DeepBGC

DeepBGC predicts biosynthetic gene clusters (BGCs) in bacterial and fungal genomes to identify and classify loci encoding microbial secondary metabolites, including small-molecule natural products relevant to antimicrobial, anticancer, and immunomodulatory activities.

Key Features:

• Deep Learning Framework: Uses a Bidirectional Long Short-Term Memory (BiLSTM) recurrent neural network to capture sequential patterns in genomic data for BGC detection.
• Vector Embedding of Protein Domains: Employs a word2vec-like vector embedding approach for Pfam protein domains to represent functional components of BGCs.
• Reduced False Positives: Reports reduced false positive rates compared to existing machine-learning tools in BGC identification.
• Novel Class Identification: Can extrapolate to identify novel BGC classes that were previously undetectable by other algorithms.
• Random Forest Classification: Incorporates Random Forest classifiers to predict BGC product class and potential chemical activity.

Scientific Applications:

• BGC discovery: Identification of putative BGCs in bacterial and fungal genomes, including previously undetectable loci.
• Natural product and drug discovery: Prioritization of BGCs that may encode natural products with antimicrobial, anticancer, or immunomodulatory activities.
• Product inference: Prediction of product class and potential chemical activity to guide downstream biochemical and pharmacological investigation.

Methodology:

Applies Pfam protein-domain word2vec-like vector embeddings as input to a BiLSTM RNN for BGC detection, followed by Random Forest classifiers to predict product class and potential chemical activity.