DrugScoreFP: profiling protein-ligand interactions using fingerprint simplicity paired with knowledge-based potential fields
© Pfeffer et al. 2008
Published: 26 March 2008
Scoring functions used in structure-based drug design are often inefficient in reliably placing near-native geometries on the first scoring rank. Furthermore, there is no means to incorporate protein-specific information which additionally captures interaction details of different experimentally observed binding modes with the target protein under consideration.
Here, we present a vector-based extension of the DrugScoreCSD formalism , called DrugScore Fingerprints (DrugScoreFP), to rescore docking results. The original DrugScore of a docked inhibitor is partitioned into per-atom scores resulting in a 1D vector. Simple distance metrics allow the determination of similarities between fingerprints of docked compounds and reference fingerprints derived from crystal structures. Furthermore, DrugScoreFP allows the generation of family-based fingerprint profiles similarly implemented in SIFT [2, 3]. Therefore, a weighted consensus vector is derived from a given set of co-crystallized inhibitors with the target protein. Thus, DrugScoreFP binding profiles capture similarities and dissimilarities with respect to drug targets for which a large amount of structural data is available.
We have applied DrugScoreFP to handle the following tasks in structure-based drug design:
The recognition of near-native docking-poses was improved compared to DrugScoreCSD and SIFT using the Wang dataset . DrugScoreFP places geometries <0.5Å rmsd on the first scoring rank in 94% of the cases. This indicates an improvement compared to the original DrugScoreCSD of 6% and SIFT of 18%. Furthermore, cross-validation studies on different consensus fingerprints were performed with respect to a trypsin dataset consisting of 61 co-crystallized ligand structures. In a leave-one-out experiment, DrugScoreFP showed better recognition rates of crystal structures than docked compounds in 75% of the cases. As a final step, GOLD was used to dock 1800 compounds from the National Cancer Institute Diversity Set (NCI; http://www.nci.nih.gov) into trypsin and HIV-1 protease. DrugScoreFP shows superior ROC-AUCs of up to 99% compared to GOLD-Score (72%) and DrugScoreCSD (85%), using a fingerprint profile constructed from 61 and 22 co-crystallized ligands as query for the trypsin and the HIV-1 protease screen, respectively.
Finally, the results prove that DrugScoreFP can be used as a powerful filter, identifying similar binding profiles. It could also be shown that DrugScoreFP is stable with respect to cross-validations. It reliably discriminates near-native poses from widely spread decoys and retrieves active compounds diluted in a large dataset almost perfectly.
- Velec HF, Gohlke H, Klebe G: J Med Chem. 2005, 48 (20): 6296-303. 10.1021/jm050436v.View ArticleGoogle Scholar
- Deng Z, Chuaqui C, Singh J: J Med Chem. 2004, 47 (2): 337-44. 10.1021/jm030331x.View ArticleGoogle Scholar
- Chuaqui C, Deng Z, Singh J: J Med Chem. 2005, 48 (1): 121-33. 10.1021/jm049312t.View ArticleGoogle Scholar
- Mpamhanga C.P., et al: J Chem Inf Model. 2006, 46 (2): 686-98. 10.1021/ci050420d.View ArticleGoogle Scholar
- Wang R, Lu Y, Wang S: J Med Chem. 2003, 46 (12): 2287-303. 10.1021/jm0203783.View ArticleGoogle Scholar