Search

The performance of quantitative structure&ndash, activity relationship (QSAR) models largely depends on the relevance of the selected molecular representation used as input data matrices. This work presents a thorough comparative analysis of two main categories of molecular representations (vector space and metric space) for fitting robust machine learning models in QSAR problems. For the assessment of these methods, seven different molecular representations that included RDKit descriptors, five different fingerprints types (MACCS, PubChem, FP2-based, Atom Pair, and ECFP4), and a graph matching approach (non-contiguous atom matching structure similarity, NAMS) in both vector space and metric space, were subjected to state-of-art machine learning methods that included different dimensionality reduction methods (feature selection and linear dimensionality reduction). Five distinct QSAR data sets were used for direct assessment and analysis. Results show that, in general, metric-space and vector-space representations are able to produce equivalent models, but there are significant differences between individual approaches. The NAMS-based similarity approach consistently outperformed most fingerprint representations in model quality, closely followed by Atom Pair fingerprints. To further verify these findings, the metric space-based models were fitted to the same data sets with the closest neighbors removed. These latter results further strengthened the above conclusions. The metric space graph-based approach appeared significantly superior to the other representations, albeit at a significant computational cost.

Activating Liver X receptors (LXRs) represents a promising therapeutic option for dyslipidemia. However, activating LXR&alpha, may cause undesired lipogenic effects. Discovery of highly LXR&beta, selective agonists without LXR&alpha, activation were indispensable for dyslipidemia. In this study, in silico approaches were applied to develop highly potent LXR&beta, selective agonists based on a series of newly reported 3-(4-(2-propylphenoxy)butyl)imidazolidine-2,4-dione-based LXR&alpha, /&beta, dual agonists. Initially, Kohonen and stepwise multiple linear regression SW-MLR were performed to construct models for LXR&beta, agonists and LXR&alpha, agonists based on the structural characteristics of LXR&alpha, dual agonists, respectively. The obtained LXR&beta, agonist model gave a good predictive ability (R2train = 0.837, R2test = 0.843, Q2LOO = 0.715), and the LXR&alpha, agonist model produced even better predictive ability (R2train = 0.968, R2test = 0.914, Q2LOO = 0.895). Also, the two QSAR models were independent and can well distinguish LXR&beta, and LXR&alpha, activity. Then, compounds in the ZINC database met the lower limit of structural similarity of 0.7, compared to the 3-(4-(2-propylphenoxy)butyl)imidazolidine-2,4-dione scaffold subjected to our QSAR models, which resulted in the discovery of ZINC55084484 with an LXR&beta, prediction value of pEC50 equal to 7.343 and LXR&alpha, prediction value of pEC50 equal to &minus, 1.901. Consequently, nine newly designed compounds were proposed as highly LXR&beta, selective agonists based on ZINC55084484 and molecular docking, of which LXR&beta, prediction values almost exceeded 8 and LXR&alpha, prediction values were below 0.

Predicting blood-brain barrier (BBB) permeability is essential to drug development, as a molecule cannot exhibit pharmacological activity within the brain parenchyma without first transiting this barrier. Understanding the process of permeation, however, is complicated by a combination of both limited passive diffusion and active transport. Our aim here was to establish predictive models for BBB drug permeation that include both active and passive transport. A database of 153 compounds was compiled using in vivo surface permeability product (logPS) values in rats as a quantitative parameter for BBB permeability. The open source Chemical Development Kit (CDK) was used to calculate physico-chemical properties and descriptors. Predictive computational models were implemented by machine learning paradigms (decision tree induction) on both descriptor sets. Models with a corrected classification rate (CCR) of 90% were established. Mechanistic insight into BBB transport was provided by an Ant Colony Optimization (ACO)-based binary classifier analysis to identify the most predictive chemical substructures. Decision trees revealed descriptors of lipophilicity (aLogP) and charge (polar surface area), which were also previously described in models of passive diffusion. However, measures of molecular geometry and connectivity were found to be related to an active drug transport component.

The applicability of the newly developed RM1 and PM6 methods implemented in the semiempirical quantum chemistry mopac2009 software package in modeling free radical scavenging activity of flavonoids was examined. Bond dissociation enthalpy (BDE) of OH groups could be calculated much faster than with DFT method but with similar quality. Despite the known shortcomings of the Trolox equivalent antioxidant capacity (TEAC) assay, we show that taking into account the hydrogen atom transfer (HAT) mechanism of free radical scavenging of flavonoids encoded by minimal BDE values (BDE(min)) and the number of OH groups (nOH), as well as experimental data, reasonable QSAR models could be developed. For TEAC values of 38 flavonoids measured by the ABTS free radical, a model based on BDE(min) and nOH was developed, having very good statistical parameters (r=0.983, r(cv)=0.976). The applicability of this model to three different data sets of flavonoids and reliability of TEAC values measured in distinct laboratories were discussed. Finally, a reasonably good model of experimental vitamin C equivalent antioxidant capacity (VCEAC) of 36 flavonoids was obtained (r=0.954, r(cv)=0.947), involving BDE(min) and nOH as descriptors. Additionally, all presented models have comparable fit and cross-validated statistical parameters, as well as significant regression coefficients.