Pep2Net: an integrated peptide knowledge base and cross-attention model for multi-task prediction of peptides

Computational peptide design is hindered by fragmented databases and models separating semantic from physicochemical features. We introduce Pep2Net, a comprehensive platform addressing these issues. First, we develop Pep2Net-DB, unifying 52 peptide resources into an ontology-inspired knowledge base. Second, we develop Pep2Net-XAtt, a multi-task model leveraging cross-attention and low-rank bilinear pooling to fuse protein language model embeddings with physicochemical descriptors. Pep2Net-XAtt achieves state-of-the-art performance in predicting cell-penetrating, anticancer toxicity, and multi-functional peptides. Accessible via Pep2Net Web, this framework provides a unified tool for large-scale peptide characterization and design.