🤖 machine learning

BrainDyn: A Sheaf Neural ODE for Generative Brain Dynamics

BrainDyn is a novel sheaf neural ordinary differential equation model that integrates anatomical brain structure with continuous-time dynamics to generate realistic brain activity across fMRI, EEG, and simulated data, enabling accurate forecasting and in silico perturbation analysis.

Original authors: Siddharth Viswanath, Panayiotis Ketonis, Chen Liu, Michael Perlmutter, Dhananjay Bhaskar, Smita Krishnaswamy

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Siddharth Viswanath, Panayiotis Ketonis, Chen Liu, Michael Perlmutter, Dhananjay Bhaskar, Smita Krishnaswamy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict how a massive orchestra will play the next few seconds of a symphony. You have a recording of what they just played, and you want to know what comes next.

Most current computer models try to do this in two ways, and both have flaws:

The "Blind Conductor" approach: These models (like standard AI) listen to every musician but ignore the fact that the violin section sits next to the violas, or that the drummer talks to the bassist. They treat every musician as an isolated individual, missing how the structure of the orchestra shapes the music.
The "One-Size-Fits-All" approach: Other models know the musicians are connected, but they assume everyone speaks the same language and uses the exact same volume. In reality, a violinist might whisper to a flutist but shout at a tuba player. Treating all connections the same way makes the music sound muddy and flat.

BrainDyn is a new computer model designed to fix this. It acts like a super-smart conductor who understands both the history of the music and the unique relationships between every single musician.

Here is how it works, broken down into simple parts:

1. The "Memory Bank" (LSTM)

First, BrainDyn doesn't just look at what a musician played right now. It looks at their recent history. It uses a "memory bank" (called an LSTM) to remember the last few seconds of a specific brain region's activity. This is like a musician remembering the melody they just played so they know how to start the next phrase.

2. The "Custom Translator" (Sheaf Restriction Maps)

This is the paper's biggest innovation. In a normal model, if the "Left Brain" talks to the "Right Brain," they just swap raw data.
BrainDyn is different. It acts like a custom translator for every single connection.

When the Left Brain talks to the Right Brain, BrainDyn uses a specific "translator" to change the message into a format the Right Brain understands best.
When the Front Brain talks to the Back Brain, it uses a different translator.
It realizes that different parts of the brain "speak" in different ways. This prevents the model from getting confused or "muddy" (a problem scientists call oversmoothing), keeping the unique character of each brain region alive.

3. The "Continuous Flow" (Neural ODE)

Most models predict the future in choppy steps (like a flipbook). BrainDyn, however, uses a Neural Ordinary Differential Equation (Neural ODE). Think of this as a smooth, flowing river rather than a series of stepping stones. It calculates how the brain activity evolves continuously over time, filling in the gaps between seconds to create a fluid, realistic prediction of what happens next.

4. The "Disagreement Detector" (Sheaf Laplacian)

How does the model know when to change the music? It looks for disagreements.
If the Left Brain and Right Brain are supposed to be in sync, but their "translated" messages don't match, BrainDyn detects this mismatch. It then uses that difference to drive the system toward a new, coordinated state. It's like a conductor noticing the violins are slightly out of tune with the cellos and gently guiding them back into harmony.

What Did They Test?

The researchers tested BrainDyn on three very different types of "orchestras":

The Slow Orchestra (fMRI): Real brain scans from healthy people showing slow blood flow changes.
The Fast Orchestra (EEG): Real brain scans from people with epilepsy, showing very fast electrical spikes.
The Simulated Orchestra (NEST): A computer simulation of 100 artificial neurons firing.

The Results:

Better Predictions: BrainDyn was better at predicting the next few seconds of brain activity than any other model, whether the data was slow (fMRI) or fast (EEG).
Handling Surprises: In the simulation, they "silenced" one neuron (like muting one musician) to see if the model could predict how the rest of the network would react. BrainDyn was the only model that could accurately predict how the disturbance would ripple through the network, even though it had never seen that specific silence before.

The Bottom Line

BrainDyn is a new way to build a "virtual brain." It doesn't just guess the future; it understands that the brain is a complex, connected system where every part has its own history and speaks a slightly different language to its neighbors. By respecting these unique connections and the smooth flow of time, it creates a much more accurate picture of how our brains work.

Technical Summary: BrainDyn - A Sheaf Neural ODE for Generative Brain Dynamics

Problem Statement

Existing generative models for brain dynamics face a fundamental trade-off between temporal expressivity and anatomical fidelity. Large language models (LLMs) and standard recurrent neural networks (RNNs) capture temporal dependencies but ignore the anatomical organization of the brain, failing to produce representations aligned with specific brain regions. Conversely, standard Graph Neural Networks (GNNs) incorporate structural connectivity but rely on simple message-passing rules (scalar summation or averaging) within a homogeneous feature space. This assumption is biologically questionable, as distinct brain regions possess different representational geometries and contents. Consequently, standard GNNs suffer from "oversmoothing," where repeated aggregation homogenizes regional activity and erases the heterogeneous dynamics essential to neural computation. Furthermore, many existing models lack a continuous-time framework, limiting their ability to model the intrinsic, fluid evolution of neural states.

