Personalized mRNA Cancer Vaccines, Playable

EP 01

Every tumor writes its own wanted posters

Cancer starts with typos in DNA. Some typos change a protein — and a changed protein fragment displayed on the cell surface is a neoantigen: a flag that exists on tumor cells and nowhere else in your body. That uniqueness is the entire strategy. Fire mutations at this gene and watch which ones actually produce a target. Translation below uses the real genetic code — no faking.

✓ good case: a missense typo makes a tumor-only flag — a clean target ✗ bad case: silent typos change nothing; shared typos exist on healthy cells — attacking them would be autoimmune friendly fire

Real standard genetic code table; one gene of 12 codons for readability. A real melanoma carries hundreds to thousands of mutations.

Why it matters: chemotherapy can't tell tumor from you. A neoantigen can — it's the only molecular difference that is 100% tumor-specific. Moderna's intismeran starts by sequencing the patient's tumor and normal blood, then diffing them, exactly like you just did.

EP 02

300 suspects, 34 posters — the algorithm's shortlist

Sequencing finds hundreds of candidate neoantigens, but the vaccine only carries 34. Which 34 is a machine-learning problem: each candidate is scored on how well it binds the cell's display machinery (MHC) and how many tumor cells actually carry it (clonality). Below are 300 candidates from a simulated tumor. Try the algorithm's pick — then try picking badly and watch predicted coverage collapse.

✓ good case: top-34 by combined score covers most tumor cells with strong binders ✗ bad case: random picks tank coverage; turning the self-filter OFF selects peptides that resemble healthy proteins — autoimmunity risk

Scores are simulated but the objective is real: rank by predicted MHC binding × clonal fraction, exclude anything resembling self. Coverage is computed live from your current selection.

Why it matters: this ranking step is where AI lives inside the drug. Neural nets trained on millions of peptide–MHC measurements predict which fragments will actually be displayed — get the shortlist wrong and the perfect delivery downstream is wasted.

EP 03

A recipe in a fat bubble — the delivery race

The 34 posters are encoded as one long mRNA — a recipe, not a finished protein. Problem: your blood is full of RNases, enzymes that shred naked RNA in seconds. The fix that earned a Nobel prize: wrap it in a lipid nanoparticle (LNP). Inject 100 copies both ways and count what survives long enough for ribosomes to read it.

RNase density 5

✓ good case: LNP shields the recipe → most copies reach cells → thousands of peptide flags get displayed ✗ bad case: naked mRNA is shredded mid-flight — this is why "mRNA vaccines" were impossible for 30 years

Conceptual simulation, not molecular dynamics. One honest detail kept: mRNA is transient — it degrades after translation and never touches your DNA. The lesson your immune system learns is what persists.

Why it matters: the same LNP + modified-nucleoside chemistry that delivered COVID vaccines is what makes a personalized drug manufacturable in weeks. You don't build 34 proteins — you print one string of RNA and let the patient's own cells do the synthesis.

EP 04

The brake — why the vaccine needs Keytruda

Trained T cells patrol below. But tumors cheat: they display PD-L1, a handshake that presses the T cell's built-in brake (PD-1) and switches it off mid-attack. Keytruda (pembrolizumab) is an antibody that blocks the handshake. Toggle the vaccine and the brake-blocker independently — the trial's whole design is hidden in this combination.

Tumor PD-L1 70%

✓ good case: vaccine + Keytruda — T cells recognize the flags AND keep their brakes released → kills climb ✗ bad case: vaccine alone against a high-PD-L1 tumor — T cells engage, get switched off grey, and the tumor survives its own discovery

Agent simulation with simplified kinetics. Real T cell exhaustion involves many checkpoints (PD-1, CTLA-4, LAG-3…); PD-1 is the one Keytruda blocks.

Why it matters: the Phase 3 compared Keytruda+vaccine vs Keytruda alone — never vaccine alone. The two mechanisms are complements: the vaccine hands out wanted posters, Keytruda un-presses the brake. Neither is sufficient; together they cut recurrence risk roughly in half in the Phase 2b five-year data.

EP 05

Run the trial yourself — 1,000 virtual patients

INTerpath-001 followed 1,137 melanoma patients after surgery: tumor removed, but invisible micro-metastases may remain. Below, 500 virtual patients per arm. Each month, each patient's residual disease may resurface. Watch the two recurrence-free survival curves separate live — then drag the tumor-heterogeneity slider and watch the vaccine's edge evaporate.

Antigen escape (tumor heterogeneity) 15%

✓ good case: low escape → hazard ratio ≈ 0.5, the shape of the real August 2026 readout ✗ bad case: crank escape to 100% — tumor clones that lack the 34 chosen antigens repopulate, curves converge, HR → 1. This is the honest failure mode the field worries about

Simulated cohort with plausible hazard shapes — not trial data. The real trial reported both endpoints met (recurrence-free and distant-metastasis-free survival); exact Phase 3 hazard ratios follow at a medical congress.

Why it matters: a stock moving +177% in a day priced one number: the hazard ratio between these two curves. And the escape slider is the next decade of oncology — solid tumors are heterogeneous, so the same platform is now being tested in lung, kidney and bladder cancer to see how general the win is.