---
license: odc-by
dataset_info:
  features:
  - name: id
    dtype: string
  - name: name
    dtype: string
  - name: sequence
    dtype: string
  - name: author
    dtype: string
  - name: designMethod
    dtype: string
  - name: evaluations
    list:
    - name: type
      dtype: string
    - name: metric
      dtype: string
    - name: valueType
      dtype: string
    - name: unit
      dtype: string
    - name: value
      dtype: string
    - name: target
      dtype: string
    - name: method
      dtype: string
    - name: measuredAt
      dtype: string
  splits:
  - name: proteins
    num_bytes: 30683901
    num_examples: 4069
  download_size: 3703301
  dataset_size: 30683901
configs:
- config_name: default
  data_files:
  - split: proteins
    path: data/proteins-*
size_categories:
- 1K<n<10K
pretty_name: Proteinbase
tags:
- biology
- proteins
- protein-binding
- ppi
---
# Proteinbase

Experimentally validated protein-protein interaction data from
[Adaptyv Bio](https://adaptyvbio.com), alongside computational predictions.
Published from <https://proteinbase.com/download>.

```python
from datasets import load_dataset

ds = load_dataset("adaptyvbio/proteinbase")["proteins"]

for ev in ds[0]["evaluations"]:
    print(ev["metric"], ev["method"], ev["value"], ev["unit"])
```

This is one unpartitioned corpus, so there is a single split named
`proteins` rather than the usual `train`/`test` division.

This snapshot is tagged `v2026-10-12`. Pin it for reproducibility:

```python
ds = load_dataset("adaptyvbio/proteinbase", revision="v2026-10-12")["proteins"]
```

## Columns

| Column | Description |
|---|---|
| `id` | Stable protein identifier, safe to treat as a primary key. |
| `name` | Optional reference label; empty if none was supplied. |
| `sequence` | Amino-acid sequence. |
| `author` | Creator handle when available. |
| `designMethod` | Slug of the design pipeline that generated the protein. |
| `evaluations` | List of measurements for this protein (see below). |

### `evaluations`

| Field | Description |
|---|---|
| `type` | `computational`, `experimental`, or `curated`. |
| `metric` | Metric slug, e.g. `kd`, `tm-score`, `expressed`. |
| `target` | Binding target slug when applicable, otherwise empty. |
| `valueType` | Parse hint: `numeric`, `boolean`, `label`, or `json`. |
| `unit` | Unit symbol such as `nM`; empty when the metric is unitless. |
| `value` | The measurement, encoded as a string (see note). |
| `method` | Tool or protocol that produced it: `esmfold`, `boltz2`, `bli`, `spr`. |
| `measuredAt` | ISO timestamp of the evaluation. |

**`value` is always a string.** It is a union of number, boolean, label, and
structured JSON in the source data, and Arrow has no union type. Parse it
according to `valueType`: `numeric` with `float()`, `boolean` against
`"true"`/`"false"`, `json` with `json.loads()`, `label` as-is.

## When the same metric appears more than once

Computational entries are deduplicated: where a tool has been run more than
once for the same protein and metric, only the most recent result is included.

Experimental entries are **not** deduplicated. Several entries sharing a
`metric` and `method` are genuine replicate measurements and should be
aggregated (e.g. averaged) rather than treated as duplicates.

Entries sharing a `metric` but differing in `method` are different tools
measuring the same quantity, and are never merged.

## License

Open Data Commons Attribution License (ODC-By). Include this text in any work
building on Proteinbase:

> "This work used Proteinbase by Adaptyv Bio under ODC-BY license"