# webLigandMPNN model assets

Browser-ready ONNX bundles for the unofficial
[`webLigandMPNN`](https://github.com/MurrellGroup/webLigandMPNN) port. These
files were converted from checkpoints trained and released by the original
[`dauparas/LigandMPNN`](https://github.com/dauparas/LigandMPNN) authors.
MurrellLab performed the web conversion and did not train the models.

## Layout

`models/` contains four LigandMPNN sequence models and the learned side-chain
packer:

- `ligandmpnn_v_32_005_25`
- `ligandmpnn_v_32_010_25` (default sequence model)
- `ligandmpnn_v_32_020_25`
- `ligandmpnn_v_32_030_25`
- `ligandmpnn_sc_v_32_002_16` (side-chain packer)

Each directory is a self-contained web bundle. `manifest.json` lists the files
needed at runtime and records their sizes and SHA-256 hashes. Additional parity
and validation JSON files document the conversion checks. `catalog.json`
indexes every hosted file and its digest. No original PyTorch `.pt` checkpoint
is included.

## Provenance and licenses

`upstream.lock.json` pins upstream LigandMPNN revision
`26ec57ac976ade5379920dbd43c7f97a91cf82de` and records the original checkpoint
hashes and source URLs without redistributing those checkpoint files. The
LigandMPNN license is included in every bundle. The packing bundle also retains
the Apache-2.0 license and notice for geometry derived from OpenFold constants.

See `THIRD_PARTY_NOTICES.md` and the source repository documentation for full
attribution, validation scope, API usage, and scientific limitations.