From small molecules to oligonucleotides: artificial intelligence for RNA-targeted drug discovery

1 Introduction

Small molecules have long been central to modern therapeutics and remain the largest class of approved medicines, owing to properties such as chemical tunability, manufacturability, oral bioavailability, tissue penetration, and access to intracellular biology (Knox et al., 2024). Yet most discovery strategies have been built around proteins with well-defined binding sites. Expanding therapy beyond this protein-centric framework has renewed interest in RNA, which regulates gene expression, cellular state, and disease pathways at multiple levels (Kovachka et al., 2024; Childs-Disney et al., 2022). Clinical successes now show that RNA function can be modulated therapeutically (Crooke et al., 2018). The approval of risdiplam, a small-molecule splicing modifier for spinal muscular atrophy, established the feasibility of selectively targeting RNA-mediated processes with drug-like molecules (Dhillon, 2020; Ratni et al., 2021; Nakatani, 2022). Building on this, the oral splicing modulators branaplam and votoplam (PTC518), which lower huntingtin (HTT) via the pseudoexon inclusion, reached phase 2 trials for Huntington’s disease (HD); votoplam met its primary endpoint, whereas branaplam’s VIBRANT-HD study was halted for neuropathy (Borowsky et al., 2026; Gao et al., 2024). In parallel, 2025 marked notable progress for oligonucleotide drugs, with FDA approval of three agents: the siRNA therapeutics fitusiran and plozasiran, and the antisense oligonucleotide (ASO) donidalorsen (U.S. Food and Drug Administration, 2026; AlShaer et al., 2025). Together with advances in oligonucleotide chemistry, safety, and delivery, these approvals underscore the expanding clinical relevance of RNA-based strategies across diverse disease areas (Li et al., 2026; Andersson et al., 2022; Roberts et al., 2020).

Despite this progress, RNA-targeted drug discovery remains substantially harder than its protein-targeted counterpart. RNA frequently adopts dynamic, heterogeneous conformations that complicate target identification and molecular design. For RNA-targeting small molecules, the scarcity of experimentally resolved RNA–ligand complexes and the limited number of persistent, well-defined ligandable sites hinder rational design. For oligonucleotides, researchers need to navigate large combinatorial spaces while balancing efficacy, specificity, toxicity, immunogenicity, chemical modification, and tissue-specific delivery. These are multidimensional optimization problems that traditional experimental and computational approaches struggle to address alone. Recent advances in artificial intelligence (AI), including deep learning (DL), foundation models, multimodal learning, and emerging agentic systems are beginning to show the potential to model the relationships among RNA sequence, structure, chemical space, and biological context, are increasingly embedded in iterative workflows spanning structure modeling, target prioritization, molecular design, safety assessment, and experimental planning. In this review, we discuss how AI and machine learning (ML) are increasingly involved in RNA drug discovery in RNA-targeting small molecules and oligonucleotide therapeutics. Figure 1 provides an overview of the methods discussed and organizes them by stages and methodological basis.

2 RNA structure and target discovery in the AI era

RNA structure is central to RNA-targeted drug discovery because conformation governs biological function and, in turn, how a therapeutic engages its target, and this holds for both modalities. For small-molecule targeting, RNA structure is critical because the three-dimensional fold defines the shape and accessibility of ligandable pockets. Unlike proteins, RNA rarely presents deep, well-defined hydrophobic cavities; instead, binding pockets often emerge from internal loops, bulges, junctions, and other higher-order structural motifs. For ASOs and siRNAs, structure matters differently: because they act through Watson–Crick hybridization, their efficacy mainly depends on target-site accessibility, so stable local secondary structure can occlude the complementary region and reduce potency, whereas open, single-stranded regions are more favorable. In both cases, accurate structure prediction is a prerequisite for locating candidate pockets on the one hand and identifying accessible hybridization sites on the other. Yet, compared with proteins, RNA structure prediction remains especially challenging: many RNAs adopt dynamic ensembles rather than a single stable conformation, contain complex motifs such as pseudoknots, and often lack strong evolutionary covariation signals because functional RNAs are frequently supported by shallow multiple-sequence alignments (Ganser et al., 2019; Rivas et al., 2020).

