Work

GraPhiT-RBP: Graph Transformers and Residue-Level Cross-Attention for Interpretable RNA–Protein Binding Prediction

Year: 2026
Type: preprint
Abstract: Accurate prediction of RNA-protein binding is essential for understanding post-transcriptional regulation, yet existing computational approaches typically aggregate features globally, which limits the spatial resolution required to map specific mechanistic determinants of binding. Here, we introduce GraPhiT-RBP, a protein-aware framework that employs graph-guided cross-attention to explicitly model relationships between RNA nucleotides and individual protein residues. By preserving molecular resolution during interaction learning, GraPhiT-RBP enables direct biological interpretation: it recovers experimentally characterized RNA-binding motifs, maps spatial attention directly to known RNA-binding domains, and predicts the disruptive effects of clinically annotated single-nucleotide variants. Evaluated across 261 transcriptome-wide datasets spanning six human cell lines, this architecture achieved superior predictive accuracy (mean AUC 0.9326) and robust zero-shot generalization to previously unseen proteins compared to eight established methods. Notably, GraPhiT-RBP extracts these transferable recognition features using 69% fewer trainable parameters than the strongest-performing baseline. Together, GraPhiT-RBP offers a highly generalizable, mechanistically interpretable framework for characterizing molecular recognition and investigating previously unexplored RNA-protein regulatory networks.
Language: English
Open Access status: green
OpenAlex ID: W7226034500
Cites: 28
Cited by:
Related to: 10