KekuleScope: improved prediction of cancer cell line sensitivity using convolutional neural networks trained on compound images

The application of convolutional neural networks (ConvNets) to harness high-content screening images or 2D compound representations is gaining increasing attention in drug discovery. However, existing applications often require large data sets for training, or sophisticated pretraining schemes for the networks. Here, we show on eight cytotoxicity IC50 data sets from ChEMBL 23 that the in vitro activity of compounds on cancer cell lines can be accurately predicted on a continuous scale from their Kekule structure representations alone by extending existing architectures (AlexNet, DenseNet-201, ResNet152 and VGG-19), which were pretrained on unrelated image data sets. We show that the predictive power of the generated models, which just require standard 2D compound representations as input, is comparable to that of Random Forest (RF) models trained on circular (Morgan) fingerprints, a combination which is considered to be the state of the art. Notably, including additional fully-connected layers further increases the predictive power of the networks by up to 10%. Analysis of the predictions generated by RF models and ConvNets shows that by simply averaging the output of the RF models and ConvNets we constantly obtain significantly lower errors in prediction (4-12% decrease in RMSE on the test set) than those obtained with either model alone, indicating that the features extracted by the convolutional layers of the ConvNets provide complementary predictive signal to Morgan fingerprints. Overall, in this work we present a set of ConvNet architectures for the prediction of compound activity from their Kekule structure representations with state-of-the-art performance, that require no generation of compound descriptors or use of sophisticated image processing techniques. The data sets and the code used are provided at this https URL.