Machine learning for RNA-targeting drug design

Targeting RNA with small molecules offers significant therapeutic potential. Machine learning could substantially accelerate preclinical drug discovery, from hit identification to lead optimization. Yet a fundamental limitation emerges: drug design machine learning models, tailored for proteins, are not readily applicable to RNAs because of fundamental differences between RNAs and proteins in both structural characteristics and interactions with small molecules. RNA-specific approaches have consequently emerged, primarily focusing on binding site identification and virtual screening. In this review, we comprehensively compare machine learning tools for RNA-targeting drug design according to the tasks they address, their methodology and their relevance in RNA-specific contexts. As open challenges will catalyze new method development, we emphasize the need for standardized, drug design-specific evaluation approaches. We provide clear guidelines to establish these standards along with a benchmark assessing the ability of current machine learning models to predict specific drug-RNA interactions.

💡 Research Summary

The paper provides a comprehensive review of machine learning (ML) methods tailored for RNA‑targeting drug design, highlighting the fundamental differences between RNA and protein targets that render protein‑centric ML models inadequate. RNA’s greater conformational flexibility, scarcity of high‑resolution structures (≈8,800 RNA entries in the PDB versus >230,000 protein entries as of 2025), and distinct physicochemical interaction patterns—dominated by electrostatics and π‑stacking rather than hydrophobic contacts—necessitate specialized approaches.

The authors map the pre‑clinical drug discovery pipeline onto six ML tasks: (1) binding site prediction, (2) pose generation, (3) pose scoring, (4) QSAR (ligand‑only scoring), (5) ligand prediction (target‑only scoring), and (6) binding affinity prediction (both ligand and target inputs). They systematically catalogue existing tools for each task, noting that the most successful binding‑site predictors are structure‑based geometric deep‑learning models such as MultimodRLBP and RLBSIF, while SMARTBind represents the first sequence‑based language model for this purpose. However, all current site‑prediction models are trained on holo (ligand‑bound) structures, limiting their applicability to the apo (unbound) structures that experimentalists typically possess.

In the docking domain, traditional protein‑focused tools (DOCK, AutoDock Vina) have been adapted for RNA, and RNA‑specific docking programs (MORDOR, rDock, RLDock) exist, yet they inherit the same limitations: rigid‑receptor assumptions, force‑field parameters ill‑suited for the highly charged RNA backbone, and prohibitive sampling costs. Deep‑learning‑based “deep docking” (e.g., EquiBind, DiffDock) and co‑folding approaches that accept RNA sequences and SMILES strings have shown promise for proteins but no comparable RNA‑specific models have been released.

Direct scoring methods are divided into three sub‑tasks. QSAR models use only ligand descriptors, ligand‑prediction models use only RNA features, and binding‑affinity models combine both. The review demonstrates that current state‑of‑the‑art models largely rely on ligand features, failing to capture RNA‑specific interaction nuances. To substantiate this claim, the authors construct a new benchmark and evaluate four recent models, finding that performance gains are driven primarily by ligand information.

Data considerations receive extensive treatment. The authors stress the need for curated datasets that distinguish apo vs. holo conformations, include multiple conformers per RNA to address flexibility, and separate quantitative binding constants (Kd, pKd) from binary activity labels. They warn against data leakage during train/validation/test splits and advocate for standardized benchmarking protocols. Evaluation metrics should go beyond ROC‑AUC and PR‑AUC to include enrichment factors and early‑recognition scores that reflect real‑world virtual‑screening efficiency.

In conclusion, the field faces three major challenges: (1) the paucity of RNA structural data and the necessity to generate multi‑conformer libraries; (2) the development of ML architectures that explicitly model RNA’s electrostatic and π‑stacking characteristics, possibly through physics‑informed transformers or equivariant networks; (3) the establishment of community‑wide benchmarks and evaluation standards. Addressing these gaps will enable ML to accelerate RNA‑targeted drug discovery from hit identification through lead optimization, ultimately expanding the therapeutic landscape beyond the single FDA‑approved non‑ribosomal RNA drug (risdiplam).