Predicting tumor cells' reactions to medical treatments has enormous significance for both medical and clinical studies. Many machine learning approaches have been previously applied to predict the effectiveness of anti-cancer drugs by analyzing genomic profiles. However, most of the current methods neglect the fact that proteins rarely act alone and often team up as "molecular machines" characterized by intricate dynamic connections to perform biological functions. Other studies indicate that subnetworks of protein-protein interaction (PPI) networks could be crucial for the development of cancer and could be used as new therapeutic targets. Recent advances in graphical representation learning inspired by deep neural networks have provided a potential solution to include PPI information into drug response prediction to improve performance.
We present a graph neural network (GNN) based approach, GNNDR, to incorporate PPI information into drug sensitivity prediction based on genomic profiles. By communicating information among neighborhoods on the PPI network, our model fuses genomic features into a structural PPI graph which encodes information relevant to drug responses. We compare our approach with state-of-the-art baselines on three publicly available datasets which include both genomic features and drug responses of cell lines and Patient Derived Xenografts (PDX). We investigate the impact of different types of interactions as well as the effectiveness of different graph pooling strategies. Experimental results demonstrate the superior performance of our model over the baselines. In particular, on the NIBR-PDXE dataset, our model outperforms the baselines in 21 out of 26 cancer-treatment pairs and achieves a MSE of 1.46 on the CCLE-GDSC dataset and a MSE of 0.11 on the NCI-ALMANAC dataset. In addition, our model provides substantial model interpretability in terms of gene similarity and gene importance via the graph pooling procedures.
We propose GNNDR, a GNN based approach which not only possesses the large learning capacity, but provides a solution to combine PPI information with genomic features for drug response prediction. Empirical evaluation demonstrates the effectiveness of adding protein information. Our approach, as far as we know, the first GNN based model for drug response prediction, provides a promising perspective for the identification of anti-cancer therapies.