Use of artificial neural networks in a QSAR study of anti-HIV activity for a large group of HEPT derivatives

Design of novel leads: ligand based computational modeling studies on non-nucleoside reverse transcriptase inhibitors (NNRTIs) of HIV-1

Researchers are on the constant lookout for new antiviral agents for the treatment of AIDS. In the present work, ligand based modeling studies are performed on analogues of substituted phenyl-thio-thymines, which act as non-nucleoside reverse transcriptase inhibitors (NNRTIs) and novel leads are extracted. Using alignment-dependent descriptors, based on group center overlap (SALL, HDALL, HAALL and RALL), an alignment-independent descriptor (S log P), a topological descriptor (Balaban index (J)) and a 3D descriptor dipole moment (μ) and shape based descriptors (Kappa 2 index ((2)κ)), a correlation is derived with inhibitory activity. Linear and non-linear techniques have been used to achieve the goal. Support Vector Machine (SVM, R = 0.929, R(2) = 0.863) and Back Propagation Neural Network (BPNN, R = 0.928, R(2) = 0.861) methods yielded near similar results and outperformed Multiple Linear Regression (MLR, R = 0.915, R(2) = 0.837). The predictive ability of the models are cross-validated using a test dataset (SVM: R = 0.846, R(2) = 0.716, BPNN: R = 0.841, R(2) = 0.707 and MLR: R = 0.833, R(2) = 0.694). It is concluded that the hydrophobicity (S log P) and the polarity (μ) of a ligand and the presence of hydrogen donor (HDALL) moieties are the deciding factors in improving antiviral activity and pharmaco-therapeutic properties. Based on the above findings, a virtual dataset is created to extract probable leads with reasonable antiviral activity as well as better pharmacophoric properties.

QSAR modeling of aromatase inhibition by flavonoids using machine learning approaches

Aromatase is a member of the cytochrome P450 family responsible for catalyzing the rate-limiting conversion of androgens to estrogens. In the pursuit of robust aromatase inhibitors, quantitative structure-activity relationship (QSAR) and classification structure-activity relationship (CSAR) studies were performed on a non-redundant set of 63 flavonoids using multiple linear regression, artificial neural network, support vector machine and decision tree approaches. Easy-to-interpret descriptors providing comprehensive coverage on general characteristics of molecules (i.e., molecular size, flexibility, polarity, solubility, charge and electronic properties) were employed to describe the unique physicochemical properties of the investigated flavonoids. QSAR models provided good predictive performance as observed from their statistical parameters with Q values in the range of 0.8014 and 0.9870 for the cross-validation set and Q values in the range of 0.8966 and 0.9943 for the external test set. Furthermore, CSAR models developed with the J48 algorithm are able to accurately classify flavonoids as active and inactive as observed from the percentage of correctly classified instances in the range of 84.6 % and 100 %. The study presented herein represents the first large-scale QSAR study of aromatase inhibition on a large set of flavonoids. Such investigations provide an important insight on the origins of aromatase inhibitory properties of flavonoids as breast cancer therapeutics.

Estimation of Anti-HIV Activity of HEPT Analogues Using MLR, ANN, and SVM Techniques

The present study deals with the estimation of the anti-HIV activity (log1/C) of a large set of 107 HEPT analogues using molecular descriptors which are responsible for the anti-HIV activity. The study has been undertaken by three techniques MLR, ANN, and SVM. The MLR model fits the train set with R²=0.856 while in ANN and SVM with higher values of R² = 0.850, 0.874, respectively. SVM model shows improvement to estimate the anti-HIV activity of trained data, while in test set ANN have higher R² value than those of MLR and SVM techniques. R_m² = metrics and ridge regression analysis indicated that the proposed four-variable model MATS5e, RDF080u, T(O⋯O), and MATS5m as correlating descriptors is the best for estimating the anti-HIV activity (log 1/C) present set of compounds.

4D-QSAR study of HEPT derivatives by electron conformational-genetic algorithm method

In this work, the EC-GA method, a hybrid 4D-QSAR approach that combines the electron conformational (EC) and genetic algorithm optimization (GA) methods, was applied in order to explain pharmacophore (Pha) and predict anti-HIV-1 activity by studying 115 compounds in the class of 1-[(2-hydroxyethoxy)-methyl]-6-(phenylthio) thymine (HEPT) derivatives as non-nucleoside reverse transcriptase inhibitors (NNRTIs). The series of NNRTIs were partitioned into four training and test sets from which corresponding quantitative structure-activity relationship (QSAR) models were constructed. Analysis of the four QSAR models suggests that the three models generated from the training and test sets used in previous works yielded comparable results with those of previous studies. Model 4, the data set of which was partitioned randomly into two training and test sets with 11 descriptors, including electronical and geometrical parameters, showed good statistics both in the regression (r2(training) )= 0.867, r2test = 0.923) and cross-validation (q (2) = 0.811, q2(ext1) = 0.909, q2(ext2) = 0.909) for the training set of 80 compounds and the test set of 27 compounds. The prediction of the anti-HIV-1 activity of HEPT compounds by means of the EC-GA method allowed for a quantitatively consistent QSAR model. In addition, eight novel compounds never tested experimentally have been designed theoretically using model 4.

