Predicting Highly Enantioselective Catalysts Using Tunable Fragment Descriptors**

Abstract

Catalyst optimization processes typically rely on inductive and qualitative assumptions of chemists based on screening data. While machine learning models using molecular properties or calculated 3D structures enable quantitative data evaluation, costly quantum chemical calculations are often required. In contrast, readily available binary fingerprint descriptors are time‐ and cost‐efficient, but their predictive performance remains insufficient. Here, we describe a machine learning model based on fragment descriptors, which are fine‐tuned for asymmetric catalysis and represent cyclic or polyaromatic hydrocarbons, enabling robust and efficient virtual screening. Using training data with only moderate selectivities, we designed theoretically and validated experimentally new catalysts showing higher selectivities in a challenging asymmetric tetrahydropyran synthesis.

Multi-Instance Learning Approach to the Modeling of Enantioselectivity of Conformationally Flexible Organic Catalysts

“…Table reports the performance of Instance-Wrapper MIL models on the three test sets described in Section . These results can also be compared with those reported by Tsuji et al, Sandfort et al, Zahrt et al, and Asahara and Miyao (see Table ).…”

“…Different popular 2D and 3D descriptors were benchmarked on the same data sets as the MIL algorithms (see above). Namely, we considered ISIDA and CircuS fragment descriptors, 2D fingerprints, and 3D descriptors available in RDKit as well as pmapper 3D atom triplet descriptors. A set of 3D RDKit descriptors includes RDF, MoRSE, WHIM, GETAWAY, and AutoCorr3D descriptors.…”

“…In this study, we used atom-centered subgraphs (atoms with first, second, and so on, coordination spheres) where the radius varied from two to five atoms. For alternative CircuS fragment descriptors, 25 we used atom-centered fragments with radii from 2 to 5.…”

Prediction of Reaction Performance by Machine Learning Using Streamlined Features: NMR Chemical Shifts as Familiar Descriptors

“…Recently, remarkable accomplishments have been reported, and predictive models for various organic reactions were gradually generated. [9][10][11][12][13][14][15][16][17][18][19] As representative examples, the Denmark and the Sigman groups applied the ML approach to predict the enantioselectivity of asymmetric imine addition, [9,13] and Doyle's work on the Buchwald-Hartwig crosscoupling showed extrapolative prediction of reaction yield. [12] These adaptations of ML in synthetic organic chemistry make reaction design more efficient by reducing time-and resource-consuming trial-and-error routines.…”

“…These ML models show a powerful predictive ability for chemical yields as well as stereoselectivities ( Scheme 1). Recently, remarkable accomplishments have been reported, and predictive models for various organic reactions were gradually generated [9–19] . As representative examples, the Denmark and the Sigman groups applied the ML approach to predict the enantioselectivity of asymmetric imine addition, [9,13] and Doyle's work on the Buchwald–Hartwig cross‐coupling showed extrapolative prediction of reaction yield [12] .…”