3D-Pharma: uma ferramenta para triagem virtual baseada em fingerprints de farmacóforos

Abstract

The pharmaceutical industry is going through a crisis of unheard proportions. Its cause can be related to the abrupt fall of new molecular entities approval by regulatory agencies, aggravated by the proximity of the expiration date of highly profitable classes of patented compounds and the increasing aggregated cost of drug design. The innovation on the drug discovery process may be one way out of this situation, and Bioinformatics and Chemoinformatics have a pivotal role on selection and rational design of drug candidates, looking for elimination of non-promising substances from the discovery pipeline before the highly costly clinical trials. In this context, Virtual Screening is regarded as a invaluable asset in rational drug design. Ligand-Based Virtual Screening is one of the oldest and most utilized techniques used in computer-aided molecular design, since chemical data is readily and widely available. Nevertheless, this work presents 3D-Pharma, a new Ligand-Based Virtual Screening method that uses fingerprints of pharmacophore triplets at atomic resolutions to build very simple and predictive models. Within 3D-Pharma the molecules are described by multiple representations that comprehend several prototropic species and conformations (multiple species, multiple mode approach). Pharmacophoric features are assigned to points that share spacial coordinates and interaction properties to heavy atoms of each molecular representation. All possible three-point pharmacophores of each representation of a moloceule are indexed in a fingerprint, and the multiple representations of a compound are concatenated into a unique fingerprint that accounts for most of its chemical and conformational diversity. The biological activity of an ensemble of active molecules are represented by a single modal fingerprint or model, validated through a new exhaustive 10-fold cross-validation scheme, which improves robustness and internal consistency of the models, as well as its predictive power. Retrospective validation studies were made with 10 datasets of active compounds and decoys gathered from the DUD database. They show the high predictive power of the models built by 3D-Pharma from three external and independent datasets of bioactive compounds (Drugs, PDB Ligands and WOMBAT), which was compared against seven state-of-the-art LBVS methods. We concluded that 3D-Pharma overperforms all other state-of-the-art LBVS tools analyzed, in terms of global accuracy as well as scaffold hopping and early recovery capacities. Furthermore, the models produced by 3D-Pharma are simple, robust, consistent and predictive.