Professor Friesner is a founder of Schrödinger and Professor of Chemistry and Director of the Center for Biomolecular Simulations at Columbia University. As chairman of Schrödinger’s Scientific Advisory Board, Professor Friesner provides strategic vision and guidance for Schrödinger's scientific advancements. In this installment of Rich's column, he describes advances in virtual screening methodology that have been incorporated into the Glide docking suite.
Since its inception several years ago, many of you have been using Glide's Extra Precision (XP) mode for enhanced docking accuracy. Glide XP is one of Schrödinger's main research and development projects, one into which we continue to invest major resources on an ongoing basis. It is also a project that I spend quite a bit of time working on myself.
The first detailed write-up of the XP methodology has just appeared in the Journal of Medicinal Chemistry. It contains a description of the XP sampling algorithms and scoring function, with the principal emphasis on the latter. As an integral part of the project, we have assembled large data sets from the literature and the Protein Data Bank, which we have used to evaluate XP performance for docking accuracy and enrichment in virtual screening. I encourage those of you who are interested in understanding the details of XP to consult this article, which can be found here.
The present discussion will give an overview of the results described in the XP publication. I will also briefly discuss some recent results that we have obtained at Columbia, in which we validate new XP binding affinity terms using explicit water molecular dynamics simulations. The paper describing these results is currently in press in the Proceedings of the National Academy of Sciences, and should be available soon.
The starting point of Glide XP scoring consists of terms that are common to empirical scoring functions present in most docking software: an atom-atom pair score that rewards contacts between lipophilic atoms on the protein and ligand, a term favoring protein-ligand hydrogen bonds, and an entropic penalty based upon the number of rotatable bonds in the ligand. XP also imposes desolvation penalties for burial of protein or ligand polar and charged groups.
The desolvation penalty is assessed by adding explicit waters to promising docked poses using a fast grid-based method, and then counting the number of water molecules in the first and second shells of each polar and charged group. These counts are compared with statistical averages of water shells for analogous groups in known active compounds, and penalties are assessed accordingly. The desolvation term plays a major role in reducing false positives in XP virtual screening.
The principal driving force for protein-ligand binding is the displacement of water molecules from the protein active site. Water molecules in hydrophobic environments have a tendency to lose orientational flexibility and hence entropy — a classic example is water at a hydrophobic wall, where the water molecules preserve their average number of hydrogen bonds, but in order to do so preferentially exclude geometries in which a hydrogen is pointing at the wall. Displacing such waters into bulk solution, and replacing them with a ligand that is well matched to the protein environment, thus yields a gain in free energy, one that is relatively small per water molecule, but can add up to a substantial value when integrated over the entire volume occupied by the ligand.
The lipophilic atom-atom pair term discussed above provides a heuristic representation of this effect, and is typically parametrized based on fitting to binding affinity data for a large number of protein-ligand complexes. As such, the calculated value will be reflective of an “average” protein active site environment. Similarly, the hydrogen bonding term captures the free energy gain upon replacement of water molecules which otherwise have to make hydrogen bonds to the protein with ligand groups that can do so with a smaller loss of entropy.
An empirical scoring function of this type can perform well for some fraction of protein-ligand complexes. However, we have found that there are environments that deviate substantially from the average, so much so, in fact, that the standard approximations become highly inaccurate. These environments are characterized by hydrophobic enclosure of the ligand, in which a cluster of hydrophobic atoms on the ligand (typically an aromatic ring, but other functional groups can also exhibit this behavior) is “surrounded” on two sides by hydrophobic protein groups (below). The enclosure implies that the water molecules displaced by the ligand in this region would have particularly unfavorable free energies, possibly due to a greater entropy loss than is usual, or even the actual loss of a hydrogen bond (or, in the most extreme case, dewetting of the cavity). Glide XP contains algorithms that recognize such regions automatically and assign additional favorable scores for binding affinity based on the geometry of the cavity and structure of the ligand.
Hydrophobic enclosure in 1aq1.
A particularly interesting structural motif occurs when hydrophobic enclosure is combined with a small number of protein groups that require hydrogen bonds. The hinge binding region in kinases is one such example of this motif, where an aromatic ring of the ligand typically makes 1-3 hydrogen bonds with protein backbone groups. XP recognizes this region and assigns additional binding affinity to ligands that form the necessary hydrogen bonds but are otherwise hydrophobic. The idea is that water molecules that make hydrogen bonds to multiple, closely spaced protein groups in a highly hydrophobic environment would experience a substantially larger than usual entropy loss when displaced by the ligand.
While we have developed and validated XP parameters by examining the performance of the methodology for a large number of virtual screening data sets, the effects postulated above should be manifested in accurate, all-atom simulations of the appropriate systems using an explicit solvation model. In collaboration with Bruce Berne’s group at Columbia, we have carried out such simulations for a number of systems identified by XP as having regions of hydrophobic enclosure. Our protocol is to remove the ligand from these systems, perform molecular dynamics simulations, and assess the distribution of water molecules via various statistical techniques (which can also be used to estimate the entropy of waters in various locations).
For the active site of the COX-2 receptor, which is highly hydrophobic, we see dewetting of the cavity even on a short simulation timescale. For the streptavidin/biotin complex, upon removal of biotin, 5 water molecules form an ice-like ring which hydrogen bonds to the protein in an enclosed region which formerly was occupied by the biotin ligand. These water molecules have very low entropies as compared to bulk; in effect, they have been frozen into position at room temperature. As a result, their displacement by biotin leads to an exceptionally large free energy of binding (~18 kcal/mol) — the largest of any complex in the PDB, despite the small size of the ligand.
These simulations demonstrate that Glide XP has a sound basis in the atomistic physical chemistry of protein-ligand complexes, as well as succeeding in explaining a wide range of empirical data. Glide XP has already been used by many groups in the pharmaceutical and biotechnology industry to facilitate both lead discovery and lead optimization efforts. Ongoing improvements of both the scoring function and sampling algorithms should make this technology even more compelling in subsequent releases.