External Prediction Set Research Articles

A Classification Study of Respiratory Syncytial Virus (RSV) Inhibitors by Variable Selection with Random Forest

Experimental pEC50s for 216 selective respiratory syncytial virus (RSV) inhibitors are used to develop classification models as a potential screening tool for a large library of target compounds. Variable selection algorithm coupled with random forests (VS-RF) is used to extract the physicochemical features most relevant to the RSV inhibition. Based on the selected small set of descriptors, four other widely used approaches, i.e., support vector machine (SVM), Gaussian process (GP), linear discriminant analysis (LDA) and k nearest neighbors (kNN) routines are also employed and compared with the VS-RF method in terms of several of rigorous evaluation criteria. The obtained results indicate that the VS-RF model is a powerful tool for classification of RSV inhibitors, producing the highest overall accuracy of 94.34% for the external prediction set, which significantly outperforms the other four methods with the average accuracy of 80.66%. The proposed model with excellent prediction capacity from internal to external quality should be important for screening and optimization of potential RSV inhibitors prior to chemical synthesis in drug development.

A Feasibility Study on Using near Infrared Spectroscopy to Classify Straw-Coal Blends

Rapid classification of biomass, coal and biomass-coal blends is very important for developing appropriate subsidy policies and methods for biomass co-firing power generation in China. The use of near infrared reflectance (NIR) spectroscopy to classify straw, coal and straw-coal blends was explored in this study. Eighty-one straw samples, 9 coal samples, 81 straw-coal blends samples with straw content from 91% to 99% (blends1) and 90 straw-coal blends samples with straw content from 1% to 30% (blends2) were prepared and separated into a calibration set and an external prediction set. Spectra were scanned by a Fourier transform-NIR spectrometer. Raw spectra of the samples were mean centred and compared using discriminant analysis. Overall, the correct classification percentages for straw, blends1, blends2 and coal samples were 98.8%, 90.1%, 83.3% and 66.7%, respectively. The results show that the NIR technique could provide excellent classification between straw and coal. The spectra of straw-coal blends with straw content less than 98% can be classified exactly from calibrations based on pure straw and the spectra of straw-coal blends with straw content greater than 20% can be well classified from calibrations based on pure coal. It is concluded that NIR is a feasible method for rapid classification of straw, coal and straw-coal blends.

Confirmation of brand identity in foods by near infrared transflectance spectroscopy using classification and class-modelling chemometric techniques — The example of a Belgian beer

Food and beverage processors require tools to monitor conformance of finished goods to their defined specification; regulatory authorities need appropriate methods for detecting retail fraud. In this report, samples (n = 275) of Belgian and other European beers were collected and analysed using near infrared transflectance spectroscopy; three class-modelling techniques (soft independent modelling of class analogy, SIMCA; potential functions techniques, POTFUN; and unequal dispersed classes, UNEQ) were employed to characterise beer types (firstly Trappist and then Rochefort) while a classification method (partial least squares discriminant analysis, PLS-DA) was applied to discriminate between two final beer classes: Rochefort 8° and Rochefort 10°. The class-models and the classification rules developed were validated by means of an external prediction set. A discussion on the appropriate use of these chemometric approaches is included. Modelling of Trappist beers met with limited success while model efficiencies for Rochefort samples were highest for SIMCA and UNEQ applications i.e. 81.4% and 84.5% respectively. The classification of beers as Rochefort 8˚ or Rochefort 10˚ was possible with an average correct classification rate of 93.4%.

Prediction of PKCθ inhibitory activity using the Random Forest Algorithm.

This work is devoted to the prediction of a series of 208 structurally diverse PKCθ inhibitors using the Random Forest (RF) based on the Mold2 molecular descriptors. The RF model was established and identified as a robust predictor of the experimental pIC50 values, producing good external R2pred of 0.72, a standard error of prediction (SEP) of 0.45, for an external prediction set of 51 inhibitors which were not used in the development of QSAR models. By using the RF built-in measure of the relative importance of the descriptors, an important predictor—the number of group donor atoms for H-bonds (with N and O)—has been identified to play a crucial role in PKCθ inhibitory activity. We hope that the developed RF model will be helpful in the screening and prediction of novel unknown PKCθ inhibitory activity.

