Short Trajectory MD with Analysis¶
The Floe Short Trajectory MD with Analysis (STMD) is for pose validation. You should start with a ligand with one or more conformers already well posed in the active site, with protein-ligand interactions in place, serving as hypotheses for ligand binding. Each pose will be evaluated by this floe to be validated (or not) as a good pose. We are not primarily asking the question “what is the pose for this ligand?” but rather “is this a good pose for this ligand?”, or with multiple poses, “which of these poses is best?” We seek an answer by running a short MD trajectory on each pose separately and comparing the results to the starting (input) pose and between the poses (if more than one).
What this Floe does¶
The structure of the STMD floe is shown in figure Structure of the STMD floe. Given the inputs of the protein and posed ligands, the complex is formed with each ligand/conformer separately, and the complex is solvated and parametrized according to the selected force fields. We refer to this ready-to-run molecular assembly as a “flask” by analogy to experiment: all the components are combined into the flask, upon which we run our experiment. A minimization stage is performed on the flask followed by a warm up (NVT ensemble) and several equilibration stages (NPT ensemble). In the minimization, warm up, and equilibration stages, positional harmonic restraints are applied on the ligand and protein. At the end of the equilibration stages, a short (default 2ns) production run is performed on the unrestrained flask. The production run is then analyzed in terms of interactions between the ligand and the active site and in terms of ligand RMSD, after fitting the trajectory based on active site C_alphas.
Just to be able to run, this floe requires ligands to have reasonable 3D coordinates, all atoms, and correct chemistry (in particular bond orders and formal charges). If the ligands already have good atomic partial charges (we recommend RESP or AM1-BCC_ELF10 charges), we recommend using these for STMD as opposed to re-charging them in the STMD floe.
Given that this floe only runs a very short timescale by default (2 ns), it is preferable that the input pose be well refined.
Although bad clashes (or poor positioning for interactions which you know are important) can be (and often are) cleaned up by even this short trajectory, it starts off the “evaluation” purpose of the floe on the wrong foot by giving a poor comparator.
Poor initial poses might even be considered outside the scope of this floe, given how short is the default timescale. This is why we strongly recommend that docked poses be subsequently minimized in the active site before input to STMD. This will resolve high gradients (usually clashes) with the protein and to allow protein-ligand interactions to optimize in the context of a good force field. It is possible that even with this pre-MD refinement, the docked-pose starting points could be reevaluated and triaged prior to the extra effort and expense of STMD. The ligand input datasets used in this tutorial are:
All the MD floes require correctly prepared protein up to “MD ready” standards. This begins with the normal prerequisites for physics-based modeling:
Protein chains must be capped,
All atoms in protein residues (including hydrogens) must be present, and
Missing protein loops resolved or capped.
Of course, protein side chain formal charges and protonation at this point determine their tautomeric state.
Additionally, cofactors and structured internal waters are also important to include, not only those in the immediate vicinity of the ligand and active site but also distally because they can have an important effect on the protein structure and dynamics over the course of the MD.
We strongly recommend using Spruce for protein preparation. The protein input dataset used in this tutorial is:
Unfortunately, proteins with covalently bound ligands or covalently bound cofactors are currently not tractable
How to use this floe¶
After selecting the Short Trajectory MD with Analysis floe in the Orion UI, you will be presented with a job form with parameters to select. In Figure STMD Job Form for ligand 35 (1 pose) you can see how we filled out the key fields of that form for the ligand 35 1-pose case described below.
Aside from the essential user-defined parameters relating to jobname, input (protein and ligand datasets as described above), and output (output and failure dataset names), all other parameters except “Flask_title” have reasonable defaults. This example is for an MCL1 protein, so after setting “Flask_title” to “MCL1”, launching the floe at this point is fine. That said, the top-level parameters you may consider changing are:
Flask_title (no default): Here is where you can put a handy short name for the protein to use in molecule titles (e.g. “Bace” instead of “beta-secretase”).
N_md_starts (default 1): This allows the user to ask for N independent starts to each ligand/pose, giving rise to N independent MD runs; this gives more sampling while keeping the simulation closer to the starting pose.