Early AI applications focused on RNA secondary structure prediction. DL methods such as SPOT-RNA, MXfold2, and UFold reframed folding as a data-driven pattern-recognition problem rather than relying only on thermodynamic models, improving prediction of base-pairing patterns, including pseudoknots and non-canonical interactions that often define ligand-binding motifs (Singh et al., 2019; Sato et al., 2021; Fu et al., 2022). These methods span a spectrum: SPOT-RNA and UFold are essentially end-to-end predictors, whereas MXfold2 integrates learned folding scores with Turner nearest-neighbor thermodynamic parameters, improving robustness on novel families (Sato et al., 2021). More recent work has shifted from task-specific folding toward RNA representation learning. RNA-FM, a self-supervised RNA language model, a network that, like language models learning word patterns from text, learns statistical patterns of nucleotides from large unannotated sequence collections and encodes them as reusable numerical embeddings, which generate transferable representations for secondary-structure prediction, contact modeling, and related tasks, letting sequence information be reused across applications, including identification of structured, potentially targetable regions (Chen et al., 2022).

Three-dimensional prediction has progressed rapidly and is especially consequential for pocket definition, since binding sites are ultimately three-dimensional. RoseTTAFoldNA extended DL prediction to protein–nucleic acid complexes, whereas RhoFold + combined RNA language-model representations with a dedicated structure module for single-chain 3D prediction (Baek et al., 2024; Shen et al., 2024). More recently, AlphaFold3 broadened modeling to biomolecular complexes containing proteins, nucleic acids, ions, modified residues, and small molecules (Abramson et al., 2024). This is directly relevant to drug discovery: a predicted model is the substrate on which pocket detection, druggability assessment, and docking operate, so pocket identification is bound by structure quality. These advances point to a shift from isolated predictors toward general frameworks integrating sequence, secondary structure, tertiary organization, molecular recognition, and cellular context, but important limitations remain and can propagate into pocket-level analysis. In the blinded, community-wide CASP15 assessment, RNA prediction methods showed substantial variation in accuracy across target classes and structural complexity (Das et al., 2023). Independent analyses likewise note that RNA prediction has not reached the reliability achieved for proteins, particularly for novel families and difficult folds (Schneider et al., 2023; Szikszai et al., 2022; Bernard et al., 2025; Liu et al., 2026). Because a mispredicted fold can create or erase an apparent pocket, structure-derived druggability assessments are best treated as testable hypotheses rather than settled predictions, with experimental data and physical modeling remaining indispensable.

Finally, many functionally relevant RNA conformations are transient, a particular challenge for pocket-based analysis, because binding-competent pockets may be cryptic, forming only in low-population states that a single static structure misses. Recent integrative and AI-based approaches address this by modeling conformational ensembles rather than single structures: NMR-guided prediction has enabled atomistic ensemble modeling of dynamic RNAs, cryo-EM increasingly resolves coexisting RNA conformational states that a single static model misses (Bonilla and Jang, 2024), and diffusion-based generative models sample diverse 3D conformations consistent with experimental observations (Shi et al., 2020; Li et al., 2025). For drug discovery, this ensemble-aware perspective is essential, a binding-competent pocket may appear in only a subset of conformers, and identifying it requires reasoning over the ensemble, a theme we return to in the context of induced fit recognition.

3 AI for RNA therapeutic design

RNA therapeutics have gained clinical momentum through both small molecules and oligonucleotides, but translating this into broadly applicable design remains challenging because RNA modulation requires coordinated optimization of molecular recognition, activity, selectivity, stability, toxicity, delivery, and disease context. AI-driven approaches are therefore being explored to integrate these variables and guide design more systematically. As summarized in Figure 1, the small-molecule and oligonucleotide branches rely on distinct methodological bases, structure- and dynamics-driven versus sequence- and data-driven. Table 1 compares representative methods in terms of input, output, structural requirements, strengths, and limitations.

TABLE 1. Comparison of each method’s inputs, outputs, methodological basis, distinguishing features, and key considerations.

Abbreviations: 2°, secondary structure; DL, deep learning; ML, machine learning; ANN, artificial neural network; GNN, graph neural network; SE(3), roto-translation–equivariant; MSA, multiple sequence alignment; NA, nucleic acid; siRNA, small interfering RNA; mRNA, messenger RNA; ASO, antisense oligonucleotide; Fpocket, geometry-based pocket-detection algorithm. Bracketed numbers are reference citations; representative tools are shown, not an exhaustive list.

