Classification of HIV protease inhibitors on the basis of their antiviral potency using radial basis function neural networks.

HIV protease inhibitors are being used as frontline therapy in the treatment of HIV patients. Multi-drug-resistant HIV mutant strains are emerging with the initial aggressive multi-drug treatment of HIV patients. This necessitates continued search for novel inhibitors of viral replication. These protease inhibitors may further be useful as pharmacological agents for inhibition of other viral replication. Classification models of HIV Protease inhibitors are developed using a data set of 123 compounds containing several heterocycles. Their inhibitory concentrations expressed as log (IC50) ranged from -1.52 to 2.12 log units. The dataset was divided into active and inactive classes on the basis of their antiviral potency. Initially a two-class problem (active, inactive) is explored using k-nearest neighbor approach. In order to introduce non-linearity in the classifier different approaches were investigated. This led to the goal of a fast, simple, minimum user input, radial basis function neural network (RBFNN) classifier development. Then the same two-class problem was resolved using the (RBFNN) classifier. A genetic algorithm with RBFNN fitness evaluator was used to search for the optimum descriptor subsets. The application of majority rules was also tested for the RBFNN classification. The best six descriptor model found by the new cost function showed predictive ability in the high 80% range for an external prediction set.

Classification of inhibitors of protein tyrosine phosphatase 1B using molecular structure based descriptors.

Loss of Protein Tyrosine Phosphatase 1B (PTP 1B) activity is known to enhance insulin sensitivity and resistance to weight gain. So potent and orally active PTP1B inhibitors could be potential pharmacological agents for the treatment of Type 2 diabetes and obesity. Classification models of PTP1B inhibitors are developed using a data set containing 128 compounds. Their inhibitory concentrations ranged from -1.59 to 1.68 log units. Initially a two-class (active, inactive) problem is tackled using a number of different methods. The data set was divided into active and inactive classes on the basis of inhibitory activity of the compounds. Molecular structure-based descriptors were calculated and used in the model development. Descriptors encoding the flexibility of the molecules were investigated. Classification models were generated using k-nearest neighbors (k-NN), linear discriminant analysis (LDA), and radial basis function neural network (RBFNN). All models are tested using an external prediction set, compounds not used anywhere during the model development procedure. A five-descriptor model is developed that produces a classification rate of 85.7% for an external prediction set. Then a three-class (active, moderately active, inactive) problem was explored. This time the data set was divided into highly active, moderate, and inactive classes on the basis of inhibitory activity of the compounds. The best classification rate achieved for an external prediction set was 85%. The classification rates achieved indicate that these models could serve as a screening mechanism, to identify potentially useful PTP 1B inhibitors. In addition multiple linear regression and computational neural network models are also developed for prediction of log IC(50) values. All QSAR models are tested using the same external prediction set.

Development of binary classification of structural chromosome aberrations for a diverse set of organic compounds from molecular structure.

Classification models are generated to predict in vitro cytogenetic results for a diverse set of 383 organic compounds. Both k-nearest neighbor and support vector machine models are developed. They are based on calculated molecular structure descriptors. Endpoints used are the labels clastogenic or nonclastogenic according to an in vitro chromosomal aberration assay with Chinese hamster lung cells. Compounds that were tested with both a 24 and 48 h exposure are included. Each compound is represented by calculated molecular structure descriptors encoding the topological, electronic, geometrical, or polar surface area aspects of the structure. Subsets of informative descriptors are identified with genetic algorithm feature selection coupled to the appropriate classification algorithm. The overall classification success rate for a k-nearest neighbor classifier built with just six topological descriptors is 81.2% for the training set and 86.5% for an external prediction set. The overall classification success rate for a three-descriptor support vector machine model is 99.7% for the training set, 92.1% for the cross-validation set, and 83.8% for an external prediction set.

Prediction of glycine/NMDA receptor antagonist inhibition from molecular structure.

The design and blood brain barrier crossing of glycine/NMDA receptor antagonists are of significant interest in pharmaceutical research. The use of these antagonists in stroke or seizure reduction have been considered. Measuring the inhibitory concentrations, however, can be time-consuming and costly. The use of quantitative structure-activity relationships to estimate IC(50) values for these receptor antagonists is an attractive alternative compared to experimental measurement. A data set of 109 compounds with measured log(IC(50)) values ranging from -0.57 to 4.5 is used. Structural information is encoded with numerical descriptors for topological, electronic, geometric, and polar surface properties. A genetic algorithm with a computational neural network fitness evaluator is used to select the best descriptor subsets. Multiple linear regression and computational neural network models are developed. Additionally, a quantitative radial basis function neural network (QRBFNN) was developed with the intent of introducing nonlinearity at a faster speed. A genetic algorithm using the radial basis function network as a fitness evaluator was also developed to search descriptor space for optimum subsets. All models are tested using an external prediction set. The nonlinear computational neural network model has root-mean-square errors of approximately half a log unit.

