CoChat
Sign Up
Cited by 10vaccines and immunoinformatics approachesMachine Learning in BioinformaticsImmunotherapy and Immune Responses

The DOI matches its Crossref record; 7 of the 7 references it cites were found in a registry or an index.

TransHLA: a Hybrid Transformer model for HLA-presented epitope detection

Open access
Lu, Tianchi, Wang, Xueying, Nie, Wan, Huo, Miaozhe, Li, Shuaicheng·2025·GigaScience
DOI 10.1093/gigascience/giaf008PMC PMC11878767

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Background: Precise prediction of epitope presentation on human leukocyte antigen (HLA) molecules is crucial for advancing vaccine development and immunotherapy. Conventional HLA–peptide binding affinity prediction tools often focus on specific alleles and lack a universal approach for comprehensive HLA site analysis. This limitation hinders efficient filtering of invalid peptide segments. Results: We introduce TransHLA, a pioneering tool designed for epitope prediction across all HLA alleles, integrating Transformer and Residue CNN architectures. TransHLA utilizes the ESM2 large language model for sequence and structure embeddings, achieving high predictive accuracy. For HLA class I, it reaches an accuracy of 84.72% and an area under the curve (AUC) of 91.95% on IEDB test data. For HLA class II, it achieves 79.94% accuracy and an AUC of 88.14%. Our case studies using datasets like CEDAR and VDJdb demonstrate that TransHLA surpasses existing models in specificity and sensitivity for identifying immunogenic epitopes and neoepitopes. Conclusions: TransHLA significantly enhances vaccine design and immunotherapy by efficiently identifying broadly reactive peptides. Our resources, including data and code, are publicly accessible at https://github.com/SkywalkerLuke/TransHLA

