# Multi-scale spiking network model of macaque visual cortex [![www.python.org](https://img.shields.io/badge/python-3.6-blue.svg)](https://www.python.org) NEST simulated [![License: CC BY-NC-SA 4.0](https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc-sa/4.0/) The code for this model is maintained at https://inm-6.github.io/multi-area-model/ ![Model overview](/getmodelfile?model=262457&file=multi-area-model-master/model_construction.png) This code implements the spiking network model of macaque visual cortex developed at the Institute of Neuroscience and Medicine (INM-6), Research Center Jülich. The model has been documented in the following publications: 1. Schmidt M, Bakker R, Hilgetag CC, Diesmann M & van Albada SJ Multi-scale account of the network structure of macaque visual cortex Brain Structure and Function (2018), 223: 1409 [https://doi.org/10.1007/s00429-017-1554-4](https://doi.org/10.1007/s00429-017-1554-4) 2. Schuecker J, Schmidt M, van Albada SJ, Diesmann M & Helias M (2017) Fundamental Activity Constraints Lead to Specific Interpretations of the Connectome. PLOS Computational Biology, 13(2): e1005179. [https://doi.org/10.1371/journal.pcbi.1005179](https://doi.org/10.1371/journal.pcbi.1005179) 3. Schmidt M, Bakker R, Shen K, Bezgin B, Diesmann M & van Albada SJ (2018) A multi-scale layer-resolved spiking network model of resting-state dynamics in macaque cortex. PLOS Computational Biology, 14(9): e1006359. [https://doi.org/10.1371/journal.pcbi.1006359](https://doi.org/10.1371/journal.pcbi.1006359) The code in this repository is self-contained and allows one to reproduce the results of all three papers. A video providing a brief introduction of the model and the code in this repository can be found [here](https://www.youtube.com/watch?v=NGAqe78vmHY&t=22s). ## Python framework for the multi-area model The entire framework is summarized in the figure below: ![Sketch of the framework](/getmodelfile?model=262457&file=multi-area-model-master/framework_sketch.png) We separate the structure of the network (defined by population sizes, synapse numbers/indegrees etc.) from its dynamics (neuron model, neuron parameters, strength of external input, etc.). The complete set of default parameters for all components of the framework is defined in `multiarea_model/default_params.py`. A description of the requirements for the code can be found at the end of this README. -------------------------------------------------------------------------------- ### Preparations To start using the framework, the user has to define a few environment variables in a new file called `config.py`. The file `config_template.py` lists the required environment variables that need to specified by the user. Furthermore, please add the path to the repository to your PYTHONPATH: `export PYTHONPATH=/path/to/repository/:$PYTHONPATH`. -------------------------------------------------------------------------------- `MultiAreaModel` The central class that initializes the network and contains all information about population sizes and network connectivity. This enables reproducing all figures in [1]. Network parameters only refer to the structure of the network and ignore any information on its dynamical simulation or description via analytical theory. `Simulation` This class can be initialized by `MultiAreaModel` or as standalone and takes simulation parameters as input. These parameters include, e.g., neuron and synapses parameters, the simulated biological time and also technical parameters such as the number of parallel MPI processes and threads. The simulation uses the network simulator NEST (https://www.nest-simulator.org). For the simulations in [2, 3], we used NEST version 2.8.0. The code in this repository runs with a later release of NEST, version 2.14.0. `Theory` This class can be initialized by `MultiAreaModel` or as standalone and takes simulation parameters as input. It provides two main features: - predict the stable fixed points of the system using mean-field theory and characterize them (for instance by computing the gain matrix). - via the script `stabilize.py`, one can execute the stabilization method described in [2] on a network instance. Please see `figures/SchueckerSchmidt2017/stabilization.py` for an example of running the stabilization. `Analysis` This class allows the user to load simulation data and perform some basic analysis and plotting. ## Analysis and figure scripts for [1-3] The `figures` folder contains subfolders with all scripts necessary to produce the figures from [1-3]. If Snakemake (Köster J & Rahmann S, Bioinformatics (2012) 28(19): 2520-2522) is installed, the figures can be produced by executing `snakemake` in the respective folder, e.g.: cd figures/Schmidt2018/ snakemake Note that it can sometimes be necessary to execute `snakemake --touch` to avoid unnecessary rule executions. See https://snakemake.readthedocs.io/en/stable/snakefiles/rules.html#flag-files for more details. ## Running a simulation The files `run_example_downscaled.py` and `run_example_fullscale.py` provide examples. A simple simulation can be run in the following way: 1. Define custom parameters. See `multi_area_model/default_params.py` for a full list of parameters. All parameters can be customized. 2. Instantiate the model class together with a simulation class instance. M = MultiAreaModel(custom_params, simulation=True, sim_spec=custom_simulation_params) 3. Start the simulation. M.simulation.simulate() Typically, a simulation of the model will be run in parallel on a compute cluster. The files `start_jobs.py` and `run_simulation.py` provide the necessary framework for doing this in an automated fashion. The procedure is similar to a simple simulation: 1. Define custom parameters 2. Instantiate the model class together with a simulation class instance. M = MultiAreaModel(custom_params, simulation=True, sim_spec=custom_simulation_params) 3. Start the simulation. Call `start_job` to create a job file using the `jobscript_template` from the configuration file and submit it to the queue with the user-defined `submit_cmd`. Be aware that, depending on the chosen parameters and initial conditions, the network can enter a high-activity state, which slows down the simulation drastically and can cost a significant amount of computing resources. ## Extracting connectivity & neuron numbers First, the model class has to be instantiated: 1. Define custom parameters. See `multi_area_model/default_params.py` for a full list of parameters. All parameters can be customized. 