AMCL: supervised contrastive learning with hard sample mining for multi-functional therapeutic peptide prediction.

Abstract

Background: Multi-functional therapeutic peptides have emerged as promising candidates in drug development and disease diagnosis due to their biocompatibility, targeting capability, and low immunogenicity. However, the identification of peptide functions through wet-lab experiments is both time-consuming and costly, necessitating efficient computational prediction methods. The field faces challenges such as long-tail distribution problems, data sparsity, and complex label co-occurrence patterns due to peptides' multi-functional nature.

Results: To address these challenges, we propose AMCL, a novel framework for multi-functional therapeutic peptide prediction. AMCL incorporates a semantic-preserving data augmentation strategy, a multi-label supervised contrastive learning mechanism with hard sample mining, and a weighted combined loss combining Focal Dice Loss (FDL) and Distribution-Balanced Loss (DBL) to alleviate class imbalance issues. Additionally, we introduce a category-adaptive threshold selection mechanism for individual functional categories. The interpretability of AMCL is demonstrated through feature space analysis and Gradient-weighted Class Activation Mapping (Grad-CAM) visualization.

Conclusions: Comprehensive experiments show that AMCL significantly outperforms existing methods across multiple key metrics, including Absolute true, Accuracy, Macro-F1, and Micro-F1, establishing a new state-of-the-art in therapeutic peptide multi-functional prediction.