Deep Learning-Based Cancer Drug Response Prediction with ScreenDL

Requirements

ScreenDL was developed using Python 3.9.13 and requires the following packages (see Installation):

numpy (>= 1.21)
pandas (>= 2.0.3)
openpyxl (== 3.1.2)
tensorflow (== 2.11.1)
tensorflow-probability (== 0.19.0)
scikit-learn (== 1.3.0)
omegaconf (>= 2.2, < 2.4)
tqdm
rdkit
deepchem (== 2.7.1)
cdrpy
scipy (== 1.8.1)

Additional Hardware & Software Requirements

Hardware

All ScreenDL models were trained on GPU nodes provided by the Utah Center for High Performance Computing equipped with either NVIDIA GTX 1080 Ti GPUs with 3584 CUDA cores and 11 GB GDDR5X memory or NVIDIA A40 GPUs with 10,752 CUDA cores and 48 GB GDDR6 memory using cuda/11.3 and cudnn/8.2.0. We note that ScreenDL can be trained using standard CPUs and does not require GPU hardware.

Git Large Files Storage (Git LFS)

In order to download and extract the datasets packaged with this library, you will need Git Large File Storage (Git LFS) installed. For install instructions, visit: Installing Git Large File Storage

Installation

Follow the steps below to setup a python virtual environment and install screendl. Please note that most of ScreenDL's dependencies are installed automatically, however, tensorflow, tensorflow-probability, and our cdrpy library require manual installation as outlined below (<15 minutes):

Setup a new conda environment:

conda create --name screendl-env python=3.9.13
conda activate screendl-env

Install tensorflow and tensorflow-probability:

pip install tensorflow==2.11.1
pip install tensorflow-probability==0.19.0

Install cdrpy:

git clone https://github.com/csederman/cdrpy.git && cd cdrpy
pip install --upgrade .

Install screendl:

cd ..
git clone https://github.com/csederman/screendl.git && cd screendl
pip install --upgrade ".[scripts]"

Running ScreenDL

ScreenDL uses Hydra for config management. Before running ScreenDL, follow the instructions below to update the demo config files and prepare the input data.

Data Preparation

All datasets used to train and evaluate ScreenDL can be found as tar.gz archives under the data/datasets directory. To prepare the data for training ScreenDL, enter the data/datasets directory (cd data/datasets) and unpack that tar.gz archives with the following commands:

cd data/datasets
mkdir <dataset name>
tar -xvzf <path to dataset archive> -C <dataset name>

For example, to unpack the CellModelPassports-GDSCv1v2 dataset, run:

cd data/datasets
mkdir CellModelPassports-GDSCv1v2
tar -xvzf CellModelPassports-GDSCv1v2.tar.gz -C CellModelPassports-GDSCv1v2

Configuring ScreenDL

ScreenDL's inputs and hyperparameters are configured using Hydra. Config files for all scripts can be found under the conf subdirectory. ScreenDL's default configuration assumes that datasets have been extracted into the data/datasets directory by following the steps outlined in Data Preparation above. If you have extracted the data according to the procedure outlined above, no additional configuration is required and you may skip this section and procede directly to Training and Evaluation.

ScreenDL can also be manually configured by updating the necessary file paths. In what follows, we outline the required manual config file updates for basic ScreenDL functionality using the CellModelPassports-GDSCv1v2 dataset as an example. Note that in order to run sweeps (multiruns) using Hydra with the default config files packaged with ScreenDL, you will need to include the additional command line arguement _datastore_="$(pwd)/data" as documented in the example commands in later sections.

Update the file paths in a given script's config file

All ScreenDL scripts are configured with a _datastore_ field that specifies the root directory for ScreenDL's outputs (and optionally inputs).
This _datastore_ field must be updated to point to the appropriate directory on the user's system.
For example, to save ScreenDL's outputs under /<path to screendl repo>/screendl/data the corresponding _datastore_ field for a given config should be set to /<path to screendl repo>/screendl/data.
This step tells ScreenDL where outputs should be stored.

Update the corresponding dataset config file with the appropriate file paths

We use Hydra's nested config system to manage the configuration of multiple datasets simultaneously.
In order for ScreenDL to read a given dataset (i.e., CellModelPassports-GDSCv1v2), the config field dataset.dir must point to the location of the corresponding extracted tar.gz archive.
Note that we recommend using expanded file paths when updating config files.
For example, for CellModelPassports-GDSCv1v2, set the dir field under conf/runners/dataset/CellModelPassports-GDSCv1v2.yaml to point to the root directory of the extracted dataset.
For example, if you extracted the CellModelPassports-GDSCv1v2 dataset to /<path to screendl repo>/screendl/data/datasets/CellModelPassports-GDSCv1v2 set the dir field in conf/runners/dataset/CellModelPassports-GDSCv1v2.yaml to /<path to screendl repo>/data/datasets/CellModelPassports-GDSCv1v2.
This step tells ScreenDL where input data is stored.

