Skip to content

Metabolic Flux Modulation of Genetic Risk

Summary

Metabolic flux modulation of genetic risk describes a non-additive relationship in which the effect of a disease-associated allele changes with the activity of a biochemical reaction. In coronary artery disease (CAD), genetically predicted organ-specific reaction fluxes revealed both amplifiers and buffers of risk alleles, implying that metabolic context can alter penetrance and that the value of targeting a pathway may depend on genotype [1].

Core Principle

Genetic variants operate within molecular networks rather than in isolation. A risk allele may alter transcription, protein function or cell behavior, while the surrounding metabolic network determines whether that perturbation is propagated, compensated or redirected.

  • Amplification means that increasing reaction flux strengthens the allele's adverse association with disease; the SNP-by-flux interaction coefficient is positive.
  • Buffering means that increasing reaction flux attenuates the allele's adverse association; the interaction coefficient is negative.
  • No interaction means that the additive effects of genotype and flux adequately describe the data at the available resolution.

This framework differs from asking whether a metabolite or reaction is associated with disease on average. A flux can have little marginal association in the whole population while having a strong effect within carriers of a particular allele.[1]

Predicted Flux Is Not Metabolite Concentration

A reaction flux is a rate of substrate conversion or transport through a biochemical reaction. It is not equivalent to the concentration of the substrate, product or enzyme. Intracellular fluxes are difficult to measure directly at population scale, so this study estimated genetically personalized flux capacity by combining imputed organ-specific gene expression with a genome-scale stoichiometric model.[1]

This should also be distinguished from experimental Lipid Metabolic Flux Analysis, which uses isotope tracing and mass spectrometry to estimate realized fluxes in a controlled biological system. Genetically predicted fluxes capture inherited differences in modeled reaction activity and intentionally omit much environmental regulation.

Genetically Personalized Flux Pipeline

Whole-body metabolic reconstructions (Harvey/Harvetta)
        |
        v
Adipose, brain, heart, liver and skeletal-muscle subnetworks
ported to the HUMAN1 human metabolic reconstruction
        |
        +--> GTEx average organ transcript abundance
        |          |
        |          v
        |       GIM3E estimates an average feasible flux state
        |
Participant genotypes
        |
        +--> PredictDB elastic-net models
                   |
                   v
          Personalized organ transcript abundance
                   |
                   v
qMTA integrates expression deviations with the average flux state
        |
        v
Genetically personalized organ-specific reaction fluxes
        |
        v
log2 transformation and standardization

The five organ models contributed 6,185 gene-annotated reactions: 721 adipose, 1,116 brain, 1,314 heart, 1,956 liver and 1,078 skeletal muscle. Correlated reactions were pruned within each organ at absolute Pearson correlation above 0.5, leaving 1,670 representative fluxes for multiple-testing correction and interpretation.[1]

Origin: The FWAS Framework (Foguet et al. 2022)

The pipeline above was introduced by Foguet, Xu, Ritchie, et al. (2022), who coined the term fluxome-wide association study (FWAS) and validated genetically personalized organ-specific flux maps against directly measured blood metabolites before the CAD-interaction analysis above was built on top of it [2].

