AMP-distillation: A knowledge distillation framework for accurate and efficient antimicrobial peptide prediction

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Immunology

Background:

• Antimicrobial peptides (AMPs) are key components of the innate immune system with broad-spectrum activity.
• Accurate identification and prediction of AMPs are challenging due to sequence diversity and imbalanced datasets.
• Existing methods struggle with the complexity and scale required for effective AMP discovery.

Purpose of the Study:

• To develop a novel computational framework, AMP-Distillation, for enhanced prediction of antimicrobial peptides.
• To leverage knowledge distillation to improve model efficiency and predictive accuracy.
• To address the challenge of severe class imbalance in AMP datasets.

Main Methods:

• Utilized a Transformer-based teacher model and a BiLSTM-based student model for knowledge distillation.
• Curated and processed protein sequences from APD3 and DADP databases using CD-HIT.
• Integrated Rotary Positional Encoding in the teacher model and guided the student with soft and ground-truth labels.

Main Results:

• AMP-Distillation achieved high performance metrics, including 99.14% accuracy, 95.13% sensitivity, and 99.47% specificity.
• The framework significantly outperformed state-of-the-art methods on imbalanced datasets.
• Reduced parameter count by nearly 50% while enhancing model generalization and stability.

Conclusions:

• AMP-Distillation offers an efficient and interpretable deep learning approach for large-scale AMP discovery.
• The framework provides a robust foundation for future antimicrobial drug design.
• Knowledge distillation effectively addresses challenges in AMP prediction, particularly data imbalance.