zonca/openfold3-mcl1-expanse
OpenFold3 MCL1 protein-ligand ensemble (20 predictions) on SDSC Expanse V100 Ensemble of 20 AlphaFold3-equivalent structure predictions for the MCL1 protein–ligand complex (official OpenFold3 example, PDB 5FDR context), generated with OpenFold3 0.4.5 on a single NVIDIA V100 32 GB GPU on SDSC Expanse (gpu-shared partition). Data 20 predicted structures (PDB): predictions/seed_{42,1337,2024,2026}/mcl1_*_model.pdb 4 seeds × 5 diffusion samples = 20 independent… See the full description on the dataset page: https://huggingface.co/datasets/zonca/openfold3-mcl1-expanse.
OpenFold3 MCL1 protein-ligand ensemble (20 predictions) on SDSC Expanse V100
Ensemble of 20 AlphaFold3-equivalent structure predictions for the MCL1 protein–ligand complex (official OpenFold3 example, PDB 5FDR context), generated with OpenFold3 0.4.5 on a single NVIDIA V100 32 GB GPU on SDSC Expanse (gpu-shared partition).
Data
- 20 predicted structures (PDB):
predictions/seed_{42,1337,2024,2026}/mcl1_*_model.pdb - 4 seeds × 5 diffusion samples = 20 independent predictions
- protein chains A–D (MCL1), ATP ligands (chains F/G/H), small-molecule ligand (chain Z)
- Per-atom confidence JSON:
*_confidences.json(plddt, pae, pde) - Aggregated confidence JSON:
*_confidences_aggregated.json(avgplddt, ptm, iptm, bespokeiptm, samplerankingscore, has_clash, ...) - Timing: per seed
timing.json - Analysis:
mcl1_ensemble_metrics.csv(all 20 predictions with confidence + ligand RMSD),mcl1_ensemble_metrics.json,analyze_ensemble.py
Run details
- Model: OpenFold3 0.4.5 (open weights, Apache-2.0), native PyTorch kernels (V100 sm_70; cuEquivariance/deepspeed not supported on V100)
- MSAs: real, via ColabFold MSA server (outbound access worked from compute node)
- GPU: 1× NVIDIA V100 32 GB, SDSC Expanse
gpu-shared, account QoSgpu-shared-normal - Wall time: 24 min 13 s (job 53383743), exit 0
Key result
Confidence ranking vs ligand-pose consistency: corr(sample_ranking_score, ligand RMSD) = −0.58 — higher-ranked predictions place the ligand more consistently (moderate effect), supporting the hypothesis that confidence can help rank ligand poses in an ensemble. Per-chain pTM ~0.87 on the top-ranked structure (chains A–D are 4 identical MCL1 copies that permute across samples, inflating global protein RMSD; use a single-chain query for cleaner 5FDR ligand-placement benchmarks).
Links
- GitHub (scripts, configs, analysis): https://github.com/zonca/openfold3-mcl1-expanse
- Zenodo dataset (archive DOIs): 10.5281/zenodo.21926059
- OpenFold3: https://github.com/aqlaboratory/openfold-3
- OpenFold3 docs: https://openfold-3.readthedocs.io
- Original example: https://huggingface.co/OpenFold/OpenFold3/tree/main/examples/common_examples/mcl1
Related
Sister dataset (AlphaFold3 via nf-core, TetR dimer+DNA): github.com/zonca/proteinfold-on-expanse
Citation
If you use this data, please cite OpenFold3 and AlphaFold3 (see repo README), and link this dataset (DOI 10.5281/zenodo.21926059).
License
Data CC-BY-4.0. OpenFold3 model Apache-2.0; AlphaFold3 cited per its paper.