QSAR modelling using combined simple competitive learning networks and RBF neural networks

The aim of this study was to propose a QSAR modelling approach based on the combination of simple competitive learning (SCL) networks with radial basis function (RBF) neural networks for predicting the biological activity of chemical compounds. The proposed QSAR method consisted of two phases. In the first phase, an SCL network was applied to determine the centres of an RBF neural network. In the second phase, the RBF neural network was used to predict the biological activity of various phenols and Rho kinase (ROCK) inhibitors. The predictive ability of the proposed QSAR models was evaluated and compared with other QSAR models using external validation. The results of this study showed that the proposed QSAR modelling approach leads to better performances than other models in predicting the biological activity of chemical compounds. This indicated the efficiency of simple competitive learning networks in determining the centres of RBF neural networks.     

Unified QSAR and network-based computational chemistry approach to antimicrobials, part 1: multispecies activity models for antifungals

There are many pathogen microbial species with very different antimicrobial drugs susceptibility. In this work, we selected pairs of antifungal drugs with similar/dissimilar species predicted-activity profile and represented it as a large network, which may be used to identify drugs with similar mechanism of action. Computational chemistry prediction of the biological activity based on quantitative structure-activity relationships (QSAR) susbtantially increases the potentialities of this kind of networks, avoiding time and resource-consuming experiments. Unfortunately, most QSAR models are unspecific or predict activity against only one species. To solve this problem we developed a multispecies QSAR classification model, in which the outputs were the inputs of the aforementioned network. Overall model classification accuracy was 87.0% (161/185 compounds) in training, 83.4% (50/61) in validation, and 83.7% for 288 additional antifungal compounds used to extend model validation for network construction. The network predicted has 59 nodes (compounds), 648 edges (pairs of compounds with similar activity), low coverage density d = 37.8%, and distribution more close to normal than to exponential. These results are more characteristic of a not-overestimated random network, clustering different drug mechanisms of actions, than of a less useful power law network with few mechanisms (network hubs).     

A neural network approach to infrared spectrum interpretation

The simple linear neural network model was investigated as a method for automated interpretation of infrared spectra. The model was trained using a database of infrared spectra of organic compounds of known structure. The model was able to learn, without any prior input of spectrum-structure correlations, to recognize and identify 76 functional groupings with accuracies ranging from fair to excellent. The effect of network input parameters and of training set composition were studied, and several sources of spurious correlations were identified and corrected.Dedicated to Professor W. Simon on the occasion of his 60th birthday     

Combining classification and regression approaches for the quantification of highly overlapping capillary electrophoresis peaks by using evolutionary sigmoidal and product unit neural networks

This is a study of the potential of neural networks built by using different transfer functions (sigmoidal, product and sigmoidal–product units) designed by an evolutionary algorithm to quantify highly overlapping electrophoretic peaks. To test this approach, two aminoglycoside antibiotics, amikacin and paramomycin, were quantified from samples containing either only one component or mixtures of them though capillary zone electrophoresis (CZE) with laser‐induced fluorescence (LIF) detection. The three models assayed used as input data the four‐parameter Weibull curve associated with the profile of the electrophoretic peak and in some cases the class label for each sample estimated by cluster analysis. The combination of classification and regression approaches allowed the establishment of straightforward network topologies enabling the analytes to be quantified with great accuracy and precision. The best models for mixture samples were provided by product unit neural networks (PUNNs), 4:4:1 (14 weights) for both analytes, after discrimination by cluster analysis, allowing the analytes to be quantified with great accuracy: 8.2% for amikacin and 5.6% for paromomycin within the standard error of prediction for the generalization test, SEP_G. For comparison, partial least square regression was also used for the resolution of these mixtures; it provided a minor accuracy: SEP_G 11.8 and 15.7% for amikacin and paramomycin, respectively. The reduced dimensions of the neural networks models selected enabled the derivation of simple quantification equations to transform the input variables into the output variable. These equations can be more easily interpreted from a chemical point of view than those provided by other ANN models. Copyright © 2007 John Wiley & Sons, Ltd.     

多肽定量构效关系与分子设计

综述了多肽定量构效关系和计算机辅助多肽分子设计方法的最新进展,重点介绍了多肽定量构效关系研究中的化学结构定量描述符和建立数学模型的统计方法,并对模拟肽学和虚拟组合多肽库在多肽分子设计中的应用进行了简要的论述.     

