multiDGD paper experiments

This is a separate repository from our tool multiDGD, collecting our code we used to produce the results presented in our manuscript. Experiments are ordered according to their appearance in the manuscript and contain code for training and analysing the model and for making the figures.

Since this was a large project, it is not possible to provide only a single bash script to execute a complete reproduction. In every experiment, however, you can find instructions on how to execute each step. There are bash scripts numbered according to the order they should be executed in. Please execute them from the folder they are in, otherwise the relative paths won't work.

If you would like to use multiDGD or scDGD, we recommend the package repositories multiDGD and scDGD.

!Note!

This repository is still a work in progress, as we are cleaning up the project to provide you with a well-structured pipeline for the experiments of our manuscript.

Running experiments

Instructions on the execution of our experiments can be found in experiments. The experiments are numbered according to their appearance in the manuscript. The bash scripts are numbered according to the order they should be executed in. Each subfolder contains more information on the experiments and what to do.

Download trained models and data

All models and processed datasets are available via Figshare. They were derived from the code presented here but are made compatible with the multiDGD package.

Download data and models:

import requests
import zipfile

figshare_url = 'https://api.figshare.com/v2/articles/23796198/files'
files = {
    'human_bonemarrow.h5ad.zip':'41740251',                # processed data 
    'dgd_human_bonemarrow.pt':'41735907',                  # model weights
    'dgd_human_bonemarrow_hyperparameters.json':'41735904' # hyperparameters
}

for file_name, file_id in files.items():
    file_url = f'{figshare_url}/{file_id}'
    file_response = requests.get(file_url).json()
    file_download_url = file_response['download_url']
    response = requests.get(file_download_url, stream=True)
    with open(file_name, 'wb') as f:
        for chunk in response.iter_content(chunk_size=8192):
            f.write(chunk)
    if file_name.endswith('.zip'):     
        with zipfile.ZipFile(file_name, 'r') as zip_ref:
            zip_ref.extractall('.')            

Load trained model:

import scanpy as sc
import multiDGD

# Load data
data = sc.read_h5ad('./human_bonemarrow.h5ad')
# Patch for broken data loader
data.obs['train_val_test'].to_csv('./_obs.csv')

# load model from the saved checkpoint
model = multiDGD.DGD.load(data=data, save_dir='./', model_name='dgd_human_bonemarrow')