Polygenic risk scores (PRS) quantify disease vulnerability by evaluating up to millions of genetic variants across an individual genome. Because historical genome-wide association studies (GWAS) skew overwhelmingly toward European cohorts, deploying these models directly onto non-European populations predictably tanks diagnostic accuracy. This failure stems from fundamental biological divergence: differences in population genetic architecture, linkage disequilibrium patterns, and variant allele frequencies. To avoid the massive capital expenditure of de novo cohort recruitment, biotech and healthtech teams routinely cut corners using transfer learning—treating legacy Western datasets as foundation models for underrepresented populations.
The Limits of Weight Transfer
Google Research rigorously tested whether this shortcut holds up. Modeling data from European participants in the UK Biobank alongside Japanese participants in Biobank Japan across eight clinical traits—including BMI, systolic and diastolic blood pressure, red and white blood cell counts, HDL/LDL cholesterol, and blood glucose—the team tracked predictive performance across scaling sample sizes. Within the UK Biobank cohort, estimated SNP-based heritability across these traits spanned 0.07 to 0.28.
Pairing Biobank Japan sample sizes against UK Biobank datasets using GWAS discovery pipelines and elastic net regression revealed a stark methodological paradox: importing European baseline weights helps only when local target data is severely bottlenecked. The moment local target cohorts expand, models trained exclusively on local data outperform transfer learning setups. In practice, co-training with foreign European data actively restricts the ceiling of predictive accuracy achievable through native cohort sampling.
Population-Specific Genetic Architecture
The inflection point where foreign data turns from an asset into a bottleneck hinges entirely on the underlying genetic architecture of each trait. Phenotypes with higher trans-ethnic genetic correlation tolerate pooled European training data across larger target sample sizes. Conversely, metabolic and lipid traits hit diminishing returns early. For these phenotypes, retaining European baseline data injects persistent structural noise rather than predictive signal once target datasets scale.
For pharma executives and clinical AI developers, the strategic takeaway is unambiguous: transfer learning functions purely as a temporary bootstrapping measure, not a scalable shortcut. Fine-tuning on Western baselines will inevitably cap diagnostic precision. Reaching commercial-grade clinical accuracy in global markets demands direct balance-sheet commitments to local cohort recruitment rather than endless algorithmic compensation for unrepresentative training data.