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…
Protein-RNA complexes are typically modeled using co-folding and docking methods, but these techniques are assessed on benchmarks that resemble the structures on which they were initially trained. This raises questions about whether their performance truly reflects generalization. In this study, researchers focus on the binding interface, which is the region responsible for biological signaling.
They create a fair benchmark that categorizes protein-RNA data into different structural and experimental groups, leading to a more rigorous test of generalization compared to aggregate averages. The findings indicate that co-folding methods are most accurate in ribosomal complexes, while accuracy varies with experimental modality for both co-folding and docking techniques.
Docking methods prove reliable only when input conformations are near the bound state and are inconsistent across RNA classes. Neither method transfers well to in vivo CLIP experiments. To tackle this issue, the researchers introduce RPFace, a model that predicts the binding preference of each protein residue and RNA nucleotide directly from sequence, without requiring experimental structures, templates, or multiple sequence alignments.
RPFace employs an ensemble of predicted secondary structures combined with a multi-modal embedding-enrichment scheme that integrates representations from RNA foundation models and structure prediction models. Compared to existing approaches, RPFace is a compact and swift model that maintains consistent performance across various structural and experimental conditions.
It demonstrates resilience to conformational changes and generalizes effectively to in vivo datasets without the need for extensive pretraining. As a result, RPFace achieves parity with state-of-the-art methods on protein interfaces without evolutionary data and surpasses them on RNA interfaces, particularly in zero-shot scenarios, all while utilizing only sequence information. The model, along with the associated code and datasets, is accessible on GitHub.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
