Activation Response (TAR) element exists in multiple conformational states that regulate viral replication [13].

Table 1: Key RNA Structural Elements and Their Functional Significance

Structural Element	Description	Biological Functions	Therapeutic Relevance
G-Quadruplexes	Four-stranded structures from G-rich sequences	Gene regulation, telomere maintenance	Cancer therapeutics, antiviral targets [14]
Pseudoknots	Nested base pairs between loop and external region	Ribosomal frameshifting, ribozyme catalysis	Antiviral targets (SARS-CoV-2 frameshift element) [16]
Riboswitches	Ligand-binding regulatory elements	Metabolic pathway regulation, gene expression	Antibacterial drug targets [14]
Tetraloops	Stable four-nucleotide hairpin loops	Folding nucleation, protein recognition	Structural motifs for engineering [13]

Methodological Advances in Studying RNA Structure and Dynamics

Experimental Techniques

Biophysical Approaches

Nuclear Magnetic Resonance (NMR) spectroscopy, particularly 19F NMR, has emerged as a powerful tool for probing RNA structures and dynamics. This method offers high sensitivity and simplicity for studying RNA folding, conformational changes, and ligand binding interactions [14]. 19F NMR requires site-specific labeling of RNA with fluorine atoms, which can be incorporated into the sugar moiety or nucleobase through chemical synthesis, chemo-enzymatic methods, or in vitro transcription [14]. Other structural methods include:

• X-ray crystallography: Provides atomic-resolution structures but requires crystallization, which can be challenging for dynamic RNA molecules [1].
• Cryogenic electron microscopy (cryo-EM): Enables structural determination of large RNA-protein complexes without crystallization [1].
• Chemical probing methods (e.g., SHAPE): Provide information on RNA secondary structure and dynamics in solution by measuring nucleotide accessibility [1].

The Scientist's Toolkit: Key Research Reagents

Table 2: Essential Research Reagents for RNA Structural Studies

Reagent / Technology	Function/Application	Key Features
19F-labeled nucleotides	Site-specific labeling for NMR studies	Enables monitoring of local environment and dynamics [14]
Chem-CLIP	Maps drug-binding pockets in RNA	Identifies druggable sites in structured RNA [16]
Mirafloxacin derivative	Targets SARS-CoV-2 frameshift element	Serves as scaffold for antiviral optimization [16]
Lipid Nanoparticles (LNPs)	RNA delivery system	Protects RNA from degradation, enhances cellular uptake [17]
GalNAc conjugates	Liver-specific RNA delivery	Targeted delivery for therapeutic applications [17]
Mono(2-ethyl-5-oxohexyl) adipate	Mono(2-ethyl-5-oxohexyl) Adipate|Plasticizer Metabolite	Mono(2-ethyl-5-oxohexyl) adipate is a specific metabolite used in human biomonitoring to assess exposure to adipate plasticizers. For Research Use Only. Not for human or veterinary use.
4-Methylcinnamic Acid	4-Methylcinnamic Acid, CAS:1866-39-3, MF:C10H10O2, MW:162.18 g/mol	Chemical Reagent

Computational and AI-Driven Approaches

Deep Learning for Structure Prediction

Machine learning, particularly deep learning, has revolutionized RNA structure prediction by leveraging large-scale sequence and structural data. These approaches can be broadly categorized into:

• Energy-based methods: Traditionally used thermodynamic parameters to predict minimum free energy structures but often struggled with accuracy [15].
• Comparative methods: Leveraged evolutionary information from multiple sequence alignments but were limited for RNAs with few homologs [1].
• Deep learning approaches: Use neural networks to learn structure-sequence relationships directly from data, significantly improving prediction accuracy [18].

ERNIE-RNA represents a recent breakthrough—an RNA language model based on a modified BERT architecture that incorporates base-pairing restrictions into its attention mechanism [1]. This model develops comprehensive representations of RNA architecture during pre-training and demonstrates remarkable capability in zero-shot RNA secondary structure prediction, outperforming conventional methods like RNAfold and RNAstructure [1].

Conformational Ensemble Modeling

DynaRNA addresses the critical challenge of capturing RNA's dynamic conformational ensembles using a diffusion-based generative model [13]. This approach employs a denoising diffusion probabilistic model (DDPM) with an equivariant graph neural network (EGNN) to directly model RNA 3D coordinates, enabling rapid exploration of RNA conformational space [13]. Unlike methods that predict single static structures, DynaRNA generates multiple conformations that represent the natural dynamic states of RNA molecules, effectively capturing rare excited states and reproducing experimental geometries without requiring multiple sequence alignment information [13].

The following diagram illustrates the integrated experimental and computational workflow for determining RNA structure and dynamics:

RNA-Targeted Therapeutic Development

Leveraging Structural Principles for Drug Design

The structured elements of RNA create unique binding pockets that can be targeted by small molecules, offering therapeutic potential for various diseases [14]. Several strategic approaches have been developed to modulate RNA function:

• Direct binding and stabilization: Small molecules can bind to specific RNA structures and stabilize or alter their conformations, thereby modulating function [14].
• Splicing modulation: Compounds like risdiplam for spinal muscular atrophy bind to RNA and affect interactions with RNA-binding proteins to modulate RNA splicing [14].
• Targeted degradation: Ribonuclease targeting chimeras (RIBOTACs) and proximity-induced nucleic acid degraders (PINADs) link target RNA to degradation machinery, reducing RNA levels [14].

Recent work on targeting the SARS-CoV-2 frameshift element exemplifies rational RNA-targeted drug discovery. Disney and colleagues identified "druggable pockets" in the structured viral RNA and used systematic chemistry, computational methods, and robotic drug discovery to develop Compound 6, which causes viral proteins to misfold and be degraded by cellular machinery [16]. This platform approach can be applied to numerous RNA-based viruses, including influenza, norovirus, Ebola, and Zika [16].

RNA-Based Therapeutics

RNA molecules themselves can serve as therapeutic agents, particularly for targeting proteins considered "undruggable" by small molecules [19]. Only approximately 15% of human proteins have binding pockets suitable for traditional small-molecule drugs, making RNA a promising alternative therapeutic modality [19]. RNAtranslator, a generative language model that formulates protein-conditional RNA design as a sequence-to-sequence translation problem, enables the design of RNA sequences that bind to specific protein targets [19]. This approach learns a joint representation of RNA and protein interactions from large-scale datasets and can generate binding RNA sequences for any given protein target, including those with no available RNA-interaction data [19].

Table 3: RNA-Based Therapeutic Modalities and Their Mechanisms

Therapeutic Modality	Mechanism of Action	Applications
Antisense Oligonucleotides (ASOs)	Bind to complementary RNA sequences, regulating gene expression through splicing alteration, mRNA degradation, or translation inhibition [12]	Genetic disorders, cancers, viral infections
Small Interfering RNAs (siRNAs)	Induce RNA interference (RNAi) by forming RISC complex, leading to cleavage and degradation of target mRNAs [12]	Genetic diseases, antiviral therapies
mRNA Therapeutics	Utilize the body's cellular machinery to produce therapeutic proteins or antigens [17]	Vaccines (e.g., COVID-19), protein replacement therapies
RNA Aptamers	Structured RNAs that bind specific molecular targets with high affinity and specificity [19]	Target validation, diagnostic and therapeutic applications

The following diagram illustrates the strategic approaches for developing RNA-targeted therapeutics:

The intricate relationship between RNA structure and dynamics fundamentally governs RNA biology, with conformational ensembles determining functional outcomes across diverse cellular processes. Advances in experimental biophysics, particularly 19F NMR and chemical probing methods, combined with revolutionary AI-driven approaches like ERNIE-RNA and DynaRNA, are providing unprecedented insights into RNA structural principles. These methodological innovations are accelerating the development of RNA-targeted therapeutics, enabling researchers to design small molecules and RNA-based therapies that modulate previously inaccessible disease pathways. As our understanding of RNA structure-dynamics-function relationships deepens, the potential for creating transformative treatments for viral diseases, cancer, genetic disorders, and other conditions continues to expand, heralding a new era in RNA-targeted drug discovery.

The understanding of RNA structure and dynamics has transitioned from viewing RNA as a simple informational molecule to recognizing it as a dynamic, multifunctional entity whose conformational ensembles directly govern its cellular functions [20]. This fundamental insight has catalyzed the development of a revolutionary class of medicines that target or utilize RNA. The commercial and clinical landscape for RNA therapeutics has expanded dramatically, moving from a niche modality to a mainstream therapeutic platform with applications across rare diseases, oncology, infectious diseases, and beyond [17] [21]. The validation of mRNA vaccines during the COVID-19 pandemic, combined with successive approvals of RNA interference (RNAi) and antisense oligonucleotide (ASO) drugs, has established RNA as a versatile and druggable target and modality [22] [23]. The global RNA therapy clinical trials market, valued at $2.82 billion in 2024, is projected to grow to $4.11 billion by 2034, reflecting a compound annual growth rate (CAGR) of 3.84% [24]. This growth is underpinned by a robust pipeline of over 5,500 active drug candidates and 2,500 clinical trials as of mid-2025 [25]. This review examines the current state of RNA-targeted therapeutics, exploring the fundamental structural principles, diverse modalities, clinical progress, and future directions that define this rapidly evolving field.

RNA Structural Dynamics: The Foundation of Therapeutic Intervention

RNA as a Dynamic Ensemble

The therapeutic potential of RNA targets is inextricably linked to their structural biology. Unlike static depictions, RNA molecules exist as dynamic ensembles of conformations, constantly sampling alternative secondary and tertiary structures on timescales from picoseconds to seconds [20]. This structural plasticity is not random but is fundamental to RNA function, enabling gene regulation through mechanisms such as riboswitches, alternative splicing, and microRNA maturation. The ensemble-based perspective is crucial for understanding how cellular cues, ligands, proteins, and pathogenic mutations influence RNA activity by shifting the equilibrium between pre-existing conformational states [20]. For instance, single-nucleotide polymorphisms (SNPs) linked to diseases like retinoblastoma and breast cancer can collapse diverse structural ensembles into single, often dysfunctional, conformations or alter junction topologies critical for tertiary folding, thereby disrupting processing and function [20].

Functional Consequences of RNA Dynamics

The structural dynamics of RNA directly enable its diverse regulatory mechanisms as shown in Table 1. Riboswitches control gene expression by undergoing ligand-induced folding into alternative secondary structures [20]. The catalytic cycles of ribozymes involve transitions between distinct tertiary structures for substrate binding, catalysis, and product release [20]. Furthermore, the recognition of RNA by proteins often requires the melting of secondary structure to expose single-stranded binding motifs, a process critical for alternative splicing factors [20]. The cellular environment itself, including processes like liquid-liquid phase separation, can influence RNA folding and its subsequent activity, adding another layer of regulatory complexity [20]. This deep understanding of RNA as a dynamic, structured polymer provides the foundational rationale for developing small molecules and oligonucleotides that specifically target these functional structures and their conformational transitions.

Table 1: Functional Consequences of RNA Structural Dynamics

RNA Class/Example	Structural Change	Functional Outcome	Therapeutic Relevance
Riboswitches	Alternative secondary structure upon ligand binding [20]	Gene regulation (ON/OFF) [20]	Target for small molecules (antibiotics) [20]
Ribozymes	Cycling through different tertiary structures [20]	Catalysis of biochemical reactions [20]	Engineered for therapeutic cleavage
pre-microRNA (e.g., let-7)	Protein-induced structural change (e.g., by LIN28A) [20]	Inhibition of maturation by blocking Dicer/Drosha recognition [20]	Target for inhibitors of silencing
HIV-1 5' Leader RNA	Changes in secondary structure [20]	Regulates switch between translation and genome packaging [20]	Target for antiviral drugs
Long Non-coding RNAs (lncRNAs)	Conformational changes upon scaffolding [20]	Assembly of ribonucleoprotein (RNP) complexes [20]	Target for modulating epigenetic states

The RNA Therapeutic Modality Landscape

The field has diversified into several distinct therapeutic modalities, each harnessing different aspects of RNA biology.

Established Modalities

• Antisense Oligonucleotides (ASOs): These single-stranded oligonucleotides modulate gene expression by binding to complementary RNA sequences through Watson-Crick base pairing, leading to target degradation (via RNase H1 recruitment) or modulation of splicing, translation, or miRNA activity [17] [22]. Approved ASOs include Nusinersen (Spinraza) for spinal muscular atrophy and Eplontersen for transthyretin amyloidosis [22].
• Small Interfering RNA (siRNA): These double-stranded RNAs induce sequence-specific degradation of complementary mRNA via the RNA-induced silencing complex (RISC) [22]. Key approved siRNAs include Patisiran (Onpattro) for hATTR amyloidosis and Inclisiran (Leqvio) for hypercholesterolemia, the latter utilizing GalNAc conjugation for efficient liver delivery [22] [21].
• Messenger RNA (mRNA): This modality involves delivering in vitro-transcribed mRNA encoding therapeutic proteins or antigens into the cytoplasm, where the host cellular machinery translates it into the desired protein [22]. Beyond the successful COVID-19 vaccines, the pipeline includes personalized cancer vaccines (e.g., mRNA-4157) and vaccines for other infectious diseases like RSV (mRNA-1345) [26] [22].

Emerging Modalities

• Self-Amplifying RNA (saRNA): Derived from alphavirus genomes, saRNA encodes both the antigen and viral replication machinery, enabling intracellular amplification of the original RNA strand and prolonged antigen expression, allowing for lower doses [26] [22].
• Circular RNA (circRNA): Engineered with a covalently closed continuous loop, circRNA lacks free ends, conferring exceptional stability and resistance to exonucleases, enabling prolonged protein expression [26]. This makes it a promising platform for vaccines and therapeutic protein expression, with candidates from Orna Therapeutics entering clinical trials [26] [23].
• RNA-Targeting Small Molecules: A growing class of traditional small molecules designed to bind specific RNA structures and alter their function, such as modulating splicing [27]. The market for these molecules is projected to grow from $2.77 billion in 2024 to $7.03 billion by 2034 [27].
• CRISPR-Based RNA Targeting: Systems like CRISPR-Cas13 can be programmed to bind and cleave specific RNA sequences, offering a powerful platform for RNA editing and knockdown without altering the genome [22].

Clinical and Commercial Landscape

The clinical pipeline for RNA therapeutics is vast and expanding rapidly, reflecting broad investment across therapeutic areas and technology platforms.

Clinical Trial and Market Analysis

The RNA therapy clinical trials market is experiencing steady growth, with the number of active drugs increasing by over 650 in the first half of 2025 alone [25]. The distribution of trials and market characteristics are summarized in Table 2.

Table 2: RNA Therapeutics Clinical Trial and Market Landscape (2024-2025)

Parameter	Market Value & Distribution	Therapeutic Area Focus	Modality Trends
Global Market Size (2024)	$2.82 Billion [24]	Rare Diseases (22% share) [24]	mRNA (35.7% share) [23]
Projected Market (2034)	$4.11 Billion [24]	Anticancer (Fastest growing) [24]	Self-amplifying RNA (22.5% CAGR) [23]
Market CAGR (2025-2034)	3.84% [24]	Neurodegenerative Diseases (Largest segment for small molecules) [27]	RNA Interference (Significant growth) [24]
Regional Leadership	North America (37% share) [24]	Infectious Diseases (Beyond COVID-19) [23]	Circular RNA (Emerging) [26]
Fastest Growing Region	Asia Pacific (4.52% CAGR) [24]	Cardiology & Metabolic Disorders [23]	RNA Splicing Modifiers (Small Molecules) [27]

Analysis of Key Commercialized Therapies

The commercial success of several RNA therapeutics has validated the entire field. siRNA therapeutics like Alnylam's Patisiran, Givosiran, and Inclisiran have demonstrated the viability of RNAi as a platform, particularly with the advent of GalNAc conjugation for subcutaneous, hepatic-targeted delivery [22] [21]. Similarly, ASOs like Spinraza and the recently approved Tryngolza (olezarsen) from Ionis have shown significant clinical impact in rare diseases [21]. The mRNA vaccines from Pfizer/BioNTech and Moderna not only addressed a global health crisis but also generated massive revenue, enabling those companies to reinvest heavily in expanding their RNA platforms into oncology and other infectious diseases [23] [21]. The revenue potential is underscored by projections that multiple RNAi and ASO therapies (Amvuttra, Leqvio, Spinraza, Wainua) are expected to exceed $1 billion in annual revenue by 2030 [21].

Enabling Technologies: Delivery and Manufacturing

Delivery Platforms

Effective delivery remains a central challenge and area of innovation. The current clinical landscape is dominated by two primary strategies:

• GalNAc Conjugation: A sugar moiety that binds with high affinity to the asialoglycoprotein receptor (ASGPR) on hepatocytes, enabling efficient siRNA and ASO delivery to the liver with subcutaneous administration [17] [21]. This technology has been pivotal for the success of drugs like Inclisiran and Givosiran.
• Lipid Nanoparticles (LNPs): These are the primary vehicle for mRNA delivery, protecting the payload from degradation and facilitating cellular uptake via endocytosis [17] [22]. Next-generation LNPs are focusing on tissue-specific targeting (e.g., to muscle or the CNS) and improving endosomal escape efficiency, a key bottleneck where currently less than 10% of LNP cargo reaches the cytosol [17] [23]. Innovations include LNPs with internal fat layers for high mRNA loading and ligands like PD-L1 binding peptides for tumor targeting [17] [26].

Other delivery modalities under investigation include extracellular vesicles, novel polymer nanoparticles, and cell-specific ligand conjugates for extrahepatic tissues like muscle and the central nervous system [17] [21].

Manufacturing and Personalized Therapeutics

The paradigm of personalized RNA therapeutics is most advanced in oncology, with personalized cancer vaccines designed based on the unique neoantigen profile of an individual's tumor [17] [26]. Manufacturing these bespoke therapies requires rapid, automated processes. Innovations have reduced production timelines for personalized vaccines from nine weeks to under four weeks [26]. However, costs remain high, exceeding $100,000 per patient, driving research into hybrid approaches that combine off-the-shelf tumor-associated antigens with patient-specific neoantigens to balance personalization with scalability [26]. The integration of artificial intelligence and closed-system automated manufacturing platforms is crucial for further streamlining production and quality control [26].

Experimental Workflows in RNA Therapeutic Development

The development of RNA therapeutics relies on a series of interconnected experimental workflows, from target identification to final formulation.

Core Methodological Workflow

The general pipeline for developing an RNA therapeutic, from initial design to in vitro validation, involves several critical stages as visualized below.

Diagram 1: RNA Therapeutic In Vitro Development Workflow. This diagram outlines the key stages from target identification through to successful in vitro validation, highlighting critical sub-tasks in design and analysis.

The Scientist's Toolkit: Essential Reagents and Technologies

The development and analysis of RNA therapeutics depend on a suite of specialized research reagents and platforms as shown in Table 3.

Table 3: Essential Research Reagent Solutions for RNA Therapeutic Development

Research Reagent / Technology	Function & Application	Specific Examples & Notes
In Vitro Transcription (IVT) Kits	Enzymatic synthesis of research-grade mRNA; template DNA is transcribed into RNA using RNA polymerase (e.g., T7, SP6).	Used for early-stage prototype mRNA synthesis for in vitro and in vivo testing [22].
Nucleotide Analogs	Chemically modified nucleotides (e.g., N1-methylpseudouridine) incorporated into RNA during synthesis to enhance stability and reduce immunogenicity [22].	Critical for improving the therapeutic properties of mRNA and siRNA [22].
Lipid Nanoparticle (LNP) Formulation Kits	Pre-formed or customizable lipid mixtures for encapsulating RNA and facilitating its delivery into cells in vitro and in vivo.	Used for screening and optimizing delivery formulations [22]. Components often include ionizable lipids, phospholipids, cholesterol, and PEG-lipids [22].
GalNAc Conjugation Reagents	Chemical linkers and activated GalNAc moieties for conjugating oligonucleotides (siRNA, ASO) to enable targeted delivery to hepatocytes.	A standard for developing liver-targeted siRNA therapeutics [21].
AI/ML Design Platforms	Software and algorithms (e.g., eSkip-Finder, Cm-siRPred) to predict optimal ASO/siRNA sequences, activity of chemical modifications, and LNP formulations [17].	Reduces discovery cycles from years to months; e.g., MIT's COMET for LNP selection [17] [23].
Cell-Based Functional Assays	Reporter assays (e.g., luciferase), qRT-PCR, flow cytometry, and Western blotting to measure gene expression knockdown (siRNA/ASO) or protein production (mRNA).	Standard for confirming functional activity and potency of the RNA therapeutic in vitro.
Dimabefylline	Dimabefylline, CAS:1703-48-6, MF:C16H19N5O2, MW:313.35 g/mol	Chemical Reagent
Strophanthidin	Strophanthidin, CAS:66-28-4, MF:C23H32O6, MW:404.5 g/mol	Chemical Reagent

Future Perspectives and Challenges

The future of RNA therapeutics is bright but requires overcoming several key hurdles. A primary challenge is expanding delivery beyond the liver. While GalNAc-conjugates excel for hepatic targets, reaching other tissues like the CNS, muscle, and lungs efficiently remains a major focus of R&D [23]. Related to delivery is the problem of endosomal escape inefficiency, which caps the bioavailability of RNA payloads and necessitates higher, potentially more toxic, doses [23]. Other persistent challenges include the cold-chain requirements for many RNA formulations, which limit their use in low-resource settings, and the high cost of goods for personalized therapies [26] [23].

Future progress will be driven by several key trends. The integration of artificial intelligence will accelerate throughout the drug development cycle, from target identification and sequence design to predicting RNA structural dynamics and optimizing LNP formulations [26] [23] [21]. The convergence of CRISPR gene editing with RNA therapies offers opportunities for enhanced immune system programming and ex vivo cell engineering [26]. Furthermore, the regulatory landscape is evolving, with the FDA releasing new guidance for therapeutic cancer vaccines in 2024, and the first commercial mRNA cancer vaccine approvals are anticipated by 2029 [26]. As the field matures, the focus will shift from proving technological feasibility to solving practical challenges in tissue targeting, long-term safety, scalable production, and ensuring equitable global access, ultimately fulfilling the promise of truly personalized and precision RNA medicine [22] [21].

Cutting-Edge Tools for Probing RNA Architecture and Interactions

The functional versatility of RNA, spanning from catalytic roles to regulatory mechanisms, is intimately linked to its ability to form intricate and hierarchical structures. Understanding these structures is crucial for elucidating molecular mechanisms and developing RNA-targeted therapeutics. This whitepaper provides an in-depth technical guide to the core experimental methodologies powering modern RNA structural biology. We examine the principles, applications, and detailed protocols of key techniques—including X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy (cryo-EM), small-angle X-ray scattering (SAXS), and single-molecule Förster Resonance Energy Transfer (smFRET). Framed within the broader context of RNA structure and dynamics research, this review equips researchers and drug development professionals with the knowledge to select and implement appropriate structural strategies for their specific challenges.

RNA molecules serve a wide range of functions, including catalysis, ligand binding, and gene regulation, which are closely linked to their complex structures [28]. The analysis of RNA structures has progressed alongside advancements in structural biology techniques, but it comes with its own set of challenges and corresponding solutions. RNA structure is hierarchical: primary sequence folds into secondary structural elements (helices, hairpins, bulges, internal loops, junctions, and pseudoknots), which further interact to form three-dimensional (3D) architectures and, in some cases, quaternary complexes [28]. While smaller non-coding RNAs (such as miRNAs and siRNAs) often rely on primary sequences, larger RNAs, including ribozymes, riboswitches, and long non-coding RNAs (lncRNAs), adopt complex tertiary structures to perform their functions [28]. This review discusses recent advances in RNA structure analysis techniques, detailing their operational protocols, inherent limitations, and appropriate use cases.

Core Experimental Techniques: Principles and Protocols

RNA Structure Probing Methods

Probing methods provide insights into RNA secondary structure and dynamics by using enzymatic or chemical reagents that react differentially with single-stranded versus double-stranded nucleotides.

• Principle: Enzymatic or chemical probes cleave or modify RNA in a structure-specific manner. These sites are detected via reverse transcription, which records truncations or mutations in cDNA for sequencing analysis [28].
• Common Probes and Their Specificity:
  • Dimethyl Sulfate (DMS): A base-specific probe that primarily modifies unpaired Adenine (A) and Cytosine (C) residues [28] [29].
  • SHAPE Reagents (e.g., 1M7, NAI): Backbone-targeting probes that react with the 2'-hydroxyl group in flexible, unconstrained RNA regions [28].
  • Nuclease S1: An enzymatic probe that cleaves single-stranded regions (limited to in vitro applications) [28].
  • Nuclease V1: An enzymatic probe that cleaves double-stranded regions (limited to in vitro applications) [28].
• High-Throughput Advancements: The integration of probing with next-generation sequencing has enabled transcriptome-wide "structurome" studies, revealing RNA structures in vivo and across various biological contexts [28]. Methods like DMS-MaPseq allow for the analysis of heterogeneous conformational states within an RNA ensemble [28].

Protocol: SHAPE (Selective 2'-Hydroxyl Acylation analyzed by Primer Extension)

• RNA Preparation: Synthesize and purify the target RNA in vitro or extract RNA from cells for in vivo studies.
• Structure Refolding: Denature the RNA at 95°C for 2 minutes and then refold by incubating in the appropriate folding buffer (e.g., containing 50 mM HEPES pH 8.0, 100 mM KCl, 5 mM MgClâ‚‚) at 37°C for 20 minutes.
• Chemical Modification: Add the SHAPE reagent (e.g., 1M7 in DMSO to a final concentration of 5-10 mM) to the folded RNA. Incubate at 37°C for 5-10 minutes. Include a no-reagent control (DMSO only).
• Reaction Quenching: Precipitate the RNA to remove excess reagent.
• cDNA Synthesis (Primer Extension): Use fluorescently or radioactively labeled DNA primers complementary to the 3' end of the target RNA. Perform reverse transcription. The enzyme will truncate at sites of SHAPE modification.
• Fragment Analysis: Separate the cDNA fragments using capillary electrophoresis. The resulting trace shows peaks corresponding to reverse transcription stops; their intensity is proportional to the reactivity at each nucleotide.
• Data Normalization and Modeling: Normalize SHAPE reactivities and use them as pseudo-free energy constraints in computational secondary structure prediction algorithms (e.g., integrated into the RNAstructure package) [30] [28].

