TVB Platform: Multi-Scale Brain Simulation

• The Virtual Brain (TVB) Platform is an open-source ecosystem for constructing, simulating, and analyzing multiscale brain network models using empirical MRI data and differential equation-based neural mass models.
• It integrates standardized, containerized MRI pipelines to extract structural and functional connectivity data, enabling robust co-simulation of mean-field and spiking neural dynamics.
• Leveraging high-performance code generation and Bayesian inference, TVB supports personalized clinical applications such as virtual neurosurgery and seizure propagation analysis.

The Virtual Brain (TVB) Platform is an open-source ecosystem for constructing, simulating, and analyzing brain network models (BNMs) across multiple spatial and temporal scales. Integrated into the EBRAINS infrastructure, TVB provides cloud-based services for processing neuroimaging data, building models rooted in individual anatomy, multiscale brain simulation, automated code generation, Bayesian clinical modeling, and collaborative research workflows. Its modular architecture accommodates the simulation of neural population dynamics, facilitates the extraction of connectome and functional matrices from MR data, and supports extensible workflows for both basic and translational neuroscience.

1. Architecture and Core Simulation Framework

TVB defines BNMs by integrating empirical structural connectomes (SCs) with neural mass models (NMMs), which mathematically describe the evolution of neural activity through coupled differential equations of the form:

$\frac{dx}{dt} = f(x, p, I)$

where $x$ are state variables (e.g., membrane potentials, firing rates), $p$ are model parameters (e.g., coupling strength, time constants), and $I$ is any exogenous input. Neural population dynamics can be specified for each brain region, with regional architectures determined from diffusion MRI tractography. Global network architecture and inter-regional delays are built from SC matrices, and neural activity is simulated using the TVB network simulator, which numerically integrates these equations to generate time series that can be mapped to empirical observables such as fMRI, EEG, or local field potentials.

2. MRI Processing and Connectome Extraction Pipelines

TVB provides containerized MRI workflows consisting of three standardized BIDS Apps. These are:

• tvb-pipeline-sc: diffusion MRI tractography for generating SCs
• tvb-pipeline-fmriprep: preprocessing of functional MRI data to derive regional BOLD time series
• tvb-pipeline-converter: integration and conversion of processed data to TVB-ready formats

These workflows process T1, diffusion, and functional MRI datasets to estimate tractograms, compute inter-regional connectivity matrices (incorporating strength and delay), and produce regionally averaged functional time series. Output products include structural connectomes, functional connectivity (FC) matrices, and forward models for EEG. The pipelines are robust, reproducible, and suitable for large cohort studies, enabling standardized model construction and facilitating the transition from raw imaging data to simulation-ready networks.

3. Multi-Scale Brain Simulation and Co-Simulation

Recognizing the multiscalar nature of brain activity, TVB integrates large-scale NMMs with microscale spiking neuronal networks. Two toolboxes—TVB-Multiscale and Parallel CoSimulation—enable bi-directional coupling between mean-field dynamics simulated in TVB and spiking networks in specialized simulators (e.g., NEST). Transformations between population firing rates and spike trains are implemented using, for example, Poisson sampling. This architecture supports multi-level modeling, in which subpopulations operating at different scales exchange information, allowing investigation into the mechanisms by which micro-scale patterns (e.g., seizure initiation in a local network) affect whole-brain dynamics.

4. High-Performance Code Generation and Acceleration

TVB employs the RateML domain-specific language, based on the LEMS standard, allowing users to concisely specify neural models in an XML-based format. TVB then generates optimized simulation code for both CPU and GPU execution. The back-end exploits Numba's vectorization and CUDA parallelization to accelerate model integration, enabling efficient exploration of high-dimensional parameter spaces and long-duration simulations. For demanding simulations, the TVB C++ back-end further accelerates performance by up to $10\times$, leveraging the Eigen library for matrix operations and providing Python bindings for seamless workflow integration (Martín et al., 29 May 2024). This dual-platform approach supports both prototyping (in TVB-Python) and production-scale simulations (in TVB C++), eliminating redundancy in code design and enhancing reproducibility.