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1093/gigascience/giaf008

Cites

7 of 7 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1: Overview of data acquisition and predictive modeling using TransHLA. (A) Data acquisition: The dataset, derived from the IEDB, features a variety of peptide sequences that bind to HLA class I or II molecules. For negative sample generation, nonoverlapping random peptide fragments were sourced by matching positive peptides to their originating proteins through sequence alignment. Then, the nonoverlapping random peptide fragments were processed with CD-HIT to achieve a reduced redundancy, resulting in the final set of negative samples. (B) With ESM2’s advanced modeling capabilities, we generated high-dimensional sequence embeddings for the peptides associated with both HLA classes. Concurrently, structural insights were obtained through ESM2’s contact map predictions, yielding structure embeddings. These 2 distinct yet complementary sets of embeddings were crafted to capture the intricate nature of peptide–HLA interaction, describes the process from data input to epitope presentation prediction. (C) The architecture of the different modules, including the CNN Module 1, CNN Module 2, and the Transformer module.
Figure 1: Overview of data acquisition and predictive modeling using TransHLA. (A) Data acquisition: The dataset, derived from the IEDB, features a variety of peptide sequences that bind to HLA class I or II molecules. For negative sample generation, nonoverlapping random peptide fragments were sourced by matching positive peptides to their originating proteins through sequence alignment. Then, the nonoverlapping random peptide fragments were processed with CD-HIT to achieve a reduced redundancy, resulting in the final set of negative samples. (B) With ESM2’s advanced modeling capabilities, we generated high-dimensional sequence embeddings for the peptides associated with both HLA classes. Concurrently, structural insights were obtained through ESM2’s contact map predictions, yielding structure embeddings. These 2 distinct yet complementary sets of embeddings were crafted to capture the intricate nature of peptide–HLA interaction, describes the process from data input to epitope presentation prediction. (C) The architecture of the different modules, including the CNN Module 1, CNN Module 2, and the Transformer module.CC BY · PubMed Central
Figure 2: This figure illustrates the experimental distribution of major HLA-I alleles in the IEDB. We plotted the alleles by type, where panel A shows the distribution for all data, B for the training dataset, C for the test dataset, and D for the validation dataset. Left represents HLA-A alleles, middle represents HLA-B alleles, and right represents HLA-C alleles.
Figure 2: This figure illustrates the experimental distribution of major HLA-I alleles in the IEDB. We plotted the alleles by type, where panel A shows the distribution for all data, B for the training dataset, C for the test dataset, and D for the validation dataset. Left represents HLA-A alleles, middle represents HLA-B alleles, and right represents HLA-C alleles.CC BY · PubMed Central
Figure 3: This figure evaluates TransHLA’s epitope prediction capabilities, benchmarked against other models using 2 key statistical metrics: AUROC and AUPRC. Panels (A) and (C) examine HLA-I class performance, with (A) showing AUROC and (C) presenting AUPRC. The graphs demonstrate TransHLA’s proficiency in distinguishing between epitope and nonepitope peptides for HLA-I, where AUROC indicates its discriminative power and AUPRC reflects the precision–recall trade-off. Panels (B) and (D) extend the analysis to HLA-II, with (B) displaying AUROC and (D) illustrating AUPRC. The performance depicted emphasizes TransHLA’s effectiveness in identifying HLA-II epitopes, highlighting its ability to differentiate between classes under imbalanced distributions. Together, the panels demonstrate TransHLA’s advantage over conventional models in epitope prediction for both HLA classes.
Figure 3: This figure evaluates TransHLA’s epitope prediction capabilities, benchmarked against other models using 2 key statistical metrics: AUROC and AUPRC. Panels (A) and (C) examine HLA-I class performance, with (A) showing AUROC and (C) presenting AUPRC. The graphs demonstrate TransHLA’s proficiency in distinguishing between epitope and nonepitope peptides for HLA-I, where AUROC indicates its discriminative power and AUPRC reflects the precision–recall trade-off. Panels (B) and (D) extend the analysis to HLA-II, with (B) displaying AUROC and (D) illustrating AUPRC. The performance depicted emphasizes TransHLA’s effectiveness in identifying HLA-II epitopes, highlighting its ability to differentiate between classes under imbalanced distributions. Together, the panels demonstrate TransHLA’s advantage over conventional models in epitope prediction for both HLA classes.CC BY · PubMed Central
Figure 4: This figure displays the PCA-based 2-dimensional feature space distribution of HLA-I and HLA-II epitope presentation test sets. The space is split into 2 regions: region I for HLA-I epitopes and region II for HLA-II epitopes. Each region shows positive samples as blue dots and negative samples as red dots, with representations from various model embeddings (A–H), including random pattern, TextCNN, TextRCNN, RNN-ATTs, DPCNN, random initial, ESM2 pretrained sequence, structure pretrained, and TransHLA.
Figure 4: This figure displays the PCA-based 2-dimensional feature space distribution of HLA-I and HLA-II epitope presentation test sets. The space is split into 2 regions: region I for HLA-I epitopes and region II for HLA-II epitopes. Each region shows positive samples as blue dots and negative samples as red dots, with representations from various model embeddings (A–H), including random pattern, TextCNN, TextRCNN, RNN-ATTs, DPCNN, random initial, ESM2 pretrained sequence, structure pretrained, and TransHLA.CC BY · PubMed Central
Figure 5: This comprehensive figure presents a series of violin plots illustrating the “helix content” structure property of peptides across various sample subsets for HLA-I and HLA-II molecules. Panel (A) delineates the helix content distribution in independent test samples for HLA-I, separated into positive and negative samples, with each subgroup’s statistical significance assessed via t-tests and annotated with corresponding P values. Panel (C) mirrors this setup for HLA-II independent test samples, highlighting the comparative helix content distributions. The external dataset distributions for HLA-I and HLA-II are respectively showcased in panels (B) and (D), emphasizing the metric’s external validity. Panels (E) and (F) delve deeper, contrasting the helix content of true positives and false negatives against true negatives and false positives within HLA-I and HLA-II datasets, respectively.
Figure 5: This comprehensive figure presents a series of violin plots illustrating the “helix content” structure property of peptides across various sample subsets for HLA-I and HLA-II molecules. Panel (A) delineates the helix content distribution in independent test samples for HLA-I, separated into positive and negative samples, with each subgroup’s statistical significance assessed via t-tests and annotated with corresponding P values. Panel (C) mirrors this setup for HLA-II independent test samples, highlighting the comparative helix content distributions. The external dataset distributions for HLA-I and HLA-II are respectively showcased in panels (B) and (D), emphasizing the metric’s external validity. Panels (E) and (F) delve deeper, contrasting the helix content of true positives and false negatives against true negatives and false positives within HLA-I and HLA-II datasets, respectively.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 48

