Evaluating the effects of high-throughput structural neuroimaging predictors on whole-brain functional connectome outcomes via network-based matrix-on-vector regression

Lu Tong, Yuan Zhang, Vince Lyzinski, Chuan Bi, Peter Kochunov, Elliot Hong, Shuo ChenBiometrics 2025 · DOI

TL;DR

This paper introduces MOAT, a novel method for analyzing the complex relationships between brain structure and function using neuroimaging data. MOAT effectively identifies specific structural features that systematically influence functional brain networks, overcoming limitations of previous approaches and revealing significant associations in a large UK Biobank dataset.

What Problem Does This Paper Address?

Understanding how the physical structure of the brain (structural imaging, SI) influences its functional activity (functional connectivity, FC) is a central goal in neuroscience. However, analyzing the vast amount of data generated by modern neuroimaging techniques presents significant challenges. Specifically, researchers face:

  • High Dimensionality: Millions of potential associations between structural features (like white matter integrity and cortical thickness) and functional connections across the entire brain.
  • Complex Network Structures: Both structural and functional data exhibit intricate network patterns, meaning relationships are not isolated but organized within specific brain regions or pathways.
  • False Positives: Traditional statistical methods often struggle to distinguish true associations from random noise, leading to an overwhelming number of false positive findings.
  • Methodological Gaps: Existing methods often treat structural and functional data separately or fail to adequately capture the systematic, network-level influences of structure on function.

This paper aims to bridge these gaps by developing a method that can jointly analyze multimodal neuroimaging data, identify specific structural predictors that systematically impact functional networks, and control for false positives.

Methods

The core of this paper is the Multi-level Network Association (MOAT) method. MOAT operates on a novel multi-level graph structure that captures both structural-functional associations and the inherent network organization of the brain.

  • Multi-level Graph Representation:
    • Level 1 (Bipartite Graph): This graph represents the direct associations between individual structural imaging (SI) measures (as predictors) and functional connectivity (FC) measures (as outcomes). Edges in this graph signify a potential association.
  • Level 2 (Unipartite Graph): This graph represents the functional connectome itself, where nodes are brain regions and edges are functional connections. This level captures the network structure of the FC data.
  • Dense Subgraph Extraction: MOAT employs a novel greedy peeling algorithm to identify "doubly-dense" multi-level subgraphs. These subgraphs are characterized by a high density of associations within a specific set of SI predictors and a corresponding set of FC connections that form a dense clique (a highly interconnected sub-network) in the functional connectome. This approach is designed to find systematic patterns of influence, rather than scattered associations.
  • Network-Level Inference: After extracting potential sub-networks, MOAT uses a tailored network-level inference framework to assess their statistical significance. This involves a novel test statistic that accounts for both the density and size of the identified multi-level subgraphs, allowing for FWER (Family-Wise Error Rate) controlled hypothesis testing.
  • Compatibility: MOAT is designed to be compatible with existing multivariate analysis tools like Canonical Correlation Analysis (CCA), allowing for further quantitative measurement of the identified associations.

Datasets

The primary dataset used in this study is from the UK Biobank.

  • UK Biobank Data: This study utilized data from 4,242 participants who had complete structural imaging (SI) and resting-state functional MRI (rs-fMRI) data.
  • SI Measures: 105 measures were collected, including 39 white matter integrity measures (assessed by fractional anisotropy, FA) and 66 cortical thickness measures.
  • FC Measures: 30,135 FC measures were derived from rs-fMRI data using the Brainnetome Atlas, representing functional connections between 246 brain regions.
  • Confounding Variables: Age, sex, educational level, and body mass index (BMI) were included as covariates.

The paper also includes a simulation study to evaluate MOAT's performance under controlled conditions. This simulation generated synthetic FC and SI data with predefined network structures and varying effect sizes, allowing for direct comparison with other methods.

