Broad-Based Quantitative Structure−Activity Relationship Modeling of Potency and Selectivity of Farnesyltransferase Inhibitors Using a Bayesian Regularized Neural Network

Quantitative design rules for protein-resistant surface coatings using machine learning

Preventing biological contamination (biofouling) is key to successful development of novel surface and nanoparticle-based technologies in the manufacturing industry and biomedicine. Protein adsorption is a crucial mediator of the interactions at the bio - nano -materials interface but is not well understood. Although general, empirical rules have been developed to guide the design of protein-resistant surface coatings, they are still largely qualitative. Herein we demonstrate that this knowledge gap can be addressed by using machine learning approaches to extract quantitative relationships between the material surface chemistry and the protein adsorption characteristics. We illustrate how robust linear and non-linear models can be constructed to accurately predict the percentage of protein adsorbed onto these surfaces using lysozyme or fibrinogen as prototype common contaminants. Our computational models could recapitulate the adsorption of proteins on functionalised surfaces in a test set with an r2 of 0.82 and standard error of prediction of 13%. Using the same data set that enabled the development of the Whitesides rules, we discovered an extension to the original rules. We describe a workflow that can be applied to large, consistently obtained data sets covering a broad range of surface functional groups and protein types.

A renaissance of neural networks in drug discovery

INTRODUCTION

Neural networks are becoming a very popular method for solving machine learning and artificial intelligence problems. The variety of neural network types and their application to drug discovery requires expert knowledge to choose the most appropriate approach.

AREAS COVERED

In this review, the authors discuss traditional and newly emerging neural network approaches to drug discovery. Their focus is on backpropagation neural networks and their variants, self-organizing maps and associated methods, and a relatively new technique, deep learning. The most important technical issues are discussed including overfitting and its prevention through regularization, ensemble and multitask modeling, model interpretation, and estimation of applicability domain. Different aspects of using neural networks in drug discovery are considered: building structure-activity models with respect to various targets; predicting drug selectivity, toxicity profiles, ADMET and physicochemical properties; characteristics of drug-delivery systems and virtual screening.

EXPERT OPINION

Neural networks continue to grow in importance for drug discovery. Recent developments in deep learning suggests further improvements may be gained in the analysis of large chemical data sets. It's anticipated that neural networks will be more widely used in drug discovery in the future, and applied in non-traditional areas such as drug delivery systems, biologically compatible materials, and regenerative medicine.

Capturing the crystal: prediction of enthalpy of sublimation, crystal lattice energy, and melting points of organic compounds

Accurate computational prediction of melting points and aqueous solubilities of organic compounds would be very useful but is notoriously difficult. Predicting the lattice energies of compounds is key to understanding and predicting their melting behavior and ultimately their solubility behavior. We report robust, predictive, quantitative structure-property relationship (QSPR) models for enthalpies of sublimation, crystal lattice energies, and melting points for a very large and structurally diverse set of small organic compounds. Sparse Bayesian feature selection and machine learning methods were employed to select the most relevant molecular descriptors for the model and to generate parsimonious quantitative models. The final enthalpy of sublimation model is a four-parameter multilinear equation that has an r(2) value of 0.96 and an average absolute error of 7.9 ± 0.3 kJ.mol(-1). The melting point model can predict this property with a standard error of 45° ± 1 K and r(2) value of 0.79. Given the size and diversity of the training data, these conceptually transparent and accurate models can be used to predict sublimation enthalpy, lattice energy, and melting points of organic compounds in general.

