MovableType: Frequently Asked Questions

Note: These questions have been updated for DEV.671 or later. If you do not have a current version, please email sales@quantumbioinc.com for download instructions. As with any new method, there are bound to be some bumps in the road, and as critical protocol and usage questions arise, we will add them to this page. If you run into difficulty with the use of MovableType (MT), please Contact Us and we’ll be happy to work through your protocol with you, review your input for completeness and correctness, and make sure that we are providing the detail MT requires to successfully use the method for your structure based drug discovery efforts. In our experience with our clients, usually this process can be accomplished in a matter of a few days to weeks as long as some structural examples can be provided to illustrate your goals. The following questions are answered below:

What is the structure preparation protocol for MT?
What is the difference between “2-step” and “3-step” MT_ScoreE [Ensemble] and which is more predictive?
How important are quality docked poses to the MT process?
How well does MT_ScoreE (w/docked poses) compare to MT_ScoreES (w/X-ray poses)?
How many ligand poses are required to obtain good predictions? Can signal be “swamped” by bad poses?
How well does MT perform in pose selection (by RMSD from X-ray poses)?
Are there “cutoff” settings which can be used to improve predictability?
Can I use induced-fit docking instead of rigid-target docking? How well does it work?
General Support Note

What is the structure preparation protocol for MT?

This question is answered in the Structure Preparation and Scaling section of the MovableType Tutorials page. As a general rule, you should treat preparation for MT as you would treat preparation for other more computationally expensive applications like QM and MD. Just because a technology is fast does not mean it doesn’t require some investment in time for structure and ensemble preparation. MovableType is an “all atom” method which means it requires target selection and structure assessment, proton addition, rotomer and explicit water site exploration, structure minimization and refinement to clean up “hot spots,” and docking to generate high quality protein:ligand landscape minima. Finally, recently support for protein flexibility has been added for the MovableType toolbox specifically to support induced-fit docking and molecular dynamics snapshots.

What is the difference between “2-step” and “3-step” MT_ScoreE [Ensemble] and which is more predictive?

The “3-step” and “2-step” MT_ScoreE are two different approaches to pose ensemble scoring in MT. These methods are broken down into two protocols (see 3-step Tutorial and 2-step Tutorial respectively):

3-step MT_ScoreE: this method is the original, published approach to MT Ensemble scoring and is – as its name implies – broken down into three steps. First, the protocol uses MT_CS [Conformational Search] in order to generate ligand conformers which exist on the ligand free energy surface with the chosen pair potential (MT-GARF or MT-AMBER) and to calculate the unbound-ZL partition function. Second, the 5 to 10 most energetically favorable conformers are chosen and passed to the MOE docker which docks each conformer semi-rigidly (some in-dock optimization is usually used, but bond rotations and rotomer flips are kept to a minimum). And third, the top 10 to 25 poses according to the docker’s built in score function are passed to the Ensemble scoring tool. The benefit of the method is that ligand poses are guaranteed to exist on the energy surface. The drawback is that the docker is “stuck” with the conformers generated even if they will not properly fit the active site.
2-step MT_ScoreE: on the other hand, the 2-step approach skips the initial MT_CS step for conformer generation, and the docker is used both to generate the conformers of interest and to dock those conformers in the active site. As the screen output will show, MT_CS is still performed in order to calculate the unbound-ZL partition function, but the conformers themselves are not used. Again, the method uses 10 to 25 docked poses which are chosen according to the docker’s built in score function. This method is less error prone and is more accommodating to alternative binding mode selection. But there are times when its predictability profile is inferior.

While there are strengths and weaknesses of both approaches, when it comes to predictability, these two approaches are highly correlative with a R2:R2 correlation better than R2=0.9 for both potentials (as shown in the figures below). However, there are several outliers which make the 3-step worth considering in one’s protocol (especially if the 2-step isn’t predictive enough for one’s purposes). In particular, CDK2, Elongin, COMT, and CA2 all show significantly better predictions (as measured by higher LOO-R2) with the 3-step protocol.

How important are quality docked poses to the MT process?

Traditionally, we think of docking/scoring as a way to generate poses and score them based on a particular function in order to find the “one” pose which represents the binding mode of a ligand within an active site. However, we know that binding occurs on a landscape and is not based on a single ligand (or protein) pose. Conventional methods for modeling pose trajectories use molecular dynamics (MD) or monte carlo simulations. MovableType can use MD trajectory snapshots as well, but the first application of MT uses a combination of global docking and localized “smearing” to model these ensembles. The software assumes that all (or most) provided poses are landscape minima and all of them are included in the MT_ScoreE [Ensemble] score (unless they are removed due to redundancy, clashes, or other noted criteria). In our experience, a few poorly docked poses will likely have minimal impact on the Ensemble results; however, given the fact that all provided poses are included in the Ensemble score, we suggest that the user removes poses which they feel do not accurately represent binding. Instead, you should treat externally docked poses just as you would treat poses prepared for an MD simulation: if a pose doesn’t make sense for MD, then it likely doesn’t make sense for MT either. The following steps should be considered the bare minimum for successful MT_{ScoreE [}Ensemble] calculations:

Review support and other best practices options for the docking function you have chosen for placement. As is well documented in the literature, each docking function has strengths and weaknesses when it comes to different target classes. In our experience, MOE (with our settings) is the best and most MT-compatible 3rd party docker available. That said, MT_ScoreE has been used “in the field” with MOE, GLIDE, GOLD, OpenEye, DOCK, and other docking functions, and the the method has been implemented to be docker agnostic. But that does not mean that pose quality has no impact on results. You should choose the right docker for your target, and then make sure that the best practices associated with that method are followed. Here is a set of links for each of the most popular dockers:
1. Recommended: MOE from Chemical Computing Group (note: we have provided a SVL script called qbDockPair.svl to optimize the docking process)
2. GLIDE from Schrodinger, Inc.
3. OEDocking from OpenEye Scientific Inc
Protons should be added prior to docking. While the MT-GARF pair potential is a united atom method, docking functions tend to do a better job with protons in place and those protons are ultimately used in MT to correctly determine heavy-atom types.
In our experience, ligand structure minimization/optimization or refinement within the active site is critical during docking. In MOE, this refinement happens prior to the final MOE GBVI/WSA driven scoring, ordering, and filtering process and we have observed the best results (both in terms of RMSD and in terms of binding affinity prediction) with this extra step in the workflow. Given the fact that we are treating these poses as landscape minima, this observation is congruent with our expectations.
Finally, with that in mind, using the qbDockPair.svl tool, we have modified the default settings in MOE in order for the final GBVI/WSA score function to be presented with the best, possible poses. Instead of only optimizing 30 poses and scoring them to generate 5 final poses (i.e. a 30:5 ratio), we have increased these defaults to 125 poses for a final set of 25 poses (i.e. 125:25). If you choose to use an alternate docking function, then you should likewise manipulate any default settings in order to increase the likelihood of success. At the end of the day, MT – like any method – is susceptible for “garbage-in/garbage-out” so your landscape minima should be the best that they can be.

Note: In our experience, your best chance at success with MT is to let the docking software you choose spend the time it needs to generate its best landscape minima instead of trying to make docking as fast as possible. MT will more than make up for the extra docking time. To better understand the impact of your poses and to aid in selecrtion, the MT_ScoreES [EndState] score – which uses “smearing” alone – for each pose is also provided along with the energy components (ZP(i), ZPL(i), etcetera) and the RMSD from the wildtype pose (if provided). Please note: unless individual poses are excluded due to clashes or redundancy, all provided poses regardless of the RMSD are used in the Ensemble score.

---------------------------------- Pose Score ----------------------------------
 S Pose RMSD      ZPb(i)    ZPnb(i)     ZLb(i)    ZLnb(i)   ZPLnb(i)     dGsolv  MTScoreES MTScoreES(s)
--------------------------------------------------------------------------------
   1      0.26    -628.64   -6877.01    -207.87    -968.75  -13924.02       9.90  -13914.12      -8.90
   2      0.56    -628.64   -6877.01    -207.80    -968.83  -13736.76       9.99  -13726.77      -8.84
   3      0.87    -628.64   -6877.01    -207.98    -969.06  -13633.82      10.13  -13623.70      -8.80
   4      1.12    -628.64   -6877.01    -204.58    -968.53  -13451.31      10.37  -13440.95      -8.74
   19     8.83    -628.64   -6877.01    -196.49    -966.74  -13437.66      10.86  -13426.80      -8.73

The relationship between “Pearson’s R” and pose count for a few representative structures is presented in the Figure below. This relationship varies depending upon the structure being considered, so we recommend that you explore this question retrospectively for a target, then apply what you have learned prospectively for the same target.

How does MT_ScoreE w/docked poses compare to MT_ScoreES w/X-ray poses?

As shown in the MT_ScoreE [Ensemble] (R2) vs. MT_ScoreES [EndState] (R2) graphs below, we tend to see strong convergence in the prediction profiles of the two scores. In the figure below, the Leave One Out statistical analysis R2’s from MT_ScoreE (w/25 docked poses/ligand) on the Y-axis are compared to the Leave One Out analysis R2’s from MT_ScoreES (w/1 X-ray pose/ligand) on the X-axis. The diagonal along the 45° represents a “perfect” agreement between the two scores. Points above that diagonal represent cases where the MT_ScoreE (w/25 docked poses/ligand) is superior, while points below that diagonal represent cases where MT_ScoreES (w/1 X-ray pose/ligand) is the better method. With an R2:R2 correlation of 0.954, this relationship is exceedingly strong. However, there are several above-diagonal cases where the Ensemble (MT_ScoreE) method is more predictive.

