Abstract
Accurate disease risk stratification is essential for personalized prevention and treatment. While Polygenic Risk Scores (PRS) aggregate effects from numerous single nucleotide polymorphisms (SNPs), they typically rely on a one-dimensional additive model. This study introduces the Polygenic Risk Vector (PRV), a multi-dimensional approach that incorporates additive and non-additive genetic effects. Using data from the UK Biobank (UKBB) and the eMERGE Network, we demonstrate that combining additive and dominant PRS models significantly improves risk stratification for cardio-metabolic diseases like Type 2 Diabetes and hypertension, particularly for individuals in high-risk groups.
Introduction
Genome-wide association studies (GWAS) have identified thousands of genetic markers associated with complex diseases. Polygenic Risk Scores (PRS) translate this information into a clinical tool by aggregating these markers into a single value representing an individual's genetic liability. However, most PRS methods assume an additive model, where each risk allele increases risk linearly. This simplification ignores non-additive effects like dominance or recessivity, potentially losing critical information for identifying high-risk individuals.
Methods
The researchers developed a multi-dimensional Polygenic Risk Vector (PRV) framework consisting of three main stages: Development, Optimization, and Validation.
Multi-Model Scoring
Instead of a single score, three separate PRSs are calculated for each individual based on different genetic architectures: Additive, Dominant, and Recessive. The coding for these models varies based on the presence of the minor allele (a).
Risk Surface Construction and Tuning
The scores are mapped onto a multi-dimensional "risk space." For a 2D PRV (e.g., Additive + Dominant), a grid is created where each cell represents a combination of percentiles for both scores. Using a training dataset, actual disease prevalence is calculated within each grid cell. A grid search then identifies the optimal combinations of scores that define "high-risk" and "low-risk" groups.
Datasets
- UK Biobank (UKBB): Data from ~500,000 individuals was used for development (70% training, 30% tuning).
- eMERGE Network (Phase III): An independent validation set of ~70,808 individuals linked to Electronic Health Records (EHR) from 11 US sites was used to test real-world generalizability.
Key Insights and Results
- Non-Redundant Information: Although Additive and Dominant scores are correlated, they provide distinct information. The "risk contours" on 2D maps are non-linear, allowing for further stratification of individuals with identical additive scores.
- Superior Stratification: For both Type 2 Diabetes (T2D) and hypertension, the 2D PRV (Additive + Dominant) consistently outperformed the standard 1D additive PRS. In the eMERGE validation set, the PRV identified high-risk groups with significantly higher Odds Ratios (OR) for disease enrichment compared to single-model scores.
- The "Tail" Matters Most: The PRV shows the most significant improvement at the extreme high end of the risk distribution, which is the most clinically relevant group for preventative interventions.
- Recessive Model Limitations: For T2D and hypertension, the recessive PRS provided negligible predictive value, suggesting these traits are primarily driven by additive and dominant effects.
Discussion and Future Work
While promising, the study focused on individuals of European ancestry to ensure consistency across datasets. Future work must validate PRVs in diverse populations to ensure equitable clinical utility. Additionally, integrating PRVs with clinical factors (age, BMI) is necessary to provide absolute risk probabilities. Future refinements could also include advanced Linkage Disequilibrium (LD) handling for SNP selection to further refine the accuracy of the individual scores within the vector.
Summary updated