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 ongoing and future research aimed at creating a viable approach for rank-ordering diverse sets of active compounds.

In a previous newsletter article, I discussed Schrödinger's efforts to develop the improved docking and scoring methods currently embodied in the XP Glide methodology. While XP Glide contains important advances compared to earlier scoring functions, it is clearly not, in and of itself, a complete solution to the docking and scoring problem. In particular, the parameterization of XP Glide has been focused on complexes for which the ligand is judged to “fit” into the receptor structure. When there are large scale induced fit effects, this assumption is typically not valid, and ligands that are incompatible with the particular receptor conformation used in a given docking study will often receive very poor scores. For flexible receptors, we have observed that as many as 50% of randomly chosen active ligands will fail to correctly fit into a single given receptor conformation. Clearly, a solution to this problem is needed before docking and scoring methods can be deployed robustly to investigate the large numbers of diverse active compounds located in a major drug discovery project.

Another problem not yet fully addressed by Glide XP is the ability to rank order compounds. To date, XP has been focused on separating “active” from “inactive” compounds. However, relative binding affinities of active compounds are not always rendered with high accuracy. For compounds that form congeneric series, MM-GBSA and FEP methods provide a useful approach to rank ordering. However, these approaches are poorly suited for highly diverse sets of compounds (e.g., those with significant structural differences, or those where the net charge on the ligand varies significantly throughout the data set). Again, progress on this problem would have a major impact on the value of docking and scoring approaches in a drug discovery context.

In this article, I will briefly outline the approach that Schrödinger is taking over the next several years to address these problems. In the initial phase of the project, we are performing studies primarily using publicly available data from the Protein Data Bank. As solutions that display efficaciousness for these data sets are developed, objective testing can be performed by pharmaceutical and biotechnology companies using proprietary in-house data sets. Note that in this approach, intellectual property concerns can generally be avoided as all that is necessary is a report of the accuracy of structural and binding affinity prediction for the proprietary data sets.

As a first step, we have chosen to study self-docking of co-crystallized complexes from the PDB. The development version of XP, augmented when necessary by quantum-polarized ligand docking (QPLD), in which QM/MM methods are used to generate charges for the ligand in the protein environment, is capable of achieving greater than 90% accuracy (RMSD under 2.0 Å) for self-docking. Our initial focus has been on data sets where there are 20 or more co-crystallized structures for a given receptor. For these cases, comparisons can be made for relative binding energies of the ligands docked into their cognate receptor conformations, and the accuracy of the docking can be rigorously evaluated. Hence, errors in the scoring function can typically be well separated from errors in the docked structures.

We are still in the process of completing these studies; however, the initial results look very promising. We have found that rank ordering of ligands is significantly improved by incorporating terms that take into account the strain energy of the protein-ligand complex, a quantity that has historically been difficult to construct models for. Our new model is performing well in preliminary tests, achieving r² values for correlations of theoretical and experimental binding affinities in the range of 0.45 - 0.65, and average errors on the order of 1 kcal/mole. While this is far from perfect, it does represent significant progress in modeling binding affinities for diverse data sets, containing a substantial number of different chemotypes – as is the case for most of the data sets derivable from the PDB as discussed above. It should be remembered, however, that these are in essence training set results, and that independent test set evaluations will be required for validation.

Our studies also indicate that the accuracy of binding affinity prediction is highly sensitive to the quality of the structures produced by the docking. This means that robust and accurate incorporation of induced fit effects is going to be necessary if rank ordering of compounds is to be extended to realistic situations involving cross-docking (as opposed to the self-docking studies carried out to date). The current Schrödinger induced fit methodology can deliver significant improvements for the structure of complexes where induced fit is important¹, but the present version is not fast enough for extensive virtual screening, and also needs to be tested on a much larger data set.

We are in the process of performing such tests, using the co-crystallized PDB complexes discussed above. If there are N complexes in a data set, there are in principle N²/2 cross-docking test cases, and of these, a substantial fraction are likely to exhibit induced fit effects. By studying this much larger data set, improving the efficiency of the algorithm, and assessing how well the induced fit complex reproduces the binding affinity score of the self-docking studies discussed previously, we believe it will be possible to create a qualitatively enhanced version of our induced fit approach which, in conjunction with improved scoring functions, will yield important progress in both rank ordering compounds for lead optimization, and in virtual screening.

Crystal structures remain the best types of data sets for testing the methodology, with regard to both the ability to accurately perform cross-docking and for evaluating the accuracy of scoring functions. As mentioned above, many biotechnology and pharmaceutical companies have proprietary data sets available, which should facilitate testing of this type. Once the reliability of cross-docking is established, there are large numbers of known active compounds in the open literature that can be profitably investigated.

We view the improvement of docking and scoring methods as a long range basic research project that is necessary for computational methods to become a full partner with experiment in drug discovery projects. In addition to the approaches described above, we are also working on incorporating the results of alternative calculations, such as MM-GBSA, linear response type methods, free energy perturbation, or molecular dynamics simulations, into the sampling and scoring processes. Ultimately a hierarchy of methods of correspondingly increasing accuracy, robustness, and computational cost is required to assemble a compelling computational platform for drug discovery. Schrödinger is committed to investing substantial resources in all of these areas and we expect significant progress to occur over the next several years as the full panoply of methods comes on line and is refined against an increasingly large and sophisticated repository of experimental data.

¹Sherman, W.; Day, T.; Jacobson, M. P.; Friesner, R. A.; Farid, R., “Novel Procedure for Modeling Ligand/Receptor Induced Fit Effects”, J. Med. Chem., 2006, 49, 534 -553.

Comments and questions on Dr. Friesner's column are welcome. Please send these via email to ask-rich@schrodinger.com, and we'll address particularly interesting topics in future newsletters.