Pocket similarity search compares the shape and chemistry of binding sites. It uses no sequence information, so it can find similar sites in unrelated proteins. Because the query can be a whole pocket or a few residues, it can also compare subpockets.
The search relies on two parts: binding sites are indexed when a structure is registered, then each search matches the query against that index.
Every structure goes through three analyses.
3decision detects pockets with fpocket, an open-source, geometry-based method. fpocket places alpha spheres (spheres that touch four atoms and contain none) in the empty spaces at the protein surface, clusters them into pockets, and scores each pocket, for example its volume and druggability. It finds pockets with a ligand and pockets where a ligand could bind.
For structures with ligands, 3decision measures how much each ligand overlaps each pocket. This is what the Pockets with ligands only option uses.
Pockets with at least 30 alpha spheres, or with a ligand, are described residue by residue. Each lining residue gets a 6-bit fingerprint of the properties its side chain exposes:
The fingerprint and the position of the residue's Cα atom form a pocket feature.
For every pair of features less than 10 Å apart, 3decision stores the distance between the Cα atoms and their relative orientation, using a local frame built from the backbone N, Cα and C atoms. These feature pairs are the searchable signature of the pocket.
The filters you set (species, proteins, structures, protein families, and their exclusions) and the Pockets with ligands only option reduce the set of pockets to compare.
Each feature pair of the query is compared with the feature pairs of every candidate pocket. Two pairs match when:
A pocket with at least 30 matching feature pairs becomes a candidate.
Candidate ranking applies to version 2.5 and later. In 2.4, candidates are not ranked, so close matches don't necessarily appear first.
The most promising candidates are evaluated first, in two phases:
This is why close matches appear early in a running search.
Many matching pairs don't guarantee a meaningful match: in a large pocket, they can come from unrelated regions. The matching pairs are clustered by the RMSD of their Cα atoms, and only the largest consistent cluster is kept.
The score is the share of the query's feature pairs that are matched in that cluster:
query feature pairs matched in the cluster
------------------------------------------
all feature pairs of the query pocket
The score is normalized by the score of the query against itself, so it ranges from 0 to 1. A higher score means that more of the query is matched, with a similar geometry.
A pocket is returned as a hit when its score is above the Match score threshold. The default, 0.2, means that at least 20% of the query's feature pairs must be matched in a consistent cluster.
When you show a hit, it is superposed on the query with a distance-weighted fit on the Cα atoms of the matched residues of the best cluster.