Local structure agreement
Compares two structures of the same protein without superposing them first (lDDT), so one badly-modeled region can't distort the score the way it can with a global alignment. Use to score a predicted structure against a reference, or to see exactly which regions of two structures actually disagree.
Local structure agreement
What it does
Takes two structures of the same protein — commonly a prediction and an experimental reference, or two structures from different methods — and scores how well their local geometry agrees, without superposing them first.
This matters because superposition-based comparison has a real weakness: the software picks the single best global rotation and translation to overlay the two structures, and then measures the leftover distance. One badly-modeled loop can drag that optimal overlay off just enough to make an otherwise-excellent model look mediocre everywhere — or, the opposite way, get averaged into invisibility across a large structure. Local structure agreement instead asks a much more local question, over and over: "for every pair of residues close together in the reference, are they still about that close together in the model?" A bad region only breaks its own answers.
How the score is computed
The method is lDDT (local Distance Difference Test), the standard superposition-free structure-comparison metric. For every pair of Cα atoms within 15 Å of each other in the reference structure, the same pair's distance in the model is checked against four thresholds — 0.5, 1, 2, and 4 Å — and scored by how many of those thresholds it still satisfies. The overall score is the average across every such pair; a per-residue score (shown in the table) averages just the pairs that residue is part of.
A score of 1.0 means every local neighborhood agreed perfectly. A residue scoring 0.0 means every one of its evaluated local relationships broke — the region moved somewhere geometrically unrelated to where it started.
Reading the output
The table lists every residue, worst first — the residues whose local neighborhoods disagree most between the two structures, which is usually what a reader actually wants to find. A residue with nothing else nearby to compare it against (rare, and typically only at the very end of a chain) shows no score rather than a misleading zero.
Input
Two structures, saved as PDB or mmCIF, of the same protein — plus which chain to use from each.
Important limitations
Cα only. This compares backbone positions, not side chains — it answers "does the fold agree locally," not "do the side-chain conformations agree."
Residues must share numbering. The two structures are matched residue-by-residue by their residue number. If one structure is numbered differently from the other (a common situation when comparing a prediction against an experimental structure with gaps or an offset), this tool can't currently reconcile that — the residues it can't match are dropped, and if too few remain, the run refuses rather than guessing.
A comparison between two specific structures, not a claim about which one is "more correct." lDDT says how much two structures agree locally; it doesn't say which one is right.