PlaNet: Predicting population response to drugs via clinical knowledge graph

PyTorch implementation of PlaNet, a geometric deep learning tool for predicting population response to drugs. PlaNet provides a new clinical knowledge graph that captures the relations between disease biology, drug chemistry, and population characteristics. Using this knowledge graph, PlaNet can take a population and drugs to be applied (e.g., a clinical trial) as an input and predict the efficacy and safety of the drugs for the population. For a detailed description of the algorithm, please see our manuscript "Predicting population response to drugs via clinical knowledge graph".

1. Installation

Environment

Run the following commands to create a conda environment:

conda create -n planet python=3.8
source activate planet

pip install torch==1.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
pip install transformers==4.4.2 datasets==2.7.1 tensorboard==2.11.0 pandas wandb scikit-learn seqeval matplotlib pyyaml seaborn anndata scanpy
pip install setuptools==58.2.0 numpy==1.22.2

pip install torch-scatter==2.0.9 torch-sparse==0.6.12 -f https://pytorch-geometric.com/whl/torch-1.10.1+cu113.html
pip install torch-geometric==2.0.0
pip install ogb==1.3.0 --no-deps

After this installation, the tokenizers library version should be 0.10.3. In case you encounter an error about the tokenizers version, go to <conda_path>/envs/<env_name>/lib/python3.8/site-packages/transformers/dependency_versions_check.py and replace the if pkg == "tokenizers": ... block with if pkg == "tokenizers": continue.

The total install time should be within 10 minutes.

Hardware requirement: 100GB RAM and GPU of 40GB memory

Data

You can download all the data (knowledge graph, clinical trial dataset, models, etc.) from here (data.zip). Unzip this, which will create a ./data directory.

2. Demo

We provide a demo notebook for loading the PlaNet knowledge graph and clinical trials data, and running the PlaNet models:

notebooks/demo.ipynb

The expected run time should be ~10 minutes.

3. Model training

Go to ./gcn_models directory. We train models to predict the efficacy, safety, and potential adverse events of a clinical trial.

To train a model for efficacy prediction, run commands in

../scripts/train_efficacy.sh

To train a model for safety prediction, run commands in

../scripts/train_safety.sh

To train a model for adverse event prediction, run commands in

../scripts/train_ae.sh

4. If you want to use PlaNet models for new clinical trial data

Overview: Running our models to predict for new clinical trials involves two steps:

• (1) parse the trial (parsing_package/parse_trial.py) so that the trial data is preprocessed and linked to the PlaNet knowledge graph
• (2) run the models to obtain safety or efficacy predictions (notebooks/predict_for_new_clinial_trial.ipynb)

Specifically, to do this, follow the steps below:

• Download all the data and resources needed for trial data parsing from here (parsing_package.zip). Unzip this and put it in the ./parsing_package directory. Install the dependencies by following ./parsing_package/README
• Go to ./parsing_package directory and run parse_trial.py to process a new clinical trial (e.g., NCT02370680)
• Finally, go to ./notebooks directory and run predict_for_new_clinial_trial.ipynb to get AE, safety, and efficacy predictions for the new clinical trial.

Citation

If you find our code and research useful, please consider citing:

@article{planet2023,
  author =  {Maria Brbi{\'c} and Michihiro Yasunaga and Prabhat Agarwal and Jure Leskovec},
  title =   {Predicting population response to drugs via clinical knowledge graph},
  year =    {2023},  
}