  1. Chaffey, 2003, Molecular biology of the cell
  2. Murphy, 2016, Janeway’s immunobiology, 10.1201/9781315533247 · DOI
  3. Abbas, 2014, Cellular and molecular immunology E-book
  4. Shao, 2020, High-throughput prediction of MHC class I and II neoantigens with MHCnuggets, Cancer Immunol Res, 8, 396, 10.1158/2326-6066.CIR-19-0464 · DOI
  5. O’Donnell, 2020, MHCflurry 2.0: improved pan-allele prediction of MHC class I-presented peptides by incorporating antigen processing, Cell Syst, 11, 42, 10.1016/j.cels.2020.06.010 · DOI · Paper page
  6. Liu, 2021, DeepSeqPanII: an interpretable recurrent neural network model with attention mechanism for peptide-HLA class II binding prediction, IEEE/ACM T Comput Biol Bioinform, 19, 2188, 10.1109/TCBB.2021.3074927 · DOI
  7. Nguyen, 2021, The pockets guide to HLA class I molecules, Biochem Soc Trans, 49, 2319, 10.1042/BST20210410 · DOI
  8. Mei, 2021, Anthem: a user customised tool for fast and accurate prediction of binding between peptides and HLA class I molecules, Briefings Bioinf, 22, bbaa415, 10.1093/bib/bbaa415 · DOI
  9. Reynisson, 2020, NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data, Nucleic Acids Res, 48, W449, 10.1093/nar/gkaa379 · DOI · Paper page
  10. Chu, 2022, A transformer-based model to predict peptide–HLA class I binding and optimize mutated peptides for vaccine design, Nat Mach Intel, 4, 300, 10.1038/s42256-022-00459-7 · DOI
  11. Tadros, 2024, Predicting MHC-I ligands across alleles and species: how far can we go?, bioRxiv, 2024
  12. Racle, 2023, Machine learning predictions of MHC-II specificities reveal alternative binding mode of class II epitopes, Immunity, 56, 1359, 10.1016/j.immuni.2023.03.009 · DOI
  13. Vaswani, 2017, Attention is all you need, Adv Neur Inf Proc Syst, 30, 10.48550/arXiv.1706.03762 · DOI · Paper page
  14. Kim, 2014, Convolutional neural networks for sentence classification, 10.3115/v1/D14-1181 · DOI · Paper page
  15. Lai, 2015, Recurrent convolutional neural networks for text classification, 10.1609/aaai.v29i1.9513 · DOI
  16. Johnson, 2017, Deep pyramid convolutional neural networks for text categorization, 562, 10.18653/v1/P17-1052 · DOI
  17. Zhou, 2016, Attention-based bidirectional long short-term memory networks for relation classification, 207, 10.18653/v1/P16-2034 · DOI
  18. Vita, 2019, The immune epitope database (IEDB): 2018 update, Nucleic Acids Res, 47, D339, 10.1093/nar/gky1006 · DOI
  19. Koşaloğlu-Yalçın, 2023, The cancer epitope database and analysis resource (CEDAR), Nucleic Acids Res, 51, D845, 10.1093/nar/gkac902 · DOI
  20. Shugay, 2018, VDJdb: a curated database of T-cell receptor sequences with known antigen specificity, Nucleic Acids Res, 46, D419, 10.1093/nar/gkx760 · DOI
  21. Nolan, 2025, A large-scale database of T-cell receptor beta (TCR$\beta$) sequences and binding associations from natural and synthetic exposure to SARS-CoV-2, Front Immunol, 10.3389/fimmu.2025.1488851 · DOI
  22. Lu, 2022, dbPepNeo2. 0: a database for human tumor neoantigen peptides from mass spectrometry and TCR recognition, Front Immunol, 13, 855976, 10.3389/fimmu.2022.855976 · DOI
  23. Xia, 2021, NEPdb: a database of T-cell experimentally-validated neoantigens and pan-cancer predicted neoepitopes for cancer immunotherapy, Front Immunol, 12, 644637, 10.3389/fimmu.2021.644637 · DOI
  24. Buchfink, 2015, Fast and sensitive protein alignment using DIAMOND, Nat Methods, 12, 59, 10.1038/nmeth.3176 · DOI · Paper page
  25. Pruitt, 2007, NCBI reference sequences (RefSeq): a curated non-redundant sequence database of genomes, transcripts and proteins, Nucleic Acids Res, 35, D61, 10.1093/nar/gkl842 · DOI
  26. Fu, 2012, CD-HIT: accelerated for clustering the next-generation sequencing data, Bioinformatics, 28, 3150, 10.1093/bioinformatics/bts565 · DOI · Paper page
  27. Chen, 2024, xTrimoPGLM: unified 100B-scale pre-trained transformer for deciphering the language of protein, 10.48550/arXiv.2401.06199 · DOI
  28. Du, 2023, UniDL4BioPep: a universal deep learning architecture for binary classification in peptide bioactivity, Briefings Bioinf, 24, bbad135, 10.1093/bib/bbad135 · DOI