3.1 RNA-targeting small molecules

The structural advances above provide a foundation for small-molecule discovery, but therapeutic design ultimately depends on how a small molecule recognizes a specific RNA sequence and how binding modulates downstream interactions, a problem complicated by the many unknown factors and the intrinsic complexity of RNA–ligand recognition. Our work on CAG (cytosine–adenine–guanine) repeat RNA illustrates the point: dimerization of the naphthyridine–azaquinolone (NA) unit produced a clear RNA-binding signal by surface plasmon resonance (SPR) that was absent for the NA monomer, shifting the binding stoichiometry from 2:1 to 1:1 and most likely reflecting an altered binding mode (Chen et al., 2024). In this case, recognition was therefore highly sensitive to ligand architecture, difficult to infer from chemical similarity alone, and to target flexibility, since the dimeric NA preferentially bound the flexible CAG overhang over a more rigid internal loop. Alongside RNA dynamics, structural heterogeneity, and limited ligand selectivity, these factors remain central bottlenecks for predicting RNA–small molecule binding.

Computational and AI-driven methods for predicting these interactions could be divided into structure-based and ligand-based approaches. Structure-based methods exploit target information, combining sequence-, structure-, and feature-based strategies to nominate accessible, functional RNA regions. Early examples include Rsite2, which uses secondary-structure-derived distance patterns to identify functional sites (Zeng and Cui, 2016), and RBind, which models RNA tertiary structure as a nucleotide-interaction network to predict candidate binding sites (Wang et al., 2018). More recent ML and DL methods build on this: Rlsite combines pre-trained RNA representations with graph attention networks to capture sequence-derived and structural features (Sun et al., 2025), fpocketR provides RNA-specific pocket detection to locate and characterize drug-like pockets (Veenbaas et al., 2025), and DRLiPS extends the framework from binding-site prediction to druggability assessment, prioritizing regions likely to support productive ligand engagement before virtual screening (Krishnan et al., 2025). In a related direction, RNAmigos2 uses graph-neural-network representations of RNA structure to predict RNA–ligand compatibility and accelerate virtual screening (Carvajal-Patiño et al., 2025). These methods differ in granularity: residue-level predictors (RBind, Rlsite) flag individual binding nucleotides, whereas pocket-based methods, fpocketR, an RNA adaptation of the geometry-based Fpocket algorithm, and DRLiPS, delineate 3D pockets and rank druggability, so residue-level output supports binding-site mapping, whereas pocket-level output provides structural sites for docking. A common limitation across these methods is their reliance on experimentally determined or modeled RNA structures. Ligand-based approaches, by contrast, exploit similarity between compounds, assuming that structurally similar ligands share similar activities, and do not require RNA structural information, but depend on the availability of known active compounds for a given target. The performance of these data-driven methods also depends on curated benchmark resources: R-BIND compiles literature-reported bioactive RNA-targeted ligands (Morgan et al., 2019), ROBIN provides a large experimentally derived library of RNA binders and non-binders from microarray screening (Yazdani et al., 2023) and Inforna maps RNA motif–small molecule interactions to enable sequence-based design (Velagapudi et al., 2014). Because these datasets remain modest and biased toward well-studied targets, generalization to novel RNA targets remains to be explored.

Because most of these methods depend on the accuracy of RNA structure information, an emerging direction predicts RNA–small molecule interactions more directly from sequence and ligand information. SMRTnet, for example, integrates RNA sequence, secondary structure, and compound representations to predict interactions, with prospective validation across disease targets (Fei et al., 2026). In a related structure-agnostic direction, SMARTBind couples an RNA language model with contrastive learning to identify small-molecule binders and their binding sites from sequence alone, reducing dependence on tertiary structure (Jiang et al., 2025). Such sequence-to-ligand models may broaden exploration of the RNA target space and couple target discovery more tightly with ligand prioritization. Furthermore, clustering can automatically label distinct binding-response patterns and kinetic profiles, providing deeper insights into binding behavior that are not captured by conventional intensity-threshold-based hit assignment (Chen et al., 2025). Together, these approaches are beginning to unite RNA targetability, conformational dynamics, and ligand binding within a more integrated in silico framework.

