The Ceiling Is Real: Why Your Polygenic Risk Scores Won’t Transfer
If you have ever tried to deploy a polygenic risk score (PRS) trained on European cohorts into an African or Hispanic patient population, you know the shape of the failure. The AUC drops. The confidence intervals widen. The clinical utility vanishes. A new preprint from the Confluence Consortium dissects exactly why this happens. They do not propose a new algorithm. Instead, they map the theoretical limits of common-variant risk prediction across four ancestries. The result reveals that the performance gap is not a bug in the scoring method. It is a feature of the underlying genetic architecture and sample size disparity.
The core finding is stark. The maximum achievable predictive performance (AUC ceiling) for breast cancer is roughly equivalent across African (AFR), East Asian (EAS), and European (EUR) populations. It hovers around 0.71. Yet, AFR requires substantially larger sample sizes to reach that ceiling. The gap we see in production today is not because the biology is different. It is because the data infrastructure is not.
The Problem
Current polygenic risk scores are heavily biased toward European ancestry. Most genome-wide association studies (GWAS) that feed these models are dominated by EUR samples. EAS lags behind. AFR and H/L (Hispanic/Latina) are severely underrepresented. When we apply a EUR-trained PRS to non-European genomes, linkage disequilibrium (LD)—the correlation between nearby genetic variants—differs enough. The tag SNPs used in the score no longer accurately mark the causal variants. The result is poor portability.
Prior literature established that PRS performance drops across ancestries. But it rarely disentangles why. Is the drop due to different genetic architectures? Different effect sizes? Or simply smaller sample sizes and weaker variant tagging? Without answering this, practitioners either abandon multi-ancestry deployment. Or they blindly tune hyperparameters hoping for a marginal lift. This paper argues that the architecture is surprisingly homogeneous. The solution lies in data volume and LD modeling. Not in changing the fundamental approach to PRS construction.
How It Works
The authors treat this as a systems problem. They estimate the capacity of the "signal" (genetic architecture) versus the noise (sampling error). They analyze summary statistics from 159,297 cases and 212,102 controls across AFR, EAS, EUR, and H/L.
1. Heritability Estimation via LDSC They use LD Score Regression (LDSC) to estimate logit-scale SNP-based heritability ($h^2$). This metric represents the variance in disease risk explained by common variants. Crucially, they compare these estimates across ancestries to see if the "signal strength" differs. They find $h^2$ ranges from 0.47 (EAS) to 0.61 (AFR). Heterogeneity tests show no significant difference ($p=0.63$). The signal is there. It is just harder to hear in some populations due to LD structure.
2. Polygenicity Modeling with GENESIS Next, they apply the GENESIS framework. This is a mixture-model approach that estimates the number of non-null susceptibility markers (polygenicity). This allows them to project the performance of "clumping-and-thresholding" (CT) PRS. This is a simple, widely used baseline method. GENESIS simulates what happens if we double or triple the GWAS sample size. It holds the architecture constant.
3. Cross-Ancestry Correlation via Popcorn They estimate genetic correlations using Popcorn. This tool accounts for LD differences. It quantifies how much the effect sizes of variants overlap between populations.
4. Cellular Context via scDRS+ Finally, they integrate GWAS hits with the Tabula Sapiens single-cell atlas using scDRS+. This maps genetic risk to specific cell types. It checks if the biological mechanisms (cellular contexts) are conserved across ancestries.
As shown in ****, the projected AUC curves for AFR, EAS, and EUR converge toward a similar asymptote (~0.71) as sample size grows.
However, the AFR curve is shifted right. It requires more data to achieve the same performance as EUR or EAS.
Numbers
The authors report several key metrics that define the boundary conditions for PRS deployment:
- Heritability: Logit-scale $h^2$ is 0.501 (EUR), 0.466 (EAS), 0.614 (AFR), and 0.588 (H/L). The lack of statistical heterogeneity suggests the underlying polygenic load is comparable.
- Projected AUC at 200k Samples: At a hypothetical 100,000 cases/controls, the projected AUC is 62.1% (EUR), 63.4% (EAS), and 61.1% (AFR). The ceiling approaches ~71% for all three. This means even with perfect data, common variants alone won't capture all risk.
- Genetic Correlation: The strongest correlation is EUR-EAS ($\rho=0.79$). The weakest is AFR-H/L ($\rho=0.26$). This confirms that while the architecture (number of variants) is similar, the tagging efficiency varies wildly due to population history.
- Polygenicity: Estimated non-null markers range from 4,446 (EAS) to 8,308 (AFR). Again, differences are not significant. This implies similar genetic complexity across groups.
The critical takeaway is in ****: the genetic correlation heatmap.
Low correlations involving AFR mean that variants significant in EUR are often not significant in AFR. This is not because they don't matter. It is because the LD structure breaks the association. This is a data problem, not a biological one.
What's Missing
The paper is rigorous in its architectural analysis. But it leaves several operational questions unanswered:
- No Empirical Validation of New Models: The authors explicitly state they do not train or validate new PRS models. They project CT-PRS performance. This is a naive baseline. Modern methods (LDpred, PRS-CS) might behave differently under these constraints. We do not know if the "sample size gap" shrinks or widens when using more sophisticated Bayesian shrinkage methods.
- H/L Non-Convergence: The GENESIS model failed to converge for Hispanic/Latina samples. This is due to small sample size and admixture complexity. This means we have zero architectural estimates for H/L beyond heritability. Any claims about H/L PRS potential are currently speculative.
- Clinical Utility Gap: The paper stops at AUC. It does not translate projected AUC into clinical metrics. Metrics like sensitivity at a fixed specificity are missing. So is the proportion of patients exceeding a risk threshold for screening. An AUC of 0.71 might be statistically sound. But it could be clinically useless if the calibration is off.
- Imputation Reference Bias: The authors acknowledge that Eurocentric genotyping arrays impair imputation in AFR. But they do not quantify the impact of switching to newer, multi-ethnic reference panels. This is a huge lever for improvement. It remains unmeasured.
Should You Prototype This
Depends.
If you are building a new PRS from scratch and have access to diverse cohorts, the GENESIS projections suggest something important. Investing in larger AFR and H/L samples yields diminishing returns on architectural discovery. But it yields massive returns on calibration. The architecture is already known to be shareable.
However, if you are trying to improve an existing EUR-centric PRS for non-European users, this paper does not offer a quick fix. The gap is structural. You cannot "prototype" your way out of small sample sizes with a better algorithm. You need data.
Code is available at https://github.com/confluence-breast-cancer-consortia/pre-confluence-analysis/tree/main. The scripts are transparent. They rely on standard tools (LDSC, Popcorn, scDRS+). If you have the compute and the data, you can reproduce these projections in a day. If you don't, this paper serves as a clear specification for what is needed. More diverse GWAS summary statistics. Specifically for AFR and H/L. Only then can we unlock the theoretical performance limits we already know are there.