Deep learning of chemical reactions

The exploration of reactions is a central topic in chemistry.

Compared to the success of machine learning for molecules, the modeling of reactions is lagging behind, especially for stereo- and regioselective reactions.

Since current efforts toward sustainable synthesis such as asymmetric organocatalysis or biocatalysis rely on the accurate prediction of enantio- and regioselective reaction pathways, new modeling approaches are needed.

The proposed project aims toward developing new, data-driven deep learning frameworks for modeling organic and enzymatic reactions, focusing on chemo-, regio-, and stereoselectivity arising through intermolecular interactions with the reagent, solvent, or catalyst.

In detail, we target the rule-free, stereochemistry-aware modeling and subsequent experimental validation of asymmetric organocatalysis to identify new enantioselective transformations, the exploration of new biocatalytic synthesis pathways including enzymatic cascades, and the accurate prediction of activation energies via developing new deep learning approaches.

We will expand molecular graph-convolutional neural networks and graph transformers to reactions in a rule-free manner, and introduce hidden three-dimensional representations to account for stereochemistry and intermolecular interactions, yielding a versatile, open-source toolbox for reaction deep learning.

This approach largely surpasses current approaches, which rely on two-dimensional representations, reaction rules, or three-dimensional input data, in offering the opportunity to model three-dimensional aspects and atom-mapping on-the-fly, for the first time, representing a significant breakthrough in this field.

Its experimental validation campaign further allows for a direct application to the identification of new asymmetric organocatalytic transformations, as well as enzymatic cascades including cofactor recycling and side-product reduction, addressing the current need for more sustainable synthesis.