Machine learning maps atoms

Machine learning software can now accurately track atoms from reactants to products without human help. On Thursday at ACS Spring 2021, a meeting of the American Chemical Society, Philippe Schwaller of IBM Research and the University of Bern described how his team took machine learning algorithms designed for interpreting language and repurposed them for the problem of atom mapping. He presented their findings in a session in the Division of Chemical Information and in a recent paper (Sci. Adv. 2021, DOI: 10.1126/sciadv.abe4166).

Machine learning models called transformers power programs like Google Translate, and Schwaller and his colleagues have recently demonstrated their ability to accurately interpret and predict chemical transformations. In an effort to understand what steps these algorithms use to predict the outcomes of reactions, Schwaller collaborated with Benjamin Hoover, also of IBM Research, to visualize connections their software was finding between atoms in reactant and product molecules. They realized that the models were working out the process of atom mapping, forming connections between individual atoms in their positions in the reactant and product molecules. The team harnessed those connections made by the algorithm to make an automated atom-mapping tool they call RXNMapper, Schwaller said during his talk on Thursday. Atom mapping is traditionally done by humans or by computers guided by set, encoded expert rules. Research has shown these atom maps are often inaccurate, especially when the reactions are complex. Schwaller said RXNMapper’s fast and accurate atom mapping can improve automated reaction prediction and synthesis planning software. The researchers showed their tool made accurate atom maps better than 99% of the time, performing as well or nearly as well as one that relies on expert rules (Nat. Commun. 2021, DOI: 10.1038/s41467-019-09440-2) and better than other atom-mapping tools. Schwaller said RXNMapper works quickly, too, processing about one reaction per 7 ms.

It remains to be seen how valuable accurate, automated atom-mapping is to the overall goal of synthesis prediction, according to experts not involved in the work. Groups that have built automated reaction prediction software, like Connor Coley’s at the Massachusetts Institute of Technology, have shown their machine learning algorithms can accurately predict the results of reactions even when they are trained on flawed atom maps. “It’s not obvious to me how much of an impediment poor atom mapping has been to success on downstream tasks,” Coley says. Schwaller acknowledges that approaches like Coley’s have demonstrated they can be accurate. Schwaller told C&EN that he thinks RXNMapper will extend machine learning tools’ abilities to make predictions by automatically generating reliable atom maps that can teach those algorithms about the movements of atoms in reactions they haven’t seen before.