Trans-Ethnic Meta-Analysis of Metformin Glycemic Response

Biostatistics

Meta-Analysis

Pharmacogenomics

A metformin pharmacogenomic marker validated in European cohorts fails to replicate across ancestries, with direct implications for companion-diagnostic design and clinical-trial diversity.

Published

March 16, 2026

The question

Metformin is the first-line drug for type 2 diabetes, yet glycemic response varies widely between patients. Several of the genetic markers used to explain that variation were discovered almost entirely in European-ancestry cohorts, which raises a question precision medicine cannot ignore: do those markers replicate in other populations, or are they European-specific?

I ran a candidate-SNP meta-analysis to find out, focusing on the two best-known markers for metformin response: rs8192675 (SLC2A2) and rs11212617 (near ATM).

Data

Three published GWAS, pulled from the EBI GWAS Catalog, totaling 22,766 participants.

Study	Ancestry	N	Role
MetGen Consortium	European	10,577	Discovery (SLC2A2)
GoDARTS / UKPDS	European	3,920	Discovery (ATM)
SUGAR-MGH	Multi-ancestry	8,269	Replication

SUGAR-MGH was the only study with full genome-wide summary statistics, so I extracted the candidate SNPs by genomic position before loading. To go beyond the two headline variants, I also ran region-based queries across the core metformin pharmacogenes, SLC22A1 (OCT1, hepatic uptake) and SLC47A1 (MATE1, renal excretion), scanning roughly a million rows of summary statistics for any signal in those loci.

Method

The core analysis is an inverse-variance-weighted meta-analysis of the continuous SLC2A2 signal, built around the QC and modeling steps that decide whether a result is trustworthy.

Allele harmonization. I confirmed effect alleles matched across studies and flipped the SUGAR-MGH ATM estimate to align reference alleles.
Fixed-effects (IVW) and random-effects (DerSimonian-Laird) models, with Cochran’s Q and I² to quantify between-study heterogeneity.
Manual derivations verified against meta::metagen(). I computed the weights, pooled estimates, τ², and Q by hand from first principles and checked them against the package output. They match to four decimals.
Fisher’s combined p-value as a sensitivity analysis, to combine evidence for the ATM SNP where an outcome-scale mismatch (binary OR versus continuous HbA1c) ruled out a formal pooled effect.
A post-hoc power analysis, to test whether a non-result was actually informative or just underpowered.

Key result

The SLC2A2 effect reverses direction between cohorts: +0.17 HbA1c units per C allele in the European discovery sample, and −0.03 in the multi-ancestry replication sample. That reversal drives the entire finding.

Forest plot showing MetGen at 0.17 and SUGAR-MGH at negative 0.03, with a non-significant pooled random-effects diamond spanning zero. — Forest plot for rs8192675 (SLC2A2). The effect reverses direction between the European discovery cohort and the multi-ancestry replication cohort, and the pooled random-effects estimate is non-significant.

The consequence shows up sharply in the two pooled models:

Model	Pooled β (HbA1c per C allele)	95% CI	p-value
Fixed-effects (IVW)	0.139	[0.10, 0.18]	1.35 × 10⁻¹³
Random-effects (DL)	0.073	[−0.13, 0.27]	0.47

Heterogeneity is extreme: Cochran’s Q = 15.07, I² = 93.4%, τ² = 0.019. A marker that looks overwhelmingly significant under the homogeneity assumption becomes statistically null once between-study variance is modeled properly.

The mechanism is visible in how the models weight the two studies. Fixed effects lets the larger European study dominate at 84.7%. Random effects, once it accounts for the heterogeneity, equalizes the contributions to roughly 52 / 48, and the signal disappears.

Grouped bar chart comparing fixed-effects and random-effects weights for the two studies, showing the European study's weight dropping and the multi-ancestry study's weight rising. — Study weight reallocation between fixed-effects and random-effects models. The European study’s weight drops from 84.7 percent to 52.3 percent once heterogeneity is modeled.

Neither SNP crossed genome-wide significance in the multi-ancestry cohort, even at the relaxed pharmacogenomic threshold. The European discovery signals sit far above the line; the replication attempts sit near zero.

Scatter plot of negative log-10 p-values showing MetGen and GoDARTS above the genome-wide threshold and SUGAR-MGH far below it. — Negative log-10 p-values by study. The European discovery signals clear the genome-wide threshold while the multi-ancestry replication points fall near zero.

Was the non-replication real, or just underpowered?

This is the question that decides whether the finding means anything, so I answered it directly with a post-hoc power calculation. Given SUGAR-MGH’s precision, its power to detect the European-sized SLC2A2 effect at genome-wide significance was only about 3%. Reaching 80% power would have required roughly 26,000 participants, about three times the sample available.

That result cuts honestly in both directions, which is the point of running it. The non-replication is not clean proof of a true ancestry difference, because the replication cohort was underpowered for a strict genome-wide test. But the effect-direction reversal and the extreme heterogeneity are not artifacts of power, and at a nominal threshold the study was well powered (94%). The most defensible reading is that a single universal effect size is not supported, and that adequately powered multi-ancestry cohorts are needed to say more.

Why it matters

This small analysis mirrors a real problem in global drug development.

Companion diagnostics. A test built on SLC2A2 genotype alone would perform well in European patients and could mislead for patients of other ancestries.
Trial diversity. It is a concrete illustration of why regulators have pushed for ancestry diversity in clinical trials. A marker validated in one population may not predict response in another.
Stratified prescribing. The heterogeneity argues for an ancestry-aware rather than universal model of metformin pharmacogenetics.

Limitations

I would rather state these than have a reviewer find them. Only two studies were available for the primary pooled estimate (k = 2), which limits the precision of the τ² estimate. The association-only files restricted me to a candidate-SNP rather than a genome-wide approach. The ATM comparison is descriptive only, because of the binary-versus-continuous outcome mismatch. And as the power analysis makes explicit, the replication cohort was underpowered for a strict genome-wide test.