Hundreds of millions of people eat crops grown on cadmium- and lead-contaminated soils, and we do not know whether the resulting chronic disease burden is the same across human populations. The agricultural side of the problem is well mapped: metals move from soil into staple grain in ways that depend on soil chemistry and crop variety, and even "low-accumulating" maize lines can still exceed legal food limits for cadmium and lead (Accumulation and Transport of Cd, Pb, As, and Cr in Different Maize Varieties in Southwest China). The human side is stuck. Every genetic-causal study of the relevant disease pathways — bone loss, kidney function, mineral metabolism — runs on European biobanks and treats the exposures as purely endogenous biomarkers (A Mendelian randomization study investigating the causal relationship between 35 blood and urine metabolite biomarkers and postmenopausal osteoporosisShort Communication; Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis). Yet cadmium toxicity is precisely a case where ancestry should matter: the exposure is environmental and geographically uneven, and the genes that handle divalent metals vary in effect across ancestral backgrounds.
The proposal: build the first multi-ancestry Mendelian randomization (MR) pipeline for exogenous heavy metal body burden. MR uses genetic variants as instruments — inherited snippets that randomize an exposure at conception, mimicking a randomized trial — to estimate whether an exposure truly causes a disease rather than merely correlating with it. I will instrument circulating cadmium and lead (and, where data permit, arsenic), estimate their causal effects on bone and kidney endpoints, and test whether those estimates and their instruments survive transfer from European to East Asian cohorts, where linkage disequilibrium — the co-inheritance pattern between a marker variant and the true causal variant — differs enough to break instruments discovered in Europe.
Core Claim
Chronic heavy metal exposure has never been instrumented in cross-ancestry MR, and the standard tools fail in ways that are predictable and fixable. The claim is threefold. First, causal estimates of metal toxicity derived from European cohorts cannot be assumed portable: European tag variants lose instrument strength in East Asian linkage disequilibrium, and the exclusion restriction (the assumption that the variant affects disease only through the exposure) may shift as gene–environment interactions change. Second, this can be addressed by nominating instruments from biologically anchored metal-handling loci — the human counterparts of the divalent-cation transporter families characterized in plants, where a single efflux pump gene was shown to cut cadmium accumulation by more than half in a functional assay (Genomic characterisation of the metal tolerance protein gene family and elucidation of functional role in heavy metal tolerance and accumulation in Coptis chinensis) — and then empirically re-validating instrument strength in each ancestry rather than importing it. Third, when metal exposure interacts with genotype, the interaction must be estimated with two-stage residual inclusion, not the more common predictor substitution, because binary disease outcomes make the latter badly biased (Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction).
Research Program
The program has three linked workstreams.
The first workstream constructs the exposure instruments. Candidate instruments come from two sources: genome-wide significant variants for blood and urinary cadmium, lead, and arsenic where such summary statistics exist, and a targeted panel of metal-transporter and metallothionein loci (the human SLC30/SLC39 transporter families, which are the animal analogues of the plant MTP efflux pumps functionally validated in Genomic characterisation of the metal tolerance protein gene family...). Biologically anchored instruments matter here because they make the exclusion restriction auditable: for a transporter with a known single function of pumping divalent cations, horizontal pleiotropy — the variant reaching the outcome through some other route — is a testable claim about biology, not an unfalsifiable assumption. For each ancestry, I recompute first-stage strength (the $F$-statistic measuring how strongly the variants predict metal burden), prune variants whose tagging relationship to the causal locus decays, and harmonize allele frequencies. The methodological template is the glaucoma study that ran identical MR machinery on European cohorts exceeding 218,000 individuals and East Asian cohorts exceeding 212,000 individuals and showed that cross-ancestry concordance is something you verify, not assume (Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations).
The second workstream estimates causal effects and their ancestry dependence. Outcomes are postmenopausal osteoporosis and chronic kidney disease endpoints in FinnGen, BioBank Japan, and the China Kadoorie Biobank, chosen because cadmium is clinically known to damage renal proximal tubules and disrupt calcium homeostasis, and because the existing endogenous-biomarker MR on these same outcome data already gives a concrete point of comparison: it found alkaline phosphatase (OR 1.26), GFR (OR 1.21), and sex hormone-binding globulin (OR 1.38) causally linked to osteoporosis, all in Europeans only (A Mendelian randomization study investigating the causal relationship between 35 blood and urine metabolite biomarkers and postmenopausal osteoporosisShort Communication). I ask whether instrumenting metal burden upstream of those biomarkers reproduces, amplifies, or contradicts that picture, and whether the estimates agree across ancestries. Bidirectional tests guard against the reverse-causation trap that the same study exposed: osteoporosis itself was shown to causally raise alkaline phosphatase (OR 1.02) and lower sex hormone-binding globulin (OR 0.98), which would silently corrupt any naive observational estimate.
