Deep self-supervised learning for biosynthetic gene cluster detection and product classification.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Natural products, crucial for pharmaceuticals, are synthesized by microbial biosynthetic gene clusters (BGCs).
• Advances in sequencing have revealed numerous undiscovered BGCs within microbial genomes and metagenomes.

Purpose of the Study:

• To develop a novel self-supervised learning approach for identifying and characterizing BGCs.
• To leverage machine learning for BGC discovery and classification from genomic data.

Main Methods:

• Representing BGCs as sequences of functional protein domains.
• Training a masked language model on these domain sequences for self-supervised learning.
• Assessing model performance in detecting BGCs and predicting their properties in bacterial genomes.

Main Results:

• The self-supervised model successfully detected BGCs in bacterial genomes.
• The model learned meaningful representations of BGCs and their protein domains.
• The approach accurately predicted BGC product classes.

Conclusions:

• Self-supervised neural networks offer a powerful framework for enhancing BGC prediction and classification.
• This method aids in the discovery of novel natural products from genomic data.
• The approach facilitates the characterization of BGCs and their functions.