Key Insights and Ideas

  • Structure-Function Link is Highly Localized: The study strongly suggests that the influence of structural brain features on functional connectivity is not diffuse across the entire brain but is concentrated within specific, organized sub-networks. This challenges the idea of a purely whole-brain diffuse pattern.
  • White Matter Integrity is Key: White matter integrity measures, particularly those in corticospinal tracts and the inferior cerebellar peduncle, were found to have a significant impact on functional connections within sensorimotor, salience, and executive networks.
  • MOAT's Advantage in Sub-network Identification: The multi-level graph structure and dense subgraph extraction are crucial for identifying these localized, systematic associations. This approach effectively filters out noise and reduces false positives compared to methods that analyze associations in isolation.
  • Biological Plausibility: The identified sub-networks and their associated structural predictors align with known neuroanatomical pathways and cognitive functions, lending strong biological interpretability to the findings.

Key Results

  • Significant FC-SI Associations: MOAT identified a statistically significant systematic association pattern (p < 0.0001) in the UK Biobank data.
  • Identified Sub-network: The primary identified sub-network (B₁) comprised 23 SI measures (3 cortical thickness, 20 FA) and 1316 FC outcomes. This sub-network corresponded to a dense clique of 79 brain regions.
  • Stronger Associations within Sub-network: FC-SI pairs within the identified sub-network showed significantly stronger associations (average correlation of 0.81) compared to those outside the network.
  • Performance in Simulations: In simulation studies, MOAT outperformed other biclustering and sparse Canonical Correlation Analysis (sCCA) methods in accurately identifying true sub-networks and controlling false positives, especially when systematic association patterns were present.

Limitations and Future Work

  • UK Biobank Sample Characteristics: The authors acknowledge that the UK Biobank participants are predominantly older Caucasians, which may limit the generalizability of the findings to other demographic groups.
  • Further Investigation: The authors suggest that further investigation and integrated analysis are needed to gain a more comprehensive understanding of FC-SI associations.
  • Future Directions: The paper highlights the potential of MOAT to guide future cognitive and psychiatric therapy by enhancing our understanding of how brain structure and function interact. The developed software package for MOAT is made publicly available.

Summary updated

Abstract

The joint analysis of multimodal neuroimaging data is vital in brain research, revealing complex interactions between brain structures and functions. Our study is motivated by the analysis of a vast dataset of brain functional connectivity (FC) and multimodal structural imaging (SI) features from the UK Biobank. Specifically, we aim to investigate the effects of SI features, such as white matter microstructure integrity (WMMI) and cortical thickness, on the whole-brain functional connectome network. This analysis is inherently challenging due to the extensive structural-functional associations and the intricate network patterns present in multimodal high-dimensional neuroimaging data. To bridge methodological gaps, we developed a novel multi-level sub-graph extraction method (dense bipartite with nested unipartite graph) within a matrix(network)-on-vector regression model. This method identifies subsets of spatially specific SI features that intensely and systematically influence FC sub-networks, while effectively suppressing false positives in large-scale datasets. Applying our method to a multimodal neuroimaging dataset of 4242 participants ffrom the UK Biobank, we evaluated the effects of whole-brain WMMI and cortical thickness on resting-state FC. Our findings indicate that the WMMI in corticospinal tracts and inferior cerebellar peduncle significantly affect functional connections of sensorimotor, salience, and executive sub-networks, with an average correlation of 0.81 ($p < 0.001$).
BibTeX
@inproceedings{tong2025evaluating,
  author = {Lu Tong and Yuan Zhang and Vince Lyzinski and Chuan Bi and Peter Kochunov and Elliot Hong and Shuo Chen},
  title = {{Evaluating the effects of high-throughput structural neuroimaging predictors on whole-brain functional connectome outcomes via network-based matrix-on-vector regression}},
  booktitle = {Biometrics},
  year = {2025},
  doi = {10.1093/biomtc/ujaf027},
}
Figure