  • Scale: personalized flux maps for skeletal muscle, adipose, liver, brain, and heart (accounting for ~66% of adult body weight) were built for over 520,000 individuals — 37,220 from INTERVAL and 487,395 from UK Biobank — "surpassing by more than two orders of magnitude" the number of personalized genome-scale metabolic models built in prior work. Organ subnetworks were extracted from the Harvey/Harvetta multi-organ reconstructions (built on Recon3D) and lifted over to HUMAN1, the more recent human genome-scale metabolic reconstruction.
  • Reaction coverage: 14,220 reaction flux values were computed per individual; after pruning strongly correlated fluxes (Pearson ρ ≥ 0.9), 4,300 relatively independent fluxes remained for association testing.
  • Validation against blood metabolites: regressing each measured blood metabolic feature (Nightingale NMR and Metabolon HD4 panels) against the 4,300 personalized fluxes in INTERVAL identified 4,312 significant flux-metabolite associations (FDR-adjusted P<10⁻⁶) spanning 229 unique metabolic features and 763 unique fluxes — liver contributed the most associations (1,301), consistent with its role in whole-body metabolic homeostasis, followed by heart (1,005), skeletal muscle (896), brain (593), and adipose (517). 83% of these associations externally replicated in UK Biobank at the same significance threshold with concordant effect direction, and effect sizes correlated at ρ=0.82 between cohorts.
  • Reconstruction choice matters: repeating the analysis with organ-specific models built from Recon3D instead of HUMAN1 found only partial overlap (1,761 of 4,312 HUMAN1-based associations replicated with Recon3D-based fluxes; effect sizes still correlated at ρ=0.72) — attributed to HUMAN1's more complete gene-reaction annotation and refined reaction reversibility relative to the older Recon3D reconstruction.
  • Biological example — the TAG-CE pathway: fluxes through a liver-centered sequence of reactions hydrolyzing triacylglycerols (TAG) to diglycerides, synthesizing phospholipids, and using them to esterify free cholesterol (the "TAG-to-cholesterol-esterification" or TAG-CE pathway) were strongly associated with HDL cholesteryl-ester percentage, reduced LDL/HDL triglycerides, and reduced HDL particle size — while a disrupting reaction (retinyl ester hydrolysis) showed the opposite association, consistent with independent clinical reports that high-dose vitamin A derivatives raise triglycerides and disrupt cholesterol esterification.
  • First CAD application: a genome-wide multi-tissue FWAS for CAD in UK Biobank (Cox regression across the 4,300 fluxes) identified 92 significant flux-CAD associations (FDR<0.05) — liver contributed the most (32), followed by adipose (26) and heart; only 31 of these replicated with Recon3D-based fluxes, again illustrating the reconstruction's influence on result stability. This 92-association FWAS is the direct predecessor to the interaction (amplifier/buffer) analysis on the same UK Biobank CAD cohort described below, which asks not just whether a flux marginally associates with CAD risk but whether its effect on risk depends on genotype at a specific risk allele.

Discovery Study Design

The discovery analysis used 459,902 UK Biobank participants of inferred European genetic ancestry. CAD was defined from linked hospital and death records, yielding 37,941 cases and 398,282 controls after excluding ischemic-heart-disease phenotypes from controls.[1]

Starting from 18,348 genome-wide significant CAD variants in a published meta-analysis, the authors retained 5,852 variants that were also genome-wide significant in the UK Biobank subset under the study's Cox model. Each variant was tested against organ-specific reaction fluxes using two complementary procedures:[1]

  1. a Cox model with an explicit SNP-dosage-by-flux interaction term;
  2. a dosage-specific test comparing the flux effect on CAD across individuals carrying zero, one or two risk alleles.

A pair was called an amplifier or buffer only when both procedures passed Benjamini-Hochberg FDR below 0.05. The tests adjusted for age, sex, genotype array and ten genetic principal components; adding BMI, systolic blood pressure, LDL cholesterol, HDL cholesterol and triglycerides did not materially change the interaction estimates.[1]

See the Reaction-Flux Genetic Interaction reference for the equations and interpretation.

Guarding Against Genetically Induced Circularity

The flux predictions themselves are derived from genetically imputed expression. A CAD risk variant could therefore correlate with a flux simply because it is in LD with an eQTL used to predict that flux, producing a form of phantom interaction.[1]

Before fitting the interaction model, the authors regressed the candidate risk allele's effect out of the predicted flux. They also compared implicated CAD variants with the metabolic eQTLs used in flux computation. Some were in strong LD, as expected near metabolic genes, but none of those eQTLs had a strong correlation with the 18 implicated reaction fluxes after adjustment.[1]

This check reduces one specific source of circularity but does not experimentally prove the modeled reaction or its direction.

