Chemical reactions often have multiple possible products. Predicting and determining the selectivity is central to synthesis. The best strategy to obtain a desired scaffold is performing multiple experiments and analyzing the results individually to determine the conditions which lead to the required regioselectivity. This approach is resource intensive. The wealth of data available in chemical reaction databases should be able to help.

In this project we want reduce experimentation cost and accelerate discovery by harnessing the power of chemical information from databases towards reaction prediction. We have developed a workflow for aromatic halogenation, an important reaction class where theoretical background is available. This will enable us to develop a robust workflow and expand the work towards other transformations where a priori knowledge is not readily available. By integrating artificial intelligence in the framework, our methodology circumvents the need to perform redundant experiments and leverages data from existing literature knowledge towards a greener and more economical research paradigm.

The first step of the project involves meticulous curation and data cleaning of entries from Reaxys. The key steps were duplicate removal, validate the reaction is a halogenation, and identify the reaction site. From 3,253 chlorination, 11,328 bromination, and 5,814 iodination reactions we build a 17,815-reaction superset of aromatic mono-halogenations.

Starting materials were featurized as Morgan fingerprints (FP, 2048-bits, radius = 2) and 12 machine learning architectures were evaluated on predicting the product (predicted-FP). The product fingerprint is compared to the product on each reaction site – generated algorithmically – and the most similar is selected as the predicted product. We achieved 79.7% accuracy on 10-fold cross validation and 79.6% on the test set. Because of the low computing power required to run the model it can be hosted and run on browser. We developed a web platform to share with industrial partners and enable direct feedback on the usability and flexibility of the developed tool. Future efforts are expanding model generalization capabilities and evaluation against literature benchmarks on the same dataset.

Henrique Magri Marcon

Appearances