The clinical momentum behind RNA-targeted medicine is no longer speculative. The approval of risdiplam, a small-molecule splicing modifier for spinal muscular atrophy, demonstrated that drug-like molecules can selectively modulate RNA-mediated processes. Building on that proof of concept, the oral splicing modulators branaplam and votoplam, which lower huntingtin through pseudoexon inclusion, reached phase 2 trials for Huntington’s disease; votoplam met its primary endpoint, while branaplam’s VIBRANT-HD study was halted due to neuropathy. In parallel, 2025 brought three new oligonucleotide approvals from the U.S. Food and Drug Administration: the siRNA therapeutics fitusiran and plozasiran, and the antisense oligonucleotide donidalorsen. Together with advances in oligonucleotide chemistry, safety, and delivery, these milestones underscore how RNA-based strategies are expanding across diverse disease areas.

Yet RNA remains a notoriously difficult drug target. Unlike proteins, RNA rarely presents deep hydrophobic cavities; instead, ligandable pockets often emerge from internal loops, bulges, junctions, and other higher-order structural motifs. Many RNAs adopt dynamic ensembles rather than a single stable conformation, contain complex motifs such as pseudoknots, and lack the strong evolutionary covariation signals that help protein structure prediction, because functional RNAs are frequently supported by shallow multiple-sequence alignments. For oligonucleotide drugs, the challenge is different but equally formidable: researchers must navigate vast 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 solve alone, and it is precisely here that deep learning, foundation models, multimodal learning, and emerging agentic systems are beginning to make their mark.

Accurate RNA structure prediction sits at the heart of both modalities. For small molecules, the three-dimensional fold defines the shape and accessibility of ligandable pockets; for antisense oligonucleotides and siRNAs, which act through Watson–Crick hybridization, efficacy depends on target-site accessibility, so stable local secondary structure can occlude the complementary region and reduce potency, while open, single-stranded regions are more favorable. Early AI applications focused on secondary structure, with deep learning methods such as SPOT-RNA, MXfold2, and UFold reframing folding as a data-driven pattern-recognition problem rather than relying solely on thermodynamic models. These tools improved prediction of base-pairing patterns, including pseudoknots and non-canonical interactions that often define ligand-binding motifs. SPOT-RNA and UFold operate as essentially end-to-end predictors, whereas MXfold2 integrates learned folding scores with Turner nearest-neighbor thermodynamic parameters, improving robustness on novel RNA families.

More recent work has shifted from task-specific folding toward general RNA representation learning. RNA-FM, a self-supervised RNA language model, learns statistical patterns of nucleotides from large unannotated sequence collections and encodes them as reusable numerical embeddings, generating transferable representations for secondary-structure prediction, contact modeling, and the identification of structured, potentially targetable regions. Three-dimensional prediction has progressed rapidly as well: RoseTTAFoldNA extended deep learning to protein–nucleic acid complexes, RhoFold+ combined RNA language-model representations with a dedicated structure module for single-chain 3D prediction, and AlphaFold3 broadened modeling to biomolecular complexes containing proteins, nucleic acids, ions, modified residues, and small molecules. Because pocket detection, druggability assessment, and docking all operate on predicted models, structure quality directly bounds what downstream analysis can achieve.

Important limitations persist, however. In the blinded, community-wide CASP15 assessment, RNA prediction methods showed substantial variation in accuracy across target classes and structural complexity, and independent analyses note that RNA prediction has not reached the reliability achieved for proteins, particularly for novel families and difficult folds. 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. Compounding the problem, many functionally relevant RNA conformations are transient: binding-competent pockets may be cryptic, forming only in low-population states that a single static structure misses. Recent integrative approaches address this by modeling conformational ensembles, with NMR-guided prediction enabling atomistic ensemble modeling, cryo-EM increasingly resolving coexisting RNA conformational states, and diffusion-based generative models sampling diverse 3D conformations consistent with experimental observations.

