By name: functional-whole-brain-models-fwbm description: "Functional Whole-Brain Models (fWBMs) - unified modeling paradigm integrating bottom-up whole-brain modeling with top-down neuroconnectionism. Defines 4 minimal criteria and 3-pillar roadmap for unifying brain structure and cognitive function. Based on arXiv:2605.18118 (May 2026). Use when studying whole-brain modeling, neural mass models, brain-inspired DNN architectures, connectome-based modeling, or cross-scale brain computation frameworks." license: Complete terms in LICENSE.txt metadata: arxiv_id: "2605.18118" published: "2026-05-18" authors: "Mario Senden, Leonardo Dalla Porta, Jan Fousek, Jorge F. Mejias, Gorka Zamora-López" categories: [q-bio.NC, q-bio.q-bio] tags: [whole-brain-modeling, neuroconnectionism, functional-whole-brain-models, neural-mass-models, connectome, brain-dynamics, cognitive-modeling, continuous-time-dynamics, brain-inspired-ai, computational-neuroscience]
Functional Whole-Brain Models (fWBMs): A Framework for Unifying Brain Structure and Cognitive Function
Source: arXiv:2605.18118 (q-bio.NC), published 2026-05-18
Authors: Mario Senden, Leonardo Dalla Porta, Jan Fousek, Jorge F. Mejias, Gorka Zamora-López
Description
This skill provides the methodology, criteria, and roadmap for constructing Functional Whole-Brain Models (fWBMs) — a new modeling paradigm that unifies two traditionally disconnected approaches in computational neuroscience: bottom-up biophysical whole-brain modeling (WBM) and top-down task-optimized neuroconnectionism. fWBMs define a class of models that simultaneously satisfy structural grounding, continuous-time dynamical realism, functional competence, and mappable observables. This skill is intended for researchers building brain models that bridge structure and function, designing brain-inspired AI architectures, or developing clinical tools for brain disorders.
1. The Problem: Two Complementary but Disconnected Traditions
Contemporary computational neuroscience features two highly successful yet disconnected modeling traditions.
1.1 Bottom-Up Whole-Brain Modeling (WBM)
WBMs start from empirical neuroanatomy — structural connectomes derived from diffusion MRI, histological tracing, or other modalities — and simulate large-scale neural dynamics using biophysically realistic neural mass or mean-field models at each node.
Strengths:
- Biophysically grounded: parameters map to known physiology (e.g., synaptic strengths, ion channel densities)
- Realistic dynamics: captures resting-state fMRI functional connectivity, spontaneous fluctuations, metastability
- Clinically relevant: can simulate structural perturbations (lesions, degeneration, stimulation)
Limitations:
- Lacks functional competence: cannot perform cognitive tasks
- Optimized for dynamics, not computation: no notion of task goals, inputs, or outputs
- Validation restricted to resting-state correlations with empirical neuroimaging
1.2 Top-Down Neuroconnectionism
Neuroconnectionist models (a term encompassing deep neural networks trained on cognitive tasks) are optimized purely for functional performance.
Strengths:
- Functional competence: excel at vision, language, navigation, working memory
- Task-aligned representations: learned representations correlate with neural data
- Mechanistic interpretability: can analyze how computations emerge from network dynamics
Limitations:
- Limited biological grounding: typically use spatially uniform, layer-wise architectures
- Discrete processing: information propagates through discrete layers, not continuous-time dynamics
- Task-specific: different architectures for different tasks, lack unified whole-brain framework
1.3 The Gap
The key observation: WBMs capture where and how brain activity emerges from structure, while neuroconnectionist models capture what the brain computes. Neither alone provides a complete account of brain function. fWBMs are proposed as the bridge.
