CoCoBind

A screen that reports a compound binds a riboswitch, without saying where, gives a medicinal chemist nothing to design against. A binding-site predictor that marks a pocket on every RNA it is shown, whether or not the compound in hand goes there, gives them a pocket they cannot trust. The two questions are usually answered by separate models, and nothing obliges their answers to agree — a sequence can be scored as a non-binder while the site head still lights up half its nucleotides.

CoCoBind closes that gap by predicting both in a single forward pass and then tying them together during training. A Noisy-OR term collapses the per-nucleotide site probabilities into the probability that some nucleotide binds, and a consistency loss penalizes any disagreement between that aggregate and the model's own interaction score. The site map therefore has to explain the interaction call, and the interaction call has to be supported somewhere on the sequence. It was developed at Macao Polytechnic University with collaborators in Suzhou, Guangzhou, Guiyang and Singapore, and published in the Journal of Medicinal Chemistry in 2026.

The model is built on frozen pretrained encoders rather than trained from scratch, which is what makes it usable on a field with very little labelled data: RNA-targeted small-molecule discovery has one landmark approved drug in risdiplam and a few thousand curated interactions to learn from. RNA-FM supplies per-nucleotide RNA representations and Ouroboros supplies molecular ones, an arrangement shared with sibling RNA–ligand predictors such as DLRNA-BERTa from a separate group.

#Key Features

• Joint interaction and site prediction: One checkpoint emits a pair-level binding probability and a per-nucleotide binding-site probability vector, rather than requiring a screening model and a pocket-finding model to be run and reconciled by hand.
• Noisy-OR consistency constraint: Site probabilities are aggregated as 1 − Π(1 − sᵢ) over unmasked positions and matched to the interaction probability under a mean-squared-error penalty, weighted at 0.5 against the task losses.
• Cross-modal cross-attention: RNA tokens act as queries and the molecule's tokens as keys and values, so every nucleotide's representation is updated with respect to the specific compound being scored.
• Sequence and SMILES input only: No experimental or predicted RNA structure is needed at inference, which matters because most therapeutically interesting RNAs are conformationally heterogeneous and poorly resolved.
• Library screening built in: A screen command ranks a compound library against one RNA sequence, and a FastAPI backend with a React front end can be run locally for single-pair prediction and batch jobs.

#Technical Details

RNA-FM embeddings, precomputed and cached rather than fine-tuned, are projected to a 256-dimensional model width for sequences up to 512 nucleotides. The compound is encoded either as a 2048-bit ECFP4 fingerprint or as an Ouroboros embedding, then reshaped into eight virtual molecule tokens. An eight-head cross-attention layer fuses the two, feeding an interaction head over the pooled representation and a site head over each nucleotide. Training combines binary cross-entropy on the interaction label, a masked binary cross-entropy site loss computed only on positive pairs, the Noisy-OR consistency term, and an optional contrastive alignment objective; optimization is AdamW with a cosine schedule and warmup.

Evaluation uses the four splits of the DeepRNA-DTI benchmark — Unseen Pair, Unseen RNA, Unseen Compound and Unseen Both — each with five cross-validation folds. The headline result is Unseen Both, where neither the RNA cluster nor the compound cluster appears in training, and where CoCoBind's interaction-prediction gain over prior methods is clearest. Structure-level analysis of predicted sites against solved complexes serves as a localization check on the site head rather than a separate benchmark.

#Applications

The intended use is hit prioritization in RNA-targeted small-molecule programs: triaging a purchasable library against a specific RNA element — a pre-miRNA hairpin, a riboswitch, a structured untranslated region — and getting back both a ranked list and a nucleotide-level hypothesis about where each candidate engages. For a wet-lab group, the site map is the part that converts a ranked list into an experiment, since it nominates positions for mutational or chemical-probing follow-up rather than leaving the prediction unfalsifiable.

#Impact

CoCoBind is a peer-reviewed demonstration that coupling the two halves of RNA–ligand prediction helps most exactly where the field is weakest: generalizing to RNA targets and chemical scaffolds absent from training. The implementation is released under the MIT license with the benchmark splits bundled in the repository, so the results are reproducible end to end, though trained checkpoints are not distributed and must be produced by rerunning training. Absolute performance in the hardest split remains modest — a constraint the benchmark's own authors attribute to the scarcity of curated RNA–compound data rather than to model design — so the practical role is candidate ranking ahead of experimental screening, not replacement of it.