2. Instantiate the model class. from multiarea_model import MultiAreaModel M = MultiAreaModel(custom_params) The connectivity and neuron numbers are stored in the attributes of the model class. Neuron numbers are stored in `M.N` as a dictionary (and in `M.N_vec` as an array), indegrees in `M.K` as a dictionary (and in `M.K_matrix` as an array). To extract e.g. the neuron numbers into a yaml file execute import yaml with open('neuron_numbers.yaml', 'w') as f: yaml.dump(M.N, f, default_flow_style=False) Alternatively, you can have a look at the data with `print(M.N)`. ## Simulation modes The multi-area model can be run in different modes. 1. Full model Simulating the entire networks with all 32 areas with default connectivity as defined in `default_params.py`. 2. Down-scaled model Since simulating the entire network with approx. 4.13 million neurons and 24.2 billion synapses requires a large amount of resources, the user has the option to scale down the network in terms of neuron numbers and synaptic indegrees (number of synapses per receiving neuron). This can be achieved by setting the parameters `N_scaling` and `K_scaling` in `network_params` to values smaller than 1. In general, this will affect the dynamics of the network. To approximately preserve the population-averaged spike rates, one can specify a set of target rates that is used to scale synaptic weights and apply an additional external DC input. 3. Subset of the network You can choose to simulate a subset of the 32 areas specified by the `areas_simulated` parameter in the `sim_params`. If a subset of areas is simulated, one has different options for how to replace the rest of the network set by the `replace_non_simulated_areas` parameter: - `hom_poisson_stat`: all non-simulated areas are replaced by Poissonian spike trains with the same rate as the stationary background input (`rate_ext` in `input_params`). - `het_poisson_stat`: all non-simulated areas are replaced by Poissonian spike trains with population-specific stationary rate stored in an external file. - `current_nonstat`: all non-simulated areas are replaced by stepwise constant currents with population-specific, time-varying time series defined in an external file. 4. Cortico-cortical connections replaced In addition, it is possible to replace the cortico-cortical connections between simulated areas with the options `het_poisson_stat` or `current_nonstat`. This mode can be used with the full network of 32 areas or for a subset of them (therefore combining this mode with the previous mode 'Subset of the network'). ## Test suite The `tests/` folder holds a test suite that tests different aspects of network model initalization and mean-field calculations. It can be conveniently run by executing `pytest` in the `tests/` folder: cd tests/ pytest ## Requirements Python 3, python\_dicthash ([https://github.com/INM-6/python-dicthash](https://github.com/INM-6/python-dicthash)), correlation\_toolbox ([https://github.com/INM-6/correlation-toolbox](https://github.com/INM-6/correlation-toolbox)), pandas, numpy, nested_dict, matplotlib (2.1.2), scipy, NEST 2.14.0 Optional: seaborn, Sumatra To install the required packages with pip, execute: `pip install -r requirements.txt` Note that NEST needs to be installed separately, see . In addition, reproducing the figures of [1] requires networkx, python-igraph, pycairo and pyx. To install these additional packages, execute: `pip install -r figures/Schmidt2018/additional_requirements.txt` In addition, Figure 7 of [1] requires installing the `infomap` package to perform the map equation clustering. See for all necessary information. Similarly, reproducing the figures of [3] requires statsmodels, networkx, pyx, python-louvain, which can be installed by executing: `pip install -r figures/Schmidt2018_dyn/additional_requirements.txt` The SLN fit in `multiarea_model/data_multiarea/VisualCortex_Data.py` and `figures/Schmidt2018/Fig5_cc_laminar_pattern.py` requires an installation of R and the R library `aod` (). Without R installation, both scripts will directly use the resulting values of the fit (see Fig. 5 of [1]). The calculation of BOLD signals from the simulated firing rates for Fig. 8 of [3] requires an installation of R and the R library `neuRosim` (). ## Contributors All authors of the publications [1-3] made contributions to the scientific content. The code base was written by Maximilian Schmidt, Jannis Schuecker, and Sacha van Albada with small contributions from Moritz Helias. Testing and review was supported by Alexander van Meegen. ## Citation If you use this code, we ask you to cite the appropriate papers in your publication. For the multi-area model itself, please cite [1] and [3]. If you use the mean-field theory or the stabilization method, please cite [2] in addition. We provide bibtex entries in the file called `CITATION`. If you have questions regarding the code or scientific content, please create an issue on github. NEST simulated        HBP logo        FZJ logo ## Acknowledgements We thank Sarah Beul for discussions on cortical architecture; Kenneth Knoblauch for sharing his R code for the SLN fit (`multiarea_model/data_multiarea/bbalt.R`); and Susanne Kunkel for help with creating Fig. 3a of [1] (`figures/Schmidt2018/Fig3_syntypes.eps`). This work was supported by the Helmholtz Portfolio Supercomputing and Modeling for the Human Brain (SMHB), the European Union 7th Framework Program (Grant 269921, BrainScaleS and 604102, Human Brain Project, Ramp up phase) and European Unions Horizon 2020 research and innovation program (Grants 720270 and 737691, Human Brain Project, SGA1 and SGA2), the Jülich Aachen Research Alliance (JARA), the Helmholtz young investigator group VH-NG-1028,and the German Research Council (DFG Grants SFB936/A1,Z1 and TRR169/A2) and computing time granted by the JARA-HPC Ver- gabegremium and provided on the JARA-HPC Partition part of the supercomputer JUQUEEN (Jülich Supercomputing Centre 2015) at Forschungszentrum Jülich (VSR Computation Time Grant JINB33), and Priority Program 2041 (SPP 2041) "Computational Connectomics" of the German Research Foundation (DFG).