Primary Results

Result level Finding
Significant SNP-flux pairs in both tests 583
Buffering pairs 353
Amplifying pairs 230
Unique interacting SNPs 279
Independent CAD risk loci 8
Risk-locus-by-flux relationships 30
Unique reaction fluxes 18
Implicated fluxes with a significant univariate CAD association 5 of 18

The two statistical procedures were highly concordant: their interaction estimates correlated at 0.998. The low number of marginally associated fluxes shows that the analysis uncovered context-specific relationships rather than merely rediscovering strong standalone metabolic predictors.[1]

Biological Mechanisms and Loci

LPA and PLG Region

The chromosome 6 region containing LPA and PLG dominated the signal, accounting for 530 of 583 significant SNP-flux pairs across four independent LD-defined risk loci.[1]

Reaction context Interaction Proposed connection
Prostaglandin E2 transport in heart and brain Amplifies selected LPA-region risk alleles Lp(a) can promote inflammatory COX2/prostaglandin signaling; greater transport capacity may propagate the inflammatory effect
Arachidonoyl-CoA elongation in adipose Buffers two alleles in the same region Elongation diverts arachidonic acid away from prostaglandin synthesis
Histamine synthesis in adipose and transport into liver Amplifies selected alleles Histamine can modulate inflammation, blood lipids and lipoprotein fractions
Polyamine transport in adipose, heart and skeletal muscle Direction depends on tissue and variant Polyamines influence inflammation, oxidative stress, vascular endothelium and muscle biology
N-acetylglucosamine 2-epimerase in liver Interacts with PLG-region alleles Its RNBP enzyme also binds renin; the balance between catalytic homodimer and renin-binding complex may connect plasminogen with the renin-angiotensin system

These mechanisms are biologically informed hypotheses built from network topology and prior literature, not direct experimental demonstrations of the human allele-flux interaction.

BCAR1/CFDP1 Region

Mitochondrial transport of stearidonoyl-carnitine in skeletal muscle amplified 38 risk variants at the BCAR1/CFDP1 locus. BCAR1 regulates vascular smooth-muscle-cell migration, proliferation and apoptosis. Stearidonoyl-carnitine links mitochondrial fatty-acid transport with stearidonic acid, an omega-3 precursor whose inflammatory and cardiovascular effects are context dependent.[1]

The proposed connection is that fatty-acid transport changes the availability of polyunsaturated substrates that regulate smooth-muscle behavior, interacting with a locus that already affects those cells.

TGFB1 Region

Two intronic TGFB1 variants were amplified by adipose Golgi UDP-diphosphatase flux. UDP hydrolysis supports UMP-dependent transport of nucleotide sugars into the Golgi, affecting glycosylation and extracellular UDP-sugar release. Both glycosylation and purinergic signaling can modify TGF-beta pathway function and adipocyte biology.[1]

SMARCA4 Region

Brain sodium-coupled galactose transport amplified 19 variants in an LD block near SMARCA4 that was distinct from neighboring LDLR risk variants. SMARCA4 participates in chromatin remodeling, inflammation, vascular calcification and myocardial proliferation, but the specific metabolic connection remains less resolved.[1]

EDNRA Region

Heart sodium-coupled exchange of homoserine and asparagine amplified a risk variant upstream of EDNRA. EDNRA mediates endothelin-1 vasoconstriction and vascular effects; circulating homoserine and asparagine have cardiovascular associations, but the mechanism joining the amino-acid transport flux to this locus remains uncertain.[1]

Coronary Atherosclerosis Versus Myocardial Infarction

The authors repeated the interaction analysis for myocardial infarction (MI), identifying 426 significant SNP-flux pairs and 26 locus-flux relationships. Of the 583 CAD pairs, 360 were also significant for MI and 91 were borderline in both MI tests.[1]

The overlap indicates shared biology, but phenotype differences mattered. Stearidonoyl-carnitine-by-BCAR1/CFDP1 and UDP-diphosphatase-by-TGFB1 interactions were not significant for MI, while MI-specific buffering appeared for pyrimidine-metabolism reactions at the PDE5A/MAD2L1 and MAP1S/FCHO1 regions.[1]

This cautions against treating coronary atherosclerosis and MI as interchangeable outcomes: MI includes plaque disruption, thrombosis and acute clinical events beyond the presence of atherosclerotic disease.

