Neoantigen Discovery Dataset (NDD)

---

Overview

In neoantigen research, different strategies serve complementary roles: threshold-based approaches are effective for filtering candidates using known rules, while deep learning methods require large-scale, standardized datasets to uncover complex patterns and improve generalization.

NDD addresses this need by providing a curated, standardized, and feature-rich dataset that enables the development, training, and benchmarking of new algorithms, while also facilitating integration with experimental evidence.

Key highlights include

• Integration of data from literature reports and public resources such as IEDB, TCGA, and CEDAR. Based on the positive data, we scientifically generate corresponding negative data, which has greatly improved the accuracy of prediction.
• Coverage of core experimental evidence alongside predicted features (binding affinity, stability, anchor mutations, driver status, etc.).
• A structured design that supports both peptide-level predictions and patient-level cohort analyses.
• Preservation of assay details to ensure reliable immunogenicity labels and reproducibility.

NDD is not just a file collection, but a community resource for advancing model development, evaluation, and integration with experimental data. The current release, NDD v0.2, marks the starting point of an ongoing effort to expand data coverage, enrich feature annotations, and promote collaboration across the research community.

---

Data Sources & Curation

NDD entries are primarily curated from original literature reports, manually reviewed and standardized for consistency. To ensure transparency and traceability, each record includes PMID/DOI identifiers, allowing users to directly access the corresponding publications.

Meanwhile, we have also sequenced and analyzed the actual human biological data from relevant hospitals, generating valid training data for therapeutic approaches ranging from mutation to immunogenicity and clinical results.

The curation process draws on the frameworks established by public databases such as IEDB and CEDAR, while also integrating cross-references to resources like TCGA. This approach ensures that NDD combines the breadth of literature-derived evidence with the rigor of structured annotation.

Standardization Principles

• Mutation nomenclature: Reported using HGVS format (e.g., KRAS p.G12D). Mutation descriptions are preserved as given in the original publications, which may reference different genome builds (commonly hg19 or hg38). In v0.1, all entries retain their original reporting style; future versions will unify genome references to improve comparability.
• Cancer type classification: Manually normalized and mapped to widely used taxonomies such as IntOGen and TCGA, ensuring consistent categorization across diverse studies.
• HLA typing: Harmonized to four-digit resolution (e.g., HLA-A02:01), with aliases and shorthand notations consolidated.

Quality Control

• Missing values are explicitly marked as NA.
• Consistency checks combine manual review with automated scripts to validate key fields.
• Future improvements will include automated QC pipelines to further minimize potential errors.
• Source traceability is preserved for every entry, ensuring reproducibility and enabling direct comparison with original publications.

Detailed documentation of the tools, versions, and parameters used for feature annotation is provided in docs/predicted-features.md.

---

Fields & Features

NDD provides information at three levels: basic fields, predicted features, and patient-level metadata. Together, these dimensions make the dataset suitable for modeling and exploratory research.

• Basic fields capture essential information from original reports, including:
  patient ID, tumor type, gene/mutation (HGVS), mutant and wild-type peptides, HLA allele, experimental method.
  These fields form the foundation of each entry.

• Predicted features extend each record with annotations widely used in neoantigen research, such as:
  binding affinity and rank (NetMHCpan), stability (NetMHCstabpan), recognition likelihood (PRIME, BigMHC_IM), cleavage and transport scores (NetChop, NetCTLpan), differential agretopicity index (DAI), anchor mutation position, TCGA cancer expression (TPM median), and driver gene status (IntOGen).
  These features make the dataset directly applicable to machine learning tasks.

• Patient-level metadata (when available) include additional information such as extended HLA typing, clinical outcomes, or treatment details. In the current release these data are limited and not yet complete, but we recognize their importance and will continue to expand and refine this part of the dataset in future versions.

Field definitions and detailed documentation are provided in the docs/ directory, including:

• field-schema.md — definitions of basic fields
• predicted-features.md — list and description of predicted features
• patient-metadata.md — definitions of patient-level metadata (when available)

---

Usage & Format

The dataset is released in tab-separated values (TSV, UTF-8 encoded) format.

