Counterpropagation Neural Network Research Articles

Development of a Structural Model for NF-κB Inhibition of Sesquiterpene Lactones Using Self-Organizing Neural Networks

A variety of sesquiterpene lactones (SLs) possess considerable anti-inflammatory activity. Several studies have shown that they exert this effect in part by inhibiting the activation of the transcription factor NF-kappaB. In the present study we elaborated on the investigation of a data set of 103 structurally diverse SLs for which we had previously developed several different QSAR equations dependent on the skeletal type. Use of 3D structure descriptors resulted in a single model for the entire data set. In particular, local radial distribution functions (L-RDF) were used that centered on the methylene-carbonyl substructure believed to be the site of attack of cysteine-38 of the p65/NF-kappaB subunit. The model was developed by using a counterpropagation neural network (CPGNN), attesting to the power of this method for establishing structure-activity-relationships. The investigations shed more light onto the influence of the chemical structure on NF-kappaB inhibitory activity.

Variable Selection and Interpretation in Structure−Affinity Correlation Modeling of Estrogen Receptor Binders

A computational approach for the identification and investigation of correlations between a chemical structure and a selected biological property is described. It is based on a set of 132 compounds of known chemical structures, which were tested for their binding affinities to the estrogen receptor. Different multivariate modeling methods, i.e., partial least-squares regression, counterpropagation neural network, and error-back-propagation neural network, were applied, and the prediction ability of each model was tested in order to compare the results of the obtained models. To reduce the extensive set of calculated structural descriptors, two types of variable selection methods were applied, depending on the modeling approach used. In particular, the final partial least-squares regression model was built using the "variable importance in projection" variable selection method, while genetic algorithms were applied in neural network modeling to select the optimal set of descriptors. A thorough statistical study of the variables selected by genetic algorithms is shown. The results were assessed with the aim to get insight to the mechanisms involved in the binding of estrogenic compounds to the receptor. The variable selection on the basis of genetic algorithm was controlled with the test set of compounds, extracted from the data set available. To compare the predictive ability of all the optimized models, a leave-one-out cross-validation procedure was applied, the best model being the nonlinear neural network model based on error back-propagation algorithm, which resulted in R2= 92.2% and Q2= 70.8%.

QSAR analysis of phenolic antioxidants using MOLMAP descriptors of local properties

Molecular maps of atom-level properties (MOLMAPs) were developed to represent the diversity of chemical bonds existing in a molecule. Chemical reactivity, being related to the ability for bond breaking and bond making, is primarily determined by the properties of bonds available in a molecule. In order to use physicochemical properties of individual bonds for an entire molecule, and at the same time having a fixed-length molecular representation, all the bonds of a molecule are mapped into a fixed-size 2D self-organizing map (MOLMAP). This article illustrates the application of MOLMAP descriptors to QSAR, with a study of the radical scavenging activity of 47 naturally occurring phenolic antioxidants. Counterpropagation neural networks (CPG NNs) were trained with MOLMAP descriptors selected using genetic algorithms to predict antioxidant activity. The model was subsequently validated by the leave-one-out (LOO) procedure obtaining a q 2 of 0.71. Random Forests were grown with the entire set of MOLMAP descriptors giving 70% of correct classifications as potent, active or inactive in a LOO experiment. Interpretations of both models in terms of discriminant variables were concordant and allowed identifying bonds and substructures that are mostly responsible for antioxidant activity. This work shows how MOLMAPs can be used for data mining of structural and biological activity data, leading to the extraction of relationships between local properties and activity.

QSAR Model for Predicting Pesticide Aquatic Toxicity

A hierarchical QSAR approach was applied for the prediction of acute aquatic toxicity. Chemical structures were encoded into molecular descriptors by an automated, seamless procedure available within the OpenMolGRID system. Finally, various linear and nonlinear regression techniques were used to obtain stable and thoroughly validated QSARs. The final model was developed by a counterpropagation neural network coupled with genetic algorithms for variable selection. The proposed QSAR is consistent with McFarland's principle for biological activity and makes use of seven molecular descriptors, namely HACA-2, HOMO-LUMO energy gap, Kier and Hall index, HA dependent HDSA-1, BETA polarizability, FHBCA fractional HBSA, and LogP. The model was extensively tested by the test set (R2= 0.79), the y-scrambling test, and sensitivity/stability tests.

