CAUSTIC

Conformation-Aware Uncertainty and Shift predicTion from proteIn Conformer ensembles

Predicts protein backbone NMR chemical shifts (H, HA, N, CA, CB, C') with per-residue uncertainties from a PDB, mmCIF or AlphaFold structure. A 741,024-parameter PaiNN equivariant graph neural network, trained on 3,433 BMRB-linked experimental structures with carbon-aggressive label-noise cleaning; NMR ensembles are predicted as the median over conformers.

Try it in your browser — no install needed.

Install

pip install caustic-nmr

Python ≥ 3.10; Linux, macOS and Windows. The model weights and calibrator are inside the package — nothing is downloaded at run time. PyTorch is a dependency (graph construction uses it; inference itself runs on ONNX Runtime).

60-second example

caustic --version                      # package version, model SHA-256, calibrator
caustic 1ubq.pdb                       # NEF to stdout
caustic 1ubq.pdb --format csv -o 1ubq_shifts.csv
caustic AF-P01112-F1-model_v6.cif      # AlphaFold model: pLDDT read from the B-factor column
caustic ensemble.pdb --ensemble median # NMR ensemble: median over conformers (default)

from caustic import predict_shifts_onnx
from importlib.resources import files

result = predict_shifts_onnx("1ubq.pdb", str(files("caustic.data") / "best_v2_carbons.onnx"))
result.mean["CA"]        # CA shifts (ppm), one per residue
result.std["CA"]         # per-residue sigma (ppm)
result.residue_names     # ['MET', 'GLN', 'ILE', ...]
result.provenance        # package version, model SHA-256, calibrator, date

Every output file carries the same provenance stamp in its header:

# caustic-nmr 0.4.0 model=best_v2_carbons.onnx sha256=ebc7bbc2fc59 calibrator=sa16_v2_carbons_slim date=2026-08-30T16:27:19Z

Inputs and outputs for ubiquitin and an AlphaFold model are in examples/.

Accuracy

Measured through this exact package (the released wheel, bundled weights, default settings) on the 735-protein held-out test split, against SPARTA+ 2.90, LEGOLAS and UCBShift2 in full mode (transfer module on), paired per residue:

MAE in ppm on the 344,487 residues all four methods predicted (731 proteins). Per-protein composite, paired, with protein-level bootstrap CIs (Δ in ppm): −25.9 % vs SPARTA+ (Δ −0.276 [−0.318, −0.243]), −35.3 % vs LEGOLAS (Δ −0.429 [−0.450, −0.412]), −11.5 % vs full UCBShift2 (Δ −0.102 [−0.127, −0.083]) — every per-nucleus CI excludes zero. 67 of 693 test structures are in UCBShift2's own reference database (where its transfer module excels); they are included, so the UCBShift2 comparison is conservative against CAUSTIC. Protocol, competitor versions, crash accounting and the fairness slices: docs/BENCHMARKS.md; regenerate everything with python benchmarks/rescore.py --bootstrap 2000.

Where the details are

Output formats

Things to know

• One chain per call (--chain); other chains, ligands and waters are not in the graph.
• Missing backbone H/HA are synthesised from geometry (X-ray and AlphaFold inputs).
• Thirty non-standard residue names (MSE, HYP, SEP, TPO, PTR, CSO, PCA, …) are mapped to their parent residue; the modification itself is invisible to the model.
• Temperature is fixed at 298 K at inference; pH is not an input.
• σ comes from the network's uncertainty head; there is no pLDDT-dependent widening and no post-hoc isotonic calibration. See LIMITATIONS.md §4–5.
• The shipped model is ONNX-only; the --backend torch path cannot load it (CHANGELOG, Known issues).

Licence

Code: MIT. Model weights and calibrator (caustic/data/): CC BY 4.0 — attribution CAUSTIC model weights, Maximilian Zinke, 2026, https://github.com/maxzinke/caustic-nmr. Training data come from the BMRB and the wwPDB; no BMRB records are redistributed (DATA.md §6).

Citation

See CITATION.cff (GitHub's "Cite this repository" button renders it).