Runs N' Poses 2026.01: download

Runs N' Poses tests the ability of protein–ligand co-folding methods to generalize to systems different from those in their training set (Have protein-ligand co-folding methods moved beyond memorisation?; code on GitHub). CAMEO scores uploaded predictions against its data.

Modeling inputs

For each target, the protein chain sequences and the ligands (SMILES; 2025.01 also CCD codes): https://cameo3d.org/benchmarks/runs-n-poses/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 protein chain has its sequence, one letter per residue, with modified residues given by their parent's letter or X. Where the target has modified residues, the chain also has a modifications list in AlphaFold 3's format (ptmType: the residue's CCD code; ptmPosition: its residue number, from 1). Tools that support modifications can model them. Either way, residues are numbered after sequence.

Other files

• ground_truth.tar.gz: per target (PDB id), receptor.cif with every protein chain of the bioassembly, and ligand_files/ with one SDF per scored ligand, named after its first chain
• annotations.csv: one row per scored ligand, including its similarity to the training data before each cutoff (similarity_2021, similarity_2025)
• msas.tar.gz: MSAs generated with colabfold search (paired and unpaired, a3m); see README.txt inside

The Runs N' Poses data is distributed under the Creative Commons Attribution 4.0 license (CC-BY-4.0).