RPI-Zero: Zero-shot RNA-Protein Interaction Prediction

RPI-Zero delivers the highest scores across eight metrics for zero-shot protein–RNA interaction prediction, using a sequence-only model evaluated on unseen RNAs, proteins, and pairs.

• Why it matters: CLIP-seq coverage is limited to about 10% of ~2,000 human RNA-binding proteins and requires antibodies, while existing predictors struggle when whole molecules are held out.
• What they did: The authors built a strict benchmark of 2.1 million examples from 11.6 million CLIP peaks targeting 209 proteins, using spliced transcripts with nucleotide-level labels and similarity-cluster splits plus length- and protein-count-matched negatives. They then trained RPI-Zero to jointly read full-length protein and RNA up to 4,096 tokens, conditioning RNA on protein via shared motif prototypes.
• The result: Compared with five state-of-the-art methods, RPI-Zero achieves the top score in eight metrics on unseen RNAs, unseen proteins, and unseen pairs, enabling interaction prediction for newly discovered or designed molecules without interaction data.

AI summary of the abstract · may contain errors · not medical advice · read the paper.