Variational Autoencoders for Metabolomics: Data Imputation, Deconfounding, and Correlation Discovery

Untargeted metabolomics faces significant statistical challenges, including compositionality, missing values, and confounding effects that can lead to spurious correlations and impede accurate metabolite relationship analysis. Here we present MetVAE, a variational autoencoder-based model that addresses these challenges. In simulations, MetVAE demonstrates superior performance over competing methods in controlling false discovery rates while maintaining high true positive rates. Applied to inflammatory bowel disease data, MetVAE identifies three metabolite clusters with disease-specific correlation patterns. Analysis of hepatocellular carcinoma data shows MetVAE can complement structural similarity measures by uncovering functional relationships between metabolites. Using a large-scale dataset of over 20,000 samples, we also demonstrate MetVAE’s scalability and effectiveness in reducing technical variation while preserving biological signals. Our results establish MetVAE as a powerful and versatile tool for advancing metabolomics research, enabling deeper exploration of disease mechanisms and metabolic interactions.