Predicting Synergistic Drug Interaction with DNN and GAT

Many complex diseases such as cancer cannot be effectively treated with one type of medication, giving rise to another treatment route that combines several drugs to achieve the desired effects. We developed deep learning models to predict a parameter that gauges such effects known as synergy score and specifically made use of relevant genetic and drug datasets. The expected outcome enables rapid identification of novel drug pairs with potential use in cancer therapy. The employed genetic datasets included gene expression, copy number variation, and somatic mutation. We applied different variations of autoencoders on these datasets, namely deep autoencoder, sparse autoencoder, and deep sparse autoencoder to reduce the dimensions and only retain what was deemed non-redundant. Alternatively, we filtered the data based on landmark gene names. As for the training drug datasets that contained associated synergy scores calculated empirically, we either used ECFP6 molecular fingerprint representations as an input for a deep neural network (DNN) or a molecular graph for a graph neural network (GNN) model. We set out to compare the performance of these two representations in appropriate deep learning models as well as determine how well each autoencoder method fared. Among different autoencoders, the best performing option was the sparse autoencoder for DNN and the deep autoencoder for GNN. After loading the processed genetic datasets into ECFP6-DNN and graph embedding-GNN model, we found that ECFP6-DNN performed better with a mean square error of 146.137, while graph embedding-GNN had a mean square error of 174.952.