Robust, quantitative tools for modelling ex-vivo expansion of haematopoietic stem cells and progenitors

Despite substantial research activity on bioreactor design and experiments, there are very few reports of modelling tools that can be used to generate predictive models describing how bioreactor parameters affect performance. New developments in mathematics, such as sparse Bayesian feature selection methods and nonlinear model-free modelling regression methods, offer considerable promise for modelling diverse types of data. The utility of these mathematical tools in stem cell biology are demonstrated by analysis of a large set of bioreactor data derived from the literature. In spite of the diversity of the data sources, and the inherent difficulty in representing bioreactor variables, these modelling methods were able to develop robust, quantitative, predictive models. These models relate bioreactor operational parameters to the degree of expansion of haematopoietic stem cells or their progenitors, and also identify the bioreactor variables that are most likely to affect performance across many experiments. These methods show substantial promise in assisting the design and optimisation of stem cell bioreactors.

CoMFA and CoMSIA analysis of tetrahydroquinolines as potential antimalarial agents

Comparative molecular field analysis (CoMFA) and comparative molecular similarity indices analysis (CoMSIA) were used on a dataset of compounds, some of them having been reported to inhibit Plasmodium falciparum protein, farnesyltransferase. The co-crystal structure of the lead molecule, BMS-214662 bound to Rat-PFT was used as a template. CoMFA yielded a good model, with r²(ncv) = 0.909, r²(cv) = 0.617 and was validated using an external set r²(pred) = 0.748). It compared favourably with CoMSIA. In the CoMFA model the steric and electrostatic fields exerted an almost equal influence on activity. The contour maps indicated the necessity for sterically large electropositive groups with electronegative tail to be present in these molecules for activity, and sterically large electronegative moieties on the sulfonamide linker. By incorporating these features some new compounds have been identified for further investigation.

N-tetrahydrothiochromenoisoxazole-1-carboxamides as selective antagonists of cloned human 5-HT2B

The serendipitous discovery of N-cyclohexyl-8-fluoro-3,3a,4,9b-tetrahydro-1H-thiochromeno[4,3-c]isoxazole-1-carboxamide as a selective human serotonin 5-HT2B antagonist with Ki of 42+/-5 nM is reported herein. A subsequent functional assay indicated little agonist activity compared to 5-HT itself.

Discovery of geranylgeranyltransferase-I inhibitors with novel scaffolds by the means of quantitative structure-activity relationship modeling, virtual screening, and experimental validation

Geranylgeranylation is critical to the function of several proteins including Rho, Rap1, Rac, Cdc42, and G-protein gamma subunits. Geranylgeranyltransferase type I (GGTase-I) inhibitors (GGTIs) have therapeutic potential to treat inflammation, multiple sclerosis, atherosclerosis, and many other diseases. Following our standard workflow, we have developed and rigorously validated quantitative structure-activity relationship (QSAR) models for 48 GGTIs using variable selection k nearest neighbor (kNN), automated lazy learning (ALL), and partial least squares (PLS) methods. The QSAR models were employed for virtual screening of 9.5 million commercially available chemicals, yielding 47 diverse computational hits. Seven of these compounds with novel scaffolds and high predicted GGTase-I inhibitory activities were tested in vitro, and all were found to be bona fide and selective micromolar inhibitors. Notably, these novel hits could not be identified using traditional similarity search. These data demonstrate that rigorously developed QSAR models can serve as reliable virtual screening tools, leading to the discovery of structurally novel bioactive compounds.

Imidazole-containing farnesyltransferase inhibitors: 3D quantitative structure-activity relationships and molecular docking

One of the most promising anticancer and recent antimalarial targets is the heterodimeric zinc-containing protein farnesyltransferase (FT). In this work, we studied a highly diverse series of 192 Abbott-initiated imidazole-containing compounds and their FT inhibitory activities using 3D-QSAR and docking, in order to gain understanding of the interaction of these inhibitors with FT to aid development of a rational strategy for further lead optimization. We report several highly significant and predictive CoMFA and CoMSIA models. The best model, composed of CoMFA steric and electrostatic fields combined with CoMSIA hydrophobic and H-bond acceptor fields, had r (2) = 0.878, q (2) = 0.630, and r (pred) (2) = 0.614. Docking studies on the statistical outliers revealed that some of them had a different binding mode in the FT active site based on steric bulk and available active site space, explaining why the predicted activities differed from the experimental activities.