Quantitative structure-retention relationship (QSRR) models for predicting the GC retention times of essential oil components

Summary Multiple linear regression (MLR) and partial least squares (PLS) analysis have been used to model gas chromatographic retention times (tR) of 77 components of the essential oil of Ottonia martiana. The genetic algorithm (GA) was used to select the variables that resulted in the best-fitting models. Appropriate models with low standard errors and high correlation coefficients were obtained. MLR and PLS analysis were performed to derive the best quantitative structure-retention relationship (QSRR) models. The predictive quality of the QSRR models was tested for an external prediction set of 15 compounds, randomly chosen from the 77 compounds. Surprisingly, the results were of approximately the same quality for MLR and PLS modeling [squared regression coefficients (R2) of 0.964 and 0.968, respectively].

Structural Contributions of Substrates to their Binding to P-Glycoprotein. A TOPSMODE Approach

A topological substructural molecular design approach (TOPS-MODE) has been used to formulate structural rules for binding of substrates of P-glycoprotein (P-gp). We first review some of the models developed in the recent literature for predicting binding to P-gp. Then, we develop a model using TOPS-MODE, which is able to identify 88.4% of substrates and 84.2% of non-substrates. When the model is presented to an external prediction set of 100 substrates and 77 nonsubstrates it identifies correctly 81.8% of all cases. Using TOPS-MODE strategy we found structural contributions for binding to P-gp, which identifies 24 structural fragments responsible for such binding. We then carried out a chemico-biological analysis of some of the structural fragments found as contributing to P-gp binding of substrates. We show that in general the model developed so far can be used as a virtual screening method for identifying substrates of P-gp from large libraries of compounds.

Recognition of tablet content by chemometric processing of differential scanning calorimetry curves—An acetaminophen example

Chemometric processing of differential scanning calorimetry (DSC) curves is rarely approached in the literature. This paper discusses and presents an example of building discriminant models against the presence of a compound in pharmaceutical formulations. The dataset containing 11 curves of pure substances and 21 signals of various OTC tablets was processed by partial least squares discriminant analysis (PLS-DA) against detection of these substances in the formulations. Good model was found in the case of acetaminophen (5 PLS factors included), due to good proportion between positive and negative training samples. The model was validated on external prediction set and was found to have following parameters: RMS equal to 0.1068 (98.8% of explained variance), RMSECV equal to 0.148 (97.7% of explained variance) and external predictive error (RMSEP) equal to 0.3918 (86.5% of explained variance).

Carbon nuclear magnetic resonance spectroscopic fingerprinting of commercial gasoline: Pattern-recognition analyses for screening quality control purposes

In this work, the combination of carbon nuclear magnetic resonance ( 13C NMR) fingerprinting with pattern-recognition analyses provides an original and alternative approach to screening commercial gasoline quality. Soft Independent Modelling of Class Analogy (SIMCA) was performed on spectroscopic fingerprints to classify representative commercial gasoline samples, which were selected by Hierarchical Cluster Analyses (HCA) over several months in retails services of gas stations, into previously quality-defined classes. Following optimized 13C NMR-SIMCA algorithm, sensitivity values were obtained in the training set (99.0%), with leave-one-out cross-validation, and external prediction set (92.0%). Governmental laboratories could employ this method as a rapid screening analysis to discourage adulteration practices.

Prediction of the Q-e parameters from radical structures

To construct reliable quantitative structure-property relationship models of Q and e parameters, density functional theory (DFT) calculations were carried out for 56 radicals with structure C1H3–C2HR3• (formed from the monomer + H•) at the B3LYP level of theory with 6-31G(d) basis set. The multiple linear regression technique was used to select the descriptors as inputs for the artificial neural network (ANN). A three-descriptor ANN for the parameter lnQ was developed that produced training set root mean square (rms) error = 0.317 and rms error = 0.313 for an external prediction set. A two-descriptor ANN for the parameter e was built with rms errors of 0.264 for the training set and 0.271 for the prediction set. Comparing to existing models, our ANN models show better or comparable statistical characteristics, which shows that calculating quantum chemical descriptors from radicals to develop ANN models and predict lnQ and e values is feasible.

