change in mechanistic pathways which can be informative and potentially provide
insight to alternative reaction pathways.
5.3.10 Model Utilization
Utilization of regression is categorized into two types: prediction and physical
interpretation [93]. Often in QSAR, one trades a meaningful model for a highly
predictive one. However, in QSSR for asymmetric catalysis, in some cases, applications of regression can lie in the sweet spot where the models have been illustrated
to be both predictive and not too complicated which allows for meaningful
interpretations [94].
Chemists have used QSSR to model, predict and optimize levels of
enantioselectivity since the turn of the twenty-first century [58, 59, 62, 85, 95,
96]. Attention to this approach continues to grow [97–101]. We select recent
examples at the frontier of QS(A/S)R applied to asymmetric catalysis to discuss
here. Examples of work on MIF-based 3D-QSSR have been comprehensively
discussed in a recent review by Denmark and co-workers and are not included
here [56].
5.4 Predicting Enantioselectivity
Sigman has pioneered modern applications of QSSR to asymmetric transformations,
publishing numerous examples of the use of regression-assisted reaction optimization and mechanistic elucidation [102–104]. Recently, in collaboration with Miller,
the Sigman laboratory has carried out multivariate modelling for the atroposelective
bromination of arylquinazolinones involving a tetrapeptide catalyst [105]. Predicted
product enantioselectivity was expressed in terms of calculated parameters: NBO
charges, IR stretches, a multidimensional steric Sterimol (L ) parameter and a
crystallographically derived parameter; main chain angles (Fig. 13). The QSSR
models derived from different conformations of peptides suggest that the multiple
conformers of the peptide β-turn contribute to the stereodetermining transition
structures. The utility of this approach is evident since the conformational complexity of the catalyst structure precludes the calculation of competing transition structures for all of the structural variants studied.
Denmark used a machine learning algorithm – a deep feedforward neural
network – to construct quantitative models to predict enantioselectivity of thiol
additions catalysed by chiral phosphoric acids (Fig. 14) [106]. His work is an
exemplar of a complete chiral ligand design workflow, consisting of (1) defining a
synthetically accessible virtual library of the chiral ligands, (2) defining a universal
training set (UTS) as a recommended starting point to synthesize and test the ligands,
(3) generating descriptors and models and (4) validating the models using internal
and external validations.
Ligand Design for Asymmetric Catalysis: Combining Mechanistic and. . .
179
insight to alternative reaction pathways.
5.3.10 Model Utilization
Utilization of regression is categorized into two types: prediction and physical
interpretation [93]. Often in QSAR, one trades a meaningful model for a highly
predictive one. However, in QSSR for asymmetric catalysis, in some cases, applications of regression can lie in the sweet spot where the models have been illustrated
to be both predictive and not too complicated which allows for meaningful
interpretations [94].
Chemists have used QSSR to model, predict and optimize levels of
enantioselectivity since the turn of the twenty-first century [58, 59, 62, 85, 95,
96]. Attention to this approach continues to grow [97–101]. We select recent
examples at the frontier of QS(A/S)R applied to asymmetric catalysis to discuss
here. Examples of work on MIF-based 3D-QSSR have been comprehensively
discussed in a recent review by Denmark and co-workers and are not included
here [56].
5.4 Predicting Enantioselectivity
Sigman has pioneered modern applications of QSSR to asymmetric transformations,
publishing numerous examples of the use of regression-assisted reaction optimization and mechanistic elucidation [102–104]. Recently, in collaboration with Miller,
the Sigman laboratory has carried out multivariate modelling for the atroposelective
bromination of arylquinazolinones involving a tetrapeptide catalyst [105]. Predicted
product enantioselectivity was expressed in terms of calculated parameters: NBO
charges, IR stretches, a multidimensional steric Sterimol (L ) parameter and a
crystallographically derived parameter; main chain angles (Fig. 13). The QSSR
models derived from different conformations of peptides suggest that the multiple
conformers of the peptide β-turn contribute to the stereodetermining transition
structures. The utility of this approach is evident since the conformational complexity of the catalyst structure precludes the calculation of competing transition structures for all of the structural variants studied.
Denmark used a machine learning algorithm – a deep feedforward neural
network – to construct quantitative models to predict enantioselectivity of thiol
additions catalysed by chiral phosphoric acids (Fig. 14) [106]. His work is an
exemplar of a complete chiral ligand design workflow, consisting of (1) defining a
synthetically accessible virtual library of the chiral ligands, (2) defining a universal
training set (UTS) as a recommended starting point to synthesize and test the ligands,
(3) generating descriptors and models and (4) validating the models using internal
and external validations.
Ligand Design for Asymmetric Catalysis: Combining Mechanistic and. . .
179