X-ray Crystallography (XRC)

X-ray crystallography has been a cornerstone technique, providing the first high-resolution 3D RNA structures, such as yeast tRNA^Phe [28].

• Principle: A crystal diffracts a beam of X-rays, producing a diffraction pattern. The electron density map calculated from this pattern allows for the atomic model building of the RNA.
• Challenges: RNA crystallization can be difficult due to molecular flexibility, negative charge, and structural homogeneity. Strategies include engineering stable constructs, using in crystallo transcription, and co-crystallization with binding partners to stabilize specific conformations.
• Output: Delivers atomic-resolution structures, crucial for understanding ligand binding and catalytic mechanisms.

Protocol: RNA Crystallography

• RNA Construct Design: Design RNA sequences with stable secondary structures (e.g., by incorporating a known structural motif like a GAAA tetraloop) to promote crystallization.
• Crystallization: Use vapor diffusion methods (sitting or hanging drop) with commercial sparse matrix screens. Crystallization conditions often include high concentrations of monovalent salts (e.g., Liâ‚‚SOâ‚„, NHâ‚„Cl) or polyamines (e.g., spermidine) to neutralize the RNA charge.
• Cryo-protection and Data Collection: Soak crystals in a cryo-protectant solution (e.g., 25% MPD or glycerol) before flash-freezing in liquid nitrogen. Collect X-ray diffraction data at a synchrotron source.
• Phasing and Model Building: Solve the "phase problem" using molecular replacement (with a known homologous structure) or experimental phasing (e.g., soaking crystals with halide ions or oligonucleotides containing anomalous scatterers like brominated bases). Iteratively build and refine the atomic model into the electron density map.

Nuclear Magnetic Resonance (NMR) Spectroscopy

NMR is powerful for studying the structure and dynamics of relatively small RNAs (< 40-50 nucleotides) in solution [28].

• Principle: NMR-active nuclei (e.g., ¹H, ¹³C, ¹⁵N) in a magnetic field absorb and re-emit electromagnetic radiation. The resulting spectrum provides information on through-bond (J-coupling) and through-space (Nuclear Overhauser Effect, NOE) interactions.
• Applications: Ideal for determining 3D structures of small RNA motifs, characterizing dynamics on timescales from picoseconds to seconds, and studying transient interactions and folding intermediates.
• Limitations: Lower throughput and limited to smaller RNAs compared to other methods.

Protocol: RNA Structure Determination by NMR

• Sample Preparation: Produce uniformly ¹³C, ¹⁵N-labeled RNA by in vitro transcription using labeled nucleotide triphosphates. The RNA sample (typically 0.2-1.0 mM) is dissolved in a suitable buffer, often with 10% Dâ‚‚O for lock signal.
• Data Collection: Acquire a suite of multi-dimensional NMR experiments (e.g., ¹H-¹H NOESY, ¹H-¹³C HSQC, HCN) to assign chemical shifts and obtain distance and dihedral angle restraints.
• Structure Calculation: Input the experimental restraints (NOE-derived distances, dihedral angles from J-couplings) into computational programs for structure calculation via simulated annealing (e.g., using XPLOR-NIH or CYANA). An ensemble of structures is generated that satisfies the experimental restraints.

Cryo-Electron Microscopy (cryo-EM)

Cryo-EM has revolutionized structural biology by enabling the determination of high-resolution structures for large, complex RNA-protein assemblies that are difficult to crystallize [28] [29].

• Principle: Purified macromolecules are frozen in a thin layer of vitreous ice, preserving their native state. An electron beam images thousands of individual particles, and computational methods classify 2D projections and reconstruct a 3D density map.
• Advantages: Requires small amounts of sample, tolerates heterogeneity, and is ideal for large complexes like ribosomes, spliceosomes, and viral RNA-protein complexes.
• Workflow: The process involves sample vitrification, automated data collection, particle picking, 2D classification, 3D reconstruction, and model building.

Small-Angle X-Ray Scattering (SAXS)

SAXS provides low-resolution structural information about RNA in solution, offering insights into overall shape, flexibility, and conformational changes [28].

• Principle: A solution of RNA scatters a collimated X-ray beam, and the scattering pattern at very low angles is recorded. This pattern is related to the pair-distance distribution function of the molecule, yielding information about its overall size (radius of gyration, R_g) and shape.
• Applications: Excellent for studying RNA folding, comparing conformational states under different conditions (e.g., with/without ligand or Mg²⁺), and validating structural models. It is often used in an integrated approach with other techniques.

Protocol: SAXS Data Collection and Analysis

• Sample and Buffer Matching: Purify the RNA to homogeneity. Dialyze the RNA sample into the desired buffer. Precisely match the buffer composition for the background measurement.
• Data Collection: Measure scattering from the RNA solution and the matched buffer blank at a synchrotron beamline. Use a range of RNA concentrations (e.g., 1-5 mg/mL) to check for interparticle interference and extrapolate to infinite dilution.
• Primary Data Analysis: Subtract the buffer scattering from the sample scattering. The Guinier analysis provides the R_g and an indication of sample quality (aggregation-free). The Kratky plot can be used to assess the degree of foldedness/flexibility.
• Ab Initio Modeling: Use programs like DAMMIF or GASBOR to generate low-resolution ab initio shape reconstructions that fit the experimental scattering curve.
• Validation and Integration: Compare the SAXS data with theoretical scattering profiles computed from high-resolution models (e.g., from crystallography or cryo-EM) to validate structures or propose conformational ensembles.

Table 1: Key Metrics for Comparing RNA 3D Structure Prediction Performance from RNA-Puzzles

Metric	Full Name	Description	Ideal Value
RMSD	Root Mean Square Deviation	Measures the average distance between equivalent atoms in superimposed structures; lower is better. [31]	< 5.0 Ã…
INF	Interaction Network Fidelity	Evaluates the accuracy of predicted base-pairing interactions (stacks, WC, non-WC). [31]	> 0.8
lDDT	local Distance Difference Test	Emphasizes local accuracy over global topology; higher is better. [31]	> 0.7
TM-score	Template Modeling Score	A scale-invariant measure for global structural similarity; higher is better. [31]	> 0.5
Clash Score	-	Measures the number of steric atomic clashes per 1000 atoms; lower is better. [31]	< 10

Single-Molecule Förster Resonance Energy Transfer (smFRET)

smFRET is a powerful technique for observing dynamic processes and conformational heterogeneity within individual RNA molecules in real-time.

• Principle: FRET efficiency is inversely proportional to the sixth power of the distance between a donor and an acceptor fluorophore attached to specific sites on the RNA. It reports on distances in the 2-8 nm range.
• Applications: Directly observes folding pathways, conformational dynamics, and transient intermediate states that are averaged out in ensemble measurements.
• Setup: Experiments are typically performed using total internal reflection fluorescence (TIRF) microscopy with surface-immobilized RNA molecules, or with confocal microscopy on freely diffusing molecules.

Protocol: smFRET to Study RNA Folding

• RNA Labeling: Chemically synthesize the RNA with specific donor (e.g., Cy3) and acceptor (e.g., Cy5) fluorophores attached to selected nucleotides (e.g., via amino-allyl modified bases).
• Surface Immobilization: Immobilize the labeled RNA on a passivated quartz slide (e.g., via a biotin-streptavidin-biotinylated DNA anchor system).
• Data Acquisition: Use a TIRF microscope to excite the donor fluorophore and simultaneously collect emission from both donor and acceptor channels with a sensitive camera (e.g., EMCCD). Record movies of thousands of individual molecules.
• Data Analysis: Identify single molecules and extract donor (I_D) and acceptor (I_A) intensities over time. Calculate FRET efficiency as E = I_A / (I_A + I_D). Analyze FRET trajectories to identify discrete states and transition rates.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Research Reagent Solutions for RNA Structural Biology

Reagent / Material	Function / Application
DMS (Dimethyl Sulfate)	Chemical probe for in vivo and in vitro mapping of unpaired A and C residues. [28]
1M7 (1-methyl-7-nitroisatoic anhydride)	A SHAPE reagent for probing RNA backbone flexibility and secondary structure. [28]
Psoralen & Derivatives	Crosslinking agent used in 2D probing methods (e.g., PARIS, SPLASH) to capture RNA-RNA interactions in vivo. [28]
RNAstructure Package	A comprehensive software suite for predicting RNA secondary structure, including facilities to incorporate SHAPE and DMS probing data as constraints. [30]
ViennaRNA Package	A standard suite of tools for RNA secondary structure prediction and analysis based on thermodynamic models. [32]
Crystallization Screens (e.g., JCSG+, Natrix)	Commercial sparse matrix screens containing a variety of precipitants, salts, and buffers to identify initial RNA crystallization conditions.
Isotope-Labeled NTPs (13C, 15N)	Essential for producing uniformly labeled RNA for NMR spectroscopy, enabling resonance assignment and structure determination.
Fluorophore-Labeled NTPs (e.g., Cy3-, Cy5-)	Used for incorporating donor and acceptor fluorophores into RNA via transcription for smFRET studies.
3,4,6-Trichlorocatechol	3,4,6-Trichlorocatechol
3,4,5-Trichlorocatechol	3,4,5-Trichlorocatechol, CAS:56961-20-7, MF:C6H3Cl3O2, MW:213.4 g/mol

Integrated Workflow Visualization

The following diagram illustrates a generalized, integrated workflow for determining RNA structure, combining multiple techniques discussed in this guide.

The experimental arsenal for RNA structure determination is powerful and diverse. No single technique can fully capture the structural complexity and dynamic nature of RNA molecules. The future of the field lies in integrative structural biology, which combines data from multiple methods—such as chemical probing, cryo-EM, X-ray crystallography, NMR, SAXS, and smFRET—to build comprehensive and accurate models of RNA architecture and conformational dynamics [28] [29]. This approach is particularly vital for studying large, flexible lncRNAs and for developing RNA-targeted therapeutics, where understanding structure-function relationships is paramount. As technologies like cryo-EM and machine learning continue to advance, they promise to further redefine our understanding of RNA structural landscapes under near-physiological conditions [29].

Ribonucleic acid (RNA) molecules are pivotal players in the central dogma of molecular biology, fulfilling essential roles in transcription, translation, catalysis, and gene expression regulation [33] [34]. The biological functions of RNA are profoundly determined by their three-dimensional (3D) structures, which in turn are dictated by their primary sequences and secondary structure interactions [33]. However, experimental determination of RNA structures through techniques like X-ray crystallography, NMR, or cryo-electron microscopy remains low-throughput and challenging due to the inherent conformational flexibility of RNA molecules [33]. As of December 2023, RNA-only structures constitute less than 1.0% of the approximately 214,000 entries in the Protein Data Bank (PDB) [33]. This structural gap has motivated the development of computational methods to predict RNA structure from sequence, a challenge that has recently been transformed by deep learning approaches.

The Computational Challenge of RNA Structure

Historical Context and Traditional Methods

Traditional computational approaches for RNA structure prediction have primarily fallen into two categories: template-based modeling and de novo prediction. Template-based methods, such as ModeRNA and RNAbuilder, rely on known structural templates from libraries but are constrained by their limited coverage [33]. De novo approaches, including FARFAR2, 3dRNA, and SimRNA, utilize thermodynamic or statistical energy functions to sample the conformational space and identify low-energy states, but this process is often computationally intensive [33].

A particular challenge in RNA structural bioinformatics has been the accurate prediction of pseudoknots—complex structural motifs where bases in a loop pair with complementary sequences outside the loop [34]. These motifs are biologically significant but computationally NP-hard to predict using traditional thermodynamic models, leading many algorithms to either exclude them or employ heuristic compromises [34].

The Data Scarcity Problem

The development of data-driven methods for RNA structure prediction has been hampered by the severe scarcity of experimentally determined structures. This scarcity presents a fundamental challenge for deep learning approaches that typically require large training datasets. With only ~5,500 RNA chains available in representative datasets (clustered at 80% sequence identity), researchers have needed innovative strategies to overcome this limitation [33].

Deep Learning Architectures for RNA Structure Prediction

Language Models for Evolutionary Insight

A transformative approach in recent RNA structure prediction methods has been the adaptation of language models pretrained on massive sequence databases. RhoFold+, a leading method, integrates an RNA language model (RNA-FM) pretrained on approximately 23.7 million RNA sequences to extract evolutionarily informed embeddings [33]. This strategy effectively leverages the statistical patterns learned from diverse RNA sequences to infer structural constraints without relying exclusively on experimentally determined structures.

These language models operate on the principle that evolutionary conservation captured in multiple sequence alignments (MSAs) contains implicit structural information. While earlier MSA-based methods like DeepFoldRNA and trRosettaRNA required computationally expensive database searches, language model approaches provide a more efficient alternative by encoding evolutionary information directly from pretrained representations [33].

End-to-End Differentiable Pipelines

Modern deep learning frameworks for RNA structure prediction increasingly employ fully differentiable architectures that directly map sequence information to 3D coordinates. RhoFold+ exemplifies this approach with its Rhoformer transformer network and invariant point attention (IPA) module, which iteratively refines structural features over multiple cycles [33]. The system integrates sequence embeddings, MSA features, and predicted secondary structures, then employs a geometry-aware structure module to optimize backbone coordinates and torsion angles [33].

Table 1: Key Deep Learning Approaches for RNA Structure Prediction

Method	Architecture	Key Innovations	Structural Output
RhoFold+	Language model + transformer	RNA-FM embeddings, invariant point attention	Full-atom 3D coordinates
KnotFold	Attention network + minimum-cost flow	Learned potentials, pseudoknot-aware algorithm	Secondary structure including pseudoknots
SCOPER	IonNet + conformational sampling	Mg²⁺ ion binding prediction, solution validation	Solution-state structures with ions
AlphaFold3	Diffusion-based	Joint biomolecular structure prediction	RNA-protein complexes

Pseudoknot-Aware Secondary Structure Prediction

KnotFold represents a significant advancement in secondary structure prediction through its novel integration of deep learning with combinatorial optimization. The method uses an attention-based neural network to predict base pairing probabilities, capturing long-range interactions and non-nested base pairs through a self-attention mechanism [34]. These probabilities are then transformed into a structural potential function, and the optimal structure is identified by solving a minimum-cost flow problem—a graph-theoretic approach that efficiently handles pseudoknots without heuristic restrictions [34].

The KnotFold potential function is defined as:

[ E(S,x) = -\sum{\substack{i < j \ S{i,j}=1}} \log \frac{P(bp{i,j}|x)}{P(bp{i,j}|\text{length})} - \sum{\substack{i < j \ S{i,j}=0}} \log \frac{1-P(bp{i,j}|x)}{1-P(bp{i,j}|\text{length})} + \lambda \sum{i < j} S{ij} ]

This formulation combines the learned pairing probabilities with a reference distribution and a sparsity constraint to identify biologically plausible structures [34].

Experimental Validation and Benchmarking

Performance on Community-Wide Challenges

Rigorous benchmarking on standardized datasets has demonstrated the superior performance of deep learning methods over traditional approaches. In retrospective evaluations on RNA-Puzzles—a community-wide blind assessment—RhoFold+ achieved an average RMSD of 4.02 Å, significantly outperforming the second-best method (FARFAR2 at 6.32 Å) [33]. Notably, RhoFold+ produced predictions with RMSD values below 5 Å for 17 of 24 targets, a level of accuracy that approaches experimental resolution for many biological applications [33].

Table 2: Quantitative Performance Comparison on RNA-Puzzles Benchmark

Method	Average RMSD (Ã…)	Average TM-Score	Targets with RMSD <5Ã…
RhoFold+	4.02	0.57	17/24
FARFAR2 (top 1%)	6.32	0.44	~40%
Best Template	-	0.48	-
Other Methods	7.50-15.20	0.31-0.41	<25%

Generalization and Cross-Family Validation

A critical test for deep learning methods is their ability to generalize to sequences with limited similarity to training examples. Analysis of RhoFold+ performance demonstrated no significant correlation between sequence similarity to training data and prediction accuracy (R² = 0.23 for TM-score, 0.11 for lDDT), indicating robust generalization capabilities [33]. For example, on target PZ7 (a 186-nucleotide Varkud satellite ribozyme), RhoFold+ achieved accurate predictions despite the most similar training structure having an RMSD of 34.48 Å to the native fold [33].

Solution-State Validation with Experimental Data

The SCOPER pipeline addresses the critical challenge of validating computational predictions against experimental solution-state data. By integrating kinematics-based conformational sampling with IonNet—a deep learning model for predicting Mg²⁺ ion binding sites—SCOPER significantly improves the fit between predicted structures and small-angle X-ray scattering (SAXS) profiles [35]. This approach acknowledges that RNA molecules exhibit conformational plasticity in solution and that ion interactions are essential for accurate structural modeling [35].

Research Protocols and Methodologies

End-to-End 3D Structure Prediction Protocol

The RhoFold+ pipeline provides a fully automated workflow for RNA 3D structure prediction:

• Input Processing: Input RNA sequence is processed through the RNA-FM language model to generate evolutionarily informed embeddings [33].

• Feature Integration: Sequence embeddings are combined with MSA features and predicted secondary structures [33].

• Transformer Processing: The Rhoformer transformer network iteratively refines structural features over ten cycles [33].

• Coordinate Generation: A geometry-aware structure module with invariant point attention generates atomic coordinates for the RNA backbone [33].

• Full-Atom Reconstruction: Complete all-atom models are reconstructed with applied structural constraints including base pairing and stacking interactions [33].

• Output: The final output includes atomic coordinates in PDB format, along with confidence estimates for different regions of the structure [33].

Workflow for End-to-End RNA 3D Structure Prediction

Pseudoknot-Inclusive Secondary Structure Prediction

The KnotFold methodology enables accurate prediction of pseudoknotted structures through a multi-stage process:

• Sequence Encoding: The input RNA sequence is processed through transformer encoder layers to generate contextual representations for each nucleotide [34].

• Base Pair Probability Calculation: An attention mechanism computes pairing probabilities between all possible base pairs through outer product operations on the sequence encodings [34].

• Potential Function Construction: The base pairing probabilities are converted into a structural potential function that includes reference distributions and sparsity constraints [34].

• Network Flow Optimization: A minimum-cost flow algorithm is applied to a bipartite graph representation of the RNA to identify the optimal structure including pseudoknots [34].

• Structure Output: The final secondary structure is returned in dot-bracket notation or other standard formats with confidence estimates [34].

Pseudoknot-Aware Secondary Structure Prediction Workflow

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Computational Tools for RNA Structure Prediction

Tool/Resource	Type	Function	Access
RhoFold+	End-to-end deep learning platform	Predicts 3D structures from sequence	Research implementation
KnotFold	Secondary structure predictor	Predicts pseudoknotted structures	Research implementation
RNA-FM	Language model	Generates evolutionary embeddings	Publicly available
SCOPER	Validation pipeline	Integrates SAXS data for solution-state validation	Research implementation
R2DT	Visualization tool	Standardized RNA secondary structure visualization	Web server
Forna	Visualization tool	Force-directed layout for RNA structures	Web server
RNAfold	Secondary structure prediction	Classical MFE-based structure prediction	Web server
Rubranol	Rubranol, CAS:211126-61-3, MF:C19H24O5, MW:332.4 g/mol	Chemical Reagent	Bench Chemicals
Methyl Syringate	Methyl Syringate, CAS:884-35-5, MF:C10H12O5, MW:212.20 g/mol	Chemical Reagent	Bench Chemicals

Biological Applications and Implications

Functional RNA Characterization

The ability to accurately predict RNA structures computationally enables researchers to characterize the vast universe of non-coding RNAs—which constitute the majority of transcribed RNAs in the human genome [33]. Structure predictions provide critical insights into functional mechanisms, including ribozyme catalysis, riboswitch ligand binding, and non-coding RNA molecular interactions [33] [34].

RNA-Targeted Drug Development

RNA structures represent promising but underexplored therapeutic targets. Deep learning methods that rapidly generate accurate structural models can significantly accelerate the identification and validation of RNA-targeted small molecules [33]. This capability is particularly valuable for targeting structured RNA elements in viral pathogens or disease-associated non-coding RNAs.

Synthetic Biology Design

In synthetic biology, computational design of RNA components with specific structural properties enables the programming of genetic circuits and regulatory systems [33]. Accurate prediction methods reduce the design-build-test cycle time for developing novel RNA-based sensors, regulators, and therapeutic devices.

Future Directions and Challenges

Despite remarkable progress, several challenges remain in the deep learning revolution for RNA structure prediction. Methods must still improve for modeling large, complex RNAs and RNA-protein complexes. The flexibility and conformational dynamics of RNA molecules present modeling challenges that go beyond single static structures [35]. Additionally, the integration of experimental data—such as chemical probing, cross-linking, and cryo-EM maps—with deep learning approaches represents a promising avenue for improving accuracy and biological relevance [35].

The field is also advancing toward more efficient methods that reduce computational requirements while maintaining accuracy, making sophisticated structure prediction accessible to non-specialists. As these tools become more robust and user-friendly, they will undoubtedly transform our understanding of RNA biology and open new frontiers in therapeutic development.

Computational microscopy has emerged as an indispensable tool for elucidating the structure and dynamics of RNA molecules, which play crucial roles in regulating biological processes from gene expression to cellular differentiation. Unlike proteins, RNA molecules exhibit high structural flexibility and complex conformational dynamics that are intimately connected to their biological functions. This technical guide explores how atomistic molecular dynamics (MD) and enhanced sampling simulations provide unprecedented insights into RNA behavior, enabling researchers to characterize conformational transitions, ligand binding mechanisms, and folding pathways that occur across multiple time scales. With applications in RNA-targeted therapeutic development and synthetic biology, these computational approaches offer a powerful complement to experimental methods by providing atomic-level resolution and access to transient states difficult to capture experimentally. We present here a comprehensive framework for implementing these techniques, including validated protocols, data presentation standards, and visualization approaches tailored for research scientists and drug development professionals.

RNA molecules adopt complex three-dimensional structures through a hierarchical folding process that begins with secondary structure formation through canonical Watson-Crick and wobble base pairing, followed by tertiary structure formation through interactions between distinct secondary structural elements [36]. The biological functions of RNAs are closely linked to their structures and dynamics, with conformational flexibility playing a particularly important role in functional RNAs such as riboswitches, ribozymes, and RNA-protein interactions [36].

Computational microscopy encompasses a suite of simulation techniques that bridge the gap between static structural snapshots and dynamic functional processes. For RNA molecules, this approach is particularly valuable due to several inherent challenges:

• Timescale Diversity: RNA dynamics span a wide range of time scales, from microseconds to seconds, reflecting the rugged free energy landscape of RNA folding, unfolding, and interactions [36].
• Ion Dependence: RNA stability and conformational plasticity are strongly influenced by the presence of cations, particularly Mg²⁺, which are essential for charge neutralization and structural integrity [35].
• Conformational Heterogeneity: RNA molecules frequently exist as ensembles of structures rather than single conformations, adopting different states in response to environmental conditions, ligand interactions, or cellular context [37].

The integration of computational microscopy with experimental validation techniques such as small-angle X-ray scattering (SAXS) has created powerful hybrid approaches for characterizing RNA solution behavior, enabling researchers to move beyond static structural models toward dynamic ensemble representations [35].

Computational Foundations

Force Fields and Solvation Models

Accurate force fields are fundamental to reliable RNA simulations. The AMBER RNA force field, particularly the bsc0χOL3 version (often called AMBER-OL3), has been widely used for MD simulations of atomistic RNA systems [36]. Recent years have seen significant refinements to improve accuracy:

• DESRES-RNA Force Field: An extensive revision of electrostatic, van der Waals, and torsional parameters based on quantum mechanical calculations and experimental information to better reproduce nucleobase stacking, base pairing, and key torsional conformers [36].
• gHBfix: Structure-specific local correction potential that selectively modifies native hydrogen bonds of RNA to improve force field performance while minimizing undesired side effects [36].
• Modified Parameters: Adjustments to phosphate Lennard-Jones parameters, backbone and glycosidic torsion parameters, and base-base/water interactions have all contributed to more accurate RNA simulations [36].

Solvation models play an equally critical role in simulation accuracy and efficiency:

• Explicit Solvent Models: Provide the most realistic representation but require substantial computational resources.
• Implicit Solvent Models: Approximate solvent as a continuum, dramatically reducing computational cost while accelerating conformational sampling due to lower solvent viscosity [36].
• GB-neck2 Model: An optimized generalized Born model that successfully reproduces Poisson-Boltzmann solvation energies and has enabled folding of small DNA and RNA hairpins to near-native structures [36].

Enhanced Sampling Techniques

Conventional MD simulations face limitations in exploring rare events or overcoming high free energy barriers. Enhanced sampling methods address these challenges:

Table 1: Enhanced Sampling Methods for RNA Simulations

Method	Key Principle	RNA Applications	Key Advantages
Replica-Exchange MD (REMD)	Multiple copies at different temperatures exchange configurations	Tetraloop folding, structural ensembles	Improved sampling of conformational space
Metadynamics	History-dependent bias potential added to overcome barriers	Ligand binding, conformational transitions	Efficient exploration of free energy landscapes
OPES (On-the-fly Probability Enhanced Sampling)	Applied biasing force with funnel-like restraint	Ligand binding to aptamers	Calculates free energy landscapes and kinetics
Simulated Tempering	Temperature variations within a single simulation	Reversible folding of tetraloops	Enhanced sampling without multiple replicas

These techniques have enabled researchers to access beyond-millisecond timescale RNA conformational transitions coupled to ligand binding, which is critical for rational design of RNA-targeted drugs [38].

