SLAE-CS and SLAE-ERR

Per-residue backbone amide ¹⁵N chemical-shift prediction from protein structure (SLAE-CS), and a per-residue estimate of its error (SLAE-ERR), both built on the SLAE all-atom structure encoder. Companion model repository of "Learning the limits of protein structure models with NMR chemical shifts" (bioRxiv, 2026). Code: WaymentSteeleLab/SLAE-CS-and-SLAE-ERR. Data: rBMRB.

Models in this repository

Folder Model Training data Runs
SLAE-CS-PANAV/seed01 … seed10 SLAE-CS on PANAV re-referenced shifts; use the mean of the 10 seeds rBMRB-PANAV 10 seeds
SLAE-ERR/ SLAE-ERR: predicts the absolute error of the SLAE-CS-PANAV ensemble for each residue rBMRB-ERROR 1 model

Each folder contains model.safetensors (the weights, 14.5 M parameters, 58 MB) and config.json (training hyper-parameters, epoch, and the file name and sha256 of the checkpoint the weights were taken from). ckpt_inventory.csv lists the sha256 of every source checkpoint, model.safetensors and config.json.

Usage

The weights are used with the code repository, which provides predict.py, the parts of the SLAE encoder package that SLAE-CS uses, and the environment (containers/slae/). A Google Colab notebook (notebooks/SLAE-CS_SLAE-ERR_colab.ipynb) installs everything on a Colab GPU and runs the test and a SLAE-CS and SLAE-ERR prediction.

hf download hws-lab/SLAE-CS-and-SLAE-ERR --local-dir weights

SLAE-CS

Predicts the backbone amide ¹⁵N chemical shift of each residue. By default all 10 seeds are run and averaged.

python predict.py --weights weights/SLAE-CS-PANAV --pdb my_protein.pdb --out my_protein_N   # 10-seed ensemble
python predict.py --weights weights/SLAE-CS-PANAV --pdb structures/ --out pred_N            # every *.pdb in a folder
python predict.py --weights weights/SLAE-CS-PANAV/seed01/model.safetensors --pdb x.pdb --out x_seed01   # one seed
  • Input: one protein per PDB file (the models were trained on single-chain proteins). Only the first model of a multi-model file is used; to predict every model of an NMR ensemble, split it first (scripts/inference_custom/split_nmr_models.py).
  • Output <out>.csv: one row per residue with protein, index, chain, resnum, resname, N_mean, N_std and the prediction of each seed (ppm). N_mean is the 10-seed mean, N_std their spread. Prolines are left out (no amide N).
  • Output <out>.json: pred [seeds][residues], mask, mean, std per protein.
  • --device auto|cuda|cpu (default auto). Inference runs at batch size 1.

SLAE-ERR

Predicts the absolute error of the SLAE-CS ensemble mean for each residue.

python predict.py --weights weights/SLAE-CS-PANAV --error-model weights/SLAE-ERR --pdb my_protein.pdb --out my_protein
  • Input: the same PDB file. predict.py first runs the 10 SLAE-CS-PANAV seeds; SLAE-ERR then uses the structure plus, for every residue, the standard deviation of the 10 predictions and the 10 predictions sorted and centred on their mean. All 10 seeds are required.
  • Output: as for SLAE-CS, plus N_abs_error (CSV) / abs_error (JSON): the predicted |N_mean − experimental shift| for each residue (ppm).

Training

  • Encoder: fine-tuned from the public SLAE autoencoder.ckpt (Chen et al., ICML 2026).
  • SLAE-CS-PANAV: 10 seeds trained on rBMRB-PANAV; for each seed, the checkpoint with the lowest validation ¹⁵N RMSE is kept.
  • SLAE-ERR: trained on rBMRB-ERROR (out-of-fold ensemble predictions for the training set) to predict the absolute error of the SLAE-CS-PANAV ensemble mean.

Full details are in the Methods of the paper; hyper-parameters are in each config.json.

Evaluation (rBMRB-PANAV test set)

Re-computed with the weights in this repository, on 407 proteins and 42,153 residues in rBMRB-PANAV test set. Residues are scored where both the experimental mask and the prediction mask (PANAV flags + tag labels) are true; prolines are excluded.

Model Metric Value
SLAE-CS-PANAV, single seeds ¹⁵N RMSE (ppm) 1.811 – 1.838
SLAE-CS-PANAV, 10-seed mean ¹⁵N RMSE (ppm) 1.745
SLAE-ERR per-residue MAE between predicted and actual absolute error (ppm) 0.760

Limitations

  • Training data. Trained on single-entity proteins from the BMRB (no bound partners, ligands or metal ions) measured at pH 4.5–8.5 and 293–313 K. Predictions outside these conditions are extrapolations.
  • Input structures. Inputs were AlphaFold2 structures. Results with experimental or other predicted structures may differ, and a structure that does not represent the measured conformational state leads to large errors (see the paper).
  • Coverage. Only the backbone amide ¹⁵N shift is predicted. Prolines are excluded.

License

The SLAE-CS weights and code are released under the Apache License 2.0. They include a fine-tuned copy of the SLAE encoder, which is also Apache-2.0 (Copyright 2025 Yilin Chen, Stanford University); its licence is included as LICENSE-SLAE. Training data: rBMRB (CC-BY-4.0).

Citation

If you use these models, please cite our paper and SLAE:

@misc{zhao_learning_2026,
    title = {Learning the limits of protein structure models with {NMR} chemical shifts},
    author = {Zhao, Cizhang and El Nesr, Gina and Chen, Yilin and Wayment-Steele, Hannah K.},
    publisher = {bioRxiv},
    doi = {10.64898/2026.11.12.766294},
    url = {https://www.biorxiv.org/content/10.64898/2026.11.12.766294v1},
    month = nov,
    year = {2026},
}
  1. Zhao, C., El Nesr, G., Chen, Y. & Wayment-Steele, H. K. Learning the limits of protein structure models with NMR chemical shifts. bioRxiv 2026.11.12.766294 (2026). doi:10.64898/2026.11.12.766294
  2. Chen, Y., Lu, T., Zhao, C., Wayment-Steele, H. & Huang, P.-S. SLAE: Strictly Local All-atom Environment for Protein Representation. in Forty-third International Conference on Machine Learning (2026).

The training data are in rBMRB; please also cite the data sources listed there (BMRB and others) when you use the dataset.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train hws-lab/SLAE-CS-and-SLAE-ERR