Metal Binding Protein Annotation (MbPA)

~101,500 proteins annotated with metal binding sites across 54 metal types from UniProt, curated into ML-ready datasets for two tasks: (1) per-residue metal binding site prediction and (2) per-protein metal type classification, with sequence-similarity-aware train/validate/test splits.

Quickstart Usage

Install HuggingFace Datasets package

Each subset can be loaded into python using the Huggingface datasets library. First, from the command line install the datasets library

$ pip install datasets

Optionally set the cache directory, e.g.

$ HF_HOME=${HOME}/.cache/huggingface/
$ export HF_HOME

then, from within python load the datasets library

>>> import datasets

Load Protein Classification Dataset

To load the per-protein multi-label metal classification dataset, use datasets.load_dataset(...):

>>> dataset_classification = datasets.load_dataset(
  path = "RosettaCommons/MbPA",
  name = "protein_classification",
  data_dir = "protein_classification")['train']

The dataset is loaded as a datasets.arrow_dataset.Dataset with multi-hot binary columns for 16 metal classes:

>>> dataset_classification
Dataset({
    features: ['accession', 'sequence', 'split', 'Mg', 'Zn', 'Fe', 'Ca', 'Mn', 'cluster', 'K', 'Na', 'Cu', 'Ni', 'Cd', 'Co', 'Hg', 'X', 'Mo', 'Pt'],
    num_rows: 101517
})

which is a column oriented format that can be accessed directly, converted into a pandas.DataFrame, or parquet format, e.g.

>>> dataset_classification.to_pandas()
>>> dataset_classification.to_parquet("classification.parquet")

To get just the training split:

>>> import pandas as pd
>>> df = dataset_classification.to_pandas()
>>> train_df = df[df['split'] == 'train']  # 79,418 proteins
>>> val_df = df[df['split'] == 'val']      #  9,823 proteins
>>> test_df = df[df['split'] == 'test']    # 12,276 proteins

Load Per-Residue Binding Site Predictions

Per-residue binding site labels provide a binary label for each residue in a protein sequence indicating whether it is a metal binding site:

>>> dataset_residue = datasets.load_dataset(
  path = "RosettaCommons/MbPA",
  name = "residue_prediction",
  data_dir = "residue_prediction")['train']

>>> dataset_residue
Dataset({
    features: ['accession', 'sequence', 'split', 'binding_labels', 'metal_labels'],
    num_rows: 101517
})

The binding_labels column is a binary string (e.g., "000100010...") of the same length as the sequence, where 1 indicates a metal binding residue. The metal_labels column is a JSON dictionary mapping each bound metal to its own per-residue binary string:

>>> import json
>>> row = dataset_residue.to_pandas().iloc[0]
>>> len(row['sequence']) == len(row['binding_labels'])
True
>>> metal_labels = json.loads(row['metal_labels'])
>>> metal_labels.keys()
dict_keys(['Mg'])

Load Cluster Assignments

Cluster assignments map each protein to a sequence-similarity cluster and data split, ensuring no data leakage between train/val/test:

>>> dataset_clusters = datasets.load_dataset(
  path = "RosettaCommons/MbPA",
  name = "cluster_assignments",
  data_dir = "cluster_assignments")['train']

>>> dataset_clusters
Dataset({
    features: ['accession', 'cluster_id', 'split'],
    num_rows: 101517
})

Dataset Details

Dataset Description

The Metal Binding Protein Annotation (MbPA) dataset is a curated collection of metal binding proteins from UniProt, organized for machine learning. The raw data comprises 54 metal-specific CSV files totaling ~116,000 rows representing ~101,500 unique proteins. Each protein entry includes its UniProt accession, amino acid sequence, annotated metal binding site positions, and bound metal types.

Curation pipeline:

1. Combine & Deduplicate: All 54 metal-specific source files are merged. Site and metal annotations (stored as parallel lists per protein) are exploded into (accession, position, metal_type) triples and deduplicated, then re-aggregated per protein to produce a unified dataset of 101,517 unique proteins with complete binding annotations.

2. Metal Class Selection: Metal types are filtered to those with at least 50 annotated proteins, yielding 16 metal classes: Mg (32,731), Zn (27,947), Fe (26,365), Ca (5,649), Mn (5,063), cluster (4,656), K (2,466), Na (2,372), Cu (1,674), Ni (1,150), Cd (352), Co (294), Hg (265), X (222), Mo (211), and Pt (53).

