QSAR Data Research Articles

Piece-wise quasi-linear modeling in QSAR and analytical calibration based on linear substructures detected by genetic algorithm

This paper introduces an alternative approach called piece-wise quasi-linear modeling methodology to split the data set into linear subsets when single linear calibration model failed to describe the whole data with desired residuals. The paper treats the linear models describing the sought-for subsets as hyperplanes in the data space. A modified genetic algorithm splits the data set into linear subsets according to a given maximum error. The proposed algorithm has successfully split a real QSAR data set into three chemically homogeneous linear subsets with very small residuals comparing with those obtained when a single linear model used to describe the data.

Theoretical analysis of peptidyl alpha-ketoheterocyclic inhibitors of human neutrophil elastase: Insight into the mechanism of inhibition and the application of QM/MM calculations in structure-based drug design.

It has been suggested from QSAR data (P. D. Edwards, D. J. Wolanin, D.A. Andisik and M. W. Davis, J. Med. Chem., 1995, 38, 76) that the inhibition of elastase by peptidyl alpha-ketoheterocyclic inhibitors can occur in two ways, the less potent inhibitors forming a non-bonded Michaelis complex and the more potent set a covalently bonded enzyme-substrate intermediate. We report QM/MM studies of both binding and reactivity that confirm these findings, showing that the activity of the least potent set of inhibitors correlates with the calculated binding energy, and that of the more potent set correlates with the stability of the intermediate. These calculations show that QM/MM methods can be successfully employed to understand complicated structure-activity relationships and might be employed in the design and assessment of new inhibitors.

Neural Networks in 3D QSAR

Abstract3D QSAR data analysis typically deals with large numbers of descriptors and various methods have been used to cope with this multivariate problem. One popular method, CoMFA, makes use of Partial Least Squares (PLS) regression. Alternative methods of data analysis to PLS have been explored including artificial neural networks (ANN). Within 3D QSAR, ANNs have been successful in producing models showing reasonable predictive capabilities. This review will examine recent studies in 3D QSAR and QSPR using ANN methods.

Toward generating simpler QSAR models: nonlinear multivariate regression versus several neural network ensembles and some related methods.

In this study we want to test whether a simple modeling procedure used in the field of QSAR/QSPR can produce simple models that will be, at the same time, as accurate as robust Neural Network Ensemble (NNE) ones. We present results of application of two procedures for generating/selecting simple linear and nonlinear multiregression (MR) models: (1) method for selecting the best possible MR models (named as CROMRsel) and (2) Genetic Function Approximation (GFA) method from the Cerius2 program package. The obtained MR models are strictly compared with several NNE models. For the comparison we selected four QSAR data sets previously studied by NNE (Tetko et al. J. Chem. Inf. Comput. Sci. 1996, 36, 794-803. Kovalishyn et al. J. Chem. Inf. Comput. Sci. 1998, 38, 651-659.): (1) 51 benzodiazepine derivatives, (2) 37 carboquinone derivatives, (3) 74 pyrimidines, and (4) 31 antimycin analogues. These data sets were parameterized with 7, 6, 27, and 53 descriptors, respectively. Modeled properties were anti-pentylenetetrazole activity, antileukemic activity, inhibition constants to dihydrofolate reductase from MB1428 E. coli, and antifilarial activity, respectively. Nonlinearities were introduced into the MR models through 2-fold and/or 3-fold cross-products of initial (linear) descriptors. Then, using the CROMRsel and GFA programs (J. Chem. Inf. Comput. Sci. 1999, 39, 121-132) the sets of I (I < or = 8, in this paper) the best descriptors (according to the fit and leave-one-out correlation coefficients) were selected for multiregression models. Two classes of models were obtained: (1) linear or nonlinear MR models which were generated starting from the complete set of descriptors, and (2) nonlinear MR models which were generated starting from the same set of descriptors that was used in the NNE modeling. In addition, the descriptor selection method from CROMRsel was compared with the GFA method included in the QSAR module of the Cerius2 program. For each data set it has been found that the MR models have better cross-validated statistical parameters than the corresponding NNE models and that CROMRsel selects somewhat better MR models than the GFA method. MR models are also much simpler than NNEs, which is the important surprising fact, and, additionally, express calculated dependencies in a functional form. Moreover, MR models were shown to be better than all other models obtained by different methods on the same data sets ("old" multivariate regressions, functional-link-net models, back-propagation neural networks, genetic algorithm, and partial least squares models). This study also indicated that the robust NNE models cannot generate good models when applied on small data sets, suggesting that it is perhaps better to apply robust methods (like NNE ones) on larger data sets.

