Why European-ancestry GWAS underserve Loudoun's South Asian community: PRS transferability in type 2 diabetes

Central Question

How well do type 2 diabetes polygenic risk scores derived mainly from European-ancestry GWAS carry over to South Asian-ancestry populations, and which measurable quantities account for the shortfall?

Background

Evidence Summary

• Accuracy, named metric: in a multi-ancestry type 2 diabetes polygenic score evaluation, incremental AUC was 0.061 in South Asian-ancestry validation cohorts against 0.143 in European-ancestry cohorts. Same score, same metric, same study (Huerta-Chagoya et al. 2026).
• The same study's best matched single-ancestry scores reached an incremental AUC of 0.02 to 0.04 in South Asian cohorts against 0.07 to 0.14 in European cohorts, and single-ancestry performance correlated positively with discovery GWAS size (Huerta-Chagoya et al. 2026).
• Odds ratio per standard deviation, a different metric, same score: 2.23 (95 percent confidence interval 2.05 to 2.42) in South Asian and 2.36 (2.32 to 2.41) in European ancestry (Huerta-Chagoya et al. 2026).
• Ancestry-matched scores are not yet the answer here: in the DIAMANTE meta-analysis the ancestry-specific South Asian genetic risk score explained under 1 percent of trait variance on the pseudo R-squared scale, and the European-derived score outperformed it, because far more variants reached genome-wide significance in the larger European meta-analysis (Mahajan et al. 2022).
• Combining training data helps: one study reported a relative improvement above 70 percent in a South Asian UK Biobank cohort when European training data (effective n about 40,000) was combined with South Asian training data (effective n about 16,000). It published no absolute R-squared for that cohort, so none is quoted here (Marquez-Luna et al. 2017).
• Allele frequency and linkage disequilibrium are not the whole explanation for this group. Averaged across traits, those two quantities account for about 37 percent of the loss of relative accuracy in South Asian ancestry, against about 86 percent in African ancestry (Wang et al. 2020).
• The architecture is weighted differently, not merely measured less precisely. In Genes & Health, partitioned scores for insulin deficiency and unfavourable fat distribution carried a greater genetic burden in British Pakistani and British Bangladeshi participants than in white European UK Biobank participants; those high on both were diagnosed 8.2 years earlier with a body mass index 3 kg/m2 lower than those low on both (Hodgson et al. 2024).
• A newer score is not automatically better for every group: the multi-ancestry score significantly outperformed the best previously published score for African or African American, Admixed American and European groups, but not for South Asian ancestry, where a prior score reached 0.068 against its 0.061 (Huerta-Chagoya et al. 2026).

Mechanistic Chain

1. 1. Discovery-sample composition. A polygenic score inherits the population it was estimated in. European ancestry supplied 51.1 percent of the effective sample size of the reference type 2 diabetes meta-analysis and South Asian ancestry 8.3 percent (Mahajan et al. 2022); field-wide the gap is larger still, at 86.3 percent against 0.8 percent (Fatumo et al. 2022). Score performance tracked discovery sample size across all five ancestry groups in the multi-ancestry evaluation (Huerta-Chagoya et al. 2026).
2. 2. Linkage-disequilibrium tagging differences. A score almost never carries the functional variant itself; it carries a nearby tag variant that travels with it. How faithfully a tag stands in for its neighbour depends on local haplotype structure, which differs between populations, so a weight estimated on a European haplotype background meets a different correlation structure (Wang et al. 2020). Handling of linkage disequilibrium was named as a methodological requirement for applying scores to non-European cohorts (Duncan et al. 2019).
3. 3. Allele-frequency differences. A variant frequent enough to detect in one population can be too rare to detect in another, and ancestry-specific variants tend to be rarer, so discovery in the under-represented group is limited to those with relatively large effects (Mahajan et al. 2022). With the tagging term, allele-frequency differences account for about 37 percent of the loss of relative accuracy in South Asian ancestry, averaged across traits (Wang et al. 2020).
4. 4. Effect heterogeneity. The cross-population correlation of effect sizes at the underlying functional variants, and the heritability ratio between populations, enter the relative-accuracy expression directly (Wang et al. 2020). For type 2 diabetes this term appears comparatively small, since association signals are mostly shared across ancestry groups despite differing allele frequencies and linkage-disequilibrium patterns (Mahajan et al. 2022). Pathway weighting does differ, though (Hodgson et al. 2024).
5. 5. Environmental and gene-environment differences. Reviews of transferability name modifiable risk factors alongside genetic ones as limits on how far a score travels (Kachuri et al. 2023). Environment is not the whole explanation: when training and testing are confined to one cohort, which reduces environmental and genotyping confounding, accuracy still falls in proportion to genetic distance (Prive et al. 2022). Cohort setting is a confounder in its own right, since ascertainment and phenotype definition differ between biobanks.
6. 6. Resulting accuracy loss. These terms compound into the measured shortfall: incremental AUC 0.061 against 0.143, same score and metric (Huerta-Chagoya et al. 2026). The loss is not a step change at a group boundary. Accuracy declines continuously with genetic distance from the training data, including inside one labelled ancestry group, which is why a group-level average is an incomplete description of that group (Ding et al. 2023).

