Classification models for identifying substances exhibiting acute contact toxicity in honeybees (Apis mellifera)$.

Nowadays, environmental and biological endpoints can be predicted with in silico approaches if sufficient experimental data of good quality are available. Since the experimental evaluation of acute contact toxicity towards honeybees (Apis mellifera) is a complex and expensive assay, the computational models that follow OECD principles for this endpoint prediction represent important alternatives for safety prioritisation of chemicals, especially pesticides. We developed and validated counter-propagation artificial neural network (CPANN) models for in silico evaluation of toxicity of pesticides towards honeybees by using new in-house software. The data set included 254 pesticides with their toxicological experimental values (acute contact toxicity after 48 h of exposure - LD50 [µg/bee]). The 2D structures of compounds were mathematically represented with 56 Dragon molecular descriptors (MDs). The two-category models were developed to separate compounds as toxic or non-toxic for two different thresholds: (i) toxic when LD50 < 1 µg/bee and (ii) toxic when LD50 < 100 µg/bee. The models give reliable predictions in an external validation set and cover a large structural space. They were applied to a structurally diverse data set of 395 experimentally untested pesticides; 19% of them were predicted as highly toxic towards bees.

Application of fragment based virtual screening towards inhibition of bacterial N-acetyglucosaminidase$.

A structure-based approach is applied for the development of inhibitors of bacterial N-acetyglucosaminidase (autolysin). Autolysins are enzymes involved in the degradation of peptidoglycan and therefore participate in bacterial cell growth and different lysis phenomena. Several studies indicate that by the inhibition of autolysins, and consequently of bacterial cell division, antibacterial activity can be obtained, thus paving the road to a novel group of therapeutics against human pathogens. As crystal structures of the autolysin E (AtlE)-ligand complexes were obtained in our laboratories, fragment-based virtual screening was the method of choice for the initial studies. Fragment libraries from various databases were merged to increase the number of compounds for the virtual screening. Twenty-four commercially available virtual hits were selected and subjected to quantitative analysis of binding interactions using the surface plasmon resonance technique. Twelve fragments showed fragment-AtlE interactions. For F1, the top hit of the virtual screening, a KD of 228 µM was determined, while other fragments displayed non-stoichiometric binding. Blind docking of potential binders uncovers three possible allosteric sites. Ligands of N-acetyglucosaminidase identified in our study represent valuable information for the further development of AtlE inhibitors, which could in future represent antibacterial agents acting by a novel mode of action.

Comparison of in silico tools for binding site prediction applied for structure-based design of autolysin inhibitors.

Autolysin E (AtlE) is a bacteriolytic enzyme which plays an important role in division and growth of bacterial cells and therefore represents a promising potential drug target. Its 3D structure has been recently elucidated. We used in silico prediction tools to study substrate or ligand (inhibitor) binding regions of AtlE. We applied several freely available tools and a commercial tool for binding site identification and compared results of the prediction. Calculation time, number of predictions and output data provided by specific software vary according to the different approaches utilized by specific method categories. Despite different approaches, binding sites in similar locations on the protein were predicted. Specific amino acid residues that form these binding sites were predicted as binding residues. The predicted residues, especially those with predicted highest conservation score, could theoretically have catalytic and binding properties. According to our results, we assume that E138, which has the highest conservation score, is the catalytic residue and F161, G162 and Y224, which are also highly conserved, are responsible for substrate binding. Ligands developed with binding specificity towards these residues could inhibit the catalysis and binding of the substrate of AtlE. The molecules with inhibitory potency could therefore represent potential new antibacterial agents.

Robust modelling of acute toxicity towards fathead minnow (Pimephales promelas) using counter-propagation artificial neural networks and genetic algorithm.

Large worldwide use of chemicals has caused great concern about their possible adverse effects on human health, flora and fauna. Increased production of new chemicals has also increased demand for their risk assessment. Traditionally, results from animal tests have been used to assess toxicity of chemicals. However, such methods are ethically questionable since they involve killing and causing suffering of the test animals. Therefore, new in silico methods are being sought to replace the traditional in vivo and in vitro testing methods. In this article we report on one method that can be used to build robust models for the prediction of compounds' properties from their chemical structure. The method has been developed by combining a genetic algorithm, a counter-propagation artificial neural network and cross-validation. It has been tested using existing data on toxicity to fathead minnow (Pimephales promelas). The results show that the method may give reliable results for chemicals belonging to the applicability domain of the developed models. Therefore, it can aid the risk assessment of chemicals and consequently reduce demand for animal tests.

Chemometrics approach for the prediction of structure-activity relationship for membrane transporter bilitranslocase.

