CoolFace
Datasetpublic

Tenstorrent/mgtbind-boltz2

MGTBind Boltz-2 Structure Predictions MGTBind Bindbridge Tenstorrent Boltz-2 ternary complex structure predictions for all 3,607 molecular glue complexes in the MGTBind database. Predicted in partnership with Bindbridge and computed on Tenstorrent AI accelerators using TT-Boltz. Results Comparison against the existing AlphaFold 3 predictions in MGTBind (all 3,607 complexes): Metric AF3 Boltz-2 Avg MG Chain ipTM 0.7443 0.7896 Delta MG Chain ipTM… See the full description on the dataset page: https://huggingface.co/datasets/Tenstorrent/mgtbind-boltz2.

sourceHugging Facemitupdated 4mo agoView on Hugging Face
0likes2.5kdownloads
Dataset Card

MGTBind Boltz-2 Structure Predictions

<table align="center"><tr> <td align="center"><a href="https://mgdb.pkumdl.cn/"><img src="logos/mgtbind.png" alt="MGTBind" height="40"></a><br><a href="https://mgdb.pkumdl.cn/"><b>MGTBind</b></a></td> <td align="center"><a href="https://bindbridge.com"><img src="logos/bindbridge.png" alt="Bindbridge" height="40"></a><br><a href="https://bindbridge.com"><b>Bindbridge</b></a></td> <td align="center"><a href="https://tenstorrent.com"><img src="logos/tenstorrent.png" alt="Tenstorrent" height="40"></a><br><a href="https://tenstorrent.com"><b>Tenstorrent</b></a></td> </tr></table>

Boltz-2 ternary complex structure predictions for all 3,607 molecular glue complexes in the MGTBind database. Predicted in partnership with Bindbridge and computed on Tenstorrent AI accelerators using TT-Boltz.

Results

Comparison against the existing AlphaFold 3 predictions in MGTBind (all 3,607 complexes):

MetricAF3Boltz-2
Avg MG Chain ipTM0.74430.7896
Delta MG Chain ipTM (Boltz-2 - AF3)-+0.0453
MG Chain ipTM > 0.6862.0%70.2%
Avg pLDDT of MGN/A74.07
Avg pLDDT of MG > 70N/A70.7%
Both criteria met (ipTM > 0.68 and pLDDT > 70)N/A56.7%

File Structure

The directory layout mirrors MGTBind's AlphaFold 3 structure convention:

{complex_id}/boltz2_structure/
    Model-1/model.cif               # Best-ranked (rank 0)
    Model-2/model.cif               # Rank 1
    Model-3/model.cif               # Rank 2
    Model-4/model.cif               # Rank 3
    Model-5/model.cif               # Rank 4
    top5_summary_confidences.json   # Per-model confidence metrics

Additional files at the repository root:

FileDescription
results.jsonComplete per-complex Boltz-2 metrics for all 3,607 predictions, including per-run confidence scores, pairwise chain ipTM matrix, and inference runtime.
results.csvFlattened CSV with one row per complex and key metrics for easy filtering (e.g. by quality thresholds).
scripts/batch_download_boltz2.pyBatch download script, mirrors MGTBind's batch_download_af3.py.

Mapping to MGTBind

Complex IDs correspond to the id column in MGTBind's `complexes.csv`. For example, complex 702 maps to 702/boltz2_structure/Model-1/model.cif.

Model-1 is the best-ranked structure (by confidence score = 0.8 ipTM + 0.2 pTM), consistent with MGTBind's AF3 ranking convention. Five diffusion samples were generated per complex.

Download

Full Dataset

bash
# Clone the full repository
git clone https://huggingface.co/datasets/tenstorrent/mgtbind-boltz2

Specific Complexes

bash
# Download specific complex IDs using the batch script
python scripts/batch_download_boltz2.py -f complex_ids.txt -o output_dir

Where complex_ids.txt contains comma-separated MGTBind complex IDs (e.g. 244,245,702).

Individual Files

Access any structure directly:

https://huggingface.co/datasets/tenstorrent/mgtbind-boltz2/resolve/main/{complex_id}/boltz2_structure/Model-{N}/model.cif

Metric Definitions

  • —MG Chain ipTM: For each complex, the molecular glue chain's interface predicted TM-score to each protein chain, averaged across protein chains.
  • —Avg pLDDT of MG: Mean pLDDT across all molecular glue heavy atoms (non-hydrogen), read from each atom's B_iso_or_equiv value in the CIF file.
  • —Confidence score: Boltz-2 ranking score, computed as 0.8 ipTM + 0.2 pTM.

The top5_summary_confidences.json per complex follows the same format as MGTBind's AF3 confidence summaries. Fields shared with AF3: model_id, ranking_score, ptm, iptm, avg_plddt, avg_plddt_of_mg. Additional Boltz-2 metrics: ligand_iptm, protein_iptm, complex_pde, complex_ipde. AF3 fields not produced by Boltz-2: has_clash, fraction_disordered.

Methodology

ParameterValue
ModelBoltz-2 via TT-Boltz
HardwareTenstorrent Blackhole AI accelerators
Precisionbfloat16
Diffusion samples5 per complex
RankingBest model selected by confidence score (0.8 ipTM + 0.2 pTM)
MSA generationMMseqs2
Input sequencesCanonical sequences from MGTBind complexes.csv

Partnership

This dataset was created through a collaboration between:

  • —[Bindbridge](https://bindbridge.com) --- Cambridge-based ag-biotech company using AI to discover molecular glues for crop protection. Bindbridge initiated the project, prepared the input data from the MGTBind database, and defined the evaluation methodology.
  • —[Tenstorrent](https://tenstorrent.com) --- All structure predictions were computed on Tenstorrent Blackhole AI accelerators using TT-Boltz, a Boltz-2 and BoltzGen implementation for inference on Tenstorrent hardware.
  • —[MGTBind](https://mgdb.pkumdl.cn/) --- Molecular Glue and Ternary Binding Database, maintained by Jianfeng Pei's group at Peking University.

Citation

If you use these structures in your research, please cite the MGTBind database:

Shuya Li, Xufan Gao, Hailin Hu, Sheng Wang, Jianfeng Pei. MGTbind: a molecular glue database with ternary complex structures, binding and functional data. Nucleic Acids Research, Volume 53, Issue D1, 2025, Pages D1500-D1507.

License

This dataset is released under the MIT License.

Tenstorrent/mgtbind-boltz2 · CoolFace