BiG-DRP: Bipartite Graph-based Drug Response Predictor

Implementation of Bipartite Graph-represented Drug Response Predictor (BiG-DRP and BiG-DRP+) as described in:

David Earl Hostallero, Yihui Li, Amin Emad, Looking at the BiG picture: incorporating bipartite graphs in drug response prediction, Bioinformatics, Volume 38, Issue 14, 15 July 2022, Pages 3609–3620, https://doi.org/10.1093/bioinformatics/btac383

Dependencies

This repository has been tested on python 3.7. To install the dependencies run the following on the terminal

pip install -r requirements.txt

Running BiG-DRP

python main.py

Running BiG-DRP+

To run BiG-DRP+, you must first run BiG-DRP while specifying the results subfolder (--folder=<folder_name>). Then run BiG-DRP with the --weight_folder specified as the results subfolder in the previous run.

python main.py --mode=train --folder=big
python main.py --mode=extra --weight_folder=big --folder=big_plus

Additional Parameters

• --split: the type of data-splitting to use (lco or lpo, default: lco)
• --dataroot: the root directory of your data (file names for input files can me modified in utils/constants.py) (default: ../)
• --outroot: the root directory of your outputs (default: ./)
• --folder: subdirectory you want to save your outputs (optional)
• --weight_folder: subdirectory for the saved weights and encodings (for BiG-DRP+ only)
• --mode: train means BiG-DRP, extra means BiG-DRP+ (default: train)
• --seed: the seed number for 5-fold CV (default: 0)
• --drug_feat: type of drug feature (desc, morgan, or mixed, default: desc)
• --network_perc: percentile used for the bipartite graph threshold (default: 1)

Data Availability

Preprocessed data can be accessed here: https://dx.doi.org/10.6084/m9.figshare.20022947

Performance Metrics

Note that when you run main.py, the output performance metrics do not correspond to the ones we presented in the paper because main.py only shows the overall performance (i.e. performance for all (drug, CCL) pairs in the test set is calculated as a whole). In the paper, we calculated the performance per drug then averaged the per-drug performances. To run the per-drug calculation, use metrics/calculate_metrics.py. Example:

python metrics/calculate_metrics.py --folder=results/big_plus/ --outfolder=results/big_plus

Note: The corresponding article to this repository reports the effective RMSE. By default, our method z-scores the lnIC50 per drug prior to training. However, since we are interested in predicting the actual lnIC50, we have to scale the predictions back to scaled_prediction = prediction*drug_stdev + drug_mean. The effective RMSE is calculated using the (raw) lnIC50s and the scaled predictions. The code in this repository calculates the RMSE using the z-scored lnIC50s and the (raw) predictions.

BibTex Citation

@article{hostallero2022looking,
  title={Looking at the BiG picture: incorporating bipartite graphs in drug response prediction},
  author={Hostallero, David Earl and Li, Yihui and Emad, Amin},
  journal={Bioinformatics},
  volume={38},
  number={14},
  pages={3609--3620},
  year={2022},
  publisher={Oxford University Press}
}