RPcontact

Knowing which protein residues touch which RNA nucleotides is the practical currency of RNA-protein biology: it constrains docking, interprets crosslinking experiments, and localizes the interfaces that drive splicing, transcription, and translation. Getting those contacts has been hard for a structural reason rather than an algorithmic one — the Protein Data Bank holds comparatively few RNA-protein complexes, and the ones it does hold are dominated by ribosomes. Structure predictors such as AlphaFold 3 and RoseTTAFoldNA can be mined for binary contacts, but their accuracy on RNA-protein interfaces still leans on homologous sequences and templates that many targets do not have.

RPcontact, from Yaoqi Zhou's group at Shenzhen Bay Laboratory with collaborators at Guangzhou National Laboratory, ShanghaiTech University, and Tongji University, attacks the data problem from the other side. Instead of scaling training data, it trains a small contact head on top of two frozen language models — ERNIE-RNA for the nucleotide sequence and ESM-2 for the protein sequence — and learns exclusively from ribosomal RNA-protein complexes, the one family the PDB covers densely. The bet is that pretrained representations carry enough transferable signal that a ribosome-only training set still generalizes.

That bet holds. Posted to bioRxiv in June 2025, RPcontact transfers without retraining to dimeric and multimeric non-ribosomal complexes, and on a non-redundant benchmark excluding AlphaFold 3 training targets it reaches an auROC of 0.77 against 0.61 for contacts read off AlphaFold 3 structures — roughly a 26% relative gain.

#Key Features

• Language models in place of alignments: Contacts are predicted from two sequences and their pretrained embeddings, with no multiple sequence alignment, template search, or homology database lookup at inference time.
• Ribosome-only training, general application: The model sees only rRNA-protein complexes during training yet is evaluated, unchanged, on structurally unrelated dimeric and multimeric complexes.
• Cross-attention contact decoder: Separate self-attention over each molecule is followed by three rounds of RNA-protein cross-attention, so intra-molecular context and inter-molecular pairing are learned in distinct stages.
• Very small trainable footprint: Only 696,433 weights are trained; the heavy representation work is done by backbones that stay fixed, and embeddings can be precomputed once per sequence.
• Graceful degradation: When precomputed embeddings are unavailable the released implementation falls back to one-hot encoding, a mode whose accuracy cost is quantified in the paper's ablation of the two language-model inputs.

#Technical Details

The embedding module concatenates 768-dimensional ERNIE-RNA nucleotide embeddings, 5120-dimensional embeddings from the 15B-parameter ESM-2 protein model, and one-hot sequence encodings; self-attention then operates in a 48-dimensional space before three iterations of cross-attention produce the contact map. Predictions are capped at 24 protein residues per nucleotide and 12 nucleotides per residue. Training used 511 rRNA-protein complexes with 49 held out for validation, a 5 Å heavy-atom contact definition, and augmentation by random windows of 32-50 positions over dense local contact maps; optimization used RMSProp with cosine annealing, batch size 32, and a learning rate of 1e-4 for 38 epochs under early stopping.

On held-out data the model reaches auROC 0.85 on the ribosomal validation set, 0.82 on 125 multimeric and 40 dimeric non-rRNA complexes, and 0.77 on a 28-target non-redundant benchmark, versus 0.61 for AlphaFold 3, 0.55 for RoseTTAFoldNA, and 0.50 for an evolutionary-coupling baseline. MCC gains over both structure predictors are significant (p = 0.005 and p = 0.015). Ablations show ERNIE-RNA outperforming RNA-FM as the RNA encoder, and both language models contributing over one-hot input alone.

#Applications

The immediate users are structural and RNA biologists who have sequences but no structure: predicted contact maps supply distance restraints for integrative modeling, prioritize residues and nucleotides for mutagenesis or crosslinking follow-up, and flag candidate riboregulation sites, as the authors demonstrate for tRNA-protein complexes. Because inference needs no homology search, screening many candidate RNA-protein pairs is cheap relative to running a full structure predictor on each one.

#Impact

RPcontact is a compact demonstration that frozen sequence foundation models can substitute for training-set scale in a data-poor structural domain, and that a narrow, biased training corpus need not produce a narrow model. The caveats are real: it predicts contacts rather than coordinates, its training signal comes entirely from ribosomes, and its behavior on RNA-protein complexes far outside the tested dimeric and multimeric sets is uncharacterized. The work is a preprint awaiting peer review, released as MIT-licensed code with inference scripts and embedding helpers.