Reconstructing the intestinal microbiota ecological network based on graph neural network and contrastive learning: Predicting key microbiota regulatory targets in Crohn's Disease.

Crohn's disease (CD) is a heterogeneous inflammatory bowel disease in which microbial, metabolic, and host factors interact dynamically. Therefore, we aimed to (i) integrate microbiota, metabolite, and host-gene data into a heterogeneous dynamic graph and (ii) use contrastive learning to prioritize microbiota-related regulatory targets in CD under limited-sample conditions. We combined a heterogeneous graph dynamic network (HGDN) with a contrastive learning module (CENet) to model microbiota-metabolite-gene interactions and to improve the robustness of target prioritization. In the current dataset, HGDN + CENet achieved higher target-prediction performance than the selected baseline models and identified coordinated microbial, metabolic, epigenetic, and inflammatory changes associated with disease activity. Animal experiments, in vitro assays, and preliminary clinical observations provided supportive evidence for selected model-prioritized targets, particularly butyrate-associated pathways. These findings should be interpreted as associative and hypothesis-generating rather than causal proof. Overall, this study suggests that graph-based multi-omics integration may help prioritize microbiota-related regulatory targets for further mechanistic and interventional validation in CD.