Predicting Biological Activity from Biosynthetic Gene Clusters Using Neural Networks.

Area of Science:

• Microbiology
• Bioinformatics
• Drug Discovery

Background:

• Microorganisms are a source of natural products for drug development.
• Identifying novel bioactive molecules is challenging due to complex processes.
• Genome mining and machine learning offer computational approaches to natural product discovery.

Purpose of the Study:

• To develop an improved computational tool for detecting multiple bioactivities from microbial genomes.
• To address limitations of existing tools, including small datasets and outdated methods.
• To facilitate the identification of novel natural products for potential therapeutic applications.

Main Methods:

• Compiled an expanded training dataset from the MIBiG database.
• Utilized the antiSMASH tool for biosynthetic gene cluster (BGC) identification.
• Introduced novel sequence-based genomic descriptors.
• Employed neural networks with class weighting to handle imbalanced data.
• Validated the tool against literature-mined test sets and compared it with existing methods.

Main Results:

• The developed tool, NPBdetect, demonstrates high confidence in detecting multiple bioactivities.
• NPBdetect outperforms existing tools in identifying potential natural products.
• The new sequence-based descriptors improve the accuracy of bioactivity prediction.
• The class weighting technique effectively addresses class imbalance issues in the model.

Conclusions:

• NPBdetect offers a more effective and reliable method for discovering natural products with diverse bioactivities.
• This tool can accelerate the identification of novel drug candidates from microbial genomic data.
• The approach advances the field of computational natural product discovery and genome mining.