The Protein Structure Ensemble Comparator triages a set of structures of one protein or complex: design candidates, prediction samples, or conformers. It scores every pair by TM-score and RMSD, groups near-identical structures into clusters, and keeps the most central member of each cluster as its representative. Use it to remove redundant designs before interface analysis or MD, or to see how much a set of predictions disagrees. It runs on CPU with Biotite.
How it works
- Selection: each structure contributes the protein chains being compared. Monomer mode uses one chain (the first protein chain unless you name one); multichain mode joins several chains in a fixed order, and every structure must have them. Waters, ligands, and other hetero groups are ignored for scoring but kept in the written structures.
- Alignment: every pair is superimposed by sequence-independent structural alignment, so designs with different sequences can be compared.
- Scores: TM-score is normalized by each structure's length, and the pair keeps the lower of the two, so a short fragment does not look like a match to a full-length protein. RMSD is over the aligned residues, and coverage is aligned residues over the longer chain. An optional reference structure, such as a target or crystal structure, is also scored against every member.
- Clustering: structures are taken in order of their mean TM-score to the rest of the set (ties by input order). Each one not yet assigned starts a cluster and claims every unassigned structure at or above the TM-score threshold. The same input always gives the same clusters.
- Representatives: each cluster keeps its medoid, the member with the highest mean TM-score to the other members. The whole set's medoid is reported too, and every structure is written superposed onto it.
Choosing a threshold
| TM-score threshold | Groups together |
|---|---|
| 0.5 | Structures with the same fold |
| 0.8 (default) | Close variants of one structure, such as near-duplicate designs |
| 0.9 and above | Near-identical copies |
TM-score is on a 0–1 scale and, unlike RMSD, does not grow with protein size, so one threshold works across ensembles of different proteins.
Inputs
PDB or CIF files, or a structure ensemble artifact from an upstream workflow step such as Boltz-2, BoltzGen, FreeBindCraft, OpenDDE, RFAntibody, or the MD Trajectory Analyzer. Only the first model of a multi-model file is compared. B-factors are kept, so predicted structures keep their pLDDT.
Outputs
| File | Contents |
|---|---|
pairs.csv | Every pair: TM-score (both normalizations and the pair score), RMSD, aligned residues, and coverage |
clusters.csv | Every structure: cluster, cluster size, whether it is the representative, scores to its representative and to the reference, and its cluster's mean TM-score and RMSD |
representatives/ | One representative structure per cluster, as PDB |
medoid.pdb | The medoid of the whole set |
aligned/ | Every input superposed onto the medoid |
summary.json | Counts, chains, threshold, cluster membership, and warnings |
A structure that PDB cannot hold (an mmCIF with chain IDs longer than one character, or more than 99,999 atoms) is still compared, but its structure files are written as mmCIF (.cif), with a warning. Those files are download only and are not passed to later workflow steps.
In a workflow, the representatives and the medoid carry TM-score and cluster-size metrics, so a filter step can keep, for example, only representatives of clusters with several members.
Not covered
- Clustering is greedy, so it depends on the threshold and is not a global optimum.
- Multichain mode joins chains in the order given; it does not try chain permutations, so swapped copies in a homo-oligomer score lower.
- A very short fragment compared with a much longer chain can fail to align and scores 0.
- No database-wide structure search, no MD-frame clustering (use the MD Trajectory Analyzer), and no small-molecule comparison (use the RMSD Calculator).
Related tools
Representatives go on to the Interaction Profiler, MM-GBSA Rescoring, or OpenMM MD.