# Third-party notices

The neural feature equations, decoding procedure, and packing algorithm follow
[dauparas/LigandMPNN](https://github.com/dauparas/LigandMPNN), copyright 2024 Justas
Dauparas, MIT. The upstream project states that its code and model parameters
are MIT-licensed. Its license is preserved in `vendor/LigandMPNN-LICENSE.txt`.
The exporter also copies the actual license from the locked upstream revision
into every generated bundle. New library and tooling code is MIT-licensed.

Packing geometry is derived, at conversion time, from the OpenFold residue
constants distributed inside LigandMPNN. This code descends from OpenFold and
AlphaFold and carries Apache-2.0 notices. The exporter retains the exact source
header and supplies `vendor/Apache-2.0.txt` with the derived geometry asset.
No OpenFold neural network is required, and no AlphaFold model weights are used.
The complete release archive includes the locked upstream sources and original
checkpoints, along with their license. Exported bundles include the derived
geometry constants and their license/notice files.

[webProteinMPNN](https://github.com/MurrellGroup/webProteinMPNN) is the architectural
reference for a library-only TypeScript/ONNX Runtime Web port. Its source code and
model artifacts are not redistributed in this repository.

[ONNX Runtime](https://github.com/microsoft/onnxruntime) is MIT-licensed. Installing
`onnxruntime-web` brings its own licenses and third-party notices; retain them
when redistributing its WASM/JavaScript binaries. The asset-copying tool copies
available license/notice files as well as runtime assets. Python export/test
packages retain their respective licenses and are not browser dependencies.

Because the ORT 1.22.0 npm archive omits the root license files, this release also
vendors `LICENSE` and `ThirdPartyNotices.txt` from Microsoft's exact `v1.22.0`
tag and copies them to `public/ort/`. Updating ORT requires matching notices.