Charge_ligands (default True): If your input ligands already have good atomic partial charges (e.g. RESP or AM1-BCC_ELF10), set this to False to have the floe use the existing ligand charges.
Ligand_forcefield (default OpenFF2.0.0): This forcefield choice has a strong impact on the results. We recommend the most recent version of the OpenFF force field from the Open Force Field Initiative.
Md_engine (default OpenMM): Gromacs is the other alternative but we recommend OpenMM because HMR works with it but not yet with Gromacs.
Hmr: Hydrogen Mass Repartitioning (HMR) gives a two-fold speedup and reduces cost. We recommend leaving it on.
We make the other top-level parameters available for expert users by turning on “Show Cube Parameters” at the bottom of the input form and then drilling down into the parameters of the desired cube in the list below.
Understanding the results for a single ligand¶
The results from the STMD floe are accessed via two main avenues: through the job output in the Jobs tab in Orion’s Floe page, and through Orion’s Analyze page. We will look at the results of two jobs run on the same MCL1 ligand; in the first case the input ligand had only a single pose and in the second case it had six slightly different docked poses.
MCL1 ligand 35: single input pose¶
First we will look at the results of the single-pose run, with default of 1 for N_md_starts: one start of one ligand with one pose, so one 2 ns MD run overall. In the Jobs tab in Orion’s Floe page, having selected the job name for your STMD job, you should land on the job results page. The left panel contains the usual Orion job information from the run, and the right panel has one tab at the top if the run was not successful or two tabs at the top if it was… we will focus on success here! Selecting the second tab called FLOE REPORT should give you a page looking similar to Figure STMD Job results page for a single pose of an MCL1 ligand.
The floe report shows a tile for each MD simulation, here there was only one ligand in the input file. The atom colors correspond to calculated B-factors, similar to Xray B-factors, depicting the mobility of those atoms in the active site over the course of the MD trajectory. This gives an immediate read-out on how much various fragments of the ligand were moving around in the active site. As a general principle greater movement suggests that that fragment is not as tightly bound in the active site, but inferences are only qualitative. Certainly fragments hanging out in water of even a tightly bound inhibitor will be expected to be more mobile than the buried parts. Other information on each tile is:
The ligand name.
The number of clusters formed by clustering the ligand positions in the MD trajectory.
The Boltzmann-weighted MMPBSA score for ligand binding over the trajectories for all poses.
The simple ensemble average BintScore [*]_ for ligand binding over the trajectories for all poses (lower score is better).
The stability of the pose relative to the starting pose (varies between 0 (no stability) and 1 (completely stable)).
Clicking on the tile drills down into the detailed analysis of that simulation, resulting in Figure Detailed results for ligand 35 (single pose):
In this MD run, the results show a single cluster for the whole trajectory. Because MD is only sampling an equilibrium, the results are not deterministic: In you run this same MD again, in principle you could get slightly different results. Nevertheless, the results are interpreted in the context of the sampling obtained.
In the graphic we see a 2D representation of the ligand binding interactions for the whole trajectory, with the default display of the Overall tab at the top of the graphic. It is an interactive graphic: selecting the Cluster 0 tab in blue will change the binding interaction representation to that corresponding to the selected cluster. Hovering over one of the interaction in the diagram lights up a strip chart on the right-hand side grey arrow showing the occupancy of that interaction over the course of the trajectory. Within the heavy frame of the graphic, we see that the interactive graph is on interactions; selecting torsions changes the depiction to show a heavy black dot in each rotatable bond. Hovering over one of these shows a radial bar graph of the occupancy of the torsion on the right-hand side. Selecting B-factor yields a depiction of the calculated B-factors for the selected cluster as in the parent tile, but additionally shows the calculated B-factor for each active site amino acid close to the ligand. To the left of the graphic is information about the clustering of the ligand trajectory, including a table giving the ensemble average MMPBSA energy and BintScore (each with standard error) for each cluster. The MMPBSA value represents a Boltzmann-weighted average over all major clusters, But for BintScore it is a simple ensmble average for the ligand as a whole. Note with only one cluster here, the Boltzmann-weighted result represents cluster 0 completely. The remaining value, “Pose Stability”, is derived from the ensemble BintScore and represents how stable the overall protein-ligand binding interactions are compared to the starting pose (varies between 0 (no stability) and 1 (completely stable)).