A recurring theme across these methods is that induced fit is not a peripheral complication but a core design problem: the binding-competent pocket is frequently shaped by the ligand rather than pre-formed in the free RNA (Koshland, 1958). Compounded by RNA’s conformational heterogeneity, recognition is governed not only by static structure but by dynamics, thermodynamics, and binding kinetics (Umuhire Juru et al., 2019). A small molecule may stabilize a pre-existing minor conformation, induce local rearrangements, or reshape the broader ensemble, so therapeutically relevant conformations may be transient, cryptic, or absent from a single static RNA structure. Ensemble-aware methods such as SHAMAN integrate molecular simulations, enhanced sampling, and small-molecule probes to identify ligandable sites across conformational ensembles (Panei et al., 2024), while related strategies, mixed-solvent simulations and AI-augmented molecular dynamics, assess whether a predicted pocket remains productive during recognition and whether binding remodels local structure (Kognole et al., 2022; Wang et al., 2022). Site- and interaction-prediction methods nominate promising regions and candidate binders, but docking and molecular dynamics are still needed to confirm how a ligand sits in a pocket and whether the complex is stable, harder for RNA than proteins, owing to its charged, flexible backbone, ion- and water-mediated contacts, and shallow polar sites, motivating dedicated engines such as rDock (Ruiz-Carmona et al., 2014). AI now assists at several points: learned scoring functions such as AnnapuRNA rescore poses more reliably than physics-based terms alone (Stefaniak and Bujnicki, 2021), diffusion-based models such as DiffDock, built on SE(3)-equivariant GNNs, reframe pose search as generative sampling with confidence estimates (Corso et al., 2023), flexible-docking methods such as DynamicBind model ligand-induced receptor changes (Lu et al., 2024), and end-to-end predictors such as AlphaFold3 generate protein–nucleic-acid–ligand complexes directly (Abramson et al., 2024). Generative chemistry designing RNA ligands de novo, a diffusion-model frontier productive for proteins, remains underexplored, constrained by scarce RNA–ligand data. Evaluating binding across an ensemble, rather than a single static pose, remains more robust for a target whose recognition is defined by induced fit.

3.2 Oligonucleotide design with AI

Unlike small molecules, which depend on structural recognition, oligonucleotides act through programmable base pairing. But potency and safety depend on more than complementarity, accessibility, strand architecture, chemical modification, nuclease stability, immune activation, off-target hybridization, and delivery, interdependent variables well suited to data-driven design (Egli and Manoharan, 2023). Chemical modification has been central to clinical maturation; bridged nucleic acids improved duplex stability and affinity (Obika et al., 1998), while phosphorothioate linkages and stereodefined chemistry contributed to nuclease resistance and pharmacological optimization (Egli and Manoharan, 2023; Oka et al., 2008). These advances also expanded the design space, where sequence, chemistry, stereochemistry, stability, activity, and safety must be considered together. This sequence-centric character is visible in Figure 1, where oligonucleotide methods cluster around sequence- and data-driven approaches rather than the structure- and dynamics-based methods that dominate small-molecule design.

Within this space, siRNAs, ASOs, and other nucleic-acid modalities have followed related but distinct computational trajectories. For siRNAs, early empirical rules provided the first practical framework for improving silencing efficacy by capturing features such as thermodynamic asymmetry, guide-strand selection, position-dependent nucleotide preferences, GC content, and target-site properties, reflecting the principle that effective RNAi requires not just complementarity but preferential guide-strand loading into RISC and productive target pairing (Ui-Tei et al., 2004; Reynolds et al., 2004). As activity datasets accumulated, BIOPREDsi showed that experimentally measured sequence–activity relationships could be learned with neural networks and applied to genome-scale library design (Huesken et al., 2005). More recent models treat efficacy as a sequence- and context-dependent interaction problem: OligoFormer integrates thermodynamic parameters, RNA-FM embeddings, and Transformer representations to predict siRNA efficacy from siRNA and mRNA sequences (Bai et al., 2024), and siRNADiscovery uses a GNN to combine empirical rules with non-empirical siRNA and mRNA features (Long et al., 2024). siRNA design has thus evolved from rule-based selection toward data-driven efficacy prediction based on sequence, target context, and interaction features.

