A docking score is an ordering device. It is reasonably good at putting plausible binders above implausible ones in the same pocket, and it is poor at saying how tightly any of them binds. That difference decides how a screen should be set up: you check whether the ordering is trustworthy for your target, you take the top of the list as a shortlist for experiments, and you never carry the number itself into a project slide as an affinity.
This article is about the practical consequences of using docking as a ranking method, assuming you already know that the energy units are not a binding free energy.
Check the ranking on compounds you already know
Before a prospective screen, run a retrospective one. Take the actives you have for this target, or published actives, mix them with a much larger set of property-matched inactives or decoys, dock everything with the same protocol, and look at where the actives land.
A worked example with round numbers. 50 known actives, 2,000 decoys, so the actives are 2.4 % of the library. If the top 1 % of the ranked list (21 compounds) contains 9 actives, that is 43 % of the top slice versus 2.4 % at random, an enrichment factor of about 18. That protocol is useful. If the top 1 % contains one active, the ranking is close to random for this pocket, and no cut you choose afterwards will fix it.
Two cautions on this exercise. Decoys that differ from the actives in molecular weight or logP make the task artificially easy, so match the properties. And an enrichment measured on one chemotype of actives does not transfer to a different chemotype.
Pick the cut by how many experiments you can afford
There is no score threshold that means "binder". What exists is a budget. If you can test 40 compounds, the question is which 40 the ranking recommends, with two adjustments:
- Spread the chemistry. Forty compounds from one cluster test one hypothesis. Cluster the top of the list and take representatives, so a failure tells you which chemotype failed rather than only that the screen failed.
- Include a reality check. Put two or three compounds from the middle of the ranking in the test set. If they perform as well as the top, you have learned that your ranking carried no information, which is worth knowing early.
Where the ranking breaks down
The ordering is least reliable in four situations, and they are common:
| Situation | Why ranking suffers |
|---|---|
| Compounds of very different size | Scores grow with the number of contacts, so large molecules rank high for the wrong reason |
| Strongly charged or highly polar ligands | Desolvation is the hardest term to get right and dominates these cases |
| Pockets with structured water | A water-mediated contact is scored as an empty gap |
| Flexible or induced-fit sites | One rigid receptor conformation cannot rank ligands that each reshape the pocket |
Size is the one to correct routinely, because it is easy: compare scores within narrow molecular-weight bands, or look at the score divided by the number of heavy atoms as well as the raw value. A ranking that is only a molecular-weight ranking is a common and avoidable outcome.
Consensus scoring helps a little, honestly
Scoring the same poses with two or three different functions and keeping the compounds that all of them rank highly does improve precision somewhat, at the price of recall: you find fewer actives but a higher proportion of your shortlist binds. That trade is usually worth it when experiments are expensive and the library is large. It is not worth presenting as a validation, because agreement between models is not evidence of correctness; three functions sharing the same blind spot agree confidently.
Look at the pose, not only at the number
The cheapest quality control is a chemist looking at the top twenty poses for fifteen minutes. The questions are chemical, not computational: is the hydrogen bond network sensible, is a charged group buried without a partner, is the ligand in a strained conformation, does the pose explain the known structure-activity relationship of the series. A score that comes from a pose a chemist rejects is a number with no support, and this review catches more bad hits than any rescoring step.
What to report
A defensible virtual screening report contains the receptor structure and its origin, the preparation decisions (protonation, tautomers, retained waters, the box), the software and version, the retrospective enrichment with the composition of the decoy set, the ranked list, the chosen cut with its reason, and the pose review notes. Numbers without those are not reproducible, and a screen that cannot be reproduced cannot be improved either.
Limits of the statement
- Enrichment measured retrospectively is an upper bound on what the prospective screen will do, because the retrospective actives are already known to bind something like this pocket.
- A good ranking still says nothing about selectivity, cell activity, stability or developability. It says this compound plausibly fits this pocket.
- Covalent binders, metal-containing sites and pockets occupied by structured water need methods built for them; a standard scoring function in those cases produces numbers that look normal and are not.
- Docking cannot tell you the pocket is the right pocket. The target hypothesis comes from biology, not from the ranking.
Frequently asked questions
Can a docking score be converted to a Kd?
No. The score is calibrated to order compounds, not to reproduce binding free energies, and the error on any single value is large compared with the differences a project cares about. Report it as a rank, not as an affinity.
How do you know whether a docking protocol works for your target?
Run it retrospectively on known actives mixed with property-matched decoys and measure the enrichment in the top slice. An enrichment factor well above one means the ordering carries information for this pocket; near one means it does not.
How many hits should go to the assay?
As many as the budget allows, chosen as cluster representatives from the top of the ranking rather than as the single best cluster, plus two or three mid-ranked compounds as a control on the ranking itself.
Does consensus scoring make docking more accurate?
It raises the proportion of true binders in a shortlist and lowers the number of actives found. It is a precision-recall trade, not a validation, because several functions can share the same systematic blind spot.
Why do large molecules rank higher?
Because most scoring functions sum favourable contacts, and more atoms make more contacts. Compare within narrow molecular-weight bands, or normalise by heavy-atom count, to see whether a ranking is more than a size ranking.