Development of multiple microbiome biomarkers using penalized regression methods

Identifying biomarkers that reflect the complex relationship between the microbiome and health outcomes in microbiome studies is essential for advancing the understanding and improving disease management. While past research was focused on a single biomarker modeling approach, this study extends that work by combining multiple taxa to identify a subset of multiple biomarkers relevant to clinical outcomes.

We extend the information theory framework for surrogate endpoint evaluation by applying LASSO and Elastic Net models to identify combinations of taxa as biomarkers for clinical outcomes. Feature selection for the biomarker's construction is done in order to maximize the goodness of fit of the predictive biomarker model. Monte Carlo cross validation is used to enhance the reliability of feature selection. The high salt diet study on mice is used to illustrate the methodology for continuous outcome (tumor size). The top 5 selected genera yielded a correlation of 0.9274 between predicted and observed tumor size, with a 67.92% reduction in uncertainty when the multiple microbiome biomarkers score is known. To illustrate the methodology for binary outcome, the CERTIFI study on Crohn's disease patients treated with ustekinumab is used. A multiple microbiome biomarkers score, constructed using the top 5 selected families, significantly improved prediction of remission 6 weeks after induction treatment (the clinical outcome of interest).

This study presents a unified approach for identifying multiple microbiome biomarkers using penalized regression for clinical outcome prediction. The proposed methods are applied to both continuous and binary outcomes. The method enhances the detection of meaningful biomarkers with potential for personalized treatment and disease management.