Assessing model fit by cross-validation.

When QSAR models are fitted, it is important to validate any fitted model-to check that it is plausible that its predictions will carry over to fresh data not used in the model fitting exercise. There are two standard ways of doing this-using a separate hold-out test sample and the computationally much more burdensome leave-one-out cross-validation in which the entire pool of available compounds is used both to fit the model and to assess its validity. We show by theoretical argument and empiric study of a large QSAR data set that when the available sample size is small-in the dozens or scores rather than the hundreds, holding a portion of it back for testing is wasteful, and that it is much better to use cross-validation, but ensure that this is done properly.

Testing non-additivity of biological activity in a combinatorial library.

Combinatorial chemistry offers new opportunities to generate and analyze QSAR data. Traditional QSAR attempts to correlate activity with structure. With combinatorial chemistry, it is possible to correlate activity directly with the reagents used in a combinatorial library. If one can determine which reagents lead to the compounds of highest activities, it may then be possible to predict active compounds in virtual libraries of 10(6) to 10(10) compounds. This would greatly facilitate library design and provide confidence that the best compounds are being considered for synthesis. An important question is whether the activity of a product molecule can be considered as a sum of its components. This is referred to as additivity between reagents. If there is non-additivity, it is necessary to identify and include the non-additive terms in the model in order to improve QSAR models. Presented here are methods for developing QSAR models relating compound activity to reagents and a method for detecting the second effects of side-chain non-additivity. If the reagents in a library are shown to be additive in their contribution to activity, simple QSAR based on additive models can be applied confidently to reagents. Testing non-additivity can also guide the synthesis of the library. If the contributions are shown to be additive then the strategy for library synthesis may be shifted to include many reagents of a given type but not to make all combinations. The result is more efficient use of resources. In the analysis of percent inhibition data of a combinatorial library an additive model using reagents as descriptors yields a R(2) of 0.43. Application of this method is probably appropriate for HTS single point data while methods employing topological or pharmacophore based descriptors would be necessary to adequately model IC50 data.

Glycosyl derivatives of 2-bromosugar of selected non-steroidal anti-inflammatory drugs. Synthesis and QSAR data

On the basis of molecular modeling and calculation of physicochemical SAR data we obtained new derivatives of 2-bromosugars 5a– d and 6 for widely used non-steroidal anti-inflammatory drugs (NSAIDs). Compounds 5a– d and 6 were synthesized as stable, crystalline compounds from the reaction of phosphoroates 1– 3 with salicylic acid ( 4a), aspirin ( 4b), diclofenac ( 4c) and indomethacin ( 4d) in an aprotic solvent. The reaction took place in the presence of silver carbonate as the activator of leaving groups with high stereoselectivity and good yields. The structures of the new derivatives of 5a– d and 6 for NSAIDs were established by spectroscopic methods 1H and 13C NMR, and elemental analyses.

GIFI-PLS: Modeling of Non-Linearities and Discontinuities in QSAR

This paper introduces to the QSAR community a novel method for modeling and understanding non-linear relationships between biological potency and chemical structure properties of molecules. The approach, GIFI-PLS, is based on ``binning'' of quantitative X-variables into categorical variables. Each categorical variable is then expanded into a set of linked 1/0 dummy variables, which enable modeling of non-linearity. By way of four QSAR data sets, it is demonstrated that GIFI-PLS is useful for modeling of non-linearity and discontinuity in QSAR, and that the predictive power of a QSAR model may improve.

Unsupervised forward selection: a method for eliminating redundant variables.

An unsupervised learning method is proposed for variable selection and its performance assessed using three typical QSAR data sets. The aims of this procedure are to generate a subset of descriptors from any given data set in which the resultant variables are relevant, redundancy is eliminated, and multicollinearity is reduced. Continuum regression, an algorithm encompassing ordinary least squares regression, regression on principal components, and partial least squares regression, was used to construct models from the selected variables. The variable selection routine is shown to produce simple, robust, and easily interpreted models for the chosen data sets.

Prediction-weighted partial least-squares regression method (PWPLS)

Prediction-weighted partial least-squares (PWPLS) is a progressive approach of partial least-squares (PLS) regression. PWPLS is a simple, efficient, and evolutionary algorithm to select appropriate predictor variables, and weight each selected predictor variable for improving predictability when there is only one dependent variable. We applied PWPLS to two QSAR data sets, and compared the predictability results of PWPLS with that of PLS.