Scrolling down exposes a strip chart and two tables detailing relevant analyses of the trajectories for all poses of the ligand. The strip chart for ligand 35 (single pose) is shown in Figure Strip Chart results for ligand 35 (single pose):
The strip chart shows a time course over the MD trajectory, maintaining always the same color scheme as in the interactive graphic: blue for cluster 0. Additionally, cluster outliers, which are ligand configurations that do not belong to any cluster, are shown in black. The chart simply shows the cluster occupancy of each frame, telling us that the trajectory spent most of the time in the blue Cluster 0, occasionally sampling outliers. It seems like quite a stable pose!
The two tables below the strip chart, shown in Cluster/Pose information for ligand 35 (single pose) describe a relationship between each cluster found in the MD for the ligand and the starting poses.
With only one pose used for this run the tables are terse, but below when we look at 6 input poses for the same ligand they will be more informative. The upper table “Cluster RMSD from each Starting Pose” describes how closely each cluster stays to the starting pose: the blue Cluster 0 sticks closely to the initial pose (1.38 Å RMSD). The second table “Cluster Percentage by Starting Pose” simply describes the occupancy that we see in the strip chart: the ligand spends 96% of its time in cluster 0. These figures tells us the blue Cluster 0 is stable and stays close to the initial pose.
MCL1 ligand 35: 6 input poses¶
Now we will look at the results of another run on the same ligand 35, but this time with 6 different input poses: 3 related poses with the methyl “up” in the upper panel of Figure Input poses for the 6-pose run and 3 related poses with the methyl “down” in lower panel of the same Figure. The “up” and “down” poses are only differentiated in the Figure for clarity; in the input file all 6 poses are together as the 6 conformers of the ligand 35 molecule. Poses 0, 3, and 5 have the methyl “down” and poses 1, 2, and 4 have the methyl “up”… this will be important later. The question we might be asking here is whether the “up” methyl or “down” methyl is preferred, and which of the input poses (if any) is preferred. And of course we want to see if the preferred cluster by MD still retains the binding interactions we thought were good enough to carry ligand 35 along up to this point.
Once the run is completed, again we go to the job results page, not shown here because it is so similar to what we saw with the single-pose example in Figure STMD Job results page for a single pose of an MCL1 ligand (above). Selecting the third tab (“FLOE REPORT”), there is still only a single tile for the single ligand; the results for all 6 poses have been aggregated and analyzed together for that ligand. The atom colors corresponding to the calculated B-factors will often be a lot “hotter” (more red) for multiple-pose inputs because trajectories for diverse poses are aggregated together, often giving higher per-atom fluctuations. Click on the tile to drill down into the detailed analysis, resulting in Figure Detailed results for ligand 35 (6 poses):
The results look quite different from the single-pose case although the binding interactions are mostly the same (the 2D representation shows a different orientation). There are now four major clusters resulting from the combined trajectories. As a reminder, the results are not deterministic: In you run this same MD again, in principle you could get slightly different results. Nevertheless, the results are interpreted in the context of the sampling obtained.
The table to the left of the graphic gives key information on each cluster. The blue cluster 0 is dominant, accounting for 45% of the trajectory and with the best (lowest) ensemble MMPBSA and Bintscore. Cluster 1 (green) is second largest at 36%, and has less good MMPBSA score and BintScore. Cluster 2 (orange) at only 13% abundance scores the worst compared to the others, while Cluster 3 (pink) scores second best by both MMPBSA and BintScore even though it is the smallest cluster at 5%. How do these clusters relate to the different input poses?
Scrolling down to the strip chart, shown below in Figure Strip Chart results for ligand 35 (6 poses), we see the time course over the MD trajectories for all starting poses concatenated and analyzed together. The strip chart and the table below it (table Cluster Percentage by Pose for ligand 35 (6 poses) both point to a clear grouping by pose: poses 0, 3,and 5 show predominantly cluster 0 occupancy (blue), and poses 1, 2, and 4 show predominantly cluster 1 occupancy (green).
