Publicly open databases of small compounds have become an indispensable tool for chemoinformaticians for collection and preparation of datasets suitable for drug discovery questions. Since these databases comprise compounds coming from structure-activity relationship (SAR) studies performed by different research groups, they are very diverse with respect to the biological assays used. In the present study we analyzed the applicability of a thoroughly curated dataset gathered from open sources for ligand-based studies, using the transient receptor potential vanilloid type 1 (TRPV1) as use case. Thorough curation of compounds according to the biological assay type and conditions led to a dataset of comparable bioactive chemicals. Subsequent exhaustive analysis of the obtained dataset using classification algorithms demonstrated that the models obtained in most of the cases possess reliable quality. Analysis of constantly misclassified compounds showed that they belong to local SAR series, where small changes in structure lead to different class labels. These small structural differences could not be captured by the classification algorithms. However application of the 3D alignment-independent QSAR technique GRIND for local, structurally related series overcomes this problem.
Two fast empirical charge models, Kirchhoff Charge Model (KCM) and Dynamic Electronegativity Relaxation (DENR), had been developed in our laboratory previously for widespread use in drug design research. Both models are based on the electronegativity relaxation principle (Adv. Quantum Chem. 2006, 51, 139-156) and parameterized against ab initio dipole/quadrupole moments and molecular electrostatic potentials, respectively. As 3D QSAR studies comprise one of the most important fields of applied molecular modeling, they naturally have become the first topic to test our charges and thus, indirectly, the assumptions laid down to the charge model theories in a case study. Here these charge models are used in CoMFA and CoMSIA methods and tested on five glycogen synthase kinase 3 (GSK-3) inhibitor datasets, relevant to our current studies, and one steroid dataset. For comparison, eight other different charge models, ab initio through semiempirical and empirical, were tested on the same datasets. The complex analysis including correlation and cross-validation, charges robustness and predictability, as well as visual interpretability of 3D contour maps generated was carried out. As a result, our new electronegativity relaxation-based models both have shown stable results, which in conjunction with other benefits discussed render them suitable for building reliable 3D QSAR models.
