Transfer learning on protein language models improves antimicrobial peptide classification.

Area of Science:

• Computational biology
• Immunology
• Bioinformatics

Background:

• Antimicrobial peptides (AMPs) are crucial for innate immunity and fighting pathogens.
• AMP classification is vital for therapeutic development, especially against antibiotic resistance.
• Limited labeled data challenges traditional AMP classifier training.

Purpose of the Study:

• To evaluate publicly available Protein Language Models (PLMs) for AMP classification.
• To benchmark PLMs against existing neural network classifiers using transfer learning.
• To investigate the impact of model scale and fine-tuning on AMP classification performance.

Main Methods:

• Utilized transfer learning with publicly available PLMs.
• Benchmarked PLM embeddings with shallow classifiers.
• Compared performance against state-of-the-art neural classifiers.
• Investigated the effect of PLM fine-tuning.

Main Results:

• Larger PLMs consistently yielded better classification performance.
• PLM embeddings with shallow classifiers achieved state-of-the-art results with minimal effort.
• Efficient fine-tuning of PLMs further boosted classification accuracy.

Conclusions:

• PLMs offer a powerful and efficient solution for AMP classification, even with limited labeled data.
• Model scale and fine-tuning are key factors for optimizing PLM performance in AMP classification.
• This approach holds significant promise for accelerating the discovery and development of novel AMPs.