  29. Xu, 2024, PTransIPs: identification of phosphorylation sites enhanced by protein PLM embeddings, IEEE J Biom Health Inf, 28, 3762, 10.1109/JBHI.2024.3377362 · DOI
  30. Mirdita, 2022, ColabFold: making protein folding accessible to all, Nat Methods, 19, 679, 10.1038/s41592-022-01488-1 · DOI · Paper page
  31. Racle, 2019, Robust prediction of HLA class II epitopes by deep motif deconvolution of immunopeptidomes, Nat Biotechnol, 37, 1283, 10.1038/s41587-019-0289-6 · DOI
  32. Nagler, 2021, Identification of presented SARS-CoV-2 HLA class I and HLA class II peptides using HLA peptidomics, Cell Rep, 35, 109305, 10.1016/j.celrep.2021.109305 · DOI
  33. Yang, 2024, MHCII-peptide presentation: an assessment of the state-of-the-art prediction methods, Front Immunol, 15, 1293706, 10.3389/fimmu.2024.1293706 · DOI
  34. Maćkiewicz, 1993, Principal components analysis (PCA), Comput Geosci, 19, 303, 10.1016/0098-3004(93)90090-R · DOI
  35. Westhof, 1984, Correlation between segmental mobility and the location of antigenic determinants in proteins, Nature, 311, 123, 10.1038/311123a0 · DOI
  36. Kim, 2021, Epitopes of protein binders are related to the structural flexibility of a target protein surface, J Chem Inf Model, 61, 2099, 10.1021/acs.jcim.0c01397 · DOI
  37. Klatt, 2020, Solving an MHC allele–specific bias in the reported immunopeptidome, JCI Insight, 5, 10.1172/jci.insight.141264 · DOI
  38. Cock, 2009, Biopython: freely available Python tools for computational molecular biology and bioinformatics, Bioinformatics, 25, 1422, 10.1093/bioinformatics/btp163 · DOI · Paper page
  39. Vihinen, 1994, Accuracy of protein flexibility predictions, Proteins Struct Funct Bioinf, 19, 141, 10.1002/prot.340190207 · DOI
  40. Chen, 2016, Xgboost: A scalable tree boosting system, 785, 10.1145/2939672.2939785 · DOI · Paper page
  41. Mariño Pérez, 2022, Visualizing protein breathing motions associated with aromatic ring flipping, Nature, 602, 695, 10.1038/s41586-022-04417-6 · DOI
  42. Anjana, 2012, Aromatic-aromatic interactions in structures of proteins and protein-DNA complexes: a study based on orientation and distance, Bioinformation, 8, 1220, 10.6026/97320630081220 · DOI
  43. Baker, 2012, Hydrogen bonding in biological macromolecules, 10.1107/97809553602060000887 · DOI
  44. Grimaldi, 2015, Stability of proteins on hydrophilic surfaces, Langmuir, 31, 1005, 10.1021/la503865b · DOI
  45. Drelich, 2011, Hydrophilic and superhydrophilic surfaces and materials, Soft Matter, 7, 9804, 10.1039/c1sm05849e · DOI
  46. Perticaroli, 2013, Secondary structure and rigidity in model proteins, Soft Matt, 9, 9548, 10.1039/c3sm50807b · DOI
  47. Mamonova, 2013, Stability and rigidity/flexibility—two sides of the same coin?, Biochim Biophys Acta, 1834, 854, 10.1016/j.bbapap.2013.02.011 · DOI
  48. Tianchi L, Xueying W, Wan N, et al. Supporting data for “TransHLA: A Hybrid Transformer Model for HLA-Presented Epitope Detection.”GigaScience Database. 2024. 10.5524/102633. · DOI

Metadata

Authors
Lu, T., Wang, X., Nie, W., Huo, M., & Li, S.
Affiliations
  • City University of Hong Kong (HK)
  • Department of Computer Science, City University of Hong Kong , Kowloon 999077 ,
ORCID
Venue
GigaScience
Publisher
Oxford University Press (OUP)
Published
2025
Volume
14
ISSN
2047-217X (ISSN-L)
DOI
10.1093/gigascience/giaf008
PMC
PMC11878767
OpenAlex
W4408141849
Cited by
10
References
7
Citations by year
2026: 5 · 2025: 5
FWCI
2.99
Type
journal article
Language
en
Open access
Licence
Locations
  • GigaScience (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • CityU Scholars (submittedVersion, cc-by)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › vaccines and immunoinformatics approaches
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Immunology and Microbiology › Immunology › Immunotherapy and Immune Responses
Keywords
HLA class I, HLA class II, ESM-2 embeddings, immunogenic epitopes, neoantigens, vaccine design, immunotherapy, IEDB, cedar
MeSH terms
Alleles, Epitopes, HLA Antigens, Humans, Software, Antigen Presentation, Epitope Mapping, Computational Biology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review