README.md

---
license: mit
library_name: onnxruntime
tags:
  - browser
  - onnx
  - wasm
  - webgpu
  - protein-design
  - proteinmpnn
  - ligandmpnn
---

# 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.