The Immune System Has a Lookup Problem

What a talk on T cell receptors reveals about the gap between reading biology and writing it

[Written by Claude. Image credit.]

You can take a biopsy from a tumour, sequence every immune cell inside it, and come away knowing almost everything about those cells except the one thing that matters: what they are attacking.

That sounds like it shouldn’t be true in 2026. We can read a genome for a few hundred dollars. We can predict how a protein folds. But the specific question of which immune cell recognises which molecular target has stayed stubbornly closed, and a seminar I sat through recently laid out both why, and what a group at MIT has done about it.

The story ends somewhere unexpected: with a research instrument that turned into a medicine, and a cancer therapy that may be given by injection rather than manufactured in a factory.

Two ways to see a cancer cell

Your immune system’s assassins are T cells. Their job is to inspect other cells and kill the ones that have gone wrong. There are two ways to do this, and they correspond to the two families of engineered cell therapy.

The first is to look at the outside. Cells wear proteins on their surface like badges on a jacket, and some cancer cells wear badges healthy cells don’t. Find such a badge and you can build a synthetic grabber for it. That grabber is a CAR — a chimeric antigen receptor — and CAR-T therapy has been genuinely transformative for several blood cancers.

Its limitation is structural. Most of what makes a cancer cell abnormal happens on the inside, where a CAR cannot see.

The second route exploits something ingenious that cells already do. Proteins inside a cell are constantly shredded into fragments, and some of those fragments get loaded onto a display stand — HLA in humans — and carried to the surface to be held up for inspection. It’s as though every cell were obliged to post snippets of its private paperwork on a public noticeboard. T cells patrol and read those noticeboards using their T cell receptor, or TCR. [Human Leukocyte Antigens (HLA) are proteins on the surface of most human cells that act as identification markers for the immune system. They help immune cells tell the difference between healthy “self” cells and foreign invaders like viruses, bacteria, or transplanted tissue.]

This is far richer. Almost any protein a cell makes can end up on display. If you could engineer TCRs deliberately, you could in principle aim the immune system at nearly anything.

Why nobody can read the code

The trouble is a combinatorial one, and the numbers are genuinely silly.

One person can display roughly a trillion different peptide-plus-stand combinations. One person carries around a hundred million different TCRs, drawn from a theoretical pool near a quadrillion. Researchers on the team put the full space of possible pairings at about 10³¹ — more combinations than there are observed stars in the universe.

Worse, the grip is weak, far weaker than an antibody’s, so real interactions are easy to miss in the lab. And each receptor is promiscuous: a single TCR responds to something like a hundred thousand different fragments. That isn’t a flaw — with 10⁸ receptors covering 10¹² targets, generalists are mathematically necessary — but it kills any hope of a clean lookup table.

So you can’t test your way through it. Can you predict it?

The chart that stops the room

Several groups have trained machine learning models on known receptor-target pairs. On targets the models had seen during training, they score around 0.70 on a standard accuracy measure, where 0.5 is coin-flipping. Modest, but real.

On a target the model has never seen an example of, performance drops to 0.50. Every model. Exactly chance.

The reason is not the algorithms. Pool the four main public databases and you get fewer than a million recorded pairs to describe a space of 10¹² possibilities — and that data is skewed five different ways. Under 4% record both halves of the receptor. About 97% of targets are viral, because viral responses are easy to study. A hundred antigens account for 70% of all pairs. Most entries don’t record which display-stand variant was involved. And over a quarter of known antigens have exactly one recorded partner.

The models weren’t learning biology. They were memorising a small, crooked neighbourhood.

Two inventions

Michael Birnbaum‘s lab at MIT built two tools that together occupy the quadrant nobody could reach: many receptors against many targets at once.

