SynapseCLR is a contrastive learning framework for navigating 3D electron microscopy data. A graphical overview of SynapseCLR pipeline and downstream applications is shown below:
The SynapseCLR repository is organized as follows:
<repo_root>/
├─ pytorch_synapse/ # SynapseCLR Python packages
├─ configs/ # Sample configuration files for pretraining SynapseCLR models
├─ scripts/ # Helper scripts
├─ notebooks/ # Notebooks for data preprocessing, interactive analysis, and reproducing paper figures
├─ data/ # (not included in GitHub; see below) Raw and processed 3D EM image chunks
├─ ext/ # (not included in GitHub; see below) External resources (e.g. other pretrained models)
├─ output/ # (not included in GitHub; see below) SynapseCLR outputs (pretrained models, extracted features, interactive analysis results)
└─ tables/ # (not included in GitHub; see below) Primary and derived resource tables
If you wish to explore the results, a good starting point is browsing notebooks in GitHub. If you wish to run the notebooks, you need to install pytorch_synapse and additionally download the contents of data, output, ext, and tables directories (not included in this GitHub repository, see below). Finally, if you wish to pretrain SynapseCLR on your own 3D EM image chunks (not necessarily synapses, mind you; mitochondria anyone?), please follow the instructions given in pytorch_synapse. You will need to preprocess your data as described in notebooks/01_data_preprocessing and maybe modify the code according to the organization of your raw dataset. Feel free to contact us should you run into any problems!
You can download SynapseCLR raw and preprocessed data, pretrained models, resource tables, and analysis results from the public SynapseCLR Terra workspace. Alternatively, the data can be directly downloaded from the following Google Bucket: gs://fc-212b2d2b-6b73-4461-87a0-62164cd9b59a. Please visit here to learn more about downloading data from Google buckets and here to learn more about Terra.
The data bucket includes the following contents:
<repo_root>/
├─ data/ # Raw and processed 3D EM image chunks
├─ ext/ # External resources (e.g. other pretrained models)
├─ output/ # SynapseCLR outputs (pretrained models, extracted features, interactive analysis results)
└─ tables/ # Primary and derived resource tables
The bioRxiv preprint for SynapseCLR can be found here. The BibTeX citation is as follows:
@article {Wilson2022.06.07.495207,
author = {Wilson, Alyssa M and Babadi, Mehrtash},
title = {Uncovering features of synapses in primary visual cortex through contrastive representation learning},
elocation-id = {2022.06.07.495207},
year = {2022},
doi = {10.1101/2022.06.07.495207},
URL = {https://www.biorxiv.org/content/early/2022/06/09/2022.06.07.495207},
eprint = {https://www.biorxiv.org/content/early/2022/06/09/2022.06.07.495207.full.pdf},
journal = {bioRxiv}
}
- Alyssa M. Wilson [email protected] (Icahn School of Medicine at Mount Sinai, New York, NY)
- Mehrtash Babadi [email protected] (Data Sciences Platform, Broad Institute, Cambridge, MA)