Abstract

This study demonstrated a spatial omics–integrated Graph Neural Network (GNN) framework for high-resolution metabolic flux estimation across heterogeneous microbial communities by jointly modeling spatially resolved transcriptomic, proteomic, and metabolomic signals with microenvironmental interaction graphs. The aim of the study is to develop a computational method capable of predicting cell-specific metabolic fluxes while capturing metabolic heterogeneity, microbial interactions, and spatially constrained biochemical dynamics. The methodology constructs multi-layer spatial graphs where nodes represent microbial taxa or single cells, edges encode spatial proximity and biochemical interaction strength, and node features embed multi-omics signals normalized through spatial-aware preprocessing. A physics-guided flux inference module constrains the GNN with stoichiometric and thermodynamic feasibility rules derived from microbial metabolic models. End-to-end training, optimized using spatiotemporal mini-batching, enables the model to learn nonlinear mappings between local omics states and emergent flux distributions under nutrient gradients, stress conditions, and community-level cross-feeding. Performance evaluation based on accuracy, F1, MAE of flux magnitudes, training dynamics, latency profiling, and model stability demonstrates that spatial-omics–guided GNNs outperform non-spatial baselines by 18–32% while enabling sub-second flux inference suitable for real-time microbial ecosystem monitoring.