An approach to the interpretation of backpropagation neural network models in QSAR studies

An approach to the interpretation of backpropagation neural network models for quantitative structure-activity and structure-property relationships (QSAR/QSPR) studies is proposed. The method is based on analyzing the first and second moments of distribution of the values of the first and the second partial derivatives of neural network outputs with respect to inputs calculated at data points. The use of such statistics makes it possible not only to obtain actually the same characteristics as for the case of traditional "interpretable" statistical methods, such as the linear regression analysis, but also to reveal important additional information regarding the non-linear character of QSAR/QSPR relationships. The approach is illustrated by an example of interpreting a backpropagation neural network model for predicting position of the long-wave absorption band of cyane dyes.     

Predicting physical-chemical properties of compounds from molecular structures by recursive neural networks

In this paper, we report on the potential of a recently developed neural network for structures applied to the prediction of physical chemical properties of compounds. The proposed recursive neural network (RecNN) model is able to directly take as input a structured representation of the molecule and to model a direct and adaptive relationship between the molecular structure and target property. Therefore, it combines in a learning system the flexibility and general advantages of a neural network model with the representational power of a structured domain. As a result, a completely new approach to quantitative structure-activity relationship/quantitative structure-property relationship (QSPR/QSAR) analysis is obtained. An original representation of the molecular structures has been developed accounting for both the occurrence of specific atoms/groups and the topological relationships among them. Gibbs free energy of solvation in water, Delta(solv)G degrees , has been chosen as a benchmark for the model. The different approaches proposed in the literature for the prediction of this property have been reconsidered from a general perspective. The advantages of RecNN as a suitable tool for the automatization of fundamental parts of the QSPR/QSAR analysis have been highlighted. The RecNN model has been applied to the analysis of the Delta(solv)G degrees in water of 138 monofunctional acyclic organic compounds and tested on an external data set of 33 compounds. As a result of the statistical analysis, we obtained, for the predictive accuracy estimated on the test set, correlation coefficient R = 0.9985, standard deviation S = 0.68 kJ mol(-1), and mean absolute error MAE = 0.46 kJ mol(-1). The inherent ability of RecNN to abstract chemical knowledge through the adaptive learning process has been investigated by principal components analysis of the internal representations computed by the network. It has been found that the model recognizes the chemical compounds on the basis of a nontrivial combination of their chemical structure and target property.     

How not to develop a quantitative structure–activity or structure–property relationship (QSAR/QSPR)

Although thousands of quantitative structure–activity and structure–property relationships (QSARs/QSPRs) have been published, as well as numerous papers on the correct procedures for QSAR/QSPR analysis, many analyses are still carried out incorrectly, or in a less than satisfactory manner. We have identified 21 types of error that continue to be perpetrated in the QSAR/QSPR literature, and each of these is discussed, with examples (including some of our own). Where appropriate, we make recommendations for avoiding errors and for improving and enhancing QSAR/QSPR analyses.     

Applicability domains for classification problems: Benchmarking of distance to models for Ames mutagenicity set

The estimation of accuracy and applicability of QSAR and QSPR models for biological and physicochemical properties represents a critical problem. The developed parameter of "distance to model" (DM) is defined as a metric of similarity between the training and test set compounds that have been subjected to QSAR/QSPR modeling. In our previous work, we demonstrated the utility and optimal performance of DM metrics that have been based on the standard deviation within an ensemble of QSAR models. The current study applies such analysis to 30 QSAR models for the Ames mutagenicity data set that were previously reported within the 2009 QSAR challenge. We demonstrate that the DMs based on an ensemble (consensus) model provide systematically better performance than other DMs. The presented approach identifies 30-60% of compounds having an accuracy of prediction similar to the interlaboratory accuracy of the Ames test, which is estimated to be 90%. Thus, the in silico predictions can be used to halve the cost of experimental measurements by providing a similar prediction accuracy. The developed model has been made publicly available at http://ochem.eu/models/1 .     

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.     

Quantum-chemical foundations of the topological substructural molecular design

The topological substructural molecular design (TOPS-MODE) approach is formulated as a tight-binding quantum-chemical method. The approach is based on certain postulates that permit to express any molecular property as a function of the spectral moments of certain types of molecular and environment-dependent energies. We use several empirical potentials to account for these intrinsic and external molecular energies. We prove that any molecular property expressed in terms of a quantitative structure-property and structure-activity relationships (QSPR/QSAR) model developed by using the TOPS-MODE method can be expressed as a bond additivity function. In addition, such a property can also be expressed as a substructural cluster expansion function. The conditions for such bond contributions being transferable are also analyzed here. Several new statistical-mechanical electronic functions are introduced as well as a bond-bond thermal Green's function or a propagator accounting for the electronic hopping between pairs of bonds. All these new concepts are applied to the development and application of a new QSAR model for describing the toxicity of polyhalogenated-dibenzo-1,4-dioxins. The QSAR model obtained displays a significant robustness and predictability. It permits an easy structural interpretation of the structure-activity relationship in terms of bond additivity functions, which display some resemblances with other theoretical parameters obtained from first principle quantum-chemical methods.     