TCRAFT solves manufacturing. Having a receptor’s sequence isn’t having the receptor — you have to physically build it, and commercial synthesis runs about $100 each. But almost all of a receptor is identical between receptors; only two short loops differ. So you order one cheap DNA strand per receptor carrying just those loops, keep the constant parts as a reusable set of plasmids, and use enzymes that leave designed four-letter “teeth” at each cut so the pieces self-sort into their correct partners. Two reactions in a single tube assemble ten thousand receptors. Under a dollar each, over 99% correct.

RAPTR solves detection, by turning a virus into an instrument. Lab lentiviruses are coated in a protein that does two separable jobs: binding to cells and fusing with them. Two mutations destroy the binding while leaving fusion intact — a delivery van that can no longer find any address but can still unload if something else gets it there. Attach a specific target to the outside, and the particle now enters only T cells whose receptor grips that target. Its cargo includes a DNA barcode saying which target it carried.

Each particle labels its own success. Mix thousands of particles with thousands of receptors, sequence the cells that light up, and every cell’s genome now contains both its receptor and the identity of what got in.

Proof, in a skin disease

They tested it on vitiligo — the condition where the immune system destroys the cells that make skin pigment. Not an arbitrary choice: melanoma is cancer of those same pigment cells, and melanoma patients who develop vitiligo-like depigmentation on immunotherapy tend to do better.

Using gentle suction blisters rather than surgery, they pulled T cells from ten patients’ lesions, recovered 3,808 complete receptor sequences, rebuilt every one, and screened them against 101 candidate targets in a single tube.

The hits held up under four independent checks. The melanocyte-reactive cells clustered in one functional state rather than scattering. They were the most clonally expanded cells in the lesions. They scored highest on an independently published signature of genuine tumour-reactive T cells. And a third of them were built from one gene segment already known for recognising the best-studied pigment antigen — established biology, recovered without being told to look for it.

For the first time, you could say which T cells in a patient’s skin were the culprits and which were bystanders.

The half that didn’t work

The talk was refreshingly honest about the other ambition: designing receptors from scratch.

Using a generative model called BoltzGen, the team produced thousands of candidate receptors for chosen targets, built them with TCRAFT, and tested them. The results: the designs fold, they express, they reach the cell surface and can fire. But most of them fire whether or not the intended target is present. They are sticky rather than specific.

That gap matters enormously. Designing a protein that binds something is now reasonably tractable. Designing one that binds only the intended target — in a body where every natural receptor cross-reacts with a hundred thousand things — is the actual clinical requirement, and it isn’t solved.

The accidental medicine

Here’s the turn I didn’t see coming.

A virus that enters only T cells of a chosen specificity is, viewed differently, a targeting system. Load it with a CAR gene instead of a barcode, inject it into a patient, and the manufacturing step happens inside the body. No cell collection, no weeks in a facility, no preparatory chemotherapy, no $400,000 bill.

That became a company, Kelonia, and then a trial. The Phase 1 results reported this year in heavily pretreated myeloma patients: all 18 evaluable patients responded, with no detectable residual disease in bone marrow at one month, and engineered cells still present at four months — all without the usual chemotherapy. Eli Lilly agreed to buy the company for up to $7 billion.

The caveats are real: Phase 1, small numbers, response rate isn’t durability, and most detail comes from conference abstracts. But the striking part isn’t the response rate. It’s that the cells were generated at all, inside a person, without clearing the way first.

What it adds up to

There’s a tidy arc here and an untidy gap.

The arc: we couldn’t read the immune system’s targeting code, so someone built the instruments to read it, proved them on a real disease, and then discovered one of the instruments was also a treatment.

The gap: reading now works — in one lab, at considerable effort. Writing doesn’t. An international consortium with $25 million and a Nobel laureate is working on closing that, and the honest status is that the designs presented so far failed the specificity test.

Which is, in its way, the more interesting state of affairs. The field has stopped pretending its predictions generalise, started generating data at the scale the problem actually demands, and given the tools away rather than licensing them. That’s usually what the phase before a solution looks like.

Leave a Reply