[ Receptor Considerations | Ligand Considerations | Chirality Considerations | Protonation Considerations | Internal Hydrogen Bonds | Pyramidal Sampling | Membrane Exposed Docking | Waters | Scoring | Ligand Strain ]

Additional Resources: Tutorial: Dock Biotin to Streptavidin Receptor | Tutorial: Dock a Ricin Inhibitor | Command Line Docs | Interactive Docking Tutorial Video | Re-dock Tutorial Video

For accurate ligand docking, the goal is to have an adequate three-dimensional model of the receptor pocket you are planning to dock ligands to. If this is the case then ICM docking has been shown to be very accurate in a number of independent assessments.

However, there are a number of pitfalls which need to be overcome to achieve accurate ligand docking. The pitfalls are that your model is not accurate overall, does not reflect the induced fit, or alternative conformations of the receptor binding pocket are missed. MolSoft has a number of in built tools to overcome some of these challenges

Some key points about ICM Ligand Docking:

• An average docking time is 0.1 seconds to 30 seconds per ligand per processor depending on the size of the ligand. The time per ligand was chosen to be the smallest possible to allow screening of very large data sets. To increase the time spent per ligand, change the thoroughness/effort parameter. For even faster docking, ideal for libraries more than a billion chemicals see our GPU enhanced methods such as RIDGE, GigaScreen and CombiRIDGE.
• ICM docking is the most accurate predictive tool of the binding geometry today. ICM docking has consistently ranked first place compared to other leading docking software in terms of accuracy. ICM has been successful in many drug design applications by scientists in academia and industry.
• ICM ligand docking procedure performs docking of the fully flexible small-molecule ligand to a known receptor 3D structure.
• The goal of the flexible docking calculation is prediction of correct binding geometry for each binder.
• ICM stochastic global optimization algorithm attempts to find the global minimum of the energy function that includes five grid potentials describing interaction of the flexible ligand with the receptor and internal conformational energy of the ligand. During the docking process a stack of alternative low energy conformations is saved.

Please click here to read more about ICM Docking.

12.1.1 Receptor Considerations

If you have only a single PDB entry for your receptor, convert the protein to an ICM object, delete water molecules and irrelevant chains. However, if you have a choice between several templates, take the following into account:

• X-ray structure is preferable to an NMR structure.
• High resolution X-ray structure ( less than 2.1A ) is much better than, say 2.5A .
• Watch out for high-B-factor regions and avoid them; sometimes crystallographers deposit fantasy coordinates with high-B-factors.
• Check the pocket for correct placement of polar hydrogens and choose correct form of histidine.
• A bound conformation of the receptor is preferable, however if you use an apo-model, an NMR structure or a model by homology, the side-chains in a pocket may be incorrect. Frequently they stick out and prevent a ligand from binding. Those stubborn side-chains can be 'tamed', (i) manually; (ii) by a side chain simulation with elevated surfaceTension; or (iii) using a method such as Dual Alanine Scanning and Refinement ( SCARE).
• A model by homology can be built with the build model command (see molecular modeling section of this manual) and used for docking.

12.1.2 Ligand Considerations

Usually a good place to start is to try to dock the known ligand(s) to the receptor model. You may also want to dock a library of compounds in order to identify lead candidates. In this case the main pitfalls are that the library is too restricted, molecules are not chemically feasible or not drug-like. For peptide docking please use the peptide docking protocol.

There is no need to convert the ligands to 3D, this is done "on-the-fly" during the docking process.

NOTE: If you are docking a ligand directly from the PDB please check the bond types and formal charges of the ligand. This is discussed in the section entitled Converting a Chemical from the PDB

12.1.3 Ligand Chirality Considerations

To sample linear chiral centers during docking you can check the option "Sample Racemic Centers" in Docking/Preferences/General. To sample chiral centers in rings it is best to generate all 2D enantiomers using Chemistry/Generate Stereoisomers. For most linear cases 'sample racemic centers' is acceptable. For most rigorous treatment one should pre-generate all stereoisomers (Chemistry/Generate Stereoisomers). If you do not select anything you get one (essentially arbitrary) stereoisomer docked.

