--- library_name: "ofoldx" tags: - "biology" - "biomolecular-design" - "protein" - "rna" - "dna" - "pipeline" - "boltzgen-adherence" - "design-generation" - "protein-design" artifact_kind: "pipeline" repo_id: "oteam/boltzgen-adherence" license: "mit" base_model: "boltz-community/boltz-2" datasets: - "boltzgen/boltzgen1_train" - "boltzgen/inference-data" pipeline_tag: "other" task: "design_generation" model-index: - name: "boltzgen-adherence" results: [] widget: - pipeline_tag: "other" task: "design_generation" example_title: "Backbone sequence design" text: "input_structure: backbone.cif\ndesign_chains: A" input_format: "structure_path" - pipeline_tag: "other" task: "design_generation" example_title: "Binder design" text: "target_structure: target.cif\ntarget_chains: A\ndesign_chains: B" input_format: "structure_path" --- # boltzgen-adherence OFoldX `pipeline` artifact for biomolecular design generation, using the `boltzgen-adherence` 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 BoltzGen design generator variant optimized for adherence to conditioning signals. Converted BoltzGen adherence generator checkpoint for controllable binder and protein design. ### Model Provenance - **Upstream Project**: BoltzGen - **Source Checkpoint**: `boltzgen1_adherence.ckpt` - **Source Release**: [https://huggingface.co/boltzgen/boltzgen-1](https://huggingface.co/boltzgen/boltzgen-1) - **Primary Paper**: [BoltzGen: Toward Universal Binder Design](https://doi.org/10.1101/2025.11.20.689494) - **Upstream License**: MIT for upstream BoltzGen code ### Model Specification | Field | Value | | ----- | ----- | | Repository | `oteam/boltzgen-adherence` | | Artifact Kind | `pipeline` | | Task | `design_generation` | | Architecture | `boltzgen-adherence` | | Entrypoint | `ofoldx.pipelines.design.DesignPipeline` | | Source Checkpoint | `boltzgen1_adherence.ckpt` | > [!NOTE] > Source checkpoint: `boltzgen1_adherence.ckpt`. ### Links - **Hub repository**: [oteam/boltzgen-adherence](https://huggingface.co/oteam/boltzgen-adherence) - **Upstream paper**: [BoltzGen: Toward Universal Binder Design](https://doi.org/10.1101/2025.11.20.689494) - **Upstream repository**: [BoltzGen](https://github.com/HannesStark/boltzgen) - **Source checkpoint release**: [https://huggingface.co/boltzgen/boltzgen-1](https://huggingface.co/boltzgen/boltzgen-1) - **Code**: [`ofoldx/pipelines/design.py`](https://github.com/OTeam-AI4S/OFoldX/tree/main/ofoldx/pipelines/design.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/boltzgen-adherence` 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/boltzgen-adherence") ``` When a matching processor is available, load it with `AutoProcessor.from_pretrained(...)` and pass the processed batch to the model. ### Interface - **Task**: `design_generation` - **Artifact kind**: `pipeline` - **Architecture**: `boltzgen-adherence` - **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 BoltzGen builds on the Boltz-2 data pipeline. The public training release includes `targets.zip` and `msa.zip` in the `boltzgen/boltzgen1_train` dataset and molecule dictionaries in `boltzgen/inference-data`; the full large-model recipe may also use additional distillation data. ### Training Procedure Upstream BoltzGen trains a diffusion objective over randomly cropped biomolecular structures with randomly selected design and conditioning regions. OFoldX converts the released adherence checkpoint; it does not run BoltzGen 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 BoltzGen work for the source checkpoint. If OFoldX supports your work, please also cite or link the OFoldX project repository. ```bibtex @article{stark2025boltzgen, author = {Stark, Hannes and Faltings, Felix and Choi, MinGyu and Xie, Yuxin and Hur, Eunsu and O'Donnell, Timothy John and Bushuiev, Anton and Ucar, Talip and Passaro, Saro and Mao, Weian and others}, title = {BoltzGen: Toward Universal Binder Design}, year = {2025}, doi = {10.1101/2025.11.20.689494}, 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 BoltzGen code. Review both OFoldX and upstream terms before redistribution or production use.