• Main file: data/ndd_v0.2.tsv
  Each row represents a mutation–HLA–peptide entry, combining both basic fields and predicted features.

• Patient-level metadata: data/patient_metadata_v0.2.tsv
  Contains information at the patient dimension, such as extended HLA typing, clinical outcomes, and treatment details when available.

Note: Patient IDs are anonymized to ensure privacy but remain consistent across files, enabling cohort-level analysis.

Limitations

• Patient-level metadata are currently limited in number and completeness, reflecting what is available in published sources.
• Certain molecular features (e.g., RNA-seq coverage, expression levels, clonality, cancer cell fraction) are not yet integrated in this release. Only TCGA cancer expression (TPM median) is included.
• Cancer type mapping to TCGA or IntOGen was performed internally during curation but is not included in the public dataset. Researchers may apply their own mapping according to their needs.

These reflect the present scope of published evidence and curation coverage. Future versions will progressively expand molecular features and strengthen patient-level annotations.

Versioning & Updates

• Current version: v0.2
  This release contains 257 curated positive neoantigen peptides, covering 25 cancer types and 46 unique HLA class I alleles.
  All entries are experimentally validated for immunogenicity and include harmonized molecular features such as binding rank, stability, anchor mutation, and expression context.
  Peptide lengths range from 8–13 amino acids, with 9-mers remaining the most common.

• Update strategy
  NDD v0.2 represents a focused, high-quality subset of experimentally confirmed positive neoantigens,
  refined to support transparent benchmarking and community collaboration.
  Future versions will expand by incorporating validated negatives, additional mutation types,
  and extended clinical metadata to enable comprehensive model training and evaluation.

• Changelog
  All modifications, new fields, and updates are documented in CHANGELOG.md,
  including patient cohort adjustments and dataset schema consistency.

• Reproducibility
  Historical drafts prior to v0.2 have been deprecated and are no longer public,
  ensuring that analyses and benchmarks reference a single, consistent dataset release.

---

Community & Contributions

NDD is an open community project, and we warmly welcome participation from researchers, clinicians, and data scientists. There are several ways to get involved:

• Feedback & suggestions
  Since the dataset has been curated largely through manual annotation and standardization, omissions or errors may remain. We sincerely welcome any feedback or corrections.

  • Submit issues directly on GitHub Issues.
  • Or contact us by email at neoantigendd@gmail.com.

• Data contributions
  You are encouraged to share new data sources, patient cohorts, or predictive features that could enrich future releases.

• Collaboration & discussion
  We invite contributions in areas such as data curation, feature development, or method benchmarking. Open discussions and new ideas are welcome.

• Pull requests (PRs)
  If you wish to directly improve documentation, metadata, or scripts, you can submit a pull request — a standard GitHub workflow for proposing changes. This allows us to review your edits and merge them into the main project.

All contributions will be acknowledged in future version updates.

---

Citation

If you use NDD in your research, please cite as:

Neoantigen Discovery Dataset (NDD), version 0.2

GitHub Repository: https://github.com/NeoDiscovery/NDD

Hugging Face Dataset: https://huggingface.co/datasets/NeoDiscovery/NDD

For formal publications, you may include the following BibTeX entry:

@misc{ndd2025,
  title        = {Neoantigen Discovery Dataset (NDD), version 0.2},
  author       = {{Neoantigen Discovery Dataset (NDD) Contributors}},
  year         = {2025},
  howpublished = {\url{https://github.com/NeoDiscovery/NDD}},
  note         = {Accessed: YYYY-MM-DD}
}

---

License & Contact

Acknowledgments

We thank the broader research community for making data and tools publicly available, which has greatly facilitated the development of NDD.
We also appreciate earlier efforts to compile and share neoantigen-related datasets, which provided valuable references during our work.

License

• Code and scripts: released under the MIT License.
• Dataset files (e.g., .tsv, .csv, .parquet): released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
  Users are free to share and adapt the dataset with proper attribution, in accordance with the license terms.

Full license texts are available in the LICENSE and DATA_LICENSE files in this repository.

Contact

For questions, feedback, or collaboration:

• GitHub Issues (preferred): open an issue
• Email: neoantigendd@gmail.com