Core Methodologies and Protocols

Enhanced Sampling for Ligand Binding

The following workflow illustrates the application of enhanced sampling methods to study RNA-ligand binding mechanisms:

Detailed Protocol for RNA-Ligand Binding Studies [38]:

• System Preparation:

  • Obtain initial RNA coordinates from experimental structures (NMR, X-ray crystallography) or prediction tools
  • Parameterize ligand molecules using appropriate force fields (GAFF, CGenFF)
  • Solvate the system in explicit water (TIP3P, TIP4P-D) or prepare for implicit solvent simulation
  • Add ions (Na+, K+, Mg2+) to neutralize charge and mimic physiological conditions

• OPES Simulation Setup:

  • Identify collective variables (CVs) that describe binding process (distance, angles, root mean square deviation)
  • Apply funnel-like restraint to prevent ligand dissociation while allowing natural binding/unbinding events
  • Implement multiple walker approach to enhance sampling efficiency
  • Set up OPES-flooding parameters to calculate residence times

• Binding Mechanism Analysis:

  • Monitor coupling between nucleotide flipping transitions and ligand binding coordinates
  • Distinguish between conformational selection and induced fit mechanisms
  • Calculate relative binding affinities and residence times for different ligands
  • Identify key interaction patterns and hydrogen bonding networks

This approach has successfully captured both conformational selection and induced fit mechanisms for theophylline RNA aptamer binding, demonstrating that mechanism preference can be determined by binding affinity [38].

Integrative Validation with SAXS

The SCOPER (Solution Conformation Predictor for RNA) pipeline integrates computational predictions with experimental validation:

Protocol for SAXS-Integrated Validation [35]:

• Initial Structure Preparation:

  • Start with sufficiently accurate initial structure (from prediction or experiment)
  • Use IonNet deep learning model to predict Mg²⁺ ion binding sites
  • Incorporate monovalent ions (Na+, K+) for charge neutralization

• Conformational Sampling:

  • Employ kinematics-based conformational sampling to explore flexibility
  • Adjust sampling intensity based on ion content (higher ion density decreases plasticity)
  • Generate ensemble of structures representing solution conformational space

• SAXS Profile Comparison:

  • Calculate theoretical SAXS profiles from structural ensembles
  • Compare with experimental SAXS data using appropriate fitting metrics
  • Avoid overfitting by carefully adjusting plasticity and ion density parameters
  • Iteratively refine ensemble composition to improve fit quality

Benchmarking against 14 experimental datasets has demonstrated that SCOPER significantly improves the quality of SAXS profile fits by including Mg²⁺ ions and sampling conformational plasticity [35].

De Novo Folding Simulations

Protocol for RNA Stem-Loop Folding [36]:

• System Setup:

  • Start from extended RNA conformations (unfolded state)
  • Select appropriate force field (DESRES-RNA with GB-neck2 recommended)
  • Configure implicit or explicit solvent based on computational resources

• Simulation Parameters:

  • Set temperature near predicted melting temperature for natural folding/unfolding
  • Use 2-fs integration time step with constraints on hydrogen-heavy atom bonds
  • Employ particle mesh Ewald for electrostatic calculations (explicit solvent)
  • Run simulations for sufficient duration to observe multiple folding events

• Analysis Metrics:

  • Calculate root mean square deviation (RMSD) relative to experimental structures
  • Monitor native base pair formation using hydrogen bonding criteria
  • Analyze clustering of folded structures to identify dominant conformations
  • Compare loop region accuracy versus stem region accuracy

This protocol has successfully folded 23 of 26 RNA stem-loops (10-36 nucleotides) into structures with native base pairs and RMSD values <2 Ã… for stem regions and <4 Ã… for loop regions [36].

Data Presentation Standards

Quantitative Comparison of Simulation Methods

Table 2: Performance Metrics for RNA Simulation Approaches

Method	System Size	Time Scale	Accuracy (RMSD)	Key Applications
Conventional MD	10-36 nt stem-loops	Nanoseconds to microseconds	<2 Ã… (stems), <5 Ã… (overall)	Stem-loop folding, local dynamics
REMD	8-14 nt tetraloops	Enhanced sampling equivalent to µs-ms	1-3 Å from experimental structures	Structural ensembles, folding pathways
OPES	RNA aptamers with ligands	Efficient sampling of binding events	Reproduces experimental affinity trends	Ligand binding mechanisms, kinetics
GB-neck2 Implicit	16 proteins, RNA hairpins	Faster sampling due to reduced viscosity	Successful folding to native structures	De novo folding, large conformational changes

RNA Motif Complexity Statistics

Table 3: Dataset Characteristics for RNA Structure Prediction [37]

Motif Type	Prevalence in RNAsolo	Average Length	Prevalence in Rfam
Internal Loops	82.4%	67 nucleotides	85.29%
3-way Junctions	9.49%	143 nucleotides	9.18%
4-way Junctions	6.38%	154 nucleotides	3.99%
8-10 way Junctions	Single instances	Several thousand nucleotides	Rare

The Scientist's Toolkit

Table 4: Essential Research Reagents and Computational Tools

Tool/Resource	Type	Function	Application Context
SCOPER	Software pipeline	Integrates conformational sampling with ion binding prediction	SAXS validation of RNA solution structures
IonNet	Deep learning model	Predicts Mg²⁺ ion binding sites	RNA structure stabilization in simulations
OPES	Enhanced sampling algorithm	Overcomes free energy barriers with biasing force	Ligand binding kinetics and mechanisms
DESRES-RNA Force Field	Molecular mechanics parameters	Describes atomic interactions in RNA	Accurate MD simulations of folding and dynamics
GB-neck2	Implicit solvent model	Approximates solvent as continuum	Accelerated sampling of conformational changes
CaCoFold-R3D	Probabilistic grammar	Predicts 3D motifs jointly with secondary structure	RNA structure prediction from sequence
Comprehensive RNA Dataset	Benchmark data	320k+ instances from validated sources	Training and testing machine learning models
Hirsutanonol	Hirsutanonol, CAS:41137-86-4, MF:C19H22O6, MW:346.4 g/mol	Chemical Reagent	Bench Chemicals

Advanced Integrative Approaches

Cotranscriptional RNA Folding

RNA folding during transcription differs fundamentally from thermodynamic folding of full-length sequences. While thermodynamic folding reaches an equilibrium ensemble of structures, cotranscriptional folding is a kinetic process where the RNA structure evolves as the chain elongates during transcription [39]. This dynamic folding pathway causes cotranscriptional structures to often deviate from thermodynamic predictions, as the system rarely reaches equilibrium. Since these effects can persist in the mature RNA's structure, understanding this kinetic process is crucial for predicting functional structures.

Computational modeling has emerged as an increasingly practical approach for investigating these dynamics, complementing resource-intensive experimental studies. Methods include:

• Kinetic Folding Algorithms: Simulate the sequential addition of nucleotides with folding at each step
• Markov State Models: Identify metastable intermediate states during transcription
• Master Equation Approaches: Model population dynamics of different structural isoforms

Machine Learning and Data-Driven Approaches

The shortage of experimentally determined high-resolution RNA 3D structures has historically limited the application of machine learning to RNA modeling [37]. However, recent initiatives have created comprehensive datasets of over 320 thousand instances from experimentally validated sources to establish new community-wide benchmarks for RNA design and modeling algorithms [37].

Promising machine learning approaches include:

• Geometric Deep Learning: Frameworks like gRNAde that handle single-state and multi-state fixed-backbone sequence design
• Diffusion Models: RiboDiffusion and RIdiffusion that learn RNA sequence distribution conditioned on fixed 3D backbone structures
• Equivariant Graph Neural Networks: Capture hierarchical structural variations efficiently, particularly in low-data settings

These approaches demonstrate promising results in capturing the geometric and topological complexities of RNA tertiary structures, though they still face limitations due to data scarcity and RNA's inherent structural flexibility [37].

Computational microscopy through atomistic and enhanced sampling simulations has transformed our ability to study RNA structure and dynamics at unprecedented spatial and temporal resolution. The integration of physical force fields with enhanced sampling algorithms, machine learning approaches, and experimental validation creates a powerful framework for investigating RNA function and designing RNA-targeted therapeutics.

Key advances include:

• Accurate prediction of ion binding sites that stabilize RNA structures
• Characterization of ligand binding mechanisms and kinetics
• Integration of computational predictions with experimental validation techniques
• Development of comprehensive datasets and benchmarks for method evaluation

As these methods continue to evolve, they will play an increasingly important role in unlocking the therapeutic potential of RNA molecules and understanding their diverse functions in cellular processes. The protocols and standards presented here provide researchers with a solid foundation for implementing these cutting-edge approaches in their own investigations of RNA structure and dynamics.

Ribonucleic acid (RNA) function is profoundly intertwined with its conformational dynamics and structural hierarchy, classified as primary sequence, secondary helices, and tertiary three-dimensional arrangement [40]. The modern interpretation of RNA moves beyond the primary structure as the sole information carrier, recognizing that secondary and tertiary structures are critically connected to RNA function, especially in interactions with ions, proteins, or other molecules [40]. However, RNA molecules are not static; they are highly dynamic, often exhibiting multiple conformations in equilibrium, a characteristic distinct from kinetics, which concerns the transition times between states [40]. This inherent dynamism presents a significant challenge: individuating these multiple concurrent structures through experimental methods alone is difficult.

Atomistic molecular dynamics (MD) simulations act as computational microscopes, providing detailed atomistic pictures of conformational ensembles [40]. Yet, these simulations can suffer from both precision and accuracy issues. Precision—the capability to produce consistent results independent of initial conditions—is limited by the accessible time scales and can be improved with enhanced sampling methods. Accuracy—the capability to reproduce experimental data—can be hampered by imperfect force fields [40]. No single experimental or computational method can fully characterize the complex structural landscape of RNA. Integrative approaches, which combine data from multiple scales and sources, have therefore emerged as a powerful strategy to overcome the limitations of individual techniques, creating robust models that more accurately predict RNA structure, dynamics, and function. This guide details the methodologies and protocols for implementing these integrative approaches, framed within the context of RNA structure and dynamics research for drug development.

Methodological Foundations: Data Types and Computational Frameworks

Integrative modeling leverages diverse experimental and computational data to refine and validate structural models.

• Experimental Data from Sequencing and Probing: RNA-sequencing (RNA-seq) data, whether from whole transcriptome (WTS) or 3' mRNA-Seq protocols, provides a foundation for understanding transcript-level information and expression quantification [41] [42]. Techniques like crosslinking and deep sequencing of RNA-RNA hybrids (RRI) experimentally map long-range interactions in genomes, such as SARS-CoV-2, revealing dynamic topologies that play regulatory roles [43]. Chemical probing data, such as SHAPE reactivity, provides experimental constraints on RNA flexibility and secondary structure.
• Computational Simulations and Predictions: Atomistic molecular dynamics (MD) simulations explicitly model RNA systems at atomic resolution using physics-based force fields, characterizing conformational sampling and interactions [40]. Complementing these, bioinformatics tools predict stable RNA-RNA interactions (RRI) using strategies ranging from dynamic programming and minimum-free energy (MFE) methods to more sophisticated accessibility-based approaches that consider intramolecular folding [43].

Core Computational Integration Strategies

Two primary philosophies exist for integrating data with simulations, each with distinct advantages.

• Ensemble Refinement Methods: In this approach, conformational ensembles generated by MD simulations are subsequently refined or corrected to enforce agreement with experimental data. This can be done post-simulation or on-the-fly, aiming to maximize the overlap between the sampled conformations and experimental observables [40]. This method is particularly useful when the experimental data is sparse or when the goal is to characterize a dynamic equilibrium between multiple states.
• Force-Field Fine-Tuning: An alternative strategy is to directly adjust the parameters of the empirical force fields used in MD simulations to better reproduce experimental observables [40]. This provides a more fundamental correction to the underlying energy model but requires careful validation to avoid overfitting to a specific system or dataset. The choice between these approaches depends on the research question, the nature and quantity of available experimental data, and the desired generality of the resulting model.

Experimental Protocols for Key Integrative Techniques

Protocol: Mapping RNA-RNA Interactions (RRI) for Integrative Modeling

Long-range RNA-RNA interactions are crucial for regulatory functions, particularly in viral RNAs like SARS-CoV-2. The following protocol outlines an experimental method for capturing these interactions, generating data that can be integrated with computational predictions [43].

• Cell Crosslinking: Treat cells with a chemical crosslinker (e.g., formaldehyde or a more specific RNA-RNA crosslinker) to covalently link spatially proximate RNA strands.
• RNA Extraction and Fragmentation: Lyse the cells and extract total RNA. Partially fragment the RNA using controlled RNase digestion.
• Hybrid Capture and Proximity Ligation: Under conditions that favor RNA-RNA duplexes, allow crosslinked fragments to form hybrids. Use a proximity ligation enzyme to join the ends of interacting RNA fragments, creating chimeric molecules.
• Library Preparation and Sequencing: Reverse transcribe the ligated products into cDNA. Prepare sequencing libraries following a standard strand-specific RNA-seq protocol [42]. Sequence the libraries on a high-throughput platform (e.g., Illumina NextSeq 500) to generate short-read data.
• Computational Analysis of RRI Data: Process the raw sequencing reads to identify chimeric sequences, which represent the ligated RNA-RNA interaction pairs. Map these reads back to the reference genome to identify the genomic locations of the interacting regions.

Protocol: An Integrative Workflow for RNA Ensemble Refinement

This protocol describes a general workflow for using experimental data to refine a structural ensemble generated from MD simulations [40].

• Initial Ensemble Generation: Perform multiple independent molecular dynamics simulations of the RNA system of interest. If necessary, employ enhanced sampling techniques to overcome significant free-energy barriers and improve the diversity of the sampled conformational space.
• Experimental Data Collection: Obtain experimental data that reports on RNA structure and dynamics. This can include WAXS data [40], NMR observables, or chemical probing data like SHAPE reactivity.
• Data Integration and Ensemble Reweighting: Use an ensemble refinement software package to reweight the conformations from the initial MD ensemble. The algorithm assigns new statistical weights to each structure to maximize the agreement between the ensemble-averaged theoretical observables and the experimental data while minimizing the deviation from the original MD distribution (often using a maximum entropy or Bayesian approach).
• Model Validation: Validate the final refined ensemble against a hold-out set of experimental data that was not used in the reweighting procedure. This ensures the model is not overfit and possesses predictive power.

The Scientist's Toolkit: Essential Reagents and Computational Tools

Table 1: Research Reagent Solutions for Integrative RNA Studies

Item Name	Function/Application	Example Use Case
NEBNext Poly(A) mRNA Magnetic Isolation Kit	Enriches for polyadenylated mRNA from total RNA by selecting molecules with poly(A) tails.	Preparation of RNA-seq libraries for studying coding transcripts, a key data source for integrative models [42].
TruSeq Stranded Total RNA Library Prep Kit	Prepares sequencing libraries from total RNA, typically incorporating ribosomal RNA depletion.	Whole transcriptome sequencing for a global view of coding and non-coding RNA species [44].
PicoPure RNA Isolation Kit	Isolves high-quality RNA from small cell numbers or sorted cell populations.	RNA extraction from fluorescence-activated cell sorted (FACS) alveolar macrophages for downstream sequencing [42].
Crosslinking Reagents (e.g., formaldehyde)	Covalently links interacting biomolecules in situ to capture transient interactions.	Stabilizing RNA-RNA interactions for crosslinking-based RRI mapping protocols [43].

Table 2: Computational Tools for RNA-RNA Interaction (RRI) Prediction

Tool	Core Strategy	Key Features	Accessibility-Based
IntaRNA	Minimum-free energy (MFE) with accessibility	Predicts intermolecular and intramolecular interactions; considers interaction accessibility using partition function.	Yes [43]
RNAup	MFE with accessibility	Calculates the energy required to make a binding site accessible for interaction.	Yes [43]
RNAcofold	Concatenation-based MFE	Combines two sequences and predicts a single secondary structure; only considers intramolecular pairs.	No [43]
RactIP	Integer programming	Predicts interactions with minimal restrictions and without concatenation; allows for longer inputs.	No [43]

Quantitative Data and Analysis in Integrative Studies

Table 3: Performance Comparison of RNA-seq Methodologies for Generating Expression Data

Metric	Whole Transcriptome (WTS)	3' mRNA-Seq	Notes and Implications
Read Localization	Distributed across entire transcript [41]	Localized to the 3' end [41]	WTS is required for isoform-level analysis; 3' mRNA-Seq simplifies quantification.
Detection of Differentially Expressed Genes (DEGs)	Detects more DEGs [41]	Detects fewer DEGs, but captures major signals [41]	WTS provides higher sensitivity; 3' mRNA-Seq is sufficient for key pathway analysis.
Required Read Depth	High (e.g., >30 million reads) [41]	Low (1-5 million reads) [41]	3' mRNA-Seq enables higher throughput and more cost-effective screening.
Performance on Degraded Samples	Poor if poly(A) tail is absent/degraded	Robust for FFPE and degraded samples [41]	3' mRNA-Seq is preferable for clinical or archival sample types.

Workflow Visualization and Logical Pathways

The following diagram illustrates the core logical workflow of an integrative approach, combining multi-scale data to arrive at a robust, validated model of RNA structure and dynamics.

The second diagram details the specific data flow and decision points within the computational integration process, highlighting the two primary strategies of ensemble refinement and force-field tuning.

Ribonucleic acid (RNA) has emerged as a compelling therapeutic target for treating a wide range of diseases, from genetic disorders and cancers to viral infections and neurodegenerative conditions. The rationale for targeting RNA extends beyond traditional protein-focused approaches, offering access to previously "undruggable" pathways and expanding the targetable space within the human genome. While only approximately 1.5% of the human genome encodes proteins, the majority is transcribed into non-coding RNAs that perform essential regulatory functions, presenting a vast landscape for therapeutic intervention [14]. RNA-targeted small molecules represent a promising class of therapeutics that combine the versatility of small molecule drugs with the specificity of genetic medicines, offering potential for oral bioavailability and blood-brain barrier penetration that remains challenging for oligonucleotide-based therapies [45].

The field has evolved significantly from early RNA-targeting antibiotics like aminoglycosides to fully synthetic compounds such as linezolid, and more recently to the first non-ribosomal RNA-targeting drug, risdiplam, approved in 2020 for spinal muscular atrophy [46]. This progression demonstrates the growing sophistication in targeting diverse RNA structures and functions. However, designing small molecules that selectively engage RNA targets presents unique challenges, including RNA's structural flexibility, highly electronegative surface, and the critical influence of metal ions and solvation on its structure [45]. This technical guide examines the complete workflow from RNA structural analysis to small molecule optimization, providing researchers with methodologies and frameworks to advance RNA-targeted therapeutic discovery.

RNA Structure Determination and Analysis

RNA Structural Hierarchy and Fundamentals

RNA molecules exhibit a hierarchical organization wherein primary sequences fold into local secondary structures (helices, loops, bulges) that subsequently assemble into complex tertiary architectures through long-range interactions. This structural complexity creates specific binding pockets and clefts that can be targeted by small molecules [45]. Understanding RNA structure is foundational to rational drug design, as the specific three-dimensional arrangement of nucleotides defines potential binding sites and determines functional mechanisms.

RNA secondary structures form through Watson-Crick base pairing (A-U, G-C) and non-canonical pairs (e.g., G-U wobble), creating elements such as stem-loops, internal loops, bulges, and junctions. These secondary elements then fold into tertiary structures stabilized by various interactions, including coaxial stacking, ribose zippers, and A-minor motifs. The folding process is hierarchical, with secondary structures forming more rapidly than tertiary conformations [47]. This structural organization creates unique binding pockets that differ significantly from protein binding sites, often featuring highly electronegative surfaces and specific metal ion coordination requirements [45].

Experimental Methods for RNA Structure Determination

Biophysical and Structural Biology Techniques

Advanced biophysical methods provide high-resolution structural information essential for rational drug design. X-ray crystallography remains the gold standard for determining atomic-resolution RNA structures, though it faces challenges in crystal formation and may not fully capture dynamic conformations [47]. Nuclear magnetic resonance (NMR) spectroscopy excels at studying smaller RNA molecules and their dynamics, with specialized approaches like selective labeling and "divide-and-conquer" strategies helping overcome size limitations [47].

19F NMR spectroscopy has emerged as a particularly powerful tool for studying RNA structure and small molecule interactions. This method involves incorporating fluorine atoms into RNA through chemical synthesis, chemo-enzymatic methods, or in vitro transcription, typically at specific positions in the sugar or base groups [14]. The technique's high sensitivity and simplicity make it valuable for probing RNA folding, conformational changes, and ligand binding properties. Key applications include monitoring RNA structural transitions, identifying small molecule binding sites, and assessing binding affinity and specificity through ligand-observed or target-observed experiments [14].

Cryo-electron microscopy (cryo-EM) has increasingly been applied to larger RNA-protein complexes, providing structural insights without the need for crystallization. Chemical probing methods, such as SHAPE (Selective 2'-Hydroxyl Acylation analyzed by Primer Extension) and DMS (Dimethyl Sulfate) mapping, provide information about nucleotide accessibility and flexibility, revealing secondary structure elements and dynamic regions [47]. These techniques are often combined in integrative approaches that leverage multiple methods to overcome individual limitations and provide comprehensive structural insights into complex RNA molecules.

Visualization and Analysis Tools

R2DT (RNA 2D Templates) represents a significant advancement in RNA secondary structure visualization, designed to produce consistent, recognizable, and reproducible layouts [48]. The platform uses a comprehensive template library with 4,612 2D templates (as of 2025) and incorporates multiple key features:

• Template-based visualization: R2DT searches input sequences against its template library using BLAST for larger RNAs, profile hidden Markov models for Rfam families, and tRNAscan-SE for tRNAs, ensuring appropriate template selection [48].
• Annotation capabilities: The software enables visualization of position-specific information such as single nucleotide polymorphisms, SHAPE reactivities, and sequence conservation as data layers [48].
• Pseudoknot visualization: R2DT can display pseudoknots, which are important structural motifs where non-nested base pairing creates knot-like structures critical for RNA function [48].
• Template-free mode: For RNAs without existing templates, R2DT can visualize structures using R2R or RNArtist layout engines, generating new templates for future use [48].

The platform is integrated into multiple biological databases and is available as a web server, API, standalone program, and embeddable web component, creating a comprehensive ecosystem for RNA 2D structure visualization and analysis [48].

Computational RNA Structure Prediction

Computational methods have dramatically advanced RNA structure prediction, complementing experimental approaches. Traditional thermodynamics-based methods like RNAfold use nearest-neighbor models and dynamic programming to predict minimum free energy structures [47]. These methods leverage parameters from experimental data, such as optical melting studies, refined through machine learning approaches [47].

Machine learning and deep learning have revolutionized RNA structure prediction. The ERNIE-RNA model exemplifies this advancement—a pre-trained RNA language model based on a modified BERT architecture that incorporates base-pairing-informed attention bias during attention score calculation [1]. This innovative approach allows the model to learn RNA structural patterns through self-supervised learning rather than relying on potentially biased structural predictions. ERNIE-RNA demonstrates remarkable capability in zero-shot RNA secondary structure prediction, achieving F1-scores up to 0.55, outperforming conventional methods like RNAfold and RNAstructure [1]. After fine-tuning, the model achieves state-of-the-art performance across various downstream tasks, including RNA structure and function predictions [1].

Molecular dynamics (MD) simulations have also seen significant advances in modeling RNA folding and dynamics. Recent research has demonstrated that combining the DESRES-RNA atomistic force field with the GB-neck2 implicit solvent model enables accurate folding simulations of diverse RNA stem-loops, achieving root mean square deviation values of less than 2 Ã… for stems and less than 5 Ã… overall in 23 of 26 tested structures [49]. This represents a milestone in computational RNA modeling, moving beyond simple stem-loops to more complex structures containing bulges and internal loops.

Table 1: Computational Methods for RNA Structure Prediction

Method Category	Representative Tools	Key Features	Applications
Thermodynamics-based	RNAfold, RNAstructure	Uses nearest-neighbor models with free energy minimization	Secondary structure prediction from sequence
Comparative Sequence Analysis	R2DT, Infernal	Leverages evolutionary conservation through multiple sequence alignments	Identifying conserved structural elements
Deep Learning Models	ERNIE-RNA, RNA-FM	Self-supervised learning on large sequence datasets; attention mechanisms	Secondary and tertiary structure prediction
Molecular Dynamics	DESRES-RNA/GB-neck2	Physics-based simulations with implicit solvent	Folding pathways, conformational dynamics
Hybrid Approaches	DREEM, MaP	Integrates chemical probing data with statistical models	RNA structural ensembles, in-cell structures

Small Molecule Targeting Strategies

RNA Target Selection and Characterization

Successful RNA-targeted drug discovery begins with careful target selection and characterization. Ideal RNA targets share several characteristics: well-defined and stable secondary and tertiary structures, functional importance in disease pathways, and the presence of specific binding pockets or clefts amenable to small molecule binding [14]. Promising target classes include riboswitches, ribosomal RNAs, internal ribosome entry sites (IRES), splicing regulatory elements, and non-coding RNAs with defined structures.

The hepatitis C internal ribosome entry site (HCV-IRES) domain IIa represents a well-characterized RNA target example. Its architecture features a complex ligand-binding pocket organized by a metal spine, containing three magnesium ions as intrinsic structural components [45]. The RNA undergoes dramatic ligand-induced conformational adaptation from an L-shaped to an extended form, creating a deep pocket resembling substrate binding sites in riboswitches [45]. Understanding such structural dynamics is crucial for effective targeting.

