A Hybrid Convolutional Network for Prediction of Anti-cancer Drug Response

The latest medical research results and clinical practice have showed that the effectiveness of existing anti-cancer drugs is highly dependent on the genomic characteristics of patients, which means that the efficacy of the same anti-cancer drugs may be very different for different patients even if they are suffering from the same cancer disease, since they usually have different genomic features. How to select appropriate anti-cancer drugs for different cancer patients is a frontier topic and challenge in the field of precision oncology. In this paper, we design a hybrid convolutional neural network (CNN) to predict the responses of anti-cancer drugs, in which the network is constructed with two input CNN branches and two output CNN+FC (full connected) branches. One input branch is to extract the genomic feature from the input data of a cancer patient's gene expression, mutation or copy number variations, and the other input branch is to extract the molecular fingerprint feature from the chemical structure data of the drug to be used for curing the cancer. In addition, attention mechanism is introduced to weight the two features according to their importance, the two weighted features are then concatenated into one vector and sent to the two output branches. For the two output branches, one is to predict the IC50 values and the other is to predict the sensitivity (or insensitivity) of cancer cell lines to anti-cancer drugs. Furthermore, the whole network system is optimized through an end-to-end training process with the joint loss function composed of two output losses. By this way, the excellent ability of CNNs in deep feature extraction and computation can be better utilized so as to better predict the IC50 and sensitivity and insensitivity of the cancer cells to anticancer drugs. Experimental results obtained in the paper show that the proposed method outperforms the existing state of the art methods in terms of the accuracy, sensitivity, and other key performance indexes.