Predictive Toxicology Modeling: Protocols for Exploring hERG Classification and Tetrahymena pyriformis End Point Predictions

Abstract

The inclusion and accessibility of different methodologies to explore chemical data sets has been beneficial to the field of predictive modeling, specifically in the chemical sciences in the field of Quantitative Structure-Activity Relationship (QSAR) modeling. This study discusses using contemporary protocols and QSAR modeling methods to properly model two biomolecular systems that have historically not performed well using traditional and three-dimensional QSAR methodologies. Herein, we explore, analyze, and discuss the creation of a classification human Ether-a-go-go Related Gene (hERG) potassium channel model and a continuous Tetrahymena pyriformis (T. pyriformis) model using Support Vector Machine (SVM) and Support Vector Regression (SVR), respectively. The models are constructed with three types of molecular descriptors that capture the gross physicochemical features of the compounds: (i) 2D, 2 1/2D, and 3D physical features, (ii) VolSurf-like molecular interaction fields, and (iii) 4D-Fingerprints. The best hERG SVM model achieved 89% accuracy and the three-best SVM models were able to screen a Pubchem data set with an accuracy of 97%. The best T. pyriformis model had an R(2) value of 0.924 for the training set and was able to predict the continuous end points for two test sets with R(2) values of 0.832 and 0.620, respectively. The studies presented within demonstrate the predictive ability (classification and continuous end points) of QSAR models constructed from curated data sets, biologically relevant molecular descriptors, and Support Vector Machines and Support Vector Regression. The ability of these protocols and methodologies to accommodate large data sets (several thousands compounds) that are chemically diverse - and in the case of classification modeling unbalanced (one experimental outcome dominates the data set) - allows scientists to further explore a remarkable amount of biological and chemical information.

In Silico Prediction of hERG Inhibition

“…The model was tested using an external test set of 66 compounds and 94% of them were predicted correctly, Table 4. Recently, Su et al [78] published a SVM model based on a dataset of 210 hERG inhibitors and 336 noninhibitors. Three groups of descriptors were applied to capture more bindingrelevant features: 228 2D MOE molecular descriptors describing various physicochemical features; 73 Vol-Surf-like molecular interaction features [83] and 5241 4D-Fingerprints.…”

Prediction of hERG K+ channel blockage using deep neural networks

“…The Pubchem set (AID376) was chosen as the external validation set for further validation of the models. This data set was used widely as test set to evaluate the quality of models by many published papers (Bo‐Han et al, ; Chavan et al, ; Q. Li et al, ; Su et al, ; Wang et al, ). The reason why we chose this data set was to compare the performance of the optimal model M15 with other models on the same data set.…”

“…The human ether‐a‐go‐go‐related gene (hERG), which is highly expressed in heart tissue, is closely connected to the occurrence of cardiotoxicity (Didziapetris & Lanevskij, ; Schyman, Liu, & Wallqvist, ; Wang, Sun, et al, ). The blockade of hERG may cause long QT syndrome and Torsade de Pointes, which lead to palpitations, fainting, or even sudden death (Schyman et al, ) such as astemizole, terfenadine, cisapride, sertindole, terolidine, droperidol, lidoflazine, and grepafloxacin (Braga et al, ; Su, Tu, Esposito, & Tseng, ). Therefore, it is essential to ensure that the drug would not cause blockade of the hERG function before entering into the market.…”

“…Random forest (RF) (Polishchuk et al, 2009), which is the "gold standard" in QSAR researches, was also included in this study for comparison. The DNN models with three hidden layers were selected as the final strategy, which not only performed well on the test set, but also showed better performance on statistical parameters than other published models (Bo-Han et al, 2010;Chavan et al, 2016;Li et al, 2008;Su et al, 2012;Wang et al, 2012) on external Pubchem validation set.…”

Extensions to In Silico Bioactivity Predictions Using Pathway Annotations and Differential Pharmacology Analysis: Application to Xenopus laevis Phenotypic Readouts

“…A large number of machine learning algorithms are available,86 but only few of them have been extensively explored in the target prediction domain. Popular examples include the similarity‐based SEA,45 the probabilistic (usually multi‐class) Naive Bayes,5,43,87,88 the distance‐based Parzen‐Rosenblatt Window,5,6 and Support Vector Machines87,89 which use hyperplanes to separate data points. Apart from the choice of algorithm, the descriptors utilised to represent molecules significantly influence the outcome of the prediction.…”