MurrellLab webports

Browser-ready model assets for unofficial web ports maintained by MurrellLab. This repository stores model files that are too large or unsuitable for the ordinary Git history of the corresponding software repositories.

webProteinMPNN

webProteinMPNN/ contains nine ONNX model bundles converted from the official ProteinMPNN checkpoints. They support ProteinMPNN, solubleMPNN, and CA-only ProteinMPNN inference through ONNX Runtime Web using WebAssembly or WebGPU.

These models were trained and released by the original ProteinMPNN authors. MurrellLab converted the checkpoints into a browser-oriented graph format; it did not train the underlying models. The source checkpoint SHA-256, upstream commit, exported-file hashes, tensor dimensions, and atom mode are recorded in each manifest.json. A collection-wide index is available at webProteinMPNN/catalog.json.

Code, usage documentation, and validation details are maintained at https://github.com/MurrellGroup/webProteinMPNN.

License and attribution

The upstream ProteinMPNN code and checkpoints are MIT licensed. The upstream notice is preserved in webProteinMPNN/LICENSE.ProteinMPNN, with additional provenance in webProteinMPNN/THIRD_PARTY_NOTICES.md. Please cite:

Dauparas J, Anishchenko I, Bennett N, et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science. 2022;378(6615):49-56. https://doi.org/10.1126/science.add2187

Intended use and limitations

The bundles are intended for research use in protein sequence design from protein backbone or CA-only coordinates. They preserve the behavior and limitations of the corresponding upstream checkpoint; conversion does not constitute new training or biological validation. Generated sequences require independent computational and experimental evaluation. These files are not a hosted inference API and are not suitable for clinical decision-making.

webLigandMPNN

webLigandMPNN/ contains four LigandMPNN sequence-design bundles and one learned side-chain packing bundle converted to ONNX for browser and Node.js inference through ONNX Runtime Web. It contains converted web assets only; the original PyTorch .pt checkpoints are not stored in this repository.

The models were trained and released by the original LigandMPNN authors. MurrellLab converted the checkpoints into browser-oriented graphs; it did not train the underlying models. Each bundle records the original checkpoint hash, locked upstream revision, converted-file hashes, dimensions, validation status, and applicable notices. A collection-wide index is available at webLigandMPNN/catalog.json.

Code, usage documentation, and validation details are maintained at https://github.com/MurrellGroup/webLigandMPNN.

License and attribution

The upstream LigandMPNN code and model parameters are MIT licensed. The locked upstream license and additional attribution are preserved under webLigandMPNN/. Packing geometry derived from OpenFold constants retains the applicable Apache-2.0 license and source notice in the packing bundle. Please cite the original LigandMPNN publication and repository as described by the upstream project.

Intended use and limitations

These bundles are intended for research use in ligand-conditioned protein sequence design and side-chain packing. Conversion does not constitute new training or independent biological validation. Generated structures and sequences require independent computational and experimental evaluation. These files are static model assets, not a hosted inference API, and are not suitable for clinical decision-making.