[Paper] Evolution With Purpose: Hierarchy-Informed Optimization of Whole-Brain Models

Source: arXiv - 2602.11398v1

Overview

This paper shows how injecting biological knowledge—specifically the hierarchical organization of brain networks—into evolutionary optimization can produce whole‑brain models that not only fit MRI data but also generalize across subjects and predict behavior. By comparing a uniform parameter model with a region‑specific one, the authors demonstrate that a curriculum‑guided evolution (HICO) yields more robust and useful models.

Key Contributions

• Hierarchy‑informed evolutionary search (HICO): A novel curriculum that respects the known ordering of brain networks during parameter optimization.
• Heterogeneous whole‑brain DMF modeling: Extends the Dynamic Mean Field model from a single global parameter set to region‑specific parameter groups (seven canonical brain regions).
• Systematic comparison of four optimization strategies: (i) simultaneous optimization, (ii) HICO (forward curriculum), (iii) reversed curriculum, and (iv) random curriculum.
• Demonstrated generalization: Only the forward curriculum (HICO) maintained performance on unseen subjects and could predict individual behavioral scores.
• Evidence that domain knowledge can act as a regularizer: Guiding evolution reduces over‑fitting despite the high dimensional, non‑convex search space.

Methodology

1. Base model – Dynamic Mean Field (DMF): A biophysically plausible whole‑brain simulator that predicts BOLD signals from structural connectivity. The baseline uses 20 shared parameters for the entire brain.
2. Heterogeneous extension: The same 20 parameters are duplicated for each of the seven canonical brain regions (e.g., visual, somatomotor, default‑mode), yielding 140 parameters that can vary across regions.
3. Evolutionary optimization: A population‑based genetic algorithm evolves parameter sets to minimize the discrepancy between simulated and empirical fMRI functional connectivity.
4. Curriculum designs:
   • All‑at‑once: Evolve all 140 parameters simultaneously.
   • HICO (forward curriculum): Start by optimizing parameters for low‑level sensory regions, then progressively add higher‑order regions following the known hierarchy.
   • Reversed curriculum: Same as HICO but start from the highest‑order regions.
   • Random curriculum: Add regions in a random order.
5. Evaluation: Models are tested on held‑out subjects for (a) fit quality (functional connectivity correlation) and (b) ability to predict behavioral scores (e.g., working‑memory, attention) using the learned parameters as features.

Results & Findings

• Fit quality: All four heterogeneous strategies achieved comparable high correlations with empirical functional connectivity (≈0.85), confirming that the added flexibility does not hurt basic model fitting.
• Generalization: When applied to new subjects, only the forward curriculum (HICO) retained high performance; the other strategies showed a noticeable drop, indicating over‑fitting.
• Behavioral prediction: Parameter vectors from HICO explained a significant portion of variance in subjects’ cognitive scores (R² ≈ 0.30), whereas the other strategies failed to produce meaningful predictions.
• Reversed vs. random curricula: Both performed worse than HICO, suggesting that the order of introducing region‑specific complexity matters—aligning it with the brain’s functional hierarchy is beneficial.

Practical Implications

• Better brain‑computer interfaces (BCIs): More reliable whole‑brain models can improve decoding of mental states for adaptive BCI systems.
• Personalized neurotechnology: Parameter sets that predict behavior open the door to tailoring neurostimulation or neurofeedback protocols to individual cognitive profiles.
• Model‑driven drug discovery: Robust simulations of whole‑brain dynamics can serve as in‑silico testbeds for assessing the impact of pharmacological interventions on network‑level activity.
• Generalizable AI optimization: The curriculum‑guided evolutionary approach is a template for other high‑dimensional, gradient‑free problems (e.g., hardware design, hyper‑parameter tuning) where domain hierarchies exist.
• Reduced computational waste: By structuring the search, HICO converges faster and avoids the massive over‑parameterization that would otherwise require extensive compute resources.

Limitations & Future Work

• Scalability to finer parcellations: The study used seven canonical regions; extending to more granular atlases may re‑introduce over‑fitting challenges.
• Dependence on a predefined hierarchy: The approach assumes the hierarchy is correct; alternative or subject‑specific hierarchies could further improve results.
• Evolutionary algorithm specifics: The paper does not explore alternative meta‑heuristics (e.g., CMA‑ES, Bayesian optimization) that might be more sample‑efficient.
• Behavioral scope: Only a limited set of cognitive scores were examined; broader phenotypic predictions (e.g., psychiatric risk) remain to be tested.
• Real‑time applicability: Current DMF simulations are offline; integrating HICO‑optimized models into real‑time systems will require speed‑up techniques (e.g., surrogate modeling).

Bottom line: By marrying evolutionary search with neuroscientific hierarchy, the authors demonstrate a practical pathway to more generalizable, behavior‑predictive whole‑brain models—an advance that resonates well beyond neuroscience into any domain where structured prior knowledge can guide complex optimization.*

Authors

• Hormoz Shahrzad
• Niharika Gajawell
• Kaitlin Maile
• Manish Saggar
• Risto Miikkulainen

Paper Information

• arXiv ID: 2602.11398v1
• Categories: cs.NE
• Published: February 11, 2026
• PDF: Download PDF