Authors: Kuniko Hunter, Rocco Moretti, Jens Meiler, David G. Harrison
Vanderbilt University Medical Center / Vanderbilt University License: CC BY 4.0
Metadata source: Immune Epitope Database (IEDB)

Dataset Summary

This dataset combines experimentally measured peptide–HLA (pHLA) binding affinities from the IEDB with computationally generated 3D structures produced using Rosetta FlexPepDock. It covers 37 HLA class I alleles (HLA-A and HLA-B), 14,077 unique peptides, and 49,268 peptide–allele pairs with associated docked structures.

Each peptide–allele pair with a recorded binding measurement (Kd or IC50) is linked to an ensemble of 25 Rosetta FlexPepDock decoy structures, along with summary Rosetta energy statistics. Pairs measured by both IC50 and Kd assays appear as separate rows in the metadata table, giving 49,488 measurement rows in total.

Intended uses include:

Training and benchmarking machine learning models for pHLA binding affinity prediction
Evaluating computational docking and scoring methods against experimental binding data

Dataset Structure

pHLA_binding/
├── README.md
├── metadata.csv                    # Master table — one row per IEDB measurement
└── structures/
    ├── A0101/
    │   ├── FHEFLSSKL.silent
    │   ├── GILGFVFTL.silent
    │   └── ...
    ├── A0201/
    └── ...

Each silent file contains the full 25-decoy ensemble for a single peptide–allele pair, including embedded Rosetta energy scores. Allele directory names use filesystem-safe notation (e.g. A0101/ for HLA-A*01:01), and silent files are named after the peptide sequence. The allele column in metadata.csv contains the standard notation (A*01:01).

There is one silent file per unique peptide–allele pair (49,268 files); the 49,488 metadata rows exceed this because 220 pairs carry both an IC50 and a Kd measurement and appear as two rows linked to the same silent file.

Metadata Fields

Column	Description
`allele_iedb`	Allele name in IEDB format (e.g. `HLA-A_01_01`)
`allele`	Standard allele notation (e.g. `A*01:01`)
`peptide`	Peptide amino acid sequence
`peptide_length`	Length of peptide in residues
`measurement_type`	Assay type. Exact string values `IC50` or `Kd` (case-sensitive)
`measurement_value`	Quantitative binding measurement
`measurement_units`	Units of measurement (nM)
`assay_method`	Assay method as reported in IEDB
`assay_response`	Full assay response description from IEDB
`pubmed_id`	PubMed ID of the source publication
`parent_protein`	Source protein of the peptide
`protein_accession`	UniProt/GenBank accession of source protein
`source_organism`	Organism of origin
`assay_pdb_id`	Experimental PDB ID from IEDB (if available)
`flagged`	Whether the entry was flagged during IEDB data cleaning
`has_structures`	Whether a structural ensemble is available for this pair
`num_pdbs`	Number of decoy structures in the silent file (typically 25)
`I_sc_best`	Best (lowest) interface score across the 25-decoy ensemble (REU). Recommended primary metric — see below
`I_sc_mean`	Mean interface score across the ensemble (REU)
`reweighted_sc_best`	Best (lowest) reweighted score across the ensemble (REU)
`reweighted_sc_mean`	Mean reweighted score across the ensemble (REU)
`total_score_best`	Best (lowest) total_score across the ensemble (REU)
`total_score_mean`	Mean total_score across the ensemble (REU)
`rosetta_best_score`	Alias of `total_score_best`, retained for backward compatibility
`rosetta_mean_score`	Alias of `total_score_mean`, retained for backward compatibility
`pdb_dir`	Relative path to the peptide silent file within the dataset

Which score to use. Three Rosetta score metrics are provided, each summarized as the best (lowest) and mean across the 25-decoy ensemble:

I_sc (interface score) — the peptide–MHC interaction energy in isolation. This is the recommended primary metric; it showed the strongest association with experimental affinity.
reweighted_sc (reweighted score) — upweights peptide-relevant energy terms but retains full-pose energy.
total_score — the complete Rosetta pose energy. Because it is dominated by the internal energy of the MHC receptor (which varies little among peptides presented by the same allele), it is the least sensitive of the three for ranking binding.

