Protein-ligand docking tends to be very good at generating binding modes that resemble experimental binding modes from X-ray crystallography and other methods (assuming we have a high quality structure…); but it is also very good at generating plausible models for ligands that don’t bind. These so-called “false positives” lead to reduced accuracy in structure-based virtual screening campaigns.
Structure-based methods are not the only way of approaching virtual screening: when all we know is the chemical structure of an active molecule, but nothing about its target (or targets), we can use ligand-based virtual screening methods, which operate on the principle of molecular similarity (Maggiora et al., 2014).
But what if we combine both methods?
There have been many attempts to combine structure-based and ligand-based virtual screening methods, but recent work by Jürgen Bajorath and co-workers (Anighoro & Bajorath, 2016) has shown promising results, by using a simplified pharmacophore derived from the X-ray crystal structure of an active ligand to re-rank binding modes generated by docking. Their pharmacophore uses four feature types: acceptor/acid, donor/base, aromatic, and hydrophobic, in addition to a simple weighting scheme that sets the weight to 1 if the atom possesses that pharmacophoric feature, and 0 if not. A simple Gaussian function dependent on the distance between the atoms in the two molecules (one an active crystallographic ligand, X, the other the docked database ligand, Y) and their van der Waals radii, all combines to give a 3D similarity score for the two molecules, F(X,Y), that is then normalized (Peltason & Bajorath, 2007).
For all four cases that they studied, (Dihydrofolate Reductase; Glucocorticoid Receptor; HIV-1 Protease; and Vascular Endothelial Growth Factor Receptor 2), using benchmark sets from DEKOIS 2.0 (Bauer et al. 2013) consisting of 40 actives and 1200 decoys, ranking by this 3D similarity score, Fnorm(X,Y), instead of the native docking score, tended to increase the overall AUC values computed from ROC curves, and in many cases, early enrichment increased also.
This effect tended to be observed regardless of the docking method used: they used AutoDock 4 and Chemical Computing Group’s MOE. For AutoDock 4, they ran 10 independent Genetic Algorithm dockings with 2.5 million energy evaluations, a population size of 150, and an RMSD clustering tolerance of 2 Å (it was surprising they didn’t use the Lamarckian GA, which has been shown to be more efficient than pure GA). For MOE, they used the triangle matcher algorithm, generating 1000 docked modes, with the 30 best London dG scores being refined using FF and re-scored using the GBVI/WSA dG scoring function.
Interestingly, when they used the best of 3 active ligands to re-rank instead of 1 active crystallographic ligand, the AUC values increased even more than when they used just 1 ligand, at least for the case they reported (GR: dexamethasone, hydrocortisone, and mifepristone >> mifepristone).
It remains to be seen if this approach is generally applicable, but these results look very promising.
Anighoro, A. and J. Bajorath (2016). “Three-Dimensional Similarity in Molecular Docking: Prioritizing Ligand Poses on the Basis of Experimental Binding Modes.” J Chem Inf Model 56(3): 580-587.
Bauer, M. R., et al. (2013). “Evaluation and optimization of virtual screening workflows with DEKOIS 2.0–a public library of challenging docking benchmark sets.” J Chem Inf Model 53(6): 1447-1462.
Maggiora, G., et al. (2014). “Molecular similarity in medicinal chemistry.” J Med Chem 57(8): 3186-3204.
Peltason, L. and J. Bajorath (2007). “Molecular similarity analysis uncovers heterogeneous structure-activity relationships and variable activity landscapes.” Chem Biol, 14(5): 489-497.