Project Rosie Worked at a Hackathon. Here's What It Would Take to Make It Real.

The hackathon version of Project Rosie did something genuinely useful:

Tumor VCF
   ↓
Annotate mutations
   ↓
Generate mutated peptides
   ↓
Predict MHC/DLA binding
   ↓
Rank candidates
   ↓
Design an mRNA construct

That is a coherent first pipeline. It is also not yet a system anyone should confuse with a clinical decision engine.

The difference is not cosmetic. The hard part of a personalized cancer vaccine is not generating an mRNA sequence. It is deciding which 10–30 targets from a tumor's enormous set of mutations are most likely to be displayed, recognized, and worth attacking.

That distinction is the roadmap.

---

The one-sentence upgrade

The MVP asks: Which mutated peptide binds this patient's MHC?

A serious pipeline asks: Which mutated peptide is present in most of the tumor, expressed, processed, displayed by this patient's MHC, recognized by a T cell, and unlikely to be an irrelevant passenger?

Every added word is another biological filter. Each removes attractive-looking candidates that are unlikely to matter in a patient.

Mutation
  ↓
Is it real and tumor-specific?
  ↓
Is it present across the tumor?
  ↓
Is the altered gene expressed?
  ↓
Can the cell process and present the peptide?
  ↓
Can a T cell recognize it?
  ↓
Does the patient actually respond?

Upgrade 1: Replace raw VAF with clonality

VAF, or variant allele frequency, is the fraction of sequencing reads that contain a mutation. It is useful evidence, but it is not the same thing as the fraction of cancer cells carrying that mutation.

Tumor purity and copy-number changes distort VAF. A mutation seen in 30% of reads could be in nearly every malignant cell; a high VAF can also be misleading in a complicated genomic region.

What we care about is cancer cell fraction (CCF):

High CCF → clonal mutation → likely shared across the tumor
Low CCF  → subclonal mutation → only one tumor branch may carry it

For a vaccine, clonal targets are often more strategically valuable. You do not want a perfect immune response against a small branch while the rest of the cancer remains untouched.

The practical upgrade is to add tumor purity and copy-number estimation, then a clonal-deconvolution tool such as PyClone-VI. The Rosie score should use CCF as a major signal, rather than treating read frequency as a direct proxy for tumor coverage.

Upgrade 2: Add tumor RNA

DNA tells us a mutation exists. It does not guarantee the tumor is making that altered protein.

If the gene is not expressed, there may be no mutant protein. No protein means no peptide. No peptide means nothing for MHC to display.

RNA sequencing improves this in three useful ways:

DNA mutation
    +
RNA expression
    +
mutant RNA support
    =
candidate with evidence it can actually exist in the tumor

This is one of the highest-value gaps between a VCF-only prototype and a serious neoantigen workflow. A candidate that is clonal, strongly expressed, and carries its mutant allele in RNA is much more compelling than one selected on DNA and binding prediction alone.

Upgrade 3: Predict presentation, not just binding

NetMHCpan answers an essential question: can a peptide plausibly bind a patient's MHC molecule?

But binding is only one station in a conveyor belt:

Mutant protein
  ↓  protein cleavage
Candidate peptide
  ↓  transport and MHC loading
Peptide–MHC complex
  ↓  stable surface display
T-cell inspection

Strong MHC affinity is therefore necessary evidence, not a guarantee of presentation. A stronger scorer should combine binding with models and evidence about antigen processing and presentation, ideally including mass-spectrometry immunopeptidomics when it becomes available.

This also corrects a tempting but misplaced idea: AlphaFold is not the first tool to add here. The full source protein does not need to sit on the cell surface. An intracellular protein can be broken into peptides and displayed by MHC. Task-specific presentation models are more relevant than a general protein-structure prediction for this ranking step.

Upgrade 4: Estimate immunogenicity

Even a peptide that is expressed and presented may not trigger a useful T-cell response.

The next layer asks whether the mutant peptide looks sufficiently different from normal self, whether the mutation improves binding compared with its normal counterpart, and whether its peptide–MHC complex appears likely to be recognized by a T-cell receptor.

No model can answer this perfectly today. That is exactly why the ranking layer should remain transparent and multi-signal rather than pretending one score is truth.

candidate score =
  clonality
  + expression
  + presentation
  + mutant-vs-normal contrast
  + immunogenicity evidence
  + functional context

The weights should be treated as hypotheses, then improved using real outcome data.

Upgrade 5: Turn the pipeline into a learning system

This is the point where Rosie becomes more than workflow automation.

For every case, the system should capture:

• tumor DNA and RNA evidence
• DLA/HLA type
• predicted and selected neoantigens
• which peptides were experimentally presented
• T-cell response measurements, where available
• vaccine composition and manufacturing QC
• tumor response, recurrence, survival, and adverse events

That creates the loop that matters:

Patient → prediction → vaccine → biological response → outcome data
    ↑                                                   ↓
    └──────────── better candidate ranking ────────────┘

The durable moat is not a wrapper around public bioinformatics tools. It is a responsibly collected dataset connecting predictions to what happened in real patients.

What changes now—and what does not

The MVP does not need to be discarded. It already establishes the useful backbone: deterministic variant processing, patient-specific peptide generation, MHC/DLA-aware ranking, and reproducible mRNA design.

The next version should earn confidence in stages:

1. Add DNA/RNA quality checks, purity, copy number, and CCF.
2. Add expression, antigen-processing, and presentation features.
3. Make the ranking explanation-first and auditable.
4. Validate top candidates experimentally before therapeutic use.
5. Build prospective clinical partnerships, oversight, and an outcomes registry.

That last part is not bureaucracy around the science. It is how the science becomes trustworthy.

Project Rosie started as a way to ask whether open tools could connect the dots from a tumor mutation to a personalized vaccine design. They can.

The next question is harder and more interesting: how do we build the evidence loop that tells us which of those designs will help a real patient?