SMARTBind

SMARTBind (Small Molecule Approach to RNA Targeting Binder Discovery) predicts which small molecules bind a given RNA — and where they bind — directly from the RNA's primary sequence, without any structural input. RNA is an increasingly attractive drug target, but identifying small-molecule binders is hard: experimental screening is expensive, and computational docking depends on RNA structures that are scarce, flexible, and difficult to model. SMARTBind sidesteps structure entirely, learning to recognize binders from sequence.

The model was developed by Shiyu Jiang, Amirhossein Taghavi, Yanjun Li, Matthew D. Disney, and colleagues at the University of Florida — spanning the Herbert Wertheim UF Scripps Institute — with collaborators at Scripps Research, and posted to bioRxiv in September 2025. It combines an RNA large language model with contrastive learning and a ligand-specific decoy-enhancement strategy, casting binder discovery as a cross-modal matching problem between RNA sequences and chemical structures.

Rather than a per-target tool that must be refit for each new RNA, SMARTBind is a fixed trained model applied prospectively to new targets. The authors used it to discover novel small molecules against the precursor of the oncogenic microRNA miR-21, then confirmed those hits in vitro and in cells.

#Key Features

• Structure-agnostic prediction: Identifies small-molecule binders and their binding sites from RNA primary sequence alone, avoiding the need for experimental or predicted RNA structures.
• RNA language model backbone: Builds on RNA-FM, a foundation model pretrained on tens of millions of non-coding RNA sequences, to encode target RNAs.
• Contrastive matching with decoy enhancement: A contrastive objective aligns RNAs with their true binders, while a ligand-specific decoy-enhancement strategy counters data scarcity and sharpens discrimination against non-binders.
• Prospectively validated: Applied to the miR-21 precursor, SMARTBind nominated novel binders that were confirmed by in vitro and cellular assays, not just retrospective benchmarks.

#Technical Details

SMARTBind couples the RNA-FM language model — a BERT-style foundation model pretrained on roughly 23 million non-coding RNA sequences — with a contrastive learning framework that embeds RNA sequences and small molecules into a shared space where true binding pairs are aligned. A ligand-specific decoy-enhancement strategy augments training with hard negatives to improve generalization under limited labeled data. From sequence alone the model ranks candidate binders and localizes their binding sites, and across multiple benchmarks and case studies it outperforms both docking-based and other data-driven methods while substantially reducing computational cost. Code is released on GitHub under an MIT license, with trained model weights deposited on Zenodo.

#Applications

SMARTBind targets RNA-focused drug discovery, enabling medicinal chemists and chemical biologists to screen compound libraries against structured RNAs — microRNAs, riboswitches, viral elements, and other regulatory RNAs — without first solving their structures. Because it operates on sequence, it can be pointed at emerging or poorly characterized targets quickly, prioritizing candidates and binding sites for experimental follow-up and lowering the cost of early-stage RNA-targeted campaigns.

#Impact

By demonstrating accurate, structure-free binder prediction backed by prospective wet-lab validation against a disease-relevant microRNA precursor, SMARTBind offers a scalable alternative to docking for RNA-targeted small-molecule discovery. The miR-21 result — novel binders confirmed in vitro and in cells — is its strongest evidence that sequence-based prediction can drive real discovery. As an openly licensed preprint with released code and weights, its broader impact will depend on independent benchmarking across more RNA classes; the paper text itself is under a non-commercial, no-derivatives license, though the released model and code are more permissive.