Year of Graduation
Hierarchical Representation Learning on Molecular Graphs using Neural Networks
Supervised learning of molecular properties has multiple potential applications in chemistry, drug discovery, and material science. Nowadays several promising machine learning models have been described. State of the art models are based on a message passing algorithm and aggregation procedure to produce graph-level embedding. An important step in developing discriminative models is to design an optimal and interpretable aggregation function. The way medicinal chemists analyze molecules inspired an algorithm proposed in this paper. A proposed method is a hierarchical approach that gradually aggregates molecular representation at different levels, starting with simple substructures, ending with large, complex scaffolds. The proposed algorithm shows better performance than previously proposed methods on the prediction of biochemical properties of molecules.