Navigating C-H Activation Selectivity Landscapes with Localised Featurisations

Juliette Schleicher, Shubhajit Das, Ruben Laplaza, Clemence Corminboeuf

C-H bond functionalisations are a central strategy for the development of efficient synthetic routes, particularly for complex molecules. Due to the myriad of available C-H bonds, regioselectivity predictions are crucial to aid and facilitate the efficient discovery of accessible pathways. Regioselectivity predictions with physics-based representations, one of the most accurate approaches [1], are often based on intermediate or even transition state structures, which prove challenging to assemble. In this context, we pivot towards using only the substrate structures for regioselectivity predictions of the direct arylation reaction, a prevalent pathway within a medicinal chemist’s toolbox [2]. We present a local, substrate-based featurisation for predicting reaction properties using computed data, thus expediting the predictive workflow by achieving independence from intermediate structures. Additionally, we introduce the database Arylation-24-TS comprising computed transition state barriers for the regioselectivity-inducing step of the direct arylation reaction.

Figure 1. Schematic outline of the predictive pipeline. The local featurisation is created based on the C-H environment in the substrate structure and the product, subsequently, ML models are trained with those featurisations on the computed reaction properties, allowing us to predict regioselectivities based on the trained models.

1. Puck van Gerwen, Ksenia R. Briling, Yannick Calvino Alonso, Malte Franke and Clemence Corminboeuf, Digital Discovery 2024, Accepted Manuscript.
2. T. Cernak, K. D. Dykstra, S. Tyagarajan, P. Vachal and S. W. Krska, Chem. Soc. Rev. 2016, 45, 546-576.

 Juliette Schleicher