A generalized supervised contrastive learning framework for integrative multi-omics prediction models

Abstract

Abstract Recent technological advances have highlighted the significant impact of the human microbiome and metabolites on physiological conditions. Integrating microbiome and metabolite data has shown promise in predictive capabilities. We developed a new supervised contrastive learning framework, MB-SupCon-cont, that (1) proposes a general contrastive learning framework for continuous outcomes and (2) improves prediction accuracy over models using single omics data. Simulation studies confirmed the improved performance of MB-SupCon-cont, and applied scenarios in type 2 diabetes and high-fat diet studies also showed improved prediction performance. Overall, MB-SupCon-cont is a versatile research tool for multi-omics prediction models.