ZINC250k

The 249,455 drug-like molecules drawn from ZINC by Gómez-Bombarelli et al. (2018), a usual benchmark of molecule generation. They are molecules, not crystals: pbc is False and x holds positions in Å, see Periodicity.

The positions are not from the source

ZINC250k is a table of SMILES. The 3D positions of this dataset were generated when it was built: hydrogens added, one conformer per molecule with RDKit 2026.03.6 (ETKDGv3, seed 42), then relaxed with the MMFF94 force field. They are plausible geometries, not DFT ones, and not the conformers of another paper.

Molecules	249,427 (249,455 in the source, 28 without conformer)
Atoms	10,923,322, 9 to 83 per molecule (44 on average), hydrogens included
Elements	H, C, N, O, F, P, S, Cl, Br, I

ZINC and redistribution

ZINC is free to use, but its terms say that significant portions of it may not be redistributed without the written permission of its authors. ZINC250k is a small sample, 0.1 % of ZINC, and its table of SMILES is published by the authors of the benchmark: it is published here on this ground. Larger parts of ZINC are not: see Datasets to build yourself.

The file can also be built again, in about half an hour on eight cores:

uv run python scripts/build_zinc250k.py datasets

Labels

The four labels (total_energy, band_gap, ...) are NaN.

properties[...]	Source column
logp	logP	octanol-water partition coefficient
qed	qed	drug-likeness, 0 to 1
sas	SAS	synthetic accessibility score
mmff_converged	—	1 if the relaxation with MMFF94 converged

Things to know

• 28 molecules are missing: RDKit found no conformer for them, mostly strained polycyclic cages. all_index.parquet gives for each row its row in the source and its SMILES.
• Two molecules are kept with mmff_converged at 0.
• Charged molecules are in the source (protonated amines, for instance) and are kept.
• No bond is stored, as for every dataset: graphs are built from the positions.
• No split. Papers draw their own.
• Licence. The table comes from the repository of the paper (Apache 2.0); the molecules come from ZINC: free to use, significant portions not to be redistributed without written permission.
• Same positions with the same RDKit. The build is reproducible with RDKit 2026.03.6; another version can give other conformers.

Format

HDF5 files of flat concatenated arrays, the format of the other datasets of materials-toolkits:

Dataset	Shape	Type	Content
lattice	(n_struct, 3, 3)	float32	lattice vectors as rows, in Å; the identity for a molecule
num_atoms	(n_struct,)	int64	number of atoms of each structure
ptr	(n_struct,)	int64	index of the first atom of each structure in x and z
x	(n_atoms, 3)	float32	fractional coordinates; positions in Å for a molecule
z	(n_atoms,)	int64	atomic numbers
pbc	(n_struct, 3)	bool	whether the structure is periodic along each lattice vector
total_energy, formation_energy_per_atom, energy_above_hull, band_gap	(n_struct,)	float32	labels, NaN when unknown
properties/<name>	(n_struct,)	float32	the properties listed above, NaN when unknown

The position of an atom in Å is x @ lattice. The lattice of a crystal is rebuilt from the lengths and the angles of its cell, in the standard orientation (first vector along x), and its fractional coordinates are wrapped into the cell. The batch array is an artefact of the writer and is not used when reading. Row i of an index file describes structure i.

Usage

from torchms.data import HDF5Dataset

dataset = HDF5Dataset.from_hub("materials-toolkits/zinc250k", "all.hdf5")
structures = dataset[:32]            # batched Structures
structures.properties                # the named properties, one value per structure

Built with torchms.

Source and checks

• 250k_rndm_zinc_drugs_clean_3.csv of chemical_vae, checked with its MD5.
• Built by scripts/build_zinc250k.py. For every molecule: its atoms are the ones of its SMILES with its hydrogens, every bond of the SMILES is shorter than 2.5 Å, no two atoms are closer than 0.5 Å, and the three properties are the ones of the table. 500 molecules were generated a second time and got the same positions.
• Repository: materials-toolkits/zinc250k, file all.hdf5.