2. Four Minimal Criteria for fWBMs
A model qualifies as a Functional Whole-Brain Model (fWBM) if and only if it satisfies all four of the following criteria:
Criterion 1: Structural Grounding
The model must incorporate empirically derived brain structure at multiple scales:
- Macroscale connectome: Structural connectivity derived from diffusion MRI tractography, polarized light imaging (PLI), or histological tract tracing, typically represented as a weighted adjacency matrix between brain regions (e.g., parcellated into 100-1000 nodes)
- Regional biology: Biophysically meaningful node models that incorporate region-specific properties (e.g., cortical vs. subcortical differences, laminar structure, receptor densities)
- Spatial embedding: White matter tract lengths, conduction delays, and distance-dependent connectivity constraints
Implementation forms:
- Diffusion MRI + tractography → structural connectivity matrix
- Histological atlases (e.g., Allen Brain Atlas, BigBrain) → regional gene expression, cytoarchitecture
- Receptor/neurotransmitter maps → regional heterogeneity in neural dynamics
Criterion 2: Continuous-Time Dynamical Realism
The model must employ continuous-time dynamical systems, not discrete processing layers:
- Neural mass models (NMMs): Mean-field reductions of local populations (excitatory/inhibitory), e.g., Jansen-Rit, Wong-Wang, reduced FitzHugh-Nagumo
- Neural field models: Spatially continuous extensions capturing wave propagation
- Conduction delays: Axonal transmission delays between regions based on tract lengths and conduction velocities
- Recurrent dynamics: Within- and between-region recurrent connectivity enabling attractor dynamics, oscillations, and transient computation
Key requirement: Information processing must emerge from ongoing dynamics, not from feedforward layer-by-layer computation. The model should exhibit spontaneous activity, multistability, and sensitivity to initial conditions — hallmarks of biological neural systems.
Criterion 3: Functional Competence
The model must be capable of performing cognitive tasks across one or more domains:
- Task input: Task-relevant stimuli must be able to drive the model's dynamics (e.g., visual input to occipital nodes, tactile input to somatosensory nodes)
- Task readout: Behavioral outputs (e.g., decision, motor response) must be decodable from model activity
- Cognitive domains:
- Working memory (parametric, spatial, verbal)
- Decision-making (perceptual, value-based, multi-attribute)
- Attention (spatial, feature-based, executive control)
- Learning and plasticity (supervised, reinforcement, unsupervised)
- Language processing (at the level of word/phrase representations)
- Navigation and spatial cognition
Evaluation: Performance metrics matched to human/animal behavioral benchmarks. The model should capture not just accuracy but also reaction times, error patterns, and cognitive biases.
Criterion 4: Mappable Observables
The model must generate predictions that can be directly compared to empirical measurements:
- BOLD fMRI: Hemodynamic forward models (e.g., Balloon-Windkessel model) convert neural activity to simulated BOLD signals for resting-state and task-based comparisons
- EEG/MEG: Electromagnetic forward models project neural currents to sensor-level signals, enabling comparison of evoked potentials, spectral power, and connectivity
- Electrophysiology: Local field potentials (LFPs), multi-unit activity (MUA), and single-unit spiking activity
- Behavioral data: Reaction times, response choices, eye movements, and other behavioral readouts
- Lesion/degeneration simulations: Structural perturbations to test causal predictions against clinical data
Validation approach:
- Resting-state: functional connectivity (FC) matrices, dynamic FC, metastability measures
- Task-based: evoked responses, representational similarity analysis (RSA), encoding models
- Perturbation: compare simulated lesion effects to patient data or pharmacological interventions
3. Three-Pillar Roadmap
The paper outlines a phased roadmap spanning short, mid, and long-term horizons.
3.1 Short-Term Horizon (Immediately Feasible)
Goal: Demonstrate feasibility by integrating existing components.
| Action | Description | Current Status |
|---|---|---|
| Simple task modules + existing WBMs | Add minimal task inputs and readouts to existing biophysical models (e.g., threshold-based decisions in TheVirtualBrain) | Multiple groups already have proofs of concept |
| Structural-functional benchmarks | Develop standardized benchmarks that jointly evaluate structural fidelity and task performance | Gap: no consensus benchmarks exist |
| Common validation protocols | Establish shared pipelines for comparing fWBMs across labs (standardized preprocessing, null models, statistics) | Adapt existing neuroimaging analysis workflows |
| Open-source repositories | Create shared code repositories with modular, extensible fWBM frameworks | TheVirtualBrain (TVB) serves as natural platform |
Immediate low-hanging fruit:
- Add a perceptual decision-making module to a resting-state WBM: present sensory evidence to sensory regions, accumulate evidence in decision-related areas (e.g., parietal), read out decision when threshold crossing occurs
- Implement working memory in a spiking neural mass model using sustained activity in prefrontal cortex nodes
3.2 Mid-Term Horizon (3-7 Years)
Goal: Build hybrid models with learned task components grounded in realistic dynamics.
