Research On Machine-learning-Based Imaging Genetics Analysis And Their Applications

The past decade has witnessed the increasing development of multimodal neuroimaging and genomic techniques.Imaging genetics,as an interdisciplinary field,aims to evaluate and characterize genetic variants in individual that influence phenotypic measures derived from structural and functional brain images.This strategy is able to reveal the complex mechanisms via macroscopic intermediates from genetic level to cognition and psychiatric disorders in humans.On the other hand,machine learning methods,considered as the powerful tools in the data-driven based association studies,can make full use of the priori-knowledge(inter correlated structure information among imaging and genetic data)into correlation modelling.Therefore,the association study can address the correlations between risk gene and brain structure or function,so as to help explore the potential to a better mechanistic understanding of behaviors or disordered brain functions.Based on the theory and technology of machine learning,this dissertation tries to solve existing problems within four aspects on imaging genetics analysis: 1)multigenetic SNPs loci to each brain ROI QT;2)single genetic SNP locus to multi-brain ROI QTs;3)multigenetic SNPs loci to multi-brain ROI QTs;4)multi-genetic SNPs loci to dynamic multi-brain ROI QTs.(1)To take full advantage of the hierarchical structure information among SNPs,this dissertation proposes a tree-guided sparse learning(TGSL)method to detect the genetic loci associated to the candidate brain ROI QT from neuroimaging.Specifically,motivated by the biological phenomenon,the hierarchical structures included individual SNPs as leaf nodes and gene groups as well as linkage disequilibrium(LD)blocks as interval nodes.Then,the constructed tree structure are imposed on the regularization of TGSL model to select the relevant features.Finally,regression is used to predict the image-derived measures using the selected SNPs features.Experimental results on the Alzheimer's Disease Neuroimaging Initiative(ADNI)database studies show that our method not only achieves better predictions,but also identifies sparse SNP patterns at the block level to better guide the biological interpretation.(2)To take full advantage of the diagnosis label and complementary information from multimodal brain imaging phenotypes,this dissertation proposes a diagnosis-guided multi-modality(DGMM)analysis method for further improving the associations between genetic risk factors and multimodal neuroimaging biomarkers.Specifically,a valuable scenario would be to discover only those QTs associated with both genetic markers(i.e.,APOE e4)and disease status to better reveal the biological pathways specific to the disease(i.e.,AD).Thus,our proposed DGMM adopts a class-similarity metrics constraint to identify the discriminating intermediate imaging QTs that can bridge the gap between one top risk gene and a disease.In addition,imaging markers consistently showing up in multimodalities may provide more insights for structural and functional characters.Accordingly,the DGMM model includes a group sparse regularization constraint to select the consistent markers across different perspectives.Finally,the accelerated proximal gradient(APG)algorithm is used to solve the proposed optimization problem.The empirical results on ADNI demonstrate that comparing to the traditional single modality analysis method,DGMM method not only help improve the performances of imaging genetic associations,but also discover robust and consistent ROIs across multi-modalities to reveal the chain from genotype to phenotype to symptom.(3)To explore the intrinsic association between multi-SNPs loci and multi-brain ROI QTs using the assistant information such as clinical scores of interest(i.e.,cognitive scores or diagnosis label),this dissertation proposes an clinical scores-guided and outcome-relevant brain imaging genetic associations via three-way sparse canonical correlation(T-SCCA)analysis.Specifically,we extend the traditional SCCA to three-way with multimodalities to discover the relationships among SNPs,imaging QTs,and cognitive and diagnostic outcomes using pairwise bi-multivariate analysis.The empirical results on ADNI demonstrate that the proposed T-SCCA model not only outperform the traditional SCCA method in terms of identifying strong associations,but also discover robust outcome-relevant imaging genetic patterns,demonstrating its promise for improving disease-related mechanistic understanding.(4)To discover the temporal changes on brain structure influenced by genetic SNPs loci in degenerative disease,this dissertation proposes a temporally-constrained group sparse canonical correlation analysis(TGSCCA)framework for identifying genetic associations with longitudinal brain phenotypic markers.Specifically,as the phenotypes associated with risk SNPs loci exhibit differences in the whole degenerative process,the TGSCCA model uses a group sparse regularization constraint to jointly select the consistent markers associated to multi-SNPs loci across multiple time-points.In addition,the characteristics of phenotypes from different time-points refer to the disease-progressive status.As the temporal changes in brain from adjacent time-points should be small,we induce the fused penalty to select the robust ROIs with consecutive changes,which are robust to noises or outliers.Finally,an efficient iteration algorithm is designed to solve the objective function.The empirical results on ADNI demonstrate that in comparison with conventional SCCA,our proposed TGSCCA method can achieve strong associations and discover multi-brain ROIs associated with multi-SNPs loci to guide disease progressive interpretation.