Monitoring and diagnosis of roller bearing conditions using neural networks and soft computing

Roller Bearings have extended use throughout the industry and their proper operation is paramount in insuring quality products. Therefore, an on-line monitoring and diagnostic system is needed to detect faulty bearings. In this work, by applying the feature selection technique to the data obtained from vibration signals, six indices were selected. Artificial neural networks and soft computing were used for nonlinear pattern recognition. An attempt was made to distinguish between normal and defective bearings. Furthermore, classification of roller bearing conditions into six different categories was conducted for the diagnostic purpose. Back propagation neural networks (BPN's), counterpropagation neural networks (CPN's), and adaptive neuro-fuzzy inference systems (ANFIS) were used for on-line monitoring and diagnosis of roller bearing conditions. All of them were able to recognize normal bearings from defective bearings with 100% success rate. In classifying the defect types, BPN obtained a success rate of 20% to 100%; CPN obtained a success rate of 31.7% to 100% while ANFIS achieved a success rate of 5% to 48%. CPN have the best performance among the three intelligent techniques. In order to monitor roller bearing conditions, a 1 × 20 × 1 CPN should be used to distinguish normal bearings from defective bearings. Furthermore, a 6 × 24 × 1 CPN can be used to diagnose the roller bearing conditions into six categories. In this manner, monitoring and diagnosis of roller bearings can be performed successfully.

Use of Structure Descriptors to Discriminate Between Modes of Toxic Action of Phenols.

Two classification models were developed based on a data set of 220 phenols with four associated Modes of Toxic Action (MOA). Counter-propagation neural networks (CPG NN) and multinomial logistic regression (multinom) were used as classification methods. The combination of topological autocorrelation of empirical π-charge and σ-electronegativity and of surface autocorrelation of hydrogen-bonding potential resulted in a 21-dimensional model that successfully discriminated between the four MOAs. Its overall predictive power was estimated to 92% using 5-fold cross-validation. Subsequently, a simple score for the distance to the training data was used to determine the prediction space of the model and used in an exploratory study on the phenols contained in the open NCI database. The use of a prediction space metric proved indispensable for the screening of such a diverse database. The prediction space covered by the proposed model is still of rather local nature which is either caused by the limited diversit...

Automatic Assignment of Absolute Configuration from 1D NMR Data

[reaction: see text] Opposite enantiomers exhibit different NMR properties in the presence of an external common chiral element, and a chiral molecule exhibits different NMR properties in the presence of external enantiomeric chiral elements. Automatic prediction of such differences, and comparison with experimental values, leads to the assignment of the absolute configuration. Here two cases are reported, one using a dataset of 80 chiral secondary alcohols esterified with (R)-MTPA and the corresponding (1)H NMR chemical shifts and the other with 94 (13)C NMR chemical shifts of chiral secondary alcohols in two enantiomeric chiral solvents. For the first application, counterpropagation neural networks were trained to predict the sign of the difference between chemical shifts of opposite stereoisomers. The neural networks were trained to process the chirality code of the alcohol as the input, and to give the NMR property as the output. In the second application, similar neural networks were employed, but the property to predict was the difference of chemical shifts in the two enantiomeric solvents. For independent test sets of 20 objects, 100% correct predictions were obtained in both applications concerning the sign of the chemical shifts differences. Additionally, with the second dataset, the difference of chemical shifts in the two enantiomeric solvents was quantitatively predicted, yielding r(2) 0.936 for the test set between the predicted and experimental values.