All scores are in Rosetta Energy Units (REU); lower is more favorable.

Dataset Statistics

Metric	Count
Peptide–allele pairs (silent files)	49,268
Total measurements (metadata rows)	49,488
Unique alleles	37
Unique peptides	14,077
IC50 measurements	13,629
Kd measurements	35,859
Flagged entries	15
Total decoy structures	1,231,700
Approximate size	~65 GB

IEDB Data Curation

Binding affinity data were retrieved from a local copy of the IEDB bulk download (mhc_ligand_full.csv, accessed April 17, 2025) and processed using a custom Python pipeline (IEDBTestPipeline.py). The curation procedure is described below.

1. Retrieval and Assay Filtering

The full IEDB MHC ligand table was read in chunks and filtered to retain only entries belonging to HLA-A or HLA-B alleles with quantitative binding measurements. Rows were retained only if they reported one of two assay response types:

Half maximal inhibitory concentration (IC50)
Dissociation constant (Kd)

Variant labels in the IEDB (e.g., dissociation constant KD (~EC50), dissociation constant KD (~IC50)) were normalized to a single canonical label (dissociation constant (KD)) prior to filtering. Entries missing either a quantitative measurement value or assay units were dropped.

2. Per-Allele Deduplication

After filtering, records were grouped by allele. Within each allele, duplicate entries for the same epitope (identified by IEDB Epitope IRI) were resolved separately for IC50 and Kd measurements using the following decision rules, applied in order:

Scenario	Action
Two entries, one lacks a PubMed ID	Retain the entry with a PubMed ID; drop the other
Two entries, both lack PubMed IDs, values differ by ≤10 nM	Retain one entry (first occurrence)
Two entries, both lack PubMed IDs, values differ by >10 nM	Drop both entries
Two entries, both have PubMed IDs, values differ by <10 nM	Retain one entry (first occurrence)
Two entries, both have PubMed IDs, values conflict (≥10 nM difference)	Flag both for manual review
More than two entries	Drop entries lacking PubMed IDs first; if >2 remain, retain the entry closest to the median measurement value
Any other ambiguous case	Flag for manual review

Entries that could not be unambiguously resolved were written to a separate flagged_IEDB_data.csv file for each allele and excluded from the cleaned dataset. These correspond to the flagged = True entries in metadata.csv.

3. Sequence-Level Filtering

After deduplication, peptide sequences containing the + character (used by IEDB to denote non-canonical or modified amino acids) were excluded from the final FASTA output and downstream structure generation. Retained peptides consist of the 20 standard amino acids and range from 7 to 15 residues.

4. Outputs per Allele

For each allele, the pipeline produced:

{allele}_cleaned_IEDB_data.csv — deduplicated, unflagged entries retained for structure modeling
{allele}_flagged_IEDB_data.csv — ambiguous entries excluded from modeling but retained for reference
{allele}_IEDB_data.fasta — peptide sequences in FASTA format for input to Rosetta FlexPepDock

5. Manual Review Checkpoint

By default, the pipeline pauses after generating per-allele cleaned and flagged CSVs to allow manual literature review of flagged entries before structure generation proceeds. In this dataset, flagged entries were reviewed and those confirmed as erroneous or unresolvable were excluded. The --skipReview flag bypasses this checkpoint when manual review is not required.

Computational Methods

3D structures were generated using Rosetta FlexPepDock with 25 decoys per peptide–allele pair. Starting structures were prepared by threading the target peptide sequence onto a same-allele, same-length template pHLA structure, followed by Rosetta FastRelax and FlexPepDock prepacking; the relaxed complex was then refined by flexible peptide docking.

