molecular docking
Molecular docking is like trying to fit a key into a lock on a computer before you ever cut the metal. Given the three-dimensional shape of a target protein's pocket and a candidate molecule, the software tries many orientations and shapes of the molecule inside the pocket to predict how it would sit there and how tightly it might hold on.
Mechanically, a docking program does two jobs. First, a search algorithm samples many possible poses, meaning positions, orientations, and internal rotations of the ligand within the binding site. Second, a scoring function rates each pose with a number meant to approximate how favorable the binding would be, so the program can rank poses and rank different molecules against one another.
Docking is fast and cheap enough to screen huge virtual libraries, which is why it underpins virtual screening and structure-based design. But it is an approximation: protein flexibility, water molecules, and the entropy of binding are handled crudely, so the predicted pose can be right while the predicted affinity is unreliable, or vice versa. Treat docking as a hypothesis generator to be confirmed by experiment, not as a measurement.
A team docks a million-compound library into a kinase ATP pocket, keeps the top-scoring few thousand poses, visually inspects them for sensible hydrogen bonds to the hinge, and orders a few hundred compounds for testing.
Docking as a funnel that narrows a vast library down to a testable shortlist.
A common pitfall is to over-interpret docking scores as true affinities. Scores are best used to rank closely related compounds or to filter obviously poor binders, not to predict a numeric IC50.