A protein is a chain of amino acids that folds into a specific three-dimensional shape, and it is usually that folded shape — not the chain — that a drug interacts with. So representing a protein for a model raises a choice.
The simplest representation is the sequence: the ordered list of amino acids, written as letters such as \(\text{MKTLLAV}\). This is cheap, reliable, and available for essentially every known protein. It captures what the protein is made of but not how it folds.
The richer representation is the three-dimensional structure, and specifically the binding pocket — the cavity or surface patch where a small molecule would sit. A pocket representation records the geometry and chemical character of that region: which amino acids line it, how large it is, whether it is deep or shallow, and whether it is greasy or water-attracting.
Historically the bottleneck was that experimental structures were slow and expensive to obtain, so most proteins had a sequence but no structure. Predicted structures have changed that balance, and this is one of the clearest places where AI has altered what is computationally available. The practical consequence for prediction is direct: a model given only sequence is guessing about geometry, while a model given a pocket can reason about fit.