Latent Koopman Dynamics of Brain Structure-Function Coupling for Identifying Adolescent Prenatal Drug Exposure

Area of Science:

• Neuroscience
• Computational Psychiatry
• Developmental Neuroscience

Background:

• Prenatal drug exposure (PDE) is linked to lasting changes in adolescent brain development.
• Understanding how structural and functional brain networks interact is crucial for identifying neurodevelopmental vulnerabilities.
• Current methods often analyze brain connectivity statically or treat different data types independently, limiting insights into structure-function relationships.

Purpose of the Study:

• To develop a novel computational framework, NeuroKoop++, for analyzing brain structure-function coupling.
• To investigate the mechanisms underlying neurodevelopmental vulnerability associated with PDE.
• To identify reliable neuroimaging biomarkers for PDE.

Main Methods:

• Developed NeuroKoop++, a graph neural network (GNN)-based multimodal framework.
• Utilized GNN encoders to extract structural connectivity (SC) and functional network connectivity (FNC) representations.
• Integrated SC and FNC using bidirectional cross-attention and a spectrally constrained Koopman operator, conditioned on cognitive scores.

Main Results:

• Applied NeuroKoop++ to the Adolescent Brain Cognitive Development (ABCD) cohort (10,199 adolescents).
• NeuroKoop++ significantly outperformed existing state-of-the-art multimodal approaches.
• Identified interpretable signatures of altered brain network organization in adolescents with PDE history.

Conclusions:

• Modeling structure-function coupling as a dynamic, cognition-modulated process enhances PDE classification.
• NeuroKoop++ provides a robust computational method for identifying neurodevelopmental biomarkers of PDE.
• This framework has broad applications for multimodal neuroimaging in pediatric and adolescent brain health research.