Combining pharmacophore fingerprints and PLS-discriminant analysis for virtual screening and SAR elucidation

The criterion of success for the initial stages of a ligand-based drug-discovery project is dual. First, a set of suitable lead compounds has to be identified. Second, a level of a preliminary structure-activity relationship (SAR) of the identified ligands has to be established in order to guide the lead optimization toward a final drug candidate. This paper presents a combined approach to solving these two problems of ligand-based virtual screening and elucidation of SAR based on interplay between pharmacophore fingerprints and interpretation of PLS-discriminant analysis (PLS-DA) models. The virtual screening capability of the PLS-DA method is compared to group fusion maximum similarity searching in a test using four graph-based pharmacophore fingerprints over a range of 10 diverse targets. The PLS-DA method was generally found to do better than the S-max method. The GpiDAPH3 and PCH fingerprints proved superior to the TGT and TGD fingerprints. Examples of SAR elucidation based on PLS-DA model interpretation of model coefficients using a reversible pharmacophore fingerprint are given. In addition, we tested the hypothesis that feature combinations coming from the analysis of two-dimensional (2D) pharmacophore fingerprints could be used to elucidate a three-dimensional pharmacophore (Williams, C. Mol. Diversity 2006, 10 (3), 311-332). This test was performed by mapping of pharmacophore triplets found by the PLS-DA model to be important for activity onto relevant ligands aligned by the protein-binding site known from X-ray complexes. The result of this analysis assists in explaining the efficiency of 2D pharmacophore fingerprints as descriptors in virtual screening.