RPFace: From biased structural benchmarks to generalizable RNA-protein interface prediction

RPFace predicts protein-RNA binding preferences from sequence alone, matching state-of-the-art on protein interfaces without evolutionary information and outperforming on RNA interfaces in zero-shot settings.

• Why it matters: Reported co-folding and docking performance may reflect benchmark bias toward training-like deposited structures, and both families fail to transfer to in vivo CLIP experiments.
• What they did: The authors release a stratified benchmark across structural and experimental groups, then introduce RPFace, a small fast model that fuses RNA foundation and structure-prediction embeddings to predict residue and nucleotide binding preferences.
• The result: Co-folding binding-site accuracy concentrates in ribosomal complexes and varies by experimental modality, while docking works only with near-bound inputs and is inconsistent across RNA classes; RPFace generalizes to in vivo datasets without extensive pretraining.

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