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

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 approaches require a predefined set of regions of interest (ROIs), usually chosen based on prior knowledge from hypothesis-driven studies, which introduces subjectivity and may overlook previously unknown patterns. To overcome this limitation, we developed a new ELA framework enhanced by automated ROI selection using genetic algorithm optimization, called ELA/GAopt. ELA/GAopt employs a genetic algorithm to select ROI subsets that maximize model fitting accuracy while accounting for individual variability. We applied ELA/GAopt to two independent resting-state functional magnetic resonance imaging datasets. In Scenario 1, we tested how well the method generalizes using the Creativity dataset (OpenNeuro: ds002330), which includes 66 healthy adults (29 males, 36 females, and one non-binary individual; average age 26.6 ± 4.3 years). ELA/GAopt was performed on a discovery set ( n = 31), and the optimized ROI sets were validated on a test set ( n = 30). Over 100 independent optimization runs, the selected ROI sets achieved significantly higher objective function values than randomly chosen ROI sets (permutation test, p < 0.05, false discovery-rate corrected). The similarity of local minimum states between discovery and test sets was also significantly higher for ELA/GAopt than for random selection ( p < 0.05). These results confirm the reproducibility and generalizability of the proposed approach. In Scenarios 2–4, we analyzed data from 293 participants in the Autism Brain Imaging Data Exchange II dataset, including 157 with autism spectrum disorder (ASD) and 136 typically developing controls (CTL) from discovery cohort A and validated findings in an independent cohort (54 ASD and 72 CTL participants). ELA/GAopt showed that ASD participants tend to visit local minima characterized by global co-activation of selected ROIs, especially within sensory-motor and visual networks. Additionally, ROI sets optimized for one group (ASD or CTL) did not generalize across groups, with significant differences in the number and structure of local minima (Mann–Whitney U test, p < 0.05, false discovery-rate corrected). These findings demonstrate that ELA/GAopt offers a robust, data-driven approach to modeling brain state dynamics, minimizing biases from predefined ROI selection. It improves reproducibility and could aid in discovering neuroimaging biomarkers for clinical conditions.