Variational Autoencoder for Anti-Cancer Drug Response Prediction

There has been remarkable progress in identifying the perplexity of genomic landscape of cancer over the past two decades, providing us a brand new way to understand cancer and anti-cancer drugs. However, it is extremely expensive and time-consuming to develop anti-cancer drugs, and the diversity of cancer genomic features and drug molecular features renders it considerably difficult to customize therapy strategy for patients. In order to facilitate the discovery of new anti-cancer drugs and the selection of drugs to provide personalized treatment strategy, we seek to predict the response of different anti-cancer drugs with deep learning models. We test our model on breast cancer cell lines at first, and then generalize our method on pan-cancer dataset. We incorporate 2 kinds of essential information for anti-cancer drug response prediction into our model, namely gene expression data of cancer cell lines and drug molecular data. In order to extract representative features from these unlabeled data, we propose variational autoencoder (VAE) which has been proved to be very powerful in unsupervised learning. Our model encode gene expression data and drug molecular data with geneVAE model and rectified Junction Tree variational autoencoder\cite{jin2018junction} (JTVAE) model respectively, and process the encoded features with a Multi-layer Perceptron (MLP) model to produce a final prediction. We reach an average coefficient of determination ($R^{2} = 0.83$) in predicting drug response on breast cancer cell lines and an average $R^{2} > 0.84$ on pan-cancer cell lines. We further explore the latent representations encoded by geneVAE and JTVAE model, and the results show these models are robust and can preserve critical features of original data.