Membrane transport proteins are essential for cellular uptake of numerous salts, nutrients and drugs. Bilitranslocase is a transporter, specific for water-soluble organic anions, and is the only known carrier of nucleotides and nucleotide-like compounds. Experimental data of bilitranslocase ligand specificity for 120 compounds were used to construct classification models using counter-propagation artificial neural networks (CP-ANNs) and support vector machines (SVMs). A subset of active compounds with experimentally determined transport rates was used to build predictive QSAR models for estimation of transport rates of unknown compounds. Several modelling methods and techniques were applied, i.e. CP-ANN, genetic algorithm, self-organizing mapping and multiple linear regression method. The best predictions were achieved using CP-ANN coupled with a genetic algorithm, with the external validation parameter QV(2) of 0.96. The applicability domains of the models were defined to determine the chemical space in which reliable predictions can be obtained. The models were applied for the estimation of bilitranslocase transport activity for two sets of pharmaceutically interesting compounds, antioxidants and antiprions. We found that the relative planarity and a high potential for hydrogen bond formation are the common structural features of anticipated substrates of bilitranslocase. These features may serve as guidelines in the design of new pharmaceuticals transported by bilitranslocase.

Comparison of criteria used to access carcinogenicity in CPANN QSAR models versus the knowledge-based expert system Toxtree.

The primary goal of this study was to describe and compare the criteria used to assess carcinogenic activity. The statistically-based predictive quantitative structure-activity relationship (QSAR) models based on the counter propagation artificial neural network (CPANN) algorithm, and knowledge-based expert systems based on a decision tree structural alert (SA) approach (Toxtree application), were considered. The integration of the QSAR (CPANN models) and SAR (Toxtree SA application) approach contributed to the mechanistic understanding of the QSAR model considered. The mapping technique inherent to CPANN Kohonen enables us to relate the similarities or dissimilarities within a congeneric set of chemicals with particular SAs for carcinogenicity. The focus of our investigations was the similarities and dissimilarities of the features used in the QSAR and SAR methods. Due to the complexity of the carcinogenic endpoint, the integration of different approaches allows the models to be improved and provides a valuable technique for evaluating the safety of chemicals.

Classification of dry-cured hams according to the maturation time using near infrared spectra and artificial neural networks.

An attempt to classify dry-cured hams according to the maturation time on the basis of near infrared (NIR) spectra was studied. The study comprised 128 samples of biceps femoris (BF) muscle from dry-cured hams matured for 10 (n=32), 12 (n=32), 14 (n=32) or 16 months (n=32). Samples were minced and scanned in the wavelength range from 400 to 2500 nm using spectrometer NIR System model 6500 (Silver Spring, MD, USA). Spectral data were used for i) splitting of samples into the training and test set using 2D Kohonen artificial neural networks (ANN) and for ii) construction of classification models using counter-propagation ANN (CP-ANN). Different models were tested, and the one selected was based on the lowest percentage of misclassified test samples (external validation). Overall correctness of the classification was 79.7%, which demonstrates practical relevance of using NIR spectroscopy and ANN for dry-cured ham processing control.

On graphical representation of trans-membrane proteins.

In the first part of this paper, we present a novel graphical representation of proteins, which starts with constructing a map of a protein that is obtained from a matrix, the elements of which record the adjacencies of pairs of amino acids in the primary structure of a protein. Starting with the novel protein map, one interprets its matrix elements as vertices of a graph, which are labelled in sequential order as in the protein sequence. The nearest vertices are connected to the nearest neighbour which has a smaller label. In the second part of this paper, we describe the construction of protein binary codes that can serve as protein descriptors. This novel graphical representation of proteins is illustrated on segments of trans-membrane proteins, which are embedded in the membrane.

QSTR modelling of the acute toxicity of pharmaceuticals to fish.

Extensive use of pharmaceuticals as human and veterinary medication raises concerns for their adverse effects on non-target organisms. The purpose of this study was to employ multiple linear regression (MLR) to predict the toxicities of a diverse set of pharmaceuticals to fish. The descriptor pool consisted of about 1500 descriptors calculated using Dragon 5.4, Spartan 06 and Codessa 2.2 software. Descriptor selection was made by the heuristic method available in Codessa 2.2. The data set was divided into training and test sets using Kohonen networks. The training set contained approximately 65% of the compounds of the full data set (99 compounds). The training set model contained eight descriptors from all dimensions, all of which were obtained from Dragon 5.4. The statistical parameters of the model for the training set are R(2 )= 0.664, F = 13.588, and R(cv)(2) (LOO) = 0.542 while it achieves R(2 )= 0.605 for the test set. The training, test and external sets have no response outliers considering the standardized residual greater than three. The external validation of the model was made with a set of pharmaceuticals obtained from several databases. The R(pred)(2) is 0.777, reflecting a relatively good predictive power for the external set.

