Deep Learning to Predict the Biosynthetic Gene Clusters in Bacterial Genomes

Area of Science:

• Genomics
• Bioinformatics
• Metabolomics

Background:

• Bacterial genomes contain biosynthetic gene clusters (BGCs) responsible for producing small molecules and secondary metabolites.
• Accurate identification of BGCs is crucial for understanding bacterial metabolism and discovering novel compounds.

Purpose of the Study:

• To develop and validate an enhanced deep learning method, e-DeepBGC, for improved detection of BGCs and their biosynthetic classes in bacterial genomes.
• To compare the performance of e-DeepBGC against the existing DeepBGC method.

Main Methods:

• Leveraged validated BGCs, protein family domains (Pfams), and associated functional information.
• Developed e-DeepBGC, an extension of the DeepBGC deep learning model.
• Applied e-DeepBGC to a large dataset of 5,666 RefSeq bacterial genomes.

Main Results:

• e-DeepBGC demonstrated reduced false positive rates in BGC identification.
• e-DeepBGC achieved increased sensitivity in detecting BGCs compared to DeepBGC.
• Identified 170,685 BGCs across 5,666 bacterial genomes, averaging 30.1 BGCs per genome.

Conclusions:

• e-DeepBGC offers a more accurate and sensitive approach for BGC detection in bacterial genomes.
• The findings provide a comprehensive summary of BGCs and their distribution across bacterial phyla.
• This advancement facilitates the exploration of bacterial secondary metabolomes and potential novel compound discovery.