cuBNM: GPU-Accelerated Brain Network Modeling.

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Biophysics

Background:

• Brain network modeling infers neural properties using simulations fitted to empirical data.
• High computational costs limit the application of these models to large cohorts and complex scenarios.
• Scalable simulation methods are crucial for advancing brain network modeling.

Purpose of the Study:

• Introduce cuBNM, a Python package for accelerating brain network model simulations using GPU parallel processing.
• Demonstrate the significant speedup achieved by cuBNM compared to CPU-based simulations.
• Showcase cuBNM's utility in optimizing group-level and individualized brain network models.

Main Methods:

• Leveraged graphics processing units (GPUs) for massively parallel simulations of brain network models.
• Implemented cuBNM as a Python package for user-friendly access to GPU acceleration.
• Applied cuBNM to optimize low- and high-dimensional models for group and individual analyses.

Main Results:

• cuBNM achieves several hundred times speedup in simulations compared to central processing units (CPUs).
• Demonstrated successful optimization of both group-level and individualized brain network models.
• Investigated test-retest reliability and heritability of simulated and empirical measures using the Human Connectome Project dataset.

Conclusions:

• cuBNM offers a highly efficient framework for large-scale brain network modeling.
• Accelerated simulations facilitate investigations across larger cohorts, denser networks, and more complex models.
• Simulated features demonstrate good reliability and heritability, supporting the utility of individualized models.