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

Researchers introduce iLoRA, a Bayesian framework that combines low-rank adaptation with latent interaction graph inference for improved domain-specific predictions. The method is evaluated on microbiome diagnosis tasks, where it outperforms standard LoRA by jointly learning prediction models and underlying biological interaction structures rather than analyzing them separately.

iLoRA represents an advancement in parameter-efficient fine-tuning by addressing a fundamental limitation of standard LoRA: its inability to model and expose the latent interactions that often drive scientific predictions. Traditional LoRA applies static, input-agnostic low-rank updates, treating feature interactions as a black box. This new Bayesian approach infers dynamic interaction graphs from input data, allowing the model to condition its weight updates on discovered relationships between variables.

The methodology emerges from growing recognition that scientific machine learning requires interpretability alongside accuracy. In microbiome research specifically, disease outcomes depend not just on species abundance but on complex microbial cross-talk—interactions that previous approaches could only analyze post-hoc through separate feature attribution methods. iLoRA integrates structure discovery into the adaptation process itself, enabling simultaneous optimization of predictive performance and biological plausibility.

The dual evaluation framework is particularly notable. Testing on interactive QA with human-annotated graphs directly measures whether the learned latent structures align with expert knowledge, while multi-cohort IBD diagnosis validates biomedical utility across different patient populations. Strong performance across both settings suggests the approach captures meaningful biological relationships rather than spurious correlations.

For the broader machine learning community, iLoRA demonstrates how Bayesian principles can enhance parameter-efficient adaptation without prohibitive computational overhead. The calibrated uncertainty estimates and graph-based interpretability could accelerate adoption of fine-tuned models in regulated domains like healthcare, where explainability requirements currently limit deployment. Future work likely extends this framework to other domains where latent structures drive outcomes.

• →iLoRA infers latent interaction graphs during fine-tuning, enabling input-conditioned LoRA updates rather than static weight adjustments.
• →The framework jointly optimizes predictive accuracy and latent structure recovery, eliminating the need for separate post-hoc interaction analysis.
• →Evaluated on microbiome diagnosis, iLoRA outperforms strong LoRA and Bayesian baselines while recovering biologically plausible interaction patterns.
• →The method provides calibrated uncertainty estimates with moderate computational overhead suitable for scientific applications.
• →Graph-conditioned adaptation represents a significant shift toward interpretable parameter-efficient fine-tuning for domain-specific prediction tasks.