Target engagement studies should assess both structural and functional aspects. Biophysical methods like surface plasmon resonance, isothermal titration calorimetry, and NMR can quantify binding affinity and kinetics, while functional assays (splicing reporters, translation assays, etc.) confirm mechanistic outcomes [14]. The increasing availability of RNA-focused small molecule screening libraries further facilitates target validation and early hit identification [46].

Computational Design and Screening

Computational approaches have become indispensable for RNA-targeted small molecule design, accelerating discovery and reducing costs. Several strategies have emerged as particularly effective:

Absolute Binding Free Energy (ABFE) Calculations using advanced molecular dynamics simulations represent a state-of-the-art approach for predicting small molecule binding affinities to RNA targets. A recently developed protocol combines the AMOEBA polarizable force field with the lambda-Adaptive Biasing Force (lambda-ABF) scheme and refined restraints for efficient sampling [45]. The AMOEBA force field employs atomic induced dipoles to include polarization effects and atomic multipoles up to quadrupoles to represent electrostatics anisotropy, providing more accurate treatment of RNA-ligand interactions [45]. This approach successfully predicted binding affinities for 19 2-aminobenzimidazole derivatives targeting the HCV-IRES IIa subdomain, demonstrating quantitative predictions for this challenging riboswitch-like RNA target [45].

Machine Learning and Deep Learning applications in RNA-targeted drug discovery are expanding rapidly. AI-powered tools can analyze large datasets to forecast RNA structures, locate potential binding regions, and simulate molecular behavior, significantly accelerating candidate identification [50]. The integration of artificial intelligence improves the precision and efficiency of development while reducing time and expense compared to traditional methods [50]. Startup-driven AI solutions specifically for RNA-targeted therapeutics represent an emerging trend in the field [51].

Molecular Docking and Virtual Screening face unique challenges for RNA targets due to their high flexibility and electrostatics. Successful strategies often incorporate RNA flexibility through ensemble docking, use physics-based scoring functions that account for solvation and electrostatic effects, and apply filters for RNA-specific physicochemical properties [47]. Specialized RNA-focused libraries can improve virtual screening success rates by enriching for compounds with appropriate properties for RNA binding [46].

Table 2: Experimental Techniques for Studying RNA-Small Molecule Interactions

Technique	Information Obtained	Throughput	Sample Requirements	Key Applications
19F NMR	Binding site identification, conformational changes, binding affinity	Medium	100-500 µM RNA, fluorine-labeled	Fragment screening, binding mechanism
Surface Plasmon Resonance	Binding kinetics (ka, kd), affinity (KD), stoichiometry	Medium-high	Low µg of RNA	Hit validation, SAR studies
Isothermal Titration Calorimetry	Binding affinity, stoichiometry, thermodynamics (ΔH, ΔS)	Low	High purity RNA and compounds	Mechanism studies, binding driving forces
Chemical Probing	Structural changes upon binding, binding site mapping	Medium	Varies by method	Binding site identification, mechanism
Fluorescence Anisotropy	Binding affinity, competition assays	High	Fluorescently-labeled RNA	High-throughput screening, competition
Native Mass Spectrometry	Stoichiometry, binding affinity, complex stability	Medium	Low µM concentrations	Complex assembly, weak interactions

Specialized Screening Libraries and Hit Identification

The development of RNA-focused small molecule libraries has evolved significantly, improving screening outcomes by enriching for compounds with appropriate properties for RNA binding [46]. Several strategic approaches have proven effective:

Fragment-Based Libraries utilize small, simple compounds (typically 100-250 Da) that explore chemical space more efficiently than larger molecules. These fragments tend to bind with higher ligand efficiency and can be optimized through structural guidance. Fragment screening often employs biophysical methods like 19F NMR, surface plasmon resonance, or X-ray crystallography to detect weak binders that might be missed in conventional assays [46].

DNA-Encoded Libraries (DELs) enable screening of extremely large compound collections (millions to billions) through taggable identification systems. DEL technology has been successfully applied to RNA targets, expanding the accessible chemical space for RNA binder discovery [47]. The approach combines the diversity of combinatorial chemistry with the sensitivity of DNA amplification, allowing efficient screening against challenging RNA targets.

Natural Product-Inspired Libraries leverage scaffolds from known RNA-binding natural products like aminoglycosides, tetracyclines, and macrolides. These compounds often possess privileged structural features for RNA recognition, including cationic groups for electrostatic interactions with the RNA backbone and rigid frameworks that pre-organize for binding [46]. Semi-synthetic derivatives can improve properties while retaining binding capabilities.

Structure-Based Design Libraries focus on compounds targeting specific RNA structural motifs, such as internal loops, bulges, G-quadruplexes, or tertiary interactions. These libraries often incorporate features known to interact with RNA, including intercalators, groove binders, and base-pair mimetics [46]. The RNA-focused small molecule drug discovery market is growing rapidly, valued at $1.68 billion in 2024 and expected to reach $5.52 billion by 2030, reflecting increased investment in this approach [50].

Experimental Protocols and Methodologies

Absolute Binding Free Energy Calculation Protocol

The accurate prediction of binding affinities is crucial for rational drug design. The following protocol describes absolute binding free energy calculations using the lambda-ABF approach with the AMOEBA polarizable force field:

System Preparation:

• Obtain the RNA-ligand complex structure from crystallography or homology modeling. For the HCV-IRES target, the PDB ID 3TZR provides the experimental structure at 2.2 Ã… resolution [45].
• Parameterize the RNA using the AMOEBA nucleic acid force field, which includes parameters for DNA and RNA derived from quantum chemistry calculations [45].
• Parameterize the ligand using the AMOEBA force field, deriving electrostatic parameters (multipoles and polarizabilities) from quantum mechanical calculations.
• Solvate the system in explicit water using a polarizable water model compatible with AMOEBA.
• Add counterions to neutralize the system and additional salt to match physiological conditions (typically 150 mM NaCl). Include essential divalent cations (Mg2+) when present in the native structure.

Equilibration and Sampling:

• Perform energy minimization using the Tinker-HP software package, which provides GPU-accelerated molecular dynamics for polarizable force fields [45].
• Conduct equilibration in the NPT ensemble (constant number of particles, pressure, and temperature) at 300 K and 1 atm pressure using a Langevin thermostat and Monte Carlo barostat.
• Apply the lambda-ABF method, which combines λ-dynamics with a multiple-walker adaptive biasing force, avoiding the discretization of the alchemical path used in traditional methods [45].
• Implement distance-to-bound-configuration (DBC) restraints to maintain the ligand in the binding site while allowing necessary conformational freedom [45].
• For systems with significant conformational changes between Apo and Holo states, combine machine learning-based collective variables with enhanced sampling simulations to capture the associated free energy barrier [45].
• Run production simulations for sufficient time to achieve convergence (typically 100-500 ns per window for complex RNA systems).

Analysis:

• Calculate the absolute binding free energy from the ABF simulations using the appropriate thermodynamic cycle.
• Estimate statistical uncertainties using block analysis or bootstrapping methods.
• Validate predictions against experimental binding data for known compounds before applying to novel designs.

19F NMR Screening Protocol for RNA-Targeted Small Molecules

19F NMR provides a sensitive method for detecting ligand binding to RNA targets. The following protocol describes a fragment screening approach:

RNA Preparation and Labeling:

• Design RNA construct containing the target structure, typically 30-50 nucleotides for manageable NMR spectra.
• Incorporate fluorine labels through:
  • Chemical synthesis of fluorinated nucleotides (e.g., 5-fluorouridine, 2-fluoroadenosine)
  • Enzymatic incorporation using fluorinated nucleotide triphosphates and T7 RNA polymerase
  • Post-synthetic modification using fluorinated reagents
• Purify labeled RNA using denaturing PAGE or HPLC, followed by electroelution or desalting.
• Fold RNA by heating to 90°C for 2 minutes in appropriate buffer, then slow cooling to room temperature in the presence of Mg2+ if required for folding.

NMR Experiments:

• Prepare RNA sample at 10-100 µM concentration in NMR buffer (typically 25 mM potassium phosphate, pH 6.5-7.5, 50-100 mM KCl, 0.1-5 mM MgCl2, in 90% H2O/10% D2O).
• Acquire 1D 19F NMR reference spectrum of RNA alone at 25°C using a cryoprobe-equipped NMR spectrometer.
• Prepare compound mixtures (10-20 fragments per pool) at 100-500 µM each in DMSO-d6, keeping final DMSO concentration ≤5%.
• Add compound pool to RNA sample, incubate 15-30 minutes, then acquire 1D 19F NMR spectrum.
• Run control experiments with compound pools alone and with non-specific RNA (e.g., tRNA) to identify non-specific binders.

Data Analysis:

• Identify binding events through:
  • Chemical shift perturbations (CSPs) of fluorine peaks
  • Line broadening indicating intermediate exchange
  • Signal intensity changes from transferred NOE
• Deconvolute active pools by testing individual compounds.
• Quantify binding affinity by titrating compound and measuring CSPs as a function of concentration, fitting to appropriate binding models.
• Validate hits using orthogonal methods (SPR, ITC, functional assays).

RNA-Focused Library Screening Protocol

Library Design and Preparation:

• Select or design a focused library incorporating known RNA-bindingprivileged scaffolds, fragment-like diversity, and target-informed designs.
• Format library in 384-well plates as 10 mM DMSO stocks, with appropriate controls and reference compounds.
• Perform quality control (LC-MS, purity assessment) on library compounds.

High-Throughput Screening:

• Express and purify target RNA using in vitro transcription or chemical synthesis.
• Develop robust binding assay (fluorescence anisotropy, FRET, ALPHA, etc.) with Z' factor >0.5.
• Screen library at single concentration (typically 10-50 µM) in duplicate, including controls on each plate.
• Identify primary hits with >3 standard deviation response from negative controls.
• Confirm hits in dose-response format (8-point dilution series) to determine IC50 or EC50 values.

Hit Validation:

• Confirm binding using orthogonal biophysical methods (SPR, ITC, NMR).
• Assess specificity using related but non-target RNA structures.
• Evaluate functional activity in biochemical or cell-based assays relevant to the target's mechanism.
• Conduct counter-screens against common off-targets and assay artifacts.

Hit-to-Lead Optimization:

• Determine structure-activity relationships through analog testing and focused library synthesis.
• Obtain structural information on binding mode (crystallography, NMR, modeling) to guide optimization.
• Optimize key properties: potency, selectivity, solubility, metabolic stability.
• Advance lead compounds to in vitro and in vivo disease models.

Visualization of Workflows and Relationships

RNA-Targeted Drug Discovery Workflow

RNA Structure Determination Methods

Small Molecule Binding Mechanisms to RNA

Table 3: Essential Research Reagents for RNA-Targeted Small Molecule Discovery

Category	Specific Reagents/Tools	Function/Application	Key Features
RNA Structure Prediction	ERNIE-RNA, RNAfold, RNAstructure	Predicting secondary and tertiary RNA structures	Base-pairing informed attention (ERNIE-RNA), thermodynamic parameters
Molecular Dynamics	Tinker-HP, AMOEBA Force Field, DESRES-RNA	Binding free energy calculations, folding simulations	Polarizable force field, GPU acceleration
Structure Visualization	R2DT, PyMOL, ChimeraX	2D and 3D RNA structure visualization	Template-based layouts, annotation capabilities
Biophysical Characterization	19F-labeled nucleotides, SPR chips, ITC reagents	Binding affinity and kinetics measurement	High sensitivity, quantitative binding parameters
Specialized Libraries	Fragment libraries, DNA-encoded libraries, Natural product collections	Hit identification for RNA targets	RNA-focused chemical space, privileged scaffolds
Structural Biology	Crystallization screens, Cryo-EM grids, NMR isotopes	High-resolution structure determination	RNA-optimized conditions, stabilization reagents

The field of RNA-targeted small molecule therapeutics is rapidly evolving, driven by advances in structural biology, computational methods, and chemical library design. The successful approval of risdiplam demonstrated that targeting non-ribosomal RNA structures with small molecules can yield effective medicines, paving the way for broader applications [46]. Current research continues to address the fundamental challenges of RNA target engagement, specificity, and functional modulation.

Several trends are shaping the future of this field. The integration of artificial intelligence and machine learning is accelerating both RNA structure prediction and small molecule design, with models like ERNIE-RNA demonstrating remarkable capability in extracting structural information from sequence data alone [1]. The growing application of advanced biophysical methods like 19F NMR provides sensitive tools for characterizing RNA-ligand interactions and guiding optimization [14]. Furthermore, the development of novel therapeutic modalities such as RIBOTACs and other targeted degraders expands the mechanisms by which small molecules can modulate RNA function [14].

The market outlook for RNA-targeting small molecule therapeutics reflects this progress, with the global market value projected to grow from $2.77 billion in 2024 to $7.03 billion in 2034, driven by increasing R&D investments and the need to treat genetic disorders, cancer, and other diseases [51]. This growth is particularly strong in the RNA splicing modification segment, which accounts for 66.76% of the current market and is expected to remain the fastest-growing segment [51].

As the field advances, key areas for continued development include expanding the structural database of RNA-small molecule complexes, improving computational methods for predicting binding affinities, and developing more sophisticated delivery strategies for tissue-specific targeting. With these advances, RNA-targeted small molecules have the potential to address currently untreatable diseases and significantly expand the druggable genome.

Navigating Challenges in RNA Structure Determination and Modeling

The field of RNA biology is at a pivotal juncture. Research has revealed that while most of the human genome is transcribed into RNA, only a minimal proportion (∼1.5%) codes for proteins [47]. This vast landscape of non-coding RNA presents unprecedented opportunities for therapeutic intervention, with RNA-targeting small molecules offering novel avenues for diseases traditionally deemed undruggable [47]. However, progress is severely hampered by a fundamental challenge: the scarcity of high-resolution RNA structural data. This scarcity stems from significant technical hurdles in RNA structure determination, including RNA's structural flexibility, polyanionic nature, and the inherent difficulties of applying techniques like X-ray crystallography and cryo-electron microscopy to RNA molecules [47]. This whitepaper outlines integrated computational and experimental strategies to overcome data scarcity, enabling robust research and drug discovery in RNA biology.

The Core Challenge: Limitations in RNA Structural Data

The foundation of rational drug design—whether targeting proteins or RNA—is high-resolution structural information. For RNA, the number of experimentally solved high-resolution structures remains limited, with a strong bias toward specific RNA classes such as ribosomal RNAs, transfer RNAs, and riboswitches [47]. This data scarcity creates a critical bottleneck for drug discovery efforts, particularly for the myriad of disease-associated non-coding RNAs that represent promising therapeutic targets.

The primary challenges in experimental RNA structure determination include:

• Structural Flexibility: RNA molecules often exist as dynamic ensembles of conformations rather than single, static structures, complicating data interpretation [47].
• Technical Limitations: Biophysical methods like X-ray crystallography face challenges with RNA crystal formation, while Nuclear Magnetic Resonance (NMR) spectroscopy is generally restricted to smaller RNA molecules [47].
• Cellular Environment Complexity: RNA structures determined in vitro may not accurately reflect their conformations in the complex cellular milieu, where interactions with proteins and other factors influence folding [52].

This structural data gap directly impedes the development of RNA-targeted small molecules, as structure-based drug design relies on detailed understanding of binding pockets and molecular interactions.

Computational Strategies for Data Augmentation

Leveraging Deep Learning for Structure Prediction

Machine learning, particularly deep learning, has emerged as a powerful approach for predicting RNA structures from sequence data, thereby augmenting limited experimental datasets. These methods can integrate diverse data sources—including sequence information, chemical probing data, and evolutionary conservation—to construct reliable structural models [47].

Table 1: Computational Approaches for RNA Structure Prediction

Method Category	Key Principles	Representative Tools/Examples
Nearest Neighbor Models	Uses dynamic programming to find the minimum free energy structure based on thermodynamic parameters [47].	RNAstructure
Deep Learning Models	Trained on established RNA structures to predict secondary and tertiary structures; can simulate multiple conformations rapidly [47].	SPOT-RNA, MXFold2, UFold
Differentiable Alignment	Implements smooth versions of alignment algorithms (e.g., Smith-Waterman) to enable end-to-end learning of multiple sequence alignments with downstream tasks [53].	SMURF (Smooth Markov Unaligned Random Field)

Ensemble Methods and Data Integration

Advanced computational strategies move beyond single-structure prediction to capture the dynamic nature of RNA biology:

Ensemble Alignment Methods: Tools like Muscle5 construct ensembles of high-accuracy multiple sequence alignments (MSAs) with diverse biases by perturbing hidden Markov model parameters and permuting guide trees [54]. This approach enables unbiased assessments of sequence homology and phylogeny by evaluating the consistency of inferences across alignment variants.

Integrative Modeling: Combines multiple experimental and computational methods to overcome individual limitations, providing comprehensive structural insights into complex RNA molecules [47]. This may include integrating chemical probing data, evolutionary information, and partial experimental constraints to refine structural models.

The following diagram illustrates how ensemble analysis provides a robust framework for confident inference when data is limited:

Synthetic Data Generation

Synthetic data generation using advanced artificial intelligence techniques provides another pathway to address data scarcity. These approaches replicate the statistical properties of real-world datasets while excluding identifiable information, enabling analyses that yield results comparable to those obtained with real data [55].

Deep Generative Models: Approaches such as Variational Autoencoders (VAEs) and Deep Boltzmann Machines (DBMs) can learn the joint distribution of various data types and generate synthetic observations after training on initial samples [56]. For single-cell RNA sequencing data, these models can be trained on pilot datasets to generate larger synthetic datasets for experimental planning and method development.

Generative Adversarial Networks (GANs): Models like recurrent time-series GANs (RTSGAN) have been used to synthesize life-log data with temporal structure, while other GAN-based approaches can generate synthetic sequencing data by introducing realistic biological variability to group-specific mean values [55].

Experimental Strategies for Enhanced Data Generation

Advanced Chemical Probing Methodologies

Experimental methods that provide medium-resolution structural information at scale can effectively bridge the gap between sequence and high-resolution structures:

eSHAPE and Related Technologies: eSHAPE is an experimental assay that generates ready-to-download datasets of RNA structures in various cell lines and tissues [52]. The technology measures nucleotide reactivity, where increased reactivity indicates a higher probability that the nucleotide is unpaired. This data serves as input to RNA folding algorithms to guide structure prediction to more accurate final structures.

Comparative Analysis: By performing eSHAPE both in vitro (without cellular factors) and in cellulo (with cellular factors), researchers can detect bases that directly interact with RNA-binding proteins, providing crucial information about RNA-protein interactions that influence structure and function [52].

Table 2: Experimental Reagents and Resources for RNA Structure Determination

Research Reagent/Resource	Function and Application
eSHAPE Assay	Measures nucleotide reactivity to determine RNA secondary structure in cellular contexts [52].
X-ray Crystallography	Provides atomic-resolution RNA structures but requires crystal formation [47].
Cryo-electron Microscopy	Enables structure determination of large RNA-protein complexes without crystallization [47].
NMR Spectroscopy	Studies smaller RNA molecules and their dynamics in solution [47].
DNA-encoded Libraries (DELs)	Identifies bioactive ligands targeting RNA through high-throughput screening [47].

The development of comprehensive, standardized datasets and benchmarks is crucial for advancing the field and making the most of limited high-resolution data:

RNA Structure Datasets: Resources like those from Eclipse Bio provide experimentally validated RNA structures in various cell lines and tissues, specifically designed for AI model training and drug development [52]. These datasets include complete packages from raw sequencing data to secondary analyses, ready for input into machine learning algorithms.

Community-wide Benchmarks: Recent efforts have created comprehensive datasets of over 320,000 instances from experimentally validated sources to establish benchmarks for RNA design and modeling algorithms [37]. These resources encompass diverse structural motifs, from internal loops to n-way junctions, enabling rigorous testing and development of computational methods.

Specialized Benchmarking Suites: For RNA 3D structure-function modeling, benchmarking suites like those built on the rnaglib Python package provide standardized datasets, splitting strategies, and evaluation metrics specifically designed for deep learning applications [57] [58].

The workflow below illustrates an integrated approach that combines experimental and computational strategies to overcome data scarcity:

Integrated Experimental-Computational Workflows

Overcoming data scarcity in RNA structural biology requires tight integration of experimental and computational approaches:

End-to-End Differentiable Pipelines

Emerging frameworks enable joint learning of multiple sequence alignments and downstream task models in an end-to-end fashion. The Learned Alignment Module (LAM) implements a smooth and differentiable version of the Smith-Waterman pairwise alignment algorithm, allowing joint optimization of alignments with subsequent analysis steps [53]. This approach demonstrates that jointly learning an alignment and the parameters of a Markov Random Field can improve contact prediction from unaligned sequences.

Cross-Modal Data Integration

Effective strategies integrate data from multiple sources and resolutions:

• Chemical Probing Integration: Using experimental data like eSHAPE reactivities as constraints in computational folding algorithms [52].
• Evolutionary Information: Leveraging multiple sequence alignments to infer structural constraints through co-evolution analysis [53] [54].
• Multi-Scale Modeling: Combining information from different structural biology techniques (cryo-EM, NMR, chemical probing) to build comprehensive models that account for RNA flexibility and cellular context [47].

Best Practices and Implementation Guidelines

Experimental Design for Data Scarcity Contexts

When working with limited high-resolution data, researchers should:

Prioritize Multi-Scale Data Collection: Invest in medium-throughput experimental data (like chemical probing) that can inform and validate computational models, even when high-resolution structures are unavailable [52].

Implement Robust Validation Strategies: Use ensemble methods like those in Muscle5 to assess confidence in inferences by measuring their consistency across alignment variants [54]. The H-ensemble confidence (HEC) metric quantifies the fraction of alignment replicates that support a particular inference, providing an unbiased assessment of confidence.

Leverage Public Data Resources: Maximize use of community resources like CompaRNA for benchmarking prediction methods [59], RNAsolo for cleaned structural data [37], and specialized benchmarks for RNA 3D structure-function modeling [57] [58].

Computational Methodology Considerations

Appropriate Benchmarking: When developing new computational methods, use standardized benchmarks and implement rigorous data splitting strategies that account for sequence and structural similarity to avoid overestimation of performance [58].

Uncertainty Quantification: Implement methods that provide confidence estimates for predictions, such as ensemble approaches that generate multiple plausible models rather than single point estimates [54].

Modular Architecture Design: Develop computational pipelines with modular components that can be updated as new experimental data becomes available, allowing for continuous improvement of models as data scarcity diminishes over time.

The challenge of data scarcity in RNA structural biology is significant but not insurmountable. By integrating computational approaches—including deep learning, ensemble methods, and synthetic data generation—with strategic experimental techniques that provide medium-resolution data at scale, researchers can build robust models of RNA structure and function. The key to success lies in the tight integration of these approaches, leveraging their respective strengths to overcome individual limitations. Community efforts to develop standardized benchmarks and share datasets are accelerating progress, enabling more effective use of limited high-resolution data. As these strategies mature, they will unlock the full potential of RNA-targeted therapeutics and advance our understanding of RNA biology, ultimately transforming treatment paradigms for a wide range of diseases.

The accurate representation of molecular systems through force fields (FFs) is the cornerstone of reliable Molecular Dynamics (MD) simulations. For RNA, a molecule whose biological function is intimately linked to its structural dynamics and interactions, the fidelity of these force fields is paramount. Imperfections in the energy models can lead to significant discrepancies between simulation outcomes and experimental observations, limiting the predictive power of in silico studies in fundamental research and drug development. This review provides an in-depth technical guide to the current state of RNA force fields, evaluating their accuracy through recent benchmark studies, detailing essential validation methodologies, and outlining persistent limitations that the field must overcome.

Current Generation of RNA Force Fields

Several families of atomistic force fields are extensively used for classical MD simulations of RNA systems. The most common are additive force fields, which use fixed atomic charges and pre-defined parameters to model bonded and non-bonded interactions [40].

The AMBER family includes the widely tested OL3 (bsc0χOL3) force field [36] [40]. Recent empirical corrections have been applied to this baseline to improve its performance. These include the gHBfix potential, which adjusts the strength of specific hydrogen bonds [40], modifications to Lennard-Jones interactions to correct over-stabilized base stacking [36], and alternative charge-derivation strategies [40]. A more extensive re-parameterization was undertaken by the D.E. Shaw Research group, resulting in the DESRES RNA force field, which was refined using quantum mechanical calculations and experimental data to better reproduce the energetics of nucleobase stacking and base pairing [36]. The CHARMM family is represented by CHARMM36 [60] [40], another widely used and benchmarked force field. Beyond these additive models, polarizable force fields exist that aim to more accurately capture electronic effects, but they come with a significantly higher computational cost [40].

Table 1: Major RNA Force Fields and Their Key Characteristics

Force Field	Base Force Field	Key Refinements	Commonly Paired Solvent Model
AMBER-OL3	AMBER (bsc0)	χOL3 correction for glycosidic torsion [36]	TIP3P, OPC, TIP4P-D [36] [40]
DESRES RNA	Not specified	Extensive reparameterization based on QM calculations and experimental data [36]	TIP4P-D [36]
CHARMM36	CHARMM	Standard nucleic acid parameters [60]	TIP3P [40]
AMBER-OL3 + gHBfix	AMBER-OL3	Structure-specific correction potential for native hydrogen bonds (gHBfix) [36]	TIP3P, OPC [61]

Assessing Force Field Accuracy: Key Metrics and Case Studies

The validity of a force field must be tested against experimental observations. Recent systematic studies have provided critical benchmarks by evaluating the performance of different FFs across various RNA systems, from simple duplexes to complex ligand-binding pockets.

