Identification and discovery of imaging genetic patterns using fusion self-expressive network in major depressive disorder

Mengqian Pang, Xiaoyun Liu, Xiaoke Hao, Meiling Wang, Chunming Xie, Li Zhang, Yonggui YuanFrontiers in Neuroscience 2023 · DOI

Abstract

In this study, we discussed the discriminant and convergence performance of the fusion self-expressive network, parameters, and atlas selection. We developed a novel "Fusion Self-Expressive Network" (FSN-MM) that integrates structural and functional brain imaging to identify genetic markers for Major Depressive Disorder (MDD). By leveraging diagnosis-guided data fusion, the model significantly improves the correlation between specific genetic variations (SNPs) and brain patterns, uncovering 15 new potential risk genes for the disorder.

Introduction: The Challenge of Imaging Genetics

Imaging genetics is a high-stakes "needle in a haystack" problem. Researchers attempt to find correlations between millions of genetic variations—specifically Single Nucleotide Polymorphisms (SNPs)—and complex brain features derived from neuroimaging. Major Depressive Disorder (MDD) research faces two primary hurdles:

  1. Modality Isolation: Most existing research analyzes brain structure (sMRI) or brain function (fMRI) in isolation, ignoring their deep interconnections.
  2. Structural Complexity: Brain imaging data is high-dimensional and noisy. Standard methods often fail to capture the underlying "subspaces" or clusters that distinguish a depressed brain from a healthy one.

Methods: The FSN-MM Framework

The proposed FSN-MM (Fusion Self-Expressive Network for Multi-Modality) model addresses these challenges through a three-stage workflow:

  • Feature Extraction: sMRI is used to derive "node" features (Gray Matter Volume), while rs-fMRI provides "edge" features (Functional Connectivity) across 116 brain regions defined by the AAL atlas.
  • Self-Expressive Fusion: This core component builds "self-expressive networks" where each subject's data is represented as a sparse combination of other subjects within the same diagnostic group (Healthy Control, Moderate MDD, or Severe MDD). These networks are iteratively fused, guided by diagnosis information to enhance disease-specific signals.
  • Association Analysis: An association model with $L_{2,1}$-norm regularization is applied to select a sparse, consistent set of brain regions across both structural and functional data.
Overview of the FSN-MM model architecture, illustrating feature extraction, fusion, and association analysis.

Key Insights: Diagnosis-Guided Data Cleaning

A critical innovation of FSN-MM is diagnosis-guided fusion. By incorporating diagnostic group information during the fusion process, the model effectively "cleans" the data. The fused network demonstrates tighter clustering within groups and superior separation between healthy and depressed subjects compared to raw data distributions.

The process of data self-reconstruction by the self-expressive method, showing improved group clustering.

Furthermore, the model identified consistent brain regions—such as the left temporal, right hippocampus, and right precuneus—as significant across both structural and functional modalities. This consistency is often lacking in traditional single-modality models. Results also suggest that functional connectivity features might hold more predictive power than structural volume alone.

Results and Discovery

The study utilized a dataset of 171 participants and 5,897 SNPs. The FSN-MM model achieved a Correlation Coefficient (CC) of 0.3845 for node features and 0.4037 for edge features when validated against the known risk SNP TPH1 rs1799913. This significantly outperformed standard single-modality methods and existing multi-modal algorithms like MSD-MM.

Beyond validation, the model discovered 15 new potential risk SNPs associated with MDD. Six of these, including CACNA1C rs11832738, are supported by existing psychiatric literature, while the remaining nine represent novel targets for future biological validation.

Conclusion

The FSN-MM framework provides a powerful mathematical approach to viewing the brain as a unified, genetically-influenced system. While limited by sample size and data availability, the integration of multi-modal imaging through self-expressive networks marks a significant step forward in identifying robust biological markers for Major Depressive Disorder. Future work may incorporate Diffusion Tensor Imaging (DTI) to further refine the "depressed connectome."

Summary updated

Abstract

In this study, we discussed the discriminant and convergence performance of the fusion self-expressive network, parameters, and atlas selection.
BibTeX
@inproceedings{pang2023identification,
  author = {Mengqian Pang and Xiaoyun Liu and Xiaoke Hao and Meiling Wang and Chunming Xie and Li Zhang and Yonggui Yuan},
  title = {{Identification and discovery of imaging genetic patterns using fusion self-expressive network in major depressive disorder}},
  booktitle = {Frontiers in Neuroscience},
  year = {2023},
  doi = {10.3389/fnins.2023.1297155},
}
Figure