PLAS-20K protein Cα MD (processed)

A processed, protein Cα view plus the ligand's heavy atoms of the PLAS-20k molecular-dynamics dataset, made for training residue-level protein models. It is not the original dataset: the water and ions are removed (the ligand keeps its heavy atoms only) and only per-residue quantities are kept. For the full simulations use the original sources below.

Source and licence

• PLAS-20k: D. B. Korlepara et al., PLAS-20k: Extended Dataset of Protein-Ligand Affinities from MD Simulations for Machine Learning Applications, Scientific Data 11, 180 (2024), https://doi.org/10.1038/s41597-023-02872-y (data record: https://doi.org/10.6084/m9.figshare.c.6742521, CC BY 4.0).
• Trajectories downloaded from MDRepo (https://mdrepo.org), collection "PLAS-20K".
• This derived dataset is shared under CC BY 4.0; please cite the PLAS-20k paper and MDRepo.

Simulations (from the paper): OpenMM 7.2, Amber ff14SB (protein), GAFF2 (ligands), TIP3P water; five independent 4-ns NPT production runs per complex, frames every 100 ps.

What was changed

• Protein atoms (mdtraj "protein" selection) and the ligand's heavy atoms (everything not protein, water, hydrogen or a single-atom ion); water, ions and hydrogens are dropped.
• Per residue and frame: Cα coordinates (Å), backbone φ and ψ (radians, NaN at chain ends), DSSP (8-state codes as int8); per frame: radius of gyration, end-to-end distance, replica (run_id) and time.
• Per ligand heavy atom and frame: coordinates (Å), in the same frames as the protein; per atom: element, bonded hydrogens, heavy-atom degree, ring membership; bond-graph hop distances. Ligands with more than 128 heavy atoms are left out (the protein is kept).
• Every frame of every replica is kept (no thinning).
• Train/validation split (15,739 / 828 complexes) by MMseqs2 sequence clusters at 30% identity: complexes of the same or similar proteins (e.g. one protein with different ligands) are always on the same side (2,064 / 162 clusters).

Contents

complexes (MDRepo runs)	16,567 of 19,466 PLAS-20K runs in MDRepo
replicas per complex	5 (most); 0 complexes have fewer
frames	3,313,400
residue rows	938,685,400
chain length	12-600 residues, median 273
complexes with a ligand	16,566
ligand heavy atoms	median 29, max 126
ligand atom rows	105,181,600

Runs left out:

reason	runs
more than one protein chain	2,236
longer than 600 (or shorter than 8) residues	663

Format

Plain NumPy files in plas20k/ (download e.g. with huggingface_hub.snapshot_download(repo_id, repo_type="dataset")):

• ca.npy float32 [R, 3], phi.npy / psi.npy float32 [R], dssp.npy int8 [R]: every system's frames flattened residue-major, R = Σ T·L.
• rg.npy, e2e.npy, time_ps.npy float32 [F], run_id.npy int32 [F]: per frame, F = Σ T.
• lig_xyz.npy float32 [Rl, 3] (Å): ligand heavy atoms, row lig_offset + t * n_lig + a for frame t and atom a (frames in the same order as the protein rows), Rl = Σ T·n_lig; no rows for a complex without a ligand.
• index.json: systems, one entry per complex with name, seq, L, T, res_offset, frame_offset, split, cluster and meta (MDRepo id and slug, PDB id, UniProt ids, ligands, force field, frames per replica). Ligand fields: n_lig (heavy atoms, 0 if none), lig_offset, lig_atoms (per atom element, n_h bonded hydrogens, degree bonded heavy atoms, ring), lig_hops (bond-graph hop distances, capped at 8, as the flat upper triangle i < j row-major, length n_lig (n_lig - 1) / 2); meta.ligand_smiles (from MDRepo) and meta.ligand_bonds (heavy-atom bonds [i, j]).
• SHA256SUMS: a checksum per file.

System s's frames: ca[s.res_offset : s.res_offset + s.T * s.L].reshape(s.T, s.L, 3). Its ligand: lig_xyz[s.lig_offset : s.lig_offset + s.T * s.n_lig].reshape(s.T, s.n_lig, 3).