How do antibodies and lectins recognize histo-blood group antigens? A 3D-QSAR study by comparative molecular field analysis (CoMFA)

The cross-reaction patterns of nine antibodies and three lectins against 12 H type 2 related oligosaccharides have been analysed by means of 3D-QSAR study. Three-dimensional descriptors of the molecular properties have been used in comparative molecular field analysis (CoMFA). Three different alignments were considered for the oligosaccharides. One, based on the superimposition of the oligosaccharide core, could be correlated to most of the antibody activities. A second alignment, based on a superimposition of the fucose residue, had to be taken into account for explaining the binding properties of Ulex europaeus isolectin I. Analysis of the QSAR data gives indications on the carbohydrate epitopes essential for antibody recognition and yields some insights about the nature of the molecular recognition. This study complements previous biochemical estimates of the H type 2 related oligosaccharide binding areas (Mollicone, R.; Cailleau, A.; Imberty, A.; Gane, P.; Pérez, S.; Oriol, R. Glycoconj. J. 1996, 13, 263–271).

The use of cluster significance analysis to identify asymmetric QSAR data sets in toxicology. An example with eye irritation data.

Cluster significance analysis is a tool that allows the identification of "embedded clusters' in QSAR datasets. It is successfully applied to an eye irritation data set to show that these data are indeed asymmetric. The method identifies five parameters that form an embedded cluster of eye irritants amongst non irritants, although full separation is not achieved. This method has considerable potential to identify potential non-linearity in toxicology data sets and for parameter reduction. It is shown also that this can be obtained relatively quickly with an analysis performed on 100,000 subsets containing the same information as an analysis on 1,000,000 subsets.

Evolutionary variable selection in regression and PLS analyses

Evolutionary and genetic algorithms are powerful tools for searching global optima of complex functions. An evolutionary approach, the MUSEUM (mutation and selection uncover models) programme, is applied to various QSAR data sets to prove the general applicability of this approach for variable selection in regression and PLS analyses. ‘Best’ regression models are found within seconds or a few minutes of calculation time, even for data sets including large numbers of variables. The MUSEUM algorithm starts from an arbitrary model and adds or eliminates variables to or from this model in a random manner. Any ‘better’ model defined by a certain fitness criterion is taken as a new breeding organism which is mutated by further variable additions, eliminations or exchanges. In this manner the models improve gradually until a global optimum or at least a good local optimum results. In most cases several different models are obtained from different runs. A systematic search for the best models indicates that in all cases the global optima and good local optima result from the evolutionary search. Most often the fit and cross-validation results of these regression models are better than the fit and cross-validation results of a PLS analysis which includes all variables of the data set. The variables contained in the best regression models are suitable as subsets for PLS analyses and some of these PLS results are even better than the best regression results.

Evolutionary variable selection in regression and PLS analyses

Evolutionary and genetic algorithms are powerful tools for searching global optima of complex functions. An evolutionary approach, the MUSEUM (mutation and selection uncover models) programme, is applied to various QSAR data sets to prove the general applicability of this approach for variable selection in regression and PLS analyses. ‘Best’ regression models are found within seconds or a few minutes of calculation time, even for data sets including large numbers of variables. The MUSEUM algorithm starts from an arbitrary model and adds or eliminates variables to or from this model in a random manner. Any ‘better’ model defined by a certain fitness criterion is taken as a new breeding organism which is mutated by further variable additions, eliminations or exchanges. In this manner the models improve gradually until a global optimum or at least a good local optimum results. In most cases several different models are obtained from different runs. A systematic search for the best models indicates that in all cases the global optima and good local optima result from the evolutionary search. Most often the fit and cross-validation results of these regression models are better than the fit and cross-validation results of a PLS analysis which includes all variables of the data set. The variables contained in the best regression models are suitable as subsets for PLS analyses and some of these PLS results are even better than the best regression results.

Analysis of linear and nonlinear QSAR data using neural networks.

The use of feed forward back propagation neural networks to perform the equivalent of multiple linear regression has been examined using artificial structured data sets and real literature data. Their predictive ability has been assessed using leave-one-out cross-validation and training/test set protocols. While networks have been shown to fit data sets well, they appear to suffer from a number of disadvantages. In particular, they have performed poorly in prediction for the QSAR data examined here, they are susceptible to chance effects, and the relationships developed by the networks are difficult to interpret. This investigation reports results for one particular form of artificial neural network; other architectures and applications, however, may be more suitable.

