Variable selection using pair-correlation method. Environmental applications

A review on machine learning methods for in silico toxicity prediction

In silico toxicity prediction plays an important role in the regulatory decision making and selection of leads in drug design as in vitro/vivo methods are often limited by ethics, time, budget, and other resources. Many computational methods have been employed in predicting the toxicity profile of chemicals. This review provides a detailed end-to-end overview of the application of machine learning algorithms to Structure-Activity Relationship (SAR)-based predictive toxicology. From raw data to model validation, the importance of data quality is stressed as it greatly affects the predictive power of derived models. Commonly overlooked challenges such as data imbalance, activity cliff, model evaluation, and definition of applicability domain are highlighted, and plausible solutions for alleviating these challenges are discussed.

Consistency of QSAR models: Correct split of training and test sets, ranking of models and performance parameters

Recent implementations of QSAR modelling software provide the user with numerous models and a wealth of information. In this work, we provide some guidance on how one should interpret the results of QSAR modelling, compare and assess the resulting models, and select the best and most consistent ones. Two QSAR datasets are applied as case studies for the comparison of model performance parameters and model selection methods. We demonstrate the capabilities of sum of ranking differences (SRD) in model selection and ranking, and identify the best performance indicators and models. While the exchange of the original training and (external) test sets does not affect the ranking of performance parameters, it provides improved models in certain cases (despite the lower number of molecules in the training set). Performance parameters for external validation are substantially separated from the other merits in SRD analyses, highlighting their value in data fusion.

Evaluation of chemometric techniques to select orthogonal chromatographic systems

Several chemometric techniques were compared for their performance to determine the orthogonality and similarity between chromatographic systems. Pearson's correlation coefficient (r) based color maps earlier were used to indicate selectivity differences between systems. These maps, in which the systems were ranked according to decreasing or increasing dissimilarities observed in the weighted-average-linkage dendrogram, were now applied as reference method. A number of chemometric techniques were evaluated as potential alternative (visualization) methods for the same purpose. They include hierarchical clustering techniques (single, complete, unweighted-average-linkage, centroid and Ward's method), the Kennard and Stone algorithm, auto-associative multivariate regression trees (AAMRT), and the generalized pairwise correlation method (GPCM) with McNemar's statistical test. After all, the reference method remained our preferred technique to select orthogonal and identify similar systems.

Quantitative Structure−Activity Relationship Modeling of Juvenile Hormone Mimetic Compounds for Culex Pipiens Larvae, with a Discussion of Descriptor-Thinning Methods

Quantitative structure-activity relationship (QSAR) modelers often encounter the problem of multicollinearity owing to the availability of large numbers of computable molecular descriptors. Sparsity of the variables while using descriptors such as atom pairs increases the complexity. Three different predictor-thinning methods, namely, a modified Gram-Schmidt algorithm, a marginal soft thresholding algorithm, and LASSO (least absolute shrinkage and selection operator), were utilized to reduce the number of descriptors prior to developing linear models. Juvenile hormone (JH) activity of 304 compounds on Culex pipiens larvae was taken as the model data set, and predictor trimming of a large number of diverse descriptors comprising 268 global molecular descriptors (topostructural, topochemical, and geometrical), 13 quantum chemical descriptors, and 915 atom pairs (substructural counts) was applied prior to linear regression by the ridge regression method. The data set (N = 304) was split into five calibration data sets of random samples of sizes 60/110/160/210/260, and the remaining 244/194/144/94/44 compounds were used for validations. LASSO was not found to be a very effective method in handling a large set of descriptors because the number of predictors retained could not exceed the number of observations. The results indicated that the modified Gram-Schmidt algorithm could be used to trim the number of predictors in the global molecular descriptor set where collinearity of the descriptors was the major concern. On the contrary, the soft thresholding approach was found to be an effective tool in subset selection from a diverse set of descriptors having both sparsity and multicollinearity, as in the case of the combined set of atom pairs and global molecular descriptors. The final model developed after variable selection was dominated more by atom pairs, which indicated the important structural moieties that affect JH activity of the compounds. The success of the method reiterates the fact that QSAR or quantitative structure-property relationship (QSPR) models can be developed for a diverse set of compounds using properly parametrized and diverse sets of descriptors, of course, with the selection of the appropriate statistical tools.

Comparison of Ridge Regression, Partial Least-Squares, Pairwise Correlation, Forward- and Best Subset Selection Methods for Prediction of Retention Indices for Aliphatic Alcohols

A quantitative structure-retention relationship (QSRR) study based on multiple linear regression (MLR) was performed for the description and prediction of Kováts retention indices (RI) of alcohol compounds. Alcohols were of saturated, linear or branched types and contained a hydroxyl group on the primary, secondary or tertiary carbon atoms. Constitutive and weighted holistic invariant molecular (WHIM) descriptors were used to represent the structure of alcohols in the MLR models. Before the model building, five variable selection methods were applied to select the most relevant variables from a large set of descriptors, respectively. The selected molecular properties were included into the MLR models. The efficiency of the variable selection methods was also compared. The selection methods were as follows: ridge regression (RR), partial least-squares method (PLS), pair-correlation method (PCM), forward selection (FS) and best subset selection (BSS). The stability and the validity of the MLR models were tested by a cross-validation technique using a leave-n-out technique. Neither RR nor PLS selected variables were able to describe the Kováts retention index properly, and PCM gave reliable results in the description but not for prediction. We built models with good predicting ability using FS and BSS as a selection method. The most relevant variables in the description and prediction of RIs were the mean electrotopological state index, the molecular mass, and WHIM indices characterizing size and shape.