Vibration control of building structures using self-organizing and self-learning neural networks

Past research in artificial intelligence establishes that artificial neural networks (ANN) are effective and efficient computational processors for performing a variety of tasks including pattern recognition, classification, associative recall, combinatorial problem solving, adaptive control, multi-sensor data fusion, noise filtering and data compression, modelling and forecasting. The paper presents a potentially feasible approach for training ANN in active control of earthquake-induced vibrations in building structures without the aid of teacher signals (i.e. target control forces). A counter-propagation neural network is trained to output the control forces that are required to reduce the structural vibrations in the absence of any feedback on the correctness of the output control forces (i.e. without any information on the errors in output activations of the network). The present study shows that, in principle, the counter-propagation network (CPN) can learn from the control environment to compute the required control forces without the supervision of a teacher (unsupervised learning). Simulated case studies are presented to demonstrate the feasibility of implementing the unsupervised learning approach in ANN for effective vibration control of structures under the influence of earthquake ground motions. The proposed learning methodology obviates the need for developing a mathematical model of structural dynamics or training a separate neural network to emulate the structural response for implementation in practice.

Use of Structure Descriptors To Discriminate between Modes of Toxic Action of Phenols

Two classification models were developed based on a data set of 220 phenols with four associated Modes of Toxic Action (MOA). Counter-propagation neural networks (CPG NN) and multinomial logistic regression (multinom) were used as classification methods. The combination of topological autocorrelation of empirical pi-charge and sigma-electronegativity and of surface autocorrelation of hydrogen-bonding potential resulted in a 21-dimensional model that successfully discriminated between the four MOAs. Its overall predictive power was estimated to 92% using 5-fold cross-validation. Subsequently, a simple score for the distance to the training data was used to determine the prediction space of the model and used in an exploratory study on the phenols contained in the open NCI database. The use of a prediction space metric proved indispensable for the screening of such a diverse database. The prediction space covered by the proposed model is still of rather local nature which is either caused by the limited diversity and size of the training set or by the high dimensionality of the descriptors.

Estimation of melting points of pyridinium bromide ionic liquids with decision trees and neural networks

Regression trees were built with an initial pool of 1085 molecular descriptors calculated by DRAGON software for 126 pyridinium bromides, to predict the melting point. A single tree was derived with 9 nodes distributed over 5 levels in less than 2 min showing very good correlation between the estimated and experimental values (R2 = 0.933, RMS = 12.61 °C). A number n of new trees were grown sequentially, without the descriptors selected by previous trees, and combination of predictions from the n trees (ensemble of trees) resulted in higher accuracy. A 3-fold cross-validation with the optimum number of trees (n = 4) yielded an R2 value of 0.822. A counterpropagation neural network was trained with the variables selected by the first tree, and reasonable results were achieved (R2 = 0.748). In a test set of 9 new pyridinium bromides, all the low melting point cases were successfully identified.

Comparison of Different Classification Methods Applied to a Mode of Toxic Action Data Set

AbstractSuccessful discrimination of compounds by mode of toxic action (MOA) is a prerequisite for process‐based quantitative structure‐activity relationship (QSAR) approaches. A data set of 115 compounds comprising nine MOA classes and 24 descriptors has been studied with several classification methods: multinomial logistic regression (multinom), linear discriminant analysis (LDA), partial least squares (PLS), and counter‐propagation neural networks (CPG NN). Variables were selected with stepwise methods and with a genetic algorithm (GA) for the CPG NN. Five‐fold cross‐validation was used for validating the models and the advantages and disadvantages of this validation method are critically discussed. Without variable selection the predictive power of the models ranges between 51% and 53% cross‐validated overall correct classification. With appropriate parameter selection the predictive power slightly increased to 52–59%.

