Predicting Solvent-Dependent Nucleophilicity Parameter with a Causal Structure Property Relationship

Abstract

Solvent-dependent reactivity is a key aspect of synthetic science, which controls reaction selectivity. The contemporary focus on new, sustainable solvents highlights a need for reactivity predictions in different solvents. Herein, we report the excellent machine learning prediction of the nucleophilicity parameter N in the four most-common solvents for nucleophiles in the Mayr’s reactivity parameter database (R 2 = 0.93 and 81.6% of predictions within ±2.0 of the experimental values with Extra Trees algorithm). A Causal Structure Property Relationship (CSPR) approach was utilized, with focus on the physicochemical relationships between the descriptors and the predicted parameters, and on rational improvements of the prediction models. The nucleophiles were represented with a series of electronic and steric descriptors and the solvents were represented with principal component analysis (PCA) descriptors based on the ACS Solvent Tool. The models indicated that steric factors do not contribute significantly, because of bias in the experimental database. The most important descriptors are solvent-dependent HOMO energy and Hirshfeld charge of the nucleophilic atom. Replacing DFT descriptors with Parameterization Method 6 (PM6) descriptors for the nucleophiles led to an 8.7-fold decrease in computational time, and an ∼10% decrease in the percentage of predictions within ±2.0 and ±1.0 of the experimental values.

Nucleophilicity Prediction Using Graph Neural Networks

“…To determine the optimal values of distance cutoff ( R cut ), number of interaction blocks ( T ), and feature dimensions of atom representations ( d of X t , Figure ) and the solvent embedding ( d sol ), Z-SchNet-CDFT with different hyperparameter combinations was built and evaluated using a fixed training:test split (90:10). The training:test split was from Nguyen’s study, with proportional data from different nucleophile types. Each hyperparameter combination was tested 10 times to get average metrics.…”

“…In addition to electronic descriptors, such as the highest occupied molecular orbital (HOMO) energy, some structural descriptors such as the buried volume were also calculated based on the optimized structure of the intermediate from the reaction of the nucleophile with a proton. Very recently, Nguyen and co-workers curated a more complex data set consisting of 904 nucleophiles in the Mayr reactivity database, covering a wide range of chemical space (17 types of nucleophiles) . An Extra Trees (ET)-based model was developed to predict the nucleophilicity parameter.…”

A machine learning approach for predicting the empirical polarity of organic solvents

“…19 For initial model screening the dataset was split randomly into 90% training and 10% test sets as described in the previous approaches. 26 Since overfitting and underfitting problems are inherent to advanced ML models, several remedies have been devised to overcome these issues. k -Fold cross validation is one such method where training data is randomly split into k number of folds, followed by training the model on k − 1 folds and validation on the remaining fold.…”

“…Several quantitative structure activity relationship studies and ML methods have proven quantum chemical and thermodynamic descriptors to be important tools for predicting various chemical and physical properties of molecules. [23][24][25][26]30 For our predictive modelling application, density functional theory (DFT) based computational approaches were used as they enable the generation of various electronic and thermodynamic descriptors in the most efficient and reliable manner. 31,32 Since the solvents employed in our study belong to 13 diverse categories, choosing appropriate descriptors that accurately represents the wholesome properties of these molecules was a big challenge.…”

“…ML methods have also found application in chemical sciences and have emerged as powerful tools to predict various chemical and physical properties of organic molecules, particularly when the dataset is large and complex. [16][17][18][19] For example, ML models have been successfully employed to predict pK a , 20 toxicity, 21 solubility, 22,23 electrophilicity, 24 nucleophilicity, 25,26 and reactivity [27][28][29] of organic molecules that provides deep insight into their reactivity pattern, drug-likeliness, and physicochemical characteristics, which can assist a chemist in a faster and more accurate drug design. Considering the vast amount of experimental data available on the empirical polarity of organic liquids, we sought out to develop a ML model to accurately predict the E T (30) values by developing a novel dataset of computationally derived topological and molecular descriptors.…”