| Action | Description | Key Challenge |
|---|---|---|
| Hybrid parameter optimization | Use gradient-based methods (e.g., backpropagation through time, BPTT) to optimize task-relevant parameters while constraining biophysical realism | Differentiable simulation of biophysical models |
| Cross-scale linking hypotheses | Develop formal relationships between microscopic (synaptic, cellular), mesoscopic (columnar, areal), and macroscopic (whole-brain, behavioral) variables | Bridging the scales is the hardest open problem |
| Shared software infrastructure | Build modular, composable frameworks where structural connectomes, node models, task modules, and forward models can be mixed and matched | Interoperability between different codebases |
| Multi-task training | Train a single fWBM on multiple cognitive tasks simultaneously to test for emergent task-general representations | Catastrophic forgetting, capacity limits |
Key enabler: Differentiable neural mass models — enabling gradient-based optimization of synaptic weights, time constants, and connectivity parameters while preserving biophysical realism. This links neuroconnectionist learning to WBM realism.
3.3 Long-Term Horizon (10+ Years)
Goal: Realize fully unified fWBMs with biophysical realism and multi-domain cognitive capability.
| Action | Description |
|---|---|
| Fully unified fWBMs | Models that satisfy all four criteria at scale: whole-brain realism, multiple cognitive domains, and comprehensive empirical validation |
| Clinical translation | fWBMs as digital twins for individual patients: predict how structural pathology (trauma, neurodegeneration, neurodevelopmental disorders) translates to cognitive deficits |
| Stimulation optimization | Closed-loop optimization of brain stimulation targets and protocols (TMS, tDCS, DBS) using personalized fWBMs |
| Developmental modeling | Track and simulate structural-functional coupling across development, aging, and learning |
| Theory building | Use fWBMs as in silico laboratories to test foundational theories of neural computation: predictive coding, free energy principle, dynamic causal models |
Ultimate vision: An fWBM trained on the same tasks humans perform, grounded in a subject-specific connectome, generating dynamics that match the subject's fMRI/EEG, and capable of predicting the cognitive effects of structural changes.
4. Scientific and Clinical Opportunities
4.1 Brain Disorders and Clinical Neuroscience
| Application | Approach |
|---|---|
| Neurodegeneration (Alzheimer's, Parkinson's) | Seed structural pathology in specific regions, simulate spread along connectome, predict cognitive deficits at each stage |
| Psychiatric disorders (schizophrenia, depression, autism) | Alter regional excitability/inhibition balance, simulate effects on large-scale dynamics and task performance |
| Stroke and traumatic brain injury | Remove or degrade specific nodes/connections, predict functional compensation and recovery trajectories |
| Neurodevelopmental disorders (ADHD, dyslexia) | Simulate altered developmental trajectories with regionally specific structural differences |
| Epilepsy | Model seizure propagation dynamics and test surgical resection strategies in silico |
4.2 Brain Stimulation Optimization
| Modality | fWBM Contribution |
|---|---|
| Transcranial Magnetic Stimulation (TMS) | Predict downstream effects of focal stimulation; optimize frequency, intensity, and target |
| Transcranial Direct Current Stimulation (tDCS) | Model current spread and network-level neuromodulation |
| Deep Brain Stimulation (DBS) | Test electrode placement and stimulation parameters for movement and psychiatric disorders |
4.3 Cognitive Science and Systems Neuroscience
| Opportunity | Description |
|---|---|
| Mechanistic hypothesis testing | Test competing theories of cognitive function (e.g., is working memory maintained by persistent activity or activity-silent synaptic traces?) in the same whole-brain model |
| Structural-functional coupling | Quantify how much of functional connectivity is explained by structural connectivity vs. task demands |
| Causal inference | Perform in silico lesion experiments impossible in humans: selective inactivation of specific connections or cell types |