RNA-Ligand Complex Stability

A comprehensive 2025 study assessed the last generation of RNA FFs on a curated set of 10 RNA–small molecule complexes from the HARIBOSS database [61]. The systems included diverse RNA topologies (double helices, hairpins) and binding modes (groove binding, intercalation). The FFs tested were OL3, its variant OL3cp with gHBfix21, the Shaw (DESRES) FF, and a version of DESRES adapted for AMBER (DES-AMBER). Each system was simulated for 1 μs, and the results were analyzed using multiple quantitative metrics [61].

Table 2: Performance of Force Fields in RNA-Ligand Complex Simulations (Adapted from [61])

Force Field	RNA Structure Maintenance	Ligand Stability (LoRMSD)	Native Contact Preservation	Key Finding
OL3	Moderate stability, some terminal fraying	Variable, significant drift in some cases	Lower contact retention	Baseline performance
OL3cp + gHBfix	Improved intra-RNA interactions, reduced fraying	Similar to or better than OL3	Better preservation of native contacts	Can sometimes distort the experimental RNA model
Shaw (DESRES)	Good structural maintenance	Variable	Moderate contact preservation	Generally effective at stabilizing RNA structure
DES-AMBER	Good structural maintenance	Variable	Moderate contact preservation	Tends to improve energetic interactions with ligands

The study concluded that while modern FFs are generally effective at stabilizing RNA structures without major distortions, they often struggle to consistently maintain stable RNA-ligand interactions over microsecond timescales. A common observation was the loss of native contacts present in the initial experimental structure and the formation of new, non-native contacts during the simulation [61]. This highlights a critical limitation: current FFs may not yet reliably reproduce the precise intermolecular interactions crucial for drug design.

Prediction of Binding Affinities for Modified RNAs

Beyond maintaining structural integrity, a force field's ability to predict binding affinities is a stringent test of its accuracy. λ-dynamics (λD), an efficient alchemical free energy method, has been used to predict the binding affinities of both canonical and modified RNAs to the Pumilio RNA-binding protein [60]. This approach screens a library of RNA modifications at different nucleotide positions to identify those that affect binding. When computed binding affinities were compared to experimental data, a high predictive accuracy was observed [60]. The study also compared the CHARMM36 and AMBER force fields, identifying the best parameter set for such binding calculations, though the specific results of that comparison were not detailed in the excerpt [60]. This demonstrates that with sufficiently advanced sampling methods, current FFs can yield quantitatively accurate predictions for biologically critical interactions, such as those involving post-transcriptional RNA modifications.

RNA Stem-Loop Folding

The capability of a force field to fold an RNA molecule from an extended conformation to its native state is one of the most demanding benchmarks. A 2025 study performed de novo folding simulations of 26 RNA stem-loops (10-36 residues) using the DESRES RNA force field and an implicit solvent model (GB-neck2) [36]. The results were highly promising:

• All 18 stem-loops without bulges or internal loops folded into structures retaining all native base pairs in the stem regions, with root mean square deviation (RMSD) values of <2 Ã… for stems and <5 Ã… for the full molecule [36].
• Five of the eight stem-loops containing bulges or internal loops were also successfully folded, with all native stem base pairs formed, though overall molecular RMSDs were higher (2.8–8.3 Ã…) [36].

This success demonstrates that a well-parameterized force field, even with an implicit solvent, can recapitulate the fundamental folding of RNA secondary structures. However, the lower accuracy for loop regions and more complex topologies points to areas requiring further refinement [36].

Essential Protocols for Force Field Validation

To ensure the reliability of MD simulations, researchers should employ a multi-faceted validation strategy. The following protocols, drawn from recent benchmark studies, provide a template for assessing force field performance.

Workflow for Benchmarking RNA-Ligand Complexes

System Setup and Simulation

• Structure Selection: Curate a diverse set of RNA-ligand complexes (e.g., from HARIBOSS) encompassing different RNA topologies and binding modes [61].
• Solvation and Ions: Solvate the system in an octahedral water box (e.g., OPC, TIP4P-D) with a 15 Ã… padding. Neutralize with K+ ions and add K+/Cl- ions to a physiological concentration of 0.15 M [61].
• Ligand Parametrization: Derive ligand parameters using the GAFF2 force field and RESP2 charges, ensuring compatibility with the RNA FF [61].
• Equilibration: Perform energy minimization, followed by step-wise thermalization to 300 K and pressure equilibration to 1 atm [61].
• Production Simulation: Run multiple, 1 μs-long, unrestrained MD simulations for each system and force field [61].

Analysis Metrics

• Root Mean Square Deviation (RMSD): Calculate for RNA heavy atoms and ligand heavy atoms (after aligning on the RNA backbone) to assess structural stability and ligand mobility [61].
• Contact Map Analysis: Define a residue-residue contact as any two heavy atoms within 4.5 Ã…. Calculate the frequency of contacts across the trajectory and compare to the starting crystal structure to identify lost or gained interactions [61].
• Helical Parameters: Use tools like pytraj nastruct to compute parameters like major groove width and twist, comparing them to experimental data [61].

Workflow for Binding Affinity Prediction

This protocol uses λ-dynamics to efficiently compute the relative binding free energies for a library of RNA modifications [60].

• System Preparation: Start with a high-resolution structure of the protein-RNA complex (e.g., Pumilio bound to its cognate RNA) [60].
• Alchemical Transformation: Set up a λ-dynamics simulation that simultaneously samples multiple alchemical states, each representing a different RNA modification at a specific nucleotide position [60].
• Free Energy Calculation: Use the λ-dynamics framework to calculate the relative change in binding free energy (ΔΔG_bind) for each modification compared to the unmodified reference [60].
• Validation: Compare the computational predictions with experimentally measured binding affinities (e.g., from in vitro assays) to determine the accuracy of the force field [60].

Table 3: Essential Computational Tools for RNA MD Simulations

Resource	Type	Function/Purpose
AMBER-OL3 / CHARMM36	Force Field	Provides energy function parameters for RNA atoms [60] [36].
GAFF2 & RESP2 Charges	Force Field	Provides parameters for small molecule ligands in complex with RNA [61].
GB-neck2	Implicit Solvent Model	Accelerates conformational sampling by approximating solvent as a continuum [36].
OPC / TIP4P-D	Explicit Water Model	Explicitly models water molecules for more realistic solvation [36] [61].
gHBfix	Correction Potential	Applies structure-specific hydrogen bond corrections to improve accuracy [36] [61].
HARIBOSS	Database	Curated source of RNA-small molecule complex structures for benchmarking [61].
λ-Dynamics	Enhanced Sampling	Efficiently calculates relative binding free energies for multiple ligands/modifications [60].
PLUMED	Software Plugin	Enables enhanced sampling methods and on-the-fly analysis [61].

Limitations and Emerging Solutions

Despite significant progress, current force fields have recognizable limitations. A primary challenge is the inconsistent stability of RNA-ligand complexes, where simulations often show a loss of native contacts and a drift in ligand position (high LoRMSD) [61]. Furthermore, accurately modeling the flexible loop regions of RNA remains more challenging than modeling the structured stem regions [36]. There is also a critical need for force fields that can accurately describe the diverse array of over 170 natural RNA chemical modifications [60], as their interactions with proteins are fundamental to epitranscriptomics.

Emerging methodologies are poised to address these limitations. Integrative approaches that combine simulations with experimental data (e.g., WAXS, NMR) can refine structural ensembles and improve accuracy [40]. Machine learning is now being leveraged to develop next-generation force fields. For instance, Grappa is a ML-based force field that predicts molecular mechanics parameters from the molecular graph, achieving state-of-the-art accuracy for small molecules, peptides, and RNA at a computational cost comparable to traditional FFs [62]. Finally, the continued development and application of polarizable force fields promise a more physical description of electrostatic interactions, which are critical for RNA-ligand and RNA-protein recognition [40].

The fidelity of force fields directly dictates the value of MD simulations in RNA research and drug development. While modern FFs like the DESRES RNA, CHARMM36, and refined AMBER-OL3 variants can reliably model RNA secondary structure folding and, in some cases, predict binding affinities with high accuracy, they still struggle with the precise stabilization of RNA-ligand complexes and the modeling of flexible regions. The path forward lies in the continuous benchmarking of FFs against diverse experimental data, the adoption of advanced sampling and integrative methods, and the harnessing of machine learning to develop more accurate and data-efficient energy models. As these tools mature, in silico predictions will become an even more powerful and indispensable component of the RNA structural biology and drug discovery pipeline.

The biological function of ribonucleic acid (RNA) is profoundly intertwined with its ability to adopt diverse and dynamic three-dimensional structures. [40] Unlike the traditional view that focused on static snapshots, the modern interpretation recognizes that RNA molecules exist as structural ensembles, sampling multiple conformations in equilibrium. [40] This dynamic nature is crucial for understanding how RNAs perform their roles, from gene regulation and catalytic activity to interactions with proteins, small molecules, and other RNAs. However, this very characteristic presents a formidable challenge for computational structural biologists: efficiently and accurately sampling the vast conformational space accessible to RNA molecules, particularly during complex transitions such as folding or ligand-induced structural rearrangements. Molecular dynamics (MD) simulations provide a powerful "computational microscope" for observing these processes at atomistic resolution, but their utility is often limited by the timescales required for adequate sampling. [40] While modern hardware can achieve simulations on the order of tens of microseconds, critical biological processes like divalent cation binding and unbinding can occur on millisecond timescales, with complex secondary structure rearrangements taking seconds or longer. [40] This gap between accessible and required simulation times represents a fundamental sampling hurdle that enhanced sampling methods are specifically designed to overcome.

The need to overcome these hurdles is not merely academic; it has direct implications for drug discovery and the understanding of human disease. Many RNAs, including long non-coding RNAs (lncRNAs), form intricate structures essential for their function, and disrupting these structures offers therapeutic potential. [29] [45] For instance, the hepatitis C internal ribosome entry site (HCV-IRES) adopts well-defined folds that are promising targets for antiviral translation inhibitors. [45] Accurately modeling the binding of small molecules to such targets requires a precise understanding of their structural dynamics and the conformational changes induced by or preceding ligand binding. This technical guide will explore the core principles, methodologies, and applications of enhanced sampling techniques for studying complex transitions in RNA, providing researchers with a framework for deploying these advanced computational tools.

Enhanced Sampling Methodologies

Enhanced sampling methods comprise a suite of computational techniques designed to accelerate the exploration of a system's free energy landscape. They work by mitigating the problem of "rare events," where transitions between important states are hindered by high energy barriers that would take an infeasibly long time to cross in a standard MD simulation. These methods can be broadly categorized into those that accelerate the entire system and those that target specific transitions.

Taxonomy of Enhanced Sampling Techniques

• Collective Variable (CV)-Based Methods: These techniques rely on the identification of a small number of order parameters, or Collective Variables (CVs), which are hypothesized to describe the essential physics of the transition of interest.

  • Adaptive Biasing Force (ABF): ABF applies a biasing force to directly counteract the system's mean force along a CV, thereby flattening the free energy landscape and facilitating diffusion across barriers. [45]
  • Metadynamics: This method deposits repulsive Gaussian hills in the free energy landscape along the CVs, which gradually fills energy wells and pushes the system to explore new regions. The sum of the hills provides an estimate of the underlying free energy.
  • Umbrella Sampling: This is a stratification technique where multiple simulations (windows) are run, each with a restraining potential that forces the system to sample a specific region along the CV. The data from all windows are then combined to reconstruct the full free energy profile using methods like the Weighted Histogram Analysis Method (WHAM).

• Alchemical Methods: These methods use a non-physical pathway to compute free energy differences, often by coupling the system to a "alchemical" parameter λ that interpolates between two states (e.g., bound and unbound ligand).

  • Lambda-Adaptive Biasing Force (λ-ABF): A powerful combination that merges alchemical transformation with the ABF method. The alchemical variable λ is treated as a CV, and an adaptive bias is applied to it, bypassing the need for discrete "lambda windows" and enabling more efficient sampling. [45]

• Temperature-Based Methods:

  • Replica Exchange MD (REMD): Also known as parallel tempering, this method runs multiple replicas of the system at different temperatures. Periodically, exchanges between replicas are attempted based on a Metropolis criterion, allowing high-temperature replicas to cross barriers and feed conformational information to low-temperature replicas.

The choice of method depends on the scientific question. Alchemical methods like λ-ABF are particularly well-suited for computing absolute binding free energies, [45] while CV-based methods are ideal for mapping conformational landscapes and mechanisms.

The Critical Role of Collective Variables and Advanced Force Fields

The success of many enhanced sampling methods hinges on the careful selection of CVs. Good CVs should distinguish between all relevant initial, final, and intermediate states and capture the slowest modes of the transformation. Poorly chosen CVs can lead to incomplete or incorrect sampling. Recent advances have integrated machine learning to automate the discovery of optimal CVs from simulation data, thereby reducing human bias and improving sampling efficiency. [45] [40]

Furthermore, the accuracy of any MD simulation is fundamentally limited by the quality of the force field—the empirical energy function that describes interatomic interactions. RNA presents unique challenges due to its highly charged backbone, complex ion atmosphere, and reliance on non-canonical interactions. [45] The use of polarizable force fields, such as AMOEBA, which account for many-body polarization effects, is emerging as a crucial development for accurately modeling RNA and its interactions with ligands and ions. [45] These force fields, while computationally demanding, provide a more realistic description of electrostatics, which is essential for predictive drug discovery campaigns.

Table 1: Key Enhanced Sampling Methods and Their Applications to RNA Systems

Method	Core Principle	Key Advantage	Typical RNA Application
λ-ABF [45]	Applies adaptive bias on an alchemical pathway (λ) connecting states.	Bypasses discrete lambda windows; efficient for binding free energies.	Absolute binding free energy calculation for small molecule inhibitors.
Metadynamics	Fills free energy minima with repulsive bias.	Intuitively explores complex landscapes and reconstructs free energy.	Characterizing ligand-induced conformational changes in riboswitches.
Umbrella Sampling	Restrains simulations to windows along a CV.	Provides high-quality free energy profiles along a pre-defined reaction coordinate.	Calculating potentials of mean force for ion binding or base pairing.
Replica Exchange MD (REMD)	Exchanges configurations between replicas at different temperatures.	Excellent for exploring broad conformational ensembles and overcoming kinetic traps.	Sampling the folding landscape of RNA hairpins and small ribozymes.

Applications to RNA Systems

Enhanced sampling techniques have been successfully applied to illuminate various aspects of RNA dynamics and interactions that were previously intractable. The following diagram illustrates a generalized workflow for applying these methods to a problem like RNA-ligand binding, integrating several of the key techniques discussed.

Characterizing RNA-Ligand Binding

Targeting RNA with small molecules is a promising frontier in drug discovery. A prime example is the inhibition of the HCV-IRES IIa subdomain by 2-aminobenzimidazole derivatives. [45] This system is exceptionally challenging for computational models because the RNA undergoes a dramatic ligand-induced conformational change from an L-shape to an extended form, creating a deep binding pocket. Furthermore, the binding site features a spine of three structurally important magnesium ions that reorganize upon ligand binding, and the inhibitors themselves carry multiple positive charges. [45]

To tackle this, a state-of-the-art protocol was developed combining the polarizable AMOEBA force field with the λ-ABF algorithm and a refined set of restraints. [45] This setup allowed for the quantitative prediction of absolute binding free energies for a series of 19 inhibitors. Crucially, to account for the large-scale conformational change, the free energy difference between the Apo (unbound) and Holo (bound) RNA structures was separately calculated using enhanced sampling simulations driven by machine learning-based CVs. [45] This integrated approach demonstrates how enhanced sampling can successfully manage the dual challenges of predicting accurate affinities and capturing complex RNA structural plasticity.

Mapping Conformational Landscapes and Folding

RNA molecules, including long non-coding RNAs (lncRNAs), are not static but explore expansive conformational landscapes. [29] Enhanced sampling is vital for characterizing these ensembles, especially for transitions between functionally distinct states. For instance, the free energy barrier associated with large-scale conformational changes between Apo and Holo forms of a riboswitch can be captured by combining machine-learned CVs with enhanced sampling. [45] This provides insights into the mechanisms of conformational selection and induced fit.

The diagram below illustrates a hypothetical multi-well free energy landscape of an RNA molecule, highlighting the sampling challenge and the goal of enhanced sampling methods.

Practical Implementation and Protocols

A Protocol for Absolute Binding Free Energy Calculation

Applying enhanced sampling to RNA-ligand systems requires a meticulous protocol. The following methodology, inspired by work on the HCV-IRES system, outlines key steps for calculating an absolute binding free energy (ABFE) for a charged, flexible small molecule binding an RNA with structural complexity. [45]

• System Preparation:

  • Obtain the Holo structure (RNA with bound ligand) from PDB (e.g., 3TZR).
  • Use PDB2PQR or similar tools to add missing hydrogen atoms and assign protonation states.
  • Place the system in a rectangular water box (e.g., TIP3P) with a buffer of at least 10 Ã…. Add neutralizing counterions (Na⁺, Cl⁻) and an additional physiological concentration of salt (e.g., 150 mM).
  • For systems with structural ions, retain crystallographic magnesium ions in the model.

• Equilibration:

  • Perform energy minimization to remove steric clashes.
  • Carry out a multi-stage equilibration using standard MD, first restraining heavy atom positions of the RNA and ligand, then gradually releasing the restraints.

• Enhanced Sampling with λ-ABF and Restraints:

  • Alchemical Path: Define the λ coordinate, where λ=0 corresponds to the fully coupled ligand and λ=1 to the fully decoupled (unbound) state.
  • Restraints: Apply a set of restraints to prevent the ligand from drifting away in the decoupled state and to improve convergence. The recently developed Distance-to-Bound-Configuration (DBC) restraints are particularly effective. [45] These include:
    • Positional restraints: A weak harmonic potential on the ligand's center of mass relative to the binding site.
    • Orientation restraints: A potential based on the quaternion representation to maintain the ligand's bound orientation.
    • Conformational restraints: Potentials to maintain the ligand's internal torsion angles close to their bound configuration.
  • Simulation: Run the λ-ABF simulation, using the Colvars module in Tinker-HP or another compatible MD engine, to compute the free energy change of decoupling the ligand from the solvated RNA complex.

• Accounting for Conformational Change:

  • If the RNA undergoes a significant conformational change upon binding, the free energy difference between the Apo and Holo states must be computed separately. This can be achieved by using a path-collective variable or machine-learned CV in a metadynamics or umbrella sampling simulation to drive the transition. [45]

• Analysis and Validation:

  • The final absolute binding free energy (ΔG°bind) is a combination of the alchemical decoupling work and the conformational free energy penalty.
  • Validate the computed affinities against experimental inhibition constants (Káµ¢ or ICâ‚…â‚€ values). Compare the simulated structural ensemble with available experimental data, such as chemical probing (e.g., SHAPE) or crystallographic B-factors.

The Scientist's Toolkit: Essential Research Reagents and Software

Table 2: Research Reagent Solutions for Enhanced Sampling of RNA Systems

Tool / Reagent	Type	Primary Function	Application Notes
AMOEBA Force Field [45]	Polarizable Force Field	Models electronic polarization and anisotropic electrostatics for more accurate RNA-ligand interactions.	Crucial for charged ligands and RNA systems with structural ions; more computationally expensive.
CHARMM36 & OL3 [40]	Additive Force Fields	Standard, well-tested non-polarizable force fields for RNA molecular dynamics.	Good starting point for many systems; less accurate for highly charged environments than polarizable models.
Tinker-HP [45]	MD Software Package	GPU-accelerated software for performing MD simulations with polarizable force fields like AMOEBA.	Enables long timescale simulations with advanced force fields.
Colvars Module [45]	Software Library	Provides a versatile interface for defining collective variables and running enhanced sampling methods.	Compatible with various MD engines; implements λ-ABF, metadynamics, etc.
PLUMED	Software Library	A comprehensive plugin for enhancing sampling and analyzing MD trajectories, used with many MD codes.	Industry standard for defining complex CVs and applying a wide range of enhanced sampling methods.
Machine-Learned CVs [45]	Computational Method	Discovers optimal low-dimensional descriptors of complex transitions from simulation data (e.g., using autoencoders).	Reduces human bias in CV selection; essential for characterizing large-scale conformational changes.

Enhanced sampling methods have evolved from niche techniques to essential tools for elucidating the complex dynamics of RNA. By overcoming the fundamental timescale limitations of standard MD simulations, they allow researchers to probe the structural transitions, folding pathways, and molecular recognition events that underpin RNA function. The integration of these methods with advanced polarizable force fields and machine learning, particularly for collective variable discovery, is pushing the frontier of what is computationally possible. As these protocols become more robust and accessible, they promise to accelerate the development of RNA-targeted therapeutics by providing unprecedented atomic-level insight into the dynamic landscapes of RNA and their interactions with small molecules. The continued refinement of these approaches, coupled with validation from experimental structural biology, will be crucial for realizing the full potential of RNA as a drug target.

The prevailing paradigm in drug discovery has long operated on a fundamental assumption: stronger binding affinity directly translates to greater functional potency. However, within the realm of RNA-targeted small molecule design, this principle is frequently subverted, creating a significant challenge for rational drug development. This disconnect between affinity and function is not merely an experimental anomaly but points to a deeper, more complex interplay between RNA structure, dynamics, and ligand binding kinetics that traditional metrics fail to capture. Riboswitches, with their well-defined structures and ligand-induced conformational changes regulating gene expression, serve as ideal model systems to dissect this paradox [63]. Recent studies synthesizing computational and experimental methods reveal that the dynamic nature of RNA and the resultant ligand-binding kinetics often prove more critical to biological activity than equilibrium binding affinity alone [63] [45]. This whitepaper examines the core mechanisms of this disconnect, leveraging cutting-edge research to provide a framework for bridging the gap in ligand design strategies.

Core Mechanisms of the Disconnect

The Primacy of Binding Kinetics

A critical insight from recent investigations is that the kinetics of ligand binding—specifically the on-rate (k~on~)—can be a more potent determinant of riboswitch activation than binding affinity. A landmark study on the Fusobacterium ulcerans ZTP riboswitch demonstrated this principle unequivocally. Researchers observed that compound 1, a synthetic analog of the native ligand ZMP, possessed weaker affinity (K~D~ ≈ 600 nM) but was a stronger riboswitch activator (T~50~ ≈ 5.8 µM) than ZMP (K~D~ ≈ 324 nM, T~50~ ≈ 37 µM) [63]. Machine learning-augmented molecular dynamics (MD) simulations were employed to dissect this phenomenon, calculating relative ligand on-rates for synthetic ligands and ZMP. The results confirmed that the values for k~on~ could distinguish between synthetic and cognate ligands and correlated with activation potency, establishing that ligand on-rates, not affinity, drive functional efficacy in this system [63].

RNA Conformational Dynamics and Induced Fit

The functional outcome of ligand binding is profoundly influenced by the inherent dynamics of the RNA target and the specific conformational changes induced by the ligand. RNA exists as an ensemble of metastable states, and ligand binding often stabilizes a particular functional conformation [63]. The disconnect arises when a ligand, despite high affinity, stabilizes a non-productive conformation or fails to induce the conformational change necessary for biological activity. For instance, targeting the hepatitis C internal ribosome entry site (HCV-IRES) IIa subdomain involves a dramatic ligand-induced conformational adaptation from an L-shaped to an extended form, creating a deep pocket [45]. The free energy barrier associated with this RNA conformational change is a key factor determining binding affinity and functional inhibition. Ligands that cannot productively navigate or induce this change may bind tightly but fail to inhibit function effectively [45].

Experimental Evidence and Quantitative Data

The following table summarizes key quantitative findings from recent studies that highlight the affinity-function disconnect across different RNA target systems.

Table 1: Documented Affinity-Function Disconnects in RNA-Targeting Ligands

RNA Target	Ligand	Binding Affinity (K~D~)	Functional Assay Result	Postulated Primary Reason for Disconnect
F. ulcerans ZTP Riboswitch [63]	Cognate (ZMP)	324 nM	T~50~ = 37 µM	Lower ligand on-rate (k~on~)
	Synthetic Compound 1	600 nM	T~50~ = 5.8 µM	Higher ligand on-rate (k~on~)
env8 Cobalamin Riboswitch [64]	Derivative 12	800 nM	Data Not Shown	Steric limitations in cryptic site
	Derivative 29	7 nM	Comparable to native MeCbl	Optimized π-stacking in cryptic site
HCV-IRES IIa [45]	2-Aminobenzimidazole Derivatives	Calculated via ABFE	Antiviral translation inhibition	Inability to overcome conformational change free energy barrier

Table 2: Key Methodologies for Investigating the Disconnect

Methodology	Application	Key Insight Provided
Machine Learning-Augmented MD [63]	Simulating ligand dissociation pathways and calculating kinetics.	Reveals atomistic details of binding mechanisms and estimates on-rates (k~on~).
Absolute Binding Free Energy (ABFE) Calculations [45]	Predicting binding affinities using polarizable force fields (AMOEBA) and enhanced sampling.	Quantitatively captures affinities and the cost of RNA conformational adaptation.
Fluorescence Displacement Assays [64]	High-throughput measurement of binding affinity for ligand libraries.	Enables QSAR analysis to identify chemical features correlating with tight binding.
X-ray Crystallography [63]	Solving high-resolution structures of RNA-ligand complexes.	Informs structure-based design and reveals unique ligand-RNA interactions.

