CoBRA

RNA has become an attractive target for small-molecule drugs: protein-coding sequence accounts for only a small fraction of the transcriptome, and folded RNAs form specific three-dimensional pockets that ligands can occupy. Finding those pockets computationally has, until recently, meant starting from a structure. The established RNA–ligand binding site predictors — Rsite, Rsite2, RBind, RNAsite, RLBind, RNABind, MVRBind, ZeSTa — all consume experimental or predicted coordinates, which restricts them to the small subset of RNAs with reliable 3D models.

CoBRA (Compound Binding site prediction for RNA) removes that dependency. Developed by Wonkyeong Jang and Woong-Hee Shin in the Department of Biomedical Informatics at Korea University College of Medicine, it labels each nucleotide in an RNA sequence as ligand-binding or not, using nothing but the sequence. Residue-level embeddings are drawn from a pretrained RNA language model held completely frozen, and a small multi-layer perceptron performs the binary classification on top of them, applying the "thin probe on a frozen backbone" pattern to a task where structural input has been the bottleneck.

The work first appeared as a bioRxiv preprint in September 2025 and was published in Briefings in Bioinformatics in January 2026. Code, trained weights, and all benchmark datasets are released under GPL-3.0.

#Key Features

• Structure-free prediction: Binding nucleotides are inferred directly from sequence, with no coordinates, distance matrices, or predicted structures required, so the method applies to RNAs that have no experimental model.
• Frozen backbone, trained head: The RNA language model is never fine-tuned; only the classifier is trained. Training the head takes under two minutes and the released checkpoint is 6.5 MB.
• Systematic backbone and objective sweep: Ten RNA language models and six loss functions were trained in combination, making the released configuration an empirically selected point rather than a default.
• Held-out cross-dataset evaluation: A single trained model is applied without retraining to four independent benchmark sets curated by other groups.
• Open release with weights: The repository ships the trained classifier, the training and benchmark splits, and a command-line pipeline that takes a FASTA file and returns per-nucleotide predictions.

#Technical Details

The released model uses RiNALMo as its embedding source, producing 1280-dimensional per-nucleotide vectors that are consumed frozen. On top sits a five-layer perceptron with hidden widths of 1024, 256, 128, and 64, each followed by layer normalization and ReLU, with dropout of 0.1, emitting two logits per residue that a softmax converts to a binding probability. Inputs are zero-padded to a maximum length of 161 nucleotides. The backbone was chosen from a candidate pool spanning non-coding-RNA models — ERNIE-RNA, RNA-FM, RNA-MSM, and RNABERT — and mRNA-specific models including SpliceBERT and 5' UTR-LM; plain binary cross-entropy outperformed class-balanced focal, Tversky, Dice, Lovász hinge, and a composite triplet-center objective.

Training data combined the TR60 and HARIBOSS sets, yielding 432 unique RNA chains after removing crystallization additives such as water, sulfate, phosphate, glycerol, ethylene glycol, and PEG, split 8:1:1 into training, validation, and internal test partitions. On the four external benchmarks the model reaches MCC / AUROC / recall of 0.557 / 0.844 / 0.650 on RB9, 0.546 / 0.816 / 0.599 on TL12, 0.293 / 0.739 / 0.531 on JL10, and 0.190 / 0.647 / 0.357 on TE18. The authors report a 22.1% relative gain in Matthews correlation coefficient and a 45.6% increase in sensitivity over structure-based state of the art; these are author-reported comparisons, not an independent evaluation. Embedding generation takes about 111 ms per sequence on an Ampere-class GPU, and inference on a benchmark set completes in well under a second.

#Applications

The intended use is early-stage RNA-targeted drug discovery: nominating candidate pockets on riboswitches, viral regulatory elements, long non-coding RNAs, and other transcripts for which no experimental structure exists, and narrowing the search space before docking or molecular dynamics. Because the classifier is small and the only heavy step is a single forward pass through the language model, the pipeline scales to transcript libraries rather than individual targets.

#Impact

CoBRA's central claim is a methodological one: for RNA–ligand site prediction, the statistical context captured in a pretrained sequence model can substitute for explicit geometry, and the trainable component needed on top is small. Its comparison of ten RNA language models on one downstream task is also a useful data point for a field where backbone selection is often made by convenience. The limitations are real and acknowledged by the authors: performance drops for binding sites in concave structural regions and at RNA dimer interfaces when only a single chain is supplied, the 161-nucleotide input cap excludes longer transcripts, and 432 training chains is a small corpus by foundation-model standards. The wide spread across benchmarks — MCC of 0.557 on RB9 against 0.190 on TE18 — indicates that generalization remains dataset-dependent.