Every immune response begins with a quiet act of recognition: a receptor on a T cell meets a fragment held up by an MHC molecule and decides whether it is friend or threat. For decades, when that receptor's target was known only as a string of genetic code, the answer could be guessed no better than by chance. Researchers at the Translational Genomics Research Institute now report, in Science Advances, that AlphaFold3, a structural AI originally built to fold proteins, can read that genetic code and anticipate the encounter with better-than-chance accuracy. Tested first on the tuberculosis bac
AlphaFold3 Predicts T Cell Recognition From Genetic Sequence Alone, Beating Chance in Tuberculosis Tests
The best anyone could do was a roll of the dice. Now, clearly better than chance.
Mark
Plainly, what did they actually do?
Mimi
They gave AlphaFold3 the sequences of an MHC, a peptide and a T cell receptor, and used the modeled structure to judge whether the receptor would recognize that antigen.
Mark
Why does the tuberculosis test matter?
Mimi
Those antigens sit in regions where little structural data exists, and the MHC molecules vary a lot across populations. It's a hard setting, and one where vaccines and diagnostics are needed.
Luke
But how well does it work? The phrase is "clearly better than chance."
Mimi
That's Altin's wording, and it's modest. The report gives no accuracy figure.
Luke
Right. And the tuberculosis test was eight MHC class II-antigen pairs. That's a small set, so I'd read this as promising, not settled.
Mark
Is it validated?
Mimi
The source says "if further validated," so not yet fully.
Luke
Also, the claim about speeding up vaccine or cancer work is a projection. Nothing here shows that has happened.
Mark
What comes next?
Mimi
A predictor of peptide-MHC class II binding affinity, using graph-based neural networks trained on PepSeq data from hundreds of thousands of peptide interactions.
Der Puls
- A three-part molecular handshake between MHC, peptide and T cell receptor decides whether the immune system sounds the alarm, yet most of these handshakes remain unmapped.
- Predicting recognition from sequence alone has been a guessing game, and researchers are now showing AlphaFold3 can beat chance on it.
- Eight tuberculosis MHC class II-antigen pairs, poorly covered by existing structural data, became the demanding proving ground for the model.
- Scientists are pushing forward by pairing AlphaFold3 with graph-based neural networks to predict peptide-MHC binding affinity, the step that comes before the receptor.
- The results are promising but still provisional, with further validation required before the method can steer vaccine and immunotherapy work.
Every immune response begins with a quiet act of recognition: a receptor on a T cell meets a fragment held up by an MHC molecule and decides whether it is friend or threat. For decades, when that receptor's target was known only as a string of genetic code, the answer could be guessed no better than by chance. Researchers at the Translational Genomics Research Institute now report, in Science Advances, that AlphaFold3, a structural AI originally built to fold proteins, can read that genetic code and anticipate the encounter with better-than-chance accuracy. Tested first on the tuberculosis bacterium, whose antigens remain largely uncharted, the method offers a modest but meaningful step toward finding the immune targets that vaccines and therapies depend upon.
The TGen team asked a deceptively simple question: if you hold only the genetic sequence of a T cell receptor and the sequence of a possible target, can you tell whether they will connect? Working at the Translational Genomics Research Institute, part of City of Hope, they report in Science Advances that the answer can now be yes, at least often enough to matter. They used AlphaFold3, the AI platform introduced in 2024.
The stakes lie in adaptive immunity. An MHC molecule displays a fragment of a virus, bacterium or tumor cell, and a T cell inspects that fragment with its receptor. When the fit is right, the T cell attacks. Because each receptor's shape is unique, the immune system can recognize billions of possible targets. Only some of these receptor-antigen pairs have been solved experimentally, leaving many known only as sequences.
John Altin, a senior author, described the old limit bluntly: without knowing the target, the best prediction was a random guess. The team now does better than chance, and he expects the model to improve. The researchers first confirmed AlphaFold3 on structures already known, then examined a larger set of immune complexes to find features linked to recognition.
