Machine learning for yield prediction for chemical reactions using in situ sensors

Davies, Joseph; Pattison, David A.; Hirst, Jonathan D.

doi:10.1016/j.jmgm.2022.108356

Cited by 4 publications

(1 citation statement)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Machine learning (ML) is a useful tool for interpreting and solving such problems. ML has found applications in predicting the bioactivities of new drugs, 12,13 optimizing reaction conditions, 14 predicting chemical yields, 15,16 and in the development of new synthetic routes to desired molecules. 17 ML has also found notable success in the area of asymmetric catalysis.…”

Section: Introductionmentioning

confidence: 99%

Describing Chiral Ligands in Palladium-catalyzed Decarboxylative Asymmetric Allylic Allylation: A Critical Comparison of Three Machine Learning Approaches

Galvin,

Aguilar Bejarano,

Rogers

et al. 2024

Preprint

View full text Add to dashboard Cite

Despite its current popularity, machine learning (ML) applied to asymmetric catalysis remains underexplored. Present strategies include direct use of existing descriptors (e.g. those originally formulated for medicinal chemistry), the development of new bespoke steric and electronic descriptors, or the use of molecular graphs. This method diversity, in the absence of user guidelines, makes selecting an optimal ML algorithm unclear. The fact that asymmetric catalysis data sets are frequently small also make interpretable ML of chiral ligand understanding difficult to realize. Herein, we present an exhaustive evaluation of reaction representations in combination with different machine learning algorithms (including linear regression, random forests, gradient boosting, and graph neural networks) using a realistic-size database compromising 103 palladium- catalyzed decarboxylative asymmetric allylic alkylation (DAAA). This database consists of the combination of three different Trost-type ligands with 54 different substrates. It is concluded that our new bespoke steric and electronic descriptors offer the best performance, while overcoming the problem of interpretability of using existing topo-electronic descriptors, and the problem of data requirements of Graph Neural Networks.

show abstract