Dreams, False Starts, Dead Ends, and Redemption: A Chronicle of the Evolution of a Chemoinformatic Workflow for the Optimization of Enantioselective Catalysts.

Catalyst design in enantioselective catalysis has historically been driven by empiricism. In this endeavor, experimentalists attempt to qualitatively identify trends in structure that lead to a desired catalyst function. In this body of work, we lay the groundwork for an improved, alternative workflow that uses quantitative methods to inform decision making at every step of the process. At the outset, we define a library of synthetically accessible permutations of a catalyst scaffold with the philosophy that the library contains every potential catalyst we are willing to make. To represent these chiral molecules, we have developed general 3D representations, which can be calculated for tens of thousands of structures. This defines the total chemical space of a given catalyst scaffold; it is constructed on the basis of catalyst structure only without regard to a specific reaction or mechanism. As such, any algorithmic subset selection method, which is unsupervised (i.e., only considers catalyst structure), should provide an ideal initial screening set for any new reaction that can be catalyzed by that scaffold. Notably, because this design strategy, the same set of catalysts can be used for any reaction that can be catalyzed with that parent catalyst scaffold. These are tested experimentally, and statistical learning tools can be used to create a model relating catalyst structure to catalyst function. Further, this model can be used to predict the performance of each catalyst candidate in the greater database of virtual catalyst candidates. In this way, it is possible estimate the performance of tens of thousands of catalysts by experimentally testing a smaller subset. Using error assessment metrics, it is possible to understand the confidence in new predictions. An experimentalist using this tool can balance the predicted results (reward) with the prediction confidence (risk) when deciding which catalysts to synthesize next in an optimization campaign. These catalysts are synthesized and tested experimentally. At this stage, either the optimization is a success or the predicted values were incorrect and further optimization is required. In the case of the latter, the information can be fed back into the statistical learning model to refine the model, and this iterative process can be used to determine the optimal catalyst. In this body of work, we not only establish this workflow but quantitatively establish how best to execute each step. Herein, we evaluate several 3D molecular representations to determine how best to represent molecules. Several selection protocols are examined to best decide which set of molecules can be used to represent the library of interest. In addition, the number of reactions needed to make accurate, statistical learning models is evaluated. Taken together these components establish a tool ready to progress from the development stage to the utility stage. As such, current research endeavors focus on applying these tools to optimize new reactions.

Development of a Computer-Guided Workflow for Catalyst Optimization. Descriptor Validation, Subset Selection, and Training Set Analysis.

Modern, enantioselective catalyst development is driven largely by empiricism. Although this approach has fostered the introduction of most of the existing synthetic methods, it is inherently limited by the skill, creativity, and chemical intuition of the practitioner. Herein, we present a complementary approach to catalyst optimization in which statistical methods are used at each stage to streamline development. To construct the optimization informatics workflow, a number of critical components had to be subjected to rigorous validation. First, the critically important molecular descriptors were validated in two case studies to establish the importance of conformation-dependent molecular representations. Next, with a large data set available, it was possible to investigate the amount of data necessary to make predictive models with different modeling methods. Given the commercial availability of many catalyst structures, it was possible to compare models generated with algorithmically selected training sets and commercially available training sets. Finally, the augmentation of limited data sets is demonstrated in a method informed by unsupervised learning to restore the accuracy of the generated models.

Cautionary Guidelines for Machine Learning Studies with Combinatorial Datasets.

Regression modeling is becoming increasingly prevalent in organic chemistry as a tool for reaction outcome prediction and mechanistic interrogation. Frequently, to acquire the requisite amount of data for such studies, researchers employ combinatorial datasets to maximize the number of data points while limiting the number of discrete chemical entities required. An often-overlooked problem in modeling studies using combinatorial datasets is the tendency to fit on patterns in the datasets (i.e., the presence or absence of a reactant or catalyst) rather than to identify meaningful trends between descriptors and the response variable. Consequently, the generality and interpretability of such models suffer. This report illustrates these well-known pitfalls in a case study, demonstrates the necessary control experiments to identify when this property will be problematic, and suggests how to perform further validation to assess general applicability and interpretability of models trained using combinatorial datasets.

Prediction of higher-selectivity catalysts by computer-driven workflow and machine learning.

Catalyst design in asymmetric reaction development has traditionally been driven by empiricism, wherein experimentalists attempt to qualitatively recognize structural patterns to improve selectivity. Machine learning algorithms and chemoinformatics can potentially accelerate this process by recognizing otherwise inscrutable patterns in large datasets. Herein we report a computationally guided workflow for chiral catalyst selection using chemoinformatics at every stage of development. Robust molecular descriptors that are agnostic to the catalyst scaffold allow for selection of a universal training set on the basis of steric and electronic properties. This set can be used to train machine learning methods to make highly accurate predictive models over a broad range of selectivity space. Using support vector machines and deep feed-forward neural networks, we demonstrate accurate predictive modeling in the chiral phosphoric acid-catalyzed thiol addition to N-acylimines.