Because macrocycles potentially hold significant promise for certain drug discovery projects, the conformational modeling of macrocycles has attracted an increasing amount of interest in recent years. In this article, Schrödinger Associate Principal Scientist Shawn Watts, MacroModel Developer Pranav Dalal, Principal Scientist Teng Lin, and MacroModel project manager John Shelley present the promising results of recent work in this area.
The approaches described in the article below will form the basis for a script designed to work with Schrödinger Suite 2011, and we anticipate the eventual creation of a graphical interface with well-validated default settings for macrocycle searches.
Background
Over the last several years significant efforts have been made to explore both natural and synthetic drug-like molecules that contain macrocycles. The reasons for doing so are multifold: Macrocyclic drug-like molecules are diverse compounds with potentially high binding affinity and target selectivity.1 Despite the fact that they often violate the “rule of five”2 guidelines for selecting orally available drug-like molecules,3 they are present in more than one hundred drugs currently in the market.1,3,4 Macrocycles are also good candidates in a number of application areas including modifying small-molecule drug-like species by ring closure and inhibiting protein-protein interactions.5
However, in spite of their relative prevalence among commercially available drugs, some researchers nevertheless feel that macrocycles are underutilized.1,4 A symposium focusing on macrocycles was part of the 2009 ACS meeting in Washington DC (described in Ref. 4) and recently a special edition of Current Topics in Medicinal Chemistry focused on in silico studies of this class of molecules.3
The advantages conferred by the introduction of a macrocyclic ring are derived in large part from their conformational characteristics. Creating a large ring within a molecule can not only limit the number and variety of energetically accessible conformations, it can dictate the molecule’s overall shape, even for compounds that would otherwise have significant flexibility. It has been argued that building in an appropriate shape can reduce the free energy cost of selecting the appropriate conformations out of a smaller population pool,4,6 even for typical small molecule drug candidates.7
Recent publications on macrocycle conformer generation
Macrocycle-containing molecules present challenges for molecular modeling because the rings and often the molecules themselves are atypically large for drug-like molecules. Two recent studies have focused on the challenge of generating three dimensional structures for macrocycles. In the first, Bonnet and coworkers8 focused on evaluating how well a number of conformational search methods, including some that they developed, performed on 19 macrocycles (Figure 1). This careful study has helped to focus the field, providing a publicly available set of molecules to use as a benchmark set. The Bonnet study found that the authors’ own methods performed the best, with MacroModel generally outperforming the remaining methods. However, the authors ran MacroModel in a non-standard manner, effectively without energy minimizations or energy filtering during the search, unintentionally degrading its performance relative to the default settings.
Figure 1: Representative compounds from Bonnet et al.8 This set included 8 macrocycles containing between 6 and 20 glycines (labeled G6 to G20), 5 cyclodextrins containing between 6 and 14 sugar rings (D6 to D14), and 6 different natural oligopeptidic macrocycles (labeled P1-P6; depicted here is macrocycle P4, Microcystin-LR).
Subsequently, Labute developed a LowModeMD method,9 which is a variation on an older idea of performing a conformation search by repeatedly alternating a short molecular dynamics simulation and an energy minimization. A distinguishing feature of his approach is that at the start of each simulation the initial randomized velocities are filtered in a manner that shifts the thermal energy away motions involving high curvature of the potential energy surface. He obtained results superior to those published by Bonnet et al.8 for the benchmark set. In both studies most of the macrocycles were subjected to 10,000 search cycles.
MacroModel search methods for macrocycles
In our hands, using MacroModel with default settings yielded results similar to the best methods reported by Bonnet et al.8 However, since MacroModel’s defaults were selected with an eye towards searching typical (i.e., relatively small) drug-like molecules or dense protein-ligand complex environments, it seemed reasonable that performance could be significantly improved by tuning the search parameters. Indeed, we have found that there are quite a few ways to run MacroModel more effectively on macrocyclic systems.
While this work is ongoing, we report here two different approaches that yield comparably good results. The first uses MacroModel’s Large Scale Low Mode (LLMOD) method10 with non-default parameters. The second is a meta-sampling approach that uses a combination of simulation cycles and LLMOD search steps. For this benchmark study we used 10,000 search cycles for either method to be consistent with previous benchmark studies.
Macrocycle-specific LLMOD conformer searches
LLMOD searches progressively build up and refine a collection of conformers during the search. At any one stage of the search, a conformation is selected from the current collection and greatly perturbed in the direction of one of the low frequency normal modes for the molecule. The perturbed structure is then energy minimized and added to the collection of conformers if it is distinct and has a sufficiently low energy. The use of low frequency normal modes ensures that changes in atom positions are coordinated in a direction that has a reasonable chance of leading to another low energy conformation.
Because LLMOD searches are typically applied to protein-ligand environments in which there is a high density of explicitly represented atoms, the default settings are such that the amplified normal mode is not permitted to displace any atom by more than 6 Å. Furthermore, it takes a significant amount of time to determine the normal modes for a protein-ligand complex. Hence the default LLMOD setting is for normal modes to be determined only once using the input structure.
Drug-like molecules with macrocycles differ from protein-ligand systems in two important ways: They are much smaller, and have a lower explicit atom density. These properties suggest that it is practical to allow larger displacements of atoms within macrocycles, and that it is both practical and advantageous to recalculate the normal modes more frequently during the search. While a few other changes have also been made, the key adjustments to the default parameters involve increasing the maximum displacement to 18 Å and recalculating the normal modes more often – for the current study, every time a new conformation is generated.
