TL;DR
Polygenic scores (PGS)—tools used to predict disease risk from DNA—often fail when applied to non-European populations because they are trained primarily on European data. This study demonstrates that the "best" way to fix this isn't universal: for some traits, a small amount of ancestry-specific data is more powerful than a massive European dataset, while for others, "borrowing" information across ancestries is still the winning move.
What Problem Does This Paper Address?
The "genomic divide" is a well-documented crisis in precision medicine. About 86% of participants in genome-wide association studies (GWAS) are of European descent. Consequently, PGS models trained on this data lose significant accuracy when applied to individuals of African, Asian, or Hispanic/Latino ancestry.
This lack of "transferability" happens because different populations have different allele frequencies and Linkage Disequilibrium (LD) patterns (how genetic variants are inherited together). While researchers have tried to bridge this gap by combining scores from different populations, we still don't fully understand the optimal way to mix-and-match data from diverse cohorts to maximize prediction accuracy for underrepresented groups.
Methods
The authors moved away from the standard practice of using "summary statistics" and instead used individual-level genotype data from the UK Biobank.
- LASSO Regression: They used L1-regularized regression (LASSO) to select genetic variants and estimate their weights. This is computationally intensive but avoids some of the assumptions required by summary-statistic methods.
- Importance Reweighting: To combat the massive sample size imbalance (e.g., 300,000 Europeans vs. 5,000 Africans), they applied a technique from survey sampling. They assigned higher "weights" to individuals from minority-ancestry groups during the model-training phase, effectively telling the algorithm to prioritize accuracy in those specific groups.
- Simulation & Empirical Validation: They first tested these strategies using simulated genetic data (where they could control the "ground truth" of genetic effects) and then validated them on 15 real-world traits.
Datasets
- UK Biobank (UKB): The primary source of real-world data, covering ~500,000 individuals.
- 15 Traits: These included anthropometric measures (Height, BMI), blood-related traits (MCV, platelet count), and diseases (Asthma, Type 1 Diabetes, Atrial Fibrillation).
- Ancestry Groups: The study focused on African (AFR), Admixed American (AMR), Central/South Asian (CSA), East Asian (EAS), and Middle Eastern (MID) ancestries, using labels from the Pan-UKBB initiative.
Key Insights and Ideas
- Trait Architecture Dictates Strategy: The "aha moment" is that there is no single "best" multi-ancestry strategy. For traits like Height, which are highly polygenic (influenced by thousands of tiny-effect variants), the sheer size of the European training set still makes it the best predictor for everyone. However, for blood traits like Mean Corpuscular Volume (MCV), a small ancestry-specific training set (n≈5,000) actually outperformed a European set 60 times its size.
- The "DARC" Effect: Traits where ancestry-specific models win often have specific genetic "hotspots" that vary wildly between populations. For example, the DARC locus has a massive effect on blood traits in African populations due to evolutionary pressure from malaria, but this signal is virtually absent or different in Europeans.
- Reweighting is a Balancing Act: Importance reweighting helps reduce bias, but it increases variance. If you weight the minority group too heavily, you effectively reduce your "effective sample size," making the model noisier.
- Adding Data Can Hurt: Surprisingly, for some traits, adding European data to an African-ancestry training set actually decreased performance on African test sets. This suggests that "more data" isn't always "better data" if the genetic architectures are sufficiently different.
Key Results
- MCV & Erythrocyte Distribution: For these traits, the African-ancestry-only PGS (trained on ~5,000 people) significantly outperformed the European-only PGS (trained on ~300,000 people) when tested on African individuals.
- Height: The massive European dataset remained the superior predictor across all ancestries, though a significant "predictive gap" remained compared to its performance in Europeans.
- Predictive Gap: In simulations, when the correlation of genetic effects between ancestries ($\rho$) was low (0.5), ancestry-specific models were almost always superior. As $\rho$ approached 1.0 (identical effects), dual-ancestry models with reweighting became more effective.
Limitations and Future Work
- UK Biobank Bias: The UK Biobank is not a representative sample; participants tend to be healthier, wealthier, and older than the general population ("healthy volunteer bias").
- Small Minority Samples: Even with the UK Biobank, sample sizes for groups like East Asians or Middle Eastern individuals were quite small, limiting the power of ancestry-specific models.
- Computational Cost: Using individual-level data with LASSO is extremely "expensive" in terms of CPU hours compared to summary-statistic methods.
- Future Direction: The authors suggest that the next frontier is moving beyond discrete "ancestry labels" (which are social constructs) toward continuous representations of genetic ancestry, which could allow for even more precise, personalized risk scores.
Summary updated