A framework for predictive modeling of microbiome multi-omics data: latent interacting variable-effects (LIVE)...

The number and size of multi-omics datasets with paired measurements of the host and microbiome is rapidly increasing with the advance of sequencing technologies. As it becomes routine to generate these datasets, computational methods to aid in their interpretation become increasingly important. Here, we present a framework for integration of microbiome multi-omics data: Latent Interacting Variable Effects (LIVE) modeling. LIVE integrates multi-omics data using single-omic latent variables (LV) organized in a structured meta-model to determine the combinations of features most predictive of a phenotype or condition.

We developed a supervised version of LIVE leveraging sparse Partial Least Squares Discriminant Analysis (sPLS-DA) LVs, and an unsupervised version leveraging sparse Principal Component Analysis (sPCA) principal components which both can incorporate covariate awarness. LIVE performance was tested on publicly available metagenomic and metabolomics data set from Crohn's Disease (CD) and Ulcerative Colitis (UC) status patients in the PRISM and LLDeep cohorts, and benchmarked against existing gut microbiome multi-omics approaches and vaginal microbiome datasests, achieving consistent and comparable performances. In addition to these benchmarking efforts, we present a detailed analysis and interpretation of both versions of LIVE using the PRISM and LLDeep cohorts. LIVE reduced the number of feature interactions from the original datasets for CD and UC from millions to less than 20,000 while conditioning the disease-predictive power of gut microbes, metabolites, enzymes, on clinical variables.

LIVE makes a distinct, complementary contribution to current methods to integrate microbiome data and offers key advantages to existing approaches in the interpretable integration of multi-omics data with clinical variables to predict to disease outcomes and identify microbiome mechanisms of disease.