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

Modeling protein-RNA complexes with co-folding and docking methods is routine, yet these methods are evaluated on benchmarks whose composition mirrors the deposited structures on which they were trained or tuned, leaving open whether their reported performance reflects true generalization. Here, we focus the evaluation on the region that carries the biological signal: the binding interface. For the first time, we construct and release a fair benchmark that stratifies protein-RNA data into distinct structural and experimental groups, providing a more stringent test of generalization than aggregate averages. On this benchmark, we find that accuracy of detecting the binding site using co-folding methods is concentrated in ribosomal complexes and varies systematically with factors such as experimental modality. For docking methods, it is reliable only when the input conformations are already close to the bound state, and it is inconsistent across RNA classes. Both families of methods fail to transfer to in vivo CLIP experiments. To address this gap, we introduce RPFace, to the best of our knowledge the first model that jointly predicts the binding preference of every protein residue and every RNA nucleotide directly from sequence, requiring no experimental structure, template, or multiple sequence alignment. RPFace combines an ensemble of predicted secondary structures with a multi-modal embedding-enrichment scheme that fuses representations from RNA foundation models and structure prediction models. Unlike existing approaches, it is a very small and fast model that delivers stable performance across structural and experimental conditions, is robust to conformational change, and generalizes to in vivo datasets without extensive pretraining. As a result, RPFace matches the state-of-the-art methods on protein interfaces with no evolutionary information, and outperforms them on RNA interfaces, with the largest gains in zero-shot settings despite using only sequence information. The model, code, and data are available at Github.