This model card was generated by the OFoldX team for an OFoldX pipeline artifact. The upstream model authors did not write this card unless explicitly stated otherwise.

OFoldX is pre-alpha research software. Check the source checkpoint, upstream release, and local validation before using the artifact for scientific or operational decisions.

Model Details

Boltz-2 affinity model with a pair-only head for protein-ligand affinity scoring.

Converted Boltz-2 affinity checkpoint for protein-ligand binding-affinity prediction.

Model Provenance

Upstream Project: Boltz-2
Source Checkpoint: boltz2_aff.ckpt
Source Release: https://huggingface.co/boltzgen/boltzgen-1
Primary Paper: Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction
Upstream License: MIT for upstream Boltz code and weights

Model Specification

Field	Value
Repository	`oteam/boltz2-affinity`
Artifact Kind	`pipeline`
Task	`binding_affinity_prediction`
Architecture	`boltz2-affinity`
Entrypoint	`ofoldx.pipelines.binding_affinity.BindingAffinityPipeline`
Source Checkpoint	`boltz2_aff.ckpt`

[!NOTE] Source checkpoint: boltz2_aff.ckpt.

Usage

The artifact depends on the ofoldx library. Install it with pip:

pip install ofoldx

Pipeline Usage

Load the artifact from oteam/boltz2-affinity with the OFoldX task pipeline. Use AutoModel or AutoProcessor only when you need lower-level control:

python

from ofoldx.pipelines import Pipeline

pipeline = Pipeline.from_pretrained("oteam/boltz2-affinity")

When a matching processor is available, load it with AutoProcessor.from_pretrained(...) and pass the processed batch to the model.

Interface

Task: binding_affinity_prediction
Artifact kind: pipeline
Architecture: boltz2-affinity
Runtime files: manifest.json, config.json, and model.safetensors when present

Training Details

OFoldX did not train these weights. This repository contains a converted checkpoint and OFoldX runtime metadata for loading it.

Training Data

The Boltz-2 affinity heads use filtered assay data from sources such as PubChem, ChEMBL, and BindingDB, with decoy generation and structural-confidence filters described by the upstream report. OFoldX does not redistribute the training set.

Training Procedure

Upstream affinity training detaches the trunk and optimizes regression and binary binder/decoy objectives. OFoldX converts the released affinity checkpoint into model.safetensors plus an OFoldX manifest; it does not run Boltz-2 training.

Evaluation

OFoldX conversion reports and contract tests validate artifact structure and checkpoint loading. Task-level scientific evaluation should be checked against the corresponding upstream model release or paper.

Limitations

This artifact is distributed for research use.
Inputs must match the model-specific processor and expected biomolecular representation.
OFoldX is pre-alpha, so APIs and artifact metadata may still change before a stable release.

Citation

Please cite the upstream Boltz-2 work for the source checkpoint. If OFoldX supports your work, please also cite or link the OFoldX project repository.

bibtex

@article{passaro2025boltz2,
  author = {Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and Reveiz, Mateo and Thaler, Stephan and Somnath, Vignesh Ram and Getz, Noah and Portnoi, Tally and Roy, Julien and Stark, Hannes and Kwabi-Addo, David and Beaini, Dominique and Jaakkola, Tommi and Barzilay, Regina},
  title = {Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction},
  year = {2025},
  doi = {10.1101/2025.06.14.659707},
  journal = {bioRxiv}
}

Contact

Please use OFoldX GitHub issues for questions or comments about this model card.

License

The Hub license metadata, when present, reflects the source checkpoint or upstream project license. The OFoldX project license is not yet finalized. The source checkpoint is associated with the upstream license noted above: MIT for upstream Boltz code and weights. Review both OFoldX and upstream terms before redistribution or production use.

oteam/boltz2-affinity

boltz2-affinity

Disclaimer