A few technical details: The frontend is written in Python (for ease of use), but to optimize efficiency, the simulation core is written in CUDA and C++. And not just simulations, but also calculation of FC, FC dynamics and goodness of fit measures is done on GPUs 🚀

17/n

As examples of the utility of these individualized models, we investigated test-retest reliability and heritability of simulated (and empirical) measures in the HCP dataset, and found simulated features to be fairly reliable and heritable.

15/n

cuBNM also supports regional heterogeneity of parameters, using map-based and node-based approaches.
To demonstrate these approaches in the preprint, we applied them to fit group-level models of the HCP data, and as expected, found them to outperform the homogenous model.

12/n

Two optimization approaches are included in cuBNM:
1️⃣ grid search
2️⃣ evolutionary optimizers (via pymoo, with CMA-ES as the featured method)
We demonstrate both in the preprint by fitting the rWW model to the group-averaged HCP data.

11/n

cuBNM currently includes five commonly used models (with rWW model featured in tutorials and the preprint).
New models can be added via YAML model specification files. And a guide for contributing new models is included in the documentation.

10/n

Using cuBNM, if you have access to a few GPUs*, you can quickly and easily create individualized BNMs for your own dataset. All you need is: resting-state BOLD & structural connectivity (from each subject’s DWI, or a template SC).

* CPUs are also supported, but are not our focus.

8/n

Why are GPUs so much faster?
GPUs are designed for parallel processing, so computations across both *simulations* and *nodes* can be massively parallelized 🧠⚙️

7/n

GPU implementation also makes it feasible to simulate denser networks with thousands of nodes: Compute time with number of nodes increases near-quadratically on CPUs, but near-linearly on GPUs.

6/n

On GPU, compared to single-core CPU, simulations can run up to 1000+ times faster (typically: 400-800x faster).
For example, running 32,768 simulations would take 3.8 *days* on a single-core CPU, but only 5.6 *minutes* on a data-center A100 GPU ⚡

5/n

Lastly, we ended with a methodological note on alternative simulation-based measures of E-I ratio, importantly highlighting that relying on model parameters may not be ideal as they are interdependent and hard to interpret in isolation.

We then rigorously tested sensitivity of our findings to modeling choices and confounds: structural connectivity, parcellation, parameter settings, noise, and more. Despite some differences, the main developmental pattern (↓E-I ratio in association areas) remained consistent.

Interestingly, the spatial pattern of E-I maturation aligned with the sensorimotor-association axis of cortical development. In addition, genes associated with regions showing ↓E-I ratio were enriched in later developmental stages compared to genes associated with regions showing ↑E-I ratio.

We replicated this pattern longitudinally in the IMAGEN dataset (n=149): Over ~5 years of follow-up (14y → 19y), association areas showed increased inhibition (↓E-I ratio), while sensorimotor areas showed ↑E-I ratio or remained stable.

We then studied how E-I ratio changes with age.

In the PNC cross-sectional data (n=752, 10-19 y), we found age-related decreases in the E-I ratio in frontal, parietal, and temporal association areas, and increases or stable E-I ratios in sensorimotor and visual regions.

So we used structural connectivity + functional imaging data from adolescents in PNC and IMAGEN datasets and built individualized biophysical network models to obtain individualized regional estimates of E-I ratio. We used the reduced Wong-Wang (rWW) model with heterogeneous regional parameters.