This site is a source code repository for in silico CRISPR/RNAi and deep learning for prediction of cancer vulnerabilities from transcriptome data.

• Requirements
• In silico CRISPR/RNAi
• Deep learning

Required modules for CRISPR/RNAi simulation.

• Python (>= 2.7.12): https://www.python.org/
• NumPy (>= 1.13.3): http://www.numpy.org/
• Pandas (>= 0.23.1): https://pandas.pydata.org/
• Networkx (>= 2.1) : https://networkx.github.io/

Required modules for deep learning.

• Scikit-learn (>= 0.19.1): http://scikit-learn.org/stable/index.html
• Theano (>= 0.9.0): http://deeplearning.net/software/theano/

To run in silico CRISPR/RNAi, a network file and basal gene expression profile are required. All data used in this study are available at the links below. You can also use your own basal expression profiles to run in silico CRISPR/RNAi.

• Downlaod regulatory networks
  • Bayesian network (Nucleic Acids Res. 2015 13;43(12))
    Breast Shuffled Inverted Liver
  • ARACNe network (Nat Genet. 2005 ;37(4))
    Breast Shuffled Inverted Liver
• Download breast cancer cell line expression profile
  cell line RNA-seq data (Cell. 2016 14;164(1-2):293-309)
• Downlaod dependency screens of breast cancer cell lines
  CRISPR-Cas9 screens (Nat Genet. 2017 ;49(12))
  CRISPR-Cas9 screens (Nature. 2019 ;568(7753))
  RNAi screens (Cell. 2016 14;164(1-2))

in silico CRISPR/RNAi Bayesian Example

cd insilico_CRISPR/Bayesian # or cd insilico_RNAi/Bayesian

# sample file directory
mkdir Bayesian_sample
cd Bayesian_sample

# Download cellline basal expression
wget http://143.248.31.34/~omics/data/DeepDependency/sample/Cellline_basal_Bayesian.txt

# Download network file 
wget http://143.248.31.34/~omics/data/DeepDependency/Networks/Bayesian_real_breast.tsv

cd ..
./run.sh

in silico CRISPR/RNAi ARACNe Example

cd insilico_CRISPR/ARACNe # or cd insilico_RNAi/ARACNe

# sample file directory
mkdir ARACNe_sample
cd ARACNe_sample

# Download cellline basal expression
wget http://143.248.31.34/~omics/data/DeepDependency/sample/Cellline_basal_ARACNe.txt

# Download network file 
wget http://143.248.31.34/~omics/data/DeepDependency/Networks/ARACNe_real_breast.tsv

cd ..
./run.sh

In silico CRISPR/RNAi output perturbed_result.txt for the example are available in output directory.

To build a model, the output of in silico CRISPR/RNAi should be matched with experimental dependency data. The whole dependency dataset is available at the link above.

The dataset is divided into training data (infile.tsv) and test data (testset.tsv). Then, the training dataset is shuffled and split into five datasets.

Deep learning model input example

cd DNN_codes
cd input

# Download sample model input
wget http://143.248.31.34/~omics/data/DeepDependency/sample/Model_input.txt

# run
./run_input.sh

Deep learning model running

cd ../codes
./run_train.sh

Edit run_train.sh and change "MAX_J" and "TOT_J" to assign the maximum nodes for parallel running and total number of model simulation.

Hyperparameters search space can be modified in SdA.py file.

Default hyperparameters search space

n_layer        = numpy.random.choice([2, 3, 4])
Hidden         = sorted([numpy.random.choice([100, 500, 1000]) for i in range(n_layer)], reverse=True)
dropout        = [numpy.random.choice([0., 0.1, 0.2, 0.3, 0.4, 0.5]) for i in range(n_layer+1)]
pretrain_lr    = numpy.random.choice([0.1, 0.01, 0.001, 0.0001])
finetune_lr    = numpy.random.choice([0.1, 0.01, 0.001, 0.0001])
L1_param       = numpy.random.choice([1e-08, 1e-04, 1e-02, 1e-01, 1, 10])
L2_param       = numpy.random.choice([1e-08, 1e-04, 1e-02, 1e-01, 1, 10])
mom            = numpy.random.choice([0., 0.1, 0.2, 0.3, 0.4, 0.5])
batch_size     = numpy.random.choice([10, 50, 100])
useRelu        = numpy.random.choice([True, False])
W_distribution = numpy.random.choice(["norm", "uni"])
insize         = 1000   #(number of features)

pretrain_epochs = 100
training_epochs = 10000

All models are saved in ./out directory. If you want to see the log file and error of each model, please check ./log and ./err directory.

Cheking a model performance.

./model_AUC.sh

The best model is moved into ./out/best_model directory and the rest of models are moved into ./out/models. AUC of the best model are plotted at Model_AUC.pdf in out directory. Also, performance of all trained models can be checked at summary.txt in out directory.

In silico CRISPR/RNAi output for the sample data is subjected to prediction.

Example of sample data prediction

cd ../

# Download Clincal samples 
wget http://143.248.31.34/~omics/data/DeepDependency/sample/Clinical_sample.tar.gz
tar zxvf Clinical_sample.tar.gz 

cd codes

# Sample prediction
./run_predict.sh

*.pred files summarize the results of five-fold cross-validation with votes and average prediction score. The default option is to save the prediction outputs in prediction directory.