Deep Learning for Drug Discovery, Property Prediction with Neural Networks on Raw Molecular Graphs

Abstract

The lengthy and expensive process of developing new medicines is a driving force in the development of machine learning on molecules. Classical approaches involve extensive work to select the right chemical descriptors to use as input data. The scope of this thesis is neural network architectures learning directly on raw molecular graphs, thereby eliminating the feature engineering step. The starting point of experimentation is a reimplementation of the previously proposed message passing neural networks framework for learning on graphs, analogous to convolutional neural networks in how it updates node hidden states through aggregation of neighbourhoods. Three modifications of models in this framework are proposed and evaluated: employment of a recently introduced activation function, a neighbourhood aggregation step involving weighted averaging and a message passing model incorporating hidden states in the graph’s directed edges instead of its nodes. The resulting models are hyperparameter optimized using a parallelized variant of Bayesian optimization. Comparison to literature benchmarks for machine learning on molecules shows that the new models are competitive with state-of-the-art, outperforming it on some datasets.