A connectome manipulation framework for SONATA circuits
- Introduction
- Install
- Framework overview
- How to run
- Examples
- Documentation
- How to contribute
- Citation
- Publications that use or mention Connectome-Manipulator
- Funding & Acknowledgment
An important way to study cortical function in silico lies in manipulations that are impossible to perform in vivo and in vitro. The purpose of the Connectome-Manipulator is to provide a general framework to apply and study various types of structural manipulations of a circuit connectome. The framework allows for rapid connectome manipulations of biophysically detailed network models in SONATA format, the standard for large-scale network models. Manipulations can be targeted to entire models, specific sub-networks, or even single neurons, ranging from insertion or removal of specific motifs to complete rewiring based on stochastic connectivity models at various levels of complexity. Important use cases include wiring a connectome from scratch in a biologically realistic way based on given features of connectivity, rewiring an existing connectome while preserving certain aspects of connectivity, and transplanting specific connectivity characteristics from one connectome to another. The resulting connectomes can be readily simulated using any simulator supporting SONATA, allowing systematic and reproducible characterization of causal effects of manipulations on network activity.
From PyPI (recommended):
pip install connectome-manipulator
Alternatively, from GitHub:
git clone https://github.com/BlueBrain/connectome-manipulator.git cd connectome-manipulator pip install .
All dependencies declared in setup.py and are available from PyPI, including one optional dependency, mpi4py (v3.1.4), which is required for parallel processing, i.e., to run parallel-manipulator. Another optional dependency, parquet-converters (v0.8.0 or higher), required for converting .parquet output files to SONATA must be installed separately, see instructions under https://github.com/BlueBrain/parquet-converters.
Recommended Python version: v3.10.8
The software famework is intended to be used on Linux/MacOS-based systems! Specifically, some dependencies, like libsonata, are currently not compatible with Microsoft Windows OS.
The Connectome-Manipulator framework is illustrated in Figure 1 and consists of the following main components:
- Connectome manipulatorAs specified in the config, applies one or a sequence of manipulations to a given SONATA connectome, and writes the manipulated connectome to a new SONATA edges file. All manipulations are separately implemented in sub-modules and can be easily extended.Details can be found in the corresponding README file in the repository: connectome_manipulation/README.md
- Model buildingAs specified in the config, builds a model from a given connectome and writes the model to a file to be loaded and used by specific manipulations requiring a model (e.g., model-based rewiring based on connection probability model). All models are separately implemented in sub-modules and can be easily extended.Details can be found in the corresponding README file in the repository: model_building/README.md
Notes:
- Some models may not even require a connectome as input.
- Some models may depend on other models as input for model building.
- Structural comparatorAs specified in the config, performs a structural comparison of the original and manipulated connectomes. Different structural parameters to compare (connection probability, synapses per connection, ...) are separately implemented in sub-modules and can be easily extended.Details can be found in the corresponding README file in the repository: connectome_comparison/README.md
The structure of the respective configuration files can be found under doc/source/config_file_structure.rst
ℹ️ More details can be also found in the accompanying publication (esp. Supplementary tables), see Citation.
Figure 1: Schematic overview of the connectome manipulation framework, consisting of the "connectome manipulator", "model building", and "structural comparator" components.
As illustrated in Figure 2, the synapses of the connectome (SONATA edges) are divided into k splits targeting disjoint sets of N post-synaptic neurons (SONATA nodes), which reduces the memory consumption and facilitates parallelization on multiple computation nodes. Each split is an edge table loaded as Pandas dataframe and comprising a list of synapses together with all synapse properties, an example is shown in Figure 3. The manipulations are then applied separately to each edge table in sequence (or alternatively, in parallel), resulting in manipulated edge tables which are then written to separate .parquet files. In the end, all .parquet files are merged into one manipulated SONATA connectome file using parquet-converters.
Figure 2: Operation principle of the "connectome manipulator", illustrating its split-based processing architecture.
Figure 3: Example of an edge table (Pandas dataframe), comprising a list of synapses together with all synapse properties.
Notes:
- Manipulations can only be applied to a single SONATA edges population at a time.
- The synapses in each loaded edge table are assumed to be sorted by post-synaptic neuron ID. Likewise, the manipulated edges tables are to be returned with synapses sorted by post-synaptic neuron ID.