ASO design poses a distinct, modality-specific problem. Unlike siRNAs, ASOs are single-stranded, and their optimal design depends strongly on mechanism of action: gapmer ASOs require architectures supporting RNase H-mediated cleavage, whereas splice-switching and steric-blocking ASOs act by occupying regulatory regions and altering RNA processing or RNA–protein interactions (Andersson et al., 2022; Egli and Manoharan, 2023). Activity therefore cannot be predicted from complementarity alone but depends on the coupling of binding position, local accessibility, transcript context, chemical modification pattern, and mechanism-specific readout. ML tools are beginning to address these requirements: eSkip-Finder uses curated exon-skipping datasets to predict sequences likely to promote exon skipping (Chiba et al., 2021), while ASOptimizer applies DL to optimize both sequence and chemical modification patterns for chemically modified, gene-regulating ASOs (Hwang et al., 2024).

Safety-aware modeling is especially important for ASOs, because toxicity can arise from off-target hybridization, protein binding, immune activation, tissue exposure, and mechanism-specific effects. Early random-forest models partly predicted hepatotoxic potential from sequence and modification pattern (Hagedorn et al., 2013). Motif-based studies then linked specific sequence motifs to LNA-gapmer hepatotoxicity. (Burdick et al., 2014), while mechanistic analyses showed that high-affinity gapmers can trigger RNase H1-dependent off-target reduction of long pre-mRNA transcripts (Bur et al., 2016). Work emphasizes that ASO toxicity emerges from synergistic effects among nucleotide sequence, chemical modifications, and the positional context of those modifications, rather than from any single factor (Bhamra et al., 2025). Across both modalities, then, computational oligonucleotide design is moving from empirical rule-based selection toward data-driven models that better capture the combined effects of sequence, target context, chemical modification, and safety.

4 Outlook

Although AI and ML are evolving rapidly, their application to RNA drug discovery is still limited by several unresolved challenges. High-quality training data remain scarce, particularly experimentally resolved RNA structures, RNA–ligand complexes, and large-scale oligonucleotide activity datasets, and existing datasets are biased toward well-studied RNA classes, chemistries, and assay conditions, limiting robustness on structurally novel RNAs or less-characterized targets. A second challenge is the gap between prediction and experimental validation: better benchmark performance does not necessarily translate into therapeutic success, especially when activity depends on cellular context, delivery, toxicity, and disease-specific mechanisms. Progress will depend not only on model accuracy but on tighter feedback between computation and experiment. Interpretability will remain essential as foundation models are applied to structure prediction, prioritization, and design, acute in RNA biology, where structural data are sparse, and many targets behave as dynamic ensembles rather than single structures. For small molecules, predictions should ideally trace to features such as local structure, tertiary contacts, and ligand-induced stabilization; for oligonucleotides, to accessibility, modification effects, off-target hybridization, and mechanism-dependent activity. Foundation models offer one route toward more generalizable representations: large-scale sequence pretraining, as in RNA-FM, already supports structure- and function-related predictions, and single-cell models such as Geneformer and scGPT indicate how cellular context could complement molecular representations (Theodoris et al., 2023; Cui et al., 2024).

Agentic systems may help integrate these steps: rather than treating target selection, candidate generation, efficacy prediction, and safety filtering separately, workflow-oriented AI, for which autonomous chemical research provides a precedent, could coordinate them into design–test–learn cycles (Boiko et al., 2023). Such systems will be most valuable when they help prioritize experiments, learn from failures, and refine design criteria, emphasizing validation and reproducibility. Ultimately, AI is unlikely to replace experimental discovery in RNA therapeutics. Its broader value may instead be as a reasoning layer linking RNA biology, design, safety evaluation, and therapeutic validation, essential for challenges beyond prediction accuracy, including delivery, tissue specificity, toxicity, and modality choice. The future impact of AI here will depend not only on more accurate models but on their ability to guide decisions across biological context, molecular modality, and translational feasibility.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The author used ChatGPT (OpenAI) solely to improve the readability and clarity of the manuscript text. No generative AI tools were used to generate scientific ideas, analyze data, draw conclusions, or create original scholarly content. The author reviewed and edited the manuscript and takes full responsibility for its final content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Urszula Katarzyna Komarnicka, University of Wrocław, Poland

Reviewed by: Amirhossein Taghavi, University of Florida, United States