Protein language model-based end-to-end type II polyketide prediction without sequence alignment

Abstract

Natural products are important sources for drug development, and the precise prediction of their structures assembled by modular proteins is an area of great interest. In this study, we introduce DeepT2, an end-to-end, cost-effective, and accurate machine learning platform to accelerate the identification of type II polyketides (T2PKs), which represent a significant portion of the natural product world. Our algorithm is based on advanced natural language processing models and utilizes the core biosynthetic enzyme, chain length factor (CLF or KSbeta), as computing inputs. The process involves sequence embedding, data labeling, classifier development, and novelty detection, which enable precise classification and prediction directly from KSbeta without sequence alignments. Combined with metagenomics and metabolomics, we evaluated the ability of DeepT2 and found this model could easily detect and classify KSbeta either as a single sequence or a mixture of bacterial genomes, and subsequently identify the corresponding T2PKs in a labeled categorized class or as novel. Our work highlights deep learning as a promising framework for genome mining and therefore provides a meaningful platform for discovering medically important natural products.