Deep learning predicts microbial interactions from self-organized spatiotemporal patterns.

Area of Science:

• Microbiology
• Computational Biology
• Machine Learning

Background:

• Microbial communities exhibit spatial patterns driven by interspecies interactions, crucial for understanding community dynamics.
• Predicting microbial interactions from spatial patterns remains a challenge, limiting our understanding of community functional dynamics.

Purpose of the Study:

• To introduce supervised deep learning as a novel tool for microbial network inference.
• To develop a method for predicting microbial interactions directly from spatiotemporal data.

Main Methods:

• An agent-based model simulated microbial community evolution under various growth and interaction scenarios.
• Deep neural networks were trained on simulated data to infer interaction coefficients.
• The model was validated using co-culture data of *Pseudomonas fluorescens* and *Escherichia coli* with different substrates.

Main Results:

• Accurate predictions of interaction coefficients were achieved in small-scale simulations (R² = 0.84).
• Deep learning models successfully predicted spatially varying interaction coefficients in larger domains without retraining.
• Context-dependent interactions (degrader-cheater and competition) were correctly identified in real biological data.

Conclusions:

• The combination of agent-based modeling and deep learning effectively infers microbial interactions from spatial data.
• This approach offers a powerful new tool for analyzing complex microbial community interactions.
• The findings highlight the potential of spatial patterns for understanding ecological relationships in microbial consortia.