The former poses correspond to the methyl “down” starting poses and the latter to the methyl “up” starting poses, which we can confirm in the Orion 3D page. While the short trajectories in this run (2 ns for each pose) do not allow interconversion between methyl “up” and “down” poses, it appears that the 3 poses in each category have collapsed to a single dominant cluster. How close is the cluster to any of the starting poses? This answered by the Table Cluster RMSD from Pose for ligand 35 (6 poses)
This table confirms that cluster 0 is quite close to the starting poses (0, 3, and 5) that contributed to it, though slightly closer to Pose 0. Cluster 1 is still within 2 Å of 5/6 poses, but closest to Pose 1 out of all.
We can visually confirm this by selecting the output dataset (in the “Data” tab of Orion) and then going to the “3D” tab. Under the list of structures for ligand 35, the starting poses are conformers under the molecule named simply “35”. Unfortunately the conformer number in this structure are off by 1, starting from “1”, compared to the analysis, which starts from 0! This bug will be fixed in a future release. The average and median for each cluster appear as a separate protein-ligand complex, labeled accordingly (for example “35_clus0_Avg” for the average of cluster 0). Selecting starting poses corresponding to “down” poses 0, 3, and 5 (i.e. conformers 1, 4, and 6) and displaying them with the average for cluster 0 (“35_clus0_Avg”) gives the upper panel in Figure Starting Poses and Cluster Averages for ligand 35. Selecting starting poses corresponding to “up” poses 1, 2, and 4 (i.e. conformers 2, 3, and 5) and displaying them with the average for cluster 1 (“35_clus1_Avg”) gives the lower panel. Interestingly, with the averages color by calculated B-factor it is obvious that the “down” cluster is markedly more stable in the active site than the “up” cluster, as well as being more energetically favorable by MMPBSA and BintScore.
These visually confirm what we had seen emerging from the analysis: the 6 poses collapse into a predominant methyl “up” and methyl “down” pose. Cluster 0 lies close to one of the starting poses, but Cluster 1 lies in between two of the starting poses. Cluster 0 has a more stable pose than Cluster 1, and the ensemble MMPBSA energies and BintScores also favor Cluster 0.
Understanding the results for a set of ligands¶
Finally we will look at how to visualize the results for a 11-ligand subset that spans the range of activities for the MCL1 dataset. Each of 11 ligands has 5 reasonable input poses from docking. The whole subset will be run in the same job in the “Short Trajectory MD with Analysis” floe. which will consist of 11 ligands * 6 poses each = 66 MD run of 2ns each. Selecting the output dataset in the “Data” tab and moving to the “Analyze” tab, the results for the entire dataset can be viewed at once as in Figure Analyze page for MCL1 dataset:
There are a lot of results showing in this page, encompassing both numerical and 3D information. The 3D info is brought in by selecting Analyze with 3D under the Layout pull-down menu at the top right. The axes of the scatterplot were selected to display the experimental deltaG (included as SD tag ‘r_user_dG.exp’ on the input ligands) on the x axis and the Boltzmann-weighted ensemble MMPBSA value on the y axis. In the 3D visualizer, select ligand 49 and unroll the list of associated structures. The point in the scatter plot corresponding to ligand 49 and the corresponding line in the spreadsheet is highlighted. In the 3D window, the 5 initial input poses for ligand 49 are under Molecule “49” and display in in gold if selected. Turn on the protein-ligand average structures for Clusters 0 and 1, which will be colored by B-factor as before. This way we can compare the poses to the representative average for each cluster, helping us to evaluate and prioritize that ligand. To call up the detailed MD analysis once again, go to the spreadsheet row for ligand 49, and under the column titled Floe_report_URL click on the little square to open up another tab in your browser with the same detailed analysis floe report as we saw above.
There is a lot of information to look at in the results from the Short Trajectory MD with Analysis floe, but this should get you started. We emphasize that a lot of the analyses can only be interpreted qualitatively at this stage, but nevertheless we feel that the sampling of both protein and ligand configurations at physiological temperatures in the context of explicit water solvation can help validate the initial input pose(s).
BintScore is a new knowledge-based scoring function from OpenEye for scoring protein-ligand interactions.