Prediction of Polygenic Risk Score by Machine Learning and Deep Learning Methods in Genome-wide Association Studies: Methodological Study

Objective: We aimed to investigate whether machine learning (ML) and deep learning (DL) methods, utilizing individual-level data from genome-wide association studies (GWAS), could serve as a viable alternative to traditional polygenic risk score (PRS) calculation methods, which rely on odds ratios as weights. PRS is widely used to estimate genetic susceptibility to diseases, but its accuracy and generalizability can be affected by variations in allele frequencies and sample sizes. Given the advancements in ML and DL techniques, we explored their potential for improving risk prediction. Material and Methods: We generated GWAS datasets using the PLINK program, simulating genetic data under various conditions by varying allele frequencies and sample sizes. This process was repeated 100 times to assess the robustness of the approaches. We applied 2 ML algorithms-Support Vector Machine and Random Forest alongside a DL approach. The predictive performance of these methods was compared to the traditional PRS calculation, which uses odds ratios as weights. Results: Our findings showed that ML and DL methods provided more consistent case-control separation than the classical approach. Additionally, they exhibited reduced bias and greater stability across different genetic conditions. Conclusion: ML and DL approaches present a promising alternative to odds ratio-based PRS calculations, offering enhanced reliability and consistency in genetic risk prediction.