RelaxDB Dataset

The RelaxDB and RelaxDB-CPMG datasets are curated data of relaxation-dispersion NMR data. This dataset was used to evaluate our model Dyna-1. Both the model and the datasets were introduced in our paper "Learning millisecond protein dynamics from what is missing in NMR spectra".

This HF datasets hosts the files for the RelaxDB data. More information on analysis from the paper, evaluation of the dataset using Dyna-1, or the Dyna-1 model itself can be found on GitHub.

Overview

• relaxdb/relaxdb_data.csv: contains the entry ids, sequences, labels, and other metadata found in in the data/ metadata.

• relaxdb/af2_pdbs: contains all of the af2 models used for input and evaluation in .pdb format

• relaxdb/split_files: contains split files for the provided datasets.

• data/metadata: contains .npy files describing the proteins for RelaxDB and RelaxDB-CPMG datasets. A.xlsx file with metadata of the curated proteins can be found on GitHub.

• data/probs: contains saved frequencies from mBMRB-Train, stored for calculating dummy baselines.

Usage

To download the dataset, you can download the metadata table quickly with the huggingface datasets package:

from datasets import load_dataset

dataset = load_dataset("gelnesr/RelaxDB")
df = dataset["relaxdb"].to_pandas()

This dataset includes protein structure files (.pdb) stored separately from the tabular data. Each dataset row contains a column (e.g. pdb_file) with the relative path to the corresponding PDB file, for example relaxdb/af2_pdbs/4267.pdb

Metadata

Column descriptions of each protein entry in relaxdb/relaxdb_data.csv:

• id: entry identifier
• dataset: relaxdb or relaxdb-cpmg
• pdb: relative path of AF2 predicted structure
• protein_name: name of protein
• uniprot_id: UniProt accesion identifier
• curation_source: data source for experimental data
• sequence: protein sequence
• label: residue-level annotations aligned to the sequence
• assignments: source of assignments if not from original source ... and others!

Labels

The labels below indicate the type and availability of relaxation data for each residue. Each vocabulary maps a single-character label to a numeric value (-1 = no data / excluded, 0 = no positive signal, 1 = positive signal for the property being scored).

Vocab for mBMRB

Scores whether a residue has missing data (. is the only positive case).

Label	Value	Meaning
A	0	no special annotation (default)
.	1	missing data

Vocab for RelaxDB

Scores chemical exchange. Here both b (mixed fast/slow) and ^ (Rex) count as positive classes, along with ..

Label	Value	Meaning
t	-1	no data due to disordered terminus
x	-1	no data; R1/R2/NOE not reported
p	0	proline (not evaluated)
A	0	no special annotation (default)
v	0	fast internal motion
.	1	missing data
b	1	mixed fast and slow motion
^	1	chemical exchange detected (Rex)

Vocab for RelaxDB-CPMG

Scores CPMG-derived exchange. Uses a distinct label set (P, N, X, Y).

Label	Value	Meaning
t	-1	no data due to disordered terminus
P	0	proline (not evaluated)
N	-1	residue assigned, but CPMG data not reported
A	0	no special annotation (default)
.	1	missing data
X	1	exchange detected via Rex criteria
Y	0	exchange inferred from unsuppressed R2

Citation

If you are using our code, datasets, or model, please use the following citation:

@article {Dyna-1,
    author = {Wayment-Steele, Hannah K. and El Nesr, Gina and Hettiarachchi, Ramith and Kariyawasam, Hasindu and Ovchinnikov, Sergey and Kern, Dorothee},
    title = {Learning millisecond protein dynamics from what is missing in NMR spectra},
    year = {2025},
    doi = {10.1101/2025.03.19.642801},
    journal = {bioRxiv}
}

Acknowledgements

We thank Katie Henzler-Wildman, Magnus Wolf-Watz, Elan Eisenmesser, J. Patrick Loria, Marcellus Ubbelink, George Lisi, Sam Butcher, and Nicolas Doucet for sharing data. We thank Martin Stone for sharing the Indiana Dynamics Database data his group curated in 2000.