PRnet: Predicting transcriptional responses to novel chemical perturbations using deep generative model

This repository hosts the official implementation of PRnet, a flexible and scalable perturbation-conditioned generative model predicting transcriptional responses to novel complex perturbations at bulk and single-cell levels.

Download model and datasets

We provide model pretrained weight and precessed dataset (LINCS_L1000 and Sci-Plex) for training and reproducibility.

If the links above fail to redirect, please download the corresponding data and pretrained model weights from: http://prnet.drai.cn/Statistics/. For the data files, download LINCS L1000 Chemical Perturbations and Sci-plex Chemical Perturbations from the Supplementary Dataset section. For the model weights, download the PRnet model from the Supplementary Dataset section. Then change the names of the datasets to Lincs_L1000.h5ad and Sci_Plex.h5ad respectively. If the dataset cannot be downloaded, please adjust your browser settings to allow insecure content on the current webpage by setting it to "Allow."

The links above are temporarily unavailable due to bandwidth limitations. Please download the corresponding data from .

To clone our model, install github and run:

git clone https://github.com/Perturbation-Response-Prediction/PRnet.git

Please download the datasets and store them in the dataset folder. Download the pretrained weights and store them in the checkpoint folder.

• PRnet/: contains the code for the model, the data, and documents.
• data: contains the utils for data processing.
• dataset: contains datasets.
• figure: contains notebooks for generate figures in our paper.
• img: contains the graphic abstract of PRnet.
• models: contains models of PRnet.
• trainer: contains trainer of PRnet.
• preprocessing: contains notebooks for processing the data.
• train and test: including test_lincs.py, train_lincs.py, train_sciplex.py and test_sciplex.py.

Step 1: Installation

We recommend using Anaconda or Dockerfile to creat environment for using PRnet. Please make sure you have installed pre-installation.

Setup the environment with Anaconda

We recommend using Anaconda to create a conda environment. You can create a python environment using the following command:

conda create -n PRnet python=3.7

Then, you can activate the environment using:

conda activate PRnet
pip install -r requirements.txt

Setup the environment with Docker

Alternatively, you can use Docker to set up the environment. Build the Docker image using the provided Dockerfile:

docker build -t prnet .

You can change the Docker image of PyTorch according to your cuda version in Dockerfile FROM pytorch/pytorch:1.12.1-cuda11.3-cudnn8-runtime.

Then run the Docker container:

docker run -dit --gpu all -v $(pwd):/workspace --name PRnet prnet

After the docker container is started, you can exit using exit. If you use the -d mode, docker will be placed in the background. You can enter the container through docker attach PRnet. After stopping the container, start the container with docker start PRnet. You also can use docker exec -it PRnet /bin/bash to enter the container.

Step 2: Test with demo datatset

Create result folders:

mkdir results
mkdir /results/lincs
mkdir /results/sciplex
mkdir /results/demo

Inference with demo dataset:

python test_demo.py --split_key demo_split
                    --data_path ./dataset/demo.h5ad
                    --results_dir  ./results/demo/

Step 3: Inference with custom datatset

Please see custom_data_preprocessing.ipynb to prepare your dataset. 'custom_data_preprocessing' is a demo which preprocesses the data from CCLE.

Step 4: Train and test with provided datatset

To train the L1000 dataset:

python train_lincs.py --split_key drug_split_4    

To test the L1000 dataset:

python test_lincs.py --split_key drug_split_4     

To train the Sci-plex dataset:

python train_sciplex.py --split_key drug_split_0         

To test the L1000 dataset:

python test_sciplex.py --split_key drug_split_0       

Step 5: Analysis the results

To analysis the L1000 dataset:

python analysis_lincs.py --split_key drug_split_4      

To analysis the Sci-plex dataset:

python analysis_sciplex.py --split_key drug_split_0            

Figures

References

paper: Xiaoning Qi, Lianhe Zhao, Chenyu Tian, Yueyue Li, Zhen-Lin Chen, Peipei Huo, Runsheng Chen, Xiaodong Liu, Baoping Wan, Shengyong Yang and Yi Zhao. Predicting transcriptional responses to novel chemical perturbations using deep generative model[J]. Nature Communications, 2024, 15(1): 1-19.

code:

License

This project is covered under the Apache 2.0 License.