Limitations of Functional‐Link Nets as Applied to QSAR Data Analysis

AbstractThis report compares the results of QSAR analyses examining a series of bis(1‐aziridinyl)‐p‐benzoquinones using multiple linear regression (MLR), neural networks and a network hybrid technique known as a functional‐link net (FUNCLINK). The FUNCLINK approach generates an expanded list of new parameters and seemed to outperform MLR as well as providing an apparently superior predictive ability. One observation from this work showed that neural networks generated higher correlation coefficients than MLR. This phenomenon, however, may be due to the inability of neural networks to predict values beyond their operating range. On examination of the results it appears that there are two disadvantages to the FUNCLINK technique. Firstly, the natural ability of neural networks to develop non‐linear relationships is removed with FUNCLINK as these must be specified. Secondly, the large number of enhanced parameters produced by FUNCLINK must increase the possibility of chance effects. From the results presented here it would appear that FUN‐CLINK adds little to the field of QSAR data analysis.

QSAR and Molecular Modelling Studies on Histamine H1‐Receptor Antagonists

AbstractIn this review article all relevant QSAR and molecular modelling studies that have been performed on both classical and non‐classical histamine H1‐antagonists are evaluated. Comparison of the results from these studies gives more insight into the physico‐chemical and conformational features that define receptor binding of H1‐antagonists. QSAR analyses indicate that at least seven classes of classical H1‐antagonists bind at the same receptor site. For these compounds a basic amino‐group, an aromatic ring (the so‐called “cis”‐ring) and a hydrophobic group (at the position of the so‐called “trans”‐ring) are essential for receptor binding. Similarities in the QSAR equations of diphenhydramines, mono‐phenyl analogues of diphenhydramines and benzimidazoles show that hydrophobic and steric factors are important for binding at the “trans”‐ring location. In literature it has been suggested that both aromatic rings of the non‐classical OPPI compounds can be superimposed in a folded conformation on the “cis”‐ and “trans”‐ring of the classical H1‐antagonist model. Although this idea remains interesting, QSAR data indicate that the proposed superimposition is probably not valid.Most molecular modelling studies performed on H1‐antagonists appear to have major draw‐backs: (1) e.g. only crystal structures were studied andor (2) antagonists were superimposed on the agonist histamine, whereas no experimental evidence is available that agonists and antagonists occupy similar receptor sites and/or (3) not all low energy conformations were considered. The most important conclusion that can be drawn from these modelling studies is that the optimal distance between the basic nitrogen atom and one of the aromatic rings is around 6 Å. Furthermore, two models of the H1‐antagonist binding site based on superimposition of antagonists on the classical semi‐rigid compound cyproheptadine are compared. It is concluded that cyproheptadine with the piperidylene ring in a boat conformation can be be used as a template for further modelling studies.

Regression analysis for QSAR using neural networks

Neural networks have been used to analyse QSAR data giving promising results. However, there is the danger of chance “correlations” and “over-fitting”. We have examined a reported analysis and shown that the size of the hidden layer can be reduced giving more efficient training while maintaining predictive performance.

Single Class Discrimination Using Principal Component Analysis (SCD‐PCA)

AbstractSingle Class Discrimination using Principal Component Analysis (SCD‐PCA) has been developed to discriminate an embedded data class. The embedded class is defined as the active class and the diffuse class, or classes as the inactives. Two basic methods are described. Both involve an initial column centring and scaling of the data to the actives space. Method I then directly obtains the directions of maximum variance in the inactives by PCA, while method II attempts to maximize the ratio of inactives to actives variance using PCA. The methods offer good potential to reveal a tightly bedded active class near the origin with the inactives dispersed. Techniques to portray the discrimination power and significance of the PCs are discussed. SCD‐PCA has been applied to artificial data sets exhibiting a variety of distributional characteristics and to two QSAR data sets. The method performed predictably on the artificial data. Good, low dimensional PC models were obtained for the QSAR data sets which were stable when cross‐validated.

QSAR and Multivariate Data Analysis of Amphiphilic Benzylamines and their Interaction with Various Phospholipids Determined by Different Methods

AbstractAmphiphilic drugs of various structures are known to interact with phospholipids. Simple benzylamines have been used as probes to test the structural influence on the degree of interaction. This interaction has been quantified with various methods as for example NMR‐binding measurements, scanning calorimetry, inhibition of Ca++‐incorporation, negative inotropic effect and antimycobacterial activity.Multiregression and multivariate data analysis have been performed and the results are discussed.