Instructions

Step 0. Dependencies

This repository requires Python 3.8.x.

Install packages listed in requirements.txt through pip install -r requirements.txt. Ideally create a virtual environment or docker image. Read on creating virtual environments here: https://docs.python.org/3.8/library/venv.html

Step 1. Simulation

If simulation is not needed, set simulated in configs/main_config.py to True, and go to step 2.

If simulation is not needed, go to PropsSimulator and follow the instructions for simulation in readme.md.

Step 2. Deconvolution

To run the steps below, In configs/main_config.py -

Change path at reference to the simulated h5ad file (output of step 1)

(Note: you can use a shell file dissect.sh which combines all following commands.) To use that file you will first need to run chmod +x dissect.sh

2.1 Prepare data

Run python prepare_data.py

The processed data will be stored in a new folder datasets in the experiment_folder

2.2 Run deconvolution

Run python dissect.py

The deconvolution result (tab seperated txt files) will be stored at dissect_fractions_[model_number].txt in the experiment_folder.

The trained model weights will be stored at model_[model_number] in the experiment_folder

2.3 Ensemble

If ensemble predictions are neeeded,

Run python ensemble.py The averaged deconvolution result (tab seperated txt files) will be stored at dissect_fractions_ens.txt in the experiment_folder.

Step 3. Quality control

Step 3.1 Run explainer

This runs GradientExplainer from python package shap (https://shap-lrjball.readthedocs.io/en/latest/generated/shap.GradientExplainer.html). This allows to caculate approximated SHAP values to attribute contribution of each gene in computing fractions. It should be noted that these are approximations and may not be correct. Nevertheless, top genes per cell type based on SHAP values should be enriched for specific cell types.

Run python explain.py

The outputs (Expected values per celltype, Raw shapely values as a numpy pickle file and plots of shap values per celltype) will be stored at shap in the experiment_folder.

Step 4. Cell type specific gene expression per sample

Run python dissect_expr.py. It uses same config file as before. Please note that dependending on the number of cell types, this step can take quite some time. Use of GPU is recommended.