Expanding the chemical space using a Chemical Reaction Knowledge Graph

In this work, we present a new molecular de novo design approach which utilizes a knowledge graph encoding of chemical reactions, extracted from the publicly available USPTO (United States Patent and Trademark Office) dataset. Our proposed method can be used to expand the chemical space by performing forward synthesis prediction on the knowledge graph and can generate libraries of de novo compounds along with a valid synthetic route. The forward synthesis prediction of novel compounds involves two steps. In a first step, a graph neural network-based link prediction model is used to suggest pairs of existing reactant nodes in the graph that are likely to react. In a second step, product prediction is performed using a molecular transformer model to obtain the potential products for the suggested reactant pairs. We achieve a ROC-AUC score of 0.861 for link prediction in the knowledge graph and for the product prediction a top-1 accuracy of 0.924. The method’s utility is demonstrated by generating a set of de novo compounds by predicting high probability reactions in USPTO. The generated compounds are diverse in nature and many exhibits drug-like properties. Further, evaluation of the potential activity using a quantitative structure–activity relationship (QSAR) model suggested presence of potential dopamine receptor D2 (DRD2) modulators among the proposed compounds. In summary, our results suggest that the proposed method can expand the easily accessible chemical space and identify novel drug-like compounds for a specific target.