QSAR and k-nearest neighbor classification analysis of selective cyclooxygenase-2 inhibitors using topologically-based numerical descriptors.

Experimental IC(50) data for 314 selective cyclooxygenase-2 (COX-2) inhibitors are used to develop quantitation and classification models as a potential screening mechanism for larger libraries of target compounds. Experimental log(IC(50)) values ranged from 0.23 to > or = 5.00. Numerical descriptors encoding solely topological information are calculated for all structures and are used as inputs for linear regression, computational neural network, and classification analysis routines. Evolutionary optimization algorithms are then used to search the descriptor space for information-rich subsets which minimize the rms error of a diverse training set of compounds. An eight-descriptor model was identified as a robust predictor of experimental log(IC(50)) values, producing a root-mean-square error of 0.625 log units for an external prediction set of inhibitors which took no part in model development. A k-nearest neighbor classification study of the data set discriminating between active and inactive members produced a nine-descriptor model able to accurately classify 83.3% of the prediction set compounds correctly.

Linear regression and computational neural network prediction of tetrahymena acute toxicity for aromatic compounds from molecular structure.

A quantitative structure toxicity relationship (QSTR) has been derived for a diverse set of 448 industrially important aromatic solvents. Toxicity was expressed as the 50% growth impairment concentration (ICG(50)) for the ciliated protozoa Tetrahymena and spans the range -1.46 to 3.36 log units. Molecular descriptors that encode topological, geometrical, electronic, and hybrid geometrical-electronic structural features were calculated for each compound. Subsets of molecular descriptors were selected via a simulated annealing technique and a genetic algorithm. From this reduced pool of descriptors, multiple linear regression models and nonlinear models using computational neural networks (CNNs) were derived and then used to predict the ICG(50) values for an external set of representative compounds. An average of 10 nonlinear CNN models with 11-5-1 architecture was found to best describe the system with root-mean-square errors of 0.28, 0.29, and 0.34 log units for the training, cross validation, and prediction sets, respectively.

QSARs for 6-azasteroids as inhibitors of human type 1 5alpha-reductase: prediction of binding affinity and selectivity relative to 3-BHSD.

Quantitative structure-activity relationships (QSARs) are developed to describe the ability of 6-azasteroids to inhibit human type 1 5alpha-reductase. Models are generated using a set of 93 compounds with known binding affinities (K(i)) to 5alpha-reductase and 3beta-hydroxy-Delta(5)-steroid dehydrogenase/3-keto-Delta(5)-steroid isomerase (3-BHSD). QSARs are generated to predict K(i) values for inhibitors of 5alpha-reductase and to predict selectivity (S(i)) of compound binding to 3-BHSD relative to 5alpha-reductase. Log(K(i)) values range from -0.70 log units to 4.69 log units, and log(S(i)) values range from -3.00 log units to 3.84 log units. Topological, geometric, electronic, and polar surface descriptors are used to encode molecular structure. Information-rich subsets of descriptors are identified using evolutionary optimization procedures. Predictive models are generated using linear regression, computational neural networks (CNNs), principal components regression, and partial least squares. Compounds in an external prediction set are used for model validation. A 10-3-1 CNN is developed for prediction of binding affinity to 5alpha-reductase that produces root-mean-square error (RMSE) of 0.293 log units (R(2) = 0.97) for compounds in the external prediction set. Additionally, an 8-3-1 CNN is generated for prediction of inhibitor selectivity that produces RMSE = 0.513 log units (R(2) = 0.89) for the external prediction set. Models are further validated through Monte Carlo experiments in which models are generated after dependent variable values have been scrambled.

Prediction of aqueous solubility of heteroatom-containing organic compounds from molecular structure.

The use of quantitative structure-property relationships (QSPRs) to predict aqueous solubilities (log S) of heteroatom-containing organic compounds from their molecular structure is presented. Three data sets are examined. Data set 1 contains 176 compounds having one or more nitrogen atoms with some oxygen (log S[mol/L] range is -7.41 to 0.96). Data set 2 contains 223 compounds having one or more oxygen atoms, with no nitrogen (log S[mol/L] range is -8.77 to 1.57). Data set 3 contains all 399 compounds from sets 1 and 2 (log S/mol/L] range is -8.77 to 1.57). After descriptor generation and feature selection, multiple linear regression (MLR) and computational neural network (CNN) models are developed for aqueous solubility prediction. The best results were obtained with nonlinear CNN models. Root-mean-square (rms) errors for training with the three data sets ranged from 0.3 to 0.6 log units. All models were validated with external prediction sets, with the rms errors ranging from 0.6 log units to 1.5 log units.

