Your immune system runs on a matching game of staggering scale. A new study shows that AlphaFold 3 T cell receptor prediction can play that game from sequence alone — well enough to be useful, and with limits the authors are refreshingly clear about.

The work comes from researchers at TGen and collaborators, and it appeared in Science Advances, a peer-reviewed journal, on October 30, 2026. The analysis code and original data are posted publicly on Zenodo.

The matching problem

Each of us carries more than 100 million unique T cell receptors. Each one is tuned to recognize a specific fragment of protein — a peptide — held up by an MHC molecule on a cell’s surface.

That three-part combination (MHC, peptide, T cell receptor) is called a triad. It is the most diverse three-way molecular interface known in nature.

Sequencing T cell receptors is now routine. Figuring out what each one recognizes is not.

Even the most sophisticated lab methods typically take months to work out the targets of a small number of receptors.

Computational methods existed, but they mostly required training on known receptors for each antigen of interest. Predicting against an unseen antigen stayed an open grand challenge.

What the researchers did

They ran AlphaFold 3 on more than 9,000 T cell receptors mapped to over 1,000 distinct epitopes across more than 70 MHC class I and class II alleles.

Step 1: Check the structures

First they benchmarked AlphaFold 3 against 284 known triad structures from the Protein Data Bank.

It did very well. Median accuracy for the receptor loop arrangement came in at 1.5 and 1.8 ångströms for structures released before the model’s training cutoff (classes I and II), and 3.2 and 2.4 ångströms for structures released after it. That beat the previous state-of-the-art AlphaFold-Multimer approach in all four comparisons.

Step 2: Find a signal

Next they pulled nearly 50 features out of each prediction — confidence scores, interface measurements, docking geometry — and asked which ones separate genuine pairings from scrambled ones.

One feature stood out: a measure of predicted alignment error at the interface between the peptide-MHC complex and the receptor. The team named it AF3-PTI-PAE and used it alone, rather than stacking features, to avoid overfitting.

How well AlphaFold 3 T cell receptor prediction performs

On held-out validation data the model reached median AUCs of 0.92 for class I and 0.81 for class II.

An AUC of 0.5 means coin-flip guessing. These numbers describe a classifier that genuinely separates real matches from decoys, against antigens it had never seen. For class II, all eight test antigens were from Mycobacterium tuberculosis and none had been structurally characterized.

Flipping the question around — given one receptor, which antigen does it recognize? — produced median AUCs of 0.86 and 0.94.

Performance did not fall off for triads that were more distant from AlphaFold 3’s training data. That is the main evidence that this generalizes rather than memorizes.

The limits, stated plainly

This is a useful screen. It is not a solved problem, and the authors say so.

• The class I validation set is small: 16 cognate triads, covering 14 unique antigens.
• Sensitivity is low at high stringency. For class II, the median sensitivity was about 13% at 100% specificity. It rose to 26% at 99% specificity and 52% at 95%.
• It is not suited to very large repertoires, where the prior chance that any given receptor matches a query antigen is tiny and false positives would swamp the result.
• Performance varies with the data source. Scores were strong for receptors identified by binding assays (median AUCs of 0.93 and 0.94) and near-random for multimer/tetramer assays (0.61 and 0.52) — which happen to contribute most of the database.
• Affinity matters. The model scores higher-affinity receptors more confidently, and natural T cell receptors tend to bind weakly, which may cap how sensitive this approach can get.
• Roughly half the genuine matches in the class II set scored no better than random, meaning AlphaFold 3 still fails to model a subset of real complexes well.

Where it is useful right now

The authors’ own recommendation is the right frame: use it to query receptors of interest against small-to-moderate sets of candidate epitopes.

That is a real job. Vaccine researchers and cancer immunotherapy teams often have a shortlist of candidate targets and a set of receptors, and want to know which pairs are worth months of bench work.

There is also an intriguing extra finding. For one well-studied HIV-specific receptor, the model distinguished genuinely cross-reactive peptides from random mutated ones at matched mutation distances. Whether that holds for other receptors is untested.

Why this is still exciting

The honest baseline for predicting T cell specificity against unseen antigens has been close to random. Moving from “essentially random” to “median AUC above 0.8” on held-out, high-quality data is a substantial step.

And the authors sketch clear paths forward: better structure models should translate directly into better predictions, and fine-tuning AlphaFold 3 specifically for this task — or adding a lightweight correction on top of it — could push performance further.

The hardest version of this problem, scanning a query receptor against billions of possible antigens, remains out of reach.

The takeaway

AlphaFold 3 T cell receptor prediction now works well enough, on antigens it has never seen, to narrow a candidate list before anyone touches a pipette. Treat it as a smart filter rather than an answer, and it can save researchers months.

That is progress worth getting excited about — with eyes open about what it cannot yet do.

Read the paper: General prediction of T cell receptor antigen specificity from sequence using AlphaFold 3