Fast Multi-Task SCCA Learning with Feature Selection for Multi-Modal Brain Imaging Genetics.

Area of Science:

• Neuroscience
• Genetics
• Machine Learning

Background:

• Brain imaging genetics integrates genetic data (SNPs) with brain imaging quantitative traits (QTs).
• Multi-task learning (MTL) and sparse canonical correlation analysis (SCCA) are common but have limitations with large datasets and multiple data types.
• Existing methods struggle with feature selection and computational challenges as the number of SNPs increases.

Purpose of the Study:

• To propose a novel Multi-Task Sparse Canonical Correlation Analysis (MTSCCA) framework.
• To identify bi-multivariate associations between SNPs and multi-modal imaging QTs.
• To leverage complementary information from different imaging modalities.

Main Methods:

• Developed MTSCCA, combining MTL and SCCA.
• Utilized G(2,1)-norm regularization for group-level sparsity on SNPs.
• Employed an L1-norm penalty for joint feature selection across multiple tasks and modalities.
• Proposed a fast optimization algorithm incorporating SNP grouping information.

Main Results:

• MTSCCA demonstrates improved performance in correlation coefficients and canonical weights patterns compared to conventional SCCA.
• The proposed method effectively handles multi-modal imaging QTs and a large number of SNPs.
• Achieved enhanced feature selection across multiple tasks and modalities.

Conclusions:

• MTSCCA offers a powerful and efficient tool for genome-wide and brain-wide imaging genetic studies.
• The method is fast, easy-to-implement, and provides superior results.
• It effectively integrates multi-modal imaging data with genetic information for comprehensive analysis.