iLoRA: Bayesian Low-Rank Adaptation with Latent Interaction Graphs for Microbiome Diagnosis

Reliable microbiome-based diagnosis is critical for precision medicine at scale in inflammatory diseases, yet current post-training pipelines in LLMs often overlook the interaction structure that governs microbial ecosystems. In inflammatory bowel disease (IBD), disease signals arise not only from species-level abundance shifts but also from latent microbe–microbe cross-talk. We propose iLoRA, a parameter-efficient Bayesian LoRA framework that infers latent interaction graphs from microbiome inputs and integrates them into adaptation, enabling joint clinical prediction and interaction discovery. Unlike correlation-based post hoc analysis, iLoRA models microbial interactions as latent variables learned end-to-end, yielding uncertainty-aware estimates of cross-talk. We evaluate iLoRA on (i) interactive question answering with human-annotated interaction graphs to quantify structural recovery and (ii) gut microbiome cohorts for IBD diagnosis. Across both domains, iLoRA consistently improves accuracy over strong LoRA baselines while producing interpretable interaction graphs consistent with annotated relations and conventional microbiome association networks.