Pep2Net unifies 52 peptide databases with a cross-attention model for multi-task peptide prediction

Summary. Researchers at Hunan University, Central South University, the University of Electronic Science and Technology of China and Shenzhen University of Advanced Technology reported in Genome Biology, published online on 14 October, Pep2Net, a platform for computational peptide characterisation. The authors describe peptide informatics as held back by fragmented databases and by models that keep language-model features and physicochemical features apart. Pep2Net-DB unifies 52 peptide resources into an ontology-inspired knowledge base, and Pep2Net-XAtt is a multi-task model that fuses protein language model embeddings with physicochemical descriptors through cross-attention and low-rank bilinear pooling.

Research context (RUO). The authors report state-of-the-art benchmark performance on several peptide property labels, including cell-penetrating and multi-functional peptide classes, and make the framework available through a web interface, Pep2Net Web. The paper is a computational methods and database study: its outputs are predictions that still need laboratory validation, and it does not report wet-lab, clinical or human-use results. Supplementary material covers the model architecture, protein language model encoder comparisons, baseline comparisons, interpretability and robustness analyses. For peptide informatics and screening groups, the contribution is a single curated knowledge base plus an open multi-task predictor rather than any product claim.