Fig. 2: VAE and PCA model construction on the TwinsUK dataset.
From: Variational autoencoders learn transferrable representations of metabolomics data
a Training and b test set metabolite correlation matrix reconstruction for a range of latent dimensionality values d. The slope of the VAE curve plateaued after d = 18. Error bars correspond to one standard deviation from bootstrapping (n = 1000 iterations). c Final VAE architecture, where μ is the mean vector and σ is the standard deviation vector that generates the latent space z. d Reconstruction MSE for latent dimensionality d = 18 on training and test sets. The box represents the interquartile range (IQR), whiskers are up to 1.5x IQR, and plotted points are outliers. The VAE preserved feature correlations substantially better than PCA. Kernel PCA-based results can be found in Supplementary Figs 3 and 4.
