Interpreting graph neural networks reveals microbial association patterns in inflammatory bowel diseases

In longitudinal gut metagenomics from 106 HMP2/iHMP individuals, graph neural networks for inflammatory bowel disease predictions rely on a densely connected module of co-abundant Bacteroidales species.

• Why it matters: Abundance-only models can separate IBD from non-IBD but interpret taxa independently, leaving unclear how between-taxa association structure contributes to predictions.
• What they did: The study inferred individual-specific microbial association networks from longitudinal metagenomics and trained GNNs, comparing them with Random Forest (AUPRC ~0.59 vs 0.446).
• The result: In an independent cohort, the Bacteroidales association module carried most predictive signal, and removing its associations reduced performance to near chance, implying coordinated association structure is informative.

AI summary of the abstract · may contain errors · not medical advice · read the paper.