How many ligand poses are required to obtain good predictions? Can signal be “swamped” by bad poses?

The answer to this question depends on the number of “good quality” poses provided by the docker and how many of these poses appear in the top-2, top-5, top-10, top-15, top-20, and top-25 of the SDF’s provided. As noted above, our goal with MT is to capture a number of landscape minima in order to correctly determine binding affinity. While a few inaccurate poses will likely not hurt the binding affinity prediction, a large number could “throw off” the prediction. Therefore, you should treat pose selection prior to MT characterization as you would treat structure preparation prior to MD and remove poses which do not accurately represent binding (since each provided pose is assumed to be a landscape minima). As a rule of thumb, we recommend starting with 10 to 25 good poses however even a few will often be enough if the poses are adequate. Some retrospective experimentation may be required depending upon the structure and the docker utilized, and then these settings can be used prospectively. The graph below shows the relationship between Pearson’s-R and the per-ligand pose count. In the case of CDK2, using poses provided by MOE, one would reach a plateau in predictive ability with the first 10 poses which are the “best” according to MOE’s built-in GBVI/WSA dG score function. While in the case of JNK1, any number of poses will do. Interestingly, too many poor poses – like in the case of thrombin depicted in the figure – will destroy the predictive ability of the method. So we recommend some verification is performed.

How well does MT perform in pose selection (by RMSD from X-ray poses)?

The MT method – like other methods using statistical mechanics – is built on the fundamental premise that binding is not dependent on a single pose and instead binding is described by myriad of poses or structures found on an energy landscape. That said, the method can be used to explore individual poses. The “Pose Score” table, depicted below, lists each of the provided poses along with individual score values. You can use this pose to help you remove poses which no longer make sense, or you can use this table to explore the overall range of poses being considered. The table is sorted by the MT_ScoreES value (which is the unscaled EndState score of that pose with the active site) and generally the lower energy scores drive the overall Ensemble score. When the wildtype pose is provided after the –ligand command line switch, and therefore when that pose chemically matches the poses provided in the dock.sdf file, the third column in the Pose Score table is the calculated RMSD of that pose.

---------------------------------- Pose Score ----------------------------------
 S Pose RMSD      ZPb(i)    ZPnb(i)     ZLb(i)    ZLnb(i)   ZPLnb(i)     dGsolv  MTScoreES MTScoreES(s)
--------------------------------------------------------------------------------
   1      0.26    -628.64   -6877.01    -207.87    -968.75  -13924.02       9.90  -13914.12      -8.90
   2      0.56    -628.64   -6877.01    -207.80    -968.83  -13736.76       9.99  -13726.77      -8.84
   3      0.87    -628.64   -6877.01    -207.98    -969.06  -13633.82      10.13  -13623.70      -8.80
   4      1.12    -628.64   -6877.01    -204.58    -968.53  -13451.31      10.37  -13440.95      -8.74
   19     8.83    -628.64   -6877.01    -196.49    -966.74  -13437.66      10.86  -13426.80      -8.73

In the table above, the top 4 poses listed are within 2 Å heavy-atom RMSD of the wildtype pose. When treating the entire CASF set, over 60% of the characterized structures have poses within 2 Å RMSD of the wild-type pose in the top-5 poses in this table and over 50% of the characterized structures have poses within 1 Å RMSD of the wild-type pose in the top-5 poses in this table.

Are there “cutoff” settings which can be used to improve predictability?

The settings for the MT_ScoreE and MT_ScoreES methods are fairly limited in part because success with the method is usually ultimately dependent upon the quality and size of the protein:ligand ensemble. The one area where some manipulation is possible which can lead to better results is in how the pocket is defined and how non-bonded cutoffs are applied. With the MT method, you don’t need the pocket to cover the entire structure and in fact often the inclusion of too much protein will “dampen” the impact of the ligand in the ligand binding calculation. This is likely an artifact of the method and one which is being actively explored. With that in mind, we have two ways to manipulate the cutoffs: the non-bonded atomic pair function cutoff and the pocket definition:

–nb-cutoff XX: Adding the –nb-cutoff XX command line argument and value to any scoring process will “throw out” any atom-atom non-bonded pairs beyond XX Angstroms. The DEFAULT value at this time is 11.0 Å.
–ligand placed_ligand.mol2 XX or –pocket placed_ligand.mol2 XX : The –ligand and the –pocket command line arguments are synonyms. The –ligand argument has always been used to define the pocket with regards to the position of the species found in the placed_ligand.mol2 file. The XX number in Angstroms can now be added in order to this command line in order to define the size of the pocket (residue extended) beyond the placed_ligand. The DEFAULT value at this time is 8.0 Å.

