RPI-Zero predicts RNA-protein binding from sequence alone, leading all eight metrics on unseen RNAs, proteins and pairs

RNA-protein interaction prediction can now target molecules absent from training: one model covers all proteins and RNA classes, reads sequence only, and needs neither structure nor interaction data for a new molecule. Earlier predictors fell into two families: one trained a separate model per protein and could not move beyond proteins with CLIP data, while the other scored pairs but targeted a single RNA class. Most were also evaluated with random splits and randomly paired negatives that left near-identical sequences on both sides of the split. On a benchmark split by whole sequence-similarity clusters, RPI-Zero has the highest value on all eight metrics on all three test slices, unseen RNA, unseen protein and unseen pair; its AUPRC is 0.753, 0.637 and 0.589, 2.7%, 7.4% and 4.6% above the second-best method on each slice, and that second-best method differs from slice to slice.

The model reads protein and RNA together, so each nucleotide's representation already depends on the protein, and it assigns a binding score to every nucleotide. RNA-X is a transformer with 14 layers and 136 million parameters, pretrained on over 100 million RNA interactions, that reads at most 4,096 tokens shared by the protein and the RNA. RPI-Zero keeps RNA-X frozen and trains only a 2-million-parameter head: the head describes each protein as a mixture of 16 shared motif prototypes, so an unseen protein is described by the same vectors as training proteins and only its weights are new; training takes 12 hours on one NVIDIA RTX PRO 6000 GPU.

The demonstrated scope is a benchmark built from CLIP data in four public databases: 11.6 million peaks, 209 proteins and 2,074,753 examples, with labels covering only the interiors of exons in one chosen transcript per gene, so binding in introns, across exon junctions and on other isoforms is not labeled. The intended reader is a researcher who needs to rank binding likelihood for unprofiled or designed molecules; the model has no cell-type input and predicts whether a protein can bind an RNA, not whether it does so in a given cell.

Protein count is limited by available CLIP data: the 209 proteins form only 181 clusters at 30% identity, and the unseen-protein and unseen-pair slices each hold only a small number of proteins, so comparisons use per-protein bootstrap intervals and corrected tests. The model is less accurate on unseen proteins and unseen pairs than on unseen RNAs. CLIP finds binding sites only on RNAs expressed in the profiled cells, so some negatives may be bound in other cells. The 4,096-token limit of RNA-X makes long transcripts scored chunk by chunk, and combining chunk scores into one score per transcript is left open.