Accurate prediction of antimicrobial peptide activity for various bacterial species based on protein language model

Abstract

In the field of antimicrobial peptide research, predicting species-specific minimum inhibitory concentration (MIC) is crucial for developing targeted and effective AMP development. This study proposes a novel approach leveraging pre-trained language models to predict the MIC of peptides against specific bacterial species. The research utilizes masked language modeling tuning with peptide sequences and diverse species information. The proposed method aims to enhance the performance of MIC predictions, thereby contributing to the screening of antimicrobial peptide design and application. Besides, this paper conducted an analysis that captures crucial amino acid tokens using the attention mechanism of the transformer. Through the ability to predict the activity of species about unseen species, this study contributes to design novel AMP about novel species.