| Autor: |
Beker W; Allchemy, Inc., Highland, Indiana 46322, United States.; Institute of Organic Chemistry, Polish Academy of Sciences, Warsaw 01-224, Poland., Roszak R; Allchemy, Inc., Highland, Indiana 46322, United States.; Institute of Organic Chemistry, Polish Academy of Sciences, Warsaw 01-224, Poland., Wołos A; Allchemy, Inc., Highland, Indiana 46322, United States.; Institute of Organic Chemistry, Polish Academy of Sciences, Warsaw 01-224, Poland., Angello NH; Department of Chemistry, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States., Rathore V; Department of Chemistry, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States., Burke MD; Department of Chemistry, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States.; Department of Biochemistry, Institute for Genomic Biology, Carle Illinois College of Medicine, and Beckman Institute, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States., Grzybowski BA; Allchemy, Inc., Highland, Indiana 46322, United States.; Institute of Organic Chemistry, Polish Academy of Sciences, Warsaw 01-224, Poland.; Center for Soft and Living Matter, Institute for Basic Science (IBS), Ulsan 44919, Republic of Korea.; Department of Chemistry, Ulsan Institute of Science and Technology (UNIST), Ulsan 44919, Republic of Korea. |
| Abstrakt: |
Applications of machine learning (ML) to synthetic chemistry rely on the assumption that large numbers of literature-reported examples should enable construction of accurate and predictive models of chemical reactivity. This paper demonstrates that abundance of carefully curated literature data may be insufficient for this purpose. Using an example of Suzuki-Miyaura coupling with heterocyclic building blocks─and a carefully selected database of >10,000 literature examples─we show that ML models cannot offer any meaningful predictions of optimum reaction conditions, even if the search space is restricted to only solvents and bases. This result holds irrespective of the ML model applied (from simple feed-forward to state-of-the-art graph-convolution neural networks) or the representation to describe the reaction partners (various fingerprints, chemical descriptors, latent representations, etc.). In all cases, the ML methods fail to perform significantly better than naive assignments based on the sheer frequency of certain reaction conditions reported in the literature. These unsatisfactory results likely reflect subjective preferences of various chemists to use certain protocols, other biasing factors as mundane as availability of certain solvents/reagents, and/or a lack of negative data. These findings highlight the likely importance of systematically generating reliable and standardized data sets for algorithm training. |
