Refactored TCRP codebase with improved organization and additional code for data transformation and model selection. The original codebase can be accessed here. For high level questions regarding TCRP, check out the TCRP FAQ.
End-to-end run is quite completely glued together yet and will require a little bit of manual work.
This part of the pipeline is not automated yet. The raw data will need to be downloaded from DepMap, and the transformed data are generated in with a jupyter notebook tcrp/data_preparation/process_sanger_drug_cell_line.ipynb
. This notebook will generate a series of pickled files and numpy compressed files that the following steps will be dependent on.
The code should all be contained in prepare_complete_run.py
. This script will create a directory that contains all of commands to sweep through all hyperpameter for all of the specific drugs. The drugs analyzed correspond to the pickle files in data/cell_line_lists
. Code to generate the pickled files still need to be included in this repository. Feel free to edit the run_name
variable to change the run name.
After the code is generated, the slurm submission scrips are created in output/{RUN NAME}/MAML_cmd
. To submit all of the slurm scripts you can run the following:
ls run_MAML_drugs*.sh | awk '{k = "sbatch "$0""; system(k); print(k)}'
Edit prepare_complete_run.py
. Change run_mode
variable to baseline
to run generate_baseline_jobs.py
. In addition, point to the correct fewshot_data_path
. This is a directory that was created in the tcrp complete run. It's simply the fewshot training and testing dataset that was used in the complete run.
The results are all embedded as logs in output/{RUN NAME}/run-logs/{DRUG}/{TISSUE}
. The log will specify the selected epoch for that hyperparameter and the correspond test performance. Additional code will be needed to gather the best performance to select the final performance for task.