Palladium catalyzed C-N cross-couplings are a real workhorse in both academia and industry. In particular, the pharmaceutical sector frequently uses such reactions to graft nitrogen functionality into organic molecules. With this often comes a quest for finding suitable reaction conditions to achieve a reasonable yield. The current state-of-the-art approach is to screen through many conditions, often in a random or empirically guided way. This is somewhat surprising, since a vast amount of chemical knowledge is present in either public, commercial or company-internal databases. These data should in principle help select promising reaction conditions. In previous work, we gathered a large amount of Pd-catalyzed cross-coupling reactions from Reaxys, Scifinder and the USPTO, cleaned them and performed statistical analysis. This resulted in cheat sheets, which are a first step to data-driven guides for immediate support of practitioners. In this work, we built on this dataset, which comprises more than 60,000 Buchwald-Hartwig reactions. We applied machine-learning (ML) methods to model the relationship between reactants/reagents and the yield. Three different methods have been investigated: i) multiple fingerprint features; ii) reaction-adapted attention-based graph neural networks; and iii) our custom chemical descriptors with tree-based models. While we find that method iii) consistently outperforms other methods on various data-splits with good metrics, in general the predictive power diminishes drastically when looking at test sets which mimic typical screening experiments. We trace this back to bias and issues with the data diversity, in particular in the space of reagents. These results are distinct from previously published works on this subject that were mainly focused on data from high-throughput-experimentation coming out of few labs, resulting in comparatively small training and test spaces. In contrast, we investigated a general-purpose reaction yield prediction for arbitrary C-N couplings from heterogeneous data sources. The lessons learned will be of importance to future developments. Moreover, we are preparing a dedicated experimental test set with realistic diversity in both reactants and reagents. We hope this will serve the community as a benchmark for future AI advancements in the prediction of reaction conditions.

Predicting C-N cross-coupling yields from literature data: Lessons from a machine learning study

ACS Spring Meeting 2021

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