Knowledge-Augmented Spectral Hypergraph Learning for Protein Complex Identification in Network-Based Drug Discovery.

Abstract

Protein complexes are fundamental modular units of cellular organization, and their accurate identification is critical for network-based drug discovery. However, most computational approaches rely on pairwise interactions or heuristic clustering, limiting their ability to capture higher-order cooperative patterns in noisy protein-protein interaction (PPI) networks and often neglecting biological constraints such as subcellular colocalization and temporal coordination. To address these challenges, we propose KSHL-PC, a knowledge-augmented spectral hypergraph learning framework for protein complex identification. KSHL-PC constructs clique-induced hypergraphs to model multi-protein cooperation and employs spectral hypergraph embedding to capture multi-scale structural dependencies. A knowledge-enhanced feature transformation module integrates Gene Ontology cellular component annotations and temporal activity signals, while a permutation-invariant discriminator enables end-to-end complex scoring without hand-crafted rules. Experiments on multiple benchmark datasets show that KSHL-PC consistently outperforms state-of-the-art methods in F1-score, AUPRC, and complex-level accuracy. The predicted complexes also exhibit strong functional coherence, subcellular colocalization, and transcriptional coordination. These results demonstrate that KSHL-PC provides an effective framework for identifying biologically meaningful protein complexes and supports downstream network-based drug target discovery. The code and datasets are publicly available at: https://github.com/weiyu1218/KSHL-PC.