A plain-English guide to how personalized cancer vaccines work, where AI enters the workflow, and why the field is moving now.
A neoantigen vaccine teaches the immune system to attack protein fragments that exist only on a patient's own tumor.
Cancer cells carry DNA mutations that healthy cells don't. Some of those mutations change the proteins a tumor makes, producing protein fragments found nowhere else in the body — neoantigens. Because they're genuinely foreign, the immune system can be taught to attack any cell displaying them. That's the premise of a personalized cancer vaccine: built per patient, from their own tumor, to point the immune system at their specific neoantigens.
Cells display chopped-up protein fragments on HLA molecules, and patrolling T cells kill any cell showing a fragment they recognize as foreign.
Every cell constantly chops up its own proteins and displays the fragments on surface molecules called HLA (the human version of the MHC). These molecules come in two flavors — MHC class I and class II — that show peptides to different kinds of T cell. This antigen-presentation system is how the body audits each cell from the outside. A displayed peptide locked into an HLA molecule is a peptide-MHC complex — the actual unit that immune cells inspect.
Patrolling T cells read these complexes with their T-cell receptors. If a receptor fits a tumor neoantigen, the T cell activates and kills the cell showing it. A vaccine works when it expands an army of T cells tuned to the tumor's neoantigens.
Because HLA genes vary so much between people, the same tumor mutation can be a perfect target in one patient and invisible in another.
HLA genes are the most variable in the human genome, so the very same tumor mutation can be a great target in one patient and invisible in another. Every pipeline therefore starts with HLA typing the patient, and a high tumor mutational burden only tells you there are more candidates to sift — not which ones will work.
Picking the few useful peptides out of hundreds is a stack of prediction problems, and immunogenicity — whether a T cell will actually respond — is the open frontier AI is racing to crack.
A tumor can present hundreds of mutated peptides, but only a few are worth putting in a vaccine. Narrowing them down is a stack of prediction problems — collectively, epitope prediction. Will the peptide be processed and presented? Will it bind this patient's HLA? And the hardest question: will a T cell actually respond — its immunogenicity?
HLA-binding prediction is now quite good; immunogenicity and predicting which receptor recognizes which target remain open, high-value problems — which is exactly where today's machine-learning and foundation-model research is concentrated. Experimental immunopeptidomics (directly measuring what's displayed) increasingly supplies the ground-truth data these models learn from.
Chosen targets are delivered — often via mRNA — and frequently paired with a checkpoint inhibitor that frees the primed T cells to act.
Once targets are chosen, they're delivered to the patient — frequently via an mRNA platform, prized because a new sequence can be manufactured quickly per patient. Vaccines are often paired with a checkpoint inhibitor, which releases the brakes on T cells so the freshly primed army can do its work.
That's the whole loop: sequence the tumor → find the mutations → predict the best neoantigens → build a vaccine → unleash the T cells. The daily brief tracks who is moving each piece forward — in the clinic, in the market, and in the algorithms.