A Toolkit for Bridging the Disconnect: Research Reagent Solutions

Table 3: Essential Research Reagents and Tools for Advanced RNA-Ligand Studies

Research Reagent / Tool	Function/Brief Explanation	Key Application in Disconnect Studies
Polarizable Force Fields (e.g., AMOEBA) [45]	Advanced computational models accounting for RNA electrostatics and polarization.	Essential for accurate absolute binding free energy calculations on charged RNA systems.
Machine Learning-Augmented Enhanced Sampling [63]	Accelerates sampling of rare events like ligand unbinding and conformational changes.	Uncovers key interaction mechanisms and dissociation pathways not visible experimentally.
Lambda-Adaptive Biasing Force (λ-ABF) [45]	An enhanced sampling method for efficient free energy calculation along an alchemical path.	Improves convergence and performance in binding affinity calculations for RNA-ligand complexes.
Small Molecule Microarrays [63]	High-throughput experimental platform for screening RNA-binding molecules.	Identifies initial hit compounds from large libraries for further characterization.
Isothermal Titration Calorimetry (ITC) [63]	Gold-standard experimental method for measuring binding affinity (K~D~) and thermodynamics.	Provides the foundational binding affinity data against which functional activity is compared.
Geometric Deep Learning (e.g., RNABind) [65]	AI framework for predicting RNA-small molecule binding sites from RNA structures.	Informs design by identifying crucial binding residues and predicting ligandable sites.

Integrated Workflow for Analysis

A robust strategy to dissect the affinity-function disconnect requires an integrated, multi-faceted workflow that synergizes computational and experimental approaches. The following diagram visualizes this comprehensive pipeline, from initial design to functional validation.

Diagram 1: Integrated workflow for identifying and resolving the affinity-function disconnect.

Computational Methodologies and Protocols

Absolute Binding Free Energy (ABFE) Calculation Protocol

Accurately predicting binding affinities for RNA-ligand systems is notoriously difficult. A state-of-the-art protocol for Absolute Binding Free Energy (ABFE) calculations, as applied to the HCV-IRES IIa RNA target, involves a multi-stage approach [45]:

• System Preparation: The RNA-ligand complex is solvated in a box of explicit water molecules, and neutralizing counterions (e.g., Mg²⁺, K⁺) are added, carefully considering ion placement, especially around critical structural metal ions.
• Polarizable Force Field: Simulations are conducted using the AMOEBA polarizable force field, which includes atomic induced dipoles to model polarization effects and atomic multipoles for electrostatic anisotropy. This is critical for capturing RNA's highly electronegative surface [45].
• Enhanced Sampling with λ-ABF: The lambda-Adaptive Biasing Force (λ-ABF) method is employed. This technique bypasses the traditional discretization of the alchemical path and uses a continuous alchemical variable, allowing for more efficient sampling [45].
• Application of Restraints: A refined set of restraints—positional, orientational, and conformational (Distance-to-Bound-Configuration, DBC)—is applied to the ligand. This prevents unrealistic ligand drift during the alchemical transformation and improves convergence [45].
• Conformational Change Free Energy: For targets with significant conformational changes between Apo and Holo states, the free energy barrier is estimated separately using enhanced sampling simulations driven by machine-learned collective variables (CVs) [45].

Machine Learning-Augmented Molecular Dynamics

To capture rare events like ligand dissociation, machine learning-augmented enhanced sampling is used [63]. This protocol involves:

• Long-Timescale MD Simulations: Running all-atom MD simulations on specialized hardware (GPUs/supercomputers).
• Collective Variable (CV) Identification: Using machine learning to identify low-dimensional CVs that accurately describe the ligand unbinding pathway.
• Enhanced Sampling: Applying biasing potentials along these ML-informed CVs to accelerate the sampling of the unbinding process, which would otherwise be prohibitively slow.
• Trajectory Analysis: Analyzing the dissociation trajectories to identify key residues involved in the binding mechanism and to compute kinetic parameters like relative on-rates [63].

Visualization of a Kinetic Binding Model

The following diagram illustrates the kinetic and conformational landscape that underpins the affinity-function disconnect, showing how different ligands can lead to divergent functional outcomes despite similar binding affinities.

Diagram 2: Kinetic model for functional and non-functional ligand binding.

The accurate prediction of ribonucleic acid (RNA) structure is a cornerstone for advancing our understanding of cellular function and for designing novel RNA-targeted therapeutics [47]. This in-depth technical guide explores the optimization of computational pipelines for predicting RNA structure and function, a critical capability within the broader context of RNA structure and dynamics research. The inherent flexibility of RNA molecules and the historical scarcity of high-resolution structural data have made this a particularly challenging domain for computational biology [33]. However, recent breakthroughs in feature extraction, model architecture, and the integration of biophysical principles are dramatically accelerating progress. This whitepaper provides a detailed examination of these state-of-the-art methodologies, offering researchers, scientists, and drug development professionals a comprehensive resource for building and refining predictive models that can bridge the gap from sequence to dynamic structure. By synthesizing advancements in deep learning, thermodynamic modeling, and advanced sampling techniques, this guide aims to equip practitioners with the knowledge to construct robust, generalizable, and highly accurate predictive pipelines for RNA.

Current Methodologies in RNA Structure Prediction

The computational prediction of RNA structure is a multi-layered problem, typically approached at the level of secondary (2D) and tertiary (3D) structure. Secondary structure prediction, which identifies canonical base pairs, has evolved from physics-based thermodynamic models to deep learning (DL) methods. Traditional algorithms, such as those implemented in RNAfold and RNAstructure, use dynamic programming to identify the Minimum Free Energy (MFE) structure based on nearest-neighbor thermodynamic parameters [47] [66]. While effective for simple structures, these methods often struggle with complex features like pseudo-knots and non-canonical base pairs. In response, DL models like SPOT-RNA, UFold, and MXfold2 have been developed, leveraging convolutional neural networks and image-like representations of sequences to achieve higher accuracy [66] [37].

Tertiary structure prediction presents a greater challenge due to the increased conformational space and limited experimental data. Early de novo methods, such as FARFAR2 and SimRNA, relied on extensive computational sampling and energy-based scoring [33]. The success of deep learning in protein structure prediction has since catalyzed the development of end-to-end DL models for RNA 3D structure. A significant innovation has been the integration of RNA language models (LMs). For instance, RhoFold+ employs a transformer-based network pre-trained on millions of RNA sequences, allowing it to extract evolutionarily informed features directly from single sequences or multiple sequence alignments (MSAs) to predict all-atom structures [33]. These methodologies underscore a paradigm shift towards data-driven approaches that learn the complex mapping from sequence to structure.

Table 1: Comparison of Representative RNA Structure Prediction Tools

Tool Name	Prediction Type	Core Methodology	Key Input Features	Notable Strengths
BPfold [66]	Secondary Structure	Deep Learning & Base Pair Motif Energy	RNA Sequence, Base Pair Motif Energy	High generalizability on unseen RNA families
RhoFold+ [33]	Tertiary Structure	Language Model & Transformer	RNA Sequence, MSA, Secondary Structure	Fully automated end-to-end pipeline; high accuracy on RNA-Puzzles
gRNAde/RiboPO [67]	Inverse Folding	Geometric Graph Neural Network	RNA 3D Backbone Structure	Conditions design on 3D structure; enables multi-state design
COMET [68]	Delivery Particle Design	Transformer-based Model	Chemical Components of LNPs	Accelerates discovery of RNA delivery formulations

Optimized Feature Extraction for RNA

The performance of any predictive model is fundamentally tied to the quality and relevance of its input features. For RNA, moving beyond the raw nucleotide sequence to incorporate biophysical and evolutionary priors is a key differentiator for modern pipelines.

Leveraging Base Pair Motif Energy

A major limitation of purely data-driven DL models is their tendency to overfit to RNA families seen during training, leading to poor performance on novel, out-of-distribution sequences [66]. To mitigate this, the BPfold framework introduces a physics-informed feature known as base pair motif energy. A base pair motif is defined as a canonical base pair (e.g., A-U, G-C) along with its locally adjacent bases, which dominate the local structural context. The BPfold team constructed a comprehensive library by enumerating the complete space of three-neighbor base pair motifs and computing their thermodynamic energy through de novo modeling of tertiary structures using the BRIQ method [66]. This energy map, which provides a thermodynamic prior for any potential base pair in an input RNA sequence, is then integrated with the sequence embedding within the model's neural network. This approach enriches the data at the base-pair level, mitigating the problem of insufficient structural data and significantly improving model generalizability, as demonstrated through cross-family validation [66].

Evolutionary and Structural Embeddings

For 3D structure prediction, constructing informative feature embeddings is critical. RhoFold+ exemplifies this by leveraging a large RNA language model, RNA-FM, which is pre-trained on ~23.7 million RNA sequences [33]. This model generates deep, evolutionarily informed embeddings that capture latent structural and functional constraints from the sequence alone. These embeddings are then combined with features derived from MSAs and predicted secondary structures. The integration of these diverse feature streams—sequence embeddings, co-evolutionary signals from MSAs, and base-pairing constraints—provides a rich informational foundation for the subsequent structure module to generate accurate 3D atomic coordinates [33].

Advanced Model Architectures and Training Paradigms

Novel neural network architectures and training strategies are pushing the boundaries of what is possible in RNA informatics, moving beyond pure prediction to generative design.

The RiboPO Framework for Inverse Folding

The inverse folding problem—designing an RNA sequence that will fold into a specific 3D structure—is a critical task for synthetic biology and therapeutics. Current geometric deep learning models like gRNAde have shown promise but often produce sequences that, while structurally accurate, are thermodynamically unstable in solution [67]. The RiboPO framework addresses this multi-objective challenge through a novel training paradigm called Reinforcement Learning from Physical Feedback (RLPF).

RiboPO fine-tunes a base model like gRNAde using Direct Preference Optimization (DPO). The core innovation lies in how preference pairs are constructed for training. For a given target backbone, the model generates multiple candidate sequences. These sequences are then evaluated using a composite reward function that couples:

• Global 3D Fidelity: Assessed via metrics like pLDDT.
• Thermodynamic Stability: Proxied by the Minimum Free Energy (MFE) of the sequence's predicted secondary structure [67].

Candidate sequences that demonstrate a superior balance of high structural accuracy and low MFE (indicating stability) are labeled as "preferred" over those that do not. By training on these preferences, the RiboPO policy learns to generate sequences that are not only structurally compliant but also ensemble-robust, significantly improving practical design success [67].

Table 2: Key Research Reagent Solutions for RNA Computational Workflows

Reagent / Resource	Type	Primary Function in the Pipeline
Base Pair Motif Library [66]	Computational Database	Provides pre-computed thermodynamic energy priors for local base pair contexts, enhancing model generalizability.
RNA-FM Language Model [33]	Pre-trained Model	Generates evolutionarily informed sequence embeddings for use as input features in 3D structure prediction.
DESRES-RNA Force Field [49]	Molecular Dynamics Parameter Set	Enables highly accurate, all-atom molecular dynamics simulations of RNA for folding validation and refinement.
AMOEBA Polarizable Force Field [45]	Advanced Molecular Model	Provides a more accurate treatment of electrostatics and polarization for binding affinity calculations in drug design.
EternaFold [67]	Secondary Structure Prediction	Used as a proxy to evaluate the thermodynamic stability and self-consistency of designed RNA sequences.

Architectures for Binding Affinity Prediction

Accurately predicting the binding affinity of a small molecule to an RNA target is paramount for drug discovery. State-of-the-art approaches now combine advanced physical models with enhanced sampling. For a riboswitch-like target, researchers have developed a protocol using the AMOEBA polarizable force field, which explicitly models many-body polarization effects crucial for RNA's highly charged environment [45]. This is coupled with the lambda-Adaptive Biasing Force (lambda-ABF) method, an alchemical free energy technique that avoids discrete lambda windows and allows for efficient sampling. Furthermore, to capture large-scale conformational changes in the RNA upon binding, machine-learned collective variables (CVs) are used within enhanced sampling simulations to overcome the associated free energy barriers [45]. This integrated pipeline achieves highly predictive absolute binding free energies, a significant advance for RNA-targeted drug discovery.

Experimental Protocols and Validation

Robust experimental validation is essential to benchmark the performance of any predictive pipeline. The following protocols detail standard methodologies for training and evaluation.

Protocol: Training a Generalizable Secondary Structure Predictor

This protocol is based on the methodology described for BPfold [66].

• Base Pair Motif Library Construction:

  • Define all possible three-neighbor base pair motifs, categorized as hairpin (BPMiH), inner chainbreak (BPMiCB), and outer chainbreak (BPMoCB) motifs.
  • For each motif, use a de novo RNA tertiary structure modeling method (e.g., BRIQ, which employs Monte Carlo sampling) to generate candidate 3D structures.
  • Calculate a normalized energy score (e.g., the BRIQ energy score, which combines physical and statistical energy) for each motif's tertiary structure and store it in the library.

• Data Preparation & Feature Generation:

  • Curate a diverse training dataset of RNA sequences with known secondary structures (e.g., from ArchiveII, bpRNA).
  • For each sequence in the training set, generate two L x L energy maps (Mμ and Mν) by querying the pre-computed base pair motif library for every potential nucleotide pair (i, j).

• Model Training with Base Pair Attention:

  • Implement a deep neural network featuring a custom-designed Base Pair Attention Block. This block should combine transformer and convolutional layers.
  • The model takes two primary inputs: the raw RNA sequence (one-hot encoded or embedded) and the base pair motif energy maps.
  • The attention mechanism is designed to effectively integrate and learn the relationship between the sequence information and the thermodynamic energy prior.
  • Train the model using standard supervised learning, with the known secondary structure as the target.

• Validation:

  • Perform both sequence-wise and family-wise cross-validation on benchmark datasets (e.g., Rfam, PDB).
  • Compare performance against state-of-the-art non-ML and DL methods using metrics like F1-score for base pair prediction to demonstrate superior accuracy and generalizability on unseen RNA families.

Protocol: Absolute Binding Free Energy Calculation for an RNA-Ligand Complex

This protocol outlines the advanced computational workflow for predicting small molecule binding affinities to RNA targets [45].

• System Preparation:

  • Obtain the high-resolution 3D structure of the RNA-ligand complex (Holo state) from PDB or modeling.
  • Parameterize the ligand and the RNA using the AMOEBA polarizable force field.
  • Solvate the system in a water box and add neutralizing counterions, paying special attention to the inclusion of divalent ions (e.g., Mg²⁺) if they are structurally important.

• Equilibration and Conformational Sampling:

  • Perform extensive molecular dynamics (MD) simulation of the Holo complex to equilibrate the system and sample relevant conformational states.
  • If a significant conformational change exists between the Apo and Holo RNA states, define Machine-Learned Collective Variables (CVs) that capture this transition.
  • Run enhanced sampling simulations (e.g., metadynamics) using these CVs to estimate the free energy difference associated with the RNA's conformational change.

• lambda-ABF Simulation:

  • Apply the lambda-ABF method, which combines alchemical transformation with an adaptive biasing force.
  • Use a refined set of restraints—including positional, orientational, and distance-to-bound-configuration (DBC) restraints—to maintain the ligand in the binding site while allowing necessary fluctuations.
  • Run the simulation to achieve sufficient sampling along the alchemical and conformational degrees of freedom.

• Free Energy Analysis:

  • Compute the absolute binding free energy (ABFE) directly from the lambda-ABF simulation data.
  • Validate the protocol by calculating ABFEs for a series of known inhibitors and comparing the predicted affinities with experimental values (e.g., ICâ‚…â‚€ or Kd), aiming for quantitative agreement.

Visualization of Workflows

The following diagrams illustrate the logical flow of two optimized pipelines described in this guide.

BPfold Model Architecture and Training

Diagram Title: BPfold Model Architecture

RiboPO Inverse Fining Training with Physical Feedback

Diagram Title: RiboPO Training with Physical Feedback

The field of RNA computational biology is undergoing a rapid transformation, driven by the sophisticated integration of feature engineering, novel model architectures, and biophysical principles. As this guide has detailed, optimizing predictive pipelines requires moving beyond purely data-driven models. The most significant gains in accuracy and generalizability are now achieved by incorporating physical priors—such as base pair motif energies and polarizable force fields—and by adopting multi-objective optimization frameworks that balance structural fidelity with thermodynamic stability. These advanced pipelines are no longer just prediction tools; they are becoming essential for the rational design of functional RNA molecules and therapeutics, as evidenced by their growing impact on RNA vaccine delivery optimization [68] and small-molecule drug discovery [45]. The continued development of standardized, large-scale datasets [37] will further fuel this progress, enabling more robust benchmarking and training. For researchers, the future lies in building and refining these integrated pipelines, which hold the key to unlocking a deeper, more dynamic understanding of RNA biology and its vast therapeutic potential.

Benchmarking Accuracy: Validating and Comparing RNA Structure Predictions

Ribonucleic Acids (RNAs) are performing a broad range of essential molecular functions in cells, many of which rely on intricate folding properties of the molecule [69]. The intrinsic dynamic flexibility and pronounced conformational heterogeneity of RNA endow it with diverse functional capabilities [13]. Establishing gold standard validation techniques is therefore paramount for elucidating the structure-function relationships that underpin RNA's roles in both normal physiological processes and disease pathways. Traditional experimental methods, including NMR, X-ray crystallography, and cryo-electron microscopy, encounter considerable limitations in resolving the complex conformational ensembles of RNA [13]. This technical guide provides a comprehensive framework for validating RNA structures and dynamics, encompassing both established experimental approaches and emerging computational technologies that offer complementary insights for the research and drug development communities.

Foundational Principles of RNA Structure

RNA Structural Hierarchy

RNA architecture is shaped by its secondary structure composed of stems, stacked canonical base pairs, and enclosing loops [69]. While stems are precisely captured by free-energy models, loops composed of non-canonical base pairs are not, nor are distant interactions linking together those secondary structure elements (SSEs) [69]. The single-stranded linear RNA molecule first folds onto itself and forms double-stranded regions by additional hydrogen bonds, mainly through Watson-Crick base pairings (A-U and G-C) and possible G-U pairing [70].

Table 1: Fundamental Components of RNA Architecture

Structural Element	Description	Functional Significance
Stems	Double-stranded regions formed by stacking of consecutive base pairs	Provide structural stability; precisely modeled by free-energy models [69]
Hairpin Loops	Single-stranded regions that a stem ends into	Often form sophisticated 3D motifs for molecular shaping [70]
Bulge Loops	Single-stranded regions interrupting a stem on one side	Create flexibility and sophisticated interaction patterns [70]
Interior Loops	Single-stranded regions interrupting a stem on both sides	Sites for potential protein/RNA binding [70]
Multi-junction Loops	Single-stranded regions where more than two stems meet	Critical for complex architectural organization [70]
Pseudoknots	Structurally complex motifs with nested base pairs	Represent topologically intricate folding patterns [70]

Graph-Based Representations of RNA Structure

Representing RNA structure as a graph has recently allowed expansion of work to pairs of SSEs, uncovering a hierarchical organization of 3D modules [69]. RNA 2D structure graphs are directed edge-labelled graphs where each node represents a nucleotide, and each edge represents an interaction (base pair or backbone) with labels according to the Leontis-Westhof geometric classification [69]. In dual graph representations, stems in the secondary structure are represented as nodes and edges come from single-stranded regions, enabling the handling of complex topologies including pseudoknots [70].

Computational Validation Techniques

Mining Structural Databases for RNA Modules

RNA modules are small and densely connected base pair patterns observed in various molecules, sometimes in multiple locations [69]. The conservation of these modules suggests evolutionary pressure to preserve specific interaction patterns. Efficient algorithms to compute maximal isomorphisms in edge-colored graphs have been developed to identify RNA modules considerably faster than previous approaches, enabling the discovery of large shared sub-structures spanning hundreds of nucleotides between ribosomes of different species [69].

Table 2: Computational Methods for RNA Structure Analysis

Method/Algorithm	Primary Function	Technical Basis	Applications
DynaRNA [13]	Dynamic RNA conformation ensemble generation	Denoising Diffusion Probabilistic Model (DDPM) with Equivariant Graph Neural Network (EGNN)	Generating conformational ensembles; capturing rare excited states; de novo folding
CaRNAval [69]	Finding identical interaction networks between RNAs	Graph matching of RNA secondary structure graphs	Identifying Recurrent Interaction Networks (RINs)
RNA 3D Motif Atlas [69]	Database of conserved 3D geometries	Collection of experimentally determined RNA modules	Structure prediction and design
Maximal Isomorphism Algorithm [69]	Computing maximal common subgraphs between RNA graphs	Edge-colored graph comparison	Building comprehensive catalog of RNA modules

DynaRNA: A Diffusion-Based Approach for Conformational Ensemble Validation

DynaRNA employs a denoising diffusion probabilistic model (DDPM) with an equivariant graph neural network (EGNN) to directly model RNA 3D coordinates, enabling rapid exploration of RNA conformational space [13]. The architecture implements a partial noising scheme where diffusion is applied only to an intermediate noise step rather than full corruption, creating a tunable balance between preserving structural information and introducing stochastic variability for sampling diverse conformations [13]. Validation studies demonstrate that DynaRNA effectively generates tetranucleotide ensembles with lower intercalation rates than molecular dynamics simulations and can capture rare excited states of complex RNA systems like HIV-1 TAR [13].

DynaRNA Architecture Workflow: The computational pipeline for generating RNA conformational ensembles using diffusion models and equivariant graph neural networks.

Experimental Methodologies for Structure Validation

Biophysical Approaches for Structural Determination

Traditional experimental techniques provide essential validation for RNA structures but face significant challenges in resolving dynamic conformational ensembles [13]. These methods often average signals from multiple conformations, making it difficult to accurately capture RNA's highly heterogeneous structural characteristics [13]. The limitations of individual techniques necessitate a multi-modal approach to validation.

Table 3: Experimental Techniques for RNA Structure Validation

Technique	Spatial Resolution	Temporal Resolution	Key Applications	Limitations
X-ray Crystallography	Atomic (~1-3 Ã…)	Static snapshots	High-resolution structure determination of stable conformations [13]	Requires crystallization; limited for dynamic ensembles [13]
NMR Spectroscopy	Atomic (~1-5 Ã…)	Millisecond-second dynamics	Solution-state structures; dynamics; local conformational changes [13]	Limited by molecular size; signal overlap
Cryo-EM	Near-atomic (~3-5 Ã…)	Static snapshots	Large RNA-protein complexes; flexible structures [13]	Heterogeneity challenges; resolution variability
Chemical Probing	Nucleotide-level	Seconds-minutes	Secondary structure mapping; ligand binding sites	Indirect structural information
SAXS	Low (~10-100 Ã…)	Milliseconds-seconds	Global shape and dimensions in solution	Low resolution; ensemble averaging

RNA-Seq Experimental Design for Functional Validation

RNA sequencing (RNA-Seq) serves as a powerful tool for validating functional aspects of RNA structure, particularly when investigating drug effects, mode-of-action, and treatment responses [71]. A thorough and careful experimental design is the most crucial aspect of ensuring meaningful results that can validate structural hypotheses [71]. Key considerations include:

Sample Size and Replication

Statistical power refers to the ability to identify genuine differential gene expression in naturally variable data sets [71]. Biological replicates are independent samples for the same experimental group that account for natural variation between individuals, tissues, or cell populations [71]. At least 3 biological replicates per condition are typically recommended, though between 4-8 replicates per sample group cover most experimental requirements in drug discovery settings [71].

Library Preparation Strategies

The selection of library preparation method depends on the research question and RNA biotype of interest. Stranded libraries are preferred for better preservation of transcript orientation information and long non-coding RNAs [72]. Ribosomal depletion techniques, including precipitating bead methods and RNaseH-mediated degradation, enhance cost-effectiveness by reducing ribosomal RNA content from approximately 80% of cellular RNA to focus sequencing on target transcripts [72].

Integrated Validation Workflow

A robust validation strategy integrates computational predictions with experimental verification through an iterative cycle. The workflow begins with computational structure prediction, proceeds through experimental validation using biophysical techniques, incorporates functional assessment via RNA-Seq, and concludes with refinement of computational models based on experimental findings.

Integrated Validation Workflow: The iterative cycle of computational prediction, experimental validation, functional assessment, and model refinement that establishes gold standard RNA structures.

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Research Reagents for RNA Structure Validation

Reagent/Category	Function	Application Notes
RNA Stabilization Reagents (e.g., PAXgene) [72]	Preserve RNA integrity during sample collection	Critical for blood samples; prevents degradation
Stranded Library Prep Kits [72]	Create strand-aware sequencing libraries	Preserves transcript orientation information
Ribosomal Depletion Kits [72]	Remove rRNA to enrich coding RNA	Increases cost-effectiveness; RNaseH methods show good reproducibility
Spike-in Controls (e.g., SIRVs) [71]	Internal standards for normalization	Assess technical variability; quality control
Chemical Probing Reagents (e.g., DMS, SHAPE)	Map RNA secondary structure	Provides nucleotide-level structural information
RNase Inhibitors	Prevent RNA degradation during processing	Essential for working with low-abundance transcripts
Quality Assessment Tools (Bioanalyzer/TapeStation) [72]	Evaluate RNA Integrity Number (RIN)	RIN >7 generally indicates sufficient integrity for sequencing

Gold standard validation of RNA structure requires integration of multiple computational and experimental approaches to overcome the limitations inherent in any single technique. Computational methods like DynaRNA enable rapid exploration of conformational space and identification of rare states [13], while graph-based approaches facilitate mining of structural databases to identify recurrent modules [69]. Experimental techniques provide essential physical constraints and validation, with RNA-Seq offering functional context in drug discovery applications [71]. The emerging paradigm combines these approaches through iterative cycles of prediction and validation, establishing rigorous standards for understanding RNA structural dynamics in both basic research and therapeutic development. As these methodologies continue to mature, they promise to unlock deeper insights into the complex relationship between RNA structure, dynamics, and biological function.

