CAMEOContinuous Automated Model EvaluatiOn

Protein complexes 2026.01: download

Complex tests the prediction of whole protein assemblies, their stoichiometry included, for complexes unlike those released before each training cutoff. CAMEO built the dataset from the PDB and scores uploaded predictions against it.

Modeling inputs

For each target, one protein sequence per entity, without copy numbers or chain names: predicting how many copies of each entity the assembly holds is part of the task: https://cameo3d.org/benchmarks/complex/2026.01/download/inputs.json.The target list, with the numeric ids that uploads refer to, comes from the API (see the example code on the benchmark page).

Each entity has its sequence, one letter per residue. The ground truth holds no modified residues: a modified residue is written as its parent amino acid (selenomethionine as methionine, phosphoserine as serine, selenocysteine as cysteine, pyrrolysine as lysine), with the atoms the parent has, or, when it has no single parent (a chromophore, for example), as UNK with its backbone atoms. The sequence gives the parent's letter, or X for UNK. A terminal cap such as ACE or NH2 is an X in the sequence, and the X must be kept, since residues are numbered after sequence (the ground truth holds the caps).

Entities are those of the PDB entry, so two entities can have the same sequence (for example, two copies that differed in the entry only by a modified cysteine); scoring matches model chains by the sequence.

Other files

The Complex data is distributed under the Creative Commons Attribution-ShareAlike 4.0 license (CC-BY-SA-4.0).