A Bayesian Group Sparse Canonical Correlation Analysis Method for Brain Imaging Genomics

Alzheimer’s disease (AD) is a highly heritable brain disorder, and thus accurate identification of risk factors could help understand its pathogenesis and prompt early diagnosis and treatment. Sparse canonical correlation analysis (SCCA) has become one of the common auxiliary techniques for identifying risk genetic variations. Most relevant models provide point estimation for canonical weights (effective size of genetic variations), assuming these effects to be fixed. Based on the Bayesian theory, we developed an interval estimation method, called the Bayesian Group SCCA method. This approach could yield interval estimation for Lasso and Group Lasso penalties. A Gibbs sampler was utilized to optimize posterior inference on the model parameters. Experimental results demonstrated that, at a certain confidence level, the interval range obtained from this method matches point estimation outcomes. Further analysis indicated that the interval estimation results can capture potentially important features that were overlooked by point estimation, thus enhancing the model’s accuracy for brain imaging genomics.