Fig. 1: Overview of our approach.
From: Variational autoencoders learn transferrable representations of metabolomics data
a VAE, linear PCA, and non-linear kernel (K)PCA models were trained using training and test sets in the TwinsUK dataset (n = 4644 samples, p = 217 metabolites). Model performance was then evaluated using Mean Squared Error (MSE) of metabolite correlation matrix reconstruction. b The SAGE method was applied to calculate the contribution of individual metabolites, sub-pathways and super-pathways to each latent dimension. c QMDiab (n = 358), AML (n = 85), and Schizophrenia (n = 201) datasets were encoded using VAE and PCA models trained on the TwinsUK data. Latent dimensions of each model were then associated with disease phenotypes.