A NEURAL NETWORK-WAVELET MODEL FOR GENERATING ARTIFICIAL ACCELEROGRAMS

In earthquake-resistant design of structures, for certain structural configurations and conditions, it is necessary to use accelerograms for dynamic analysis. Accelerograms are also needed to simulate the effects of earthquakes on a building structure in the laboratory. A new method of generating artificial earthquake accelerograms is presented through adroit integration of neural networks and wavelets. A counterpropagation (CPN) neural network model is developed for generating artificial accelerograms from any given design spectrum such as the International Building Code (IBC) design spectrum. Using the IBC design spectrum as network input means an accelerogram may be generated for any geographic location regardless of whether earthquake records exist for that particular location or not. In order to improve the efficiency of the model, the CPN network is modified with the addition of the wavelet transform as a data compression tool to create a new CPN-wavelet network. The proposed CPN-wavelet model is trained using 20 sets of accelerograms and tested with additional five sets of accelerograms available from the U.S. Geological Survey. Given the limited set of training data, the result is quite remarkable.

Multivariate data analysis in classification of vegetable oils characterized by the content of fatty acids

A novel method for fast and efficient determination of classes of oil samples in routine analyses performed in food control laboratories is proposed. It is based on counterpropagation neural network, which offers a possibility for automatic classification. The fatty acid composition of 132 samples of different edible vegetable oils from the market, including pumpkin, sunflower, peanut, olive, soybean, rapeseed, corn and some mixed oils has been determined. Gas chromatography (GC) was used for qualitative and quantitative determination of methyl esters of palmitic, stearic, oleic, linoleic, linolenic, eicosanoic and eicosenoic acids. The resulting fatty acid composition was evaluated by different chemometrics methods: (i) principal component analysis (PCA), (ii) the clustering method based on the Kohonen neural network, and (iii) the counterpropagation neural network method applied for the classification of oil samples. The first two methods were both successful to distinguish clusters of different oil samples. However, they are not suitable for an automatic prediction of oil classes, because they require a visual inspection of resulting classes and consequently the final decision by an expert. The counterpropagation neural network method was found to be a good model for the classification of different edible vegetable oils on the basis of their composition regarding seven fatty acids. After training with 95 oil samples, the network was able to ascribe correct classes to all samples. The assessment of the model by the leave-one-out cross validation procedure demonstrated 94 correct predictions (i.e. 98.95%) for 95 samples of known type of oil. For a comparison, the linear discriminant analysis (LDA) was performed, which resulted in 92 (i.e. 96.84%) correct predictions. The remaining 37 samples of mixed oils and oils of unknown origin were used to test the model for its ability to predict the composition of mixtures and to show the implementation of the model for predicting the most probable class of an unknown sample.

Structure‐Mutagenicity Study of 12 Trimethylimidazopyridine Isomers Using Orbital Energies and “Spectrum‐Like Representation” as Descriptors.

The set of 12 trimethylimidazopyridine isomers with mutagenic potency toward two strains of Salmonella was treated in this study. Ten isomers with known mutagenic properties were taken to build the models. Fifteen molecular orbital energies, or a “spectrum-like” representation of 3D structures, were taken as descriptors. As modeling techniques the multiple linear regression and the counter propagation neural network were applied. Models were tested with the recall ability test and the leave-one-out cross-validation tests. For two isomers, which have not been synthesized yet, we report predicted values for both mutagenic potencies obtained with different models. The best models were found when unoccupied molecular orbital energies are among the descriptors.

Classification of Potential Endocrine Disrupters on the Basis of Molecular Structure Using a Nonlinear Modeling Method.

A methodology for the classification of endocrine disruption chemicals is proposed. It is based on a data set of 106 substances extracted from the list of 553 chemicals that were inspected by the European Union Commission for the scientific evidence of their endocrine disruption activity. The substances belong to different categories defined in the EU Commission report: (i) literature evidence for certainly active as endocrine disrupters, (ii) for potentially active, (iii) for less probable active lacking evidence, and (iv) for certainty nonactive. 3D molecular coordinates were calculated using the AM1or the PM3 optimization method. From 3D coordinates an extensive set of molecular descriptors was calculated. The classification model based on the counterpropagation neural network was constructed and evaluated. This is the first time that the counterpropagation neural network is applied for the classification of compounds regarding their literature evidence for the endocrine disruption activity. The devel...