- Optionally, processing can be resumed from an earlier (incomplete) run, by re-using all .parquet files that already exist instead of re-computing them.
- By default, all .parquet files will be deleted after successfull completion, i.e., after the manipulated SONATA connectome file has been generated. Optionally, these temporary .parquet files can be kept as well.
Usage: connectome-manipulator [OPTIONS] COMMAND [ARGS]... Connectome manipulation tools. Options: --version Show the version and exit. -v, --verbose -v for INFO, -vv for DEBUG [default: 0] --help Show this message and exit. Commands: build-model Extract and build models from existing connectomes. compare-connectomes Compare connectome structure of two circuits. manipulate-connectome Manipulate or build a circuit's connectome.
Usage: connectome-manipulator manipulate-connectome [OPTIONS] CONFIG
Manipulate or build a circuit's connectome.
Options:
--output-dir PATH Output directory. [required]
--profile Enable profiling.
--resume Resume from exisiting .parquet files.
--keep-parquet Keep temporary parquet files.
--convert-to-sonata Convert parquet to sonata and generate circuit
config
--overwrite-edges Overwrite existing edges file
--splits INTEGER Number of blocks, overwrites value in config file
[default: 0]
--target-payload INTEGER Number of gid-gid pairs to consider for one block.
Supersedes splits when a parquet based
configuration is used [default: 20000000000]
--parallel Run using a parallel DASK job scheduler
-a, --parallel-arg TEXT Overwrite the arguments for the Dask Client with
key=value
--help Show this message and exit.
Just running serially you can do something like this:
connectome-manipulator -v manipulate-connectome wiring_config.json \
--output-dir PATH_TO_OUTPUT --profile --convert-to-sonata --splits 1
Running splits in parallel (with Dask) you can use the parallel-manipulator executable that will set up Dask automatically (and switch the --parallel flag by default to True, too):
srun --nodes 10 --tasks-per-node=2 --cpus-per-task=20 --constraint=clx --mem=0 \
parallel-manipulator -v manipulate-connectome wiring_config.json \
--output-dir PATH_TO_OUTPUT --profile --convert-to-sonata --splits=100
Please note that this feature will require at least 4 MPI ranks. Dask will use 2 ranks to manage the distributed cluster. We recommend to use a high number for --cpus-per-task to create Dask workers that will be able to process a lot of data in parallel.
When processing with parallel-manipulator, one may pass the flag --target-payload to determine how big the individual workload for each process should be. The default value of 20e9 was determined empirically to run on the whole mouse brain with 75 million neurons. We recommend to use this value as a starting point and scale it up or down to achieve the desired runtime characteristics.
Details on the CONFIG file structure can be found under doc/source/config_file_structure.rst
Errors may occur for many different reasons and are not always easy to track. Most common errors are that an allocation gets "killed", either due to a time limit or due to an out-of-memory error. Here we provide a few hints on how to avoid or track errors that may occur:
- Use the "verbose" mode (
-vflag) which will produce a lot of log output. - Look into the log files: there is usually one master log file and individual log files for all data splits, all of which can be found in the
/logssubfolder of the output circuit folder. - Use a small connectome to start with.
- Use a simple operation to start with, such as
null_manipulation(see examples). - Run serially to start with, before switching to parallel processing.
- Start with a single data split.
- But: In case of memory errors, use more than a single data splits, even when running serially (!), which will reduce the memory consumption as individual splits will be processed one after the other.
- When running in parallel, use
--tasks-per-nodein the SLURM configuration to define how many tasks (=splits) will be executed on a single node; reducing this number may reduce the risk of out-of-memory errors. - In general: Increasing memory allocation and/or allocation time may help.
- For high performance: Allocate many nodes and use parallel processing together with a relatively large number of data splits depending on the network size (i.e., aim for a few hundered post-synaptic neurons per data split).
Usage: connectome-manipulator build-model [OPTIONS] CONFIG
Extract and build models from existing connectomes.
Options:
--force-reextract Force re-extraction of data, in case already existing.
--force-rebuild Force model re-building, in case already existing.
--cv-folds INTEGER Optional number of cross-validation folds, overwrites
value in config file
--help Show this message and exit.
Details on the CONFIG file structure can be found under doc/source/config_file_structure.rst
Usage: connectome-manipulator compare-connectomes [OPTIONS] CONFIG
Compare connectome structure of two circuits.
