rdaggs / neugenes

Brain mapping for uncovering new circuitry. Built for Ying-Hui Fu Neurogenetics lab at UCSF under Pu-Yun Shih

NeuGenes

Colocalization analysis, Brain Registration to Allen CCFv3, Protein signaling for axonal tracing

Installation

Clone the repo

git clone https://github.com/rdaggs/neugenes.git
cd neugenes

Then run (we recommend using a python virtual environment)

pip install --upgrade pip
set -e
python3.11 -m venv venv
source venv/bin/activate

Update the setuptools

# if errors arise with pip, try this
pip install --upgrade setuptools

Run the following command to integrate the neccessary build distribution of tensorflow 2.15 (dependent machine architecture):

pip install tensorflow-2.15.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
pip install tensorflow-2.15.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl 
pip install tensorflow-2.15.0-cp310-cp310-macosx_10_15_x86_64.whl
pip install tensorflow-2.15.0-cp310-cp310-macosx_12_0_arm64.whl 

Download preliminary structures (or wait and let processor interface directly with API for downloading the structures)

unzip resolution_25.zip -d model.config.mask_zip_file

Data Preprocessing

Some of the image cleaning and preprocessing is handled internally. The output of standard nikon microscopy equipment (xx.nd2) is not supported yet so you'll need to process and save as tiff, jpeg, png etc.

python model.flag_bad_scans.py

Given you have multiple datasets, condition your dataset to only compare the axially corresponding scans. i.e. if coronal axis of brain scans in control = [456,228,111] and stress = [461,219,21], this alternate preprocessing step can eliminate the final scan for a more precise comparison.

python model.condition_coronal_correlation.py [--dir DATASET_CONTROL] [--dir DATASET_TREATMENT] [--threshold 27] 

Inference

Run

python -u main.py [--dir DATASET_DIRECTORY] [--structures BLA,PVT,SCN.....]

for instance:

python -u main.py -u --input_dir Npsr1-cre_Ai14      --structures BLA,PVT,HPF,LS
python -u main.py -u --input_dir DREADD_m2.7_CONTROL --structures FULL_BRAIN

Please note, FULL_BRAIN requires about 2.5x as much time as LIGHT_BRAIN