Dechlorination of chlorinated compounds by Trametes versicolor ATCC 200801 crude laccase and quantitative structure-activity relationship of toxicity

Chlorinated compounds constitute an important class of xenobiotics. Crude laccase was produced using Trametes versicolor ATCC (200801) in potato dextrose broth, with wheat bran as an inducing medium, and its ability to dechlorinate eight compounds was determined. The compounds were 2-chlorophenol, 4-chlorophenol, 2,4-dichlorophenol, 2,6-dichlorophenol, 2,4,5-trichlorophenol, 2,4,6-trichlorophenol, heptachlor and pentachlorophenol. A range of parameters for the dechlorination of some compounds was tested, including incubation period, pH, initial substrate concentration, temperature, and enzyme quantity. The oxygen consumption was determined during each dechlorination process, under pre-determined optimum conditions. The changes in chemical structure of the compounds were also determined, by using FTIR analysis, following dechlorination of test chlorophenolics. Strong interactions were found to lead to the reactivity of hydroxyl groups in some cases and chlorine atoms were released from the benzene ring. The changes in compound toxicity were monitored before and after enzymatic treatment, using Microtox. Quantitative structure-activity relationships for the toxicity of the chlorinated compounds were developed. Consequently, the toxic activity of the test compounds was controlled by electrophilic index and electronic properties.

Application of ant colony optimization in development of models for prediction of anti-HIV-1 activity of HEPT derivatives

Quantitative structure-activity relationship models were derived for 107 analogs of 1-[(2-hydroxyethoxy) methyl]-6-(phenylthio)thymine, a potent inhibitor of the HIV-1 reverse transcriptase. The activities of these compounds were investigated by means of multiple linear regression (MLR) technique. An ant colony optimization algorithm, called Memorized_ACS, was applied for selecting relevant descriptors and detecting outliers. This algorithm uses an external memory based upon knowledge incorporation from previous iterations. At first, the memory is empty, and then it is filled by running several ACS algorithms. In this respect, after each ACS run, the elite ant is stored in the memory and the process is continued to fill the memory. Here, pheromone updating is performed by all elite ants collected in the memory; this results in improvements in both exploration and exploitation behaviors of the ACS algorithm. The memory is then made empty and is filled again by performing several ACS algorithms using updated pheromone trails. This process is repeated for several iterations. At the end, the memory contains several top solutions for the problem. Number of appearance of each descriptor in the external memory is a good criterion for its importance. Finally, prediction is performed by the elitist ant, and interpretation is carried out by considering the importance of each descriptor. The best MLR model has a training error of 0.47 log (1/EC(50)) units (R(2) = 0.90) and a prediction error of 0.76 log (1/EC(50)) units (R(2) = 0.88).

Comparative structure-toxicity relationship study of substituted benzenes to Tetrahymena pyriformis using shuffling-adaptive neuro fuzzy inference system and artificial neural networks

The purpose of this study was to develop the structure-toxicity relationships for a large group of 268 substituted benzene to the ciliate Tetrahymena pyriformis using mechanistically interpretable descriptors. The shuffling-adaptive neuro fuzzy inference system (Shuffling-ANFIS) has been successfully applied to select the important factors affecting the toxicity of substituted benzenes to T. pyriformis. The results of the proposed model were compared with the model of linear-free energy response surface and also the principal component analysis Bayesian-regularized neural network (PCA-BRANN) trained using the same data. The presented model shows a better statistical parameter in comparison with the previous models. The results of the model are promising and descriptive. Five descriptors of octanol-water partition coefficient (logP), bond information content (BIC0), number of R-CX-R (C-026), eigenvalue sum from Z weighted distance matrix (SEigZ) and fragment based polar surface area (PSA) selected by Shuffling-ANFIS reveal the role of hydrophobicity, electronic and steric interactions in the mechanism of toxic action. Sequential zeroing of weights (SZW) as a sensitivity analysis method revealed that the hydrophobicity and electronic interactions play a major role in toxicity of these compounds. Satisfactory results (q(2)=0.828 and RMSE=0.348) in comparison with the previous works indicate that the Shuffling-ANFIS-ANN technique is able to model a diverse chemical class with more than one mechanism of toxicity by using simple and interpretable descriptors. Shuffling-ANFIS can be used as powerful feature selection technique, because its application in prediction of toxicity potency results in good statistical and interpretable physiochemical parameters.

