Automating region selection with genetic algorithms for energy landscape analyses of brain dynamics.

Abstract

Understanding brain dynamics is essential in cognitive neuroscience. Energy landscape analysis (ELA), which characterizes brain activity using a pairwise maximum entropy model, is a powerful tool for analyzing these dynamics but has traditionally relied on the subjective, manual selection of small regions of interest (ROIs) to satisfy mathematical constraints. We developed ELA/GAopt, a framework that uses genetic algorithms to automate ROI selection from whole-brain atlases in a data-driven manner. ELA/GAopt's consistent identification of reproducible ROI subsets was validated across three independent datasets. In multi-site clinical data, the framework identified and replicated autism-specific dynamics characterized by global co-activation within sensory-motor and visual networks. These results demonstrate that ELA/GAopt provides a systematic, objective approach for characterizing condition-specific brain dynamics, thereby establishing a methodological foundation for future externally validated biomarker studies and the systematic exploration of brain state transitions.