It is common in medicinal chemistry literature to use rigid-target docking methods to explore the binding positions of novel synthesized ligands inside a known target. In order to carry out these tasks, an X-ray crystallographic structure of the target, which is often in presence of a previously reported ligand, is needed. When several crystallographic structures are present, one of them is selected. The next task is to demonstrate that the selected protein structure is reliable to accomplish docking experiments. A common test to justify the selection is to try to reproduce the orientation of the ligand that is inside the selected crystallographic structure: self-docking according to our definition in the Materials and Methods section.
In our experiment, each ligand from Table 1
was self-docked inside its own binding site. The top 10 poses (according to the scoring function energies) were selected to further structural analysis by calculating the root mean square deviation (RMSD) of each pose against its conformation in the crystal structure. We used three different RMSD classifications for docking solutions: (a) good solution when RMSD ≤ 2.0 Å [8
], (b) acceptable solutions when RMSD is between 2.0 and 3.0 Å, and (c) bad solutions when RMSD ≥ 3.0 Å. For one ligand-protein pair, when a pose is classified as a good solution, this means that the scoring function reproduced the crystallographic binding orientation. To verify that RMSD classifications were adequate, compounds a1
) were docked in their respective targets (MAO-B, thrombin and B-RAF, respectively). Selected docking solutions are represented in Figure 1
(three of the ten poses are displayed) including good, acceptable and bad solutions according to above mentioned RMSD criteria. It is clear that an RMSD < 2.0 Å corresponds to good docking solutions. On the other hand, docking solutions with RMSD between 2.0 and 3.0 Å deviate from the position of the reference, but they keep the desired orientation. Finally, docking solutions with RMSD > 3.0 Å are completely wrong.
The self-docking experiments were performed for the 30 different ligands from Table 1
into their own binding sites. Each docking was performed in triplicate to have a most representative exploration of the docking solutions. For each protein-ligand complex, the ten better obtained poses according to scoring functions (in Glide HTVS, Glide SP, Glide XP, and AutoDock methods) were analyzed by using the RMSD with respect to the reference complex. The best solution s2
was defined for each complex (and for each docking method) as the docking solution with RMSD < 2.0 Å at the best scoring position. At the same time, the best solution s3
was defined for each complex (and for each docking method) as the docking solution with RMSD < 3.0 Å at the best scoring position. s2
only accounts for good solutions, while s3
accounts for good or acceptable solutions. For each s2
values, the specific scoring position between 1 and 10 was annotated to construct a plot of the percentage of recurrence of the best solution at each scoring position for the targets MAO-B, thrombin and B-RAF; Figure 2
shows these plots for the used docking methods. For a protein-ligand complex, when none of the scoring positions contain a solution that complies with s2
, then it is annotated as a bad solution (BS), which indicates that there are no good poses for this case.
Self-docking results in Figure 2
allow the analysis of the recurrence of the best solution in the top first scoring position for different targets by using different docking methods. At first, it is clear that none of the docking methods obtained the best solution (under s2
criteria) at top first scoring position for all the ligands in interaction with the three targets under study. This means that it is not completely reliable to select the top first scoring solution in the simplest docking experiment: the self-docking.
If a rigorous criterion is selected, where the best solution s2
is defined, the best solution in the top 1st scoring position have different recurrences according to the target and selected docking method. With regard to this, it is possible to extract some conclusions from the bottom part of the Figure 2
The recurrence of the top first scoring position was higher for B-RAF inhibitors by using all the docking methods.
The recurrence of the top first scoring position was higher when the docking methods Glide SP and XP were used.
In the self-docking study, Autodock was the worst method to get the best solution at the top first scoring position.
If the less rigorous criterion is selected, where the best solution s3
is defined (Figure 2
top), there are no big changes in the analysis; except for the ability of AutoDock for detecting the best solutions at the top first scoring position for B-RAF inhibitors. This means that AutoDock yielded acceptable solutions for this specific target.
Other important point is the recurrence of BSs when self-docking is performed. When the best solution s2
is defined (Figure 2
bottom), there are many BSs for different combinations of docking method with a specific target. For instance, there were many BSs when Glide XP was applied to study MAO-B inhibitors. This means that there were many MAO-B X-ray crystallographic structures that give no good solutions when the simplest self-docking was applied. It should be recalled that s2
excludes acceptable solutions; therefore, when the best solution s3
is defined for self-docking (Figure 2
top), there are less BSs.