Options:
--force-recomp-circ1 Force re-computation of 1st circuit's comparison data,
in case already existing.
--force-recomp-circ2 Force re-computation of 2nd circuit's comparison data,
in case already existing.
--help Show this message and exit.
Details on the CONFIG file structure can be found under doc/source/config_file_structure.rst
Examples can be found under examples/ in the repository.
The full documentation (API reference, CONFIG file structure, ...) can be found on Read the Docs.
Contribution guidelines can be found in CONTRIBUTING.md in the repository.
If you use this software, we kindly ask you to cite the following publication:
Christoph Pokorny, Omar Awile, James B. Isbister, Kerem Kurban, Matthias Wolf, and Michael W. Reimann (2024). A connectome manipulation framework for the systematic and reproducible study of structure–function relationships through simulations. Network Neuroscience. DOI: 10.1162/netn_a_00429
@article{pokorny2024connectome,
author = {Pokorny, Christoph and Awile, Omar and Isbister, James B and Kurban, Kerem and Wolf, Matthias and Reimann, Michael W},
title = {A connectome manipulation framework for the systematic and reproducible study of structure--function relationships through simulations},
journal = {Network Neuroscience},
year = {2024},
publisher={MIT Press},
doi = {10.1162/netn_a_00429}
}
- Michael W. Reimann, Sirio Bolaños-Puchet, Jean-Denis Courcol, Daniela Egas Santander, et al. (2024) Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part I: Anatomy. eLife, 13:RP99688. DOI: 10.7554/eLife.99688.1
- James B. Isbister, András Ecker, Christoph Pokorny, Sirio Bolaños-Puchet, Daniela Egas Santander, et al. (2024) Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part II: Physiology and Experimentation. eLife, 13:RP99693. DOI: 10.7554/eLife.99693.1
- Daniela Egas Santander, Christoph Pokorny, András Ecker, Jānis Lazovskis, Matteo Santoro, Jason P. Smith, Kathryn Hess, Ran Levi, and Michael W. Reimann. (2024) Heterogeneous and higher-order cortical connectivity undergirds efficient, robust and reliable neural codes. bioRxiv, 2024.03.15.585196. DOI: 10.1101/2024.03.15.585196
- Christoph Pokorny, Sirio Bolaños-Puchet, András Ecker, James B. Isbister, Michael Gevaert, Joni Herttuainen, Henry Markram, and Michael W. Reimann. Impact of simplified network structure on cortical activity. Bernstein Conference, 2022, Berlin.
- Kerem Kurban, Christoph Pokorny, Julian Budd, Alberto Antonietti, Armando Romani, and Henry Markram. Topological properties of a full-scale model of rat hippocampus CA1 and their functional implications. Annual meeting of the Society for Neuroscience, 2022, San Diego.
- Christoph Pokorny, Omar Awile, Sirio Bolaños-Puchet, András Ecker, Daniela Egas Santander, James B. Isbister, Matthias Wolf, Henry Markram, and Michael W. Reimann. A connectome manipulation framework for the systematic and reproducible study of the structure-function relationship through simulations. Bernstein Conference, 2023, Berlin.
- Christoph Pokorny, Omar Awile, James B. Isbister, Kerem Kurban, Matthias Wolf, and Michael W. Reimann. A connectome manipulation framework for the systematic and reproducible study of structure-function relationships through simulations. FENS Forum, 2024, Vienna.
- Christoph Pokorny, Omar Awile, James B. Isbister, Kerem Kurban, Matthias Wolf, and Michael W. Reimann. A connectome manipulation framework for the systematic and reproducible study of structure-function relationships through simulations. Bernstein Conference, 2024, Frankfurt.
- Daniela Egas Santander, Christoph Pokorny, András Ecker, Jānis Lazovskis, Matteo Santoro, Jason P. Smith, Kathryn Hess, Ran Levi, and Michael W. Reimann. Efficiency and reliability in biological neural network architectures. Annual meeting of the Society for Neuroscience, 2024, Chicago.
The development of this software was supported by funding to the Blue Brain Project, a research center of the École polytechnique fédérale de Lausanne (EPFL), from the Swiss government’s ETH Board of the Swiss Federal Institutes of Technology.
Copyright (c) 2024 Blue Brain Project/EPFL