On the ChEMBL Platform, a Large-scale Evaluation of Machine Learning Algorithms for Drug Target Prediction

Abstract

Deep learning is currently the most successful machine learning technology in a wide range of application fields, and it has recently been used to forecast possible therapeutic targets and screen for active compounds in drug discovery research. However, it is unclear whether deep learning can outperform existing computational methods in drug discovery tasks due to the lack of large-scale studies, the compound series bias that is common in drug discovery datasets, and the hyperparameter selection bias that comes with the large number of potential deep learning architectures. As a result, we compared the outcomes of different deep learning methods to those of other machine learning and target prediction methods on a large-scale drug development dataset. We employed a stacked cluster-cross-validation technique to avoid any biases from hyperparameter selection or compound series. We discovered that (i) deep learning methods beat all competing methods, and (ii) deep learning's prediction performance is often comparable to that of tests conducted in wet labs (i.e., in vitro assays).