Several lessons are provided from the self-docking experiment. The reliability on the top first scoring position as the best solution depends on the target under study and the docking method. In our study, B-RAF seems to be the more adequate target for self-docking; meanwhile, MAO-B seems to have difficulties during self-docking. On the other hand, Glide SP and XP seem to be the most adequate methods for self-docking. Noteworthy, it is no gain when the more precise XP version is used; in fact, SP is better for MAO-B and B-RAF inhibitors.
When the best solution is not found at the top first scoring position, it is possible to find it at the second, third or other position, but, in several cases, there are no good solutions. It is not clear why several crystallographic structures fail during self-docking, and the consequences of that for a cross-docking experiment. However, it sounds logical and prudent not to take a structure for performing cross-docking if we know in advance that it failed when self-docking was done.
The most common practice in docking experiments is the inclusion of novel ligands inside a binding site occupied by a previous reported molecule: cross-docking according to our definition in Materials and Methods section. To test if cross-docking deteriorates the recurrence of the best solution in the top first scoring position for different targets by using different docking methods, we used the three compounds that yielded the best solutions in the self-docking experiments applied to each target for performing additional cross-docking experiments. The selected compounds were b1, h1, j1 (MAO-B inhibitors), b2, g2, h2 (thrombin inhibitors), b3, f3, and g3 (B-RAF inhibitors). They were cross-docked against the remaining nine receptor crystal structures, and the obtained poses were compared with the conformation and orientation of the compound in the original crystal structure by means of RMSD measures.
Each cross-docking was performed in triplicate to have a most representative exploration of the docking solutions. For each protein-ligand complex, the ten better obtained poses according to scoring functions (in Glide HTVS, Glide SP, Glide XP and AutoDock methods) were analyzed by using the RMSD with respect to the reference complex. The best solutions s2 were defined for each complex (and for each docking method) as in previous self-docking.
shows the ability of the docking methods to find a good solution in cross-docking experiments inside the binding sites of the crystallographic structures of the studied targets. Dark gray squares represent self-docking, light gray squares represent instances with at least one good solution, and black squares represent instances with only BSs. The analysis reflects some interesting points:
There are compounds that are well oriented in most of the crystal structures, but there are others that are conflictive. For instance, compound h2 had good orientations in almost all the crystallographic structures of thrombin, but b2 and g2 cross-docking experiments failed in most of the instances.
The selection of the docking method was crucial in some cases; for instance, h1 had good orientations in almost all the crystallographic structures of MAO-B when AutoDock was used, but almost all cross-docking experiments for this compound failed when Glide HTVS, SP and XP were used.
B-RAF was a conflictive target for cross-docking. All its inhibitors were bad oriented in almost all crystallographic structures by using all docking methods. Particularly, compound b3 failed in all the instances.
The recurrence of the best solution at the first scoring position was also analyzed for cross-docking experiments. For this, the specific scoring positions between 1 and 10 (considering s2
criteria) were annotated to construct a plot of the percentage of recurrence of the best solution at each scoring position for the targets MAO-B, thrombin, and B-RAF; Figure 4
shows these plots for the used docking methods. BSs were also annotated for the cases that have no good poses.
Once again, it is clear that none of the docking methods obtained the best solution during cross-docking experiments (under s2
criteria) at top first scoring position for all the ligands in interaction with the three targets under study. This means that it is not completely reliable to select the top first scoring solution when cross-docking is achieved without an additional analysis. If a rigorous criterion is selected, when the best solution s2
is defined, the best solution at the top 1st scoring position have different recurrences according to the compound, target and selected docking method. With regard to this, it is possible to conclude from the bottom part of the Figure 4
that the recurrence of the top first scoring position was higher for a few combinations of compound and docking method. For instance, the MAO-B inhibitors b1
had around 65% and 80% of recurrences of the top first scoring position when Glide HTVS and AutoDock are used, respectively. On the other hand, the thrombin inhibitors g2
had around 50% of recurrences of the top first scoring position when Glide XP and Glide HTVS are used, respectively. The remaining cases had small recurrences of the top first scoring position. In general, it is possible to see that cross-docking was characterized by a poor recurrence of the top first scoring position and many BSs. It suggests that it is not reliable to select the top first scoring position as the best solution when cross-docking is performed.