QSAR and QSPR studies of a highly structured physicochemical domain

The relevance of terms other than linear when deriving quantitative structure-activity relationship/quantitative structure-property relationship (QSAR/QSPR) models has been rarely considered so far. In this study, the impact of quadratic and interacting terms has been taken into account. The first effect of including such highly structured terms is a significant extension of the parametric domain that moves from the initial N to N(N + 3)/2 parameters. This substantial enlargement over the conventional linear boundaries involves a higher computational cost due to the increased combinatorial number of resulting theoretical QSAR/QSPR models. To face this issue, novel genetic-algorithm-based software, MGZ (multigenetic zooming), was developed and used for both variable selection and model building. To speed up the entire process of domain searching, MGZ was supported with multiple independent evolving populations and genetic storms to further QSAR/QSPR analyses. In addition, a novel fitness function was developed to score models on the basis of their inner predictive capability, assessed on the training set, structure complexity, and presence of nonlinear terms. The models were further validated by monitoring model redundancy and performing intensive randomization runs. The Selwood data set was used as a reference set to derive QSAR models. Furthermore, a QSPR study was conducted on the solubility data set of a large array of organic compounds. The results reported in the present paper demonstrate that our approach is successful in finding linear models, which are at least as good as the models previously derived using standard statistical approaches, and in deriving new nonlinear models with good statistical figures.     

离子液体的定量结构-性质/活性研究

本文系统介绍了离子液体定量结构-性质/活性相关(QSPR/QSAR)的研究方法和步骤,综述了QSPR/QSAR在离子液体的熔点、有机物在离子液体中的无限稀释活度系数、离子液体的表面张力、离子液体的电导率、有机物在离子液体中的溶解度、离子液体的黏度以及离子液体的生物毒性和降解性等方面的最新研究进展,总结了该方法的优缺点,并对未来的研究趋势进行了展望。     

Neural networks convergence using physicochemical data

An investigation of the neural network convergence and prediction based on three optimization algorithms, namely, Levenberg-Marquardt, conjugate gradient, and delta rule, is described. Several simulated neural networks built using the above three algorithms indicated that the Levenberg-Marquardt optimizer implemented as a back-propagation neural network converged faster than the other two algorithms and provides in most of the cases better prediction. These conclusions are based on eight physicochemical data sets, each with a significant number of compounds comparable to that usually used in the QSAR/QSPR modeling. The superiority of the Levenberg-Marquardt algorithm is revealed in terms of functional dependence of the change of the neural network weights with respect to the gradient of the error propagation as well as distribution of the weight values. The prediction of the models is assessed by the error of the validation sets not used in the training process.     

On the topological sub-structural molecular design (TOSS-MODE) in QSPR/QSAR and drug design research

A recently introduced graph-theoretical approach to the study of structure-property-activity relationships is presented. The theoretical approach and the computational strategy for the use of the TOSS-MODE approach are given with details. Several QSPR and QSAR applications are reviewed including the study of physical properties of organic compounds, diamagnetic susceptibilities, and biological properties. The applications of the TOSS-MODE approach to discrimination of active/inactive compounds, the virtual screening of compounds with a desired property from databases of chemical structures, identification of active/inactive fragments and its relationships with 2D/3D pharmacophores, and to the design of novel compounds with desired biological activities are also reviewed.     

QSAR as a random event: criteria of predictive potential for a chance model

The CORAL software (http://www.insilico.eu/coral) was suggested as a tool to build up quantitative structure–property/activity relationships (QSPRs/QSARs). This software is based on conception “a QSPR/QSAR model should be interpreted as a random event.” This is reflection of fact: different distributions into the training set (substances involved in modeling process) and the validation set (substances, which are not known at the moment of the modeling process) give models with significant dispersion in the statistical quality of the QSPR/QSAR. Results of experiments with the software and possible ways of further improvement of this software are discussed. The most attractive new ways to estimate predictive potential of the CORAL model seem to be the following ones: (i) index of ideality of correlation and (ii) correlation contradiction index. These can be also proposed as criteria of predictive potential for arbitrary QSPR/QSAR.