Project Rosie: How Do You Turn a Tumor Mutation Into a Personalized Cancer Vaccine?

The easiest way to remember Project Rosie is this:

A tumor has mutations. Some create tiny, tumor-only flags. Rosie works out which flags are most worth teaching the immune system to attack, then turns the winners into an mRNA-vaccine design.

That is the whole idea. The biology underneath it is more complicated, but the story becomes intuitive once you follow it one step at a time.

---

First: what are we trying to make the immune system see?

Cancer cells are your own cells that have accumulated mutations. Some mutations change the proteins a cancer cell makes. When the cell breaks those changed proteins into tiny fragments, it can create a fragment that normal cells do not have.

That tumor-only fragment is called a neoantigen.

Normal cell:  ... A B C D E F G ...
Cancer cell:  ... A B C D Y F G ...
                         ↑
                    mutation

The personalized-vaccine idea is:

Find the tumor's unique mutations
        ↓
Find which ones make useful neoantigens
        ↓
Choose the best handful
        ↓
Put instructions for them into an mRNA vaccine
        ↓
Train T cells to hunt cells displaying those neoantigens

Project Rosie is the system for the middle part: deciding which handful is worth choosing.

Step 1: start with a map of mutations

The tumor is sequenced, ideally alongside normal tissue from the same patient. The result can be represented in a VCF—a Variant Call Format. Think of it as a structured list of DNA letters that changed in the tumor:

Chromosome 3, position 12345: A → G
Chromosome 7, position 82921: C → T
Chromosome 11, position 98211: G → A

Rosie's original MVP deliberately began here:

Tumor sample → sequencing → VCF → Project Rosie

A VCF tells us where a mutation occurred. It does not say whether the mutation changes a protein or whether the immune system can see anything from it.

Step 2: translate the DNA change

That is the job of VEP, the Variant Effect Predictor.

VCF
 ↓
"Chromosome 17, A → G"
 ↓  VEP
"This falls in gene X and changes amino acid A into amino acid B"

VCF tells us where the mutation happened. VEP tells us what the mutation changed.

For each protein-changing mutation, Rosie generates short altered protein fragments—peptides—that could contain the mutation. Now we have candidate neoantigens. But most will not make good targets.

Step 3: can the immune system inspect the peptide?

Cells use molecules called MHC as tiny display stands. They hold peptides at the cell surface so T cells can inspect them. Humans call these molecules HLA; dogs use the related DLA system.

Mutated protein inside cancer cell
           ↓
Broken into small peptides
           ↓
MHC / HLA / DLA holds up one peptide
           ↓
Peptide displayed on cell surface
           ↓
T cell inspects it

The whole mutated protein does not need to be on the cell surface. It can be inside the cell. What matters is whether a small piece reaches the MHC display stand.

Step 4: predict which peptides fit

Every person—and every dog—has a particular set of HLA or DLA molecules. A peptide that fits one patient's MHC may not fit another's.

NetMHCpan estimates that fit:

Mutated peptide + this patient's HLA/DLA
                 ↓
           NetMHCpan
                 ↓
"Likely to bind" or "unlikely to bind"

Binding is powerful evidence, not the final answer. A peptide can bind well and still not be made in enough quantity, displayed by the tumor, or recognized by a T cell. That is why a vaccine pipeline needs ranking, not a single yes/no test.

Step 5: rank the candidates

The original Rosie MVP combined mutation annotation, MHC/DLA binding, and a transparent score. A stronger future version adds more evidence:

Does the mutation exist across most cancer cells?       → clonality
Is the altered gene being made in the tumor?            → RNA expression
Can the peptide be processed and displayed?             → presentation
Does it look different from normal self?                → immunogenicity
Is the mutation biologically consequential?             → functional context

The aim is not to claim that one score knows the truth. It is to turn thousands of possibilities into a short, explainable shortlist that can be tested and reviewed.

Thousands of tumor mutations
          ↓
Hundreds of altered peptides
          ↓
Dozens of plausible displayed peptides
          ↓
A small set of high-priority neoantigens

Step 6: turn the winners into an mRNA design

Once targets are selected, Rosie creates an mRNA construct that encodes them. An mRNA vaccine is not the antigen itself. It is an instruction set: cells read it, make the selected antigen fragments, and give the immune system a chance to learn what to recognize.

Selected neoantigens
        ↓
Multi-target mRNA sequence
        ↓
Manufacturing specification
        ↓
Vaccine dose

What Project Rosie is really building

The short version is:

Tumor VCF → understand mutations → generate neoantigens
          → rank the strongest targets → design personalized mRNA

The long-term version is a learning loop:

Patient → tumor data → candidate ranking → vaccine design
   ↑                                           ↓
   └──── clinical and biological outcomes ────┘

Every properly studied case can teach the system which peptides were truly presented, which T-cell responses occurred, which tumors responded, and which predictive signals mattered most.

That is the vision behind Project Rosie: not simply using AI to write an mRNA sequence, but building the evidence system that gets progressively better at answering one hard question:

Which unique features of this cancer should we teach this patient's immune system to attack?

Project Rosie is a research and software concept. Personalized cancer vaccines require rigorous laboratory validation, clinical oversight, and regulatory review before therapeutic use.