Computational evolution of new catalysts for the Morita–Baylis–Hillman reaction

Abstract

We present a de novo discovery of an efficient catalyst of the methanol-mediated Morita– Baylis–Hillman (MBH) reaction by searching chemical space for molecules that lower the es- timated barrier of the rate-determining step using a genetic algorithm (GA) starting from randomly selected tertiary amines. We performed five independent GA searches that resulted in 448 unique molecules, for which we were able to locate 435 true transition states at semiem- pirical level of theory. The predicted activation energies of all 435 molecules were all lower than that of 1,4-diazabicyclo[2.2.2]octane (DABCO), which is a popular catalyst of the MBH reaction. Virtually all the molecules contain an azetidine N as the catalytically active site, which is discovered by the GA since it is either not found in the initial population or discarded early only to be rediscovered as the search progresses. Many of the GA searches also intro- duce a substituent with a hydrogen bond donor that helps to stabilize the transition state and thus lower the barrier. Two molecules are selected for further study based on their synthetic accessibility as predicted by the retrosynthesis package Manifold. For these two molecules, we compute the entire free energy reaction profile at the DFT level and show that their rate- determining barriers are 1.7 and 2.4 kcal/mol lower than that of DABCO. Azetidines have not been used as catalysts for the MBH reaction so we experimentally verified our predictions. One of the molecules was successfully synthesized and is indeed more active than DABCO with an eight-fold increase in the rate-constant corresponding to a roughly 1 kcal/mol lower barrier, in good agreement with the predictions. We believe this is the first experimentally verified de novo discovery of an efficient catalyst using a generative model.

Augmenting Genetic Algorithms with Machine Learning for Inverse Molecular Design

“…There recently has been an uptake in interest for evolutionary optimization in chemistry, with successful applications to diverse problems including the design of mechanosensitive conductors, 31 polymers, [32][33][34] drug-like molecules, [35][36][37][38] and catalysts. [39][40][41][42][43] Given the success of both evolutionary and machine learning in materials science, it is natural to investigate the combination of both approaches. While being an incipient area of research still in its infancy, efforts have been made to explore the synergies and some very promising advances have already been achieved.…”