GitHub repository hosting the code used in the paper Robust Evaluation of Deep Learning-based Representation Methods for Survival and Gene Essentiality Prediction on Bulk RNA-seq Data.
Use conda and poetry to easily set up the repository :
conda create -n omics_rpz_env python=3.9
conda activate omics_rpz_env
pip install "poetry==1.4.0"
make install-all # or make install-M1 for M1 Macs
pre-commit installThe data used in our study is available for download on this bucket (WIP - will be available shortly).
Once downloaded, make sure to update the PREFIX variable in omics_rpz/constants.py to point to the location of downloaded folder benchmark_data.
Several tasks are available on the repository. The benchmark script can be run to perform CV or in a train/test settings.
python ./tools/benchmark.py task=survival_prediction_tcga_task prediction_model=cox_model representation_model=pca
python ./tools/benchmark.py task=gene_essentiality prediction_model=linear_regression representation_model=pcaHyperparameters per model per task are stored in the config files /conf/representation_model/ as nested dicts :
default:
_target_: omics_rpz.transforms.AutoEncoder
repr_dim: 128
hidden: [512, 256]
dropout: [0.5, 0.5]
bias: True
num_epochs: 10
batch_size: 32
learning_rate: 0.0005
split_data: False
device: cpu
TCGAPredictionTask:
num_epochs: 50
OS:
repr_dim: 64
dropout: [0, 0]
num_epochs: 100
batch_size: 128
By default, the default configuration is used and overriden with task-specific parameters if the task class name (given by the _target_ field in conf/task/ files) is an entry in the conf file. For TCGAPredictionTask, task-specific parameters are overriden by label-specific parameters as shown above.
Values given with command line arguments (e.g. python ./tools/benchmark.py representation_model=auto_encoder representation_model.hidden=[1024,512]) override both default and task-specific parameters.
To perform hyperparameter-tuning with Optuna on TCGA prediction tasks, modify the config.yaml file with the tuning options, and add the option --multirun in the command line.
One example on purity prediction task:
python ./tools/benchmark.py task=purity_prediction_tcga_task prediction_model=linear_regression representation_model=pca --multirunYou can also specify the hyperparameter range or set in the command line. For example:
python ./tools/benchmark.py task=purity_prediction_tcga_task prediction_model=linear_regression representation_model=pca 'representation_model.repr_dim=range(16, 511)' --multirun- Because of a hydra code choice (detailed here), to change samplers, 2 lines must be changes in the config file:
defaults / override /hydra/sweeper/samplerandhydra / sweeper / sampler / _target_. See config.yaml for more details. - To easily filter all trials of a given study in MLflow, we may set study_name=<descriptive_string>, either in the config.yaml file or via command line. Then, in MLflow, we filter using params.study_name="descriptive_string".
More info on the hydra-optuna plugin here
The repeated_holdout.py script executes benchmark.py as many times as necessary to finish a repeated holdout experiment.
This script performs num_repeated_holdout (10 is the current default value from config.yaml) runs of:
- Randomly split out train/val and test sets with a different
test_split_random_seed(fromconfig.yaml). - Perform hyperparameter search with cross-validation to find the best hyperparameters (executing our traditional
benchmark.py, with--multirunandtrain_test_no_cv=False). - Use the best hyperparameters to measure performance on the test set (using our traditional
benchmark.py, withtrain_test_no_cv=True).
These runs generate num_repeated_holdout test-set performance metrics. We report mean (SD) and 95% empirical confidence intervals.
Logs and .csv results from each experiment are kept in folders with the following format: outputs/repeated_holdout_{study_name}_{start_time}_{k}, where k is the run index (from 0 to num_repeated_holdout - 1). Final performance metrics are stored in outputs/repeated_holdout_{study_name}_{start_time}_test_metrics.
Example of how to run this pipeline (here only the HP hidden is searched over a space of two choices):
python ./tools/repeated_holdout.py task=survival_prediction_tcga_task prediction_model=mlp_prediction representation_model=auto_encoder +"representation_model.hidden=choice([1024], [512])" study_name="survival_repeated_holdout"You can use ray to // the run of experiments by seeds by specifying use_ray=True. Make sure to specify the following arguments to match your machine configuration:
num_cpus_for_ray: number of CPUs that you want to give access to Ray globally (summed over all repetitions). Try to leave room for buffer jobs by specifying around 12 CPUs less than that you have on your machine.num_gpus_for_ray: number of GPUs that you want to give access to Ray globally (summed over all repetitions).cpu_per_repeat: number of CPUs per // threadgpu_per_repeat: number of GPUs per // thread
By not specifying correclty these arguments, ray may struggle to launch in //. Make sure what you are requesting is consistent (no more CPU than what you have, no GPU if you don't have GPU or CUDA installed, NB_REPEATS*cpu_per_repeat < num_cpus_for_ray < full number of CPUS)
To track your results with MLFlow, run:
mlflow ui --backend-store-uri ./outputs/mlflow/ --port 8889Ray automatically starts a dashboard when ray.init() is called if the dependencies required for the dashboard are installed (which is the case in the latest lock file).
The dashboard by default is created on the port 8265.
You can add libraries to the dependencies of this project by using poetry
poetry add optunaThe results shown here are in whole or part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga. We thank Oussama Tchita, Omar Darwiche Domingues and Thomas Chaigneau for their valuable contributions and comments to strengthen our pipeline and coding best practices. We thank Gilles Wainrib for initial ideas and discussions, Nicolas Loiseau for his advice and statistical expertise, Floriane Montanari, Benoît Schmauch, Gilles Wainrib and Jean-Philippe Vert for their detailed proofreading and insightful comments.