External Validation in All of Us

Validation used genetically predicted fluxes in 118,058 European-ancestry participants from the All of Us Research Program. There were 14,117 CAD cases and 75,204 controls.[1]

For the 583 UK Biobank CAD pairs, interaction estimates correlated at 0.796 and dosage-specific estimates at 0.80 between cohorts. A total of 253 pairs were significant by at least one validation test, 131 more were borderline, and 548 of 583 (94%) had the same interaction direction.[1]

For MI, cross-cohort correlations were 0.90 and 0.91, with 410 of 426 pairs (96%) directionally consistent. These results support reproducible direction and effect structure, but not every discovery pair passed the original two-test significance rule in the smaller validation cohort.[1]

Implications for Precision Medicine

If experimentally validated, gene-by-flux interactions would imply that:

  • a metabolic intervention could have different efficacy across genotype groups;
  • a pathway with no average population effect could matter strongly in genetically susceptible people;
  • buffering reactions might identify protective compensatory mechanisms;
  • amplifying reactions might reveal contexts in which a risk allele becomes especially penetrant;
  • combined genotype and modeled network state could improve mechanism-based stratification.

These are translational hypotheses. The current evidence is observational and model-based, so it does not establish that changing the reaction flux will change clinical outcomes.

Relationship to Adjacent Concepts

Limitations

  • Fluxes are simulated from genetically imputed transcription, not directly measured reaction rates.
  • Diet, lifestyle, BMI, medication, substrate availability and post-translational enzyme regulation are not represented in the personalized flux inputs.
  • Coding and splicing variants are not modeled explicitly because systematic quantitative effects on enzyme activity are unavailable.
  • The discovery and validation analyses described here were restricted to inferred European genetic ancestry.
  • Interaction testing at this scale is sensitive to prediction error, LD, phenotype definitions and multiple-testing choices.
  • Regressing the risk SNP out of predicted flux reduces, but may not eliminate, dependencies induced by shared genetic predictors.
  • The inferred mechanisms for non-coding risk loci remain provisional because the variant-to-function gap is unresolved.
  • No direct experimental perturbation validated the predicted human allele-by-reaction interactions.
  • Hazard interaction on the multiplicative model scale is not automatically equivalent to biological interaction or clinical effect modification on an absolute-risk scale.

Reproducibility Checklist

A comparable study should report:

  • metabolic network reconstruction and organ-specific extraction procedure;
  • expression reference, imputation model and transcript-to-flux integration algorithm;
  • reaction annotation, scaling and correlated-flux pruning thresholds;
  • ancestry inference and cohort inclusion criteria;
  • disease definitions, follow-up dates and control exclusions;
  • origin and within-cohort validation of candidate risk variants;
  • residualization of genotype effects from predicted flux;
  • full Cox model, covariates, stratification and quadratic terms;
  • dosage-specific procedure and weighting;
  • multiple-testing universe and FDR threshold;
  • LD rule used to collapse SNPs into independent loci;
  • marginal flux associations alongside interaction results;
  • cross-cohort effect correlations, direction agreement and significance replication;
  • clear separation of statistical findings from proposed biochemical mechanisms.

Citations

[1] Foguet et al. (2025), "Metabolic reaction fluxes as amplifiers and buffers of risk alleles for coronary artery disease" [2] Foguet, C., Xu, Y., Ritchie, S.C., Lambert, S.A., Persyn, E., Nath, A.P., Davenport, E.E., Roberts, D.J., Paul, D.S., Di Angelantonio, E., Danesh, J., Butterworth, A.S., Yau, C., & Inouye, M. (2022). Genetically personalised organ-specific metabolic models in health and disease. Nature Communications, 13, 7356.