Deep Learning Models for prediction of hERG inhibitors and kinase inhibitors

Term: 7/2016 - 12/2018

The Institute of Bioinformatics, Johannes Kepler University Linz (JKU), has recently developed successful Deep Learning models for the prediction of biological assays, concretely toxicity and activity assays. To this end and for the performance under this Research Plan, (1) chemical structures and their associated measurements have to be preprocessed by a computational pipeline to obtain a numerical description of the instances. This includes the selection of chemical descriptors and fingerprints which influence the final performance of the models. (2) Deep Learning architectures have to be tested and evaluated where breadth and depth of layers, different activation and loss functions, learning and regularization methods are considered. Furthermore, (3) a large number of hyperparameters, such as the learning rate, have to be tested. (4) The performance of different architectures and hyperparameters has to be evaluated in a cross-validation procedure. (5) The resulting performance values are tested against a competing machine learning method, RandomForest. (6) The computational pipeline should provide probabilistic outputs, that is a probability for a certain compound to be active in a specific assay. To this end, a sigmoid curve has to be fitted (Platt scaling).