Development of polygenic scores for the Mexican population for obesity, diabetes, and dyslipidemias

Abstract

Genetic prediction estimates a risk for a diseases using genetic data, aiding with earlier diagnoses, prevention, and targeted treatments. Polygenic scores (PGS) estimate risk by multiple single nucleotide polymorphisms (SNPs), the most common form of genetic variation. Though each SNP has a small effect, their joint effect provides key insights into the risk for common diseases influenced by multiple genetic factors. A key limitation of PGS is that the majority have been trained on European populations, leading to a significant drop in predictive accuracy when applied to non-European groups. This study aimed to address this issue by improving the accuracy of PGS for type 2 diabetes (T2D), BMI, Triglycerides, Total Cholesterol, HDL, and LDL levels in the Mexican population through a series of strategies. We implemented various established methods for constructing PGS, including techniques that have shown success in non-European populations and ensemble models combining ancestrybased PGS scores to optimize accuracy across diverse populations. Our key innovation lies in applying shrinkage to the ancestry-based PGS according to each individual’s ancestry proportions, prioritizing ancestry-based scores that are genetically closer to the individual and enhancing the relevance of matched ancestry data. Our results showed no improvement, and in some cases, a decrease in accuracy when using multi-ethnic or Mexican training data, likely due to the underrepresentation of non-European individuals and the small sample size of the Mexican GWAS. However, notable exceptions included LDL and Triglycerides predictions, where the Mexican GWAS outperformed the European GWAS. This outcome may be attributed to genetic loci associated with lipid levels unique to Mexicans, some linked to Amerindian ancestry which explain a greater variance than the loci captured in the European GWAS. Moreover, ensembles incorporating both ancestry adjustment and the Mexican-based PGS underperformed compared to the European baseline model, whereas those excluding the Mexican-based PGS exceeded the European baseline’s performance. Ensembles constructed with Lassosum and LDpred2 fell short of the PRScsx ensemble’s results, suggesting an advantage to jointly modeling multiple populations rather than treating them separately. Introducing ancestry adjustment in PRScsx (RAW4) maintained accuracy and, in some traits, even improved it for subgroups with predominantly African ancestry, showing promise in the proposed ancestry-based shrinkage approach. However, despite these improvements, disparities in accuracy persisted across population subgroups, especially for individuals with a high proportion of African ancestry. These results highlight the current challenge of generalizability gaps in PGS models, even for methods designed for diverse populations like PRScsx. Future studies could focused on developing a sophisticated Bayesian framework for ancestry adjustment, refining ancestry estimation methods, or incorporating a Native American component to better capture the genetic diversity in the Mexican population.