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

Area of Science:

• Biotechnology
• Computational Biology
• Drug Discovery

Background:

• Multi-functional therapeutic peptides offer advantages in drug development and diagnostics.
• Wet-lab identification of peptide functions is resource-intensive, necessitating computational approaches.
• Challenges include data sparsity and complex label co-occurrence in peptide data.

Purpose of the Study:

• To develop an efficient computational framework for predicting multi-functional therapeutic peptides.
• To address data sparsity, long-tail distribution, and label co-occurrence issues.
• To improve the accuracy and efficiency of therapeutic peptide function prediction.

Main Methods:

• Proposed AMCL framework utilizing semantic-preserving data augmentation.
• Implemented multi-label supervised contrastive learning with hard sample mining.
• Employed a weighted combined loss (Focal Dice Loss and Distribution-Balanced Loss) and category-adaptive thresholding.
• Assessed interpretability using feature space analysis and Grad-CAM visualization.

Main Results:

• AMCL framework demonstrated superior performance in multi-functional therapeutic peptide prediction.
• Achieved significant improvements across key metrics: Absolute true, Accuracy, Macro-F1, and Micro-F1.
• Established a new state-of-the-art in the field of therapeutic peptide multi-functional prediction.

Conclusions:

• AMCL effectively addresses the challenges of predicting multi-functional therapeutic peptides.
• The proposed methods significantly enhance prediction accuracy and establish a new benchmark.
• AMCL offers a powerful computational tool for accelerating drug discovery and diagnostics.