NRPStransformer, an Accurate Adenylation Domain Specificity Prediction Algorithm for Genome Mining of Nonribosomal Peptides

Nonribosomal peptides serve as pivotal sources for drug discovery. Accurate prediction of the substrate specificity of adenylation domains in nonribosomal peptide synthetases is crucial for genome mining of nonribosomal peptides, yet current prediction methods fall short in accuracy. In this work, we analyzed 4,100 adenylation domains from documented nonribosomal peptide synthetases and found that the flavodoxin-like subdomain universally governs substrate specificity in all bacterial adenylation domains and that its phylogenetic analysis can correlate the sequences of adenylation domains and their substrate specificity. Leveraging the sequences within the flavodoxin-like subdomain, we developed a substrate specificity prediction algorithm using a protein language model, achieving 92% overall prediction accuracy for 43 frequently observed amino acids, significantly improving the prediction reliability. The efficacy of our prediction tool was validated through targeted genome mining, which led to the discovery of novel antimicrobial peptides. Our work lays a foundation to understand the sequence-to-function relationship of the bacterial adenylation domain and will facilitate the exploitation of nonribosomal peptides. NRPStransformer is available at http://www.nrpstransformer.cn.