Genome-wide Association Analysis To Detect Biomarkers Of Alzheimer’s Disease Based On Deep Learning

Genome-wide Associatiion Analysis(GWAS)is widly used to detect the biomarkers of the Alzheimer’s Disease(AD)by evaluating the associations between AD phenotypes and AD genotypes.There exist three ptoblems with this GWAS:1)ignoring the diagnostic information of labels;2)coarsely modelling the complex associations between phenotypes and genotypes with linear mapping;3)high dimensionality of the genotypesTo solve the three problems abovementioned,we proposed a GWAS based on the deep learning(deep-GWAS)to identify biomarkers,including ROI marker and SNP marker.The deep-GWAS approximated the high-dimensional phenotypes to genotypes by adopting the nonlinear model with the strategy which was supposed to thin the parameters.Moreover,supervised learning was used to construct a unified deep learning framework combined with genotype data and disease diagnostic information,finally achieving the prediction of clinical diagnostic information and detection for the related biomarkers of AD.Specifically speaking,our deep-GWAS consist of three associations including association between imaging measurement and disease label,association between imaging latent feature and genetic variants,and the association between genetic latent feature and disease labelIn this paper,we performed our experiments on the ADNI dataset.There were 708 samples in the ADNI data set,each of which contained Magnetic Resonance Imaging(MRI)scan and Single Nucleotide Polymorphism(SNP)of gene data.The 708 samples were consisted of 198 Normal Controls,152 MCInc(Mild Cognitive Impairment,MCI non converter),194 MCIc(MCI converter)and 164 AD.In our study,708 samples were divided into 2 classes,in which NC/MCInc was a class,denoted as 0;AD/MCIc is a class,denoted as 1.After preprocessing MRI and SNP respectively,Finally,each sample includes the volume of the Region of Interest(ROI)in dimension 93 and SNP vector in dimension 501584.It should be noted that our purpose in this study was to detect AD related markers,including ROI markers and SNP markers.Therefore,after modelling the three associations in deep-GWAS,respectively,corresponding markers were finally detected based on our marker detection method.The ROI accuracy of 708 samples is 0.820±0.025.Experimental results showed that the ROI markers including:hippocampal formation right,hippocampal formation left,entorhinal cortex right,amygdala left,subthalamic nucleus right,angular gyrus right,fornix right,anterior limb of internal capsule left,middle temporal gyrus left,etc.The SNP accuracy of 708 samples is 0.70±0.15.The SNP markers including TOMM40,DDX60L,LHFPL2,PHACTR3,LOC105374660,FGD6,LOC112268261,NAALADL2,LOC105374660,LOC100506974,CACNB2,etc.Among them,hippocampal,entorhinal cortex,amygdala and other brain tissues have been found to have morphological changes with the progression of AD,indicating that our marker detection method is effective.Therefore,the SNP markers we detected were associated with AD.Our contributions are as follows with deep-GWAS1)Modelling nonlinearity the super-high dimensional SNP as function using in GWAS;2)Mapping directly the super-high dimensional SNP into the representation space of diagnostic information;3)It provides a strategy to thin the enormous amount of parameters;4)It provides a way to integrate traditional GWAS with deep learning;...