SIMBA-GNN: mechanistic graph learning for microbiome prediction.

Area of Science:

• Microbiology
• Computational Biology
• Systems Biology

Background:

• Predicting microbial community assembly requires models beyond taxonomic correlations.
• Understanding interspecies interactions is crucial for accurate microbiome predictions.

Purpose of the Study:

• To develop a novel computational model, SIMBA (simulation-augmented graph neural network), for predicting microbial community composition.
• To integrate mechanistic insights from metabolic simulations with graph neural networks for enhanced predictive power.

Main Methods:

• Developed SIMBA, a graph neural network combining metabolic simulations and edge-aware graph transformers.
• Inferred cross-feeding probabilities and functional similarity from pairwise metabolic simulations.
• Employed a multi-stage training pipeline including self-supervised learning and supervised pretraining.

Main Results:

• SIMBA successfully predicts microbial presence and relative abundance across individuals.
• The model leverages a mechanistic prior derived from metabolic simulations.
• Enabled hypothesis-driven exploration of microbial ecosystems.

Conclusions:

• SIMBA provides a powerful framework for predicting microbial community dynamics.
• Integrating mechanistic and machine learning approaches advances microbiome research.
• The model facilitates deeper insights into gut microbial ecosystem assembly and function.