What’s Changed

Added

• #43 Preprint is out now! Linking it in the documentation.
• #42 Added authors and licenses to the python scripts.
• #43 Added --no_hyperparameter_tuning flag for quick runs without hyperparameter tuning: hpam_split takes this as argument
• #43 Added --final_model_on_full data flag: if True, a final/production model is saved in the results directory. If hyperparameter_tuning is true, the final model is tuned, too. The model can later be loaded using the implemented load functions of the drevalpy models.
  • New process FINAL_SPLIT: splits the full dataset for each model class into train, validation, and optionally early stopping. This is done per model class and not overall because here, we no longer need across-model compatibility but want to train on the maximum amount of data (which might vary between models due to different feature availability)
  • New process TUNE_FINAL_MODEL: trains the final model(s) with all hyperparameter combinations
  • Added process EVALUATE_AND_FIND_MAX_FINAL: re-uses the EVALUATE_AND_FIND_MAX process to find the best hpam combination (evaluated on the validation dataset)
  • New process TRAIN_FINAL_MODEL: uses the best hpam combination to train the final model and save it
• #43 Added ProteomicsElasticNet, SingleDrugProteomicsRandomForest to list of known models
• #38 Reporting all package versions
• #38 Added UNZIP module for loading and unzipping the drug response datasets instead of handling this in LOAD_RESPONSE: UNZIP_RESPONSE, UNZIP_CS_RESPONSE (for cross-study datasets).
• #38 Added icon
• #30 Added the possibility of a leave-tissue-out (LTO) split

Changed

• #53 Changed to large runner for the GitHub Actions because of Docker → Singularity conversion.
• #42 Moved all publishDir directives to modules.config.
• #44 Fixed drevalpy versions in conda and docker to 1.3.5: now supporting Python 3.13
• #38 Support for AWS: changed the structure of load response and parameter check to conform more to Nextflow best practices.
• #44 Since drevalpy 1.3.5., the split_early_stopping function is no longer private.
• #39 Template update to version 3.3.1
• #38 Changed the defaults for test_mode from LPO to LCO and dataset_name from GDSC to CTRPv2 to better match the publication.
• #35 , #38 Introducing assets/NO_FILE for empty file handling in the visualization process.
• #30 Changed pipeline overview svg to Figure 1 from paper

Removed

• #30 Simplified visualization: multiple short processes were creating overhang → more efficient in one process.
• #44 Removed the --no_refitting parameter in load_response. It was no longer needed because of the new, more nextflow-y preprocess workflow
• #44 Removed redundant code in the visualization python script. Possible because of a new wrapper function in drevalpy 1.3.5.
• #38 Removed PARAMS_CHECK process: now handled by the schema and the utils_nfcore_drugresponseeval_pipeline subworkflow.
• #38 Removed the --curve_curator flag which was true by default. It is now the no_refitting flag which is false by default.

Fixed

• #44 casting a path to a string in bin/consolidate_results.py for drevalpy 1.3.5 compatibility.
• #43 casting drug to str in bin/collect_results.py because there were issues if all drugs were pubchem IDs and were treated as numeric values.
• #43 forgot to add the dataset_name in bin/load_response.py, made the tissue identifier optional. This was causing problems for custom datasets.
• #38 passing rand_modes in quotes to bin/consolidate_results.py because otherwise, if more than one mode was passed, it was not recognized as a list.
• #30 Added the path to the data directory to COLLECT_RESULTS because from there, we get the drug and cell line names for visualization.
• #30 Fixed handling of when ‘None’ was passed as randomization mode to CONSOLIDATE_RESULTS.

Dependencies

Parameters

Full Changelog: https://github.com/nf-core/drugresponseeval/compare/1.0.0…1.1.0