Multivariate image analysis-thin layer chromatography (MIA-TLC) for simultaneous determination of co-eluting components

This paper addresses the solution of peak overlapping, as a fundamental problem in TLC, by multivariate analysis of the images recorded by a digital camera. We report the results of our study on the application of multivariate image analysis (MIA) for simultaneous determination of several species on thin layer chromatography (TLC) sheet for the first time. An imaging system, composed of a dark cabinet, a digital camera and a multivariate image analysis program, was prepared for recording the images of TLC plates after development of a multi-component solution. The written program was able to produce 2- and 3-dimensional chromatograms of the solutions, which were subsequently used as inputs of partial least squares, as an efficient multivariate calibration method. The ability of the proposed MIA-TLC method for simultaneous determination of the co-eluting components was validated by analysis of ternary synthetic mixtures of indicators of highly overlapped chromatograms (i.e., methyl yellow, bromocresol green and creseol red) and a real mixture of nifedipine and its photo-degradation product. By application of different strategies like principal component analysis and variable selection, models were obtained that could estimate the concentration of indicators in the external prediction set with relative errors of lower than 10% and in most cases lower than 5%.

The importance of molecular structures, endpoints’ values, and predictivity parameters in QSAR research: QSAR analysis of a series of estrogen receptor binders

Quantitative structure-activity relationship (QSAR) methodology aims to explore the relationship between molecular structures and experimental endpoints, producing a model for the prediction of new data; the predictive performance of the model must be checked by external validation. Clearly, the qualities of chemical structure information and experimental endpoints, as well as the statistical parameters used to verify the external predictivity have a strong influence on QSAR model reliability. Here, we emphasize the importance of these three aspects by analyzing our models on estrogen receptor binders (Endocrine disruptor knowledge base (EDKB) database). Endocrine disrupting chemicals, which mimic or antagonize the endogenous hormones such as estrogens, are a hot topic in environmental and toxicological sciences. QSAR shows great values in predicting the estrogenic activity and exploring the interactions between the estrogen receptor and ligands. We have verified our previously published model for additional external validation on new EDKB chemicals. Having found some errors in the used 3D molecular conformations, we redevelop a new model using the same data set with corrected structures, the same method (ordinary least-square regression, OLS) and DRAGON descriptors. The new model, based on some different descriptors, is more predictive on external prediction sets. Three different formulas to calculate correlation coefficient for the external prediction set (Q2 EXT) were compared, and the results indicated that the new proposal of Consonni et al. had more reasonable results, consistent with the conclusions from regression line, Williams plot and root mean square error (RMSE) values. Finally, the importance of reliable endpoints values has been highlighted by comparing the classification assignments of EDKB with those of another estrogen receptor binders database (METI): we found that 16.1% assignments of the common compounds were opposite (20 among 124 common compounds). In order to verify the real assignments for these inconsistent compounds, we predicted these samples, as a blind external set, by our regression models and compared the results with the two databases. The results indicated that most of the predictions were consistent with METI. Furthermore, we built a kNN classification model using the 104 consistent compounds to predict those inconsistent ones, and most of the predictions were also in agreement with METI database.

Use of Self-Training Artificial Neural Networks in a QSRR Study of a Diverse Set of Organic Compounds

For a very diverse set of toxicologically compounds, the gas chromatographic Kovats retention indices have been modeled using chemometric methods. First, a genetic algorithm–multiple linear regression (GA–MLR) model has been obtained using molecular descriptors. Then, 15 selected descriptors in the GA–MLR model have been used as input for a self-training artificial neural network (STANN). STANN has been developed as a faster and more accurate non-linear method in our laboratory. After optimization, a 15-9-1 STANN was generated for prediction of retention indices of these organic compounds. The predictive quality of the STANN model was tested for an external prediction set and also five leave-multiple-outs cross-validation sets. Obtained results showed the ability of developed STANN model for predicting retention indices of various compounds. Also, obtained results indicate that in this QSRR study, genetic algorithm is a suitable method for selecting the molecular descriptors.

