README.md

---
license: mit
task_categories:
  - tabular-regression
tags:
  - protein
  - enzymes
  - pH
  - regression
  - biology
pretty_name: Optimal pH (EpHod pHopt)
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: seqs
      dtype: string
    - name: labels
      dtype: float64
    - name: Accession
      dtype: string
    - name: Organism
      dtype: string
    - name: EC Number
      dtype: string
  splits:
    - name: train
      num_bytes: 3461335
      num_examples: 7124
    - name: valid
      num_bytes: 372844
      num_examples: 760
    - name: test
      num_bytes: 954440
      num_examples: 1971
  download_size: 4258985
  dataset_size: 4788619
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: valid
        path: data/valid-*
      - split: test
        path: data/test-*
---

# Optimal pH (`optimal_ph`)

Enzyme optimum-pH (`pHopt`) regression dataset. Each row is a single enzyme
with its UniProt accession, amino acid sequence, and the experimentally
reported pH at which its catalytic activity is maximal. This is a
re-release of the exact train/valid/test split used in the EpHod paper
(Gado *et al.*, 2025), mirrored from the authors' Zenodo deposit and
reformatted for the Hugging Face `datasets` library.

- **Task:** regression (predict `labels` ∈ ℝ from `seqs`)
- **Input:** single-chain protein sequence (amino acids, length 32–1022)
- **Target:** optimum pH (typically ~1.0 to ~12.5, centered near 7)
- **n = 9,855** enzymes over 7,124 train / 760 valid / 1,971 test

## Columns

| column      | dtype  | description                                         |
|-------------|--------|-----------------------------------------------------|
| `seqs`      | string | single-letter amino acid sequence                   |
| `labels`    | float  | experimental pHopt (mean over BRENDA entries)       |
| `Accession` | string | UniProt accession                                   |
| `Organism`  | string | source organism (species) from UniProt              |
| `EC Number` | string | Enzyme Commission number                            |

## Splits

| split | rows  | note                                                       |
|-------|-------|------------------------------------------------------------|
| train | 7,124 | mmseqs2 clusters at 20% identity, 90% of non-test clusters |
| valid | 760   | 10% of non-test clusters                                   |
| test  | 1,971 | held-out 20% random sample of the full dataset             |

The split is inherited verbatim from the EpHod paper. The test set spans the
same pHopt distribution as training (not identity-held-out), so performance
on this split does **not** measure generalisation to dissimilar sequences by
itself. For that, use the paper's `Test <20% to Train` annotation (not
included here; available in the Zenodo deposit).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("GleghornLab/optimal_ph")
print(ds)
print(ds["test"][0])
```

## Known caveat: sequence-level leakage

The authors' published split contains **24 byte-identical sequences that
appear in more than one split** (20 train↔test, 4 valid↔test). These are
orthologs, paralogs, or reviewed/unreviewed pairs with different UniProt
accessions but identical amino acid strings. Some pairs have conflicting
pHopt labels or different EC annotations, which puts a ceiling on test RMSE
for any model that memorises training examples.

9,855 accessions collapse to 9,774 unique sequences; 24 of the duplicates
cross split boundaries.

If you need a leakage-free split, de-duplicate by sequence or re-cluster
train+valid+test jointly before training.

## Construction

This Hub dataset was built from the authors' raw `pHopt_data.csv` (Zenodo
deposit linked below) with the following one-shot transform:

