BGC-MAC and BGC-MAP: Attention-Based Models for Biosynthetic Gene Cluster Classification and Product Matching.

Area of Science:

• Biochemistry
• Bioinformatics
• Machine Learning

Background:

• Natural products synthesized by biosynthetic gene clusters (BGCs) are crucial sources of therapeutics.
• Accurate annotation of BGCs is vital for exploring natural product diversity but remains challenging due to incomplete understanding of biosynthetic pathways.
• Current annotation methods face limitations in accuracy and interpretability.

Purpose of the Study:

• To develop advanced deep learning models for improved BGC annotation.
• To enhance the classification of BGCs by natural product class and associate BGCs with specific product structures.
• To provide an explainable AI framework for deeper insights into BGC-natural product relationships.

Main Methods:

• Developed two deep learning models: BGC-multihead attention classifier (BGC-MAC) and BGC-multihead attention product-matcher (BGC-MAP).
• Trained models on experimentally validated BGC-natural product pairs.
• Utilized cross-attention mechanisms for explainable AI, identifying key protein domains and BGC-substructure relationships.

Main Results:

• BGC-MAC outperformed existing tools (antiSMASH, DeepBGC) in classifying BGCs by natural product class.
• BGC-MAP demonstrated potential in prioritizing candidate BGCs or identifying products from BGCs.
• The models' explainable AI features revealed biosynthetic logic without prior annotations.

Conclusions:

• BGC-MAC and BGC-MAP offer a data-driven, explainable AI framework for BGC annotation.
• These models deepen biosynthetic insights and accelerate the discovery of novel natural products.
• The developed software is publicly available to facilitate research.