A computational framework for fitting biophysical basal-ganglia network models, applied to Parkinsonian beta...

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning in Neuroscience

Background:

• Fitting detailed spiking neural network models to experimental data is computationally intensive, limiting large-scale parameter searches.
• Previous methods relied on manual tuning, simplified models, or small network sizes, hindering comprehensive analysis.
• Automated fitting of biophysically detailed models was computationally intractable on conventional hardware.

Purpose of the Study:

• To develop an integrated computational pipeline for automated, data-driven fitting of biophysically detailed spiking neural network models.
• To make parameter searches for large-scale neural circuits tractable on a single Graphics Processing Unit (GPU).
• To enable routine parameter identifiability analysis for complex neural models.

Main Methods:

• A pipeline was created coupling a just-in-time (JIT) compiled spiking neural network model (subthalamic nucleus, globus pallidus externa, globus pallidus interna) in JAX with covariance-matrix-adaptation evolution strategy (CMA-ES) optimization.
• A fixed-indegree connectivity scheme was employed to ensure fitted configurations are transferable across different network sizes.
• The pipeline was tested on an NVIDIA L4 GPU, utilizing JIT compilation and kernel fusion for acceleration.

Main Results:

• JIT compilation and kernel fusion accelerated simulations by ~736-fold compared to un-jitted Python, with a further 2-fold speedup on GPU versus CPU.
• A 1000-trial optimization completed in approximately 17 minutes, with fitted configurations scaling across a 100-fold range of network size.
• The pipeline recovered a parkinsonian basal ganglia configuration with elevated beta-band power (~29 Hz) in the subthalamic nucleus, consistent with MPTP-primate models.

Conclusions:

• The developed pipeline offers an accessible and efficient tool for fitting biophysically detailed neural network models.
• Automated fitting and parameter identifiability analysis are now feasible on a single cloud GPU.
• The study provides a proof-of-concept for applying the pipeline to basal ganglia circuit modeling, revealing specific weight changes associated with a parkinsonian state.