Skip to content

MetaboXcan

Summary

MetaboXcan extends the MetaXcan/S-PrediXcan approach from gene expression to the metabolome. It trains genetic predictors of over 500 plasma metabolites from matched genotype and metabolomics data in the Metabolic Syndrome in Men Study (METSIM), then uses those predictors (alongside GTEx expression predictors) to run four complementary association analyses against a GWAS trait — gene-to-trait (TWAS), gene-to-metabolite (M-TWAS), metabolite-to-trait (MWAS), and expression-aggregated metabolite-to-trait (g-MWAS) — organizing results into metabolic pathways and gene-metabolite networks for interpretation. Applied to chronic kidney disease (CKD), it recovered established CKD genes and metabolites and proposed a glycine-availability-centered mechanistic axis [1].

Motivation

Metabolite GWAS (mGWAS) and gene expression TWAS each provide partial mechanistic insight into GWAS loci, but directly measuring both metabolomic and disease-phenotype data in the same large cohort is costly and rare. Genetically predicting metabolite levels — analogous to how TWAS genetically predicts expression — lets researchers test metabolite-trait associations using only GWAS summary statistics, at population scale, and even in contexts where direct measurement is impractical (e.g., specific tissues, developmental stages), while also reducing the environmental noise (diet, medication) that affects directly-measured metabolomics.

Metabolite Prediction Models

Trained on 6,136 METSIM participants (1,391 quality-controlled plasma metabolites from the Metabolon DiscoveryHD4 platform, missing values imputed via softImpute), with 400 individuals held out for testing:

  • Metabolite heritability (GCTA) ranged 0–91% (median 17%), with 91% (1,267/1,391) significantly heritable.
  • Genome-wide lasso regression (sparse) outperformed genome-wide ridge regression (fully polygenic): lasso produced usable predictors (R > 0.10, p < 0.05) for 42% (580/1,391) of metabolites with median held-out R = 0.221, versus 22% (301/1,391) for ridge with median R = 0.142 — suggesting metabolite genetic architecture is comparatively sparse.
  • Lasso predictors outperformed the existing OmicsPred Bayesian-ridge metabolite predictors on held-out METSIM data despite METSIM's smaller training sample (6,136 vs. 8,153 individuals).
  • Cross-ancestry/cross-cohort generalization was confirmed in the Insulin Resistance Atherosclerosis Study–Classic (IRASC, N=184, individuals of Mexican descent): held-out performance (R = 0.42) matched the METSIM test set, with lasso again outperforming ridge and OmicsPred.

Four Association Analyses

  1. TWAS (gene-to-trait): run via multi-tissue S-PrediXcan/MultiXcan.
  2. M-TWAS (gene-to-metabolite): same underlying software as MultiXcan, applied to metabolites; phenotype-agnostic and distributed pre-computed with the software.
  3. MWAS (metabolite-to-trait): direct predicted-metabolite-to-trait association, using the same software as S-PrediXcan.
  4. g-MWAS (expression-based metabolite-to-trait): aggregates TWAS gene p-values for genes associated with a given metabolite, connecting expression evidence to metabolite-trait signal.

Results are further organized into Metabolon pathway groupings and a gene-metabolite interaction network to support biological interpretation.

Application: Chronic Kidney Disease

Applied to CKDGen GWAS summary statistics:

  • TWAS: 4 Bonferroni-significant genes (p < 2.3×10⁻⁶), including PDILT (p=3.8×10⁻⁹, likely reflecting the flanking kidney-function gene UMOD) and SPATA5L1 (p=2.3×10⁻⁷, likely reflecting GATM, which encodes the rate-limiting enzyme in creatine/homoarginine synthesis and has prior CKD associations).
  • MWAS: 5 Bonferroni-significant metabolites — N-acetylglycine, glycine, γ-glutamylglycine, homoarginine, propionylglycine — all previously implicated in CKD or related phenotypes.
  • g-MWAS: 7 Bonferroni-significant metabolites including homoarginine, whose driving gene cluster included the GATM locus, consistent with GATM's established biosynthetic role.
  • The combined pattern nominated a glycine-availability-centered biochemical axis spanning oxidative stress, cellular energetics, and vascular signaling as a candidate CKD mechanism for downstream investigation.

Availability

Source code: github.com/hakyimlab/metaboxcan. Requires user-supplied GWAS summary statistics, LDSC heritability estimate, and sample size; uses precomputed metabolite/expression weights from predictdb.org and Metabolon pathway annotations.

See Also

  • MetaXcan / sMultiXcan — the underlying TWAS methodology MetaboXcan extends to metabolites.
  • OmicsPred — an existing metabolite genetic-predictor resource that MetaboXcan's lasso models outperformed in held-out validation.

Citations

[1] Nyasimi, F., Sumner, S., Liang, Y., Yin, X., Park, Y., Brown, A., et al., & Im, H.K. (2026). MetaboXcan: A multiomic framework linking genetically predicted metabolites, gene expression, and complex traits. bioRxiv.