| Individual differences | Link individual variation in connectome structure to variation in cognitive performance |
| Comparative neuroscience | Build connectome-specific fWBMs for different species to study evolution of cognitive capacity |
4.4 Artificial Intelligence
| Opportunity | Description |
|---|---|
| Brain-inspired architectures | Extract design principles from fWBMs: continuous-time recurrent computation, spatial embedding, conduction delays, and noise |
| Energy efficiency | Spiking network implementations of fWBMs can inform neuromorphic computing |
| Robustness | Biological noise and degenerate solutions may inspire more robust AI systems |
| Lifelong learning | Structural plasticity mechanisms in fWBMs may inform continual learning approaches |
5. Implementation Methodology
5.1 Architecture Components
A complete fWBM implementation requires these composable modules:
fWBM = StructuralConnectome + NodeModel + TaskInterface + ForwardModel
- StructuralConnectome: Adjacency matrix (N×N) with optional edge weights (tract count, fractional anisotropy), tract lengths, and inter-region conduction delays
- NodeModel: Per-region dynamical system, either:
- Neural mass models: Reduced population models (e.g., Wong-Wang for excitatory/inhibitory, Jansen-Rit for cortical columns)
- Mean-field models: Single-population reduction (e.g., reduced FitzHugh-Nagumo, Kuramoto oscillators)
- Spiking models: Izhikevich, AdEx, or Hodgkin-Huxley neurons (more computationally expensive)
- Hybrid models: Different node types for different regions (e.g., thalamic nodes use different dynamics than cortical nodes)
- TaskInterface: Input encoding (how stimuli enter the model) and output decoding (how behavior is read from model activity)
- Input: stimulus features → regional activation patterns
- Output: activity from task-relevant regions → decision/response classifier
- ForwardModel: Mapping from neural activity to observable signals
- Hemodynamic (Balloon-Windkessel → BOLD)
- Electromagnetic (lead-field → EEG/MEG)
5.2 Model Design Decisions
| Decision | Options | Trade-off |
|---|---|---|
| Node model complexity | Spiking → NMM → reduced NMM | Biophysical detail vs. computational tractability |
| Connectome resolution | 100 → 1000+ nodes | Anatomical detail vs. parameter identifiability |
| Delay inclusion | Zero-delay → realistic delays | Accuracy vs. simulation stability |
| Noise | None → Poisson → Ornstein-Uhlenbeck | Biological realism vs. signal-to-noise ratio |
| Learning | Fixed weights → Hebbian → BPTT | Biological plausibility vs. functional performance |
5.3 Training and Optimization Pipeline
- Initialize with empirical structural connectome and standard biophysical parameters
- Validate resting-state dynamics — ensure spontaneous activity matches empirical FC, power spectra, and metastability measures
- Inject task structure — add input encoding and output decoding mechanisms
- Optimize task parameters — adjust local parameters (synaptic strengths, time constants, thresholds) to achieve task performance
- Validate task activity — compare simulated task-evoked responses to empirical neuroimaging data
- Iterate — structural parameters may also be adjusted within plausible bounds
5.4 Key Computational Tools
| Tool | Purpose |
|---|---|
| TheVirtualBrain (TVB) | Established platform for large-scale WBM simulations; natural candidate for fWBM extension |
| NEST / Brian2 | Spiking network simulators for detailed node models |
| PyTorch / JAX | Automatic differentiation for gradient-based optimization of task parameters |
| DiffEq libraries (torchdiffeq, Diffrax) | Differentiable ODE solvers enabling BPTT through neural mass dynamics |
| NiBabel / Nilearn | Neuroimaging data loading, manipulation, and analysis |
6. Evaluation Metrics and Benchmarking
6.1 Structural Fidelity Metrics
| Metric | Description |
|---|---|
| Structural connectome similarity | Correlation between model connectome and empirical connectome |
| Regional property matching | Distribution of node model parameters matches empirical distribution (e.g., cortical thickness, myelination) |
6.2 Dynamical Fidelity Metrics
| Metric | Description |
