The Structural Underpinnings of Functionally Identified Neuronal Assemblies

Data and analysis code for the manuscript The Structural Underpinnings of Functionally Identified Neuronal Assemblies (Wagner-Carena et al., Allen Institute, 2026). This project and repository are also internally referred to as "Hebb's Vision."

Overview

This repository contains the data and analysis code needed to reproduce the figures in the manuscript. Using a dataset that links in vivo optical physiology to postmortem electron-microscopy connectivity in mouse V1, we extract neural assemblies from higher-order correlations in activity and relate them to the underlying structural network. Key findings:

• Neurons in V1 organize into assemblies that reliably encode natural visual stimuli.
• Over a quarter of V1 neurons join no assembly and are less integrated in the network.
• We find no evidence for stronger direct excitatory connections within assemblies than between them.
• Structural cross-inhibition exceeds within-assembly inhibition, and is correlated with lower co-activity — suggesting assemblies are delineated by mutual inhibition.

Preprint

A preprint is available on bioRxiv: https://www.biorxiv.org/content/10.1101/2025.04.24.649900v2

Note: the bioRxiv version above does not yet reflect revisions made since May 2025.

Citation

If you use this code or data, please cite:

Wagner-Carena, J., Kate, S., Riordan, T., et al. Hebb's Vision: The Structural Underpinnings of Hebbian Assemblies bioRxiv 2025.04.24.649900. https://doi.org/10.1101/2025.04.24.649900

License

This project is licensed under the terms described in License.txt.

Data

The large data files (calcium imaging, assembly extractions, electron-microscopy connectivity, etc.) live under data_files/ and are tracked with Git LFS. You must have Git LFS installed before cloning so that these files download instead of small pointer placeholders:

git lfs install
git clone https://github.com/AllenInstitute/HebbsVision

Pulling the full dataset downloads approximately <X> GB.

Setup Part 1: Clone the Repository

Clone https://github.com/AllenInstitute/HebbsVision into a directory of your choice (see the Data section above — Git LFS must be installed first so the large files in data_files/ download correctly).

Setup Part 2: Environment Setup

First, ensure that anaconda (conda) is installed on your system. Then, to create the environment and install the conda and pip packages, run:

conda env create -n HebbsVision -f conda_files.yaml

In the case of failure installing NeuroAnalysisTools, due to the lack of a requirements.txt, you must create the environment first and then download NeuroAnalysisTools from https://github.com/zhuangjun1981/NeuroAnalysisTools. Once it exists in a local directory, you can install it locally from pip with pip install -e ./<path_to_NeuroAnalysisTools>. The same workaround can be employed with V1DD-Physiology (https://github.com/AllenInstitute/v1dd_physiology) if a similar error occurs.

The LSMM library must be installed via a local wheel file at the moment, and the .whl is included in the root of the repository. Install it with:

pip install lsmm_data-0.1.2-py2.py3-none-any.whl

Assembly Extraction and Hyperparameter Tuning (Optional)

Note that this step can take some time, so for those simply interested in reproducing our analysis of the extracted assemblies, this repository includes all the hyperparameter files that were generated using Scan 1–4 data following the instructions in Mölter et al. 2024: "Similarity Graph Clustering for Neural Assembly Detection" p. 175, and we also include the file generated from Scan 1–3 (the scan on which the remaining analysis is focused) using the selected best hyperparameters.

Figure 1

To recreate panel B of figure 1 (the only one displaying data from our analysis), open FigureCode/Figure1/Figure1.ipynb and execute all the cells in order. It will create files in the FigureCode/Figure1/ directory.

• Figure 1B uses FigureCode/Figure1/3D_Scan_Images.png, FigureCode/Figure1/Example_Activity_Traces.png, and FigureCode/Figure1/raster_plot_red.png.

Figure 2

To recreate panels B, C, D, and the data for panel E of figure 2, open FigureCode/Figure2/Figure2.ipynb and execute all the cells in order. It will create files in the FigureCode/Figure2/ directory.

• Figure 2B uses FigureCode/Figure2/Assemblies_Intersection_Upset_Plot.png.
• Figure 2C uses FigureCode/Figure2/Assemblies_Plotted_In_Recording_Space_Same_Plot.png.
• Figure 2D uses FigureCode/Figure2/Spatial_Distribution_of_Cells_by_Assembly.png.
• The data for Figure 2E are printed in the .ipynb notebook, following the cell where compare_assemblies_spatial_distribution is called.

Figure 3

To recreate all panels in Figure 3, open FigureCode/Figure3/Figure3.ipynb and execute all the cells in order. It will create files in the FigureCode/Figure3/ directory.

• Figure 3A uses FigureCode/Figure3/correlations_assemblies_vs_random_ensembles_raincould_plot.png
• Figure 3B uses FigureCode/Figure3/sparsity_with_Gini_coefficient_by_assembly_and_random_ensembles.png
• Figure 3C uses FigureCode/Figure3/oracle_scores_dff_all_sets_raincloud.png
• Figure 3D uses FigureCode/Figure3/assembly_balanced_clip_id_percentage_decoder_MLPClassifier.png
• Figure 3E uses FigureCode/Figure3/random_ensemble_balanced_clip_id_percentage_decoder_MLPClassifier.png
• Figure 3F uses FigureCode/Figure3/Trigger_Frame_Assembly_4_real_assemblies_mean_trigger_frame.png, FigureCode/Figure3/Trigger_Frame_Assembly_4_random_ensembles_mean_trigger_frame.png, and FigureCode/Figure3/Trigger_Frame_Assembly_4_difference_squared_in_mean_trigger_frame_assembly_minus_random.png

Figure 4

To recreate all panels in Figure 4, open FigureCode/Figure4/Figure4_Master.ipynb and execute all the cells in order. It will create files in the FigureCode/Figure4/ directory.

• Figure 4A uses FigureCode/Figure4/connectivity_plot_final.png
• Figure 4B uses FigureCode/Figure4/draft_figures/Betweenness_Centrality_All.png
• Figure 4C uses FigureCode/Figure4/draft_figures/Outdegree_Centrality_All.png
• Figure 4D uses FigureCode/Figure4/draft_figures/A_No_A_Prob_Conn_by_Conn_Type_v2.png and FigureCode/Figure4/draft_figures/Prob_Conn_by_Conn_Type_v2.png
• Figure 4E uses FigureCode/Figure4/draft_figures/Nonzero_PSD_by_Conn_with_side_plot.png
• Figure 4F uses FigureCode/Figure4/draft_figures/A_No_A_Prob_Conn_by_Conn_Type_E_Chains_v2.png and FigureCode/Figure4/draft_figures/Prob_Conn_by_Conn_Type_E_Chains_v2.png
• Figure 4G uses FigureCode/Figure4/draft_figures/Nonzero_PSD_by_Conn_E_Chain_with_side_plot.png
• Figure 4H uses FigureCode/Figure4/draft_figures/A_No_A_Prob_Conn_by_Conn_Type_I_Chains_v2.png and FigureCode/Figure4/draft_figures/Prob_Conn_by_Conn_Type_I_Chains_v2.png
• Figure 4I uses FigureCode/Figure4/draft_figures/Nonzero_PSD_by_Conn_I_Chain_with_side_plot.png

Additional tests referenced in the statistical table are printed out throughout the .ipynb files above as they are run.