Abstract: The development of neuroimaging technology and molecular genetics has produced a large number of imaging genetic data,which has greatly promoted the study of complex mental diseases.However,because the dimensions of the data are too high and the correlation measure is based on the assumption that data obey Gaussian distribution,traditionalalgorithms often fail to explain the dependencies between two types of data.In order to solve the shortcomings of traditional algorithms,this paper proposed a method for correlation analysis of a large number of SNP and fMRI data.This method guides fused lasso to perform feature selection by constructing a network structure of features,and uses higher-order statistics to extract statistically significant variables.Thus,biomarkers associated with mental illness are identified.The experimental results show that the distribution of typical vector values obtained by the algorithm in simulation data are almost consistent with the real data,and the correlation coefficient obtained is the closest to the correlation coefficient in the real dataset.The average correlation coefficient of the proposed algorithm is up to 81%,which is about 20% higher than L1-SCCA and about 3% higher than FL-SCCA.Compared with the other two algorithms in real data,the proposed algorithm can find more genes and brain regions that have potential effects on schizophrenia.The experimental results show that the proposed algorithm can effectively identify risk genes and abnormal brain regions within a reasonable time.

Key words: Correlation analysis, Feature selection, Higher-order statistics, Image genetics, Sparse representation