Imaging Genetics Study Based On A Sparse Regression Model For Biomarker Detection Of Alzheimer’s Disease

Imaging genetics is a novel interdisciplinary filed,which aims to study the influence of genetic genotype on individuals by using brain neuroimaging.In imaging genetics,associations between neuroimaging phenotypes(e.g.,Quantitative Traits,QTs)and genotype data(e.g.,Single Nucleotide Polymorphism,SNP)are analyzed to evaluate the effects of genetic variations on brain function and structure,and detect disease-related biomarkers.The detected biomarkers can be used as an effective clue for understanding the internal mechanism of the disease,disease prediction and diagnosis.Nowadays,imaging genetics is widely used for biomarker detection of Alzheimer’s Disease(AD).However,there are some problems exist in imaging genetics:1)Missing data issue since high measurement costs or poor data quality of multimodality data;2)Complementary information among multimodality imaging QT data is ignored;3)SNP group information and QT structure information is neglected;4)Information of longitudinal imaging QTs is discarded;5)It is difficult to analyze the complex association between imaging QT and SNP by using a simply linear model.To solve the above-mentioned problems,we proposed a Multimodal and Temporal Group Sparse Regression Model for Biomarker Detection of AD.In this study,the innovations are listed as follows:(1)In clinical,complementary information can be achieved by using multimodality images,such as MRI and PET.However,there may be missing data in multimodality images.Therefore,in order to make full use of the complementary information among multimodalities and deal with missing data to improve the performance of biomarker detection(i.e.,risk SNP and brain imaging QT),a Multitask Sparse Regression model based on Latent Feature Representation(MTSRLR)is proposed.First of all,in order to solve the problem of missing data,we decompose the multimodality data into two parts,one consists of complete multimodality data,which are projected into a common latent space to learn the common features of all modalities,and the other one consists of incomplete multimodality data,which are projected into a separate latent space to learn the unique features of each modality(i.e.,modalityspecific).Then,the association between the learned latent feature representation and SNP data is analyzed for AD-related biomarker detection.We use the alternative convex search method to solve the objective function.Experimental results show that the MTSRLR model outperforms several state-of-the-art methods in terms of the Area Under the Curve(AUC).Moreover,the detected genes and phenotypes of AD have been confirmed in some previous studies,thereby further identifying the effectiveness of our approach.(2)A method based on Temporal Group Sparse Regression and Additive Model(T-GSRAM)is proposed,which can be used to analyze the association between longitudinal phenotype and SNP for AD biomarker detection.Firstly,a nonparametric additive model is introduced to consider the complex association between image QT and SNP,which is more flexible than a linear model.Secondly,longitudinal QT data are used to identify the changes of image genetic association over time.Moreover,the SNP information of group and individual levels are incorporated in the proposed method to boost the power of biomarker detection.Finally,a sparse learning method is used in the T-GSRAM model to select the relevant features,and an effective algorithm is applied to solve the T-GSRAM model.We evaluated our method using simulation data and real data obtained from ADNI.Experimental results show that the T-SGRAM model outperforms several state-of-the-art methods in terms of the AUC.Moreover,the detected genes and phenotypes of AD have been confirmed in previous studies,thereby further identifying the effectiveness of our approach and helping the understanding of the genetic basis over time during disease progression.