IEDB Assay Export

This repository contains a mirrored IEDB XML export in iedb_export.zip. The added Parquet files make the export loadable in the Hugging Face Dataset Viewer and the datasets API.

The default table is assay-level: each row represents one B-cell, T-cell, MHC-binding, or MHC-ligand-elution assay extracted from the XML. Reference-level article metadata and epitope-level metadata are repeated on each assay row so the table can be filtered and viewed directly.

IEDB states that its database content is freely available and licensed under Creative Commons Attribution 4.0 International; users should cite IEDB when using the resource.

Splits

Split	Rows
train	6,705,467
test	745,344
total	7,450,811

The split is deterministic and epitope-aware: sha256(epitope_id) % 10 == 0 goes to test; buckets 1 through 9 go to train. If an assay lacks an epitope ID, the script falls back to reference ID, assay ID, then XML filename.

Dataset Statistics

Metric	Value
XML files parsed	26,785
XML parse-error files	0
Assay rows	7,450,811
Unique references with assays	26,785
Unique epitopes with assays	2,320,500
Uncompressed XML bytes	26.32 GB
Compressed XML bytes	605.58 MB

Assay category counts:

Assay category	Rows
MhcLigandElution	4,635,502
BCell	1,419,468
MhcBinding	825,450
TCell	570,391

Top qualitative measurements:

Measurement	Rows
Positive	5,239,946
Negative	1,967,175
Positive-Low	118,346
Positive-High	69,648
Positive-Intermediate	55,696

Top epitope chemical types:

Chemical type	Rows
Peptide from protein	6,773,619
Peptide, no natural source	591,652
Discontinuous protein residues	51,083
Other Non-Sequence	14,279
Carbohydrate fragment	5,650

Load With datasets

from datasets import load_dataset

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

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

Load one split directly:

from datasets import load_dataset

train = load_dataset("LiteFold/IEDB", split="train")
test = load_dataset("LiteFold/IEDB", split="test")

Stream rows without materializing the full table locally:

from datasets import load_dataset

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

Filter to positive T-cell assays:

from datasets import load_dataset

train = load_dataset("LiteFold/IEDB", split="train")
positive_tcell = train.filter(
    lambda row: row["assay_category"] == "TCell" and row["is_positive"]
)

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

Main Columns

Column	Description
xml_file	Source XML file inside iedb_export.zip.
reference_id, pubmed_id, article_year, article_title, journal_title, authors	Reference/article metadata.
epitope_id, epitope_name	IEDB epitope identifiers and names.
epitope_structure_type, epitope_chemical_type	Parsed XML structure type and chemical type.
linear_sequence, linear_sequence_length	Linear peptide/protein sequence when present.
discontinuous_residues	Discontinuous residue string when present.
starting_position, ending_position	Natural-sequence coordinates when present.
source_organism_id, source_molecule_genbank_id	Source organism and molecule identifiers parsed from epitope structure.
assay_category, assay_id, assay_type_id	Assay type and identifiers.
qualitative_measurement, is_positive, quantitative_measurement	Assay outcome fields.
host_organism_id, host_sex, host_age, disease_state	Host and disease context when present.
mhc_allele_id, mhc_allele_types_present	MHC information when present.
cell_type, cell_tissue_type, cell_culture_conditions	Effector-cell context when present.
antigen_evidence_code, immunogen_evidence_code, assay_comments	Additional assay curation metadata.
split_bucket	Deterministic hash bucket; bucket 0 is test.

Source Files Used

• iedb_export.zip

The XML export is preserved in this repository. The preparation script is included at scripts/prepare_iedb_dataset.py.