Google researchers evaluated how transfer learning from European genetic data can improve disease risk prediction in underrepresented populations. The study, published on the Google AI Blog, used data from nearly 200,000 Japanese individuals in Biobank Japan and hundreds of thousands of European individuals in the UK Biobank to assess polygenic risk scores (PRSs)—AI models that predict disease risk from genetic variants—across eight clinical traits including BMI, blood pressure, and cholesterol levels. The research revealed a counterintuitive finding: while transfer learning from European cohorts significantly improves prediction accuracy in small non-European populations, it actually degrades performance once target population sample sizes exceed approximately 15,000 individuals. This effect results from differences in genetic architecture between populations—co-training with external European data limits the accuracy gains achievable from larger target-population-specific datasets. The researchers tested three approaches: direct variant transfer with elastic net modeling, meta-analysis combining genetic associations from both populations, and PRS-CSx, a method designed to handle genetic linkage disequilibrium differences across populations. The findings provide practical guidance for healthcare systems and researchers developing genomic risk scores for diverse populations. Rather than always incorporating larger European datasets, optimal model performance requires balancing transfer learning benefits against population-specific genetic differences. This work addresses a significant equity challenge in precision medicine, where historical genome-wide association studies have overwhelmingly focused on European populations, causing substantial prediction accuracy drops for non-European groups.