A ligand's-eye view of protein binding

Docking tools created for structure-based design and virtual screening have also been used to automate ligand alignment for comparative molecular field analysis (CoMFA). Models based on such alignments have been compared with those obtained based solely on shared ligand substructures, but such comparisons have generally failed to distinguish between conformational specification (alignment in the internal coordinate space) and embedding in a shared external frame of reference (Cartesian alignment). Here, large sets of inhibitors were docked into two cyclooxygenase and two reverse transcriptase crystal structures, and the poses generated were evaluated in terms of the CoMFA models they produced. Realigning the conformers obtained by docking by rigid-body rotation and translation to overlay their common substructures improved model statistics and interpretability, provided the protein structure used for docking was reasonably appropriate to the ligands being considered.

Neural networks in analytical chemistry

This chapter covers a part of the spectrum of neural-network uses in analytical chemistry. Different architectures of neural networks are described briefly. The chapter focuses on the development of three-layer artificial neural network for modeling the anti-HIV activity of the HETP derivatives and activity parameters (pIC50) of heparanase inhibitors. The use of a genetic algorithm-kernel partial least squares algorithm combined with an artificial neural network (GA-KPLS-ANN) is described for predicting the activities of a series of aromatic sulfonamides. The retention behavior of terpenes and volatile organic compounds and predicting the response surface of different detection systems are presented as typical applications of ANNs in chromatographic area. The use of ANNs is explored in electrophoresis with emphasizes on its application on peptide mapping. Simulation of the electropherogram of glucagons and horse cytochrome C is described as peptide models. This chapter also focuses on discussing the role of ANNs in the simulation of mass and 13C-NMR spectra for noncyclic alkenes and alkanes and lignin and xanthones, respectively.

Classification of small molecules by two- and three-dimensional decomposition kernels

MOTIVATION

Several kernel-based methods have been recently introduced for the classification of small molecules. Most available kernels on molecules are based on 2D representations obtained from chemical structures, but far less work has focused so far on the definition of effective kernels that can also exploit 3D information.

RESULTS

We introduce new ideas for building kernels on small molecules that can effectively use and combine 2D and 3D information. We tested these kernels in conjunction with support vector machines for binary classification on the 60 NCI cancer screening datasets as well as on the NCI HIV data set. Our results show that 3D information leveraged by these kernels can consistently improve prediction accuracy in all datasets.

AVAILABILITY

An implementation of the small molecule classifier is available from http://www.dsi.unifi.it/neural/src/3DDK.

Modeling robust QSAR

Quantitative Structure Activity Relationship (QSAR) is a term describing a variety of approaches that are of substantial interest for chemistry. This method can be defined as indirect molecular design by the iterative sampling of the chemical compounds space to optimize a certain property and thus indirectly design the molecular structure having this property. However, modeling the interactions of chemical molecules in biological systems provides highly noisy data, which make predictions a roulette risk. In this paper we briefly review the origins for this noise, particularly in multidimensional QSAR. This was classified as the data, superimposition, molecular similarity, conformational, and molecular recognition noise. We also indicated possible robust answers that can improve modeling and predictive ability of QSAR, especially the self-organizing mapping of molecular objects, in particular, the molecular surfaces, a method that was brought into chemistry by Gasteiger and Zupan.

Prediction of cytotoxicity data (CC(50)) of anti-HIV 5-phenyl-1-phenylamino-1H-imidazole derivatives by artificial neural network trained with Levenberg-Marquardt algorithm

A Levenberg-Marquardt algorithm trained feed-forward artificial neural network in quantitative structure-activity relationship (QSAR) was developed for modeling of cytotoxicity data for anti-HIV 5-phenyl-1-phenylamino-1H-imidazole derivatives. A large number of descriptors were calculated with Dragon software and a subset of calculated descriptors was selected with a stepwise regression as a feature selection technique. The 28 molecular descriptors selected by stepwise regression, as the most feasible descriptors, were used as inputs for feed-forward neural network. The neural network architecture and its parameters were optimized. The data were randomly divided into 31 training and 11 validation sets. The prediction ability of the model was evaluated using validation data set and "one-leave-out" cross validation method. The root mean square errors (RMSE) and mean absolute errors for the validation data set were 0.042 and 0.024, respectively. The prediction ability of ANN model was also statistically compared with results of linear free energy related model. The obtained results show the validity of proposed model in the prediction of cytotoxicity data of corresponding anti-HIV drugs.