All decoy structures and their associated Rosetta energy scores are stored in per-peptide Rosetta silent files (structures/{allele}/{peptide}.silent), with scores embedded alongside the decoy coordinates. Three score metrics — the interface score (I_sc), reweighted score (reweighted_sc), and total_score — are summarized as best (lowest) and mean values across the 25-decoy ensemble and included in metadata.csv, all in Rosetta Energy Units (REU). Individual decoy structures can be extracted from silent files using extract_pdbs from the Rosetta toolkit.

Known Limitations and Caveats

Binding measurement caveats:

IC50 values show large spikes at 20,000 nM, 50,000 nM, and 70,000 nM, consistent with assay detection limits. These values should be treated as censored (i.e. greater than or equal to the reported value) rather than exact measurements.
Kd values show analogous saturation at 5,000 nM, 10,000 nM, and 20,000 nM for the same reason. In particular, approximately 71% of Kd measurements (25,571 records) are reported at exactly 20,000 nM. Users training on Kd data should account for this large censored fraction; only about 27% of Kd records represent quantitative affinities.
IC50 and Kd are not directly comparable and have not been normalized or converted between each other. Both are retained with measurement_type indicating which was recorded.

Structure caveats:

All 3D structures are computationally generated; no experimental validation has been performed for the majority of pairs. (A structural validation against 52 matched crystal structures is reported in the accompanying manuscript.)
Any RMSD values embedded in silent files are calculated relative to the threaded template input, not to any experimental reference structure, and should not be interpreted as a measure of model accuracy.
The best-scoring Rosetta decoy is not guaranteed to represent the native-like conformation.

Coverage caveats:

Allele representation is highly uneven: A*02:01 accounts for ~~9,100 pairs (~~18% of the dataset), while several alleles have fewer than 10 pairs.
The dataset covers HLA class I only (HLA-A and HLA-B). HLA-C and class II alleles are not included.
The assay_pdb_id column references experimental structures deposited in the PDB as recorded by IEDB; these structures are not included in this dataset.

Citation

If you use this dataset, please cite:

[Manuscript citation — to be added upon publication]

Dataset DOI: 10.57967/hf/9669

Experimental binding data are sourced from the Immune Epitope Database (IEDB), which is also available under CC BY 4.0. Please also cite IEDB:

Vita R, Mahajan S, Overton JA, et al. The Immune Epitope Database (IEDB): 2018 update. Nucleic Acids Research. 2019;47(D1):D339–D343. https://doi.org/10.1093/nar/gky1006

License

This dataset is released under CC BY 4.0. You are free to share and adapt the material for any purpose, provided appropriate credit is given.

Usage Example

import pandas as pd
from pathlib import Path

# Load metadata
df = pd.read_csv("metadata.csv")

# Detection-ceiling values to treat as censored
IC50_CEILINGS = {20000, 50000, 70000}

# Filter to unflagged, quantitative IC50 entries (excluding detection ceilings)
df_clean = df[
    ~df["flagged"]
    & (df["measurement_type"] == "IC50")
    & (~df["measurement_value"].isin(IC50_CEILINGS))
    & (df["measurement_value"] < 70000)
]

print(f"{len(df_clean)} quantitative IC50 entries")
print(f"Alleles: {df_clean['allele'].nunique()}")
print(f"Unique peptides: {df_clean['peptide'].nunique()}")

# Interface score (recommended primary metric) vs measured affinity
print(df_clean[["allele", "peptide", "measurement_value", "I_sc_best"]].head())

# Locate the silent file for a specific peptide-allele pair
silent_file = Path("structures/A0101/FHEFLSSKL.silent")
print(f"Silent file exists: {silent_file.exists()}")

# To extract individual PDB structures, use Rosetta's extract_pdbs:
# extract_pdbs.linuxgccrelease -in:file:silent structures/A0101/FHEFLSSKL.silent

Downloads last month: 12

Total file size:

65.2 GB