Docking asks a deceptively simple question: if this small molecule met this protein, where would it sit, and how well would it fit? The answer is used to rank millions of compounds before any of them is made in a lab, which makes it one of the most consequential calculations in early drug discovery — and one of the easiest to misuse.
The two halves of the calculation
Every docking run combines a search problem with a scoring problem.
The search explores the space of possible poses: the compound’s position, its orientation, and the rotation of each of its rotatable bonds. This space is enormous — even a molecule with six rotatable bonds has far more distinct arrangements than can be listed one by one — so the search is a heuristic. Genetic algorithms, Monte Carlo methods and incremental construction are all attempting the same thing: find low-energy arrangements without visiting every possibility.
The scoring function then puts a number on each pose. It is a physics-flavoured approximation, combining terms for van der Waals contact, electrostatics, hydrogen bonding and the energy cost of freezing a flexible ligand into one shape. Because it is an approximation, the score is closer to a ranking signal than to a measured binding free energy.
Why poses are often easier than affinities
Search usually succeeds at a modest task: the best-scoring pose frequently lands close to the experimental pose, especially when the protein’s binding site is well defined and reasonably rigid. Affinity is harder. Scores correlate loosely with measured binding, and small changes that a chemist would consider trivial can change the ranking.
This is why docking is trusted as a filter — a way to reduce a large library to a smaller, more plausible set — and not as a substitute for an assay.
The assumptions that decide your result
Three assumptions do most of the damage:
A rigid receptor. Many protocols treat the protein as fixed, or let only a few side chains move. Real binding sites rearrange. When the true pose needs a side chain to shift, rigid docking can miss it entirely.
A bound-water transplant. Water molecules in a site can be structural — they can be part of the interaction network rather than obstacles to it. Removing all of them, or keeping all of them, changes the answer.
A single protonation state. The compound and the protein each have a charge state that depends on pH. Choosing the wrong one changes the electrostatics, sometimes decisively.
Add the fact that many compounds are scored in a conformation they could not realistically adopt, and the practical rule becomes clear: docking narrows a field, it does not decide it.
Using docking honestly
A defensible docking study states, before the run, how poses will be judged. If a crystal structure or a known co-crystallised ligand exists, re-docking that ligand is the calibration step — if the method cannot reproduce a pose that is already known, the scores it produces for novel compounds mean very little.
It also reports the search space, the protonation states, the treatment of water, and whether the receptor was flexible. Those four details are what make a docking result reproducible by someone else — and what let a reader decide how much weight it deserves.
Where it sits in a pipeline
In a realistic workflow, docking follows target validation and dataset curation, runs alongside property and toxicity prediction, and precedes synthesis. Its value is the compounds it removes from consideration, which is exactly why the assumptions above are worth stating plainly rather than hiding behind a table of scores.