Evaluating Polygenic Score Transferability for Lipid Traits in Underrepresented Populations: Evidence from Samoan Cohorts

Modern precision medicine relies on polygenic scores (PGS). These are mathematical aggregates that sum the effects of thousands of genetic variants. They aim to predict an individual's predisposition to complex traits like cholesterol levels. While these scores work well in European populations, their utility elsewhere is uncertain. Pacific Islander populations comprise only 0.002% of GWAS (genome-wide association studies) participants as of 2024. Consequently, the predictive power of these scores in these groups remains unmapped.

A recent study involving Samoan cohorts addresses this gap. Researchers investigated how well existing multi-ancestry genetic scores for lipid traits function in Samoan adults. Their findings reveal a striking technical dependency. Success depends not just on biology, but on how the mathematical model handles genetic variants.

The failure of curated shortcuts

Transferring polygenic scores across ancestries is difficult because genetic "architecture" is not universal. Two factors cause scores to degrade in new groups. First, allele frequencies (how common a mutation is) vary. Second, linkage disequilibrium (the pattern in which certain markers are inherited together) differs.

Current practice often uses a "pruning-and-thresholding" approach. This method selects a small subset of genetic variants to reduce noise. However, the authors demonstrate that this shortcut can fail underrepresented groups. When testing a curated LDL-C score (PGS000889) on Samoan data, the researchers found a variant matching rate of only ~9%. Because the score looked for the "wrong" markers, the predictive performance dropped to near zero. This suggests that many reported failures in non-European populations may stem from restrictive, curated models. These models fail to capture the actual genetic landscape of the population.

Benchmarking through genome-wide harmonization

To test if scores could work when built correctly, researchers used genome-wide coverage. They implemented the PRS-CS approach (PGS000888), which uses approximately 1.24 million variants.

The researchers' workflow included several critical stages: 1. Genomic Imputation: They used a Samoan-specific reference panel to fill in missing genetic data. This ensured the genotypes accurately reflected the local population. 2. Coordinate Lifting: They used the UCSC liftOver tool to translate variant weights to the current genomic standard (GRCh38). 3. Variant Harmonization: They aligned global weights with the specific variant dosages found in the Samoan cohorts. 4. Statistical Modeling: Performance was measured using linear mixed models. They used the PGS as the primary predictor and adjusted for age, sex, and ancestry.

By ensuring a 99.6–99.7% matching rate, the researchers bypassed the "missing marker" problem. This allowed the genome-wide score to perform meaningfully.

Predictability varies by lipid trait

The study results, shown in, show that predictability is not uniform.

The authors report that HDL-C (the "good" cholesterol) had the highest performance. It showed an incremental $R^2$ (the extra variance explained by the score) of 5.0% to 15.0%. Total cholesterol (TC) followed, with an incremental $R^2$ of 5.0–10.7%.

LDL-C (the "bad" cholesterol) predictions were successful only with the genome-wide PRS-CS model. This achieved an incremental $R^2$ of 5.7–8.6%. In contrast, triglycerides (TG) showed the lowest performance, ranging from 3.5% to 7.0%. The authors suggest this lower signal for triglycerides may result from two factors. First, there were higher rates of monomorphic variants (sites where everyone has the same value). Second, environmental factors play a huge role. The Samoan cohorts spanned a period of rapid dietary transition. This "noise" from changing lifestyles may have hidden the genetic signal.

Limits of the Samoan benchmark

Several caveats remain. The sample size of 4,342 participants is large for a Pacific Islander study. However, it is still modest compared to European studies. This produces wider confidence intervals for the results.

The study also highlights a potential "gene-environment" issue. The data covers twenty years of rapid modernization in Samoa. It is difficult to tell if performance changes are due to genetics or lifestyles. Finally, the researchers note these scores came from multi-ancestry studies without direct Samoan representation. While the scores are transferable, they are not yet fully optimized for Pacific Islander genetic nuances.

The verdict: Coverage is king

The takeaway for researchers is clear: depend on genome-wide coverage, not curated subsets. Pruning methods used in Europe are likely to fail in underrepresented populations. The study proves that meaningful transferability is possible. LDL-C prediction rates in Samoans were comparable to those in East Asian populations. However, this requires strict variant harmonization. Practitioners should mandate the reporting of variant matching rates in all future assessments.