Understanding RNA three-dimensional (3D) structure is fundamental to deciphering its diverse biological functions, from gene regulation to therapeutic applications. Community-wide blind assessments have emerged as the gold standard for objectively evaluating the state of computational structure prediction methods. Two primary initiatives—RNA-Puzzles and the Critical Assessment of Structure Prediction (CASP)—provide rigorous frameworks for comparing RNA structure modeling techniques through blinded experiments. These benchmarks reveal the capabilities and limitations of current methodologies, drive algorithmic innovations, and establish best practices for the research community. This whitepaper examines the performance outcomes, methodological insights, and evolving challenges identified through these collaborative efforts, providing researchers and drug development professionals with a comprehensive technical reference for navigating the landscape of RNA structural bioinformatics.

Benchmarking Initiatives: Experimental Frameworks and Protocols

RNA-Puzzles: A Community-Wide Experiment for RNA 3D Structure Prediction

RNA-Puzzles is a dedicated collective experiment established specifically for blind assessment of RNA 3D structure prediction methods [73] [74]. The project operates through a carefully designed protocol:

• Confidential Sequence Distribution: Experimental groups provide RNA sequences from solved but unpublished structures to organizers, who confidentially distribute them to participating modeling groups [31].
• Prediction Timeframes: Participating groups have defined submission windows—typically 3-4 weeks for human-expert predictions and 48 hours for fully automated server predictions [74].
• Double-Blind Design: The process is double-blind; modeling groups don't know the identity of the structural biologists, and the structures are absent from public databases during prediction [31].
• Standardized Assessment: Once experimental structures are published, all predictions are systematically evaluated against reference structures using multiple quantitative metrics [73].

The initiative aims to determine capabilities and limitations of computational methods, identify bottlenecks hindering progress, and guide users in selecting appropriate tools for specific problems [74].

CASP: Comprehensive Structure Prediction Assessment

CASP (Critical Assessment of Structure Prediction) is a broader community-wide experiment assessing structure prediction methods for proteins and nucleic acids [75]. While historically protein-focused, CASP has expanded to include RNA assessment categories:

• Biennial Evaluation Cycle: CASP runs every two years, with CASP16 conducted in 2024 [75].
• Diverse Prediction Categories: The RNA assessment occurs within the "Nucleic acid (NA) structures and complexes" category, which includes RNA and DNA single structures and complexes with proteins [75].
• Collaborative Integration: CASP maintains close collaboration with RNA-Puzzles, leveraging its expertise for RNA-specific assessments [75].
• Expanding Scope: Recent CASP experiments have introduced new categories including macromolecular conformational ensembles and integrative modeling with sparse experimental data [75].

Quantitative Performance Assessment

RNA-Puzzles Round V: Large-Scale Assessment Outcomes

The most recent comprehensive analysis from RNA-Puzzles Round V evaluated predictions from 18 groups for 23 diverse RNA structures across 5 functional categories [31]. The assessment revealed substantial variation in prediction accuracy:

Table 1: RNA-Puzzles Round V Performance Summary (23 Puzzles)

RNA Functional Category	Number of Puzzles	Best Achieved RMSD (Ã…)	Key Challenges Identified
RNA Elements	4	~5.0 (for most difficult)	Non-Watson-Crick pairs, helical stacking
Aptamers	2	Variable across targets	Ligand-binding region accuracy
Viral Elements	4	Variable across targets	Overall architecture reproduction
Ribozymes	5	Variable across targets	Catalytic residue positioning
Riboswitches	8	Variable across targets	Ligand interaction accuracy

The analysis demonstrated that while overall folds could often be captured, precise reproduction of complex structural features remained challenging [31]. Three of the top four modeling groups in this RNA-Puzzles round also ranked among the top four in the CASP15 contest, indicating consistency in methodological performance across different assessment frameworks [31].

CASP15 Ensemble Modeling Challenges

CASP15, conducted in 2022, introduced a novel category for evaluating conformational ensemble predictions for both proteins and RNA [76]. The results provided important insights into the state of RNA ensemble modeling:

• Partial Success: Participants achieved full or partial success in reproducing ensembles for four of the nine targets [76].
• Experimental Data Integration: An experimentally derived flexibility ensemble enabled identification of a single accurate RNA structure model [76].
• Persistent Challenges: Difficulties included handling sparse or low-resolution experimental data and limited methods for modeling RNA/protein complexes [76].

Standardized Assessment Metrics

Both initiatives employ multiple complementary metrics to evaluate prediction quality:

Table 2: Standardized Assessment Metrics for RNA Structure Prediction

Metric	Description	Strengths	Limitations
RMSD (Root Mean Square Deviation)	Measures average distance between corresponding atoms in superimposed structures [73]	Intuitive geometric measure	Global measure insensitive to local accuracy
DI (Deformation Index)	Incorporates base-pair and base-stacking prediction accuracy [73]	RNA-specific interaction fidelity	Less intuitive geometric interpretation
INF (Interaction Network Fidelity)	Quantifies accuracy of non-canonical interaction networks [31]	Direct functional relevance	Complex calculation
TM-score	Scale-invariant measure of global fold similarity [31]	Less length-dependent than RMSD	May overlook local details
lDDT (local Distance Difference Test)	Emphasizes local structural accuracy [31]	Local structure emphasis	May underestimate global fold quality
Clash Score	Measures steric clashes per thousand residues [73]	Stereochemical quality assessment	Doesn't address fold correctness

Methodological Insights and Bottlenecks

Critical Challenges in RNA Structure Prediction

Analysis of benchmark performances has identified consistent methodological challenges:

• Helix Identification and Assembly: Correct identification of helix-forming pairs and proper coaxial stacking between helices represents a fundamental bottleneck [31].
• Non-Watson-Crick Interactions: Accurate prediction of non-canonical base pairs and tertiary interactions remains particularly challenging [31].
• Entanglement Avoidance: Modelers frequently struggle to avoid topological knots and entanglements during helix assembly [31].
• Ligand Binding Sites: Predicting RNA-ligand interaction geometries proves difficult, especially for aptamers and riboswitches [31].
• Conformational Dynamics: Capturing RNA structural heterogeneity and multiple functional states presents significant challenges [76] [13].

Emerging Methodological Trends

Recent benchmarks reveal shifting methodological approaches:

• Deep Learning Integration: While only one of 18 modeling approaches in recent RNA-Puzzles was primarily deep learning-based [31], new specialized architectures like DynaRNA show promise for conformational ensemble generation [13].
• Large Language Model Applications: RNA-specific large language models (RNA-LLMs) are being explored for secondary structure prediction, though comprehensive benchmarking reveals significant challenges in low-homology scenarios [77].
• Hybrid Approaches: Successful predictors often combine multiple methodologies, integrating comparative modeling with physics-based refinement and experimental constraints [31].

Experimental Protocols and Workflows

Standardized RNA Structure Prediction Pipeline

The following diagram illustrates the generalized workflow for community RNA structure prediction benchmarks:

Specialized Workflow for Conformational Ensemble Prediction

For the emerging category of conformational ensemble prediction, specialized workflows have been developed:

Table 3: Essential Research Resources for RNA Structure Prediction Benchmarking

Resource Category	Specific Tools/Platforms	Primary Function	Access Information
Benchmarking Platforms	RNA-Puzzles, CASP-NA category	Community-wide blind assessment	rnapuzzles.org, predictioncenter.org
Assessment Metrics	RNA-Puzzles Toolkit, CASP-RNA evaluation suite	Standardized structure comparison	RNA-Puzzles Toolkit, CASP-RNA GitHub
Specialized Modeling Tools	SimRNA, RNAComposer, DynaRNA	Automated 3D structure prediction	Publicly available (varies by tool)
Reference Datasets	RNA3DB, RNANet, RNAsolo	Curated RNA structural data	Public databases
Large Language Models	RNA-FM, RNABERT, ERNIE-RNA, RiNALMo	RNA sequence representation learning	GitHub repositories (varies by model)

Future Directions and Strategic Recommendations

Emerging Priorities in RNA Structure Assessment

Based on benchmark performances, several strategic priorities have emerged:

• Ensemble Modeling Advancement: Methods for predicting conformational ensembles require further development, particularly for capturing functional states and dynamics [76] [13].
• RNA-Protein Complex Prediction: Current methods show limited effectiveness for modeling RNA-protein complexes, representing a critical frontier [76].
• Hybrid Method Integration: Combining computational predictions with sparse experimental data (NMR, SAXS, crosslinking) shows promise for challenging systems [75].
• Generalization Capabilities: Improving performance on novel folds with low homology to known structures remains a priority [77] [31].
• Standardized Benchmarking Infrastructure: Initiatives like the comprehensive RNA 3D structure-function benchmarking suite (rnaglib) provide essential infrastructure for reproducible model comparison [58].

Recommendations for Research Applications

For researchers and drug development professionals applying these methods:

• Method Selection: Choose prediction methods based on target RNA type and required accuracy, recognizing that performance varies substantially across RNA functional categories [31].
• Quality Assessment: Utilize multiple complementary metrics (RMSD, INF, lDDT) for comprehensive model evaluation, as no single metric fully captures biological relevance [31].
• Experimental Integration: Incorporate available experimental constraints to improve prediction accuracy, especially for complex systems [76].
• Ensemble Considerations: For functional studies, consider conformational ensembles rather than single structures, particularly for dynamic RNA systems [13].

Community-wide benchmarks through RNA-Puzzles and CASP provide indispensable frameworks for objectively assessing RNA structure prediction methodologies. Systematic evaluation across diverse RNA targets has revealed both significant progress and persistent challenges, with accurate prediction of non-canonical interactions, ligand binding sites, and conformational ensembles representing key frontiers. As methodological approaches evolve—particularly through deep learning and hybrid methodologies—these community benchmarks will continue to drive innovation, establish performance standards, and guide researchers in selecting appropriate tools for specific biological and therapeutic applications. The ongoing development of more sophisticated assessment metrics, specialized categories for complexes and ensembles, and standardized benchmarking infrastructure will further enhance our ability to decipher the complex relationship between RNA structure and function.

Ribonucleic acid (RNA) structure determination is a cornerstone of molecular biology, crucial for understanding gene regulation, cellular functions, and enabling advances in RNA-targeted therapeutics and synthetic biology. The conformational flexibility of RNA molecules has made experimental determination of their three-dimensional structures challenging, with RNA-only structures comprising less than 1.0% of the Protein Data Bank (PDB) as of December 2023 [33]. This scarcity has driven the development of computational methods for predicting RNA structures from sequence data, which generally fall into three categories: template-based modeling, physics-based (thermodynamic) approaches, and deep learning methods [33]. This review provides a comprehensive technical comparison of these methodologies, examining their underlying principles, performance benchmarks, and experimental protocols within the broader context of RNA structure and dynamics research.

Methodological Foundations

Template-Based Modeling Approaches

Template-based methods rely on comparative modeling principles, predicting structures by transferring spatial arrangements from known RNA structures with sequence similarity.

Core Protocol: The standard workflow begins with identifying template structures through database searches using tools like BLAST or Infernal. The query sequence is then aligned to template structures, followed by model building through spatial restraint satisfaction and loop refinement. Finally, model selection and validation are performed using statistical potential functions [33].

Key Limitations: Template-based methods are fundamentally constrained by the limited size of RNA structure databases. The BGSU representative sets contain only 782 unique sequence clusters from 5,583 RNA chains after redundancy reduction, creating significant coverage gaps [33]. Performance degrades sharply when sequence similarity to known structures falls below 30%, making these methods unsuitable for novel RNA folds.

Physics-Based (Thermodynamic) Methods

Physics-based approaches utilize energy minimization principles, predicting structures by identifying conformations with the lowest free energy states.

Core Protocol: These methods employ nearest-neighbor parameters derived from experimental measurements of oligonucleotide thermodynamics. The folding landscape is explored using algorithms like stochastic gradient descent or Monte Carlo sampling. Energy evaluation incorporates both canonical and non-canonical base pairing energies, with subsequent structure refinement through molecular dynamics simulations [66].

Implementation Examples: ViennaRNA RNAfold, RNAstructure, and EternaFold represent prominent implementations of this paradigm [66]. These methods effectively predict simple secondary structures with nested base pairs but struggle with complex topological features like pseudoknots and non-canonical interactions.

Deep Learning Approaches

Deep learning methods leverage neural networks to learn the mapping between RNA sequences and their structures directly from data, either through evolutionary information or sequence patterns.

Core Protocol: The standard pipeline involves feature extraction through multiple sequence alignments (MSAs) or language model embeddings. These features are processed through deep neural architectures such as transformers or convolutional networks. The model then generates geometric constraints (distances, angles) or directly outputs atomic coordinates, followed by full-atom model reconstruction and refinement [33].

Architectural Innovations: RhoFold+ exemplifies modern implementations, integrating an RNA language model (RNA-FM) pretrained on ~23.7 million sequences with a transformer network (Rhoformer) and structure module employing invariant point attention for coordinate optimization [33]. BPfold introduces base pair motif energy, computing thermodynamic energy for complete three-neighbor base pair motifs and integrating this physical prior through a custom base pair attention mechanism [66].

Table 1: Key Methodological Characteristics

Method Category	Theoretical Basis	Representative Tools	Primary Outputs
Template-Based	Comparative homology modeling	ModeRNA, RNAbuilder	Full-atom 3D models
Physics-Based	Free energy minimization	RNAfold, RNAstructure, EternaFold	Secondary structure & 3D models
Deep Learning	Pattern recognition from data	RhoFold+, BPfold, AlphaFold3	Geometric constraints or direct coordinates

Performance Benchmarking

Experimental Design for Method Evaluation

Rigorous benchmarking requires specialized datasets and standardized evaluation metrics. The RNA-Puzzles competition has served as a community-wide standard for assessing 3D structure prediction methods, providing blind testing on experimentally determined structures [33]. For comprehensive evaluation, benchmarks should implement multiple splitting strategies:

• Sequence-wise splitting: Standard random division by individual sequences
• Family-wise splitting: Separation by RNA families to test generalizability
• Structural similarity splitting: Clustering by structural similarity (TM-score > 0.5) to prevent data leakage [58]

Standard evaluation metrics include Root Mean Square Deviation (RMSD) for global structure comparison, Template Modeling Score (TM-score) for topological similarity, and Local Distance Difference Test (LDDT) for local atomic accuracy [33].

Quantitative Performance Comparison

Comprehensive benchmarking on RNA-Puzzles demonstrates the superior performance of deep learning methods. RhoFold+ achieved an average RMSD of 4.02 Ã… across 24 single-chain RNA targets, outperforming the second-best method (FARFAR2) by 2.30 Ã… [33]. On 17 of 24 targets, RhoFold+ achieved RMSD values below 5 Ã…, with an average TM-score of 0.57 compared to 0.41-0.44 for other top performers [33].

For secondary structure prediction, BPfold demonstrates exceptional generalizability in cross-family validation on Rfam datasets (10,791 RNAs), addressing a critical limitation of earlier deep learning methods that suffered performance degradation on unseen RNA families [66].

Table 2: Performance Benchmarking on Standardized Tests

Method Category	Representative Method	Average RMSD (Ã…)	Average TM-score	Generalizability Assessment
Physics-Based	FARFAR2 (top 1%)	6.32	0.41-0.44	Moderate (energy functions transferable)
Template-Based	Best single template	>10.0 (estimated)	0.42 (average)	Poor (requires structural templates)
Deep Learning	RhoFold+	4.02	0.57	High (R²=0.23 TM-score vs similarity)
Hybrid DL/Physics	BPfold	N/A (secondary structure)	N/A (secondary structure)	High (effective cross-family prediction)

Technical Protocols

Deep Learning Workflow: RhoFold+ Implementation

RNA 3D Structure Prediction with RhoFold+

Detailed Protocol:

• Input Processing: Input RNA sequences are processed through RNA-FM, a large language model pretrained on ~23.7 million RNA sequences, to extract evolutionarily informed embeddings [33].
• Feature Integration: Concurrently, MSAs are generated by searching extensive sequence databases. Embeddings and MSA features are integrated and fed into the Rhoformer transformer network for iterative refinement over ten cycles [33].
• Structure Generation: The structure module employs geometry-aware attention and invariant point attention (IPA) to optimize local frame coordinates and torsion angles for RNA backbone atoms [33].
• Constraint Application: Structural constraints including predicted secondary structure and base pairing are applied during full-atom coordinate reconstruction [33].
• Model Output: The final output consists of all-atom 3D coordinates in PDB format suitable for molecular visualization and further analysis.

Hybrid Approach: BPfold Implementation

Detailed Protocol:

• Base Pair Motif Library Construction: Enumerate the complete space of three-neighbor base pair motifs (BPMiH, BPMiCB, BPMoCB) representing all possible local structural contexts for canonical base pairs [66].
• Energy Calculation: Compute thermodynamic energy for each motif using BRIQ, a de novo RNA structure modeling method that combines density functional theory physical energy with statistical energy calibrated by quantum mechanics [66].
• Energy Map Generation: For an input RNA sequence of length L, construct two L×L energy maps (Mμ for outer base pair motifs, Mν for inner base pair motifs) representing the thermodynamic landscape [66].
• Neural Network Processing: Process RNA sequence features and energy maps through the base pair attention block, which integrates transformer and convolutional layers to effectively learn the relationship between sequence and thermodynamic priors [66].
• Secondary Structure Prediction: Output base pairing probabilities and determine the final secondary structure including pseudoknots and non-canonical pairs.

Physics-Based Method: FARFAR2 Implementation

Detailed Protocol:

• Fragment Library Assembly: Generate a library of RNA backbone fragments from known high-resolution structures [33].
• Monte Carlo Sampling: Perform large-scale fragment assembly simulations using Monte Carlo methods to explore the conformational space [33].
• Energy Evaluation: Score sampled structures using Rosetta's all-atom energy function, which combines physical and knowledge-based terms [33].
• Ensemble Selection: Cluster low-energy structures and select representative models from the largest clusters to address the rugged energy landscape [33].
• Refinement: Apply all-atom energy minimization to remove steric clashes and optimize local geometry.

The Scientist's Toolkit

Table 3: Essential Research Reagents and Computational Resources

Resource Category	Specific Tools	Primary Function	Application Context
Benchmarking Platforms	RNA-Puzzles, rnaglib	Standardized performance assessment	Method validation & comparison
Structure Databases	PDB, BGSU representative sets	Source of experimental structures	Template-based modeling, training data
Sequence Databases	Rfam, RNACentral	Evolutionary information source	MSA construction, language model training
Energy Parameters	Turner rules, BRIQ energy	Thermodynamic energy calculation	Physics-based methods, hybrid approaches
Neural Architectures	Transformers, Invariant Point Attention	Geometric deep learning	3D coordinate generation
Analysis Tools	US-align, CD-HIT	Structural & sequence similarity	Dataset preparation, result evaluation

Discussion and Future Directions

The comparative analysis reveals a clear trajectory in RNA structure prediction methodology. Deep learning approaches, particularly those integrating physical priors and evolutionary information, have demonstrated superior performance in both accuracy and computational efficiency. RhoFold+ achieves remarkable performance with an average RMSD of 4.02Ã… on RNA-Puzzles targets, significantly outperforming physics-based (6.32Ã… for FARFAR2) and template-based methods [33]. However, the integration of physical principles, as exemplified by BPfold's base pair motif energy, appears crucial for generalizability across diverse RNA families [66].

Critical challenges remain in several areas. The scarcity of high-quality experimental structures continues to limit method development, with RNA structures representing less than 1% of the PDB [33]. Addressing RNA flexibility and conformational dynamics requires moving beyond static structure prediction. Modeling RNA-protein complexes and RNA-ligand interactions presents additional challenges due to interface complexity. The field would benefit from standardized benchmarking frameworks like those proposed in recent work [58] to ensure rigorous evaluation and comparability across methods.

Future methodological developments will likely focus on multi-scale modeling approaches that integrate local base pair interactions with global architectural principles. The successful integration of physical priors with deep learning, as demonstrated by BPfold, suggests a promising path toward more generalizable and physically plausible predictions. As these methods mature, they will increasingly support drug discovery efforts targeting RNA and enhance our fundamental understanding of RNA biology in cellular regulation and disease mechanisms.

In the field of computational structural biology, accurately predicting and evaluating the three-dimensional conformation of biomolecules is paramount. For RNA, whose function is intimately tied to its dynamic structure, assessing the quality of predicted models is a critical step in both methodological development and practical application. While numerous assessment metrics exist, Root-Mean-Square Deviation (RMSD), Template Modeling score (TM-score), and Local Distance Difference Test (LDDT) represent three cornerstone evaluation methods. Each provides a unique perspective on model quality, from local atomic precision to global topological similarity. This guide provides an in-depth technical examination of these core metrics, framing them within the context of RNA structural research and drug development. We detail their underlying methodologies, present comparative analyses, and offer practical protocols for their application, serving as a foundational resource for researchers navigating the complex landscape of model quality assessment.

Theoretical Foundations of Core Metrics

Root-Mean-Square Deviation (RMSD)

RMSD is one of the most traditional and widely used metrics for quantifying the difference between two atomic coordinate sets, such as a predicted model and a native reference structure. Its calculation is mathematically straightforward, providing a single value representing the average magnitude of atomic displacements.

• Calculation Protocol: The standard protocol for calculating RMSD involves a least-squares superposition of the model onto the native structure to minimize the RMSD value, followed by the calculation of the square root of the average squared distances between corresponding atoms. The formula for calculating RMSD after optimal superposition is:

  RMSD = √[ (1/N) × Σᵢ (dᵢ)² ]

  where N is the number of atoms being compared, and dáµ¢ is the distance between the i-th pair of equivalent atoms after superposition. For RNA, comparisons are typically made using the phosphorus (P) atoms or all heavy atoms of the backbone.

• Interpretation and Limitations: A lower RMSD value indicates a closer match to the native structure. However, a significant limitation of RMSD is its high sensitivity to local errors and its tendency to increase with the length of the molecule. A large local error in a small region can disproportionately inflate the overall RMSD, providing a potentially misleading assessment of global fold accuracy [78]. Furthermore, RMSD values are expressed in Ã…ngströms (Ã…) and do not have a fixed upper bound, making it difficult to interpret their absolute significance without context.

Template Modeling Score (TM-score)

The TM-score was developed to address several shortcomings of RMSD, specifically by providing a more nuanced assessment of global fold similarity that is less sensitive to local variations and is normalized for protein length [79].

• Calculation Protocol: TM-score is a superposition-based metric that weighs smaller distances more heavily than larger ones, making it more reflective of the global fold than local errors. It is calculated using the following formula:

  TM-score = max [ (1/L) × Σᵢ [1 / (1 + (dᵢ/d₀)²)] ]

  where L is the length of the native structure, dᵢ is the distance between the i-th pair of equivalent residues (typically Cα for proteins, P or C4' for RNA), and d₀ is a length-dependent scale factor that normalizes the score. The "max" indicates that the structural alignment is optimized to maximize this score.

• Interpretation and Significance: The TM-score ranges between (0, 1], where a score of 1 indicates a perfect match. Empirically, scores below 0.17 indicate random structural similarity, while scores above 0.5 generally indicate that two structures share the same fold in protein SCOP/CATH classifications [79]. This normalization makes TM-score highly useful for comparing the quality of models for molecules of different lengths. It has been successfully adapted for use in RNA structure assessment [80] [78].

Local Distance Difference Test (LDDT)

LDDT is a superposition-free metric that evaluates the local consistency of a model by comparing inter-atomic distances within a defined cutoff, making it ideal for assessing models without the need for global alignment.

• Calculation Protocol: LDDT is computed by comparing distances between all atom pairs (or a specific subset, like Cα or P atoms) in the model and the native structure within a specified radius (e.g., 15 Ã…). The protocol involves:

  • For a given residue, all atom pairs within the cutoff radius are identified in the native structure.
  • The distance differences for these pairs between the model and native are calculated.
  • The fraction of these pairs with a difference below a set of predefined thresholds (commonly 0.5, 1, 2, and 4 Ã…) is computed.
  • This fraction is averaged over all residues and all four thresholds to produce the final LDDT score.

• Key Advantages: As a local and superposition-free score, LDDT is robust for evaluating models with domain movements and is less sensitive to insertion/deletion errors than superposition-based methods. Its score range of [0, 1] and local nature make it a valuable component of the CASP assessment criteria for proteins, and it is increasingly applied to RNA structures [80] [81] [78].

Table 1: Core Characteristics of RMSD, TM-score, and LDDT

Metric	Score Range	What is Measured	Superposition Required?	Key Advantage
RMSD	[0, ∞)	Mean distance between corresponding atoms	Yes	Intuitive, direct measure of atomic-level deviation
TM-score	(0, 1]	Length-normalized, global fold similarity	Yes	Robust to local errors; allows cross-length comparison
LDDT	[0, 1]	Preservation of local inter-atomic distances	No	Evaluates local quality; insensitive to domain shifts

Comparative Analysis and Practical Interpretation

A comprehensive model assessment requires understanding how these metrics complement each other. Research on protein models has shown that the distributions and correspondences between these scores can be highly heterogeneous [81].

Score Distributions and Correlations: Empirical analyses reveal that RMSD (often transformed to a [0,1] scale as tRMSD = 1/(1+(RMSD/10)²)) and TM-score distributions on large model sets can show a bimodal character, while LDDT tends to spread values more evenly. The correspondence between any two scores is not linear; for instance, a single TM-score value can correspond to a wide range of RMSD values, and vice-versa [81]. This underscores the importance of using multiple metrics.

Selecting the Best Model: Different metrics can prioritize different models. RMSD may favor a model with good overall atom positioning but an incorrect global fold, whereas TM-score is designed to penalize incorrect global topology more heavily. LDDT, being local, might identify a model with correct local environments even if domains are mis-oriented. Consequently, the "best" model often depends on the biological question or application at hand.