Modeling of acetylcholinesterase inhibition by tacrine analogues using Bayesian-regularized Genetic Neural Networks and ensemble averaging

Acetylcholinesterase inhibition was modeled for a set of 136 tacrine analogues using Bayesian-regularized Genetic Neural Networks (BRGNNs). In the BRGNN approach the Bayesian-regularization avoids overtraining/overfitting and the genetic algorithm (GA) allows exploring an ample pool of 3D-descriptors. The predictive capacity of our selected model was evaluated by averaging multiple validation sets generated as members of diverse-training set neural network ensembles (NNEs). The ensemble averaging provides reliable statistics. When 40 members are assembled, the NNE provides a reliable measure of training and test set R values of 0.921 and 0.851 respectively. In other respects, the ability of the nonlinear selected GA space for differentiating the data was evidenced when the total data set was well distributed in a Kohonen Self-Organizing Map (SOM). The location of the inhibitors in the map facilitates the analysis of the connection between compounds and serves as a useful tool for qualitative predictions.

QSAR study on tetrahydroquinoline analogues as plasmodium protein farnesyltransferase inhibitors: a comparison of rationales of malarial and mammalian enzyme inhibitory activities for selectivity

The quantitative structure-activity relationships of Plasmodium falciparum and Rat protein farnesyltransferase (PFT) inhibitory activities of 6-cyano-1-(3-methyl-3H-imidazoly-4-ylmethyl)-3-substituted-1,2,3,4-tetrahydroquinoline (THQ) analogues are investigated in order to explore the similarities/deviations between the two enzymes for these analogues. The structure space of a ligand (BMS-214662) bound to Rat-PFT (PDB code 1SA5) has been used as the conformational space of the compounds under investigation. The study has been carried out using the combinatorial protocol in multiple linear regression with several 2D- and 3D-descriptors from molecular operating environment (MOE) representing the physicochemical and electronic features of the compounds. The molecular potential energy and partially charged van der Waals surface areas have taken part in the PFT models. They suggested in favor of molecular arrangement with minimum energy and low positively/negatively charged surfaces for optimum Pf-PFT inhibitory activity. Furthermore, less hydrophobic compounds are preferred for the activity. The Rat-PFT inhibitory activity models suggested in favor of more negatively as well as more positively charged surface area descriptors for the better activity. The PLS analysis carried out on the descriptors of the Pf-PFT and Rat-PFT models suggested that among the parameters, the partially charged surface areas in the range -0.20 to -0.15 (PEOE_VSA-3) and -0.30 to -0.25 (PEOE_VSA-5), hydrophobicity (a_hyd, logP(o/w) and SlogP_VSA4), and electronic energy (PM3_Eele) of the molecules hold promise for modulating the Pf-PFT/R-PFT inhibitory activities of the compounds. This suggested the possibility of modulating the Pf-PFT/R-PFT inhibitory activities and bringing about selectivity in the THQ analogues for the malarial parasite enzyme.

Network models in drug discovery and regenerative medicine

Network motifs and modelling paradigms are attracting increasing attention as modelling tools in drug design and development, and in regenerative medicine. There is a gradual but inexorable convergence between these hitherto disparate disciplines. This review summarizes some very recent work in these areas, leading to an understanding of the complementary roles networks play and factors driving this convergence: network paradigms can be excellent ways of modelling and understanding drug molecules and their action, an understanding of the robustness and vulnerabilities of biological targets may improve the efficacy of drug design and discovery, drug design has an increasingly large role to play in directing stem cell properties, stem cell regulatory networks can be modelled in useful ways using network models at a reasonable level of scale, and the network tools of drug design are also very useful for the design of biomaterials used in regenerative medicine.