5. Personalized and Disease-Specific BNMs

TVB facilitates the generation of simulation-ready models for both healthy volunteers and patient populations. Datasets comprising SC matrices, FC matrices, EEG/MEG projection matrices, and cortical surface reconstructions are standardized for direct upload into TVB. This capability supports clinical workflows, including “virtual neurosurgery” and pre-surgical planning in tumor, stroke, or epilepsy scenarios (Schirner et al., 2021). The platform also extends to animal models, such as mice, via the TVB Mouse Brains service, enabling disease and plasticity modeling across species.

An emerging direction is the use of virtual brain twins, in which individual-specific models are constructed from personal neuroimaging data. Recent advances include amortized inference for privacy-preserving personalization, allowing robust individualized modeling with reduced computational burden (Baldy et al., 26 Jun 2025).

6. Bayesian Inference and Clinical Modeling

Clinical translation is achieved through the Bayesian Virtual Epileptic Patient (BVEP) service, which fits TVB’s Epileptor model (consisting of five coupled state variables) to patient EEG/fMRI data for seizure onset and propagation analysis. Bayesian inference—using MCMC (NUTS) and ADVI—estimates region-specific excitability posteriors:

$P(\theta|data) \propto P(data|\theta) P(\theta)$

Posterior excitability maps allow classification into epileptogenic zones (EZ), propagation zones (PZ), or healthy zones (HZ), directly informing clinical decisions (e.g., resective surgery planning).

7. Collaborative Infrastructure and Educational Resources

TVB is embedded within EBRAINS’s cloud services, supporting collaborative, reproducible neuroscience. Users access containerized JupyterLab environments for interactive simulation and analysis. Datasets, models, and code assets are discoverable via EBRAINS’s KnowledgeGraph, utilizing the openMINDS metadata standard. RESTful APIs and containerized workflows ensure secure, versioned deployment across teams.

Educational support includes interactive tutorials, Jupyter notebooks, lectures, and documentation, as well as curated training spaces such as the INCF TVB training and TVB EduPack (Schirner et al., 2021). Materials are designed to facilitate competency in model building, simulation, data analysis, and interpretation for both novice and expert users.

Table: Key TVB Ecosystem Components

Component	Purpose	Example Tools/Features
BNMs & Simulator	Neural mass modeling, multi-scale sim	TVB network simulator, RateML
MRI Processing Pipelines	Connectome extraction	tvb-pipeline-sc, tvb-pipeline-fmriprep, tvb-pipeline-converter
Multiscale Co-simulation	Spiking/mean-field integration	TVB-Multiscale, Parallel CoSimulation
Code Generation/Acceleration	High-performance simulation	RateML, TVB C++, CUDA/Numba
Clinical Modeling	Bayesian inference of pathology	BVEP, Epileptor, simulation-ready BNMs
Collaboration/Education	Cloud workflows, training	EBRAINS Collaboratory, KnowledgeGraph, INCF training

8. Significance and Future Directions

TVB’s impact is marked by its ability to link experimental data and computational models for rigorous hypothesis-testing across brain states and disease conditions. The platform’s extensibility and integration with high-throughput HPC architectures (including modular frameworks for GPU acceleration and parallelization (Vlag et al., 2023)) allow comprehensive exploration of parameter spaces and individual variability in functional dynamics. The introduction of anonymized personalization and amortized inference further increases the clinical applicability and scalability for large cohort analysis (Baldy et al., 26 Jun 2025).

The collaborative, secure, and open-access infrastructure sets a precedent for generalizability, clinical translation, and reproducibility in neuroscience. A plausible implication is the further convergence of multiscale modeling, advanced visualization, and individualized brain modeling within integrated research ecosystems, enabling robust data-driven calibration, validation, and intervention studies in both basic and translational neuroscience.