12.1.4 Ligand Protonation Consideration

The protonation state can be set in the Docking/Preferences/General dialog box. Choose the drop down option in the field labeled Charge Groups. To use MolSoft's pKa model choose the option auto, which uses built-in prediction of Ka/Kb to charge and protonate/deprotonate appropriate groups. If you do not want to use the pKa model then charges of some ionizable basic groups can be set, it currently accepts the following values: NH2, NH, NT for primary secondary and tertiary aliphatic amines. There are also options for imidazole and amidine.The none option still charges acids unless neutralAcids flag is set to no in the docking table file .dtb or .tab.

To determine protonation at a range of pH values you can use the options in Chemistry/Set Formal Charges/ and choose the option Auto using Pka Model.

12.1.5 Internal Hydrogen Bonds

Internal hydrogen bonds are taken into account during docking by default. You can check in the docking_project_name.dtb table file to make sure that l_internalHB field is 'yes'.

12.1.6 Pyramidal Inversion Sampling

If the parameter l_samplePyramid is set to yes in the Docking Project .dtb file:

• During Monte Carlo sampling, the algorithm includes steps to invert the pyramid geometry at nitrogen atoms.
• This specifically targets uncharged aliphatic amines and sulfonamides.

12.1.7 Membrane-Exposed Docking

(l_membraneScore)

If you are docking to a membrane-exposed site, set the l_membraneScore flag in the docking project .dtb file to yes.

Score Term	Modification Detail
SolEl (Solvation Electrostatic)	Calculated with all Born radii set to a fixed high value, assuming high-dielectric medium (water) is distant.
Donor/Acceptor Desolvation	Calculated excluding receptor desolvation, as receptor atoms are presumed to be already buried in the membrane.
Hydrophobic Term	Excludes buried receptor surface area from the calculation.

12.1.8 Handling Waters in the Binding Pocket

Before docking, you will need to convert your PDB structure into an ICM object. During this step, you can choose how water molecules are handled. The options are to keep all waters, delete all waters, or remove only loose waters.

Loose waters are defined as water molecules that do not make any hydrogen bonds with the protein or a co-crystallized ligand. Tight waters are those that make three or more hydrogen bonds and are often more structurally stable.

Before docking, crystallographic water molecules in or near the binding pocket should be reviewed carefully. In most cases, bulk or weakly ordered waters should be deleted, as docking methods implicitly account for solvent effects and these waters may block ligand placement.

Waters that may be safely removed include:

• Waters located at the pocket entrance or solvent-exposed regions
• Waters with high B-factors or partial occupancy
• Waters not involved in stable hydrogen-bond networks
• Waters displaced by known ligands in co-crystal structures

Waters that may be worth retaining include:

• Conserved waters observed across multiple crystal structures
• Waters that bridge key protein–ligand interactions
• Waters forming stable hydrogen-bond networks with low B-factors
• Waters deeply buried in the pocket with limited exchange

When waters are retained, they should be treated consistently during docking and scoring. If the role of a water molecule is uncertain, it can be useful to run docking with and without selected waters and compare docking scores and poses.

12.1.9 Scoring

After docking and screening you can make a hitlist of the results which reports the following scores. The goal of scoring in virtual ligand screening is to ensure maximal separation between binders and non-binders , and not to rank a small number of binders according to their binding energies. The scores can be linearly related to binding energy estimates, but the transformation parameters need to be calculated from several reference points (see the Learn and Predict chapter).

• Physics-based Score. The Score is the ICM VLS Score which is a GBSA/MM-type scoring function augmented with a directional hydrogen bonding term (see Neves et al 2012). A ICM score -of 32 and lower are generally considered good scores - but it depends on the receptor (e.g. exposed pockets or pockets with metal ions may have higher scores than -32).

  An important parameter in a VLS run is the score threshold. A docked conformation for a given ligand is stored by the ICM VLS procedure only if its binding score is below this threshold.

  The threshold can be chosen in several ways:

  1. If a co-crystal structure is available, remove the ligand, re-dock it, and use the resulting score as a reference.
  2. If known ligands exist, dock them and examine their scores. A threshold set slightly above the typical scores of these ligands is usually a good starting point.
  3. If no ligands are known, run a short pre-simulation using approximately 1000 compounds from the screening database. Based on the score distribution, set the threshold to retain roughly 1 percent of the ligands.

