Physics-constrained neural ordinary differential equation models to discover and predict microbial community dynamics

Area of Science:

• Microbiology
• Computational Biology
• Systems Biology

Background:

• Microbial communities are vital for ecosystem functions.
• Mechanistic and machine learning models have limitations in predicting microbiome dynamics.
• Metabolite competition and cross-feeding are key drivers of microbial interactions.

Purpose of the Study:

• To develop a flexible and accurate model for microbial community dynamics.
• To overcome limitations of traditional mechanistic and machine learning models.
• To provide insights into direct biological interactions within microbiomes.

Main Methods:

• Developed a physics-constrained machine learning model named Neural Species Mediator (NSM).
• Integrated a mechanistic model of metabolite dynamics with a machine learning component.
• Validated the NSM model on experimental datasets.

Main Results:

• The NSM model demonstrated higher accuracy than standalone mechanistic or machine learning models.
• The NSM model provided improved interpretability of biological interactions.
• The model successfully captured metabolite-mediated interactions in microbial communities.

Conclusions:

• Embedding neural networks into mechanistic models enhances prediction performance.
• The NSM offers a powerful tool for understanding and controlling microbial communities.
• This hybrid approach improves upon existing predictive modeling frameworks for microbiomes.