The End of All Cancers

For sixty years, drug discovery has chased proteins. We argue that the protein layer is not the right address for cancer cure, and we name three pillars whose simultaneous satisfaction is what a cure for ALL cancers actually requires.

Pillar I — DNA

Cancer is, at the level of molecular causation, a disease of somatic DNA mutation. Every other layer in the cell — RNA, protein, post-translational modification, chromatin — is dynamic, redundant, and erased on timescales of hours to days. Only DNA is the low-entropy, heritable information persisting through every mitotic division of a malignant lineage. The cause lives at the DNA layer. So must the cure.

Pillar II — Programmability

Cancer is mosaic. Each tumor carries 4–5 driver mutations on average, drawn from a long tail of ~150 recurrent driver genes across ~748 census genes. No per-target small-molecule program scales to that tail at $1–2B per drug. A platform whose targeting layer is programmable — one chassis, one swap, days to retarget — converts the tail from constant rediscovery to engineering.

Pillar III — Elimination

Every durable cancer cure that has ever been demonstrated works by removing the disease-driving cell from the population. Surgery, definitive radiation, allogeneic HSCT, CAR-T in B-ALL — all eliminations. Every inhibition modality, by contrast, plateaus against resistance. Inhibition extends life; elimination cures.

If you could target any cell carrying any DNA signature and eliminate it, and you could retarget the system to any new signature in days at marginal cost, you would have constructed the architecturally complete answer to cancer. DNA gives you the address. Programmability gives you the scale. Elimination gives you the cure.

The persistence asymmetry

The case for DNA as the right cause-layer rests on a single physical fact: DNA is the only low-entropy heritable molecular substrate in the cell.

• 01mRNA half-life: median ~10 hours in mammalian cells (Schwanhäusser 2011, Nature).
• 02Protein half-life: median ~36 hours; cyclin and p53 turn over in minutes.
• 03Post-translational modifications: reversible on seconds-to-hours timescales.
• 04DNA mutations: persist for the entire lifespan of the cell and all its descendants.

Every other molecular species in the cell is a transient product of the DNA template. Any therapy that ablates the product leaves the template intact and productive. The unchanged DNA continues to remake the protein the drug just suppressed.

Why the field routed around DNA

Reading DNA inside a living cell was prohibitively expensive until recently — the Human Genome Project cost ~$2.7B and took thirteen years. The tools we had read proteins, not DNA. The FDA's review apparatus, grant funding, and patent law were all protein-shaped. Medicine searched at the protein layer not because that was where cancer lived, but because that was where the light was.

Recognition is a deterministic-endpoint problem

For a given DNA target sequence, we can deterministically design a recognition module that binds this sequence — derivable from Watson-Crick complementarity. There is no equivalent for a small molecule that binds this protein pocket and ignores all other pockets. DNA targeting is engineering. Protein targeting is craft.

The information-theoretic accounting is decisive. A 20-nucleotide guide RNA carries 40 bits — a ~256× excess over the 31.5 bits required to uniquely address the human genome. Antibody paratopes are bounded at 15–25 bits — barely enough to distinguish one protein from the others.

Hijacking a search engine

We are not reinventing recognition. CRISPR/Cas systems already constructed an end-to-end programmable DNA-sequence search-and-engage machine over bacterial co-evolution with phage. FinalDose uses the same machinery to gate a cytotoxic payload — the recognition stage is identical; the conformational change at the end of recognition becomes the molecular switch that releases the cell-destroying payload.

The cost arithmetic

A single chassis is reused across all programs, and only a short guide sequence (~20 nt) changes between indications. This compresses per-target discovery cost by 30–100× relative to small-molecule platforms, reducing the marginal cost of addressing each additional driver alteration to roughly $0.5–2M and the time to clinical candidate from years to weeks.

Inhibition has a structural ceiling

Any therapeutic that operates by inhibition must drive the entire mosaic distribution below a therapeutic threshold simultaneously, and must hold it there for as long as the disease lives. As long as the substrate persists, evolution operates on it.

Elimination beats evolution

The empirical pattern in oncology is unambiguous. The therapies that have produced durable cures all operate by removing the disease-driving cell population:

• 01Surgery in early-stage solid tumors — definitive resection.
• 02Definitive radiation in localized disease — local elimination.
• 03Allogeneic HSCT in leukemias — reset of the hematopoietic compartment.
• 04CD19 CAR-T in B-ALL and DLBCL — elimination of the malignant B-cell clone.
• 05Cisplatin in testicular germ-cell cancer — ~90% cure rate in good-risk disease.

Cancer evolves around inhibition because inhibition allows the cell to survive long enough to evolve. Elimination removes the cell. There is no surviving substrate for selection to act on.

Each pillar is necessary. None alone is sufficient. The convergence is what makes the architecture complete.

DNA recognition + programmable retargeting + cell elimination. FinalDose is the only modality that achieves this. No prior modality occupies it.

Won't perfect protein design close the gap?

If AlphaFold + RFdiffusion + PROTACs + computational antibody design were 100% solved tomorrow, wouldn't the protein layer become sufficient? The answer is no. In the limit of perfect design, the residual cancer-driver gap shrinks from ~75% (current) to ~50–60% — but does not close. The remainder is bounded by target biology, not design quality: flat protein-protein interfaces, intrinsically disordered regions, highly homologous paralogs, and loss-of-function mutations where there is simply nothing to target.

FDCas is the only modality whose addressable target set is bounded by delivery — a moving frontier — rather than by target biology — a fixed physical constraint.

Delivery is the field-wide constraint

Wilhelm 2016 reported median 0.7% of injected nanoparticle dose reaching the tumor across a 232-study meta-analysis. Our differentiation lives in what happens after delivery, not in solving delivery itself. We start where delivery is already solved — HPV, HBV, NMIBC, IPMN, HCC — and the systemic-pan-cancer delivery problem matures alongside the platform.

Single-driver targeting invites resistance

97% of all cancers have the same ancestral mutation shared across every single cell. For the remaining 3%, the mitigation is multi-driver targeting — stacking guides at different recurrent drivers in the same indication.

"Cancer" as a single disease name was always an absurd umbrella over hundreds of distinct molecular conditions. Ten years out, the productive frame may not be "the patient has cancer" but "the patient has KRAS G12C disease, with PIK3CA E545K and TP53 R175H co-occurring." Each driver is its own indication.

The pricing model collapses. A programmable platform amortizes most cost across all indications: ~$200M to validate the chassis, then ~$0.1M of guide-design per additional indication on top. The first FDCas drug is the expensive one. The fifth is an order of magnitude less.