AMPDeep

AMPDeep predicts the hemolytic activity of antimicrobial peptides using transformer-based deep learning and transfer learning for accurate sequence-based classification.

Key Features:

• Deep Learning Framework: A transformer-based model automatically extracts sequence features for hemolytic activity prediction.
• Pre-trained Protein Language Model: Self-supervised pre-training via masked amino acid prediction on unlabeled protein sequences provides contextual peptide embeddings.
• Transfer Learning: The pre-trained model is fine-tuned on small peptide datasets and leverages knowledge from other protein and peptide databases to address data scarcity.
• Dataset Utilization: Employs three distinct datasets for therapeutic and antimicrobial peptides and a combined dataset to mitigate sequence similarity issues.
• Performance Optimization: Uses hyper-parameter optimization and selective fine-tuning to achieve state-of-the-art performance on multiple hemolysis datasets using only peptide sequences.

Scientific Applications:

• AMP screening and development: Predicts potential hemolytic effects to inform design and screening of antimicrobial peptides for therapeutic applications.
• Non-hemolytic candidate identification: Supports identification of non-hemolytic AMPs to accelerate discovery of safer antimicrobial agents.

Methodology:

Uses a transformer-based protein language model pre-trained by self-supervised masked amino acid prediction on unlabeled sequences, then applies transfer learning by fine-tuning on small peptide datasets (three therapeutic/antimicrobial datasets and a combined dataset), with hyper-parameter optimization, selective fine-tuning, and evaluation on multiple hemolysis datasets.