Cross-Species Protein Interaction Prediction by Hierarchical Contrastive Pretraining

Shiyi LIU, Buwen Liang, Yuetong Fang, Zixuan JIANG, Ruohao Guo, Renjing Xu

Keywords: Cross-species protein interaction, Hierarchical learning, Interpretability

TL;DR: HIPPO improves both intra- and cross-species protein interaction prediction with hierarchical supervision and interpretable embeddings.

Abstract: Protein-protein interaction prediction is fundamental for understanding cellular processes, yet most existing approaches struggle with both intra-species accuracy and cross-species generalization. We present HIPPO, a hierarchical contrastive framework for protein-protein interaction prediction across organisms. HIPPO integrates amino acid sequences, biological hierarchies, and functional annotations into a unified representation learning objective. By aligning proteins not only at the sequence level but also according to their hierarchical relationships, HIPPO enforces embeddings that reflect the multi-level organization of protein functions. This structured supervision enables more accurate predictions within species while also facilitating transfer to unseen proteins and species. To capture global network context, protein embeddings are propagated through interaction graphs using graph neural architectures. Experiments on benchmark datasets demonstrate that HIPPO achieves consistent state-of-the-art performance, with substantial improvements in both intra-species and cross-species prediction. Crucially, extensive interpretability analyses reveal that hierarchical supervision highlights conserved motifs, binding residues, and post-translational modification regions, yielding biologically grounded interpretability and improving the reliability of protein interaction discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 7729