This is a community model for drug response prediction curated as part of the IMPROVE project.
Enables prediction of drug response with drug and cancer features.
Pan-drug, pan-cancer, single drug response prediction model.
DualGCN consists of two Graph Convolutional Networks (GCN) branches, one for protein-protein interaction information (which includes Copy Number Variation and Gene Expression), and one for the chemical structure of the applied drug. Each branch has built-in dropout and batch normalization, where the dropout rate is one of the considered hyperparameters. At the end of the branches, the obtained features are concatenated and fed into a fully connected network (FCN) with three hidden layers, aiming to do the regression analysis between the output of the two branches and the drug response values.
A data scientist can train the provided untrained model on their own data or use the trained model to classify the provided test samples. The provided scripts use the IMPROVE Benchmark Dataset.
The model-related files are located in Model and Data Clearinghouse (MoDaC), and the source code is on GitHub.
- Cancer features: Copy Number Variation, Gene Expression
- Drug features: Drug SMILES
- Other features: Protein-Protein Interaction
Predicted drug response values for a given sample and drug pair and metrics if ground truth is provided.