Graph-based representation learning of neuronal dynamics and behavior
7 affiliations
- 1Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Clayton, VIC, Australia
- 2Cortical Labs, Melbourne, VIC, Australia
- 3Department of Computer Science, The University of British Columbia, Vancouver, BC, Canada
- 4Neural Dynamics Laboratory, Department of Medicine, The University of Melbourne, Melbourne, VIC, Australia
- 5Department of Biochemistry and Pharmacology, University of Melbourne, Melbourne, VIC, Australia
- 6CIFAR Global Scholars Program, Toronto, ON, Canada
- 7Queen Square Institute of Neurology, University College London, London, United Kingdom
Abstract
Understanding how neuronal networks reorganize in response to stimuli and give rise to behavior is a central challenge in neuroscience and artificial intelligence. Existing methods, however, often struggle to capture the evolving structure of neural connectivity and its relationship to behavior, particularly in dynamic, uncertain, or high-dimensional settings with sufficient resolution and interpretability. We introduce the Temporal Attention-enhanced Variational Graph Recurrent Neural Network (TAVRNN), a framework for modeling time-varying neuronal connectivity by integrating probabilistic graph learning with temporal attention. TAVRNN learns single-unit latent dynamics while maintaining interpretable population-level representations, enabling identification of behaviorally relevant connectivity patterns. It generalizes across neural systems differing in scale, behavioral structure, and supervision regime, achieving state-of-the-art classification and clustering performance. We validate TAVRNN on three datasets: (1) electrophysiological recordings from a freely behaving rat, (2) primate somatosensory cortex recordings during reaching task, and (3) biological neurons in the DishBrain platform interacting with a virtual game environment. TAVRNN outperforms state-of-the-art dynamic embedding techniques and reveals previously unreported relationships between adaptive behavior and evolving network topology. These findings establish TAVRNN as a powerful, scalable, generalizable, and modality-agnostic approach for modeling neural dynamics across experimental and synthetic biological systems, with applicability across neural recording platforms and behavioral paradigms.
Related
- Neuronal FICD shapes motivated behavior through modulation of dopamine receptor dynamics
- DAG-HEART: Directed Acyclic Graph-Guided Health Equity-Aware Representation Transfer Learning Framework for Breast Cancer
- GRASP: Graph Representation Learning with Assay Supervision for Molecular Properties
- Extracting interpretable single-cell metabolic states with graph-guided representation learning
- Cross-species representation learning aligns mouse and human neural dynamics and tracks clinical drug efficacy
BibTeX
@article{khajehnejad2026graphbased,
title = {{Graph-based representation learning of neuronal dynamics and behavior}},
author = {M. Amin Khajehnejad and Forough Habibollahi and Ahmad Khajehnejad and Chris R. French and Brett J. Kagan and Adeel Razi},
journal = {Communications Biology},
year = {2026},
month = jul,
doi = {10.1038/s42003-026-10721-y},
url = {https://doi.org/10.1038/s42003-026-10721-y},
note = {CC BY},
}
RIS
TY - JOUR
TI - Graph-based representation learning of neuronal dynamics and behavior
AU - M. Amin Khajehnejad
AU - Forough Habibollahi
AU - Ahmad Khajehnejad
AU - Chris R. French
AU - Brett J. Kagan
AU - Adeel Razi
PY - 2026
DA - 2026/07/29
JO - Communications Biology
DO - 10.1038/s42003-026-10721-y
UR - https://doi.org/10.1038/s42003-026-10721-y
AB - Understanding how neuronal networks reorganize in response to stimuli and give rise to behavior is a central challenge in neuroscience and artificial intelligence. Existing methods, however, often struggle to capture the evolving structure of neural connectivity and its relationship to behavior, particularly in dynamic, uncertain, or high-dimensional settings with sufficient resolution and interpretability. We introduce the Temporal Attention-enhanced Variational Graph Recurrent Neural Network (TAVRNN), a framework for modeling time-varying neuronal connectivity by integrating probabilistic graph learning with temporal attention. TAVRNN learns single-unit latent dynamics while maintaining interpretable population-level representations, enabling identification of behaviorally relevant connectivity patterns. It generalizes across neural systems differing in scale, behavioral structure, and supervision regime, achieving state-of-the-art classification and clustering performance. We validate TAVRNN on three datasets: (1) electrophysiological recordings from a freely behaving rat, (2) primate somatosensory cortex recordings during reaching task, and (3) biological neurons in the DishBrain platform interacting with a virtual game environment. TAVRNN outperforms state-of-the-art dynamic embedding techniques and reveals previously unreported relationships between adaptive behavior and evolving network topology. These findings establish TAVRNN as a powerful, scalable, generalizable, and modality-agnostic approach for modeling neural dynamics across experimental and synthetic biological systems, with applicability across neural recording platforms and behavioral paradigms.
SN - 2399-3642
ER -