---
license: cc-by-4.0
library_name: onnx
pipeline_tag: other
tags:
  - protein-nmr
  - chemical-shifts
  - graph-neural-network
  - painn
  - onnx
---

# CAUSTIC model weights

Model parameters and the post-prediction calibrator for
[CAUSTIC](https://github.com/maxzinke/caustic-nmr), which predicts protein backbone
chemical shifts (H, HA, N, CA, CB, C') from a 3D structure.

**You do not need to download these by hand.** The same bytes are bundled inside the
`caustic-nmr` Python package (`pip install caustic-nmr==0.4.0`). This repository is
the citable, tagged, DOI-carrying home of the exact files a package version uses.

| File | Bytes | SHA-256 |
|---|---|---|
| `best_v2_carbons.onnx` | 3,045,952 | `ebc7bbc2fc59327a50384105207958948cd90d7b5c7ea1ec2906b473e02948b2` |
| `sa16_calibrator_v2.json` | 846 | `32d277df600a8e6b2f84e2f7ccaaaf6d9de1332664df61e51b36f31fdf267bf4` |

Tag `v0.4.0` of this repository corresponds to `caustic-nmr` 0.4.0.

## What the files are

- `best_v2_carbons.onnx` — PaiNN equivariant graph neural network (741,024 parameters,
  ONNX opset 17) trained on BMRB-linked experimental structures with carbon-aggressive
  label-noise cleaning. Architecture, features and training recipe:
  [docs/METHOD.md](https://github.com/maxzinke/caustic-nmr/blob/main/docs/METHOD.md).
- `sa16_calibrator_v2.json` — per-nucleus global offsets and cysteine CB modifiers applied
  after prediction (the "slim" SA16 v2 calibrator).

Training data, split protocol and licences:
[docs/DATA.md](https://github.com/maxzinke/caustic-nmr/blob/main/docs/DATA.md).
Benchmark protocol and numbers:
[docs/BENCHMARKS.md](https://github.com/maxzinke/caustic-nmr/blob/main/docs/BENCHMARKS.md).

## Licence

These files are released under **CC BY 4.0** (see
[LICENSE-WEIGHTS](https://github.com/maxzinke/caustic-nmr/blob/main/LICENSE-WEIGHTS)).
The package code is MIT. Attribution: *CAUSTIC model weights, Maximilian Zinke, 2026,
https://github.com/maxzinke/caustic-nmr*.

## How to cite

Zinke, M. *CAUSTIC: conformation-aware uncertainty and shift prediction from protein
conformer ensembles*. Zenodo. https://doi.org/10.5281/zenodo.22213167 (concept DOI, resolves to the
latest version). Machine-readable metadata:
[CITATION.cff](https://github.com/maxzinke/caustic-nmr/blob/main/CITATION.cff).