A segmented principal component analysis–regression approach to quantitative structure–activity relationship modeling

The major problem associated with application of principal component regression (PCR) in QSAR studies is that this model extracts the eigenvectors solely from the matrix of descriptors, which might not have essentially good relationship with the biological activity. This article describes a novel segmentation approach to PCR (SPCAR), in which the descriptors are firstly segmented to different blocks and then principal component analysis (PCA) is applied on each segment to extract significant principal components (PCs). In this way, the PCs having useful and redundant information are separated. A linear regression analysis based on stepwise selection of variables is then employed to connect a relationship between the informative extracted PCs and biological activity. The proposed method was first applied to model the aqueous toxicity of aliphatic compounds. The effect of the number of segments on the prediction ability of the method was investigated. Finally, a correlation analysis was achieved to identify those descriptors having significant contribution in the selected PCs and in aqueous toxicity. The proposed method was further validated by the analysis of Selwood data set consisting of 31 compounds and 53 descriptors. A comparison between the conventional PCR algorithm and SPCAR reveals the superiority of the latter. For external prediction set, SPCAR represented all requirements to be considered as predicted model whereas PCR did not. In addition, a comparison was made between the models obtained by SPCAR and those reported previously.

Structure-Activity Relationships for Serotonin Transporter and Dopamine Receptor Selectivity

Antipsychotic medications have a diverse pharmacology with affinity for serotonergic, dopaminergic, adrenergic, histaminergic and cholinergic receptors. Their clinical use now also includes the treatment of mood disorders, thought to be mediated by serotonergic receptor activity. The aim of our study was to characterise the molecular properties of antipsychotic agents, and to develop a model that would indicate molecular specificity for the dopamine (D(2)) receptor and the serotonin (5-HT) transporter. Back-propagation artificial neural networks (ANNs) were trained on a dataset of 47 ligands categorically assigned antidepressant or antipsychotic utility. The structure of each compound was encoded with 63 calculated molecular descriptors. ANN parameters including hidden neurons and input descriptors were optimised based on sensitivity analyses, with optimum models containing between four and 14 descriptors. Predicted binding preferences were in excellent agreement with clinical antipsychotic or antidepressant utility. Validated models were further tested by use of an external prediction set of five drugs with unknown mechanism of action. The SAR models developed revealed the importance of simple molecular characteristics for differential binding to the D(2) receptor and the 5-HT transporter. These included molecular size and shape, solubility parameters, hydrogen donating potential, electrostatic parameters, stereochemistry and presence of nitrogen. The developed models and techniques employed are expected to be useful in the rational design of future therapeutic agents.

Quantitative Structure−Property Relationship Estimation of Cation Binding Affinity of the Common Amino Acids

The quantitative structure-property relationship (QSPR) methodology is applied to estimate the binding affinity of lithium, sodium, potassium, copper, and silver cations to the 20 common amino acids. The proposed model, nonlinearly derived from computational neural networks (CNN), contains seven descriptors and was validated by an external prediction set. Good results are obtained with correlation coefficients, R(2), and root-mean-square errors (rms) (kJ/mol) of 0.998 (3.89), 0.999 (2.86), and 0.997 (3.90) for the training, prediction, and validation sets, respectively. Five of the descriptors of the model correspond to the amino acids and the other two to the cations; they encode information clearly related to the cation-amino acid interactions responsible for the binding affinity values analyzed. A detailed analysis of results shows that, despite the different nature of the bonding between the metal cations and the amino acids, the neural networks used are capable of predicting accurately the property studied.

QSPR Study of the Distribution Coefficient Property for Hydantoin and 5‐Arylidene Derivatives. A Genetic Algorithm Application for the Variable Selection in the MLR and PLS Methods

AbstractA quantitative structure‐property relationship (QSPR) analysis has been performed on the 5‐arylidene derivatives of hydantoin. Modeling the distribution coefficient property of these compounds as a function of the theoretically derived descriptors was established by multiple linear regressions (MLR) and partial least squares (PLS) regression. The genetic algorithm (GA) was used for the selection of variables, which resulted in the best‐fitted models. After the selection of the variables, the MLR and PLS methods were applied with a leave‐one‐out cross validation, for building the regression models. The predictive quality of the QSPR models was tested for an external prediction set of 9 compounds randomly chosen among 48 compounds. The PLS regression method was applied to model the structure‐distribution coefficient relationship more accurately. This is, to the best of our knowledge, the first report of a QSPR study with distribution coefficient (log D65, octanol/water) relationships. However, the results surprisingly demonstrated almost identical qualities for the MLR and PLS modelings, according to the squared regression coefficients R2, which were 0.975 and 0.976 for MLR and PLS, respectively.

