SIMBA-GNN: mechanistic graph learning for microbiome prediction

Predicting how gut microbial communities assemble and change requires models that capture the underlying mechanisms driving interspecies interactions, not just taxonomic correlations. We present SIMBA, a simulation-augmented graph neural network that integrates mechanistic insights from metabolic simulations with edge-aware graph transformers to predict microbial community composition. Using a high-fiber dietary cohort mapped to metabolic networks, we ran thousands of pairwise simulations to infer cross-feeding probabilities, pathway activity fingerprints, and microbe-microbe functional similarity. These signals instantiate a global microbe-metabolite-pathway graph for learning. A custom heterogeneous graph transformer incorporates scalar edge attributes into attention. It is trained through a multi-stage pipeline combining self-supervised learning, supervised pretraining on simulated graphs, and fine-tuning on experimental microbial abundance data. Each individual's microbiome is represented as a sample-specific instantiation of the shared mechanistic graph derived from metabolic simulations, where only the set of microbes detected in that individual varies. SIMBA learns from this mechanistic prior to predict microbial presence and relative abundance across individuals, enabling hypothesis-driven exploration of microbial ecosystems.