The real test came with tuberculosis. Eight MHC class II-antigen pairs from the bacterium were largely unexplored, and the disease is hard to ignore: the pathogen carries many antigens, and human MHC molecules vary widely across populations. The model modeled these complexes accurately enough to predict recognition, a result that could help speed infectious disease, cancer immunotherapy and vaccine work if further validation holds.
A second effort is underway. Kamel Lahouel and colleagues, working with Altin and Cristian Tomasetti, are combining AlphaFold3 with graph-based neural networks trained on hundreds of thousands of peptide interactions from Altin's PepSeq assays. The aim is a predictor of binding affinity between peptides and MHC class II molecules, the step that comes before the receptor enters the picture.
The question the TGen team set out to answer is a narrow one with wide consequences: if you have only the genetic sequence of a T cell receptor and the sequence of a possible target, can you tell whether the two will connect? Researchers at the Translational Genomics Research Institute, part of City of Hope, report in Science Advances that they can, using AlphaFold3, the AI platform introduced in 2024.
The problem sits at the center of adaptive immunity. Three molecules must work together. An MHC molecule picks up a fragment of a virus, bacterium or tumor cell and holds it out for inspection. That fragment is the peptide. A T cell then examines it with its receptor, the TCR. If the receptor fits both the MHC and the fragment it carries, the T cell registers a threat and attacks. Scientists call this three-part assembly the MHC:peptide:TCR complex.
Each receptor has a unique structure, which is how the immune system distinguishes among billions of potential antigens. Researchers have worked out experimentally the structures of some receptors and the antigens they recognize, but many others are known only as genetic sequences. Predicting whether a given receptor would recognize an arbitrary antigen has been a significant challenge.
John Altin, an associate professor in TGen's Immunology and Microbiome Division and a senior author, put the old situation plainly. If you knew nothing about the target antigen and wanted to know whether a TCR sequence was specific to it, the best available result was a random guess. Now, he said, the team can do clearly better than chance. He called it remarkable that a general structural model works on such a specialized task, and said he expects it to improve.
AlphaFold transformed structural biology by predicting three-dimensional molecular structures from sequences. AlphaFold3 was built to handle interactions among multiple molecules, which suits the three-part complex. The team first tested it on triads whose structures were already known experimentally. After confirming its accuracy, they analyzed a larger set of immune complexes to find structural features linked to recognition and non-recognition of specific antigens.
Then came the harder test: high-quality sequences from previously uncharacterized complexes, including eight MHC class II-antigen pairs from the bacterium that causes tuberculosis. The disease is a demanding choice. The pathogen carries many antigens, the MHC molecules that present them vary considerably among individuals and populations worldwide, and new vaccines and diagnostics are sorely needed. Altin said these antigens lie in large unexplored or underexplored regions of this space, which the group is keen to examine.
The researchers found AlphaFold3 modeled these complexes accurately enough to predict whether a TCR would recognize antigens known only by sequence. If the approach is further validated, it could speed work on infectious disease, cancer immunotherapy and vaccine development by helping scientists pick promising immune targets more efficiently.
A second effort is already under way. Kamel Lahouel, an assistant professor in TGen's Early Detection and Prevention Division and a co-author, said the team is also interested in the step that precedes the TCR: the binding affinity between a peptide and an MHC class II molecule. Teams led by Lahouel, Altin and Cristian Tomasetti are combining AlphaFold3's structural modeling with graph-based neural networks, trained on data from hundreds of thousands of peptide interactions generated by Altin's PepSeq assays, to build a binding-affinity predictor.
Bemerkenswerte Zitate
Now we can do clearly better than chance.— John Altin, Ph.D., associate professor, TGen
The team is also interested in predicting binding affinity between the peptide and an MHC class II molecule, the step before the TCR gets involved.— Kamel Lahouel, Ph.D., assistant professor, TGen (paraphrased)
Möchten Sie die ganze Geschichte?
Das Original lesen bei News-Medical ↗