SNPRS:Stacked Neural network for predicting Polygenic Risk Score

In recent years, polygenic risk scores (PRS) have increasingly been used to predict disease susceptibility from genome-wide association studies (GWAS) outcomes. However, these models are limited by overfitting and potential overestimation of the effect size of correlated variants. To address these issues, this study presents a novel Stacked Neural Network Polygenic Risk Score (SNPRS) approach. SNPRS combines outputs from multiple neural network models trained using genetic variants selected across a wide range of p-value thresholds, thus capturing a broader spectrum of genetic variants and more accurately determining the effect size of variant combinations. We tested SNPRS using real data from the UK Biobank to predict the genetic risk of breast and prostate cancer. The results show that SNPRS outperforms conventional models and a single deep neural network model, suggesting that it can significantly enhance the predictive accuracy and relevance of PRS in genetic research.