Of 41 peptides predicted by algorithms to bind a key immune molecule, only 5 actually triggered immune responses in vaccinated melanoma patients — exposing a major gap in cancer vaccine prediction tools.
5 of 41 workedOnly 12% of computer-predicted HLA-binding peptides actually induced T cell responses in vaccinated melanoma patients
What the researchers found
Starting with 41 peptides predicted by in silico algorithms as the highest-affinity binders to HLA-A*0201, the study found a dramatic funnel of attrition:
- 41 peptides predicted as strong binders
- 19 actually showed strong binding to HLA-A2
- 10 formed stable HLA-A2-peptide complexes and induced CD8+ T cells in transgenic mice
- Only 5 induced T cell responses in PBMCs from ESO-vaccinated melanoma patients
The 5 immunogenic peptides shared features not predicted by algorithms: strong binding, high complex stability, and multiple large hydrophobic and aromatic amino acids. This reveals that current prediction tools capture only part of what makes a peptide immunogenic in humans.
Why it matters
Personalized cancer vaccines depend on correctly predicting which tumor peptides will trigger an immune response. If prediction algorithms have a high false-positive rate — as this study shows — researchers waste time and money testing peptides that won't work. Improving prediction accuracy could accelerate cancer vaccine development by focusing resources on the peptides most likely to be immunogenic in actual patients.
How the study worked
Researchers used in silico algorithms to predict the top 41 HLA-A*0201-binding 8-11mer peptides from the NY-ESO-1 tumor antigen. They then tested binding strength and kinetic complex stability using biochemical assays. Immunogenicity was assessed in HLA-A2-transgenic mice and in peripheral blood mononuclear cells from melanoma patients who had received ESO vaccination. Results were compared against the algorithm predictions.
What this study cannot tell us
The study focused on a single tumor antigen (NY-ESO-1) and a single HLA allele (HLA-A*0201), which may not represent the full diversity of antigen-HLA combinations in cancer. The patient cohort was limited to melanoma patients who had received ESO vaccination. The 5 immunogenic peptides identified may not generalize to other tumor antigens or HLA types. Mouse transgenic models don't perfectly replicate human immune responses.
How to read the evidence
This is a rigorous translational study combining computational prediction with biochemical validation and human immune cell testing from actual cancer patients. Published in the Journal of Biological Chemistry, it provides strong evidence for the gap between prediction and immunogenicity, though findings are limited to one antigen-HLA combination.
When this study was published
Published in 2017, this study remains highly relevant as the field of personalized cancer vaccines continues to grapple with prediction accuracy. The call for improved algorithms has driven subsequent machine learning approaches that incorporate the features this study identified.
The bigger picture
This study highlights a critical bottleneck in personalized cancer immunotherapy: the gap between computational prediction and clinical reality. As cancer treatment moves toward neoantigen vaccines tailored to individual tumors, the accuracy of peptide prediction algorithms becomes a rate-limiting step. The finding that hydrophobic and aromatic amino acid composition predicts immunogenicity better than current algorithms offers a path toward improved tools.
Questions still open
- Can the chemical features identified (hydrophobic/aromatic amino acids, complex stability) be incorporated into next-generation prediction algorithms?
- Does this prediction gap apply equally to neoantigens from individual patient tumors, or is it specific to shared tumor antigens?
- Would machine learning models trained on actual human immunogenicity data outperform current binding-prediction algorithms?
Common questions
Why don't computer predictions match what actually works in cancer patients?
What does this mean for personalized cancer vaccines?
Read the original research
In silico and cell-based analyses reveal strong divergence between prediction and observation of T-cell-recognized tumor antigen T-cell epitopes.
The Journal of biological chemistry, 292(28), 11840-11849
Citation
Schmidt, Julien; Guillaume, Philippe; Dojcinovic, Danijel; Karbach, Julia; Coukos, George; Luescher, Immanuel. (2017). In silico and cell-based analyses reveal strong divergence between prediction and observation of T-cell-recognized tumor antigen T-cell epitopes.. The Journal of biological chemistry, 292(28), 11840-11849. https://doi.org/10.1074/jbc.M117.789511