Predicting fungal secondary metabolite activity from biosynthetic gene cluster data using machine learning

Area of Science:

• Mycology
• Bioinformatics
• Machine Learning

Background:

• Fungal secondary metabolites (SMs) are crucial for ecological roles and possess valuable medicinal/industrial properties.
• Biosynthetic gene clusters (BGCs) encode SM production, but predicting SM bioactivity from BGCs is challenging.
• Fungi are a rich source of natural products and drugs, yet most biosynthetic pathways remain uncharacterized.

Purpose of the Study:

• To adapt and evaluate machine learning models for predicting fungal SM bioactivity from BGCs.
• To assess the impact of training data size and composition (fungal vs. fungal and bacterial BGCs) on prediction accuracy.
• To highlight the need for systematic efforts to link fungal BGCs to their corresponding SMs and bioactivities.

Main Methods:

• Adapted existing machine learning models trained on bacterial BGC data for fungal BGC analysis.
• Trained models on two datasets: fungal BGCs only (314 BGCs) and combined fungal (314 BGCs) and bacterial BGCs (1,003 BGCs).
• Evaluated model performance using balanced accuracies for predicting antibacterial, antifungal, and cytotoxic/antitumor activities.

Main Results:

• Models trained solely on fungal BGCs achieved balanced accuracies between 51% and 68%.
• Incorporating bacterial BGC data into training yielded marginal improvements, with accuracies ranging from 56% to 68%.
• Low prediction accuracies were attributed to the limited size of available fungal BGC and bioactivity data.

Conclusions:

• Current machine learning approaches have limited applicability to fungal SM studies due to data scarcity.
• Systematic efforts are urgently required to identify fungal BGCs, their products, and bioactivities.
• Expanding knowledge of fungal SMs is critical for discovering novel drugs and understanding fungal biology.