Quantitative Structure–Property Relationship Studies for Predicting Flash Points of Organic Compounds using Support Vector Machines

AbstractA Quantitative Structure–Property Relationship (QSPR) model was developed to predict the flash points of organic compounds. The widely used group contribution method was employed, and a new collection of 57 functional groups were selected as the molecular descriptors. The new chemometrics method of Support Vector Machine (SVM) was employed for fitting the possible quantitative relationship that existed between these functional groups and flash points. A total of 1282 organic compounds of various chemical families were used and randomly divided into a training set (1026) and an external prediction set (256).The optimum parameters of the SVM were obtained by employing the leave‐one‐out cross‐validation method. Simulated with the final optimum SVM, the results show that most of the predicted flash point values are in good agreement with the experimental data, with the average absolute error being 6.894 K, and the root mean square error being 11.367 for the whole dataset, which are lower than those obtained by previous works. Moreover, by employing the convenient group contribution method as well as the large modeling dataset, the presented model is also expected to be simple to apply and with a wide applicability range.

Accurate Prediction of Aquatic Toxicity of Aromatic Compounds Based on Genetic Algorithm and Least Squares Support Vector Machines

AbstractQuantitative Structure–Toxicity Relationship (QSTR) plays an important role in ecotoxicology for its fast and practical ability to assess the potential negative effects of chemicals. The aim of this investigation was to develop accurate QSTR models for the aquatic toxicity of a large set of aromatic compounds through the combination of Least Squares Support Vector Machines (LS‐SVMs) and a Genetic Algorithm (GA). Molecular descriptors calculated by DRAGON package and log P were used to describe the molecular structures. The most relevant descriptors used to build QSTR models were selected by a GA‐Variable Subset Selection procedure. Multiple Linear Regression (MLR) and nonlinear LS‐SVMs methods were employed to build QSTR models. The predictive ability of the derived models was validated using both the test set, selected from the whole data set by the Kennard–Stone Algorithm, and an external prediction set. The model applicability domain was checked by the leverage approach and the external prediction set was used to verify the predictive reliability of the models. The results indicated that the proposed QSTR models are robust and satisfactory, and can provide a feasible and promising tool for the rapid assessment of the toxicity of chemicals.

QSPR Prediction of pKa for Benzoic Acids in Different Solvents

AbstractA Computational Neural Network (CNN) derived model is proposed for the pKa prediction of benzoic acids in different solvents. The system studied contains 519 pKa values corresponding to 136 benzoic acids determined in water and in 8 organic solvents. The benzoic acids were described by the usual molecular descriptors and the solvents by a number of physical properties and by several parameters of the most widely used polarity solvent scales. The model is composed of seven descriptors – five of them corresponding to the solute and the other two to the solvents – and was validated by an external prediction set. The three sets of values needed in the analysis, training, prediction, and cross‐validation, have the same squared correlation coefficient (0.998) and Root‐Mean Square Error (RMSE) (0.21) values. The robustness of the model is also given for the statistical results for small subsets, such as ortho‐ and non‐ortho‐substituted acids, those obtained in protic or aprotic solvents, those obtained in each solvent, and even those for the set of pKa values of one specific acid in several solvents. The descriptors encode information about the chemical nature of the solutes and the solvents that is clearly related to the interactions present in the dissociation process in solution. The derived model also has the ability to predict pKa values of a larger set of aromatic neutral acids, containing both phenols and benzoic acids.

Artificial Neural Networks in ADMET Modeling: Prediction of Blood–Brain Barrier Permeation

AbstractA supervised artificial neural network (ANN) model has been developed for the accurate prediction of the Blood–Brain Barrier (BBB) partition (in Log BB scale) of chemical compounds. A structural diverse set of 108 compounds of known experimental Log BB value was chosen for this study. The molecules were defined by means of a non‐supervised neural network using our CODES program. This program codifies each molecule into a set of numerical parameters taking into account exclusively the information of its chemical structure from its Simplified Molecular Input Line System (SMILES) code. The model obtained averages 83% of accuracy in the training set and of 73% in the external prediction set. The model is able to predict correctly the behavior of a very heterogeneous series of compounds in terms of the BBB permeation. The results indicate that this approach may represent a useful tool for the prediction of Absorption, Distribution, Metabolism, Excretion and Toxicity (ADMET) properties. CODES© is available free of charge for academic institutions.