Cancer Drug Response Prediction via a Hybrid Graph Convolutional Network

This work has been accepted to ECCB2020 and will also appear in Bioinformatics soon.

DeepCDR is a hybrid graph convolutional network for cancer drug response prediction. It takes both multi-omics data of cancer cell lines and drug structure as inputs and predicts the drug sensitivity (binary or contineous IC50 value).

• Keras==2.1.4
• TensorFlow==1.13.1
• hickle >= 2.1.0

DeepCDR can be downloaded by

git clone https://github.com/kimmo1019/DeepCDR

Installation has been tested in a Linux/MacOS platform.

We provide detailed step-by-step instructions for running DeepCDR model including data preprocessing, model training, and model test.

Step 1: Data Preparing

Three types of raw data are required to generate genomic mutation matrix, gene expression matrix and DNA methylation matrix from CCLE database.

CCLE_mutations.csv - Genomic mutation profile from CCLE database

CCLE_expression.csv - Gene expression profile from CCLE database

CCLE_RRBS_TSS_1kb_20180614.txt - DNA methylation profile from CCLE database

The three types of raw data genomic mutation file, gene expression file and DNA methylation file can be downloaded from CCLE database or from our provided Cloud Server.

After data preprocessed, the three following preprocessed files will be in located in data folder.

genomic_mutation_34673_demap_features.csv -- genomic mutation matrix where each column denotes mutation locus and each row denotes a cell line

genomic_expression_561celllines_697genes_demap_features.csv -- gene expression matrix where each column denotes a coding gene and each row denotes a cell line

genomic_methylation_561celllines_808genes_demap_features.csv -- DNA methylation matrix where each column denotes a methylation locus and each row denotes a cell line

We recommend to start from the preprocessed data. Please note that each preprocessed file is in csv format, of which the column and row name are provided to speficy mutation location, gene name, methylation location and corresponding Cell line.

Step 2: Drug feature representation

Each drug in our study will be represented as a graph containing nodes and edges. From the GDSC database, we collected 223 drugs that have unique Pubchem ids. Note that a drug under different screening condition (different GDSC drug id) may share the same Pubchem id. Here, we used deepchem library for extracting node features and gragh of a drug. The node feature (75 dimension) corresponds to a stom in within a drug, which includes atom type, degree and hybridization, etc.

We recorded three types of features in a list as following

drug_feat = [node_feature, adj_list, degree_list]
node_feature - features of all atoms within a drug with size (nb_atom, 75)
adj_list - adjacent list of all atoms within a drug. It denotes the all the neighboring atoms indexs
degree_list - degree list of all atoms within a drug. It denotes the number of neighboring atoms 

The above feature list will be further compressed as pubchem_id.hkl using hickle library.

Please note that we provided the extracted features of 223 drugs from GDSC database, just unzip the drug_graph_feat.zip file in data/GDSC folder

Step 3: DeepCDR model training and testing

Here, we provide both DeepCDR regression and classification model here.

python run_DeepCDR.py -gpu_id [gpu_id] -use_mut [use_mut] -use_gexp [use_gexp] -use_methy [use_methy] 
[gpu_id] - set GPU card id (default:0)
[use_mut] - whether use genomic mutation data (default: True)
[use_gexp] - whether use gene expression data (default: True)
[use_methy] - whether use DNA methylation data (default: True)

One can run python run_DeepCDR.py -gpu_id 0 -use_mut True -use_gexp True -use_methy True to implement the DeepCDR regression model.

The trained model will be saved in data/checkpoint folder. The overall Pearson's correlation will be calculated.

python run_DeepCDR_classify.py -gpu_id [gpu_id] -use_mut [use_mut] -use_gexp [use_gexp] -use_methy [use_methy] 
[gpu_id] - set GPU card id (default:0)
[use_mut] - whether use genomic mutation data (default: True)
[use_gexp] - whether use gene expression data (default: True)
[use_methy] - whether use DNA methylation data (default: True)

One can run python run_DeepCDR_classify.py -gpu_id 0 -use_mut True -use_gexp True -use_methy True to implement the DeepCDR lassification model.

The trained model will be saved in data/checkpoint folder. The overall AUC and auPRn will be calculated.

We also provided the external patient data downloaded from Firehose Broad GDAC. The patient data were preprocessed the same way as cell line data. The preprocessed data can be downloaded from our Server.

The preprocessed data contain three important files:

mut.csv - Genomic mutation profile of patients

expr.csv - Gene expression profile of patients

methy.csv - DNA methylation profile of patients

Note that the preprocessed patient data (csv format) have exact the same columns names as the three cell line data (genomic_mutation_34673_demap_features.csv, genomic_expression_561celllines_697genes_demap_features.csv, genomic_methylation_561celllines_808genes_demap_features.csv). The only difference is that the row name of patient data were replaced with patient unique barcode instead of cell line name.

Such format-consistent data is easy for external evaluation by repacing the cell line data with patient data.

As GDSC database only measured IC50 of part cell line and drug paires. We applied DeepCDR to predicted the missing IC50 values in GDSC database. The predicted results can be find at data/Missing_data_pre/records_pre_all.txt. Each record represents a predicted drug and cell line pair. The records were sorted by the predicted median IC50 values of a drug (see Fig.2E).

If you have any question regard our code or data, please do not hesitate to open a issue or directly contact me ([email protected])

This project is licensed under the MIT License - see the LICENSE.md file for details