Imagine trying to navigate Tokyo using a highly detailed map of London. While both cities share fundamental urban concepts—roads, intersections, and transit systems—the specific landmarks, street layouts, and signs do not match. If you try to use the London map to find a specific building in Tokyo, you will quickly find yourself lost.
This is the core intuition behind the cross-ancestry portability gap in genomics.
Over the last two decades, Genome-Wide Association Studies (GWAS) have allowed scientists to build Polygenic Risk Scores (PRS)—statistical models that aggregate the effects of millions of genetic variants to predict an individual's likelihood of developing a disease. However, because the vast majority of historical genomic data has been collected from individuals of European descent, these predictive "maps" are heavily optimized for European genetic architectures. When clinicians attempt to export these models to non-European or admixed populations, their predictive power degrades sharply.
Why the Map Fails: The Mechanics of the Gap
The drop in accuracy is not caused by a fundamental difference in the biology of the diseases themselves. Instead, it stems from how genetic information is packaged and distributed across human histories. Two primary statistical phenomena drive this decay: Linkage Disequilibrium (LD) patterns and allele frequencies.
1. The "Passenger and Driver" Problem (Linkage Disequilibrium)
Most variants measured on a standard genetic chip are not the actual biological causes (the "drivers") of a disease. Instead, they are nearby variants that happen to be inherited alongside the causal mutation (the "passengers"). This statistical correlation between neighboring genetic variants is known as Linkage Disequilibrium (LD).
Because human populations split and migrated across the globe at different times, chromosomes have undergone different amounts of historical recombination. In a European population, a specific passenger variant might be tightly bound to a disease-driving mutation. But in an African population—which has a much longer evolutionary history and more generations of genetic shuffling—that same passenger and driver may have parted ways. As a result, a PRS model tracking the passenger variant will successfully predict disease in Europeans but completely miss the signal in individuals of African descent.
2. Shifting Allele Frequencies
A genetic variant that is common and highly predictive in one population might be exceedingly rare, or entirely absent, in another. If a PRS model relies heavily on a variant that only exists in European cohorts, its mathematical weight becomes useless when applied to a population where that variant does not occur. As Y.X. Liu et al. demonstrated in their study on bone mineral density, Variability in performance of genetic-enhanced DXA-BMD prediction models across diverse ethnic and geographic populations: A risk prediction study, as the genetic distance between the training cohort and the target population grows, the predictive power of the model can decay so severely that it becomes less effective than a simple population average.
The Problem of Absolute Scale
The portability gap does not just reduce accuracy; it shifts the entire baseline of measurement.
In a study by Lars G. Fritsche et al., titled On cross-ancestry cancer polygenic risk scores, the authors highlighted that polygenic score distributions shift dramatically between different ancestries. A raw score of "10" might place a European individual in the highest 5% of risk for a disease, but that exact same raw score might represent an average, baseline risk for an East Asian individual.
Because there is no universal absolute scale, clinicians cannot apply a single global threshold for screening or diagnosis. Without population-specific recalibration, applying a Eurocentric model globally results in systemic misclassification—either overdiagnosing or underdiagnosing individuals based purely on their ancestral background.
Methodological Solutions: Rebuilding the Maps
To build equitable clinical tools, researchers are developing mathematical frameworks to bridge this portability gap without waiting decades for global biobanks to fully catch up in size.
Fine-Mapping and Variant Selection
One intuitive fix is to stop tracking "passenger" variants and isolate the actual "drivers." Burcu F. Darst et al. explored this in Evaluating Approaches for Constructing Polygenic Risk Scores for Prostate Cancer in Men of African and European Ancestry. They compared massive genome-wide models containing millions of variants against a highly curated set of just 269 fine-mapped variants.
Surprisingly, the smaller, high-quality variant set performed just as well or better across both African and European ancestries. By stripping away the population-specific "noise" of passenger variants, the model focused purely on shared biological causes, drastically improving cross-ancestry portability.
Modeling Local Genetic Correlation
Instead of treating the genome as a uniform sequence, newer methods model regional differences in genetic architecture. Jiacheng Miao et al. introduced a framework called X-Wing in Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics.
X-Wing isolates specific genomic regions where genetic effects are highly correlated across populations and applies a summary-data-only weighting scheme. By statistically accounting for local LD differences, the authors achieved up to a 119% improvement in predictive power in non-European populations, proving that the portability gap can be mitigated by treating regional architecture as a variable rather than a constant.
Deconstructing Admixed Genomes
The challenge becomes even more complex in admixed populations (such as Hispanic/Latino or African American individuals), whose genomes are mosaics of multiple ancestral origins. A single individual might have a chromosome segment of European origin right next to a segment of African origin.
Admixed Genome Mosaic:
[--- African Ancestry Segment ---][--- European Ancestry Segment ---]
| |
Requires African-tuned Requires European-tuned
LD Reference Panel LD Reference Panel
To address this, Quan Sun et al. developed GAUDI, a penalized-regression method detailed in Improving polygenic risk prediction in admixed populations by explicitly modeling ancestral-differential effects via GAUDI. GAUDI uses a "fused lasso" framework to dynamically adjust weights based on the local ancestry of specific chromosomal segments. By explicitly modeling these ancestral-differential effects within an individual's DNA, the model prevents the signal from being washed out by mismatched reference panels.
The Path Forward
As demonstrated by Sonia Moreno–Grau et al. in Polygenic risk score portability for common diseases across genetically diverse populations, simply applying the "best" European model globally cuts predictive accuracy by half or more for critical conditions like heart disease. However, by intentionally selecting ancestry-optimized models and utilizing frameworks like those developed by Injeong Shim et al. in Clinical utility of polygenic scores for cardiometabolic disease in Arabs, researchers are successfully recovering this lost accuracy.
The cross-ancestry portability gap is ultimately a reminder that in precision medicine, a model is only as precise as its training data is representative. Resolving it requires a dual approach: aggressively diversifying global genetic biobanks while refining the statistical tools necessary to translate complex, multi-ethnic genetic architectures into universal clinical insights.
Key Foundational Work
Analysis of the Genetic Basis of Disease in the Context of Worldwide Human Relationships and Migration (2013) — contributed 1 concept elements to this direction.
Genetic effects influencing risk for major depressive disorder in China and Europe (2017) — contributed 1 concept elements to this direction.