PepBind3D

Curated peptide-HLA class I binding affinities paired with Rosetta FlexPepDock structural ensembles. 112,561 peptide-allele pairs across 95 HLA-A, -B and -C alleles, each with 25 docked decoy structures and per-decoy interface energies.

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

---

Dataset Summary

The interaction between peptide antigens and class I human leukocyte antigens (HLA) is a central determinant of CD8⁺ T cell recognition, and a foundational target for cancer immunotherapy, vaccine design and autoimmune disease research. Computational prediction of peptide-HLA (pHLA) binding has progressed rapidly with sequence-based machine learning, but structure-aware approaches remain limited by the scarcity of paired experimental affinity data and three-dimensional structural information.

This dataset pairs the two: every peptide-allele pair carries an experimental binding measurement and an ensemble of 25 Rosetta FlexPepDock decoys with their energy terms. Each pair gives a distribution over peptide conformations rather than a single pose. Pairs measured by both IC50 and Kd appear as separate rows, giving 118,985 rows over 112,561 pairs.

Intended uses:

• Training and benchmarking machine learning models for pHLA binding affinity prediction, particularly models that consume explicit structural features
• Evaluating computational docking and scoring methods against experimental binding data
• Downstream structural analysis: anchor residue mapping, conformational diversity, interface comparison between binders and non-binders

---

Dataset Statistics

Metric	Count
Peptide-allele pairs (silent files)	112,561
Total measurements (metadata rows)	118,985
Unique alleles	95 (HLA-A, -B, -C)
Unique peptides	25,622
Peptide lengths	7-15 residues
IC50 measurements	21,234
Kd measurements	97,751
Censored measurements	52,462 (44%)
Flagged entries	15
Self-templated pairs	370 (0.33%)
Total decoy structures	2,814,025
Approximate size	~155 GB

---

Dataset Structure

PepBind3D/
├── README.md
├── metadata.csv                        # Master table - one row per IEDB measurement
└── structures/
    ├── A0101/
    │   ├── F/
    │   │   └── FHEFLSSKL.silent
    │   ├── G/
    │   │   └── GILGFVFTL.silent
    │   └── ...
    ├── A0201/
    └── ...

Each silent file holds the 25-decoy ensemble for one peptide-allele pair, with Rosetta energy scores embedded alongside the coordinates. Paths use filesystem-safe allele names (A0101/ for HLA-A*01:01) and group files by the peptide's first residue; metadata.csv carries the standard form (A*01:01) and a pdb_dir column with the full path for every row.

There is one silent file per pair (112,561); the 118,985 rows exceed this because 6,423 pairs carry both an IC50 and a Kd measurement.

---

Quickstart

The structures total ~155 GB, so start with metadata.csv (40 MB) and pull only the silent files you need.

pip install huggingface_hub pandas

1. Get the metadata table.

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("kunikohunter/PepBind3D", "metadata.csv",
                       repo_type="dataset")
df = pd.read_csv(path)
print(len(df), "measurements over", df[["allele", "peptide"]].drop_duplicates().shape[0], "pairs")

2. Separate quantitative from censored measurements. Roughly 44% of rows are reported at or above an assay detection ceiling. These are not affinities and must not be used as numbers: a peptide recorded at 20,000 nM bound too weakly to measure, so the value is a floor, not a result. They are still information. They identify peptides that did not bind, which is the label a binder/non-binder model needs, and I_sc separates them from quantitative binders on its own (AUROC 0.68 for IC50, 0.64 for Kd). Which subset you want depends on the task: exclude the censored rows when regressing on affinity, keep both when classifying.

CEILINGS = {"IC50": ({20000, 50000, 70000}, 70000),
            "Kd":   ({5000, 10000, 20000},  20000)}

def split_by_censoring(df, assay):
    """-> (quantitative affinities, censored non-binders)"""
    exact, top = CEILINGS[assay]
    sub = df[(df.measurement_type == assay) & ~df.flagged]
    censored = sub.measurement_value.isin(exact) | (sub.measurement_value >= top)
    return sub[~censored], sub[censored]

ic50, ic50_censored = split_by_censoring(df, "IC50")   # 18,113 and 3,114 rows
print(ic50[["allele", "peptide", "measurement_value", "I_sc_best"]].head())

3. Download the structures for one pair. pdb_dir gives the path for every row.

row = ic50.iloc[0]
silent = hf_hub_download("kunikohunter/PepBind3D", row.pdb_dir, repo_type="dataset")

To pull a whole allele instead, use allow_patterns:

from huggingface_hub import snapshot_download
snapshot_download("kunikohunter/PepBind3D", repo_type="dataset",
                  allow_patterns="structures/A0201/*")

4. Extract PDB coordinates from a silent file with Rosetta:

extract_pdbs.linuxgccrelease -in:file:silent A0201/G/GILGFVFTL.silent

---

IEDB Data Curation

Binding affinity data were retrieved from a local copy of the IEDB bulk download (mhc_ligand_full.csv, accessed April 14, 2025) and processed with a custom Python pipeline (IEDBTestPipeline.py).