Methodology: BrainDyn

The authors propose BrainDyn, a sheaf neural ordinary differential equation (neural ODE) framework designed to model continuous-time dynamics on structured brain graphs. The architecture integrates three core components:

1. Memory-Based Node Stalks

To capture the history-dependent nature of neural activity, BrainDyn encodes the recent temporal history of each brain region (node) using a Long Short-Term Memory (LSTM) network over a sliding temporal window. The resulting hidden states serve as stalks (vector spaces associated with nodes), representing the intrinsic temporal structure and latent dynamical state of each region.

2. Learnable Sheaf Restriction Maps

Instead of aggregating features in a single shared space, BrainDyn employs a cellular sheaf structure.

Restriction Maps: For each edge connecting two nodes, learnable linear maps ( $\rho$ ) project the node features into an edge-specific shared space. This allows different connections to implement structured, feature-wise transformations, reflecting the hypothesis that communication between brain regions involves transformation, not just summation.
Attention Mechanism: Learnable attention coefficients modulate the contribution of neighboring nodes, allowing the model to selectively emphasize or suppress specific dynamical modes based on context.
Sheaf Laplacian: The model computes a discrepancy between neighboring nodes after their features have been aligned via restriction maps. This discrepancy is aggregated back to the nodes via the Sheaf Laplacian ( $L_F = \delta^\top \delta$ ), which generalizes the classical graph Laplacian by incorporating edge-dependent transformations before measuring inconsistency.

3. Continuous-Time Evolution via Neural ODE

The aggregated signals from the sheaf message passing are fed into a neural ODE. The vector field of the ODE is defined by a Multi-Layer Perceptron (MLP) acting on the sheaf-processed node representations. This setup governs the continuous-time evolution of neural activity, driven by both intrinsic dynamics and the heterogeneous interactions captured by the sheaf structure.

Training and Regularization

The model is trained to minimize the Mean Squared Error (MSE) between predicted and observed neural trajectories. To ensure biological plausibility and generalizability, two regularization terms are added:

Sparsity Loss ( $L_{sparse}$ ): Penalizes the magnitude of edge discrepancy signals to encourage parsimonious communication patterns.
Prior Graph Loss ( $L_{prior}$ ): Encourages learned interactions to remain consistent with a prior graph (constructed via Granger causality from the input context), suppressing unsupported interactions while reinforcing known structural/functional connections.

Key Contributions

Sheaf-Based Neural Dynamical Model: The paper introduces the first framework combining cellular sheaves with neural ODEs to model continuous-time neural dynamics on structured networks.
Expressive Interaction Modeling: By utilizing learnable restriction maps, BrainDyn enables structured, feature-wise transformations across connections, allowing brain regions to maintain distinct representational geometries while communicating.
Generative Modeling of Neural Dynamics: The model generates brain-like trajectories that capture neural transients and temporal evolution, reflecting underlying interactions across regions.
Comprehensive Evaluation: The framework is evaluated across three distinct modalities: resting-state fMRI (PNC dataset), scalp EEG with focal epilepsy (TUSZ dataset), and simulated spiking network activity (NEST simulator).

Empirical Results

BrainDyn was evaluated against representative baselines, including sequence models (CNN-LSTM, BIOT), dynamic graph models (EvolveGCN), and graph neural ODEs (ODEBRAIN, RiTINI).

Forecasting Performance: On both PNC fMRI and TUSZ EEG datasets, BrainDyn achieved the lowest error rates (MSE, MAE, and normalized Dynamic Time Warping distance) compared to all baselines. Notably, it outperformed modality-specific models, demonstrating robustness across different temporal and spatial scales.
Ablation Studies: Removing the LSTM encoder or the sheaf structure (reverting to a standard graph) resulted in significant performance degradation, confirming that both temporal encoding and heterogeneous edge transformations are critical.
Perturbation Generalization: In NEST simulations, BrainDyn was trained on unperturbed activity and tested on out-of-distribution perturbed dynamics (silencing a single neuron). It successfully extrapolated to these perturbed states, outperforming the strongest baseline (ODEBRAIN) in forecasting the subsequent trajectory. This suggests the model learns the underlying generative dynamics rather than merely fitting observed time series.

Significance and Claims

The authors position BrainDyn as a step toward a "principled and expressive virtual brain." The significance of the work lies in its ability to:

Bridge the Gap: It addresses the limitations of both sequence models (lack of structure) and standard GNNs (oversmoothing and lack of continuous-time dynamics).
Biological Plausibility: By modeling communication as transformation rather than simple summation, the architecture aligns with biological hypotheses regarding heterogeneous regional representations.
In Silico Utility: The model's ability to generalize to perturbed, out-of-distribution dynamics establishes its potential as a testbed for in silico perturbation analysis and intervention studies, supporting downstream tasks in understanding neurological disorders and brain state transitions.

The paper concludes that combining sheaf-based interaction modeling with continuous-time dynamics provides a robust framework for learning generalizable representations of brain activity across modalities.