Tissue Specificity

Secondary in a methods brief, since the subject is a statistical model rather than a pathway. One cited note: the partitioned scores most strongly associated with type 2 diabetes and younger diagnosis age in British Pakistani and British Bangladeshi participants were insulin deficiency, implicating pancreatic islet beta cells, and unfavourable fat distribution, implicating adipose tissue (Hodgson et al. 2024).

Counterarguments / Limitations

• Multi-ancestry scores narrow the gap. They do: the multi-ancestry score beat its matched single-ancestry counterpart for South Asian ancestry, and in coronary artery disease a multi-ancestry score lifted the South Asian relative effect size from 0.97 to 1.19 of the European reference (Patel et al. 2023). But in type 2 diabetes it did not significantly outperform the best previously published score in South Asian ancestry, where a prior score reached 0.068 against its 0.061 (Huerta-Chagoya et al. 2026). The gap is narrowing, not closed, and not monotonically.
• "South Asian" is internally heterogeneous. Correct, and a genuine limit on every figure here. The reference meta-analysis separated Sri Lankan, Bangladeshi and South Indian study populations from North Indian and Pakistani ones along its own axes of genetic variation (Mahajan et al. 2022), and no verified source reports a transferability metric for any one subgroup. The asthma literature is worse on this point: its nearest cohort is labelled Central and South Asian combined (Tsuo et al. 2022). A single South Asian number averages over real structure.
• Cohort setting differs from Loudoun. It does. The South Asian evidence comes overwhelmingly from British Pakistani and British Bangladeshi participants in Genes & Health (Finer et al. 2019) and from South Asian UK Biobank participants, recruited in another country with a different health system, ascertainment and environment. Loudoun's largest recorded South Asian Census group is Asian Indian, at 45,468 residents, not the predominant origin group in those cohorts. No verified source evaluates a type 2 diabetes score in a South Asian-ancestry population recruited in the United States.
• Accuracy metrics differ across studies, so the figures here should not be pooled. Incremental AUC, variance explained on the liability scale, odds ratio per standard deviation, pseudo R-squared and relative accuracy answer different questions and do not share a scale, which is why every value above carries its metric name. Even the denominator shifts: in a blood-pressure score evaluation, residual variance, the denominator of variance-explained statistics, differed across self-reported background groups by up to 25 percent (Kurniansyah et al. 2023). Two numbers from two studies are not a trend.

Validation Criteria

• A South Asian-ancestry type 2 diabetes GWAS large enough that an ancestry-matched score clears the pseudo R-squared below 1 percent reported for the current one, separating the sample-size term from the tagging and frequency terms.
• An evaluation of existing type 2 diabetes scores in a South Asian-ancestry cohort recruited in the United States, reporting the same metric as a European-ancestry cohort in the same study, so cohort setting is not confounded with ancestry.
• A trans-ancestry genetic correlation for type 2 diabetes between European and South Asian ancestry, which would bound the effect-heterogeneity term directly rather than leaving it as a residual.
• Transferability metrics reported separately for South Asian subgroups, at minimum separating North Indian and Pakistani from South Indian, Bangladeshi and Sri Lankan study populations, since the reference meta-analysis already resolves that structure.
• An allele-frequency and linkage-disequilibrium attribution computed for type 2 diabetes specifically in South Asian ancestry rather than averaged across traits.
• Replication of the partitioned-score finding on insulin deficiency and fat distribution in a cohort independent of Genes & Health.