Structure-Based versus Property-Based Approaches in the Design of G-Protein-Coupled Receptor-Targeted Libraries

Abstract

In this work, two alternative approaches to the design of small-molecule libraries targeted for several G-protein-coupled receptor (GPCR) classes were explored. The first approach relies on the selection of structural analogues of known active compounds using a substructural similarity method. The second approach, based on an artificial neural network classification procedure, searches for compounds that possess physicochemical properties typical of the GPCR-specific agents. As a reference base, 3365 GPCR-active agents belonging to nine different GPCR classes were used. General rules were developed which enabled us to assess possible areas where both approaches would be useful. The predictability of the neural network algorithm based on 14 physicochemical descriptors was found to exceed the predictability of the similarity-based approach. The structural diversity of high-scored subsets obtained with the neural network-based method exceeded the diversity obtained with the similarity-based approach. In addition, the descriptor distributions of the compounds selected by the neural network algorithm more closely approximate the corresponding distributions of the real, active compounds than did those selected using the alternative method.

Design, Synthesis and Biological Evaluation of Ligands Selective for the Melanocortin-3 Receptor

“…Successful approaches to the design of focused compound libraries have recently been reported for the melanocortin 4 receptor and other GPCRs [73,[85][86][87][88]. They range from 2D simulation algorithms, to the analysis of ligand-receptor spatial arrangements and to neural network learning QSAR systems.…”

Status of HTS Data Mining Approaches

“…Constructing libraries targeted for one of these classes and refinements thereof is of common interest now. Recently, neural network techniques were used for designing GPCR targeted libraries [110,111] and serine protease targeted libraries [112]. A first SVM approach predicting GPCR ligands has also been published [113].…”

Bayesian Model for the Classification of GPCR Agonists and Antagonists

“…These training-set molecules are then analyzed to develop a decision rule that can be used to classify new molecules (the test set) into one of the two classes. 6 Various machine-learning techniques have already been suggested to be used to increase the chances of identifying novel GPCR ligands; for example, back-propagation neural networks with BCUT descriptors, 7 a combination of structure-based and property-based parameters, 8 pattern recognition techniques, 9 similarity searching and dynamic compound mapping, 10 or self-organizing neural networks with RDF descriptors. 11 Besides a wide variety of these available methods, Bayesian concepts and methodology has existed for many years to analyze structure activity data or to predict chemical properties; however, its popularity as a tool for substructural analysis within drug discovery and structure-activity analysis is somewhat recent.…”

“…For most of published works so far, only representative subgroups belonging to either class A and/or class B GPCR ligands were used. The main objective of these published works was either to classify GPCR and non-GPCR binding ligands 9 or to distinguish target-and family-selective GPCR antagonists. 10 Our model will be able to distinguish the mode of action of GPCR binding ligands, whether they are agonistic or antagonistic.…”