Artificial neural networks based on genetic input selection for quantification in overlapped capillary electrophoresis peaks

The application of multilayer perceptron artificial neural networks (MLP ANN) based on genetic input selection for quantification of the unresolved peaks in micellar electrokinetic capillary chromatography (MECC) is reported. An optimization strategy for genetic input selection was also proposed. When the corresponding CE peaks cannot be resolved completely only by separation techniques, MLP ANN based on genetic input selection can be a suitable tool to resolve the problem. Both the spectra and the electrophoretograms of the unseparated analytes were used as the multivariate input data. The two kinds of the data were suitable for quantification of overlapped CE peaks by MLP ANN based on genetic input selection. The study also shows that the applying of genetic input selection in MLP ANN can improve the precision of quantification in both completely and partially overlapped CE peaks to some extent.     

人工神经网络研究硝基苯化合物结构与毒性的关系

基于简单的化学基团描述符,应用人工神经网络研究了硝基苯类化合物对黑呆头鱼的毒性构效关系,并与多元线性回归相比较,结果显示了人工神经网络处理非线性问题的优越性.     

Artificial neural network‐based drug design for diabetes mellitus using flavonoids

Diabetes mellitus is a chronic metabolic disease involving the failure to regulate glucose blood levels in the body and has been linked with numerous detrimental complications. Studies have shown that these complications can be linked to the activities of aldose reductase (AR), an enzyme of the polyol pathway. Flavonoids have been identified as good AR inhibitors (ARIs) and are also strong antioxidants with radical scavenging (RS) activity. As such, flavonoids show potential to become a better class of ARIs because they are able to concurrently address the oxidative stress issue. In this article, we carried out quantitative structure‐activity relationship analysis of flavones and flavonols (members of flavonoid family) using artificial neural networks. Three computer experiments were conducted to study the influence of hydrogen (H), hydroxyl (? OH), and methoxyl (? CH₃) functional groups on eight substitution sites of the lead flavone molecule and to predict potential ARIs. Of 6561 possible flavones and flavonols, in experiment 1, we predicted 69 potent ARIs, and in experiment 2, we predicted 346 compounds with strong RS activity. In experiment 3, we combined these results to find overlapping compounds with both strong AR inhibition and RS activity and we are able to predict 10 potent compounds with strong AR inhibition (IC₅₀ < 0.3 μM) and RS activity (IC₂₅ < 1.0 μM). These 10 compounds show promise of being good therapeutic agents in the prevention of diabetic complications and is suggested to undergo further wet bench experimentation to prove their potency. © 2010 Wiley Periodicals, Inc. J Comput Chem, 2011     

Selection of quasi-optimal inputs in chemometrics modeling by artificial neural network analysis

Instrumentation spectra used for chemometrics analysis are often too unwieldy to model, as many of the inputs do not contain important information. Several mathematical methods are used for reducing the number of inputs to the significant ones only. Artificial neural networks (ANN) modeling suffers from difficulties in training models with a large number of inputs. However, using a non-random initial connection weight algorithm and local minima avoidance and escape techniques can overcome these difficulties. Once the ANN model is trained, the analysis of its connection weights can easily identify the more relevant inputs. Repeating the process of training the ANN model with the reduced input set and the selection of the more relevant inputs can proceed until a quasi-optimal, small, set of inputs is identified. Two examples are presented—finding the minimal set of wavelengths in benchmark diesel fuel NIR spectra, and in spectra generated in a recent work, modeling of “artificial nose” sensor array. In the last example, 1260 inputs were reduced to optimal sets of <10 inputs. Causal index calculation can analyze the influence of each of selected wavelengths on the predicted property. Some of the resulting minimal sets are not unique, depending on the ANN architecture used in the training. The accuracy of the resulting ANN models is usually better, and more robust, than the original large ANN model.     

Growing neural networks for a multivariate calibration and variable selection of time-resolved measurements

In this study, an algorithm for growing neural networks is proposed. Starting with an empty network the algorithm reduces the error of prediction by subsequently inserting connections and neurons. The type of network element and the location where to insert the element is determined by the maximum reduction of the error of prediction. The algorithm builds non-uniform neural networks without any constraints of size and complexity. The algorithm is additionally implemented into two frameworks, which use a data set limited in size very efficiently, resulting in a more reproducible variable selection and network topology.

The algorithm is applied to a data set of binary mixtures of the refrigerants R22 and R134a, which were measured by a surface plasmon resonance (SPR) device in a time-resolved mode. Compared with common static neural networks all implementations of the growing neural networks show better generalization abilities resulting in low relative errors of prediction of 0.75% for R22 and 1.18% for R134a using unknown data.