If the less rigorous criterion is selected, where the best solution s3
is defined (Figure 4
top), there are subtle changes in the analysis. More combinations of compound and docking method yielded the best solution at the top first scoring position; however, it is necessary to remember that this improvement is due to the addition of acceptable cases with no optimal orientation in many cases.
Another important point is the recurrence of BSs when cross-docking is performed. When the best solution s2
is defined (Figure 4
bottom), there are many BSs for different combinations of compounds with docking methods. For instance, there were many BSs when all the docking methods were applied to study the three B-RAF inhibitors. This suggests that a special care should be taken when cross-docking is performed for novel B-RAF inhibitors. On the other hand, the number of BSs is low for the thrombin inhibitor h2
by using all the docking methods. This suggests that there are compounds more likely to be well oriented inside different conformations of the binding site. The inclusion of acceptable solutions (s3
criterion) reduces the number of BSs (Figure 4
top), at the expense of lower quality of the orientations.
Several lessons are provided from the cross-docking experiment. We observed for self-docking that the reliability on the top first scoring position as the best solution depends on the target under study and the docking method, but further experiments show that it is not possible to rely on the top first scoring when cross-docking is done. Previously, we observed that B-RAF seemed to be the more adequate target for self-docking; but further experiments show that B-RAF is the less adequate target when cross-docking was done. MAO-B had difficulties during self-docking, but it was the target with a better performance during cross-docking.
The main reason of the success rate reduction for cross-docking method is that, even though the structures correspond to the same protein (MAO-B, thrombin or B-RAF in our work), the global shape of the binding site varies for different crystal structures of the same protein due to different orientations of the side chains of several residues. To observe the principal differences between different structures of the same protein, the crystallographic MAO-B, thrombin and B-RAF structures were aligned, and the RMSD values for comparison between residues were calculated to study how different are the binding sites (results are represented in the Electronic Supplementary Material, Figures S1–S3
). In general, RMSD values were below 0.5 Å for MAO-B and thrombin structures, and were above 1.0 or 1.5 Å for B-RAF in most of the structural alignments.
We selected the residues that displayed different conformations among the studied crystallographic structures due to the presence of different ligands and compared their RMSD values. In MAO-B crystal structures, the residues with higher RMSD values were F103, I199 and Y326 (Supplementary Material, Figures S1B–D
); however, values below 2 Å were observed for I199 and Y326 because a small displacement of side chains of these residues in the structure with PDB code 1S2Y. Additionally, values below 3 Å were observed for F103 because higher displacements of its side chain in the structures with PDB codes 1S2Y and 4A79. These observations explain the bad performance of the ligand h1
and the protein structure with PDB code 1S2Y during the rigid cross-docking.
In thrombin crystal structures, the residues with higher RMSD values were W86 and E232 (Supplementary Material, Figures S2B–D
); however, values below 1.5 Å were observed for W86 because a small displacement of its side chain in the structure with PDB code 4LXB. Additionally, values below 3 Å were observed for E232 because higher displacements of its side chain in the structures with PDB codes 3SV2, 1AHT, and 4UDW. These observations explain the bad performance of the protein structure with PDB code 1AHT during the cross-docking of ligands b2
. Despite the displacement of the side chain of E232 in the structure with code 1AHT, the ligand in this structure (h2
) had good performance when it is cross-docked in other structures, indicating that displacements do not always negatively affect cross-docking.
Finally, in B-RAF crystal structures we found the largest numbers of residues with high RMSD values: I463, K483, E501, L505, and F583 (Supplementary Material, Figures S3B–D
). Values below 2.5 Å were observed for K483 and L505 due to displacements of their side chains when the crystallographic structures were compared. High RMSD values were also observed for I463 because its longer displacement in the structures with PDB codes 3D4Q and 3PRF. This residue is located at the glycine-rich loop of B-RAF, which present backbone conformational changes for adapting to ATP or inhibitor binding [14
]. The largest displacements in B-RAF were observed for E501 (backbone and side-chain) and F583 (side-chain) with RMSD values above 4 Å for their side chains when comparing different crystallographic structures. E501 is the conserved glutamate at the αC-helix; this helix has a high mobility involved in controlling the dynamic equilibrium between active and inactive functional forms of protein kinases. F583 is part of the hydrophobic wall of the ATP binding site which adapts side chain orientation to different inhibitors. These observations reveal that B-RAF binding site changes more than the ones for MAO-B and thrombin; this explains the bad results of B-RAF in cross-docking.