Nonlinear predictive modeling of MHC class II-peptide binding using Bayesian neural networks

Methods for predicting the binding affinity of peptides to the MHC have become more sophisticated in the past 5-10 years. It is possible to use computational quantitative structure-activity methods to build models of peptide affinity that are truly predictive. Two of the most useful methods for building models are Bayesian regularized neural networks for continuous or discrete (categorical) data and support vector machines (SVMs) for discrete data. We illustrate the application of Bayesian regularized neural networks to modeling MHC class II-binding affinity of peptides. Training data comprised sequences and binding data for nonamer (nine amino acid) peptides. Peptides were characterized by mathematical representations of several types. Independent test data comprised sequences and binding data for peptides of length < or = 25. We also internally validated the models by using 30% of the data in an internal test set. We obtained robust models, with near-identical statistics for multiple training runs. We determined how predictive our models were using statistical tests and area under the receiver operating characteristic (ROC) graphs (A(ROC)). Some mathematical representations of the peptides were more efficient than others and were able to generalize to unknown peptides outside of the training space. Bayesian neural networks are robust, efficient "universal approximators" that are well able to tackle the difficult problem of correctly predicting the MHC class II-binding activities of a majority of the test set peptides.

3D-QSAR analysis of antimalarial farnesyltransferase inhibitors based on a 2,5-diaminobenzophenone scaffold

With annual death tolls in the millions and emerging resistance to existing drugs, novel therapies are needed against malaria. Wiesner et al. recently developed a novel class of antimalarials derived from farnesyltransferase inhibitors based on a 2,5-diaminobenzophenone scaffold. The compounds displayed a wide range of activity, including submicromolar, against the multi-drug resistant Plasmodium falciparum strain Dd2. In order to investigate quantitatively the local physicochemical properties involved in the interaction between drug and biotarget, we used the 3D-QSAR methods CoMFA and CoMSIA to study some of the series, including the screened lead compound 2,5-bis-acylaminobenzophenone, 28 cinnamic acid derivatives, 29 N-(3-benzoyl-4-tolylacetylaminophenyl)-3-(5-aryl-2-furyl)acrylic acid amides, and 34 N-(4-substituted-amino-3-benzoylphenyl)-[5-(4-nitrophenyl)-2-furyl]acrylic acid amides. We found that steric, electrostatic, and hydrophobic properties of substituent groups play key roles in the bioactivity of the series of compounds, while hydrogen bonding interactions show no obvious impact. We built several highly predictive 3D-QSAR models, including a CoMSIA one composed of steric, electrostatic, and hydrophobic fields, with r(2)=0.94, q(2)=0.63, and r(pred)(2)=0.63. The results provide insight for optimization of this class of antimalarials for better activity and may prove helpful for further lead optimization.

Modeling of farnesyltransferase inhibition by some thiol and non-thiol peptidomimetic inhibitors using genetic neural networks and RDF approaches

Inhibition of farnesyltransferase (FT) enzyme by a set of 78 thiol and non-thiol peptidomimetic inhibitors was successfully modeled by a genetic neural network (GNN) approach, using radial distribution function descriptors. A linear model was unable to successfully fit the whole data set; however, the optimum Bayesian regularized neural network model described about 87% inhibitory activity variance with a relevant predictive power measured by q2 values of leave-one-out and leave-group-out cross-validations of about 0.7. According to their activity levels, thiol and non-thiol inhibitors were well-distributed in a topological map, built with the inputs of the optimum non-linear predictor. Furthermore, descriptors in the GNN model suggested the occurrence of a strong dependence of FT inhibition on the molecular shape and size rather than on electronegativity or polarizability characteristics of the studied compounds.

