Co-informed prediction of microbiome-metabolome interactions through novel transfer learning models - ABSTRACT The gut microbiome plays a critical role in the pathogenesis of human diseases through their metabolites, acting as a dynamic interface between our environment and health. Recent paired microbiome-metabolome studies enable the simultaneous analysis of hundreds of microbes and metabolites, uncovering their associa- tions with various conditions in unprecedented detail. Integrative analyses of paired microbiome-metabolome data sets require multivariate models that can account for complex correlation structures between microbes and metabolites. However, these multivariate models face the challenge of low statistical power for detecting microbiome-metabolome interactions due to small sample sizes and weak biological signals. Therefore, the overarching goal of the proposed research is to develop novel statistical methods for the powerful detection of weak microbiome-metabolome interactions with limited data. In Aim 1, we will develop a powerful transfer- learning framework for high-dimensional regression models with composition covariates. Unlike most existing transfer-learning methods that borrow information from external data sets, our framework leverages the simi- larities across metabolites within a single cohort. This will help minimize the potential risk of negative transfer, given the different sequencing platforms and bioinformatic pipelines used in different studies. In Aim 2, we apply our transfer-learning method to the human microbiome project (HMP), the integrative human microbiome project (iHMP), and an external paired microbiome-metabolome data set. Our analyses will reveal biologically relevant microbiome-metabolome interactions for both healthy people and patients diagonosed with inflammatory bowel disease (IBD), advancing our understanding of the microbiome-IBD link. Furthermore, we will analyze biopsy- specific microbiome-metabolome interactions for IBD patients, revealing potential spatial heterogeneity for IBD- associated microbiome and metabolites. This proposed research is distinguished by its originality in introducing novel, powerful, and flexible transfer-learning frameworks and novel applications of microbiome-metabolome interactions in both healthy and IBD populations.