  You can edit the Score threshold in the Docking/Preferences dialog box or edit the .dtb file.

• RTCNN Score is a Neural Network Score. RTCNN is a Radial Convolutional Neural Net including layers that do Topological (chemical graph) convolutions and 3D Radial convolutions. The RTCNN score is different from the regular ICM Score (see above) - it does not use any molecular mechanics or physical energy terms. Instead, the score is trained to recognize native-like complexes versus decoys directly, based only on geometries of putative complexes. So, for example, if internally the neural net recognizes a hydrogen bod, it is because it is observed during training that the configurations of atoms that we call hydrogen bonds are present to a greater extend in native-like structures. This is the same for other favorable or unfavourable interactions - it is all learnt from examples, not from any input terms.

  The RTCNN score performs extremely well on our benchmarks and the CASF benchmark (Su et al J.Chem.Inf.Model 2019). Please find details about this benchmark here: RTCNN_performance.pdf and read more about RTCNN Score here.

  When analyzing a hitlist from virtual screening we would recommend to plot Score vs RTCNN and prioritize those in the quadrant that have low Score and low RTCNN score. One key difference you may observe between RTCNN score and the regular Score is that the RTCNN score has a lower penalty for small induced fit ligand-receptor clashes. The lower the RTCNN score the better the predicted ligand-receptor interaction.

• mfScore is the statistical potential of mean force which was derived according to the published methodology (see Muegge J.Med.Chem 2006)

Note about comparing Neural Network Score (RTCNN) and Physics Based (Score): In a library screening scenario, typically only a very small fraction (<<1%) of compounds are true actives. As a consequence, and due to the inherent noise in scoring, there will be many false positives for any score. They will be almost certainly outnumbering true positives in the VLS hitlist, even though in a less demanding setting of, say, 50/50 or 10/90 active/inactive compound mix the score might have nicely separate them. In this context, the most efficient use of two scoring functions is to look for consensus - because if one score is good but the other one is poor, it is a lot likelier that the molecule is a false positive of the first one, rather than a false negative of the second one. In our evaluations, enrichments that are afforded by the two scores individually are comparable (although depending on the specific target one might perform somewhat better or worse than the other). Therefore, unless there is any data such as scores for known binders, we recommend to apply cutoffs of comparable stringency to both. Lastly, one should also pay some attention to absolute values of scores. Unless screening for low-affinity fragments, compounds with scores weaker than -25. generally are not worth consideration. For most pockets, numbers should really be better than -30. , with large deep pockets often going beyond -40.

Note about why Scores may differ between docking runs: It is important to understand that unless there is a single stable binding pose for a given ligand, docking will produce a variety of poses. This is consistent with physics - only reasonably high-affinity ligands in reality would stay bound in a single configuration, poor binders will drift around through a variety of weakly-bound configurations. One should look first at the pose convergence and only after convergence is achieved one would look at score convergence. When there isn't a strong single most favorable pose, top docked pose may vary from run to run. Many poses may need to be re-scored to get consistent best-scoring poses because one or even three poses may not be enough to always include top-scoring pose according to VLS score.

12.1.10 Ligand Strain

Docking using ICM-Pro + VLS reports global strain of the ligand in the hitlist (see Eintl). The lower the Eintl value is the lower the strain - we suggest cutoff between 5 and 10, depending on how strict or permissive you want to be. More permissive settings may be justified by the noise in strain evaluation and lack of full receptor relaxation (induced fit would potentially let the ligand relax further). Also larger ligands may generally give larger strains. This value represents the energy difference between the relaxed (unbound) ligand and its conformation in the docked (bound) state - so lower values are better. It also contributes to the physics-based docking score (the Physics-Based Score value).

To calculate Strain, an ensemble of low-energy ligand conformers is generated in solution using the MMFF94s force field. The lowest-energy conformer is used as the unstrained reference. Strain is then defined as the energy difference between this reference conformation and the docked pose, with both energies computed using the same force field.

Analysis of specific torsion can be performed via torsion profile calculation at force-field level with MolMechanics/Torsion Scan, or via Crystallography Open Database search with appropriate substructure (Search Tab, select COD search, draw substructure) and statistics calculation with Chemistry/Torsion analysis.