cuBNM: GPU-Accelerated Brain Network Modeling

Reviewer #1 (Public review):

Summary:

The manuscript introduces cuBNM, a GPU‑accelerated Python package for whole‑brain modeling. The authors demonstrate that running simulations on GPUs provides substantial benefits in computational speed, cost-efficiency, and scalability compared to traditionally used CPUs, making large‑scale and individualized brain network modeling computationally feasible. The usage of cuBNM has been demonstrated by running optimization of group-level and individualized low- and high-dimensional models. By investigating the test-retest reliability and heritability of simulated and empirical measures in the Human Connectome Project dataset, the authors showed that simulated features were fairly reliable and significantly heritable.

Strengths:

This study is timely and presents an important contribution to the field of whole-brain computational modeling. A major strength is that the authors go beyond introducing a GPU-accelerated framework by demonstrating its utility through comprehensive benchmarking and biologically relevant applications, including individualized model fitting, comparisons of homogeneous and heterogeneous models, and analyses of test-retest reliability and heritability.

The computational performance is evaluated comprehensively, assessing speed, computational cost, energy consumption, and scalability across different simulation settings. The Human Connectome Project dataset is used to demonstrate that the software enables individualized whole-brain modeling in large datasets.

Finally, the software is modular, open-source, and well-documented, and can facilitate the broader adoption of GPU-accelerated whole-brain modeling within the neuroscience community.

Weaknesses:

The test-retest reliability and heritability are estimated using high-quality Human Connectome Project data. The manuscript would benefit from discussion and/or demonstrations regarding how the software performs under more challenging conditions, such as clinical datasets, shorter data acquisitions, higher-motion datasets, or multi-site datasets.

Apart from demonstrating the benefits of GPUs over CPUs, the manuscript would benefit from a more direct comparison between cuBNM and other whole-brain modeling software, such as The Virtual Brain.

The manuscript demonstrates that heterogeneous models improve the fit to empirical functional connectivity. However, the biological interpretation of this improvement could be expanded. The heterogeneous models are also more complex than homogeneous models, and some improvement in model fit may be explained by the increased model flexibility.

In whole-brain brain network modeling, different parameter combinations can result in similar empirical functional connectivity measures. The manuscript would benefit from a discussion of how this influences the interpretation of individualized parameter estimates.

Reviewer #2 (Public review):

Summary:

The authors aim to address a major problem in brain network modeling: the high computational cost of simulating and fitting brain activity models, particularly for large samples, individualized models, and broad parameter searches. They introduce cuBNM, an open software package that uses graphics processing units to accelerate model simulation, fitting, and calculation of simulated brain activity features.

The manuscript is primarily a methods and software contribution, rather than a paper providing novel neurobiological insights. The authors demonstrate the tool using human imaging data, showing examples of group-level and individualized model fitting, comparisons between homogeneous and heterogeneous model parameterizations, and analyses of repeated-measurement stability and genetic influences of simulated features. They also provide speed and scaling tests to support the claim that the software can make large-scale and individualized brain network modeling more practical for the field.

Strengths:

A major strength of this work is that it addresses a clear computational bottleneck in brain network modeling. The authors provide an open software package that combines a user-friendly Python interface with an accelerated back-end, making large numbers of simulations and model fits more practical for other researchers.

The demonstrations are broad and relevant to real use cases. The authors show group-level and individualized model fitting, different optimization strategies, and comparisons between homogeneous and heterogeneous models, rather than limiting the paper to a narrow technical benchmark. The benchmarking and openness of the work further increase its value. The comparisons across hardware and network sizes give readers a practical sense of the tool's speed and scalability, while the availability of code, documentation, tutorials, and containers should make the method easier for the community to test and adopt.

Weaknesses:

(1) The benchmarking provides solid evidence for substantial speed improvements within the authors' implementation, but the generality of the performance claims is more limited. The largest reported speed-ups are measured relative to a single central processing unit thread, and the study does not fully benchmark cuBNM against other optimized brain modeling frameworks. This makes the results useful as evidence of strong acceleration in the tested setting, but less definitive as a general comparison across available implementations.

(2) The comparison between homogeneous and heterogeneous models is informative, but it is not fully controlled for model complexity. The best-fitting node-based heterogeneous model has more free parameters than the homogeneous and map-based alternatives, so its improved fit may partly reflect greater flexibility rather than a more biologically valid parameterization. As a result, the model comparison supports the conclusion that this parameterization fits better under the current setup, but not necessarily that it is generally superior or more biologically realistic.

(3) The reliability and heritability analyses are valuable demonstrations of what scalable individualized modeling can enable, but they do not establish the simulated features as validated biological mechanisms. Because these simulated features are derived from individualized structural and functional imaging data, their stability across repeated measurements and genetic influences may partly reflect information already present in the empirical inputs or fitting targets. These results therefore support a more cautious conclusion: the simulated features retain stable and genetically structured variation, but their biological interpretation remains model dependent.

(4) The empirical demonstrations are narrower than some of the broader claims made in the manuscript. Most analyses rely on one human imaging dataset, one cortical parcelation, one main brain model, and a specific fitting objective, while broader claims refer to diverse populations, dense networks, high-dimensional models, and biological applications. The current results show that cuBNM is a useful and scalable tool in the tested setting, but the extent to which the findings generalize across datasets, model classes, network resolutions, or clinical contexts remains to be established.