--- library_name: "ofoldx" tags: - "biology" - "biomolecular-design" - "protein" - "rna" - "dna" - "pipeline" - "boltz2-affinity" - "binding-affinity-prediction" - "affinity-prediction" - "protein-ligand" artifact_kind: "pipeline" repo_id: "oteam/boltz2-affinity" license: "mit" base_model: "boltz-community/boltz-2" pipeline_tag: "other" task: "binding_affinity_prediction" model-index: - name: "boltz2-affinity" results: [] widget: - pipeline_tag: "other" task: "binding_affinity_prediction" example_title: "Protein-ligand affinity scoring" text: "complex_structure: complex.cif\nligand_chain: L" input_format: "structure_path" --- # boltz2-affinity OFoldX `pipeline` artifact for protein-ligand binding-affinity prediction, using the `boltz2-affinity` architecture. ## Disclaimer 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](https://huggingface.co/boltzgen/boltzgen-1) - **Primary Paper**: [Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction](https://doi.org/10.1101/2025.06.14.659707) - **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`. ### Links - **Hub repository**: [oteam/boltz2-affinity](https://huggingface.co/oteam/boltz2-affinity) - **Upstream paper**: [Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction](https://doi.org/10.1101/2025.06.14.659707) - **Upstream repository**: [Boltz-2](https://github.com/jwohlwend/boltz) - **Source checkpoint release**: [https://huggingface.co/boltzgen/boltzgen-1](https://huggingface.co/boltzgen/boltzgen-1) - **Code**: [`ofoldx/pipelines/binding_affinity.py`](https://github.com/OTeam-AI4S/OFoldX/tree/main/ofoldx/pipelines/binding_affinity.py) - **Project repository**: [https://github.com/OTeam-AI4S/OFoldX](https://github.com/OTeam-AI4S/OFoldX) - **Issues**: [https://github.com/OTeam-AI4S/OFoldX/issues](https://github.com/OTeam-AI4S/OFoldX/issues) ## Usage The artifact depends on the [`ofoldx`](https://github.com/OTeam-AI4S/OFoldX) library. Install it with pip: ```bash 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](https://github.com/OTeam-AI4S/OFoldX/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.