QSARs and activity predicting models for competitive inhibitors of adenosine deaminase

Combinations of multiple linear regressions, genetic algorithms and artificial neural networks were utilized to develop models for seeking quantitative structure-activity relationships that correlate structural descriptors and inhibition activity of adenosine deaminase competitive inhibitors. Many quantitative descriptors were generated to express the physicochemical properties of 70 compounds with optimized structures in aqueous solution. Multiple linear regressions were used to linearly select different subsets of descriptors and develop linear models for prediction of log(k(i)). The best subset then fed artificial neural networks to develop nonlinear predictors. A committee of six hybrid models - that included genetic algorithm routines together with neural networks - was also utilized to nonlinearly select most efficient subsets of descriptors in a cross-validation procedure for nonlinear log(k(i)) prediction. The best prediction model was found to be an 8-3-1 artificial neural network which was fed by the most frequently selected descriptors among these subsets. This prediction model resulted in train set root mean sum square error (RMSE) of 0.84 log(k(i)) and prediction set RMSE of 0.85 log(k(i)) (both equivalent of 0.10 in normal range of log(k(i))) and correlation coefficient (r(2)) of 0.91.

Predicting activities without computing descriptors: graph machines for QSAR

We describe graph machines, an alternative approach to traditional machine-learning-based QSAR, which circumvents the problem of designing, computing and selecting molecular descriptors. In that approach, which is similar in spirit to recursive networks, molecules are considered as structured data, represented as graphs. For each example of the data set, a mathematical function (graph machine) is built, whose structure reflects the structure of the molecule under consideration; it is the combination of identical parameterised functions, called "node functions" (e.g. a feedforward neural network). The parameters of the node functions, shared both within and across the graph machines, are adjusted during training with the "shared weights" technique. Model selection is then performed by traditional cross-validation. Therefore, the designer's main task consists in finding the optimal complexity for the node function. The efficiency of this new approach has been demonstrated in many QSAR or QSPR tasks, as well as in modelling the activities of complex chemicals (e.g. the toxicity of a family of phenols or the anti-HIV activities of HEPT derivatives). It generally outperforms traditional techniques without requiring the selection and computation of descriptors.

A novel simple QSAR model for the prediction of anti-HIV activity using multiple linear regression analysis

A quantitative-structure activity relationship was obtained by applying Multiple Linear Regression Analysis to a series of 80 1-[2-hydroxyethoxy-methyl]-6-(phenylthio) thymine (HEPT) derivatives with significant anti-HIV activity. For the selection of the best among 37 different descriptors, the Elimination Selection Stepwise Regression Method (ES-SWR) was utilized. The resulting QSAR model (R (2) (CV) = 0.8160; S (PRESS) = 0.5680) proved to be very accurate both in training and predictive stages.

Quantitative structure–activity relationship studies on HEPTs by supervised stochastic resonance

Quantitative structure-activity relationship studies (QSAR) on HEPTs were performed by using a new approach--supervised stochastic resonance (SSR) in this paper. Errors in physicochemical properties have great effects on variable selection and the predictive capability of QSAR models but errors-in-variables were seldom discussed in QSAR. In this paper, based on the theory of stochastic resonance (SR), SSR was proposed and employed to the problem. In SSR, errors and abundant variables were regarded as noise and the relevant descriptors as signals. In the nonlinear systems involved in the SR, the signal and the noise interact harmonically and the signal was consequently enhanced. Therefore, the correlation between the relevant variables and a specified activity of a series molecule was improved by SSR. It is demonstrated that the obtained QSAR models for HEPT analogues by SSR were comparable to those by published methods in their stability and predictivity. SSR is an efficient and promising approach to QSAR studies.

