Predicting CYP2C19 catalytic parameters for enantioselective oxidations using artificial neural networks and a chirality code

Machine Learning Classification of One-Chiral-Center Organic Molecules According to Optical Rotation

In this study, machine learning algorithms were investigated for the classification of organic molecules with one carbon chiral center according to the sign of optical rotation. Diverse heterogeneous data sets comprising up to 13,080 compounds and their corresponding optical rotation were retrieved from Reaxys and processed independently for three solvents: dichloromethane, chloroform, and methanol. The molecular structures were represented by chiral descriptors based on the physicochemical and topological properties of ligands attached to the chiral center. The sign of optical rotation was predicted by random forests (RF) and artificial neural networks for independent test sets with an accuracy of up to 75% for dichloromethane, 82% for chloroform, and 82% for methanol. RF probabilities and the availability of structures in the training set with the same spheres of atom types around the chiral center defined applicability domains in which the accuracy is higher.

Machine learning to predict the specific optical rotations of chiral fluorinated molecules

A chemoinformatics method was applied to the assignment of absolute configurations and to the quantitative prediction of specific optical rotations using a data set of 88 chiral fluorinated molecules (44 pairs of enantiomers). Counterpropagation neural networks were explored for the classification of enantiomers as dextrorotatory or levorotatory. Regression models were trained using multilayer perceptrons (MLP), random forests (RF) or multilinear regressions (MLR), on the basis of physicochemical atomic stereo (PAS) descriptors. New descriptors were also derived considering the common structural features of the data set (cPAS descriptors), which enabled RF models to predict the whole data set with R = 0.964, mean absolute error (MAE) of 9.8° and root mean square error (RMSE) of 12.5° in leave-one-pair-out cross-validation experiments. The predictions for the 30 compounds measured in chloroform were obtained with R = 0.971, MAE = 9.1° and RMSE = 12.5°, which compares favorably with quantum chemistry calculations reported in the literature.

Learning a Local-Variable Model of Aromatic and Conjugated Systems

A collection of new approaches to building and training neural networks, collectively referred to as deep learning, are attracting attention in theoretical chemistry. Several groups aim to replace computationally expensive ab initio quantum mechanics calculations with learned estimators. This raises questions about the representability of complex quantum chemical systems with neural networks. Can local-variable models efficiently approximate nonlocal quantum chemical features? Here, we find that convolutional architectures, those that only aggregate information locally, cannot efficiently represent aromaticity and conjugation in large systems. They cannot represent long-range nonlocality known to be important in quantum chemistry. This study uses aromatic and conjugated systems computed from molecule graphs, though reproducing quantum simulations is the ultimate goal. This task, by definition, is both computable and known to be important to chemistry. The failure of convolutional architectures on this focused task calls into question their use in modeling quantum mechanics. To remedy this heretofore unrecognized deficiency, we introduce a new architecture that propagates information back and forth in waves of nonlinear computation. This architecture is still a local-variable model, and it is both computationally and representationally efficient, processing molecules in sublinear time with far fewer parameters than convolutional networks. Wave-like propagation models aromatic and conjugated systems with high accuracy, and even models the impact of small structural changes on large molecules. This new architecture demonstrates that some nonlocal features of quantum chemistry can be efficiently represented in local variable models.

Modeling, optimization and experimental studies of supported nano-bimetallic catalyst for simultaneous total conversion of toluene and cyclohexane in air using a hybrid intelligent algorithm

This study reveals the simultaneous deep oxidation of toluene and cyclohexane over optimal supported bimetallic catalysts over almond shell based activated carbon. To the best of our knowledge, this study is the first to construct a hybrid intelligent model to predict and determine an optimal supported bimetallic catalyst to oxidize aromatic and aliphatic compounds in air. The effects of preparation and operating parameters, including oxidation temperature, initial concentration of VOCs, structure of the catalyst and metal oxide content, on VOCs conversion were studied by modeling a database containing 50 data points derived from our previously published study, by an artificial neural network (ANN). Reported experimental data were predicted by a feed-forward network with 11 neurons and tansig function in the hidden layer. The non-linear network demonstrated stronger influence of oxidation temperature and cobalt content on the complete conversion of toluene and cyclohexane in the mixture. A hybrid model containing a genetic algorithm (GA) and an ANN were employed to realize the optimum catalyst at constant operating conditions for the complete conversion of toluene and cyclohexane in air. A well dispersed optimal alloy catalyst with 8 wt% metal oxide content (2.5 wt% copper oxide and 5.5 wt% cobalt oxide) over activated carbon was synthesized by heterogeneous deposition–precipitation for the complete conversion of toluene (model = 95.50%, experimental = 96%) and cyclohexane (model = 91.88% and experimental = 91%), simultaneously. Characterizations of the optimal catalyst were carried out by XRD, TEM, ICP, FESEM and BET analyses to justify its highest performance.

This study reveals the simultaneous deep oxidation of toluene and cyclohexane over optimal supported bimetallic catalysts over almond shell based activated carbon.

Improving quantitative structure-activity relationship models using Artificial Neural Networks trained with dropout

Dropout is an Artificial Neural Network (ANN) training technique that has been shown to improve ANN performance across canonical machine learning (ML) datasets. Quantitative Structure Activity Relationship (QSAR) datasets used to relate chemical structure to biological activity in Ligand-Based Computer-Aided Drug Discovery pose unique challenges for ML techniques, such as heavily biased dataset composition, and relatively large number of descriptors relative to the number of actives. To test the hypothesis that dropout also improves QSAR ANNs, we conduct a benchmark on nine large QSAR datasets. Use of dropout improved both enrichment false positive rate and log-scaled area under the receiver-operating characteristic curve (logAUC) by 22-46 % over conventional ANN implementations. Optimal dropout rates are found to be a function of the signal-to-noise ratio of the descriptor set, and relatively independent of the dataset. Dropout ANNs with 2D and 3D autocorrelation descriptors outperform conventional ANNs as well as optimized fingerprint similarity search methods.

Prediction of cytochrome P450 mediated metabolism

Cytochrome P450 enzymes (CYPs) form one of the most important enzyme families involved in the metabolism of xenobiotics. CYPs comprise many isoforms, which catalyze a wide variety of reactions, and potentially, a large number of different metabolites can be formed. However, it is often hard to rationalize what metabolites these enzymes generate. In recent years, many different in silico approaches have been developed to predict binding or regioselective product formation for the different CYP isoforms. These comprise ligand-based methods that are trained on experimental CYP data and structure-based methods that consider how the substrate is oriented in the active site or/and how reactive the part of the substrate that is accessible to the heme group is. We will review key aspects for various approaches that are available to predict binding and site of metabolism (SOM), what outcome can be expected from the predictions, and how they could potentially be improved.