Prediction of surface tension, viscosity, and thermal conductivity for common organic solvents using quantitative structure-property relationships.

Predictive models for the surface tension, viscosity, and thermal conductivity of 213 common organic solvents are reported. The models are derived from numerical descriptors which encode information about the topology, geometry, and electronics of each compound in the data set. Multiple linear regression and computational neural networks are used to train and evaluate models based on statistical indices and overall root-mean-square error. Eight-descriptor models were developed for both surface tension and viscosity, while a nine-descriptor model was developed for thermal conductivity. In addition, a single nine-descriptor model was developed for prediction of all three properties. The results of this study compare favorably to previously reported prediction methods for these three properties.

Prediction of C60 solubilities from solvent molecular structures.

Models predicting fullerene solubility in 96 solvents at 298 K were developed using multiple linear regression and feed-forward computational neural networks (CNN). The data set consisted of a diverse set of solvents with solubilities ranging from -3.00 to 2.12 log (solubility) where solubility = (1 x 10(4))(mole fraction of C60 in saturated solution). Each solvent was represented by calculated molecular structure descriptors. A pool of the best linear models, as determined by rms error, was developed, and a CNN model was developed for each of the linear models. The best CNN model was chosen based on the lowest value of a specified cost function and had an architecture of 9-3-1. The 76-compound training set for this model had a root-mean-square error of 0.255 log solubility units, while the 10-compound cross-validation set had an rms error of 0.253. The 10-compound external prediction set had an rms error of 0.346 log solubility units.

Classification of multidrug-resistance reversal agents using structure-based descriptors and linear discriminant analysis.

Linear discriminant analysis is used to generate models to classify multidrug-resistance reversal agents based on activity. Models are generated and evaluated using multidrug-resistance reversal activity values for 609 compounds measured using adriamycin-resistant P388 murine leukemia cells. Structure-based descriptors numerically encode molecular features which are used in model formation. Two types of models are generated: one type to classify compounds as inactive, moderately active, and active (three-class problem) and one type to classify compounds as inactive or active without considering the moderately active class (two-class problem). Two activity distributions are considered, where the separation between inactive and active compounds is different. When the separation between inactive and active classes is small, a model based on nine topological descriptors is developed that produces a classification rate of 83.1% correct for an external prediction set. Larger separation between active and inactive classes raises the prediction set classification rate to 92.0% correct using a model with six topological descriptors. Models are further validated through Monte Carlo experiments in which models are generated after class labels have been scrambled. The classification rates achieved demonstrate that the models developed could serve as a screening mechanism to identify potentially useful MDRR agents from large libraries of compounds.

Prediction of IC50 values for ACAT inhibitors from molecular structure.

A quantitative structure-activity study is performed on several series of compounds derived from N-chlorosulfonyl isocyanate to develop models that relate their structures to IC50 activity for inhibition of acyl-CoA:cholesterol O-acyltransferase (ACAT). Numerical descriptors are used to encode topological, electronic, and geometric information from the molecular structures of the inhibitors. A data set of 157 compounds showing triglyceride- and cholesterol-lowering activity is used to develop successful linear regression models and nonlinear computational neural network models. The models are validated using an external prediction set.

Prediction of inhibition of the sodium ion-proton antiporter by benzoylguanidine derivatives from molecular structure.

The use of quantitative structure-activity relationships to predict IC50 values of 113 potential Na+/H+ antiporter inhibitors is reported. Multiple linear regression and computational neural networks (CNNs) are used to develop models using a set of information-rich descriptors. The descriptors encode information about topology, geometry, electronics, and combination hybrids. A five-descriptor CNN model with root-mean-square (rms) errors of 0.278 log units for the training set and 0.377 log units for the prediction set was developed. Examination of data set subclasses showed that systematic structural variations were also well-encoded resulting in 100% accuracy of prediction trends. An experiment involving a committee of five CNNs was also performed to examine the effect of network output averaging. This showed improved results decreasing the training and cross-validation set rms error to 0.228 log units and the prediction set rms error to 0.296 log units.

Prediction of acute mammalian toxicity of organophosphorus pesticide compounds from molecular structure.

