SPSGL: uncovering psychiatric network mechanisms via structural-prior guided synaptic graph learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Functional magnetic resonance imaging (fMRI) is vital for studying brain functional connectivity (FC) in psychiatric disorders.
• Static FC models struggle to capture the dynamic nature of brain plasticity and often overlook network structural differences or weak connections.

Purpose of the Study:

• To develop a novel deep learning framework, SPSGL, for constructing advanced brain connectivity patterns from fMRI signals.
• To improve the modeling of dynamic FC and identify neurobiological meaningful connections for psychiatric research.

Main Methods:

• SPSGL transforms fMRI time series into frequency-domain functional brain graphs.
• It utilizes a biologically inspired gated edge-update mechanism and multi-head attention guided by structural priors.
• The framework integrates Orthonormal Clustering Readout (OCRead) for adaptive multi-scale graph representation and functional parcellation.

Main Results:

• SPSGL demonstrated superior performance across five psychiatry-related computational tasks compared to existing methods.
• The model identified task-relevant functional connections and hub regions linked to aberrant coupling in default mode, sensorimotor, and subcortical networks.
• It uncovered shared brain network factors across diverse psychiatric conditions, suggesting potential neuroimaging biomarkers.

Conclusions:

• SPSGL offers a unified, interpretable, and high-performing framework for fMRI-based brain connectivity analysis.
• The framework advances mechanistic understanding and holds potential for clinical translation in mental health research.
• Publicly available code facilitates further research and application.