Abstract
Owing to their diverse biological activities, multifunctional therapeutic peptides (MTPs) have emerged as a promising direction in precision medicine and targeted therapy. However, the inherent multilabel characteristics and severe class imbalance of MTPs pose major challenges to accurate computational prediction. To address these issues, we propose a deep learning model named Multisource Enhanced Therapeutic Peptide Function Prediction via Adapter Network (METFAN), which integrates multisource feature representations with tailored optimization and aggregation strategies. Specifically, METFAN combines local sequence features captured by a multiscale TextCNN with global semantic embeddings from two pretrained protein language models, ESM2 and ProtT5. Because the raw embeddings of ESM2 and ProtT5 perform poorly across most functional categories, we designed a feature optimization module to refine them, enhancing sensitivity and discriminative capacity while preserving robustness in certain classes. In addition, a feature aggregation network effectively integrates heterogeneous features, absorbs complementary strengths, and reduces redundancy. Experimental results show that METFAN outperforms state-of-the-art methods, achieving a sample-level accuracy of 0.623 and a label-level F1-score of 0.522. Moreover, METFAN demonstrates superior robustness and generalizability under severe label imbalance. Overall, METFAN provides a novel and effective framework for MTP prediction and a solid foundation for peptide function screening and mechanistic studies. The data and code are publicly available at https://github.com/szlstart/METFAN.