Development and Validation of a Pharmacophore‐Based QSAR Model for the Prediction of CNS Activity

Medicinal Chemistry and Molecular Modeling Department, CEREP, Courtaboeuf, France.
ChemMedChem (Impact Factor: 2.97). 02/2009; 4(2):204-9. DOI: 10.1002/cmdc.200800282
Source: PubMed


A QSAR model for the prediction of CNS activity was developed and validated based on data from an in-house database of “drug-like” compounds. The model has demonstrated its applicability for novel chemical structures and its usefulness for the design of CNS-focused compound libraries.
A QSAR model aimed at predicting central nervous system (CNS) activity was developed based on the structure–activity relationships of compounds from an in-house database of “drug-like” molecules. These compounds were initially identified as “CNS-active” or “CNS-inactive”, and pharmacophore 3D descriptors were calculated from multiple conformations for each structure. A linear discriminant analysis (LDA) was performed on this structure–activity matrix, and this QSAR model was able to correctly classify approximately 80 % in both a learning set and a validation set. For validation purposes, the LDA model was applied to compounds for which the blood–brain barrier (BBB) penetration was known, and all of them were correctly predicted. The model was also applied to 960 other in-house compounds for which in vitro binding tests were performed on 20 receptors known to be present at the CNS level, and a high correlation was observed between compounds predicted as CNS-active and experimental hits. Finally, the model was also applied to a set of 700 structures with known CNS activity or inactivity randomly chosen from public sources, and more than 70 % of the compounds were correctly classified, including novel CNS chemotypes. These results demonstrate the applicability of the model to novel chemical structures and its usefulness for designing original CNS-focused compound libraries.

Download full-text


Available from: Nicolas Froloff, Sep 15, 2015
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Bioinformatics concerns itself with developing data-driven tools for extracting biological information from high-throughput datasets (e.g. gene expression data, protein interaction data). Until recently, most of these tools focus on analyzing a single source of data. However, all high-throughput measurements are noisy and incomplete. Therefore, we can improve the accuracy of our data driven predictions by combining various high-throughput measurements. In this thesis, we take a couple of steps towards data integration, which are reported in this thesis. We compare and introduce several computational methods for integrating different measurements.
    Preview · Article ·
  • Source

    Full-text · Article · Nov 2008 · Value in Health
  • [Show abstract] [Hide abstract]
    ABSTRACT: QSAR - Quantitative Structure-Activity Relationships, obtained by machine learning from data sets of molecules of measured properties, are key tools in modern day virtual screening protocols. These use computers, applying such models to predict the expected properties of not yet synthesized or tested compounds, then select molecules most likely to display wanted properties. Nowadays, QSAR became a specific field of machine learning, addressed by modelers, statisticians and members of the artificial intelligence community. Often, QSARs are developed for their own sake, just in order to prove that a novel machine learning technique is able to extract "knowledge" from a data set. However, QSARs should be effective tools, helping the chemist to rationally focus on synthesis/testing the candidate molecules most likely to fulfill expectations, as far as state-of-the-art may tell. Yet, experimental chemists typically have little insight into QSAR modus operandi, regarding it as a mathematical black box which returns a property value upon input of a molecular structure. This is unfortunate, because insufficient understanding prevents effective use. By combining theoretical insights with practical examples from the own experience and anecdotal aspects relating to the everyday use of QSAR models in both industrial and academic chemistry practice, the author wishes to contribute to a better understanding and a more effective use of these essential chemoinformatics tools. The goal of this paper is to bridge the cultural gap between experimentalists and model builders, by presenting QSAR in a different light. It focuses on practical aspects of machine learning, and of learning in chemistry in general: modeling in chemistry cannot be discussed while ignoring its psychological aspects, for any tool needs to first gain wide acceptance of its user community. QSARs would be better accepted by experimentalists once they understand that these are, like the entire body of experimental know-how, just empirical and fallible rules, extracted from a set of known examples.
    No preview · Article · Nov 2010 · Revue Roumaine de Chimie
Show more