Modeling of activity of cyclic urea HIV-1 protease inhibitors using regularized-artificial neural networks

Artificial neural networks (ANNs) were used to model both inhibition of HIV-1 protease (K(i)) and inhibition of HIV replication (IC90) for 55 cyclic urea derivatives using constitutional and 2D descriptors. As a preliminary step, linear dependences were established by multiple linear regression (MLR) approaches, selecting the relevant descriptors by genetic algorithm (GA) feature selection. For ANN models non-linear GA feature selection was also applied. Non-linear modeling of K(i) overcame the results of the linear one using four properties, keeping in mind standard Pearson R correlation coefficients (0.931 vs. 0.862) and leave one out (LOO) cross-validation analysis (Q(LOO)2 = 0.703 vs. 0.510). On the other hand, IC90 modeling was insoluble by a linear approach: no predictive model was achieved; however, a non-linear relation was encountered according to statistic results (R = 0.891; Q(LOO)2 = 0.568). The best non-linear models suggested the influence of the presence of nitrogen atoms and the molecular volume distribution in the inhibitor structures on the HIV-1 protease inhibition as well as that the inhibition of HIV replication was dependent on the occurrence of five-member rings. Finally, inhibitors were well distributed regarding its activity levels in a Kohonen self-organizing map built using the input variables of the best non-linear models.

GRID formalism for the comparative molecular surface analysis: application to the CoMFA benchmark steroids, azo dyes, and HEPT derivatives

Shape analysis is a powerful tool in chemistry and drug design, and molecular surface defines shape in the molecular scale. In the current publication we presented a novel formalism for the comparative molecular surface analysis (s-CoMSA). The method enables both quantitative modeling of 3D-QSAR and finding possible pharmacophoric sites. The method provides very predictive models for the CBG activity of the benchmark steroid series, tinctorial properties of the heterocyclic azo dyes and anti-HIV activity of the HEPT series.

Quantitative study of the structure-retention index relationship in the imine family

The Kováts retention index is one of the most popular descriptors of the performance of organic compounds in gas chromatography (GC). The mathematical modeling of this index is an interesting and open problem in analytical chemistry. In this paper, two models for the prediction of the Kováts retention index are presented. Topologic, topographic and quantum-chemical descriptors were used as structural descriptors. Multiple linear regression (MLR) analysis provides the first model using the forward stepwise procedure for the variable selection. For the second one, an ensemble of artificial neural network (ANN) was constructed using the pruning algorithm. Both methods were validated by an external set of compounds, by the Golbraikh and Tropsha method and by the leave-one-out (LOO) and the leave many out (LMO) procedures. The R2, RMScv and Q2, values for the training sets were 0.884, 0.589 and 0.830 for NN and 0.974, 0.417 and 0.970 for MLR models, respectively. The robustness of both models was demonstrated. Both portrait the chromatographic performance of the sample but in this case, the results of MLR equation are better than the NN ones. The MLR model is recommended because of its simplicity.

Predicting permeability coefficient in ADMET evaluation by using different membranes-interaction QSAR

Membrane-interaction quantitative structure activity relationship (MI-QSAR) analysis was applied to a data set with 18 compounds in 18 different membranes. MI-QSAR was used to estimate the ADMET properties including the transport of organic solutes through biological membranes. The most important descriptors are the aqueous solvation free energy, FH2O, and diffusion coefficient for all membranes. The correlation coefficient, r2, and cross-validation correlation coefficient, q2, for DMPG membrane is 0.850 and 0.770, respectively. The relationship between FH2O and permeability is nonlinear. But the detail effect of aqueous solvation free energy and diffusion coefficient to the permeability depends on the type of membrane. The final models also support the solution-diffusion mechanism of transport is important in membrane.

A 4D-QSAR study on anti-HIV HEPT analogues

We used the 4D-QSAR method coupled with the PLS analysis and uninformative variable elimination or its variants for the investigations of the antiviral activity of HEPT, a series of conformationally flexible molecules that bind HIV-1 reverse transcriptase. An analysis of several Hopfinger's and SOM-4D-QSAR models indicated that both methods yield comparable results. Generally, charge descriptors provide better modeling efficiency. We have shown that the method properly indicates the mode of interaction revealed by X-ray studies. It also allows us to calculate highly predictive QSAR models.

Computational Studies and Drug Design for HIV-1 Reverse Transcriptase Inhibitors of 3′,4′-di-O-(S)-camphanoyl-(+)-cis-Khellactone (DCK) Analogs

Molecular docking and molecular dynamics simulation were applied to study the binding mode of 3',4'-di-O-(S)-camphanoyl-(+)-cis-khellactone (DCK) analogs anti-HIV inhibitors with HIV-1 RT. The results suggest that there is a strong hydrogen bond between DCK O16 and NH of Lys101, and that DCK analogues might act similarly as other types of HIV-1 RT inhibitors. The investigation about drug resistance for DCK shows no remarkable influence on the most frequently observed mutation K103N of HIV-1 RT. Based on the proposed mechanism, some new structures were designed and predicted by a SVM model. All compounds exhibited potent inhibitory activities against HIV replication in H9 lymphocytes with EC50 values lower than 1.95 microM. The rationality of the method was validated by experimental results.