Predicted Structures · LiteFold
AlphaFoldDB
AlphaFoldDB prediction index — open database of predicted protein 3D structures with confidence scores, providing structural coverage for known protein sequences at scale.
What AlphaFoldDB is
AlphaFoldDB is a comprehensive tabular index of predicted protein three-dimensional structures and confidence metrics compiled from AlphaFold DB. Published by LiteFold, the resource indexes 246,689,516 structural predictions to expand structural coverage across known protein sequences for computational biology, medicine, and genomics.
The index encompasses sequences ranging in length from 5 to 4,186 residues, with an approximate median length of 278 residues and an average length of 328.55 residues. It reflects metadata across 48 bulk archive files, which include 16 proteome archives, 30 global-health archives, and two Swiss-Prot archives. Most indexed entries belong to version 6 of the underlying prediction models, giving researchers direct cross-references between UniProt accession identifiers and AlphaFold entry keys.
How it was made
The dataset is distributed as 14 Parquet files divided into a training partition of 222,017,452 rows across 12 files and a testing partition of 24,672,064 rows across two files. Partitioning is assigned deterministically by calculating the hash of each UniProt accession modulo 10; accession hashes resolving to bucket 0 are placed in the test split, while buckets 1 through 9 form the training split.
Each row contains the UniProt accession, the AlphaFold DB identifier, the latest model version number, the first and last residue indices in UniProt numbering, and the derived sequence length. Long sequences split across multiple segments are documented via parsed fragment numbers and boolean flags indicating multi-fragment predictions.
Key facts
- Total records
- 246,689,516
- Format
- Parquet (14 files)
- Modality
- Tabular, text
- Publisher
- LiteFold
- Split sizes
- Train: 222,017,452 rows; Test: 24,672,064 rows
- Sequence length range
- 5 to 4,186 residues (mean: 328.55; median: ~278)
- Version 6 records
- 241,070,489
- Bulk archives covered
- 48 (16 proteome, 30 global-health, 2 Swiss-Prot)
What you can use it for
- UniProt to structure cross-referencing
- Biologists can map UniProt sequence accessions to corresponding AlphaFold DB model identifiers and online entry URLs at proteome scale.
- Version-specific structure filtering
- Researchers can isolate entries by model release version, such as querying the 241,070,489 records generated under model version 6.
- Protein sequence length and fragmentation analysis
- Bioinformaticians can filter proteins by sequence length or identify fragmented predictions requiring multi-part structural assembly.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/AlphaFoldDB at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 152,131,733 rows of default/train
alphafold_id
string_text- min
- 12
- median
- 16
- mean
- 15.64
- max
- 19
first_residue_index
int- min
- 1
- median
- 1
- mean
- 1
- max
- 1
fragment_number
int- min
- 1
- median
- 1
- mean
- 1
- max
- 1
is_fragmented_prediction
bool- False100%
last_residue_index
int- min
- 16
- median
- 278
- mean
- 328.94
- max
- 2,700
latest_version
int- min
- 6
- median
- 6
- mean
- 6
- max
- 6
sequence_length
int- min
- 16
- median
- 278
- mean
- 328.94
- max
- 2,700
split_bucket
int- min
- 1
- median
- 5
- mean
- 5.00
- max
- 9
uniprot_accession
string_text- min
- 6
- median
- 10
- mean
- 9.64
- max
- 13
Strengths and limits
Good at
- Features a deterministic hash-based partition on UniProt accessions, facilitating consistent cross-validation and benchmarking.
- Stored in Parquet format, allowing high-throughput tabular querying with engines such as DuckDB, PyArrow, Polars, and Spark.
Watch out for
- Contains tabular metadata and identifiers rather than raw 3D atomic coordinates.
- Excludes or leaves fragment numbers unparsed for 5,619,027 rows with nonstandard or missing suffix identifiers.
- Upper sequence limit is capped at 4,186 amino acids across indexed models.
Questions
Can I look inside AlphaFoldDB without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is AlphaFoldDB under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What fields are included in each record?
Each row provides the UniProt accession, AlphaFold DB identifier, latest model version, initial and terminal residue indices, derived sequence length, fragment number, fragmentation flag, and the deterministic split bucket.
How is the train and test split determined?
Splits are assigned deterministically using the formula hash(uniprot_accession) modulo 10, assigning bucket 0 to test and buckets 1 to 9 to train.
How can the full coordinate models be retrieved from an entry?
The AlphaFold DB identifier in each row allows researchers to generate direct AlphaFold DB web entry URLs or match corresponding files within the 48 bulk archive sets.
Can this table be queried without loading the entire dataset into memory?
Yes, the Parquet files can be streamed iteratively or queried in place using columnar engines like DuckDB, PyArrow, Polars, or Spark.
How to cite it
Varadi, Mihaly, Stephen Anyango, Mandar Deshpande, and others. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Research 50, no. D1 (2022): D439–D444. doi: 10.1093/nar/gkab1061.
Written from AlphaFoldDB's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.