Counter propagation artificial neural network categorical models for prediction of carcinogenicity for non-congeneric chemicals.

One of the main goals of the new chemical regulation REACH (Registration, Evaluation and Authorization of Chemicals) is to fill the gaps on the toxicological properties of chemicals that affect human health. Carcinogenicity is one of the endpoints under consideration. The information obtained from (quantitative) structure-activity relationship ((Q)SAR) models is accepted as an alternative solution to avoid expensive and time-consuming animal tests. The reported results were obtained within the framework of the European project 'Computer Assisted Evaluation of industrial chemical Substances According to Regulations (CAESAR)'. In this article, we demonstrate intermediate results for counter propagation artificial neural network (CP ANN) models for the prediction category of the carcinogenic potency using two-dimensional (2D) descriptors from different software programs. A total of 805 non-congeneric chemicals were extracted from the Carcinogenic Potency Database (CPDBAS). The resulting models had prediction accuracies for internal (training) and external (test) sets as high as 91-93% and 68-70%, respectively. The sensitivity and specificity of the test set were 69-73 and 63-72% correspondingly. High specificity is critical in models for regulatory use that are aimed at ensuring public safety. Thus, the errors that give rise to false negatives are much more relevant. We discuss how we can increase the number of correctly predicted carcinogens using the correlation between the threshold and the values of the sensitivity and specificity.

Data-driven model for the prediction of protein transmembrane regions.

We present a novel approach combining mathematical methods and artificial neural networks to predict the transmembrane regions of transmembrane proteins, considering protein sequence information alone. We have focused on developing a data-driven model based on a non-linear modelling method, the counter-propagation artificial neural network, and on mathematical descriptors defining the sequence information of transmembrane proteins with known three-dimensional structures. The developed model has proven to be promising in predicting protein transmembrane regions, with an error below 10% for the external validation set. In combination with available experimental data the model can give us a better understanding of transmembrane proteins.

Spectral representation of reduced protein models.

We consider a spectrum-like two-dimensional graphical representation of proteins based on a reduced protein model in which 20 amino acids are grouped into five classes. This particular grouping of amino acids was suggested by Riddle and co-workers in 1997. The graphical representation is based on depicting sequentially the amino acids on five horizontal lines at equal separations. One-letter codes, B, O, U, X and Y, to which numerical values 1 to 5 have been assigned, are suggested as labels for the fictional amino acids that represent all the amino acids within each group. The approach is illustrated on ND6 proteins of eight species having from 168 to 175 amino acids. While visual inspection of the novel spectral graphical representations of proteins may reveal local similarities and dissimilarities of protein sequences, arithmetic manipulations of spectra offer an elegant route to graphic visualization of the degree of similarity for selected pairs of proteins.

On novel representation of proteins based on amino acid adjacency matrix.

A novel characterization of proteins is presented based on selected properties of recently introduced 20 x 20 amino acid adjacency matrix of proteins in which matrix elements count the occurrence of all 400 possible pair-wise adjacencies obtained by reading protein primary sequence from the left to the right. In particular we consider the characterization based on the sum and the difference of the rows and the corresponding columns, which characterize proteins by a pair of 20-component vectors. The approach is illustrated on a set of ND6 proteins of eight species.

Representation of proteins as walks in 20-D space.

A novel representation of proteins was introduced. It is independent of arbitrary decisions with respect to the choice of labels to be assigned to the 20 natural amino acids. The approach is based on an assignment of 20 unit vectors in 20-dimensional vector space to the 20 natural amino acids. Proteins are then represented by a walk, that is, a sequence of steps in the 20-dimensional space analogous to a walk in the (x, y) plane in the case of binary strings. A straightforward numerical characterization of proteins is obtained from the distance matrix associated with the walk representing the protein in 20-dimensional space combining the information on the Euclidean distance between various amino acids in protein sequence. The Line Distance matrix offers additional numerical characterization of proteins, while the lengths of steps of the walk in 20-D space allow construction of a "protein profile," which represents distribution of average lengths of the steps and their powers.

Chemometric approach in quantification of structural identity/similarity of proteins in biopharmaceuticals.