1. Retrieval and Assay Filtering

The full IEDB MHC ligand table was read in chunks and filtered to entries belonging to HLA-A, HLA-B or HLA-C 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 (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

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 rules 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 per allele and excluded from the cleaned dataset. These are the flagged = True entries in metadata.csv.

3. Sequence-Level Filtering

Peptide sequences containing the + character (used by IEDB to denote non-canonical or modified amino acids) were excluded from structure generation. Retained peptides consist of the 20 standard amino acids, 7 to 15 residues long.

---

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)
allele_compact	Filesystem-safe allele form used in paths (e.g. A0101)
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
source_version	Internal curation-batch label; carries no meaning for downstream use
flagged	Whether the entry was flagged during IEDB data cleaning
self_templated	Whether the pair's own crystal structure was available as a threading template. Exclude these when measuring structure-prediction accuracy
has_structures	Whether a structural ensemble is available for this pair
num_pdbs	Number of decoy structures in the silent file (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 / reweighted_sc_mean	Best and mean reweighted score (REU)
total_score_best / total_score_mean	Best and mean total_score (REU)
pep_sc_best / pep_sc_mean	Best and mean peptide score (REU)
pdb_dir	Relative path to the peptide silent file within the dataset

Which score to use. Each score is 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. REU is not a binding free energy - these rank poses within a modeling framework and are not thermodynamic quantities.

---

Computational Methods

Structures were generated with Rosetta FlexPepDock, 25 decoys per pair. Each peptide was threaded onto a length-matched template from a local MHC template database (SimpleThreadingMover), trimmed, given its receptor, relaxed (FastRelax, 5 repeats, ref2015), prepacked, and refined by flexible peptide docking (-pep_refine -nstruct 25 -ex1 -ex2aro). No score cutoff was applied: all 25 decoys are retained.

Where an allele had no experimental receptor structure, the α1/α2 domains were modeled with AlphaFold2 from the IPD-IMGT/HLA protein alignment.

All structures were generated with Rosetta 2024.09+release.06b3cf8.

FlexPepDock refinement holds the MHC backbone fixed, so decoys within a pair differ only in the peptide; MHC Cα coordinates are identical across an ensemble.

---

Validation

From the accompanying manuscript:

• Structural accuracy. 76 pairs have a matching experimental crystal structure. Measured on ensembles re-docked with self-matching templates excluded, the best-scoring decoy reaches a median peptide-backbone RMSD of 1.21 Å (IQR 0.93-1.68, 87% within 2 Å); the best decoy of each ensemble reaches 0.99 Å.
• Score-affinity relationship. Spearman ρ between best-decoy I_sc and log affinity is 0.31 (IC50, n = 18,113) and 0.18 (Kd, n = 48,395), censored values excluded. The relationship is real but modest: these scores are intended as input features for a downstream model, not as standalone affinity predictors.
• Binder discrimination. I_sc separates censored from quantitative measurements with AUROC 0.678 (IC50) and 0.639 (Kd), so the censored measurements carry information rather than being merely excluded.

---

Known Limitations and Caveats

Binding measurements

• IC50 values pile up at 20,000, 50,000 and 70,000 nM, and Kd values at 5,000, 10,000 and 20,000 nM, consistent with assay detection ceilings. Treat these as censored (≥ the reported value) rather than exact.
• Values above the highest ceiling are censored too, and testing only for the exact ceiling values keeps them: 1,366 IC50 rows sit above 70,000 nM and 8,651 Kd rows above 20,000 nM, including a 1,000,000 nM placeholder used for peptides with no measurable binding. Censored rows are ~44% of the dataset overall, and half of the Kd rows.
• Kd pools three different assays, distinguishable through assay_method. The competitive radioligand subset is ~1% censored; the two fluorescence subsets are 71-74% censored and centered about one log unit stronger. Stratify on assay_method or model the censoring explicitly.
• IC50 and Kd are not directly comparable and have not been converted between each other. Both are retained, with measurement_type recording which.

Structures

• All structures are computationally generated, and the best-scoring decoy is not guaranteed to be the native-like conformation.
• RMSD values embedded in silent files are computed against the threading template, not against any experimental reference, and are not a measure of model accuracy.
• self_templated marks 370 pairs (0.33%) whose own crystal structure was in the threading template library, so their structures were built from real coordinates of that exact peptide. They are the most accurate structures here, and they must be excluded when measuring structure-prediction accuracy or the result is inflated.

Coverage

• Allele representation is uneven: A*02:01 alone accounts for 9.2% of pairs, and seven alleles have fewer than five.
• HLA-C is included but sparsely sampled (1,861 pairs) and limited to 9- and 10-mer peptides, because the template library holds only 9- and 10-mer HLA-C structures.
• HLA class II alleles are not included, and peptides carrying post-translational modifications or non-standard residues were excluded before structure generation.

---

Citation

If you use this dataset, please cite:

[Manuscript citation - to be added upon publication]

Dataset DOI: https://doi.org/10.57967/hf/10466

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

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

Template and reference structures come from the RCSB PDB (public domain, CC0).

---

Code

Curation, structure generation and the validation analyses: https://github.com/kunikohunter/PepBind3D (archived at https://doi.org/10.5281/zenodo.22796306)

---

License

Released under CC BY 4.0: reuse and modification are permitted, including commercially, as long as the dataset is cited.