```python
import pandas as pd
from datasets import Dataset, DatasetDict

SPLIT_MAP = {"Training": "train", "Validation": "valid", "Testing": "test"}
KEEP_COLS = ["seqs", "labels", "Accession", "Organism", "EC Number"]

df = pd.read_csv("pHopt_data.csv", index_col=0)
df = df.rename(columns={"Sequence": "seqs", "pHopt": "labels"})
df = df.drop(columns=["Sample Weight", "Test <20% to Train"])

dsd = DatasetDict({
    new_name: Dataset.from_pandas(
        df[df["Split"] == raw_name][KEEP_COLS].reset_index(drop=True),
        preserve_index=False,
    )
    for raw_name, new_name in SPLIT_MAP.items()
})
dsd.push_to_hub("GleghornLab/optimal_ph")
```

## Source data

- **BRENDA** (accessed 2022-05-25): 49,227 pH-optimum entries; 11,174 had
  UniProt sequence mappings. Multiple entries per sequence were averaged if
  within 1.0 pH unit, else dropped. Sequences <32 or >1022 aa were dropped,
  leaving 9,855.
- **UniProt**: canonical sequences for the retained accessions.
- **Split**: 20% random holdout → test; remaining clustered with MMseqs2 at
  20% identity, 90/10 cluster split → train/valid.

## License

Released under MIT to match the [EpHod code release](https://github.com/jafetgado/EpHod).
Underlying BRENDA and UniProt data are subject to their own terms
(BRENDA academic licence; UniProt CC BY 4.0); cite the sources below.

## Citations

EpHod paper (Nature Machine Intelligence, 2025):

```bibtex
@article{Gado2025EpHod,
  title   = {Deep learning prediction of enzyme optimum pH},
  author  = {Gado, Japheth E. and Knotts, Matthew and Shaw, Amber M. and
             Marks, Debora and Gauthier, Nicholas P. and Hopf, Thomas A. and
             Beckham, Gregg T.},
  journal = {Nature Machine Intelligence},
  year    = {2025},
  doi     = {10.1038/s42256-025-01026-6},
  url     = {https://www.nature.com/articles/s42256-025-01026-6}
}
```

bioRxiv preprint:

```bibtex
@article{Gado2023EpHodPreprint,
  title   = {Deep learning prediction of enzyme optimum pH},
  author  = {Gado, Japheth E. and Knotts, Matthew and Shaw, Amber M. and
             Marks, Debora and Gauthier, Nicholas P. and Hopf, Thomas A. and
             Beckham, Gregg T.},
  journal = {bioRxiv},
  year    = {2023},
  doi     = {10.1101/2023.06.22.544776},
  url     = {https://www.biorxiv.org/content/10.1101/2023.06.22.544776v2}
}
```

Zenodo data deposit (v1):

```bibtex
@dataset{Gado2024EpHodData,
  title     = {EpHod: Deep learning prediction of enzyme optimum pH (dataset)},
  author    = {Gado, Japheth E. and Knotts, Matthew and Shaw, Amber M. and
               Marks, Debora and Gauthier, Nicholas P. and Hopf, Thomas A. and
               Beckham, Gregg T.},
  publisher = {Zenodo},
  year      = {2024},
  doi       = {10.5281/zenodo.14252615},
  url       = {https://doi.org/10.5281/zenodo.14252615}
}
```

BRENDA (primary data source):

```bibtex
@article{Chang2021BRENDA,
  title   = {BRENDA, the ELIXIR core data resource in 2021: new developments
             and updates},
  author  = {Chang, Antje and Jeske, Lisa and Ulbrich, Sandra and Hofmann,
             Julia and Koblitz, Julia and Schomburg, Ida and Neumann-Schaal,
             Meina and Jahn, Dieter and Schomburg, Dietmar},
  journal = {Nucleic Acids Research},
  volume  = {49},
  number  = {D1},
  pages   = {D498--D508},
  year    = {2021},
  doi     = {10.1093/nar/gkaa1025}
}
```

UniProt (sequences):

```bibtex
@article{UniProt2023,
  title   = {UniProt: the Universal Protein Knowledgebase in 2023},
  author  = {{The UniProt Consortium}},
  journal = {Nucleic Acids Research},
  volume  = {51},
  number  = {D1},
  pages   = {D523--D531},
  year    = {2023},
  doi     = {10.1093/nar/gkac1052}
}
```