inMTSCCA: An Integrated Multi-task Sparse Canonical Correlation Analysis for Multi-omics Brain Imaging Genetics

Abstract

Identifying genetic risk factors for Alzheimers disease (AD) is an important research topic. To date, different endophenotypes such as imaging-derived endophenotypes and proteomic expression-derived endophenotypes have shown the great value in uncovering risk genes compared to case-control studies. Biologically, a co-varying pattern of these different omics derived endophenotypes could result from the shared genetic basis. However, existing methods mainly focus on the effect of endophenotypes, and that of cross-endophenotype associations remains largely unexploited. In this paper, we used both endophenotypes and their cross-associations of multi-omics to identify genetic risk factors, and proposed two integrated multi-task sparse canonical correlation analysis (MTSCCA) methods, i.e., pairwise endophenotype correlation guided MTSCCA (pcMTSCCA) and high-order endophenotype correlation guided MTSCCA (hocMTSCCA). pcMTSCCA employed pairwise correlations between MRI-derived, plasma-derived, and cerebrospinal fluid (CSF) derived endophenotypes as an additional penalty. hocMTSCCA used high-order correlations among these multi-omics for regularization. To figure out genetic risk factors at individual and group levels, as well as altered endophenotypic markers, we introduced sparsity-inducing penalties in both models. We compared pcMTSCCA and hocMTSCCA with three related methods on both simulation data and real neuroimaging, proteomic analytes, and genetic data. The results showed that our methods obtained better or comparable canonical correlation coefficients (CCCs) and feature subsets than benchmarks. Most importantly, the identified genetic loci and heterogeneous endophenotypic markers showed high relevance. Therefore, jointly using multi-omics endophenotypes and their cross-endophenotype associations are promising to reveal genetic risk factors, and both methods are qualified for this complicated task. The source code and manual of inMTSCCA is available at: https://ngdc.cncb.ac.cn/biocode/tools/BT007330.