Figure 1.
Sampling of the HIV-1 protease binding pocket by ritonavir and a conserved water.
Top: One key water molecule hydrogen bonds (black lines) with both HIV-1 PR flexible flaps and protease inhibitor Ritonavir. Middle: Standard docking begins with translation of the inhibitor from its centroid, by up to 5 Å (green sphere). Protein centric water docking also includes up to 4 Å translation of water (red sphere). Bottom: Grey mesh indicates sampling space covered after ligand rotation. Image was prepared using Pymol. The structure shown was downloaded from the protein databank (PDB ID: 1HXW).
Table 1.
Comparison of protocols for the 2 benchmark studies presented in this paper.
Figure 2.
RMSD and rank comparisons between standard and protein-centric water docking of HIV-1 PR/PI.
Left panel: RMSD of top scoring Rosetta model. 69 models fall below the diagonal (improved RMSDs) while 30 lie above it. Red dashed lines represent the 2 Å RMSD metric for successful docking. Predictions in the lower-right quadrant turn from failures to successes up on water docking. Upper-left quadrant contains predictions that succeeded without water docking and fail with water docking. Right panel: rank of the lowest scoring Rosetta model with RMSD under 2 Å. Where multiple HIV-1 cross-docking predictions achieved the same rank with and without water docking, these points are replaced with text indicating the number of overlapping points.
Table 2.
Summary statistics describing the CSAR dataset.
Figure 3.
RMSD and rank comparisons between standard and ligand-centric tight water docking of CSAR dataset.
Left panel: RMSD of top scoring Rosetta model. 106 models fall below the diagonal (improved RMSDs) while 82 lie above it. Red dashed lines represent the 2 Å RMSD metric for successful docking. Predictions in the lower-right quadrant turn from failures to successes up on water docking. Upper-left quadrant contains predictions that succeeded without water docking and fail with water docking. Right panel: rank of the lowest scoring Rosetta model with RMSD under 2 Å. Where multiple CSAR docking predictions achieved the same rank with and without water docking, these points are replaced with text indicating the number of overlapping points.
Figure 4.
RMSD and rank comparisons between standard and ligand-centric loose water docking of HIV-1 PR/PI.
Left panel: RMSD of top scoring Rosetta model. 159 models fall below the diagonal (improved RMSDs) while 129 are above it. Red dashed lines represent the 2 Å RMSD metric for successful docking. Predictions in the lower-right quadrant turn from failures to successes up on water docking. Upper-left quadrant contains predictions that succeeded without water docking and fail with water docking. Right panel: rank of the lowest scoring Rosetta model with RMSD under 2 Å. Where multiple CSAR docking predictions achieved the same rank with and without water docking, these points are replaced with text indicating the number of overlapping points.
Table 3.
Mean values for top models from Rosetta CSAR docking results.
Figure 5.
RMSD vs Rosetta interface score for CSAR predictions.
Each plot contains the top 100 Rosetta models by total score for both standard (red) and water (blue) docking for particular CSAR datapoint. Each plot is identified by its CSAR label (e.g ‘set1_91’). CSAR labels are followed by rank before and after water docking. Ranks of ‘n/a’ indicate that no model below 2 Å RMSD was sampled by Rosetta. Each set of 3 plots represent the largest rank changes seen in that category. Successes are defined as ranks that decrease and failures as ranks that increase.
Figure 6.
Probability of changes in CSAR docking success upon replication of docking study.
Success is measured as whether the top scoring model by interface score has a ligand pose within 2.0 Å RMSD of the native pose. As sampling size increases, the probability that resampling with would change the outcome of docking decreases. Equations for the best-fit lines are available in Table S4 in File S1.
Table 4.
P-values calculated using a one-tailed binomial distribution.
Figure 7.
Docking results for CSAR complex ‘set1_120’.
Top left: Experimental structure of PDB: 1IUP, coded as CSAR datapoint ‘set1_120’. Waters (Oxygen only) are shown as red spheres. Black lines represent polar contacts predicted by PyMOL. Top right and bottom row: native ligand (lines) and waters (spheres) are shown in grey for comparison. Docked waters are shown as sticks (note that Rosetta adds hydrogens). Docked ligands are shown in cyan, yellow, and green. For each study the models were sorted by total score, then interface energy. The first model with RMSD <2.0 Å is depicted. Its position in the sorted list (rank) and its RMSD to native are shown.
Figure 8.
Docking results for CSAR complex set1_181.
Top: experimental structure with ligand in blue, water as a red sphere, and polar contacts as black dashed lines. 22 polar contacts are predicted by PyMOL, 4 of which contact the water molecule. Middle: Top scoring model from docking without water. Native ligand and water in grey, Rosetta model in cyan. PyMOL predicts 16 polar contacts. Bottom: Lowest RMSD model from docking with loose waters. Rosetta model shown in green. No model within the top 100 by total energy score has RMSD <2.0 Å (hence rank is ‘n/a’). Shown is the lowest RMSD structure. PyMOL predicts 11 polar contacts (1 with water).
Figure 9.
Relation between binding pocket crowdedness and the improvements in CSAR model ranking when water is docked.
Crowdedness is calculated as the number of ligand/protein contacts divided by the total number of ligand atoms. Datapoints with rank changes between −10 and 10 were omitted to focus on data where water docking makes a large impact on results. Note that below a crowdedness threshold of 2, addition of water rarely worsens rank.