3. Per-Protein Labels: A multi-hot binary label matrix is constructed with one column per metal class. 8,952 proteins bind two or more distinct metals.

4. Per-Residue Labels: For each protein, a binary label array of length equal to the sequence is created, marking each binding residue. Site positions follow UniProt's 1-indexed convention and are converted to 0-indexed. Metal-specific per-residue labels are also provided. Out of ~43.7M total residues, ~403K (0.92%) are annotated as metal binding.

5. Sequence Similarity Clustering: Proteins are clustered using k-mer (k=3) MinHash signatures with Locality-Sensitive Hashing (128 hashes, 32 bands) to identify similar sequence pairs, followed by connected-component extraction on the similarity graph. This produces ~26,600 clusters. Clusters larger than 500 members are subdivided to enable balanced splitting.

6. Stratified Train/Val/Test Split: Clusters are stratified by their primary metal type and assigned to train/val/test splits at approximately 78%/10%/12%, ensuring that no cluster spans multiple splits (preventing data leakage from homologous sequences) and that all 16 metal classes are represented in every split.

License: cc-by-4.0

Dataset Sources

• Source Data: UniProt metal binding annotations
• Metadata: Cross-references to UniProt, AlphaFold, PDB, and NCBI Taxonomy IDs

Dataset Statistics

Statistic	Value
Total proteins	101,517
Total metal classes	16
Total sequence-similarity clusters	26,614
Total binding residues	403,376
Total residues	43,652,571
Binding residue ratio	0.92%
Multi-metal proteins	8,952
Mean sequence length	430
Median sequence length	371
Min sequence length	10
Max sequence length	34,350

Split Distribution:

Split	Proteins	Fraction
Train	79,418	78.2%
Validation	9,823	9.7%
Test	12,276	12.1%

Metal Class Distribution:

Metal	Proteins	Train	Val	Test
Mg	32,731	25,541	3,065	4,125
Zn	27,947	21,693	2,642	3,612
Fe	26,365	21,443	2,457	2,465
Ca	5,649	4,509	538	602
Mn	5,063	3,951	582	530
cluster	4,656	3,497	498	661
K	2,466	1,812	353	301
Na	2,372	1,559	188	625
Cu	1,674	1,389	135	150
Ni	1,150	900	113	137
Cd	352	—	—	—
Co	294	—	—	—
Hg	265	—	—	—
X	222	—	—	—
Mo	211	—	—	—
Pt	53	—	—	—

Exported Files

File	Description	Size
mbpa_protein_classification.csv	Per-protein multi-label metal classification with split assignments	46 MB
mbpa_residue_prediction.csv	Per-residue binding site labels with metal-specific annotations	134 MB
mbpa_cluster_assignments.csv	Protein-to-cluster and cluster-to-split mappings	1.6 MB
mbpa_dataset_stats.json	Summary statistics for the dataset	<1 KB

Uses

Per-residue metal binding site prediction: Given a protein sequence, predict which residues are involved in metal binding. This is a sequence labeling task analogous to secondary structure prediction, with extreme class imbalance (<1% positive).

Per-protein metal type classification: Given a protein sequence, predict which metal(s) it binds. This is a multi-label classification task across 16 metal types, with significant class imbalance (Mg: ~32K vs. Pt: 53).

Transfer learning and benchmarking: The sequence-similarity-based splits ensure that models are evaluated on their ability to generalize to novel protein families, not memorize homologous sequences.

Out-of-Scope Use

This dataset is derived from UniProt annotations, which may be incomplete. Absence of a binding annotation does not guarantee that a residue does not bind metals. The "cluster" and "X" metal classes are placeholders from the source data and may not represent specific metal ions. Caution should be used when interpreting predictions for rare metal classes (e.g., Pt with only 53 proteins).

Source Data

Metal binding annotations were obtained from UniProt, the Universal Protein Resource, which provides comprehensive, high-quality, and freely accessible protein sequence and functional information. The 54 metal-specific source files cover metals across the periodic table, from common biological cofactors (Mg, Zn, Fe, Ca, Mn, Cu) to rarer metal binders (Au, Pt, Ir, Os, Ru). Metadata cross-references to AlphaFold predicted structures, PDB experimental structures, and NCBI taxonomy identifiers are provided separately.

Dataset Card Authors

Jie Chen, Shajesh Sharma