Kevin J. Wischnewski

Brain network modeling uses computer simulations to infer about latent neural properties at micro- and mesoscales by fitting brain dynamic models to empirical data of individual subjects or groups. However, computational costs of (individualized) model fitting is a major bottleneck, limiting the practical feasibility of this approach to larger cohorts and more complex models, and highlighting the need for scalable simulation implementations. Here, we introduce cuBNM, a Python package which leverages parallel processing of graphics processing units to massively accelerate simulations of brain network models. We show running simulations on graphics processing units is several hundred times faster compared to central processing units. We demonstrate the usage of cuBNM by running optimization of group-level and individualized low- and high-dimensional models. As examples of the utility of individualized models, we investigated test-retest reliability and heritability of simulated and empirical measures in the Human Connectome Project dataset. We found simulated features were fairly reliable and significantly heritable, suggesting their biological plausibility. Overall, cuBNM enables large-scale simulations of brain network models, opening new avenues for studying latent neural processes across diverse populations, dense networks, and high-dimensional models, which was previously impractical due to computational constraints.

The excitation-inhibition ratio is a key functional property of cortical microcircuits which changes throughout an individual’s lifespan. Adolescence is considered a critical period for maturation of excitation-inhibition ratio. This has primarily been observed in animal studies. However, there is limited human in vivo evidence for maturation of excitation-inhibition ratio at the individual level. Here, we developed an individualized in vivo marker of regional excitation-inhibition ratio in human adolescents, estimated using large-scale simulations of biophysical network models fitted to resting-state functional imaging data from both cross-sectional ( n = 752) and longitudinal ( n = 149) cohorts. In both datasets, we found a widespread decrease in excitation-inhibition ratio in association areas, paralleled by an increase or lack of change in sensorimotor areas. This developmental pattern was aligned with multiscale markers of sensorimotor-association differentiation. Although our main findings were robust across alternative modeling configurations, we observed local variations, highlighting the importance of methodological choices for future studies.

Personalized modeling of the resting-state brain activity implies the usage of dynamical whole-brain models with high-dimensional model parameter spaces. However, the practical benefits and mathematical challenges originating from such approaches have not been thoroughly documented, leaving the question of the value and utility of high-dimensional approaches unanswered. Studying a whole-brain model of coupled phase oscillators, we proceeded from low-dimensional scenarios featuring 2–3 global model parameters only to high-dimensional cases, where we additionally equipped every brain region with a specific local model parameter. To enable the parameter optimizations for the high-dimensional model fitting to empirical data, we applied two dedicated mathematical optimization algorithms (Bayesian Optimization, Covariance Matrix Adaptation Evolution Strategy). We thereby optimized up to 103 parameters simultaneously with the aim to maximize the correlation between simulated and empirical functional connectivity separately for 272 subjects. The obtained model parameters demonstrated increased variability within subjects and reduced reliability across repeated optimization runs in high-dimensional spaces. Nevertheless, the quality of the model validation (goodness-of-fit, GoF) improved considerably and remained very stable and reliable together with the simulated functional connectivity. Applying the modeling results to phenotypical data, we found significantly higher prediction accuracies for sex classification when the GoF or coupling parameter values optimized in the high-dimensional spaces were considered as features. Our results elucidate the model fitting in high-dimensional parameter spaces and can contribute to an improved dynamical brain modeling as well as its application to the frameworks of inter-individual variability and brain-behavior relationships.