Chemically-meaningful molecular similarity metrics play a crucial role in drug discovery including virtual-high throughput screening, characterizing chemical diversity in compound libraries, modeling quantitative structure-activity relationship, and analogue search. As molecules can be viewed as 2D graphs, graph edit distance (GED) between them serves as a chemically relevant and interpretable metric of 2D molecular similarity. GED as a metric of chemical similarity has a distinct advantage over fingerprint and substructure-based similarity metrics in terms of being appropriately sensitive to molecular changes. GED also reflects medicinal chemist’s intuition better than difficult-to-interpret fingerprint-based similarity measures. However, the computation of exact graph edit distance is NP-hard. Approximate algorithms that scale polynomially with the number of atoms are generally not sufficiently accurate for chemical applications, limiting their routine usage in drug discovery workflows.<br/>Here, we address that limitation by proposing a novel graph neural network-based approach to predict GED between chemical graphs. Our approach scales linearly with the number of atoms, accurately predicting GED at a fraction of the computational cost compared to computing it exactly. In addition, our approach outperforms existing machine-learning-based  GED predictors on non-trivial train-test splits, demonstrating transferability to previously unseen areas of chemical space. We show that consideration of bond type information is critical to achieve high accuracy in predicting GEDs sensitive to bond types. Specifically, we show the applicability of the model to drug-like molecules enabling applications in drug discovery.

Graph neural network approach to accurately and scalably predict graph edit distance in drug-like chemical spaces

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.