Prediction of selectivity of α 1-adrenergic antagonists by counterpropagation neural network (CP-ANN)

A quantitative structure-selectivity relationships of series of structurally diverse α 1-adrenergic antagonists was performed by using counter-propagation neural network (CP-ANN). The theoretical molecular descriptors have been calculated and selected using CODESSA program. The results obtained for a highly non-congeneric set of molecules have confirmed the potential of use of CP-ANN approach in prediction of relative activity (selectivity) of α 1-adrenergic antagonists.

Modeling of structure–mutagenicity relationships: counter propagation neural network approach using calculated structural descriptors

Counter propagation artificial neural network was applied for modeling the mutagenicity of 95 aromatic and heteroaromatic amines collected from the literature. Molecules were represented by topostructural, topochemical, geometrical and quantum chemical descriptors. A sphere exclusion algorithm was used for rational division of the dataset into training and test sets. The initial training set was improved by a step-wise inclusion of two outliers from the test set. Recall ability of the final model is good ( R 2=0.986) as well as prediction ability in respect to the test set ( R 2=0.816). Validity of the best model obtained in the study was confirmed by randomization test and test with exchanged training and test sets. Study demonstrated the capabilities of CP ANN in analyzing the similarities between compounds and identifying of outliers. It was shown that CP ANN is a powerful tool for modeling the structure–mutagenicity relationships of the compounds considered.

Classification of potential endocrine disrupters on the basis of molecular structure using a nonlinear modeling method.

A methodology for the classification of endocrine disruption chemicals is proposed. It is based on a data set of 106 substances extracted from the list of 553 chemicals that were inspected by the European Union Commission for the scientific evidence of their endocrine disruption activity. The substances belong to different categories defined in the EU Commission report: (i) literature evidence for certainly active as endocrine disrupters, (ii) for potentially active, (iii) for less probable active--lacking evidence, and (iv) for certainty nonactive. 3D molecular coordinates were calculated using the AM1or the PM3 optimization method. From 3D coordinates an extensive set of molecular descriptors was calculated. The classification model based on the counterpropagation neural network was constructed and evaluated. This is the first time that the counterpropagation neural network is applied for the classification of compounds regarding their literature evidence for the endocrine disruption activity. The developed classification model is proposed as a tool for a preliminary assessment of potential endocrine disrupters, which would help the assessors to make the priority list for a large amount of chemicals that have to be tested with more expensive in vitro and in vivo methods.

Structure-based predictions of 1H NMR chemical shifts using feed-forward neural networks.

Feed-forward neural networks were trained for the general prediction of 1H NMR chemical shifts of CH(n) protons in organic compounds in CDCl3. The training set consisted of 744 1H NMR chemical shifts from 120 molecular structures. The method was optimized in terms of selected proton descriptors (selection of variables), the number of hidden neurons, and integration of different networks in ensembles. Predictions were obtained for an independent test set of 952 cases with a mean average error of 0.29 ppm (0.20 ppm for 90% of the cases). The results were significantly better than those obtained with counterpropagation neural networks.

Structure-mutagenicity study of 12 trimethylimidazopyridine isomers using orbital energies and "spectrum-like representation" as descriptors.

The set of 12 trimethylimidazopyridine isomers with mutagenic potency toward two strains of Salmonella was treated in this study. Ten isomers with known mutagenic properties were taken to build the models. Fifteen molecular orbital energies, or a "spectrum-like" representation of 3D structures, were taken as descriptors. As modeling techniques the multiple linear regression and the counter propagation neural network were applied. Models were tested with the recall ability test and the leave-one-out cross-validation tests. For two isomers, which have not been synthesized yet, we report predicted values for both mutagenic potencies obtained with different models. The best models were found when unoccupied molecular orbital energies are among the descriptors.