README.md

---
pretty_name: AlphaFoldDB Prediction Index
license: cc-by-4.0
tags:
- biology
- proteins
- protein-structure
- alphafold
- alphafolddb
- parquet
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/*.parquet
  - split: test
    path: data/test/*.parquet
---

# AlphaFoldDB Prediction Index

AlphaFoldDB is an open database of predicted protein 3D structures with confidence scores, massively expanding structural coverage for known protein sequences.

## Splits

| Split | Rows | Parquet files |
|---|---:|---:|
| train | 222,017,452 | 12 |
| test | 24,672,064 | 2 |
| total | 246,689,516 | 14 |

The split is deterministic: `hash(uniprot_accession) % 10 == 0` goes to `test`; buckets `1` through `9` go to `train`.

## Dataset Statistics

| Metric | Value |
|---|---:|
| Rows | 246,689,516 |
| Minimum sequence length | 5 |
| Approximate median sequence length | 278 |
| Mean sequence length | 328.55 |
| Maximum sequence length | 4,186 |
| Rows without parsed fragment number | 5,619,027 |

Latest-version distribution:

| Latest version | Rows |
|---|---:|
| 1 | 5,271,725 |
| 2 | 347,302 |
| 6 | 241,070,489 |

The mirrored `download_metadata.json` describes 48 bulk archive files: 16 proteome archives, 30 global-health archives, and 2 Swiss-Prot archives.

## Load With `datasets`

```python
from datasets import load_dataset

ds = load_dataset("LiteFold/AlphaFoldDB")
print(ds)

row = ds["train"][0]
print(row)
```

Load one split directly:

```python
from datasets import load_dataset

train = load_dataset("LiteFold/AlphaFoldDB", split="train")
test = load_dataset("LiteFold/AlphaFoldDB", split="test")
```

Stream rows without materializing the full table locally:

```python
from datasets import load_dataset

streamed = load_dataset("LiteFold/AlphaFoldDB", split="train", streaming=True)
first_row = next(iter(streamed))
```

Construct an AlphaFold DB entry URL from a row:

```python
entry_url = f"https://alphafold.ebi.ac.uk/entry/{row['alphafold_id']}"
```

Filter to current v6 entries:

```python
from datasets import load_dataset

train = load_dataset("LiteFold/AlphaFoldDB", split="train")
v6_train = train.filter(lambda row: row["latest_version"] == 6)
```

For large jobs, prefer streaming or process the Parquet files with a columnar engine such as DuckDB, PyArrow, Polars, or Spark.

## Columns

| Column | Description |
|---|---|
| `uniprot_accession` | UniProt accession from `accession_ids.csv`. |
| `alphafold_id` | AlphaFold DB identifier, for example `AF-Q5VSL9-F1`. |
| `latest_version` | Latest available AlphaFold DB model version for the entry. |
| `first_residue_index` | First residue index in UniProt numbering. |
| `last_residue_index` | Last residue index in UniProt numbering. |
| `sequence_length` | Derived as `last_residue_index - first_residue_index + 1`. |
| `fragment_number` | Parsed `F<number>` suffix from `alphafold_id`, nullable when the suffix is absent or nonstandard. |
| `is_fragmented_prediction` | Whether `fragment_number` is greater than 1. |
| `split_bucket` | Deterministic bucket from `hash(uniprot_accession) % 10`; bucket 0 is test. |


# Citation

```
@article{varadi2022alphafolddb,
  title   = {{AlphaFold} Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models},
  author  = {Varadi, Mihaly and Anyango, Stephen and Deshpande, Mandar and others},
  journal = {Nucleic Acids Research},
  volume  = {50},
  number  = {D1},
  pages   = {D439--D444},
  year    = {2022},
  doi     = {10.1093/nar/gkab1061}
}
```