mtVAE
Summary¶
Gomari, Schweickart, Cerchietti, et al. (2022) trained a variational autoencoder (VAE) on 217 metabolite measurements from 4,644 TwinsUK blood samples to learn a nonlinear, low-dimensional latent representation of the metabolome. Latent dimensions were interpretable in terms of metabolic pathways (via Shapley Additive Global Importance, SAGE, feature attribution) and — critically — generalized to three unseen clinical datasets (type 2 diabetes, acute myeloid leukemia treatment response, schizophrenia), where individual VAE dimensions associated with disease status more strongly than any dimension from linear PCA or nonlinear kernel PCA (KPCA) trained the same way [1].
Model and Baselines¶
- Architecture: 217-dimensional input/output layer (one per metabolite), 200-dimensional intermediate layer, 18-dimensional latent layer; hyperparameters (intermediate dimensionality, learning rate=0.001, KL-divergence weight=0.01) selected via Keras Tuner on an 85/15 TwinsUK train/test split. Latent dimensionality (d=18) was chosen where reconstruction MSE of the metabolite correlation matrix (CM-MSE) plateaued.
- Baselines: linear PCA and kernel PCA (cosine, sigmoid, RBF, polynomial kernels), fit on the same training data and truncated to the same d=18 dimensions.
- Reconstruction: on training data, KPCA reconstructed the raw data matrix best; on held-out test data, PCA and VAE both outperformed KPCA on sample-wise reconstruction, with PCA edging out VAE. However, for correlation-matrix reconstruction (CM-MSE) — arguably the more biologically relevant metric, since it reflects preserved metabolite-metabolite relationships rather than raw value fidelity — both nonlinear methods (VAE and KPCA) outperformed linear PCA on both training and test sets.
Pathway Interpretability¶
SAGE feature-importance scores were computed at metabolite, sub-pathway (e.g., "TCA Cycle," "Sphingolipid Metabolism"), and super-pathway (e.g., "Lipid," "Amino acid") granularity. Almost every metabolic sub-pathway mapped predominantly onto a single VAE latent dimension — e.g., dimension 12 captured both glycolysis/gluconeogenesis/pyruvate metabolism and alanine/aspartate metabolism (biologically linked via the glucose-alanine cycle), and dimension 14 captured ketone bodies, fatty acid metabolism, and the TCA cycle jointly. PCA and KPCA dimensions, by contrast, non-specifically mixed unrelated sub-pathways, particularly in their top (highest-variance) components.
Generalization to Unseen Disease Cohorts¶
VAE, PCA, and KPCA encoders trained only on TwinsUK were applied (frozen) to three independent metabolomics datasets, and each latent dimension was tested by t-test against a clinical outcome:
- Type 2 diabetes (QMDiab, n=358): VAE dimension 12 showed the strongest association (p=1.6×10⁻²²), stronger than the best PCA dimension (p=2.1×10⁻²⁰). Top-ranked metabolites (glucose, serine, mannose, glycine) and sub-pathways matched established T2D biology (impaired gluconeogenesis consuming glucogenic amino acids). Dimension 12 also correlated strongly with HbA1c (p=6.2×10⁻⁴⁵ vs. 1.1×10⁻³⁰ for the best PCA dimension and 6.0×10⁻³⁸ for the best KPCA dimension).
- Acute myeloid leukemia treatment response (n=85): VAE dimension 12 (p=0.027) outperformed the best KPCA dimension (RBF, p=0.12) by an order of magnitude, though the result would not survive multiple-testing correction and is flagged as suggestive. Driving sub-pathways (lysolipid, polyunsaturated fatty acids) are mechanistically linked to daunorubicin's membrane-targeting action.
- Schizophrenia (n=201): VAE dimension 11 (p=3.0×10⁻⁶) outperformed the best KPCA dimension (RBF, p=3.6×10⁻⁶). Top contributors were phenylalanine/tyrosine metabolism (dopamine precursor pathway) and creatine, both with independent prior literature support in schizophrenia.
In all three cohorts, the strongest VAE-dimension association was comparable in magnitude to the strongest single-metabolite univariate association, but — unlike a single metabolite — a latent dimension aggregates a system-level, multi-pathway signal.
Availability¶
Source code: github.com/krumsieklab/mtVAE (archived at Zenodo).
See Also¶
- McMLP — another metabolomics deep learning model that also relies on population-scale reference data to generalize to unseen cohorts, applied to dietary-intervention response rather than static disease-state classification.
- AUTOENCODIX — a general-purpose multi-architecture autoencoder framework spanning the same representation-learning space as this domain-specific metabolomics VAE.