Training and Evaluation

Note that, as detailed in Installation, running benchmarking experiments requires the additional dependencies installed by running pip install .[scripts]. To perform basic training and evaluation of ScreenDL, run:

python scripts/runners/run.py

The typical CPU runtime for basic training of ScreenDL is <30min for a single train/test split.

This script will generate several output files under the directories configured above:

predictions.csv: The raw predictions generated for each tumor-drug pair. This file is used as input for subsequent analyses.
scores.json: The train/validation/test metrics for the ScreenDL model.

To train and evaluate ScreenDL on all train/val/test folds, run:

python scripts/runners/run.py -m _datastore_="$(pwd)/data"

Running with ScreenAhead

To train and evaluate ScreenDL with ScreenAhead tumor-specific fine-tuning, run:

python scripts/runners/run_screenahead.py

The typical runtime for training of ScreenDL with ScreenAhead is <30min for a single train/test split.

This script will generate several output files under the directories configured above:

predictions.csv: The raw predictions generated by ScreenDL for each tumor-drug pair. This file is used as input for subsequent analyses.
predictions_sa.csv: The raw predictions generated by ScreenDL-SA (ScreenDL with tumor-specific fine-tuning) for each tumor-drug pair. This file is used as input for subsequent analyses.
scores.json: The train/validation/test metrics for the ScreenDL model.

To train and evaluate ScreenDL with ScreenAhead tumor-specific fine-tuning on all train/val/test folds, run:

python scripts/runners/run_screenahead.py -m _datastore_="$(pwd)/data"

Running PDX/PDXO Experiments

PDXO Validation

To run the full PDXO validation pipeline, run the following command:

python scripts/experiments/pdxo_validation.py -m _datastore_="$(pwd)/data"

This command will train an ensemble of 10 ScreenDL models using different subsets of cell lines and both perform domain-specific fine-tuning and ScreenAhead tumor-specific fine-tuning under leave-one-out cross-validation. To train a single model, run:

python scripts/experiments/pdxo_validation.py

PDX Validation

To run the PDX validation pipeline, run the following command:

python scripts/experiments/pdx_validation.py -m _datastore_="$(pwd)/data"

This command will train an ensemble of 10 ScreenDL models using different subsets of cell lines and perform domain-specific fine-tuning and ScreenAhead tumor-specific fine-tuning under leave-one-out cross-validation. To train a single model, run:

python scripts/experiments/pdx_validation.py

Benchmarking Experiments

Note that running benchmarking experiments requires the additional dependencies installed by running pip install .[scripts].

HiDRA

Cell line benchmarking

Clone the HiDRA repo and checkout the cdrpy-benchmarking branch.

git clone https://github.com/csederman/HiDRA.git && cd HiDRA
git checkout cdrpy-benchmarking

To train HiDRA on a single fold:

HIDRA_ROOT="<path to HiDRA repo>" python scripts/runnners/run.py \
    model=HiDRA-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

To run HiDRA on all training folds, use:

HIDRA_ROOT="<path to HiDRA repo>" python scripts/runnners/run.py -m \
    _datastore_="$(pwd)/data" \
    model=HiDRA-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

PDXO/PDX benchmarking

HIDRA_ROOT="<path to HiDRA repo>" python scripts/experiments/pdx_benchmarking.py -m \
    _datastore_="$(pwd)/data" \
    model=HiDRA-legacy \
    dataset=CellModelPassports-GDSCv1v2-HCI \
    dataset.preprocess.norm=global

More about HiDRA

For more information on HiDRA, checkout the original publication: HiDRA: Hierarchical Network for Drug Response Prediction with Attention

DualGCN

Cell line benchmarking

Clone the DualGCN repo and checkout the cdrpy-benchmarking branch

git clone https://github.com/csederman/DualGCN.git && cd DualGCN
git checkout cdrpy-benchmarking

To train DualGCN on a single fold:

DUALGCN_ROOT="/<path to DualGCN repo>/code" python scripts/runnners/run.py \
    model=DualGCN-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

To run DualGCN on all training folds, use:

DUALGCN_ROOT="/<path to DualGCN repo>/code" python scripts/runnners/run.py -m \
    _datastore_="$(pwd)/data" \
    model=DualGCN-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

More about DualGCN

For more information on DualGCN, checkout the original publication: DualGCN: a dual graph convolutional network model to predict cancer drug response

DeepCDR

Cell line benchmarking

Clone the DeepCDR repo and checkout the cdrpy-benchmarking branch

git clone https://github.com/csederman/DeepCDR.git && cd DeepCDR
git checkout cdrpy-benchmarking

To train DeepCDR on a single fold:

DEEPCDR_ROOT="/<path to DeepCDR repo>/prog" python scripts/runnners/run.py \
    model=DeepCDR-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

To run DeepCDR on all training folds, use:

DEEPCDR_ROOT="/<path to DeepCDR repo>/prog" python scripts/runnners/run.py -m \
    _datastore_="$(pwd)/data" \
    model=DeepCDR-legacy \
    dataset=CellModelPassports-GDSCv1v2 \
    dataset.preprocess.norm=global

PDXO/PDX benchmarking

DEEPCDR_ROOT="/<path to DeepCDR repo>/prog" python scripts/experiments/pdx_benchmarking.py -m \
    _datastore_="$(pwd)/data" \
    model=DeepCDR-legacy \
    dataset=CellModelPassports-GDSCv1v2-HCI-Mutations \
    dataset.preprocess.norm=global

For more information on DeepCDR, checkout the original publication: DeepCDR: a hybrid graph convolutional network for predicting cancer drug response

Running with IMPROVE

Citing ScreenDL

[1]: Casey Sederman, Chieh-Hsiang Yang, Emilio Cortes-Sanchez, Tony Di Sera, Xiaomeng Huang, Sandra D. Scherer, Ling Zhao, Zhengtao Chu, Eliza R. White, Aaron Atkinson, Jadon Wagstaff, Katherine E. Varley, Michael T. Lewis, Yi Qiao, Bryan E. Welm, Alana L. Welm, Gabor T. Marth, A precision oncology-focused deep learning framework for personalized selection of cancer therapy, bioRxiv 2024.12.12.628190; doi: https://doi.org/10.1101/2024.12.12.628190

References

Models

[1]: Qiao Liu, Zhiqiang Hu, Rui Jiang, Mu Zhou, DeepCDR: a hybrid graph convolutional network for predicting cancer drug response, Bioinformatics, Volume 36, Issue Supplement_2, December 2020, Pages i911–i918, https://doi.org/10.1093/bioinformatics/btaa822

[2]: Ma, T., Liu, Q., Li, H. et al. DualGCN: a dual graph convolutional network model to predict cancer drug response. BMC Bioinformatics 23 (Suppl 4), 129 (2022). https://doi.org/10.1186/s12859-022-04664-4

[3]: Jin, I. & Nam, H. HiDRA: Hierarchical Network for Drug Response Prediction with Attention. J. Chem. Inf. Model. 61, 3858–3867 (2021). https://doi.org/10.1021/acs.jcim.1c00706

Name		Name	Last commit message	Last commit date
Latest commit History 138 Commits
conf		conf
data		data
pub		pub
screendl		screendl
scripts		scripts
.gitattributes		.gitattributes
.gitignore		.gitignore
LICENSE		LICENSE
MANIFEST.in		MANIFEST.in
README.md		README.md
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt
scripts-requirements.txt		scripts-requirements.txt
setup.cfg		setup.cfg
setup.py		setup.py
test-requirements.txt		test-requirements.txt

License

csederman/screendl

Folders and files

Latest commit

History

Repository files navigation

Deep Learning-Based Cancer Drug Response Prediction with ScreenDL

Contents

Requirements

Additional Hardware & Software Requirements

Hardware

Git Large Files Storage (Git LFS)

Installation

Running ScreenDL

Data Preparation

Configuring ScreenDL

Update the file paths in a given script's config file

Update the corresponding dataset config file with the appropriate file paths

Training and Evaluation

Running with ScreenAhead

Running PDX/PDXO Experiments

PDXO Validation

PDX Validation

Benchmarking Experiments

HiDRA

Cell line benchmarking

PDXO/PDX benchmarking

More about HiDRA

DualGCN

Cell line benchmarking

More about DualGCN

DeepCDR

Cell line benchmarking

PDXO/PDX benchmarking

Running with IMPROVE

Citing ScreenDL

References

Models

Datasets

About

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages