Discovering a cancer vaccine.
An AI discovery engine re-derived the cancer biology we already know, then found the targets every cancer shares and cannot hide from the immune system.
An AI discovery engine has drafted a single, potentially safe cancer vaccine that could cover roughly 85% of tumours.
Its targets are not the famous cancer genes but the moves a cancer cannot avoid, and they stay put as the cancer spreads: present in 94% of the cells that break away and seed new tumours.
Everything here was done in computation (in-silico), using public data. What follows is how it got there.
Heather, a general intelligence
Heather is Aphorion's general intelligence. It reads raw measurements directly, finds the hidden rules inside them, and knows when something is genuinely new rather than guessing. It is the same instrument that, given raw motion, recovers a law of physics. Here it is pointed at cancer: gene activity and dependency across roughly a thousand tumours and as many healthy tissues, all public data, with no expert features supplied. Every claim below is read off what Heather found, then checked against independent patient cohorts.
It also knows what it does not know. When the search turned up a "perfect" target that healthy cells also make, the engine threw it out rather than report it, which is what keeps the coverage numbers later believable rather than flattering.
First, it re-discovered biology we already knew
Before trusting Heather on anything new, we made it prove itself on answers already known. Run blind, it re-derived a stack of established cancer biology on its own, melanoma drivers, breast-cancer programs, the antigens real vaccines use, even the immune molecule that displays the KRAS mutation, and only afterward was any of it checked against the textbooks.
The one move every cancer makes
A healthy cell does its job: a liver cell handles toxins, a gut cell absorbs nutrients. A cancer cell abandons the job to do one thing: copy itself. Heather found that this switch, from doing your function to replicating yourself, is a single direction shared across every tumour it examined, regardless of which mutation flipped it.
In the data that move has a concrete shape: the cell turns up its copy-making machinery (DNA replication, ribosome building, the cell cycle, with the master regulator E2F1 near the top) and turns down whatever specialised job its tissue used to do.
Why a single vaccine could even work
Here is what makes one vaccine thinkable. A cancer's genome can be a mess: anywhere from a couple of dozen mutations to more than twelve thousand, and no two tumours carry the same set.
If the thing you had to aim at grew with that mess, no fixed vaccine could keep up. It doesn't.
That gap, between a genome that varies tenfold and a state that barely moves, is the whole reason a general, off-the-shelf vaccine, instead of one built for every patient, is even on the table.
Why a vaccine, and not a cure
We didn't start with a vaccine. We first asked whether the same map could describe a cure: a single change that turns a cancer cell back into a normal one.
On the data that change exists, and it is surprisingly simple: it acts on that same growth program (E2F1 and the replication machinery), it is close to a single move, and it is nearly identical across very different cancers.
A vaccine steps around the wall. Instead of forcing the change ourselves, we teach the immune system to recognise the cells that have made the move and clear them, leaving healthy tissue alone. The same map that described the cure becomes the instructions we hand the immune system: this is what to look for.
Targets are bottlenecks, not famous genes
The deeper idea is a change of question. Cancer research usually asks what molecule is expressed? Heather's framework asks what can the cancer not avoid doing?
That is why some of the strongest targets are genes no one had flagged: NDUFA4L2, switched on almost only in kidney cancer, or COL10A1 and MYB, which separate tumour from normal tissue better than the classic markers do.
Telling cancer from healthy tissue: the safety problem
The danger in any cancer vaccine is hitting healthy tissue that divides: bone marrow, the gut lining, skin. Heather found that "dividing fast" alone cannot tell cancer from healthy renewal, because healthy renewing tissue copies itself too. Two readings together do separate them:
- How far the cell has abandoned its job. Cancer de-matures and stops doing its tissue's work (a kidney tumour, say, switching off the normal-tubule protein UMOD); healthy renewal keeps its identity even while dividing.
- Whether one clone has taken over. A cancer is a single lineage that has swept the tissue; healthy renewal is a patchwork of many.
A vaccine target that fires only when both are true spares the healthy dividing tissue that a single signal would endanger. Reassuringly, both readings can be taken from one biopsy, and together they separate cancer from healthy tissue almost perfectly in the data.
Even what cancer breaks becomes a target
There is a catch worth answering. Cancer does not only switch things on; much of what it does is break things: lose genes, drop safeguards. You would think a loss leaves nothing to aim at.
Kidney cancer is the clearest case: it has no useful shared mutation, yet the very thing it breaks forces a brand-new protein onto its surface for a therapy to grab. It is not a one-off, and folding in these gains is what lets coverage climb toward near-complete.
How far this gets us
The honest way to talk about reach is to count. We lined up about 1,600 real tumours and, for each one, asked whether it carries at least one target we could safely aim at. Coverage then builds up one kind of target at a time, climbing as the story does:
- Targets found only in cancer. The cancer's own mutation, and proteins it reawakens that healthy tissue never makes. Safe by nature, and about as far as such targets reach. 67–80%
- The cancer's own driver mutation, for the cancers built on one. Safe for the same reason: that exact mutant protein is in no healthy cell. pancreas 90% · colon 73%
- A target the cancer's own damage exposes. When a cancer breaks one of its safeguards, a hidden program switches on and pushes a new protein onto the cell surface, a handle even for cancers with no mutation or cancer-only protein to aim at. the hard cases
- Programs the cancer over-expresses that healthy tissue shows only faintly. Reachable, with a careful dose. → 95–100%
So the picture is not one magic target but a ladder: every cancer shares something the immune system can be taught to see, and which rung it sits on follows what it shares. One caveat runs through all of it: these numbers mean the target is present, not that the immune system will respond to it. Proving the response is the laboratory step that comes next.
A vaccine that adapts
Cancer escapes a fixed target by dropping it. The encouraging part, from the data, is that the exits are few. Under immune pressure tumours don't scatter; they fall back onto the same small set of states.
A short, predictable set of exits is what makes pre-empting them possible, so the vaccine is run as a strategy, not a single shot:
- Immunise against the target the tumour is using now.
- Work out the one move the surviving cancer cells are then forced to make.
- Prime immunity against that next state before the tumour reaches it.
- Repeat, covering the escape path move by move until the exits run out.
The vaccine we designed
This is not a sketch. Heather wrote two concrete target lists from raw public data, and we locked them before scoring each on data it had never seen, so the design had to call the numbers before it saw the answers. The intended form is one multivalent mRNA vaccine paired with a tumour-sequencing test that picks which targets a given tumour carries, general rather than per-patient. That form is hypothesis; the lists below are not.
KRAS G12D/G12V/G12R/G12C · TP53 R175H/R248W/R273C · BRAF V600E · PIK3CA E545K · NRAS Q61 · SMAD4 · FBXW7
Scored on MSK-MET, 24,755 tumours from a different hospital and a different sequencing machine, none used to build the list:
- 86.5% pancreatic (design said ~90%)
- 68.1% colorectal (design said ~73%)
- 76.2% thyroid
- 4.9% kidney — cold, exactly where the theory says this list cannot reach
XAGE1B · PIWIL1 · CSAG3 · GAPDHS · MAGE family · SSX family — genes switched on in tumours and testis, silent in ordinary tissue
Built on one half of the tumours, then read on the other half it had never seen:
- 92% of unseen tumours carry at least one target (20-target list)
- 18/20 targets stay silent in healthy tissue
That is the strongest evidence in this whole report: not a number we fit, but a prediction the data agreed with afterward.
It follows the tumour, even through metastasis
A vaccine target is only useful if the cancer keeps it. Checked against two large patient cohorts (25,775 tumours in one, multi-region maps of lung tumours in the other), the founding events Heather points to persist from the primary tumour into metastasis.
94% of the clones that actually go on to spread are defined by the tumour's founding events, not by mutations picked up on the way.MSK-MET and TRACERx cohorts
Those founding events sit in the trunk of the tumour, present in every region rather than in late-growing branches (the major lung drivers each show up in 90–96% of regions). So aiming at the founding bottleneck catches the very lineage that goes on to spread, which is exactly what a preventive vaccine needs.
What the framework predicts
A strong idea predicts things that become obvious only once you see it. Seven of Heather's predictions were tested on public data; several hold outright, and the ones that only partly hold sharpened the design rather than breaking it:
- Different mutations, same target. Colorectal cancers driven by completely different mutations land in the same state and share the same vaccine targets, so targets can be chosen without knowing the mutation.
- The same move across cancer families. The copy-yourself switch appears independently in carcinomas, sarcomas and leukaemias: a shared signature of transformation, not a quirk of one tissue.
- Cancers converge under pressure. Under immune attack, tumours of every type fall back onto one shared escape state that is limited and targetable.
- The overlooked targets are the strong ones.Non-famous targets Heather surfaced out-discriminate classical tumour markers, and some are visible only in the one cancer where they sit on a bottleneck.
- A handful of targets goes a long way. Because the shared state is so simple, three well-chosen targets already cover most tumours; the set of possible immune targets is far smaller than the genome's variety suggests.
The honest catch
This is a map, not a medicine. Every result here is computed from public data. It shows that targets are present, shared, and kept as cancer spreads; it does not yet show that the immune system mounts a response against them, or that the safety window holds in a living person. Those are laboratory and clinical questions that data alone cannot answer.
The decisive next steps are physical. Demonstrating that the proposed targets are actually displayed on tumour cells and provoke a real immune response, and that immune pressure drives the convergence the framework predicts, requires the bench, not the laptop.
Bottom line
Every cancer makes one shared move, visible to the immune system and kept as the cancer spreads. That is the basis for a general preventive vaccine, offered here as an in-silico blueprint, with the cures left to the experiments it points to.
Questions and answers
Is this a vaccine I can get today?
No. It is an in-silico blueprint computed from public cancer data. It shows that shared targets are present and preserved as cancer spreads, but not that a vaccine against them produces an immune response or is safe in people. Those are laboratory and clinical steps that come next.
How can one vaccine cover many different cancers?
Because every cancer makes the same underlying move, from doing its tissue's job to copying itself. The state that move produces is small and shared across patients, so a handful of shared targets, rather than a different design per patient, can reach most cancers.
How much of cancer could it reach?
Counting real tumours, intrinsically safe targets that appear only in cancer reach about 67 to 80 percent. Adding the cancer's own driver mutation (KRAS covers about 90 percent of pancreatic and 73 percent of colorectal tumours) and induced programs pushes coverage toward 95 to 100 percent.
Why a vaccine rather than a cure?
The single change that would reverse a cancer cell acts on its growth machinery, and switching that off harms every dividing cell, much like chemotherapy. A vaccine sidesteps that by teaching the immune system to recognise and clear the cells that made the move, sparing healthy tissue.
How would the vaccine be administered (hypothetically)?
As an illustration, not a clinical protocol: a single multivalent mRNA vaccine in a lipid nanoparticle, the same platform behind recent COVID vaccines, given as an intramuscular prime followed by boosters. A companion sequencing test would read which shared targets a person's tumour or risk profile carries and pick the matching set from one off-the-shelf construct. For loss-driven cases with no peptide target, such as kidney cancer, the matching form would instead be a cell therapy against the induced surface protein.
How is it kept safe for healthy tissue that also divides?
Targets in the safe core appear in no normal tissue at all, including marrow, gut and skin. Where that is not possible, recognition is gated by two signals together, loss of the tissue's identity and a single clone having taken over, which healthy renewal does not show.
What data and tools produced this?
Public tumour-dependency, expression and mutation cohorts, from about 1,000 to 25,775 tumours, read with Heather, a general intelligence that finds shared structure in raw data. Every result was checked against independent patient cohorts, and the engine first re-derived known cancer biology as a calibration.
More questions? Email [email protected].
The full technical treatment is not published openly. It is shared on request with researchers, clinicians, and partners.
In-silico. Every result here is computed from public data (cancer dependency, expression and mutation cohorts). It shows the targets are present, shared across patients, and preserved as cancer spreads. It does not show that a vaccine against them works in a person: immune recognition and the safety window remain laboratory and clinical questions. A blueprint to test, not a treatment.