Meta-sampling
The second MacroModel method is a meta-sampling method that employs both an LLMOD search and a simulation cycling strategy. Each simulation cycle consists of three steps:
1. A short stochastic dynamics simulation at high temperature
2. A short simulated annealing run during which the temperature is rapidly lowered
3. A minimization
Unlike Labute’s method, no special filtering of the velocities is used. Although the characteristics of a meta-sampling calculation will likely change in future versions of the algorithm, this approach currently uses the conformations generated by a calculation with 5,000 simulation cycles to seed a 5,000 step LLMOD search performed with the macrocycle-appropriate settings described above.
Comparing Search Methods for Macrocycles
To assess the performance of the newly developed MacroModel search strategies for macrocycles, we compared the results of these search methods on the benchmark set against those reported by Bonnet et al. for the SOS2 method8 and Labute for LowModeMD9. Better search methods should generally find lower energy conformers, if not the lowest energy conformer. Table 1 lists the energy of the lowest energy conformer for each of these four search techniques.
Table 1: Lowest energies found for all macrocycles in the benchmark set. Bold text is used for the lowest energy conformer(s) found for each macrocycle. The progressively more intense background colors correspond to progressively larger energy differences between the best and second best methods. Light green represents a difference of 1.0-2.5 kcal/mol, green represents a difference of 2.5-5.0 kcal/mol, and dark green represents a difference of greater than 5.0 kcal/mol. The two MacroModel methods are not compared against each other. All searches consisted of 10,000 search cycles except for some LowModeMD runs, marked by an asterisk (*) , in which cases the search was terminated automatically as part of the normal functioning of that product. MMFFs vacuum energies are reported for all test cases.
In most cases, both MacroModel methods have lower energies than the competing methods. For 9 of the 11 macrocycles for which the energy differences are larger than 1 kcal/mol, at least one MacroModel technique was able to find a lower energy conformer than the published studies. In 8 of these 11 cases, both MacroModel techniques were able to find a lower energy conformer. In cases with energy differences smaller than 1 kcal/mol, differences may be due to minor differences in processing between the different studies, such as using different convergence thresholds.
Coverage of conformational space is also important for drug discovery. One measure of conformational coverage is the number of unique conformers found during a search (Table 2). With the exception of about half the polyglycines in the benchmark set, MacroModel generally finds more conformations than the SOS2 method. Overall MacroModel and LowModeMD perform similarly well in terms of the number of conformations generated. However, their relative performance varies dramatically from system to system.
Table 2: Number of distinct macrocycle conformations for each macrocycle in the benchmark set. See Table 1 for background information on these calculations. Conformers are considered distinct if the RMSD values for optimally superimposed orientations are greater than 0.75 Å. The table only counts conformers whose relative energy is no more than 20 kcal/mol greater than that method’s lowest energy conformer for the relevant molecule. Bold text is used to highlight the method(s) that produce the most conformers for each macrocycle. MacroModel methods are not compared against each other.
Figure 2 shows the lowest energy conformations for P6, Tolybyssidin A, as obtained by various methods. (For P6, the SPE method outperformed SOS2 so Bonnet et al. provided the structure from that method.) The energy of the conformer found by LLMOD is more than 7 kcal/mol lower than that found by the other published methods. This conformer is also qualitatively different: It has beta-sheet like hydrogen bonding, and no side chains lie within the central region of the molecule. Conformers from the other two methods have more complex backbone geometries with side chains that penetrate into the central region. Because the lowest energy conformers generated by these methods can be qualitatively very different, the choice of search method can have significant consequences.
Figure 2: The lowest energy conformers and their MMFFs vacuum energies for macrocycle P6, Tolybyssidin A.
Summary and Future Work
For a common benchmark set of macrocycles and a common force field, two MacroModel-based search methods generally outperform the methods published by Bonnet et al. (namely, SOS2)8 and Labute (LowModeMD).9 We also show that the difference in results can correspond to qualitatively different low energy conformers.
High quality searches of macrocycles can be performed in recent versions of MacroModel using LLMOD, particularly if the size of the search steps is increased and the normal modes are updated during the search. To facilitate macrocycle conformations searches using MacroModel, a new macrocycle conformational sampling script will be available in the script center for Suite2011. To make macrocycle searches as easy as possible, we anticipate the eventual creation of a dedicated graphical interface with well-validated default settings.
References
(1) Driggers, E. M.; Hale, S. P.; Lee, J.; Terrett, N. K. Nat. Rev. Drug Discov. 2008, 7, 608.
(2) Lipinski, C. A.; Lombardo, F.; Dominy, B. W.; Feeney, P. J. Adv. Drug Deliv. Rev. 2001, 46, 3.
(3) Oyelere, K. A. Curr. Top. Med. Chem. 2010, 10, 1359.
(4) Drahl, C. Chem. & Eng. News 2009, 87, 54.
(5) Johnson, V. A.; Singh, E. K.; Nazarova, L. A.; Alexander, L. D.; McAlpine, S. R. Curr. Top. Med. Chem. 2010, 10, 1380.
(6) Brandt, W.; Haupt, V. J.; Wessjohann, L. A. Curr. Top. Med. Chem. 2010, 10, 1361.
(7) Avolio, S.; Summa, V. Curr. Top. Med. Chem. 2010, 10, 1403.
(8) Bonnet, P.; Agrafiotis, D. K.; Zhu, F.; Martin, E. J. Chem. Inf. Model 2009, 49, 2242.
(9) Labute, P. J. Chem. Inf. Model 2010, 50, 792.
(10) Kolossváry, I.; Keseru, G. M. J. Comp. Chem. 2001, 22, 21.