For small-molecule discovery, the review highlights how AI-driven methods divide into structure-based and ligand-based approaches. Structure-based tools exploit target information to nominate accessible, functional RNA regions: Rsite2 uses secondary-structure-derived distance patterns to identify functional sites, RBind models RNA tertiary structure as a nucleotide-interaction network, Rlsite combines pre-trained RNA representations with graph attention networks, fpocketR provides RNA-specific pocket detection, and DRLiPS extends the framework from binding-site prediction to druggability assessment. RNAmigos2 uses graph-neural-network representations of RNA structure to predict RNA–ligand compatibility and accelerate virtual screening. Ligand-based approaches, by contrast, exploit similarity between compounds and do not require RNA structural information, but depend on the availability of known active compounds. Curated benchmark resources such as R-BIND, ROBIN, and Inforna underpin these data-driven methods, though the datasets remain modest and biased toward well-studied targets, limiting generalization to novel RNA targets.

An emerging direction predicts RNA–small molecule interactions more directly from sequence and ligand information. SMRTnet integrates RNA sequence, secondary structure, and compound representations to predict interactions, with prospective validation across disease targets, while 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. 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. 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 any single static structure. Ensemble-aware methods such as SHAMAN integrate molecular simulations, enhanced sampling, and small-molecule probes to identify ligandable sites across conformational ensembles, while AI-augmented molecular dynamics assess whether a predicted pocket remains productive during recognition.

Oligonucleotide design follows a distinct, sequence-centric trajectory. Unlike small molecules, oligonucleotides act through programmable base pairing, but potency and safety depend on accessibility, strand architecture, chemical modification, nuclease stability, immune activation, off-target hybridization, and delivery, interdependent variables well suited to data-driven design. Chemical modification has been central to clinical maturation, with bridged nucleic acids improving duplex stability and affinity, and phosphorothioate linkages and stereodefined chemistry contributing to nuclease resistance and pharmacological optimization. For siRNAs, early empirical rules captured features such as thermodynamic asymmetry, guide-strand selection, position-dependent nucleotide preferences, and GC content, and BIOPREDsi showed that experimentally measured sequence–activity relationships could be learned with neural networks for genome-scale library design. More recent models treat efficacy as a sequence- and context-dependent interaction problem: OligoFormer integrates thermodynamic parameters, RNA-FM embeddings, and Transformer representations, while siRNADiscovery uses a graph neural network to combine empirical rules with non-empirical features. For antisense oligonucleotides, whose optimal design depends strongly on mechanism of action, tools such as eSkip-Finder predict sequences likely to promote exon skipping, and ASOptimizer applies deep learning to optimize both sequence and chemical modification patterns.

Safety-aware modeling is especially critical for antisense oligonucleotides, where 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, motif-based studies linked specific sequence motifs to locked nucleic acid gapmer hepatotoxicity, and mechanistic analyses showed that high-affinity gapmers can trigger RNase H1-dependent off-target reduction of long pre-mRNA transcripts. Recent 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. Looking forward, the review identifies high-quality training data as a persistent bottleneck, particularly experimentally resolved RNA structures, RNA–ligand complexes, and large-scale oligonucleotide activity datasets, and stresses that better benchmark performance does not necessarily translate into therapeutic success when activity depends on cellular context, delivery, toxicity, and disease-specific mechanisms. Agentic systems that coordinate target selection, candidate generation, efficacy prediction, and safety filtering into design–test–learn cycles may help, but the author concludes that AI is unlikely to replace experimental discovery in RNA therapeutics. Its broader value may instead lie as a reasoning layer linking RNA biology, design, safety evaluation, and therapeutic validation, guiding decisions across biological context, molecular modality, and translational feasibility.

Subject of Research: Artificial intelligence methods for RNA-targeted drug discovery, spanning RNA structure prediction, small-molecule binding-site identification, and oligonucleotide therapeutic design

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

Article References: Chen, Q. (2026). From small molecules to oligonucleotides: artificial intelligence for RNA-targeted drug discovery. Frontiers in Chemical Biology, 5, Article 1915343. https://doi.org/10.3389/fchbi.2026.1915343

Image Credits: AI Generated

DOI: 10.3389/fchbi.2026.1915343

Keywords: RNA drug discovery, artificial intelligence, machine learning, small molecules, oligonucleotides, siRNA, antisense oligonucleotides, RNA structure prediction, AlphaFold 3, RNA language models, drug design, deep learning