Improving image-based compound activity prediction with convolutional neural networks

Predicting the biological activity of a chemical compound in-silico is a crucial step in the drug design process, because measuring the activity in-vitro or in-vivo is time consuming and cost intense. In this work, we evaluate different machine learning models that leverage high-throughput images to predict biological activity. In con- trast to quantitative structure-activity relationship (QSAR) models that are based on the chemical structure, the image-based models include biological information and are independent of chemicals structure. This allows the image-based models to de- tect novel, potentially unexpectedly active chemical structures. In this work, we compare two different approaches for image-based compound activity prediction: The first approach is a deep neural network (DNN) trained on precalculated image descriptors, whereas the second approach is based around a convolutionalneuralnetwork(CNN)whichistraineddirectlyonthehigh-throughput images. We show that the CNN significantly outperforms the DNN with an average area under the curve (AUC) of 0.60 (+-0.26) over all assays on an external test set. A significant and high predictivity (AUC>0.9) is achieved for 16.92% or 11 assays. Descriptor-based DNNs achieve an average AUC of 0.51 (+- 0.14) and no significant and high predictivity. Using these highly accurate predictive models, we were able to annotate 12.000 compounds in 11 assays, which amounts to 130.000 newly obtained data points almost equivalent to assay measurements. Instead of measuring a new com- pound for activities, a high-throughput image can be taken of the cells with that specific compound applied to them. This image can further be processed to predict different assays or effects of yet unknown compounds to speedup the drug design process.