Table 2: Metric Performance Against Desirable Properties for Model Assessment

Desirable Property	RMSD	TM-score	LDDT
Promotes complete models	Poor	Good	Good
Handles flexible regions/domains	Poor	Moderate	Excellent
Independent of protein/RNA length	No	Yes	Yes
Assesses realistic stereochemistry	Indirectly	Indirectly	Indirectly
Sensitive to local structural variations	Excellent	Moderate	Excellent

Application in RNA Structural Biology

The evaluation of RNA 3D structures introduces specific challenges due to RNA's greater flexibility, complex charge distribution, and stabilization by base pairing and stacking, rather than a hydrophobic core.

Adaptation of Metrics for RNA: While RMSD, TM-score, and LDDT were originally developed for proteins, they have been adapted for RNA. TM-score can be calculated for RNA using phosphorus atoms or other backbone references [79] [78]. Similarly, LDDT calculations for RNA can focus on key atoms to evaluate the local nucleotide environment. However, these general metrics may not fully capture RNA-specific features.

The Role of RNA-Specific Metrics: To address the limitations of general metrics, RNA-oriented scores have been developed. The Interaction Network Fidelity (INF) score assesses the accuracy of base-pairing and base-stacking interactions, which are crucial for RNA stability [78]. The Mean of Circular Quantities (MCQ) measures the dissimilarity of torsion angles, providing a detailed view of backbone conformation accuracy [78]. For a thorough assessment of an RNA model, it is recommended to use a combination of general metrics (RMSD, TM-score, LDDT) and RNA-specific metrics (INF, MCQ) [80] [78].

Integrated Tools for Assessment: Tools like RNAdvisor have been developed to provide a unified platform for the automated calculation of a wide array of these metrics and scoring functions for RNA 3D structures, significantly enhancing the efficiency and comprehensiveness of model evaluation [80] [78].

Experimental Protocols and Workflows

Standard Protocol for Calculating RMSD, TM-score, and LDDT

This protocol describes a standard workflow for evaluating a predicted structural model against a known native structure using the three core metrics.

• Step 1: Input Preparation. Obtain the coordinate files for the predicted model and the native reference structure in PDB or mmCIF format. Ensure the two structures contain the same sequence for a residue-by-residue comparison. If not, a sequence-based structural alignment tool (like TM-align) is a prerequisite.
• Step 2: Structural Superposition for RMSD/TM-score. For RMSD and TM-score, the model must be superposed onto the native structure. This is typically done using a least-squares fitting algorithm on a defined set of equivalent atoms (e.g., P atoms for RNA). The rotation and translation matrix that minimizes the RMSD is applied to the entire model.
• Step 3: RMSD Calculation. After superposition, calculate the RMSD using the standard formula on the set of equivalent atoms. The output is a value in Ã…ngströms.
• Step 4: TM-score Calculation. Using the superposed structures, calculate the TM-score. The algorithm will identify the optimal residue mapping if not predefined and compute the length-normalized score. The output is a value between 0 and 1.
• Step 5: LDDT Calculation. For LDDT, no superposition is needed. The calculation involves comparing all relevant inter-atomic distances in the model and native structure within a 15 Ã… radius for each residue, using thresholds of 0.5, 1, 2, and 4 Ã…. The output is a value between 0 and 1.
• Step 6: Analysis and Interpretation. Synthesize the results from all three metrics. A high-quality model should ideally have a low RMSD, a high TM-score (>0.5), and a high LDDT (>0.7), though discrepancies can reveal specific strengths and weaknesses in the model.

<85 chars

Protocol for Integrated RNA Model Quality Assessment

For a more comprehensive evaluation specific to RNA, this protocol incorporates RNA-specific metrics and unified tools.

• Step 1: Dataset and Tool Setup. Prepare your set of RNA model structures and the native structure. Install an integrated assessment tool like RNAdvisor 2, which provides a unified platform for calculating numerous metrics and scoring functions [80].
• Step 2: Configuration. Configure RNAdvisor 2 to calculate a suite of metrics. A recommended set includes:
  • General Metrics: RMSD, CLASH score (for steric clashes).
  • Protein-Inspired Metrics: TM-score, GDT-TS, CAD-score, LDDT.
  • RNA-Oriented Metrics: INF (for base pairing/stacking), MCQ (for torsion angles) [78].
• Step 3: Automated Calculation. Run the RNAdvisor 2 tool. It will automatically parse the structures, perform necessary alignments, and compute all selected metrics.
• Step 4: Meta-Metric Analysis. RNAdvisor 2 introduces the concept of meta-metrics and meta-scoring functions, which unify diverse evaluation criteria into more robust indicators of quality. Analyze these composite scores for a holistic view [80].
• Step 5: Comparative Benchmarking. Use the tool's benchmarking capabilities to compare your model's scores against those of other models or established decoy sets to contextualize its relative quality.

Table 3: Key Software Tools for Structural Quality Assessment

Tool Name	Type	Primary Function	Application Note
RNAdvisor 2 [80]	Unified Platform	Automated calculation of a wide range of RNA metrics & scores	Integrates metrics (RMSD, INF), scoring functions, and meta-metrics. Web server and CLI available.
TM-score [79]	Standalone Executable	Calculate TM-score and RMSD for proteins/RNA	Requires pre-aligned structures; C++ and Fortran source available.
Bio.PDB [82]	Python Module	Parser and structure analysis within Biopython	Provides foundational classes for reading PDB files and implementing custom metric calculations.
PyMOL [83]	Molecular Viewer	Visualization and basic measurement	Command-line align and rms_cur commands can be used for manual calculation and visualization of fits.
AMBER	MD Suite	Molecular Dynamics Simulations	Used in refinement; force fields like χOL3 can test model stability [84].

The accurate assessment of structural models is a critical component of computational biology, directly impacting the reliability of scientific conclusions and the success of downstream applications like drug design. RMSD, TM-score, and LDDT form a triad of essential metrics, each contributing a unique and valuable perspective on model quality. RMSD offers a direct measure of atomic-level precision, TM-score provides a robust, length-normalized evaluation of global fold, and LDDT gives a superposition-free insight into local environment accuracy. For RNA structures, these should be complemented with RNA-specific metrics such as INF and MCQ for a truly comprehensive evaluation. As the field progresses, unified platforms like RNAdvisor 2 and the development of meta-metrics are paving the way for more integrated, informative, and automated model assessment, empowering researchers to better judge the quality and utility of their structural predictions.

The advent of deep learning (DL) has revolutionized de novo RNA secondary structure prediction, with models often surpassing the performance of traditional thermodynamics-based algorithms on common benchmarks [85]. However, the practical utility of these statistical models critically hinges on a factor distinct from their peak performance: their generalizability to novel RNA folds. Cross-family generalizability, the ability of a model trained on sequences from one set of RNA families to accurately predict structures for sequences from entirely different families, represents a significant hurdle in the field [85]. This challenge is rooted in the core of how these models learn; instead of inferring physical laws, they act as statistical learners of sequence-structure correlations present in the training data. Consequently, when a model encounters a test sequence with low similarity to its training set, its performance can degrade rapidly [85]. This technical guide examines the evidence for this limitation, outlines methodologies for its quantitative assessment, and discusses pathways toward creating more robust predictive models for RNA structure, a cornerstone of understanding RNA dynamics and function.

The Generalizability Challenge in de novo Deep Learning Models

The Core Problem: Performance-Decay with Declining Sequence Similarity

The exceptional performance of de novo DL models is heavily dependent on the statistical distribution of the data they are trained on. These models are not learning the fundamental biophysics of RNA folding but are instead learning to associate sequences with structures based on patterns seen during training [85]. This reliance makes them susceptible to a major limitation: a pronounced drop in predictive accuracy when applied to sequences that are evolutionarily distant from those in the training set.

A 2023 quantitative study systematically evaluated this phenomenon by developing a series of DL models (SeqFold2D) and testing them under varying levels of sequence similarity between training and test datasets [85]. The study defined three levels of stringency:

• Cross-sequence: No identical sequences between sets.
• Cross-cluster: All sequences share less than 80% identity.
• Family-wise: Training and test sequences are from completely different RNA families.

The findings were stark. While models demonstrated excellent expressive capacity and high performance on test sequences from the same families used for training, their generalizability—the performance gap between training and test sets—degraded rapidly as sequence similarity decreased [85]. This inverse correlation between performance and generalizability was observed collectively across a range of learning-based models, indicating it is a fundamental characteristic of the current statistical learning approach rather than a flaw in a specific model architecture.

Quantitative Evidence from Recent Studies

The table below summarizes key findings from recent research investigating the generalizability of RNA structure prediction models:

Table 1: Quantitative Evidence of Generalizability Challenges in RNA Structure Prediction

Study Focus	Key Finding on Generalizability	Impact on Model Performance
SeqFold2D Model Analysis [85]	Model generalizability degrades rapidly as sequence similarity between training and test sets decreases.	Performance gap between seen and unseen data widens significantly in cross-family validation scenarios.
Basic CNN + RNAstructure [85]	Excellent performance on test data from the same RNA families as training data.	Markedly worse performance observed for test sequences from different families than the training set.
Neural Networks & Thermodynamic Data [85]	Models fail to generalize over sequences of different lengths and, more critically, over sequences with structures absent from training data.	Highlights non-data-agnostic behavior, indicating models learn dataset-specific correlations rather than underlying folding principles.

Methodologies for Quantifying Generalizability

To systematically evaluate and benchmark the cross-family generalizability of an RNA structure prediction model, a rigorous experimental framework is required.

Defining Similarity Levels and Dataset Curation

The first step is to curate training and test datasets with precisely defined relationships. The three-level hierarchy provides a robust framework [85]:

• Cross-sequence Level: The baseline, ensuring no identical sequences are shared between training and test sets.
• Cross-cluster Level: A more stringent test, where sequences in the training and test sets share less than 80% sequence identity. Tools like CD-HIT-EST are used to cluster sequences and select representatives to enforce this separation [85].
• Family-wise Level: The most rigorous test of generalizability, where training and test sets are composed of sequences from completely different RNA families, ensuring the model encounters truly novel folds.

Experimental Protocol for Benchmarking

The following workflow provides a detailed protocol for conducting a generalizability assessment:

Diagram: Workflow for Generalizability Assessment

Step 1: Data Preparation and Partitioning

• Source a large, curated RNA structure database such as bpRNA [85].
• Filter the dataset to remove sequences with high similarity using a tool like CD-HIT-EST with an 80% identity threshold to create distinct sequence clusters [85].
• Partition the clusters into training and test sets. For family-wise validation, ensure clusters from entire RNA families are exclusively in either the training or test set.

Step 2: Model Training and Validation

• Select/Design a DL model for RNA secondary structure prediction. The SeqFold2D study used a minimal two-module architecture (Seq2Seq and Conv2D) without post-processing to probe intrinsic model characteristics [85].
• Train the model on the designated training set. Monitor performance on a held-out validation set from the training family distribution.
• Record performance metrics (e.g., F1 score, precision, recall) on the validation set.

Step 3: Cross-Family Testing and Analysis

• Predict secondary structures for all sequences in the various test sets (cross-sequence, cross-cluster, family-wise).
• Calculate the same performance metrics for each test set.
• Compute the generalizability gap for each metric, defined as the difference between the performance on the training/validation set and the performance on each test set.

Quantitative Analysis via Sequence and Structure Alignment

To move beyond a simple performance score and gain mechanistic insight, generalizability should be quantitated against direct measures of similarity.

• Sequence Identity: Use pairwise sequence alignment tools (e.g., BLAST) to compute the sequence identity between each test sequence and its closest match in the training set. The dependence of the generalizability gap on this identity score can then be plotted [85].
• Structure Identity: Use RNA structure alignment tools to compute a structure identity score (e.g., using metrics like F1-score on base pairs) between the predicted and known structures for test sequences. This reveals how structural divergence correlates with performance decay [85].

Table 2: Key Reagents and Computational Tools for Generalizability Research

Resource Category	Specific Tool / Reagent	Function in Generalizability Research
RNA Structure Datasets	bpRNA [85]	A large-scale repository of annotated RNA secondary structures for training and benchmarking models.
Sequence Clustering Tool	CD-HIT-EST [85]	Clusters RNA sequences by identity to create non-redundant training and test sets for robust evaluation.
Computational Prediction Tools	IntaRNA, RNAup [43]	Tools for predicting RNA-RNA interactions; useful for comparison and understanding interaction-specific generalizability.
Visualization Software	Sashimi plots (in IGV) [86]	Enables quantitative visualization of RNA-Seq read alignments and isoform expression across samples.
Alignment & Analysis	Pairwise sequence & structure alignment algorithms [85]	Quantifies sequence and structure identity, providing the basis for analyzing generalizability dependence.

Pathways Toward Improved Generalizability

Overcoming the generalizability challenge is an active area of research. Several promising pathways are emerging, though a complete solution remains on the horizon.

• Incorporation of Biophysical Constraints: Integrating RNA folding thermodynamics or chemical mapping data (e.g., SHAPE reactivity) as constraints during model training or prediction can ground the model in physical reality rather than purely statistical correlations [43].
• Exploitation of Co-evolutionary Information: While de novo models use a single sequence, leveraging evolutionary information from multiple sequence alignments of homologous RNAs can provide powerful constraints on possible structures, a strategy used by comparative sequence analysis methods [85].
• Advanced Regularization and Data Augmentation: Employing more sophisticated regularization techniques during training and generating synthetic RNA sequences with known structures (e.g., via inverse folding) can help models learn a more robust and generalizable understanding of the folding landscape [85].
• Multi-Task and Self-Supervised Learning: Training models on auxiliary tasks or using self-supervised pre-training on large corpora of unlabeled RNA sequences could force the model to learn more fundamental representations of RNA sequence and structure.

The issue of cross-family generalizability is a central challenge that must be addressed for de novo deep learning models to become universally reliable tools in RNA structural biology and drug development. Quantitative evidence clearly demonstrates that current models, while powerful, are primarily sophisticated statistical learners whose performance is tightly coupled to the similarity between their training data and the target of inquiry. By adopting standardized, rigorous benchmarking protocols that include family-wise validation and by quantitating performance against sequence and structure similarity, researchers can better diagnose and understand the limitations of their models. The future of robust computational RNA structure prediction lies in moving beyond pure pattern recognition and integrating the statistical power of deep learning with the foundational principles of RNA biophysics and evolution.

Conclusion

The field of RNA structure and dynamics is undergoing a transformative shift, driven by the convergence of deep learning, powerful simulations, and integrative experimental methods. Tools like RhoFold+ are demonstrating that accurate de novo 3D structure prediction is now achievable, while advanced MD simulations and selective labeling techniques provide unprecedented insights into dynamic conformational ensembles. The key takeaway is that a multi-faceted approach is essential; no single method can fully capture the complexity of RNA biology. The successful application of these advancements is already evident in the burgeoning pipeline of RNA-targeted therapeutics, including small molecules, splice modulators, and RNA degraders. Looking forward, future progress will hinge on improving the accuracy of physical force fields, expanding the library of solved RNA structures for training and validation, and further refining AI models to predict not just structure but also functional dynamics and interactions. This integrated knowledge base is poised to accelerate the rational design of novel RNA-targeted therapies, unlocking new treatment paradigms for a wide range of diseases previously considered undruggable.

References

• 1. ERNIE-RNA: an RNA language model with structure ... [https://www.nature.com/articles/s41467-025-64972-0]
• 2. Advances and opportunities in RNA structure experimental ... [https://www.nature.com/articles/s41592-022-01623-y]
• 3. 10.4: RNA Structure [https://bio.libretexts.org/Bookshelves/Computational_Biology/Book%3A_Computational_Biology_-_Genomes_Networks_and_Evolution_(Kellis_et_al.)/10%3A_RNA_Folding/10.04%3A_RNA_Structure]
• 4. Nucleic acid tertiary structure [https://en.wikipedia.org/wiki/Nucleic_acid_tertiary_structure]
• 5. Chapter 8 Color Palette | Single Cell Multi-Omics Data ... [https://bookdown.org/ytliu13207/SingleCellMultiOmicsDataAnalysis/color-palette.html]
• 6. RNA structure experimental analysis--chemical modification [https://pubmed.ncbi.nlm.nih.gov/24034333/]
• 7. Quantitative Understanding of SHAPE Mechanism from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6311105/]
• 8. Frontiers | Viral replication modulated by hallmark conformational ensembles: how AlphaFold-predicted features of RdRp folding dynamics combined with intrinsic disorder-mediated function enable RNA virus discovery [https://www.frontiersin.org/journals/virology/articles/10.3389/fviro.2025.1501616/full]
• 9. Identification of conserved RNA regulatory switches in living cells using RNA secondary structure ensemble mapping and covariation analysis | Nature Biotechnology [https://www.nature.com/articles/s41587-025-02739-0]
• 10. RNA structure prediction/design using RNA-As-Graphs [https://pmc.ncbi.nlm.nih.gov/articles/PMC7999136/]
• 11. RAG: RNA-As-Graphs web resource - BMC Bioinformatics [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-5-88]
• 12. RNA research for drug discovery: Recent advances and ... [https://www.sciencedirect.com/science/article/abs/pii/S0378111925001301]
• 13. DynaRNA: accurate dynamic RNA conformation ensemble ... [https://www.nature.com/articles/s42003-025-08875-2]
• 14. 19F NMR in RNA structural biology: exploring ... [https://www.sciencedirect.com/science/article/abs/pii/S0223523425004477]
• 15. Deep dive into RNA: a systematic literature review on ... [https://link.springer.com/article/10.1007/s10462-024-10910-3]
• 16. Breakthrough in RNA-targeted drug discovery offers hope ... [https://www.news-medical.net/news/20251006/Breakthrough-in-RNA-targeted-drug-discovery-offers-hope-against-viral-diseases.aspx]
• 17. The current and future landscape of RNA-based therapies ... [https://www.nature.com/articles/s43856-025-01166-1]
• 18. RNA structure prediction using deep learning [https://www.sciencedirect.com/science/article/pii/S0010482525001957]
• 19. RNAtranslator: Modeling protein-conditional RNA design as ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12510665/]
• 20. The roles of structural dynamics in the cellular functions of RNAs [https://pmc.ncbi.nlm.nih.gov/articles/PMC7656661/]
• 21. RNA Therapeutics As We Enter 2025 Looking Beyond The ... [https://www.advancingrna.com/doc/rna-therapeutics-as-we-enter-2025-looking-beyond-the-horizon-of-first-generation-success-0001]
• 22. Advances in RNA-based therapeutics: current breakthroughs ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12592931/]
• 23. RNA Therapeutics Market Size, Share & 2030 Trends Report [https://www.mordorintelligence.com/industry-reports/rna-therapeutics-market]
• 24. RNA Therapy Clinical Trials Market Size 2025 to 2034 [https://www.precedenceresearch.com/rna-therapy-clinical-trials-market]
• 25. RNA H1 2025 Landscape Review [https://beacon-intelligence.com/landscape-reviews/rna-h1-2025-landscape-review/]
• 26. Current Progress and Future Perspectives of RNA-Based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12153701/]
• 27. RNA Targeting Small Molecules Therapeutics Global ... [https://finance.yahoo.com/news/rna-targeting-small-molecules-therapeutics-100900139.html]
• 28. RNA structure determination: From 2D to 3D - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11197651/]
• 29. Challenges and opportunities in technologies and methods ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12492894/]
• 30. RNAstructure [https://rna.urmc.rochester.edu/RNAstructure.html]
• 31. RNA-Puzzles Round V: blind predictions of 23 ... [https://www.nature.com/articles/s41592-024-02543-9]
• 32. The RNA workbench: best practices for RNA and high ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5570170/]
• 33. Accurate RNA 3D structure prediction using a language ... [https://www.nature.com/articles/s41592-024-02487-0]
• 34. Accurate prediction of RNA secondary structure including ... [https://www.nature.com/articles/s42003-024-05952-w]
• 35. Predicting RNA structure and dynamics with deep learning ... [https://www.sciencedirect.com/science/article/pii/S0006349524041055]
• 36. Molecular Dynamics Simulations of RNA Stem-Loop Folding ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12593105/]
• 37. Comprehensive datasets for RNA design, machine ... [https://www.nature.com/articles/s41598-025-07041-2]
• 38. Enhanced Sampling of Ligand Binding Coupled to RNA ... [https://chemrxiv.org/engage/chemrxiv/article-details/689d8334728bf9025e5dc531]
• 39. Computational modeling of cotranscriptional RNA folding [https://www.sciencedirect.com/science/article/pii/S2001037025002235]
• 40. and Interactions Revealed through Atomistic Simulations RNA Dynamics [https://arxiv.org/html/2507.08474v1]
• 41. Whole Transcriptome vs 3' mRNA-Seq: How to Choose ... [https://www.lexogen.com/blog/choosing-between-whole-transcriptome-total-rna-seq-and-3-mrna-seq/]
• 42. A Beginner's Guide to Analysis of RNA Sequencing Data [https://pmc.ncbi.nlm.nih.gov/articles/PMC6096346/]
• 43. Experimental and computational methods for studying the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10799312/]
• 44. Systematic comparison and assessment of RNA-seq ... [https://www.nature.com/articles/s41598-020-76881-x]
• 45. Targeting RNA with small molecules using state-of-the-art ... [https://www.nature.com/articles/s42003-025-08809-y]
• 46. The evolution and application of RNA-focused small molecule ... [https://pubs.rsc.org/en/content/articlehtml/2025/cb/d4cb00272e]
• 47. Discovery of RNA‐Targeting Small Molecules - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC12375692/]
• 48. R2DT: a comprehensive platform for visualizing RNA ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11806352/]
• 49. RNA folding: new model could change future drug design [https://www.drugtargetreview.com/news/190250/rna-folding-new-model-could-change-future-drug-design/]
• 50. RNA Targeting Small Molecule Drug Discovery Research ... [https://finance.yahoo.com/news/rna-targeting-small-molecule-drug-145600353.html]
• 51. RNA Targeting Small Molecules Therapeutics Global ... [https://www.businesswire.com/news/home/20251114138503/en/RNA-Targeting-Small-Molecules-Therapeutics-Global-Research-Report-2025-2034-Market-to-Grow-with-Rising-RD-Investments-Genetic-Disorder-Incidence-and-Cancer-Prevalence---ResearchAndMarkets.com]
• 52. RNA Structure Datasets for AI & Drug Development [https://eclipsebio.com/datasets/rna-structure/]
• 53. End-to-end learning of multiple sequence alignments with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9805565/]
• 54. Muscle5: High-accuracy alignment ensembles enable ... [https://www.nature.com/articles/s41467-022-34630-w]
• 55. Synthetic data production for biomedical research - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC12066231/]
• 56. Synthetic single cell RNA sequencing data from small pilot ... [https://www.nature.com/articles/s41598-021-88875-4]
• 57. A Comprehensive Benchmark for RNA 3D Structure ... [https://arxiv.org/abs/2503.21681]
• 58. A Comprehensive Benchmark for RNA 3D Structure ... [https://arxiv.org/html/2503.21681v3]
• 59. CompaRNA - on-line benchmarks of RNA structure ... [https://genesilico.pl/comparna/]
• 60. Accurate in silico predictions of modified RNA interactions ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12439594/]
• 61. Success and Limitations of Current Force Fields for the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12621256/]
• 62. Grappa – a machine learned molecular mechanics force field [https://pubs.rsc.org/en/content/articlelanding/2025/sc/d4sc05465b]
• 63. Machine Learning‐Augmented Molecular Dynamics ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12304844/]
• 64. Designing small molecules targeting a cryptic RNA binding ... [https://www.nature.com/articles/s41589-025-02018-8]
• 65. Identifying RNA-small Molecule Binding Sites Using ... [https://www.sciencedirect.com/science/article/abs/pii/S0022283625000762]
• 66. Deep generalizable prediction of RNA secondary structure ... [https://www.nature.com/articles/s41467-025-60048-1]
• 67. RiboPO: Preference Optimization for Structure [https://arxiv.org/html/2510.21161v2]
• 68. How AI could speed the development of RNA vaccines and ... [https://news.mit.edu/2025/how-ai-could-speed-development-rna-vaccines-and-other-rna-therapies-0815]
• 69. Finding recurrent RNA structural networks with fast ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8191989/]
• 70. RNA motif discovery: a computational overview - Biology Direct [https://biologydirect.biomedcentral.com/articles/10.1186/s13062-015-0090-5]
• 71. RNA-Seq Experimental Design Guide for Drug Discovery [https://www.lexogen.com/blog/planning-for-success-a-strategic-design-guide-for-rna-seq-experiments-in-drug-discovery/]
• 72. Commentary: a review of technical considerations for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12522984/]
• 73. RNA-Puzzles: A CASP-like evaluation of RNA three ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3312550/]
• 74. RNA-Puzzles: Group [https://www.rnapuzzles.org/]
• 75. Home - CASP16 [https://predictioncenter.org/casp16/]
• 76. Ensemble structure modeling challenges in CASP15 - PubMed [https://pubmed.ncbi.nlm.nih.gov/37872703/]
• 77. Comprehensive benchmarking of large language models ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11982019/]
• 78. RNAdvisor: a comprehensive benchmarking tool for the measure and prediction of RNA structural model quality - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10939302/]
• 79. TM-score: Quantitative assessment of similarity between protein structures [https://zhanggroup.org/TM-score/]
• 80. RNAdvisor 2: A unified platform for RNA 3D model quality ... [https://www.nature.com/articles/s41598-025-23326-y]
• 81. Comparative analysis of methods for evaluation of protein models against native structures - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6419896/]
• 82. The Biopython Structural Bioinformatics FAQ [https://biopython.org/wiki/The_Biopython_Structural_Bioinformatics_FAQ]
• 83. PyMOL Command Reference [https://pymol.org/pymol-command-ref.html]
• 84. When does molecular dynamics improve RNA models ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12513224/]
• 85. Sequence similarity governs generalizability of de novo ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10138783/]
• 86. Quantitative visualization of alternative exon expression ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4542614/]