Additional information was obtained when various characteristics of the protein binding sites and ligands were analyzed for each MAO-B, thrombin, and B-RAF PDB structures. Protein binding site volumes (BSV), volume depth (VD) values, and averaged VDs were obtained for the protein binding sites by using the software POCASA [47
]. VD is the volume depth value, which is determined by summing the depth of all pocket points, where the depth of every pocket point is defined as the shortest distance from pocket point to probe surface. Average VD is the average of the depth of every pocket points. Binding site volumes were calculated considering a probe radius of the probe sphere of 1 Å, single point flag of 12, protein depth flag of 18, and grid size of 1 Å [47
]. Number of rotatable bonds and molecular weights of the inhibitors inside each target were calculated by using Molinspiration web (http://www.molinspiration.com/cgi-bin/properties
). The calculated characteristics for protein binding sites and ligands are reported in Table 2
MAO-B has a small binding site (low BSV) in a deeper zone of the protein (high VD and averaged VD). There are big differences in BSVs among different PBD structures (BSDs are 97 and 289 Å3
for PDB structures 4A79 and 1OJ9, respectively), which could lead to difficulties in cross-docking experiments. However, analysis of the Figure 3
does not give evidence that the differences in MAO-B BSVs caused the failures in cross-docking. Major problems were found for compound h1
(from PDB 1S2Y), which has the lower molecular weight and only two rotatable bonds. Compounds b1
had major problems in the PDB structure where h1
is bound: 1S2Y. While b1
are bigger ligands with more rotatable bonds, they have good solutions in other MAO-B structures with smaller BSVs (for instance, in PDB 2VZ2). Therefore, we consider that the problem in cross-docking for the studied MAO-B inhibitors is mainly due to orientations of the residue side chains in the binding site (as we observed previously in Figure S1 of the Supplementary Material
), instead of BSVs or ligand characteristics.
Thrombin has a bigger binding site (BSV above 280 Å3 in most of PDB structures) at the surface of the protein (low VD and averaged VD). BSVs are similar, being between 280 and 370 Å3 in the majority of the PDB structures, with the only exception of the PDB structure 4UDW. Thrombin inhibitors have many rotatable bonds, with the exception of compound h2. It is just this compound the best one in cross-docking experiments, while compounds b2 and g2 (with 10 and six rotatable bonds respectively) fail in most of cross-docking experiments. Therefore, we consider that the problem in cross-docking for the studied thrombin inhibitors is mainly due to the number of rotatable bonds of the ligands, instead of different BSVs or different orientations of the residue side chains in the thrombin binding site.
B-RAF has three size ranges for the binding site. The smaller range has BSVs between 180 and 220 Å3
(PDB structures 5CSW, 4YHT, 3D4Q, 2FB8, and 3PRF), the intermediate range has BSVs between 310 and 360 Å3
(PDB structures 4XV9, 4E26, and 3C4C), and the bigger range has BSVs around 440 Å3
(PDB structures 5CSX and 4KSP). VD and averaged VD values indicate that B-RAF binding site is not too deep from the surface. The big differences in BSVs among different PBD structures could explain the observed difficulties in B-RAF cross-docking experiments. B-RAF inhibitor b3
has many rotatable bonds and was bound to one of the biggest binding sites (PDB: 5CSX, BSV = 438 Å3
). It had no solutions when it was cross-docked in the other binding sites with lower or similar BSVs. Compounds f3
have less rotatable bonds and were bound to smaller size binding sites. In general, they also fail in most of the cross-docking experiments, but they have solutions mainly when cross-docking was done inside other PDB structures in the same BSV range, such as 4YHT, 3D4Q, 2FB8, and 3PRF. This analysis suggests that the number of rotatable bonds and the different BSVs are the key effects that lead to problems in B-RAF cross-docking experiments. At the same time, different orientations of the residue side chains in the binding site (as we observed previously in Figure S3 of the Supplementary Material
) cause the differences in B-RAF BSVs.