Energy Landscape Analysis with Automated Region-of-Interest Selection via Genetic Algorithms

Abstract

Understanding brain dynamics is essential for advancing cognitive and clinical neuroscience. Energy landscape analysis (ELA), based on the pairwise maximum entropy model, is a powerful framework to characterize brain activity as transitions among discrete states defined by regional activity patterns. However, traditional ELA relies on the subjective manual selection of a small subset of regions of interest (ROIs) from whole-brain parcellations to satisfy mathematical constraints, which limits the scope and reproducibility of ELA results due to subjective subset selection. To overcome this, we developed ELA/GAopt, a meta-framework that utilizes a genetic algorithm to automate the selection of ROI combinations from the entire atlas search space by optimizing a user-defined objective function. In this study, we implemented a representative objective function balancing model fitting accuracy with the inter-individual variability of model parameters. We applied ELA/GAopt to three independent resting-state functional magnetic resonance imaging datasets. In Scenario 1, using the Creativity dataset (OpenNeuro: ds002330, n = 61), the ROI sets identified by ELA/GAopt achieved significantly higher objective function values and pattern reproducibility than randomly selected ROI sets (p < 0.05). Additional validation with the large-scale Human Connectome Project Young Adult (HCP-YA) dataset (n = 270) confirmed the robustness of our framework in high-dimensional settings. Stability analysis using Jaccard and Hamming metrics demonstrated that ELA/GAopt consistently identified reproducible ROI subsets across independent optimization runs. In Scenarios 2-4, we analyzed data from the Autism Brain Imaging Data Exchange II dataset using a site-disjoint validation design to mitigate findings were robust against multi-site artifacts. ELA/GAopt identified ASD-specific dynamics, where participants tend to visit local minima characterized by global co-activation of selected ROIs within sensory-motor and visual networks. These signatures were replicated in an independent cohort consisting of different scanning sites. Furthermore, ROI sets optimized for one group (ASD or typically developing controls) did not generalize to the other, highlighting distinct neurodynamic architectures. These results demonstrate that ELA/GAopt provides a reproducible, data-driven pathway for characterizing condition-specific brain dynamics, serving as a methodological basis for future, harmonization-aware and externally validated biomarker studies.