Discovering Novel Chemical Reactions

Abstract

Accurately predicting chemical reactions can facilitate the search for optimal synthe sis routes in a chemical reaction network and as a consequence expedite the lengthy drug discovery process. As an effort in this direction, this work aims to explore AstraZeneca’s chemical knowledge graph by two complementary analyses. In a first part, graph theory related statistics is employed as a means to gain insights about the chemical reaction graph at AstraZeneca. Significant differences are observed be tween this internal reaction graph and the one based on the public dataset of United States patents as well as other reaction graphs discussed in literature. Secondly, a link prediction model is applied to and evaluated on AstraZeneca’s chemical reaction graph, in order to suggest new potential chemical reactions. In order to successfully accomplish this task, an existing link prediction model is adapted and trained. The test results are then compared to heuristic baselines, showing that the proposed implementation substantially exceeds what can be achieved with heuristic methods. One of the contribution from this research is a comparison between different ways to sample the ground truth class of non-existing links for training and evaluation. The choice of method for this task is shown to have an impact on the final predictions. Finally, a set of promising, predicted reactions are suggested and is currently under further investigation at AstraZeneca.