A personalized cancer vaccine is built from the mutations inside one patient's own tumor, turned into short protein fragments and injected to teach the immune system what to attack. Most of those fragments do nothing. In a new analysis of 352 cancer patients published in Frontiers in Immunology ↗, only 10 to 20 percent of the peptides chosen for these vaccines actually provoked an immune response. The other four in five were, in effect, wasted slots.
The team assembled a clinically annotated record of 2,317 short peptides, each derived from a single-letter change in a tumor's DNA, and each tested in a patient who had received a personalized peptide vaccine delivered on dendritic cells. For every peptide they had a result from an IFN-gamma ELISPOT assay, a standard lab test that counts how many immune cells fire when shown a given fragment. A peptide counted as immunogenic, meaning it genuinely woke up T cells, if it at least doubled the ELISPOT signal after vaccination.
The instructive part is what did not predict success. Whether a peptide worked had nothing to do with where the mutation sat in the protein, and nothing to do with any particular sequence motif. The field has spent years treating neoantigen selection as a problem of picking the right mutation. This dataset says the mutation's exact identity is close to irrelevant once it is in the vaccine.
What separated the hits from the duds was plainer chemistry. Peptides that were more hydrophobic, meaning more water-repelling and greasy, were significantly more likely to provoke a response (P = 0.0005, a strong statistical signal). So were peptides that bound more tightly to HLA molecules, the surface proteins that hold a fragment up for passing T cells to inspect (P = 0.0014 by one widely used predictor, NetMHC; P = 0.028 by another, MHCflurry). Peptides that stayed bound to HLA longer scored better too (NetMHCstab, P = 0.043), as did those rated more likely to be displayed by presentation predictors (mixMHCpred, P = 0.012; the presentation arm of MHCflurry, P = 0.009).
In plain terms, a neoantigen works when it is sticky enough to grab the immune system's display platform and hold there long enough to be seen.
Why the dud rate matters
None of these predictors is new on its own. The contribution is showing, across hundreds of real patients rather than cell lines, that combining the physicochemical features (hydrophobicity above all) with the existing binding predictors discriminates immunogenic peptides better than any single tool does. Most selection pipelines today lean heavily on HLA binding scores alone, which is part of why they miss so often. A pipeline that loads twenty peptides into a vaccine and gets two or three responders is paying full manufacturing cost for a handful of working components.
Peptide cancer vaccines have had a hard month in the clinic. A different peptide cancer vaccine, Elicio's ELI-002, missed its primary endpoint ↗ in pancreatic cancer this June, though its cleanly resected subgroup did better. Work like this dataset sits one step upstream of those trials. The question is not whether a vaccine platform works, but whether the specific peptides loaded into it were ever likely to fire. peptidemodel tracks the anticancer ↗ peptide space, where the gap between a sequence that binds in a predictor and one that moves a patient's T cells is exactly the gap this analysis is trying to close.