A quantitative structure-activity relationship (QSAR) investigation was done for the acute oral mammalian toxicity (LD50) of a set of 54 organophosphorus pesticide compounds. The compounds were represented with calculated molecular structure descriptors, which encoded their topological, electronic, and geometrical features. Feature selection was done with a genetic algorithm to find subsets of descriptors that would support a high quality computational neural network (CNN) model to link the structural descriptors to the -log(mmol/kg) values for the compounds. The best seven-descriptor non-linear CNN model found had an rms error of 0.22 log units for the training set compounds and 0.25 log units for the prediction set compounds.

Prediction of fathead minnow acute toxicity of organic compounds from molecular structure.

Interest in the prediction of toxicity without the use of experimental data is growing, and quantitative structure-activity relationship (QSAR) methods are valuable for such predictions. A QSAR study of acute aqueous toxicity of 375 diverse organic compounds has been developed using only calculated structural features as independent variables. Toxicity is expressed as -log(LD(50)) with the units -log(millimoles per liter) and ranges from -3 to 6. Multiple linear regression and computational neural networks (CNNs) are utilized for model building. The best model is a nonlinear CNN model based on eight calculated molecular structure descriptors. The root-mean-square log(LD(50)) errors for the training, cross-validation, and prediction sets of this CNN model are 0.71, 0.77, and 0.74 -log(mmol/L), respectively. These results are compared to a previous study with the same data set which included many more descriptors and used experimental data in the descriptor pool.

Prediction of human intestinal absorption of drug compounds from molecular structure.

The absorption of a drug compound through the human intestinal cell lining is an important property for potential drug candidates. Measuring this property, however, can be costly and time-consuming. The use of quantitative structure-property relationships (QSPRs) to estimate percent human intestinal absorption (%HIA) is an attractive alternative to experimental measurements. A data set of 86 drug and drug-like compounds with measured values of %HIA taken from the literature was used to develop and test a QSPR mode. The compounds were encoded with calculated molecular structure descriptors. A nonlinear computational neural network model was developed by using the genetic algorithm with a neural network fitness evaluator. The calculated %HIA (cHIA) model performs wells, with root-mean-square (rms) errors of 9.4%HIA units for the training set, 19.7%HIA units for the cross-validation (CV) set, and 16.0%HIA units for the external prediction set.

Neural network classification and quantification of organic vapors based on fluorescence data from a fiber-optic sensor array.

Computational neural networks have been developed to classify and quantify nine organic vapors. The neural network analyses used data that consisted of the change in fluorescence from a sensor array that consisted of 19 fiber optics with immobilized dye in polymer matrices. Plots of change in fluorescence intensity versus time were measured as pulses of analyte were presented to the sensor array. Descriptors were calculated from the intensity vs time plots, and they were used to build neural network models that accurately classified and quantified each of the nine analytes. Most of the data were used to train the neural networks (training set members), some were used to assist termination of training (cross-validation set members), and some were used to validate the models (prediction set members). Classification rates approaching 100% were achieved for the training set data, and 90% of the members in the prediction set were correctly classified. In addition, 97% of the prediction set observations were assigned a correct relative concentration.

Prediction of reduced ion mobility constants of organic compounds from molecular structure.

Quantitative structure-property relationships (QSPRs) are used to develop mathematical models that accurately predict the reduced ion mobility constants (K(0)) for a set of 168 organic compounds directly from molecular structure. The K(0) values are taken from an unpublished database collected by G. A. Eiceman, Chemistry Department, New Mexico State University. The data were collected using a Graseby Ionics environmental vapour monitor (EVM) gas chromatography/ion mobility spectrometer. Standardized conditions with controlled temperature, pressure, and humidity were used, and 2,4-lutidine was used as an internal standard. K(0) values were measured for all monomer peaks. The best model was found with a feature selection routine which couples the genetic algorithm with multiple linear regression analysis. The set of six descriptors was also analyzed with a fully connected, feed-forward neural network. The model contains six molecular structure descriptors and has a root-mean-square error of about 0.04 K(0) unit. The descriptors in the model lend insight into some of the important molecular features that influence ion mobility. The model can be utilized for prediction of K(0) values of compounds for which there are no empirical K(0) data.

Simulation of the 13C nuclear magnetic resonance spectra of trisaccharides using multiple linear regression analysis and neural networks.

Predictive models are developed for the 13C NMR chemical shifts of the carbon atoms comprising the central rings of 46 trisaccharide compounds. Thirty-nine trisaccharides are used as a training set for development of models using regression analysis and computational neural networks, and seven compounds are used as an external prediction set. The descriptors used in the models are developed directly from the molecular structures of the trisaccharides. Three different methods of descriptor selection are compared. The dependence of the models on the geometries of the trisaccharides is explored. The models developed with geometric descriptors are better than those developed without geometric descriptors, although the latter models are still of a comparable quality. Overall, the best model found is a neural network based on descriptors selected by multiple linear regression.