The third workstream models gene-by-metal interaction with the estimator choice that Gorfine and colleagues proved necessary. For a binary outcome, two-stage residual inclusion (2SRI) keeps the measured exposure in the second stage and adds the first-stage residuals as an extra covariate to absorb confounding:
$$\mathrm{logit}\,P(Y=1) = \beta_X X + \beta_G G + \beta_{GX}\,(G \times X) + \beta_R \hat{R},$$
where $Y$ is disease status, $X$ the metal burden, $G$ the instrument, and $\hat{R}$ the residual from regressing $X$ on $G$. In logistic models the odds ratio is non-collapsible — conditioning on the residuals changes the estimand even without confounding — and Gorfine et al. showed this makes the alternative predictor-substitution estimator severely biased while 2SRI retains better Type-I error control (Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction). I extend their analysis, which used body mass index in a pooled 44,500-case / 52,235-control colorectal cancer dataset, to exogenous toxicants and to multi-ancestry settings, with a simulation study calibrated to real metal-exposure distributions before any cohort analysis.
The agricultural literature supplies the exposure geography, not the genetics. Field trials establish that grain cadmium varies severalfold across cultivars on the same contaminated soil and that the ranking is unstable against soil conditions — salinity shocks can remobilize cadmium that amendments had sequestered (Nengzhan Zheng et al.) — so I use these data to parameterize plausible within-region exposure contrasts and to identify the high-exposure regions where gene–metal interaction, if it exists, should be detectable.
Baselines
The naive reference point is an ordinary observational regression of disease outcomes on measured blood or urinary metal levels, stratified by ancestry group. This is what most environmental epidemiology does, and it is the correct null to beat because it requires no genetics at all. It will be run on the same cohorts, with the same outcome definitions and the same confounder sets, so any gap is attributable to the causal machinery rather than to data. Beating it means producing estimates that survive bidirectional and pleiotropy checks where the observational estimate does not; the precedent is direct, since Gorfine et al. found the naive observational log-odds ratio for body mass index on colorectal cancer was roughly 0.14 per 5-unit increase against an instrumented estimate of about 0.23 — the naive model understated the effect by around 40% (Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction).
The strongest existing method is the European-only two-sample biomarker MR pipeline of the osteoporosis study (A Mendelian randomization study... postmenopausal osteoporosisShort Communication): inverse-variance-weighted MR with MR-Egger and MR-PRESSO sensitivity analyses, exposure GWAS from UK Biobank (n = 363,228), outcomes from FinnGen R12 with R11 replication. It reports alkaline phosphatase at OR 1.26 for postmenopausal osteoporosis. It is the strongest competitor because it already works on the same outcome data with the same sensitivity machinery; my proposal differs only in the exposure (exogenous metals versus endogenous metabolites) and the ancestry scope, so it must beat it on exactly those two axes. I will run it unchanged on my data as the primary comparator — same instruments-per-exposure budget, same IVW/MR-Egger/MR-PRESSO stack, same genome-wide significance threshold — and additionally run it once with European-derived metal instruments transplanted unmodified into East Asian cohorts. Beating it means one of two results: either the metal instruments recover causal effects that the endogenous-biomarker version attributes to the biomarkers alone, or the transplanted instruments fail (low F, heterogeneity, discordant estimates) while ancestry-revalidated instruments succeed. The second result is the more publishable one, because it is the first quantification of ancestry-bias failure specifically for toxicant instruments.
The key ablation isolates the ancestry-revalidation step: identical pipeline, identical outcomes, identical instrument-discovery rules, with the only difference being whether instruments are re-selected and strength-tested within each ancestry or imported wholesale from the European discovery cohort. This ablation is the proposal's central ingredient, and it directly measures how much of any cross-ancestry discordance is weak-instrument artifact rather than genuine biological heterogeneity. A second ablation swaps 2SRI for predictor substitution in the interaction models, reproducing the bias that Gorfine et al. established for body mass index and testing whether it worsens, as theory predicts, for a skewed toxicant exposure.
Why not just…?
Why not just regress disease on measured metal levels in diverse cohorts? Because confounding and reverse causation eat the estimate. The naive log-odds ratio of 0.14 versus the instrumented 0.23 in Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction is a measured instance of the naive model understating an environmental effect, and the osteoporosis MR study showed disease itself feeds back onto the biomarkers an epidemiologist would regress on — osteoporosis raises alkaline phosphatase (OR 1.02) after the disease starts (A Mendelian randomization study... postmenopausal osteoporosisShort Communication). The observational baseline stays in the design as the null to beat, but no one should expect it to be competitive; the proposal's decisive comparison is against proper MR, not against regressions.
