Graph auto-encoding brain networks with applications to analyzing large-scale brain imaging datasets.

Area of Science:

• Neuroimaging
• Network Science
• Computational Neuroscience

Background:

• Human brain connectomes, represented as networks, are crucial for understanding cognition.
• Analyzing high-dimensional, non-Euclidean brain network data and its relation to traits is challenging.
• Existing methods often rely on simplified network summaries like topological features or PCA.

Purpose of the Study:

• To develop a novel deep learning model for characterizing brain graph population distributions.
• To infer relationships between brain connectomes and human cognitive traits.
• To introduce Graph AuTo-Encoding (GATE) as an advanced analytical tool.

Main Methods:

• Developed a nonlinear latent factor model named Graph AuTo-Encoding (GATE).
• Applied GATE to large-scale datasets: Adolescent Brain Cognitive Development (ABCD) and Human Connectome Project (HCP).
• Investigated structural brain connectomes and their association with cognitive functions.

Main Results:

• GATE demonstrated superior prediction accuracy, statistical inference, and computational efficiency compared to existing methods.
• Identified a stronger association between structural connectomes and diverse cognitive traits.
• Highlighted the effectiveness of deep learning for brain network analysis.

Conclusions:

• GATE provides a powerful nonlinear approach for modeling brain graph distributions and their trait associations.
• The findings underscore the significant role of the structural connectome in cognitive abilities.
• This method advances the analysis of complex brain network data in neuroscience research.