Installing the pipeline

Clone the repo from github:

git clone [email protected]:znamlab/2p-preprocess.git

Next, navigate to the repo directory and run following commands to install the package and dependencies:

cd 2p-preprocess
conda env create -f environment.yml

conda activate 2p-preprocess
pip install -e .
pip install -U "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install optax==0.1.2
conda deactivate

This should install the dependencies and create conda environments for suite2p and for the repo itself. Environments are created in each users home directory. Note that the version of torch generated issues and the version of optax conflicted with python=3.8. The python version is enforced because suite2p does not want to bring their dependencies forward.

run_suite2p.sh and run_suite2p_gpu.sh contain example scripts that first runs the standard run_suite2p pipeline and then applies neuropil correction using the AST model. If running neuropil correction using the AST model, using a GPU node is recommended.

To start the slurm job, navigate to the 2p-preprocess directory. Put the steps you want to run to y, and the steps you don’t want to run to n, e.g.:

--run-suite2p n --run-neuropil y --run-dff y

and run thesbatch script, passing the session details as environment variables, e.g.:

sbatch --export=PROJECT=depth_mismatch_seq,SESSION=BRAC9057.4j_S20240517,CONFLICTS=overwrite,TAU=0.7 run_suite2p_gpu.sh

ASt model

The Asymmetric Student's t-model for neuropil correction is described here. The python implementation in this repository uses JAX for automatic differentiation and rapid GPU computation. If run on a node without GPU, it should revert to CPU.

About

Some code in this repository (extractdff_gmm, ast_model.py) originates from a different code base and is reused with permission of the original author, Maxime Rio.