Why not just run European-only MR and extrapolate? This is the serious alternative, because European biobanks are bigger and better phenotyped, and it would be cheaper. The corpus gives three converging reasons it fails: the developers of the GSMR causal-network method explicitly state their causal estimates are bounded by the European summary data beneath them (Causal associations between risk factors and common diseases inferred from GWAS summary data); the mineral-endocrine MR on FGF23 and aortic stenosis leaves cross-ancestry validity of its therapeutic implication open for the same reason (Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis); and European-calibrated genetic criteria have already been shown to misfire clinically in non-White populations for diabetes diagnostics (Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing). Metal toxicity is a worse case than any of these, because the exposure itself is geographically concentrated in exactly the regions where European instruments are weakest. The transplant ablation settles this empirically rather than by argument: if transplanted instruments perform as well as revalidated ones, the extrapolation alternative wins and I will report that.
Why not just fix the food chain — passivation amendments and low-accumpling varieties — and skip the human modeling? Because the agronomic literature itself documents the ceiling of that approach. On severely contaminated Yunnan soil, varieties that passed every relative low-accumulation criterion still produced grain exceeding statutory human-consumption limits for cadmium and lead (Accumulation and Transport of Cd, Pb, As, and Cr in Different Maize Varieties in Southwest China), and immobilized metals are not removed but parked: salinity changes from irrigation or seawater intrusion drive calcium and magnesium ion exchange that redissolves sequestered cadmium (Nengzhan Zheng et al.). Remediation reduces but does not eliminate dietary exposure, so the residual exposure still has to be risk-assessed in the populations actually eating the grain — which are overwhelmingly non-European. The two lines of work are complements: agronomy sets the exposure distribution; this proposal says who is harmed by it and by how much.
Proposed Experiments and Decisive Checks
The first experiment is the portability test, run on the osteoporosis and kidney endpoints with the glaucoma cross-ancestry design as template (Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations). Report, per ancestry: first-stage F-statistics for each instrument set, IVW causal odds ratios with 95% intervals, MR-Egger intercept and MR-PRESSO global pleiotropy tests, Steiger directionality, and between-ancestry heterogeneity. The pre-registered success criterion is ancestry-revalidated instruments achieving F > 10 and concordant or quantifiably divergent causal estimates where transplanted European instruments do not. A null outcome everywhere — metals showing no causal effect in either ancestry with well-powered instruments — is itself a usable result, mirroring the glaucoma study's confirmation of null causality across ancestries.
The second experiment is the interaction study with the 2SRI/predictor-substitution ablation on binary kidney-disease endpoints, preceded by the simulation study calibrated to real exposure distributions. The simulation's decisive output is a bias-versus-exposure-skew curve showing whether toxicant exposures, which are far more right-skewed than body mass index, push predictor substitution past usable bias in the regime Gorfine et al. flagged.
The third component links the exposure geography to the causal estimates: using the cultivar and soil-mobility findings to define high- and low-exposure regions, and testing whether gene–metal interaction terms are detectable precisely where exposure contrasts are large. The GF-5 hyperspectral soil-mapping pipeline — which reached validation $R^2$ of 0.6029 for zinc and 0.3038 for nickel over an opencast coal mine (Using a Two-Stage Scheme to Map Toxic Metal Distributions Based on GF-5 Satellite Hyperspectral Images at a Northern Chinese Opencast Coal Mine) — shows what regional exposure mapping can achieve and also its transferability ceiling, since those models omit soil-moisture and topography covariates that vary across regions; I treat it as a feasibility demonstration, not a turnkey input.
The honest feasibility check comes first, before any of this: whether genome-wide summary statistics for blood or urinary cadmium, lead, or arsenic exist for East Asian cohorts at usable sample sizes. No reported result found on Lacuna for a GWAS of urinary or blood cadmium, lead, or arsenic in East Asian biobanks. If such statistics do not exist, the transporter-anchored instrument panel becomes the primary design rather than a supplement — smaller but testable — and a pilot GWAS of urinary metals within an East Asian biobank becomes a deliverable of the project rather than an input to it. That contingency is stated up front because it is the single fact most likely to determine whether the full program runs or a scoped version does; either way, the transplant ablation and the 2SRI-versus-predictor-substitution comparison can be completed on whatever instrument sets are actually obtainable.
References
- Accumulation and Transport of Cd, Pb, As, and Cr in Different Maize Varieties in Southwest China
- A Mendelian randomization study investigating the causal relationship between 35 blood and urine metabolite biomarkers and postmenopausal osteoporosisShort Communication
- Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis
- Genomic characterisation of the metal tolerance protein gene family and elucidation of functional role in heavy metal tolerance and accumulation in Coptis chinensis
- Unveiling challenges in Mendelian randomization for gene–environment interaction
- Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
- Responses of Soil Cadmium Desorption under Different Saline Environments and Its Controlling Factors
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
- Using a Two-Stage Scheme to Map Toxic Metal Distributions Based on GF-5 Satellite Hyperspectral Images at a Northern Chinese Opencast Coal Mine