Bayesian Optimisation of Large-Scale Biophysical Networks.

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Systems neuroscience

Background:

• The human brain's structure-function relationship is known but not fully understood.
• Large-scale biophysical models integrate structural data (e.g., diffusion tractography) with neuronal dynamics to study brain networks.
• Current models face challenges in parameterization, simulation cost, and reliable comparison.

Purpose of the Study:

• To introduce Bayesian optimization as a method for assessing biophysical brain network models.
• To provide a principled framework for incremental modeling and model comparison.
• To efficiently navigate complex parameter spaces and identify regions of high functional similarity.

Main Methods:

• Adaptation of a Bayesian optimization technique for costly, high-dimensional, non-convex problems.
• Application to a biophysical network model with five key parameters.
• Comparison of different structural connectivity estimation methods from diffusion tractography.
• Validation against real magnetoencephalography (MEG) data for functional connectivity.

Main Results:

• Bayesian optimization converged to regions of high functional similarity with MEG data using few samples.
• The method effectively explored the parameter space without getting trapped in local extrema.
• A map of uncertainty was built and exploited across the parameter space.
• One structural connectivity estimation method demonstrated superior simulation results.

Conclusions:

• Bayesian optimization is an effective tool for parameterizing and comparing complex biophysical brain network models.
• This approach facilitates the experimental aspect of computational modeling.
• The findings suggest specific structural connectivity methods yield more accurate brain simulations.