RefDNN: a reference drug based neural network for more accurate prediction of anticancer drug resistance

We introduce a Reference Drug-based Neural Detwork model (RefDNN) for predicting anticancer drug resistance and identifying biomarkers contributing drug resistance.

RefDNN can predict accurately whether a cell line is resistant to a new drug on which the model was not trained.

RefDNN requires gene expression profiles and drug molecular fingerprint data.

For more detail, please refer to Choi. et al. "RefDNN: a reference drug based neural network for more accurate prediction of anticancer drug resistance." Scientific Reports 10 (2020):1-11

• Latest update: 21 April 2022

---

SYSTEM REQUIERMENTS:

• RefDNN requires system memory larger than 24GB.

• If you want to use tensorflow-gpu, GPU memory of more than 8GB is required.

---

Installation:

• We recommend to install via Anaconda (https://www.anaconda.com/)

• After installing Anaconda, please create a conda environment with the following commands:

git clone https://github.com/mathcom/RefDNN.git
cd RefDNN

~ only CPU ~
conda env create -f environment.yml

~ using GPU ~
conda env create -f environment_gpu.yml

---

Tutorials:

• We provide several source codes for tutorials.

  • 1_nested_cv_RefDNN.py : Nested Cross-validation for RefDNN evaluation

    • 1-1_nested_cv_elasticnet.py

    • 1-2_nested_cv_randomforest.py

  • 2_lococv_on_GDSC.py : Leave-one-Cancer-out Cross-validation for evaluating RefDNN on the GDSC dataset

  • 3_lodocv_on_GDSC.py : Leave-one-Drug-out Cross-validation for evaluating RefDNN on the GDSC dataset

  • 4_biomarker_identification.py : Biomarker identification of reference drugs using RefDNN trained by '1_nested_cv_RefDNN.py'

• These tutorial files are available for reproducibility purposes using the default values.

• An user can open them using the following commands:

conda activate RefDNN

python 1_nested_cv_RefDNN.py data/response_GDSC.csv data/expression_GDSC.csv data/fingerprint_GDSC.csv -o output_1_GDSC

python 1-1_nested_cv_elasticnet.py data/response_GDSC.csv data/expression_GDSC.csv data/fingerprint_GDSC.csv -o output_1-1_GDSC

python 2_lococv_on_GDSC.py data/response_GDSC.csv data/expression_GDSC.csv data/fingerprint_GDSC.csv data/cell_annotation_GDSC.txt -o output_2_lococv_GDSC

python 3_lodocv_on_GDSC.py data/response_GDSC.csv data/expression_GDSC.csv data/fingerprint_GDSC.csv -o output_3_lodocv_GDSC

python 4_biomarker_identification.py output_1_GDSC -o output_4_GDSC

conda deactivate

---

HELP MESSAGES:

• For each tutorial code, the option parameter '-h' shows help message.

• The following is the help message of '1_nested_cv_RefDNN.py':

$ python 1_nested_cv_RefDNN.py -h

usage: 1_nested_cv_RefDNN.py [-h] [-o outputdir] [-b batchsize]
                                  [-t numtrainingsteps]
                                  [-s numbayesiansearch] [-k outerkfold]
                                  [-l innerkfold] [-v verbose]
                                  responseFile expressionFile
                                  fingerprintFile

positional arguments:
  responseFile          A filepath of drug response data for TRAINING
  expressionFile        A filepath of gene expression data for TRAINING
  fingerprintFile       A filepath of fingerprint data for TRAINING

optional arguments:
  -h, --help            show this help message and exit
  -o outputdir          A directory path for saving outputs
                        (default:'output_1')
  -b batchsize          A size of batch on training process. The small size is
                        recommended if an available size of RAM is small
                        (default: 64)
  -t numtrainingsteps   Number of training steps on training process. It is
                        recommended that the steps is larger than (numpairs /
                        batchsize) (default: 5000)
  -s numbayesiansearch  Number of bayesian search for hyperparameter tuning
                        (default: 20)
  -k outerkfold         K for outer k-fold cross validation (default: 5)
  -l innerkfold         L for inner l-fold cross validation (default: 3)
  -v verbose            0:No logging, 1:Basic logging to check process, 2:Full
                        logging for debugging (default:1)

---

EXAMPLE:

• This example shows how to run the nested cross validation on the CCLE dataset.

• Since the default argument values of source code produce very very long process time(about 24 hours), we modify some values.

$ python 1_nested_cv_RefDNN.py data/response_CCLE.csv data/expression_CCLE.csv data/fingerprint_CCLE.csv -o output_1_CCLE -s 10 -t 1000 -b 32

• The example requires about 30 minutes if GPU is exploited.

• This example will make a directory named 'output_1_CCLE' and save a result file named 'metrics_hyperparameters.csv' in the directory.

• The result file contains 3 metrics and hyperparameter values computed by the nested cross validation.

• When running the example, logging information like below is found.

[START]
[DATA] NUM_PAIRS: 5724
[DATA] NUM_DRUGS: 12
[DATA] NUM_CELLS: 491
[DATA] NUM_GENES: 18926
[DATA] NUM_SENSITIVITY: 2322
[DATA] NUM_RESISTANCE: 3402
[TIME] [1] 2019-4-12 15:16:18
[TIME] [2] 2019-4-12 15:16:18
[OUTER] [1/5] NOW TUNING THE MODEL USING BAYESIAN OPTIMIZATION...
[OUTER] [1/5] BEST_HIDDEN_UNITS : 6
[OUTER] [1/5] BEST_LEARNING_RATE_FTRL : 1.000e-01
[OUTER] [1/5] BEST_LEARNING_RATE_ADAM : 6.827e-02
[OUTER] [1/5] BEST_L1_REGULARIZATION_STRENGTH : 7.324e+00
[OUTER] [1/5] BEST_L2_REGULARIZATION_STRENGTH : 1.052e-01
[OUTER] [1/5] BEST_TRAINING_ACCURACY : -0.908
[OUTER] [1/5] NOW TRAINING THE MODEL WITH BEST PARAMETERS...
[OUTER] [1/5] BEST_TEST_ACCURACY : 0.907
[OUTER] [1/5] BEST_TEST_AUCROC : 0.926
[OUTER] [1/5] BEST_TEST_AUCPR : 0.940
[TIME] [3] 2019-4-12 15:22:11
[OUTER] [2/5] NOW TUNING THE MODEL USING BAYESIAN OPTIMIZATION...

(...intermediate omission...)

[OUTER] [5/5] NOW TRAINING THE MODEL WITH BEST PARAMETERS...
[OUTER] [5/5] BEST_TEST_ACCURACY : 0.898
[OUTER] [5/5] BEST_TEST_AUCROC : 0.944
[OUTER] [5/5] BEST_TEST_AUCPR : 0.951
[TIME] [3] 2019-4-12 15:45:36
[TIME] [4] 2019-4-12 15:45:36
[FINISH]

---