Genetic neural network modeling of the selective inhibition of the intermediate-conductance Ca2+-activated K+ channel by some triarylmethanes using topological charge indexes descriptors

Selective inhibition of the intermediate-conductance Ca(2+)-activated K(+ )channel (IK (Ca)) by some clotrimazole analogs has been successfully modeled using topological charge indexes (TCI) and genetic neural networks (GNNs). A neural network monitoring scheme evidenced a highly non-linear dependence between the IK (Ca) blocking activity and TCI descriptors. Suitable subsets of descriptors were selected by means of genetic algorithm. Bayesian regularization was implemented in the network training function with the aim of assuring good generalization qualities to the predictors. GNNs were able to yield a reliable predictor that explained about 97% data variance with good predictive ability. On the contrary, the best multivariate linear equation with descriptors selected by linear genetic search, only explained about 60%. In spite of when using the descriptors from the linear equations to train neural networks yielded higher fitted models, such networks were very unstable and had relative low predictive ability. However, the best GNN BRANN 2 had a Q ( 2 ) of LOO of cross-validation equal to 0.901 and at the same time exhibited outstanding stability when calculating 80 randomly constructed training/test sets partitions. Our model suggested that structural fragments of size three and seven have relevant influence on the inhibitory potency of the studied IK (Ca) channel blockers. Furthermore, inhibitors were well distributed regarding its activity levels in a Kohonen self-organizing map (KSOM) built using the inputs of the best neural network predictor.

Predictive Bayesian neural network models of MHC class II peptide binding

We used Bayesian regularized neural networks to model data on the MHC class II-binding affinity of peptides. Training data consisted of sequences and binding data for nonamer (nine amino acid) peptides. Independent test data consisted of sequences and binding data for peptides of length </=25. We assumed that MHC class II-binding activity of peptides depends only on the highest ranked embedded nonamer and that reverse sequences of active nonamers are inactive. We also internally validated the models by using 30% of the training data in an internal test set. We obtained robust models, with near identical statistics for multiple training runs. We determined how predictive our models were using statistical tests and area under the Receiver Operating Characteristic (ROC) graphs (A(ROC)). Most models gave training A(ROC) values close to 1.0 and test set A(ROC) values >0.8. We also used both amino acid indicator variables (bin20) and property-based descriptors to generate models for MHC class II-binding of peptides. The property-based descriptors were more parsimonious than the indicator variable descriptors, making them applicable to larger peptides, and their design makes them able to generalize to unknown peptides outside of the training space. None of the external test data sets contained any of the nonamer sequences in the training sets. Consequently, the models attempted to predict the activity of truly unknown peptides not encountered in the training sets. Our models were well able to tackle the difficult problem of correctly predicting the MHC class II-binding activities of a majority of the test set peptides. Exceptions to the assumption that nonamer motif activities were invariant to the peptide in which they were embedded, together with the limited coverage of the test data, and the fuzziness of the classification procedure, are likely explanations for some misclassifications.

An in silico approach for screening flavonoids as P-glycoprotein inhibitors based on a Bayesian-regularized neural network

P-glycoprotein (P-gp), an ATP-binding cassette (ABC) transporter, functions as a biological barrier by extruding cytotoxic agents out of cells, resulting in an obstacle in chemotherapeutic treatment of cancer. In order to aid in the development of potential P-gp inhibitors, we constructed a quantitative structure-activity relationship (QSAR) model of flavonoids as P-gp inhibitors based on Bayesian-regularized neural network (BRNN). A dataset of 57 flavonoids collected from a literature binding to the C-terminal nucleotide-binding domain of mouse P-gp was compiled. The predictive ability of the model was assessed using a test set that was independent of the training set, which showed a standard error of prediction of 0.146+/-0.006 (data scaled from 0 to 1). Meanwhile, two other mathematical tools, back-propagation neural network (BPNN) and partial least squares (PLS) were also attempted to build QSAR models. The BRNN provided slightly better results for the test set compared to BPNN, but the difference was not significant according to F-statistic at p=0.05. The PLS failed to build a reliable model in the present study. Our study indicates that the BRNN-based in silico model has good potential in facilitating the prediction of P-gp flavonoid inhibitors and might be applied in further drug design.