As an example, imagine the MT_ScoreE [Ensemble] calculation calls in the tutorials above. The following settings will change the above noted DEFAULT values to 8.0 for the non-bonded cutoff and 4.5 for the pocket cutoff. These values will often yeild good results for certain sets:

 ${DIVCON_INSTALL}/bin/qmechanic Bace_030215.pdb --ligand Bace_030215_CAT_4p.mol2 4.5 -h amberff14 --mtdock *_dock.pdb  --mtscore ensemble --nb-cutoff 8.0 --np 4 -v 2

The following table is provided as a guide for several combinations and for several sets. In each case, the Pearson-R value is being reported for the set. As you can see, spending a little time to “tune” these settings retrospectively for the various sets in your area(s) of interest can go a long way to providing you with improved results when you use these settings prospectively. Note: this table lists those structures which work well (CDK2, JNK1, p38, and so on), and those which work less well for rigid-docking with 25 MOE-provided poses (BACE, MCL1, and TYK2). Some additional tuning and including additional conformers/poses may take the improvement further. Likewise, inclusion of additional protein [target] sampling may also be required. Finally, in all cases 25 poses were provided, yet we know that in some cases (thrombin) the quality of the poses provided deteriorates.

MT-AMBER (Ensemble)
–nb-cutoff XX	–ligand mol2 XX	BACE	CDK2	JNK1	MCL1	p38	PTP1B	thrombin	TYK2
13.0	13.0	0.30	0.74	0.62	0.18	0.59	0.72	0.44	0.18
20000.0	15.0	0.22	0.73	0.72	0.22	0.55	0.73	0.57	0.05
8.0	4.5	0.22	0.75	0.71	0.34	0.64	0.71	0.32	0.08
MT-AMBER (EndState)
–nb-cutoff XX	–ligand mol2 XX	BACE	CDK2	JNK1	MCL1	p38	PTP1B	thrombin	TYK2
13.0	13.0	0.28	0.64	0.63	0.19	0.63	0.75	0.73	0.03
20000.0	15.0	0.22	0.73	0.72	0.21	0.55	0.73	0.57	0.04
8.0	4.5	0.29	0.26	0.63	0.30	0.62	0.75	0.53	0.01
MT-GARF (Ensemble)
–nb-cutoff XX	–ligand mol2 XX	BACE	CDK2	JNK1	MCL1	p38	PTP1B	thrombin	TYK2
13.0	13.0	0.30	0.74	0.63	0.19	0.59	0.72	0.33	0.12
20000.0	15.0	0.22	0.73	0.72	0.22	0.55	0.74	0.53	0.05
8.0	4.5	0.23	0.75	0.71	0.36	0.63	0.70	0.24	0.06
MT-GARF (EndState)
–nb-cutoff XX	–ligand mol2 XX	BACE	CDK2	JNK1	MCL1	p38	PTP1B	thrombin	TYK2
13.0	13.0	0.27	0.64	0.64	0.20	0.63	0.75	0.68	0.05
20000.0	15.0	0.22	0.73	0.72	0.22	0.55	0.74	0.53	0.05
8.0	4.5	0.29	0.67	0.65	0.32	0.65	0.77	0.47	0.03

This analysis has been expanded in order to include the CASF set. In the first two figures, the best settings for each case in MT-GARF and the best settings for each case in MT-AMBER are each graphed against the default settings. While in many cases the defaults will be effective, in individual cases (e.g. Factor Xa, TGT, CrtM, and so on) some experimentation with settings will improve your results. In the third figure, the “best” settings for MT-AMBER are put up against the “best” settings for MT-GARF, and again, we see that the methods are highly correlative but with crucial outliers suggesting that some retrospective experimentation can make significant impact to one’s predictive capabilities on a case-by-case basis.

Can I use induced-fit docking instead of rigid-target docking? How well does it work?

Traditionally, MovableType used “rigid-target / flexible-ligand” docking. This protocol was based on a limitation within the method which only allowed for very limited (if any) target-protein movement. With DEV.671 this limitation has been removed and “bound protein” poses are now supported. This addition is illustrated with the induced-fit tutorial. As depicted in the figures below, in most of the CASF cases considered, induced-fit and rigid-target docking are highly correlative. However, there are clear outliers in which one or the other method is superior. Going forward, an identical approach can be applied to bound-complex “snapshots” from molecular dynamics calculations: simply substitute these snapshots for the docked-protein:ligand poses in the induced-fit tutorial.

General Support Note

No tool will be 100% predictive, 100% of the time. But if you do not see results comparable to those provided by our internal calculations and our publications, we want to hear from you so we can help you with the process. Use the Contact Us link and we’ll follow up quickly.