Genetic Epistasis Analysis And Genetic Risk Prediction For Alzheimer's Disease

Recently,many loci that displayed small effects in Alzheimer's disease(AD)have been discovered by large-scale genome-wide association studies(GWASs).However,combining these new discovered loci with known disease genes,we can only explain part of the heritability in AD.This is because GWASs have only analyzed common variants,leaving epistasis,structure variants and rare variants untouched.Therefore,we mainly focused on the interaction between genetic loci,namely epistasis.Through the discovery of genetic interactions related to AD and its pathological hallmarks,we gained insights into the pathogenesis and genetic mechanism of AD.We demonstrated that genetic interactions can be used to predict the risk of AD,and the combined risk scores(CRSs)of SNPs and SNP-SNP interactions can better predict the risk of AD compared to polygenic risk scores(PRSs).We applied traditional machine learning models,deep learning models and genetic risk scores to predict individual's disease risk.The convolution neural network(CNN)which included both long-range and short-range interactions showed the best accuracy in predicting the risk of AD.In chapter 1,we summarized the genetic mechanism of AD,and the application of genetic risk scores and deep learning models in genetic risk prediction.In chapter 2,we screened for genetic interactions associated with ?-amyloid(A?)or tau protein through genome-wide interaction analysis.Abnormal deposition of A? and tau protein are the pathological hallmarks of AD.Interactions identified by the ROSMAP dataset(N = 2,090)were further validated using the ADNI dataset(N = 1,550).A total of2803 SNP-SNP interactions for tau protein and 464 SNP-SNP interactions for A? were discovered.Genetic interactions for tau protein mainly lie in genes related to axonal function,such as axon regeneration,axon guidance and axon development.Genetic interactions for A? mainly lie in genes involved in nervous system development,cell determination and cation transmembrane transport.The interacting genes were usually found to be co-expressed,indicating their concerted function in the same biological pathway.Furthermore,through analyzing the relationships between the candidate interactions and the hippocampus volume,the entorhinal cortex volume,and the fractional anisotropy in certain brain regions,we can gain insights about how brain atrophy and white matter loss were related to AD.In chapter 3,genetic interactions related to AD(using the clinical diagnosis of AD as phenotype)were identified based on three db Ga P datasets(N = 10,389).To evaluate if genetic interactions could be used to predict the risk of AD,each sample in the ROSMAP and ADNI datasets was assigned an epistasis risk score(ERS).We found that ERSs could serve as an indicator of AD risk.At the same age,individuals with a higher ERS were more likely to develop AD.Combined risk score(CRS)which includes both additive effects and epistatic effects can be obtained by combining ERS with PRS.CRSs can better predict the risk of AD compared to PRSs.In chapter 4,we constructed a deep learning model to predict the risk of AD.Based on the raw genotype data,machine learning methods can model not only the effect of each locus,but also the epistatic effect between loci.All models were trained on db Ga P datasets(N = 10,389).For different models,five-fold cross-validation or validation on the ROSMAP dataset were performed to select optimal parameters.Using ADNI as the testing dataset,we evaluated the predictive power of PRSs,CRSs,XGBoost,CNN and multi-layer perceptron based on multiple metrics.CNN was the best model under any metrics that was adopted.The important loci in CNN model did not display significant P value in GWAS,which indicated that these loci may interact with other loci to contribute to AD risk.In conclusion,we performed a systems genetic analysis to study the genetic interactions underling AD and pathological hallmarks of AD,and proved that epistasis can be used as a predictor of genetic risk,finally,constructed a CNN model which can better predict the disease risk.