Improving therapeutic synergy score predictions with adverse effects using multi-task heterogeneous network embedding
The accepted version link: [https://academic.oup.com/bib/article/24/1/bbac564/6958504]
How to use it:
Our Test Environment (Windows or Linux, the higher environment version beyond this might be also feasible):
- Python 3.6.13
- Pytorch 1.10.2
- CUDA tool kit 11.3.1
- dgl-cuda11.3
Step1. Download our collected and compiled dataset from https://drive.google.com/file/d/11xEGdgZLXlcxUfjajbrA3eSgIMT0-evr/view?usp=sharing, and put it into your specified file folder as the original data folder. The original data folder is used to generate the model input for each independent repeat.
Main Raw Data:
- CCLE_expression.csv: Downloaded original DepMap gene expression data.
- drugcomb_alldruginfo_dict.pickle: Collected information of involved drugs.
- drugcomb_synergy_score.csv: Collected drug-drug-cell line synergy score samples.
- twosides_side_effect.csv: Collected drug-drug adverse effect samples.
- drug_target_interaction.csv: Collected drug-target interaction samples.
- target_target_interaction.csv: Collected target-target interaction samples.
Step2. Run data_processing_updated_1.ipynb and data_processing_updated_2.ipynb step/block by step/block to generate the model input for independent repeats from the Raw Data (in the original data folder). In addition, we have provided a group of data (generated by these two .ipynb files) for an independent repeat named as 'fold1' in above link).
Step3. Follow the instruction described in each following .py file, to read the model input generated by above two .ipynb files (e.g., files in provided 'fold1') and run&evaluate different variants described in the manuscript (after training the model, these .py files will automatically evaluate the model performance on the test set):
- NEW_HNEMA_evaluation_ECFP6_CCLE_withAE.py: the implementation of standard Muthene using ECFP6 + selected 60 cell lines (described by CCLE gene expression data) with the adverse effect prediction module.
- NEW_GIN_evaluation_ECFP6_CCLE_withAE.py: the implementation of the Muthene variant using GIN (encoding drug molecular graphs) + selected 60 cell lines (described by CCLE gene expression data) with the adverse effect prediction module.
- NEW_HNEMA_evaluation_ECFP6_CCLE_withoutAE.py: the implementation of Muthene variant using ECFP6 + selected 60 cell lines (described by CCLE gene expression data) without the adverse effect prediction module.
A running example (including training and evaluation):
python NEW_HNEMA_evaluation_ECFP6_CCLE_withAE.py --root-prefix './fold1/'
(--root-prefix: the file folder storing the generated model input, e.g., fold1)
How to predict different types of synergy scores:
There is a bulit-in hyper-parameter in each above .py file named predicted_te_type, set it to 1, 2, 3, 4 can make the model predict ZIP, Loewe, HSA, and Bliss, respectively.
Instead of the framework name in the manuscript (i.e., Muthene), we use HNEMA (Heterogeneous Network Embedding with Meta-path Aggregation) here to define the function.