|---|---|
| Functional connectivity (FC) correlation | Pearson/Spearman correlation between simulated and empirical FC matrices |
| Edge functional connectivity (eFC) | Distribution of pairwise correlations between regions |
| Dynamic FC (dFC) | Sliding-window FC variability, metastability index |
| Power spectra | Band-limited power (delta, theta, alpha, beta, gamma) across regions |
| Functional connectivity dynamics (FCD) | Second-order FC: correlation between FC states over time |
| Graph-theoretic measures | Small-worldness, modularity, hub structure, rich-club organization |
6.3 Functional Performance Metrics
| Metric | Description | Example |
|---|---|---|
| Task accuracy | Percent correct on cognitive tasks | Perceptual decision: % correct choices |
| Reaction times (RT) | Simulated decision latencies | Match human RT distributions |
| RT-accuracy trade-off | Speed-accuracy characteristic | Compare to human speed-accuracy curves |
| Representational similarity (RSA) | Correlation of representational geometries between model and brain | IT cortex representations vs. model activity |
| Encoding model performance | How well model activity predicts neural responses | Variance explained in voxel responses |
| Lesion behavior | Performance degradation under simulated damage | Match patient neuropsychological profiles |
6.4 Unification Metrics
| Metric | Purpose |
|---|---|
| Structure-function coherence | How task performance depends on structural connectome fidelity |
| Explanatory scope | Number of empirical phenomena (across scales and modalities) simultaneously explained |
| Predictive validity | Out-of-sample prediction of novel empirical findings |
7. Activation Keywords
- functional whole-brain models, fWBM, functional WBM
- whole-brain modeling, whole brain model, large-scale brain model
- neuroconnectionism, neuroconnectionist model
- brain structure-function coupling, structure-function unification
- neural mass model, mean-field model, biophysical brain model
- connectome-based modeling, connectome-informed neural network
- continuous-time brain dynamics, brain dynamical system
- brain-inspired AI, brain-grounded AI
- cognitive modeling, brain task modeling
- brain stimulation optimization, TMS modeling, DBS modeling
- brain disorder modeling, clinical brain simulation
8. Related Skills
multi-objective-optimisation-oscillatory-snn: Multi-objective optimization for oscillatory spiking neural networks — complementary numerical techniquesadaptive-spiking-neuron-asn: Adaptive spiking neuron implementations for vision and language — can serve as node model componentspiking-neural-network-analysis: Analysis framework for SNN papers — useful for evaluating fWBM node model dynamics
References
- Senden, M., Dalla Porta, L., Fousek, J., Mejias, J. F., & Zamora-López, G. (2026). Functional Whole-Brain Models: A New Framework for Unifying Brain Structure and Cognitive Function. arXiv:2605.18118 [q-bio.NC].
- Sanz-Leon, P., Knock, S. A., Spiegler, A., & Jirsa, V. K. (2015). Mathematical framework for large-scale brain network modeling. NeuroImage, 111, 385-430.
- Deco, G., Jirsa, V. K., & McIntosh, A. R. (2011). Emerging concepts for the dynamical organization of resting-state activity. Nature Reviews Neuroscience, 12(1), 43-56.
- Yamins, D. L., & DiCarlo, J. J. (2016). Using goal-driven deep learning models to understand sensory cortex. Nature Neuroscience, 19(3), 356-365.
Notes
- The fWBM framework is proposal-based — no single fWBM currently satisfies all four criteria simultaneously. The roadmap lays out a phased path toward that goal.
- The four criteria are deliberately minimal: any model satisfying them qualifies, regardless of specific implementation choices.
- Short-term integrations (Criterion 1 + 2 + task input/readout) are immediately feasible with existing tools and are likely to appear in the literature within 1-2 years.
- The framework is agnostic to node model choice — spiking models, neural masses, and reduced mean-field models are all valid depending on the research question.
- For clinical applications, subject-specific connectomes from patient MRI are essential for personalized fWBMs.
Last updated: 2026-05-26