TL;DR
Polygenic Risk Scores (PRS) are less accurate when applied to populations different from the one they were trained on, limiting their global use. This study shows that tailoring PRS models to specific ancestries significantly improves their performance, paving the way for more equitable and effective genetic risk prediction across diverse populations.
What Problem Does This Paper Address?
Polygenic Risk Scores (PRS) are powerful tools that aggregate the effects of many genetic variants to estimate an individual's risk for common diseases. However, most PRS are developed using data primarily from individuals of European ancestry. This creates a significant "portability" problem: these European-derived PRS often perform poorly when applied to individuals from other ancestral groups (e.g., African, East Asian, South Asian). This disparity limits the clinical utility of PRS on a global scale and risks exacerbating existing health inequities. The core challenge this paper addresses is how to improve the accuracy and reliability of PRS across diverse populations.
Methods
The researchers employed a multi-pronged approach to evaluate and improve PRS portability:
- Model Selection: They first identified the best-performing PRS models for each of 14 medical conditions by comparing publicly available scores and by training new models using three state-of-the-art PRS methods: LDpred2, PRS-CSx, and SNPnet.
- Training and Testing: A large European cohort from the UK Biobank (UKBB) was used for training these models (285,000 individuals). The models were then tested on diverse individuals from the UKBB (18,000 individuals across African, European, East Asian, and South Asian ancestries).
- Ancestry-Specific Optimization: To address the portability issue, they re-evaluated PRS performance by selecting the best-performing model for each specific ancestry group, rather than relying solely on European-derived models.
- Independent Validation: The selected optimal PRS models were further validated using an independent dataset, the Biobank of the Americas-Genomelink (BBofA-GL), comprising 8,417 individuals of European ancestry.
- PRS-PheWAS: They conducted PRS-phenome-wide association studies to explore how PRS for specific conditions were associated with other traits, revealing potential cross-trait interrelationships.
Datasets
- UK Biobank (UKBB): A large prospective cohort from the UK used for both training (285,000 European individuals) and testing (18,000 individuals across African, European, East Asian, and South Asian ancestries) of PRS models.
- Biobank of the Americas-Genomelink (BBofA-GL): An independent cohort of 8,417 European ancestry individuals used for validation of the developed PRS models.
Key Insights and Ideas
- European PRS don't travel well: The study strongly confirms that PRS derived from European populations have significantly reduced predictive power in non-European ancestries. This isn't just a minor dip; it's a substantial loss of accuracy.
- Ancestry-specific models are key: The most impactful finding is that selecting PRS models tailored to an individual's ancestry dramatically improves predictive performance. This suggests that genetic architectures and variant effects differ enough across populations to warrant specialized models.
- Cross-trait connections revealed: The PRS-Phewas analysis highlighted interesting biological interplays, particularly between cardiometabolic conditions and between autoimmune components, suggesting shared genetic underpinnings.
- Democratizing PRS: By demonstrating how to improve PRS performance in diverse groups, this work offers a practical strategy to make PRS more equitable and accessible, moving towards "democratizing the implementation of PRS."
Key Results
- When European-derived PRS were applied to non-European groups, predictive performance dropped significantly: 51.3% for South Asians, 46.6% for East Asians, and 39.4% for Africans, compared to Europeans.
- By selecting ancestry-specific optimal PRS models, the researchers observed substantial improvements in phenotype variance explained: 1.62 for Africans, 1.40 for South Asians, and 0.96 for East Asians.
- For specific conditions like CAD, hypertension, hypothyroidism, HDL-C, LDL-C, and T2D, ancestry-specific models allowed performance in diverse populations to become much more comparable to that seen in Europeans.
- The study successfully validated PRS models in an independent cohort, reinforcing their findings.
- The PRS-Phewas identified 65 significant phenotype associations, with 43 indicating increased disease risk and 22 indicating decreased risk, across 14 conditions.
Limitations and Future Work
- Phenotyping Differences: The study acknowledges disparities in PRS effect sizes between the UKBB (using ICD10 codes) and BBofA-GL (using self-reported questionnaires) for some conditions. While self-reported data is becoming more common, it offers less granularity than clinical data.
- Admixed Individuals: The study did not train or evaluate PRS predictability in admixed individuals, which presents a significant layer of complexity due to mixed ancestry haplotypes and challenges in representing LD for these populations.
- Data Diversity: The authors emphasize that the "gold standard" solution to portability problems involves increasing the diversity of participants in genetic studies. Future work should focus on training new PRS models beyond European ancestries and integrating ancestry deconvolution methods.
- Future Research: The work enables future studies to establish best practices for evaluating PRS models, democratize their implementation, and apply them to translational research.
Summary updated