We present a chemometrics study in which we show the identity or degree of similarity of 3D protein structures of various G-CSF (Granulocyte Colony-Stimulating Factor) isolates. The G-CSF isolates share the same amino acid sequence, but the preparation was carried out by somehow diverse technologies. The comparison of 3D structures was made on the basis of 2D NMR NOESY (Nuclear Overhauser Enhancement Spectroscopy) spectra of proteins. In searching for the most appropriate criteria to determine the identity or degree of similarity of selected spectral regions of different isolates, two methods for quantitative evaluation of identity/similarity were used. The first method compares all peaks in the two investigated protein spectral regions; the extent of peaks that overlap is determined. The second method includes spectral invariants originating from graph theory. The criteria of identity/similarity were calculated from graphs, derived from a collection of up to 200 peaks of investigated 2D NMR spectral region. The peaks were linked into a graph according to the sequential nearest neighborhoods. According to the first method all peaks were relevant, considering that spectral noise was previously removed; the largest similarity was found between the protein of a commercially available G-CSF drug and one of the three new isolates produced in the laboratory. The second method indicated that the pairwise similarity of the three new isolates is larger than the similarity of any of the new isolates with the commercially available drug. This is an expected result taking into account that the new isolates are produced by the same technology, while the commercial product has additives for long-term storage that could not be completely compensated. The proposed measure of similarity may help the developers of biosimilar products to optimize the controllable parameters of the production technology and eventually to argue the identity of the new isolate in comparison with the originator commercial product.

Novel numerical and graphical representation of DNA sequences and proteins.

We have introduced novel numerical and graphical representations of DNA, which offer a simple and unique characterization of DNA sequences. The numerical representation of a DNA sequence is given as a sequence of real numbers derived from a unique graphical representation of the standard genetic code. There is no loss of information on the primary structure of a DNA sequence associated with this numerical representation. The novel representations are illustrated with the coding sequences of the first exon of beta-globin gene of half a dozen species in addition to human. The method can be extended to proteins as is exemplified by humanin, a 24-aa peptide that has recently been identified as a specific inhibitor of neuronal cell death induced by familial Alzheimer's disease mutant genes.

Enzyme binding selectivity prediction: alpha-thrombin vs trypsin inhibition.

In the present work we explore the possibility of an in-depth computational analysis of available experimental X-ray structures in the specific case of a series of alpha-thrombin and trypsin complexes with their respective inhibitors for the development of a novel scoring function based on molecular electrostatic potential computed at the contact surface in the enzyme-inhibitor molecular complex. We subsequently employ the chemometrical approach to determine which are the interactions in the large volume of data that determine the resulting experimental binding constant between ligand and receptor. The results of the model evaluated with molecules in the independent validation set show that a reasonable average error of 1.30 log units of the difference between experimental and calculated binding constants was achieved in the system thrombin-trypsin, which is comparable with those of methods from the literature. Furthermore, by a careful preparation of the Kohonen top layer in the artificial neural network approach that is normally perceived as a "black box device", we have been able to follow the implications of the structure of the inhibitor-enzyme complex for the inhibitor's binding constant. The method appears to be suitable for evaluation of selectivity in structurally similar enzymatic systems, which is currently an important problem in drug design.

Prediction of selectivity of alpha1-adrenergic antagonists by counterpropagation neural network (CP-ANN).

A quantitative structure-selectivity relationships of series of structurally diverse alpha1-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 alpha1-adrenergic antagonists.

Application of counterpropagation artificial neural network for modelling properties of fish antibiotics.

The present study focuses on fish antibiotics which are an important group of pharmaceuticals used in fish farming to treat infections and, until recently, most of them have been exposed to the environment with very little attention. Information about the environmental behaviour and the description of the environmental fate of medical substances are difficult or expensive to obtain. The experimental information in terms of properties is reported when available, in other cases, it is estimated by standard tools as those provided by the United States Environmental Protection Agency EPISuite software and by custom quantitative structure-activity relationship (QSAR) applications. In this study, a QSAR screening of 15 fish antibiotics and 132 xenobiotic molecules was performed with two aims: (i) to develop a model for the estimation of octanol--water partition coefficient (logP) and (ii) to estimate the relative binding affinity to oestrogen receptor (log RBA) using a model constructed on the activities of 132 xenobiotic compounds. The custom models are based on constitutional, topological, electrostatic and quantum chemical descriptors computed by the CODESSA software. Kohonen neural networks (self organising maps) were used to study similarity between the considered chemicals while counter-propagation artificial neural networks were used to estimate the properties.

Novel matrix invariants for characterization of changes of proteomics maps.

Previous studies on mathematical characterization of proteomics maps by sets of map invariants were based on the construction of a set of distance-related matrices obtained by matrix multiplication of a single matrix by itself. Here we consider an alternative characterization of proteomics maps based on a set of matrices characterizing local features of an embedded zigzag curve over the map. It is shown that novel invariants can well characterize proteomics maps. Advantages of the novel approach are discussed.