MolCraftDiffusion Zoo — Data
Corpora and run inputs for MolCraftDiffusion, stored as ASE SQLite databases. Corpora are keyed by corpus and processing, so models that share one share the download.
Assets resolve by name, not by path:
pip install molcraftdiffusion
export HF_TOKEN=hf_xxx # these repos are private
MolCraftDiff zoo fetch --model kgdiff
MolCraftDiff generate examples/kgdiff_generate.yaml
Files land in $MOLCRAFT_ASSETS (default ~/.cache/molcraft/zoo), laid out
exactly as the paths below.
How things are named
| Prefix | Meaning | Example |
|---|---|---|
data/<corpus>/<variant> |
a corpus — never named after a model | data/crossdocked/pockets10a |
inputs/<name> |
a run input — reference structure, single target | inputs/templates |
| (model name) | weights and their sidecars | kgdiff/pretrained |
A corpus belongs to nobody, so CrossDocked pockets are
data/crossdocked/pockets10a — read by KGDiff, IPDiff and DiffSBDD alike, and
downloaded once.
Code: https://github.com/pregHosh/MolCraftDiffusion
Corpora
| asset | size | licence |
|---|---|---|
data/chembl3d/stereo-200 |
868.0 KB | CC-BY-SA-3.0 (ChEMBL) |
data/crossdocked/diffdec |
923.2 KB | CC0-1.0 |
data/crossdocked/diffint |
371.4 MB | CC0-1.0 |
data/crossdocked/diffpharma |
4.1 GB | CC0-1.0 |
data/crossdocked/diffpharma-smoke |
4.1 MB | CC0-1.0 |
data/crossdocked/pmdm |
440.0 KB | CC0-1.0 |
data/crossdocked/pockets10a |
1.3 MB | CC0-1.0 |
data/formed/3k |
32.5 MB | MIT |
data/formed/full |
1.4 GB | MIT |
data/geom/ase |
618.8 MB | CC0-1.0 (GEOM-Drugs) |
data/geom/ase-gcdm |
488.0 KB | CC0-1.0 |
data/geom/graph3d-flowmol3 |
102.8 MB | CC0-1.0 |
data/geom/graph3d-val |
102.9 MB | CC0-1.0 |
data/iflp_mof/linkers |
33.6 MB | MIT |
data/qm9/ase |
253.0 MB | CC0-1.0 |
data/qm9/ase-300 |
368.0 KB | CC0-1.0 |
data/qm9/ase-4k |
7.4 MB | CC0-1.0 (QM9) |
data/qm9/graph3d |
191.7 MB | CC0-1.0 |
data/qm9/graph3d-400-ditmc |
552.0 KB | CC0-1.0 |
data/qm9/graph3d-400-nextmol |
552.0 KB | CC0-1.0 |
data/qm9/graph3d-4k |
5.2 MB | CC0-1.0 |
data/tmc/ligandiff |
341.3 MB | CC0-1.0 |
data/tmc/ligandiff-multi |
3.5 GB | CC0-1.0 |
data/zinc/difflinker |
666.1 MB | free-for-academic (ZINC) |
Run inputs
| asset | size | licence |
|---|---|---|
inputs/apo2mol/8tv6 |
644.0 KB | MIT |
inputs/difflinker/fragments |
2.3 KB | free-for-academic (ZINC) |
inputs/ligandiff-multi/reference |
385.0 B | CC0-1.0 |
inputs/ligandiff/reference |
615.0 B | CC0-1.0 |
inputs/loqi/stereo-ref |
216.4 KB | CC-BY-SA-3.0 (ChEMBL) |
inputs/silvr/reference |
1.2 KB | CC0-1.0 |
inputs/templates |
4.7 KB | MIT |
Provenance
- QM9 — Ramakrishnan et al. 2014 (CC0).
- GEOM-Drugs — Axelrod & Gómez-Bombarelli, Harvard Dataverse.
- CrossDocked2020 — Francoeur et al. 2020; CC0 1.0 per the upstream LICENSE.
- ChEMBL3D — derived from ChEMBL (CC BY-SA 3.0).
- ZINC — via DiffLinker's preprocessed release.
- FORMED, IFLP-MOF — produced for the MolCraftDiffusion manuscript.
The Apo2Mol apo/holo corpus is not here: its upstream dataset is
access-gated. MolCraftDiff zoo recipe explains how to obtain it.
- Downloads last month
- 18