The missing link in my vaccine

I have never smoked. My cancer began with a genetic accident: two genes, KIF5B and RET, fused together, and the protein made at the join keeps the cell switched on. Since February 2026 I have been taking a targeted pill (a TKI) aimed at RET. The tumour has shrunk.

With every pill in this class, the same thing happens over time: the cells find a way around it. Not if. When.

I want to add a personalised vaccine now: a preparation that shows my immune system the fragment of protein that exists only in the tumour. It isn’t a preventive vaccine. It’s a way of making the tumour visible. And I want it while the tumour is small. That’s my reasoning, not a promise.

I wrote about the kinds of personalised vaccine in part one , and about the role of tissue in part two .

You can’t see what’s inside a cell from the outside. The cell cuts its proteins into short fragments and lifts some of them to the surface, on HLA molecules, which present them to the immune system; T cells can recognise what is displayed there. A vaccine makes sense only if the fragment I train the immune system with actually ends up on display.

Step 5 has a method with a name: immunopeptidomics. You take cells that make the fusion protein, pull the HLA molecules off them, wash the peptides off with acid, and read them by mass spectrometry, the instrument that identifies each peptide by its mass. Not what could be displayed. What was displayed.

There are three different questions here, and they’re easily confused. Prediction tells me what could be displayed. Immunopeptidomics tells me what was displayed. T-cell assays tell me what T cells react to. Step 5 is the middle one — the only one I can answer without tissue and without having my own lymphocytes drawn.

How far the first is from the third has been measured. In an international effort, several teams predicted epitopes — the fragments an immune cell can recognise — from the same sequencing data; the 608 prioritised peptides were then tested for T-cell recognition, on samples from the same patients. 37 proved immunogenic — 6%. The paper explicitly separates the links: first display on HLA, only then recognition. Their filter for display and recognition together removed 98% of the non-immunogenic peptides but kept only 45% of the immunogenic ones — those two figures belong to the 286 peptides for which all five presentation measurements existed, not to the full 608. On a separate cohort of 310 peptides with 4 immunogenic ones, the display-only filter removed 97% of the non-immunogenic peptides and preserved 3 of the 4; adding recognition to it took the removal to 99% and preserved the same 3 of 4 (Wells, Cell 2020 ).

Two limits of that study matter for me, and the authors state both themselves: it tested only class I peptides, and only ones arising from point mutations and small insertions-deletions. Fusion proteins were left out, and so was the whole class II arm. My peptide is a fusion peptide, on class II. So this work tells me how far prediction can drift from measurement. It does not tell me how far mine has.

A filter that throws out almost all the noise but loses more than half the signal on the way doesn’t tell me where the chain broke for my peptide: unrecognised can mean never displayed, or displayed and still unseen by any T cell. Measuring display — immunopeptidomics — is what separates the two. It isn’t infallible either: a review of the method, written by people who run it, warns that matching spectra against the enormous space of non-canonical sequences can itself produce false positives (Chong, Nat Biotechnol 2022 ). What it does do is ask the question I need answered: does this fragment actually reach the surface?

I need the cells, and I have no usable tissue. There’s one narrow way round it — a minigene: a short piece of DNA carrying just the join region, put into a cell line with my HLA type. The result is valid for the cell model; it isn’t a measurement on my tumour. But it’s the only route I can see that consumes no tissue.

In case step 5 sounds like it settles more than it does, here’s what stays open even if it works:

• It doesn’t say whether the peptide is displayed the same way in my tumour of today, eight months into treatment, nor whether the whole protein is processed like the short piece in the construct.
• It doesn’t say whether my T cells see it, nor whether they can tell the tumour fragment from the near-identical stretch in the normal protein.
• It says nothing about killing the tumour, about safety alongside the pill I take, about dose, or about clinical benefit.
• A negative result isn’t proof that the peptide isn’t displayed, unless the laboratory shows what it could have detected: without a detection limit, the absence of a signal isn’t biological absence.

Step 5 is not the sign-off for manufacturing. It is the one thing I cannot choose what goes in the vial without — and whether my T cells then see it is the second of the four things I ask for at the end.

Why not just send tissue?

No mutations were found. That doesn’t mean there are none — it means the question was never answered.

And there’s no lesion left to biopsy. The treatment worked too well. That’s the best problem to have, and it’s still a wall.

I ran the raw data from the two laboratories through public bioinformatics tools, using AI models to write the code, explain the output and check both. Everything that follows is my own analysis. It is exploratory, and it needs independent confirmation.

The study used different prediction tools from mine (Castillo, Front Immunol 2025 ). What it actually measured is less than it looks: it added the peptides from the outside onto cells from two healthy, matched donors and counted the lymphocytes that reacted. It did not measure physical binding to my molecules. It did not measure whether a cell makes the fragment and puts it on its own surface. It measured nothing in me. Even the molecule the fragments would be presented on is, in the study’s own record, a prediction. Its bar for calling a response was set in advance, and it is not a lenient one: more than 20 spots per million cells, and a mean spot count at least three times that of the unstimulated control. Alongside the junction fragments it ran a standardised nine-letter peptide of no biological relevance and a pool of 42 cytomegalovirus peptides. What the article does not print is the spot counts themselves, so how far past that three-times line the positive fragments actually landed cannot be read from it. That is an argument for an experiment that is both better controlled and fully reported, not for discarding the result.

So I can’t say which fragment is the right one. What I can say is narrower than that: the study tested eight fragments spanning the join and published one read-out for each, so there is no ranking of its own to reverse — but the fragment my prediction put first measured negative in both donors, and of the two that did respond, one is a fragment I had scored as a non-binder. That is enough to disqualify prediction as the sole criterion for choosing. There’s another reason I don’t simply take the fragment with the best result: it’s also the closest to the normal protein of the whole set, a single letter away — and not from one normal protein but from two. The strongest experimental signal sits on the thinnest safety footing. No calculation resolves that tension, and I’m not going to resolve it by picking a favourite.

One more thing, so it doesn’t get lost: these fragments are class I, a different arm from the class II peptide the laboratory chose. Their results say nothing about it, for or against. The gap between calculation and measurement isn’t noise, and only a laboratory can close it.

One piece of the prediction does argue for running that measurement, on the class I side, and it splits in two. Delivery looks good: the TAP transporter, which carries a fragment to where HLA molecules are loaded, scores the junction fragment VKEDPKWEF at +1.025, 4th of the 89 nine-letter fragments in the window and the best of the eight that cross the join; the combined transport-and-cutting score also comes 4th. Cutting is a different story: on the proteasome component alone the same fragment is mid-pack, 29th of 89, and the strongest predicted cut in the neighbourhood falls precisely on the junction itself, which is a way to destroy the fragment rather than to release it. What these two scores do have going for them is that they don’t depend on which HLA allele you assume, so the uncertainty about my alleles doesn’t touch them. Taken together: prediction says the piece is likely to be delivered if it survives cutting, and what stays uncertain is whether it is cut out intact and then stays bound long enough to be displayed. Only the measurement can settle that. Those two steps belong to the class I route — class II molecules are loaded by a different path — so this is one favourable piece of prediction about the junction region, not a verdict on the class II peptide itself.

The same raw data also told me what my immune system looks like inside the tumour.

One team built a vaccine on the junction of a fusion, in a patient with a different cancer type, proved by mass spectrometry that the peptide is naturally displayed, and followed the patient for over 21 months without relapse (Bauer, Nat Commun 2022 ). In resected melanoma, a personalised mRNA vaccine added to an immune checkpoint inhibitor (an immunotherapy that lifts the brake off T cells) gave, at the 5-year update, a hazard ratio (HR) of 0.510 for recurrence-free survival (KEYNOTE-942, J Clin Oncol 2026 ). And the route itself exists on paper: personalised multi-peptide vaccination is registered as expanded access, listed as available, sponsored by a German university centre (NCT05014607 ). Expanded access is not a trial; it is a named-patient route. It is also the same German university programme I list below as blocked for me: it requires fresh-frozen tissue, which I do not have, and it does not make vaccines against oncogenic fusions, with one exception: the fusion-junction vaccine this section opens with, in another cancer type. So it is a door I cannot walk through — and still proof that the door is built.

Different cancer, different situation, different vaccine. My disease is advanced, not resected, and that number is not mine. But the mechanism is real and it has been measured.

And a join like mine has already been met by human T cells once. In a patient with anaplastic thyroid cancer carrying a CCDC6-RET fusion, lymphocytes taken from inside the tumour reacted to the fusion neoantigen; the break on the RET side is the same exon 12 break as mine, and the 25-letter stretch they tested is identical to mine at 14 of its 25 positions, with all 11 differences falling on the partner half (Garza, J Clin Endocrinol Metab 2026 ). Three things travel with that, and none of them is small: it is a different cancer, it is a different experiment — T cells already inside a tumour, not a vaccine given to anyone — and the bar for calling a response was a 1.5-fold rise in a single replicate, weaker than the three-fold bar in the lung study above. It also does not say which half the reactive piece sits on, the RET stretch we share or the partner half we do not: nobody localised it. What it does say is that this is not a one-off observation in a single paper.

The measurement itself is not hypothetical, and I want to be accurate about that rather than dramatic. Immunopeptidomics has been published on samples of a few tens of milligrams, using an ion-mobility step built for exactly the problem of too little material, with each candidate peptide then confirmed against a synthetic standard; the same work ran the method on cell lines as well as on tissue, which is the route open to me (Minegishi, Commun Biol 2022 ). So the obstacle is not that nobody knows how. It is that I have not found the step offered on its own, to someone who arrives with data instead of tissue, and the groups who publish it run it inside their own pipelines. One approach went out on 25 October and has had no reply. I am not writing it down as a route until someone answers.

Tissue does not grow back in paraffin. Data does not improve on its own. The question is simple: do I build the vaccine with what I have, or not.

One more thing, stated plainly. The clone of cells carrying the fusion is, most likely, the largest population in my tumour. Whether there are other useful targets as well, I do not know: exactly the sequencing that would have answered that is the one that failed.

For most people, AI means a chat. For me it means two tools paid for month after month: Claude, from Anthropic, and Codex, from OpenAI. With them I have run code, worked through the raw sequencing data, and had one model check the other’s work. The hard part isn’t getting an answer; it’s knowing when an answer is good enough to act on: every claim tied to a source, every number recomputed, every conclusion run past the second model. What these tools could not do is repair broken data: no algorithm reconstructs reads that were never produced.

Everything marked as computed by me is my own analysis, at an indicative level, and needs independent confirmation. It does not replace a laboratory result. Accredited results are marked as such.