Epistatic Features and Machine Learning Improve Alzheimer's Risk Prediction Over Polygenic Risk Scores

Background. Polygenic risk scores (PRS) are linear combinations of genetic markers weighted by effect size that are commonly used to predict disease risk. For complex heritable diseases such as late-onset Alzheimer's disease (LOAD), PRS models fail to capture much of the heritability and generalize poorly across populations. Objective. Construct a paragenic risk score that, in addition to single-marker PRS information, incorporates epistatic interaction features and machine learning methods to predict lifetime risk for LOAD. Methods. We identified epistatic interactions between SNP loci using an evolutionary algorithm (Crush-MDR) guided by shared pathway information, and estimated risk via an ensemble of machine learning models (gradient boosting and deep learning) rather than logistic regression. We compared the paragenic model to a PRS model trained on the same dataset. Results. The paragenic model achieved an AUC of 0.83 with near-clinically significant matched sensitivity/specificity of 0.75 under 10-fold cross-validation, and remained significantly more accurate than PRS on an independent holdout dataset. The paragenic model maintained accuracy within APOE genotype strata. Conclusion. Paragenic models that incorporate epistatic features and machine learning can improve lifetime disease risk prediction for complex heritable diseases such as LOAD over PRS models.

Stephen Hermes, Janet Cady, Steven Armentrout, James O'connor, Sarah Carlson, Carlos Cruchaga, Thomas Wingo, Ellen Mcrae Greytak