| Objective:To explore the predictive effect of machine learning（ML）and polygenic risk score（PRS）on statistical modeling of genetic risk of Alzheimer’s disease（AD）and to investigate the effect of different single nucleotide polymorphisms（SNPs）inclusion strategies on model prediction performance based on different SNPs thresholds of genome-wide summary data.And to provide a more rapid and effective statistical modeling strategy for genetic risk prediction of AD based on genome-wide high-dimensional data.Methods:The SNPs were divided according to different thresholds(1×10-8、1×10-7、1×10-6、1×10-5、1×10-4、1×10-3),and statistical modeling was performed for AD genetic risk prediction based on PRS,Least Absolute Shrinkage And Selection Operator（LASSO）regression,Elastic Net（EN）regression,Ridge regression,Random Forest（RF）,and Extreme Gradient Boosting（XGBoost）models under four different SNPs inclusion strategies:APOE+clumping,APOE+no-clumping,original+clumping,and original+no-clumping.Ten-fold cross-validation was used,and area under curve（AUC）,sensitivity,specificity,and accuracy were used to measure model prediction accuracy.Results:A total of 369 subjects in the Alzheimer’s Disease Neuroimaging Initiative（ADNI）were included,of which 159 were AD patients and 210 were cognitively normal individuals.EN regression and LASSO regression achieved the best predictive performance（AUC=0.774）for the subset of SNPs at P<1×10-4,and the worst predictive performance was achieved by Ridge regression at the subset of SNPs at P<1×10-3（AUC=0.580）.In the threshold range from 1×10-8 to 1×10-5,the AUC of most algorithms did not show a more obvious decreasing trend,and only PRS and Ridge regression showed a decrease in AUC,When SNPs with P<1×10-4 and P<1×10-3 were added,the AUC of most algorithms showed a decrease.For all methods,in the original inclusion strategies,the results did not show an increase in AUC compared to the no-clumping inclusion strategies.Whereas,after incorporating APOE into the model,the AUC of clumping was higher than that of no-clumping inclusion strategies for all methods except PRS and Ridge regression.In the clumping inclusion strategies,the performance of APOE is significantly better than that of the original inclusion strategies,while in the no-clumping inclusion strategies,the performance of APOE in all methods except PRS is the same as that of the original inclusion strategies.Conclusions:APOE+no-clumping inclusion strategies combined with ML algorithm（EN regression and LASSO regression）is an ideal statistical modeling strategy for AD genetic risk prediction,which can identify individuals with high genetic risk of AD and provide a scientific basis and new perspective for AD precision medicine research and clinical treatment. |
