M.C. Liu

Application of artificial neural networks for prediction of the retention indices of alkylbenzenes

Abstract Artificial neural networks (ANN) with extended delta–bar–delta (EDBD) learning algorithms were used to predict the retention indices of alkylbenzenes. The data used in this paper include 96 retention indices of 32 alkylbenzenes on three different stationary phases. Four parameters: temperature, boiling point, molar volume and the kind of stationary phase, were used as input parameters. These three stationary phases are: PEG, SE-30, SQ. The 96 group data were randomly divided into two sets: a training set (including 64 group data) and a testing set (including 32 group data). The structures of networks and the learning times were optimized. The best network structure is 4–7–1. The optimum number of learning time is about 20 000. It is shown that the maximum relative error is no more than 3%. The result illustrated that the prediction performance of ANN in the field of investigating the retention behaviors of alkylbenzenes is very satisfactory.

QSAR study of neuraminidase inhibitors based on heuristic method and radial basis function network

Neuraminidase (NA) is a critical enzyme of the influenza virus and many inhibitors targeting this enzyme are quite efficient anti-influenza agents. In this paper, quantitative structure-activity relationship (QSAR) method was used to predict the activity of different kinds of 46 NA inhibitors. Heuristic method (HM) and radial basis function network (RBFNN) were first used to build linear and nonlinear models, respectively. The prediction results were in agreement with the experimental value. The proposed model is simple and can be extended to other QSAR investigations.

QSAR study of natural, synthetic and environmental endocrine disrupting compounds for binding to the androgen receptor

A large data set of 146 natural, synthetic and environmental chemicals belonging to a broad range of structural classes have been tested for their relative binding affinity (expressed as log (RBA)) to the androgen receptor (AR). These chemicals commonly termed endocrine disrupting compounds (EDCs) present a variety of adverse effects in humans and animals. As assays for binding affinity remains a time-consuming task, it is important to develop predictive methods. In this work, quantitative structure–activity relationships (QSARs) were determined using three methods, multiple linear regression (MLR), radical basis function neural network (RBFNN) and support vector machine (SVM). Five descriptors, accounting for hydrogen-bonding interaction, distribution of atomic charges and molecular branching degree, were selected from a heuristic method to build predictive QSAR models. Comparison of the results obtained from three models showed that the SVM method exhibited the best overall performances, with a RMS error of 0.54 log (RBA) units for the training set, 0.59 for the test set, and 0.55 for the whole set. Moreover, six linear QSAR models were constructed for some specific families based on their chemical structures. These predictive toxicology models, should be useful to rapidly identify potential androgenic endocrine disrupting compounds.

LARGE ARTIFICIAL NEURAL NETWORKS APPLIED TO THE PREDICTION OF RETENTION INDICES OF ACYCLIC AND CYCLIC ALKANES, ALKENES, ALCOHOLS, ESTERS, KETONES AND ETHERS

Abstract Artificial neural networks (ANN) with extended delta-bar-delta (EDBD) back propogation learning algorithms were used to predict the retention indices of 184 organic compounds. These compounds include acyclic and cyclic alkanes, alkenes, alcohols, esters, ketones and ethers. The networks architecture and parameters were optimized to give maximum performance. The best network is 2–6–1, the optimum learning epoch is 2000. In the process of the study, cross-validation and leave-20%-out were used. The results show that the prediction performance of ANN operating with such non-linear systems is remarkably good.

Quantitative structure-toxicity relationships (QSTRs): A comparative study of various non linear methods. General regression neural network, radial basis function neural network and support vector machine in predicting toxicity of nitro- and cyano- aromatics to Tetrahymena pyriformis §

Prediction of toxicity of 203 nitro- and cyano-aromatic chemicals to Tetrahymena pyriformis was carried out by radial basis function neural network, general regression neural network and support vector machine, in non-linear response surface methodology. Toxicity was predicted from hydrophobicity parameter (log Kow) and maximum superdelocalizability (Amax). Special attention was drawn to prediction ability and robustness of the models, investigated both in a leave-one-out and 10-fold cross validation (CV) processes. The influence that the corresponding changes in the learning sets during these CV processes could have on a common external test set including 41 compounds was also examined. This allowed us to establish the stability of the models. The non linear results slightly outperform (as expected) multilinear relationships (MLR) and also favourably compete with various other non linear approaches recently proposed by Ren (J. Chem. Inf. Comput. Sci., 43 1679 (2003)). § Presented at CMTPI 2005: Computational Methods in Toxicology and Pharmacology Integrating Internet Resources (Shanghai, China, October 29–November 1, 2005).

Neural network-molecular descriptors approach to the prediction of properties of alkenes

Abstract A molecular descriptors set of five parameters (W, P, w,p, s) including three grades of structural information was set up as an alkene molecule descriptor to predict its normal boiling point, density and refractive index by a neural network. The five parameters are W based on the distance matrix of a molecule, P the polarity number, w representing the absolute contribution of a double bond to the whole size of a molecule, p indicating the absolute contribution of a double bond to the shape of a molecule, and s representing enantiomer of alkenes respectively. The resulting estimates show average accuracies of 0.95% with a maximum deviation of 10%.

Predicting the Standard Enthalpy (ΔHo f) and Entropy (So) of Alkanes by Artificial Neural Networks

Abstract Artificial Neural Networks (ANNs) with Extended Delta-Bar-Delta (EDBD) back propagation learning algorithm have been developed to predict the standard enthalpy and entropy of 87 acyclic alkanes. Molecular weight, boiling point and density of the compounds were used as input parameters. The networks architecture and parameters were optimized to give maximum performances. The best network was a 3-6-2 ANN, and the optimum learning epoch was about 1320. The results show that the maximum relative errors of enthalpy and entropy are less than 3%. They reveal that the performances of ANNs for predicting the enthalpy and entropy of alkanes are satisfying.

Prediction of Programmed-temperature Retention Values of Naphthas by Artificial Neural Networks

Abstract It is proposed for the first time a method of prediction of the programmed-temperature retention times of components of naphthas in capillary gas chromatography using artificial neural networks. People are used to predict the programmed-temperature retention time using many formulas such as the integral formula, which requires that four parameters must be determined by calculation or experiments. However the results obtained by the formula are not so good to meet the demand of industry. In order to predict retention time accurately and conveniently, artificial neural networks using five-fold cross-validation and leave-20%-out methods have been applied. Only two parameters: density and isothermal retention index were used as input vectors. The average RMS error for predicted values of five different networks was 0.18, whereas the RMS error of predictions by the integral formula was 0.69. Obviously, the predictions by neural networks were much better than predictions by the formula, and neural networks need fewer parameters than the formula. So neural networks can successfully and conveniently solve the problem of predictions of programmed-temperature retention times, and provide useful data for analysis of naphthas in petrochemical industry.