When a virus enters the body or a cancer cell appears, T cells, the immune system's fighters, take the stage. Each T cell has its own receptor, a lock-like molecule that searches for the right key: a piece of protein (peptide) on the surface of the diseased cell. If the key fits, the T cell activates and destroys the enemy. But there are billions of locks and keys, and predicting which receptor will work against a specific peptide is a daunting task.

Computer models try to solve this by reading amino acid sequences — the letters that make up proteins. The new model ImmSET works like an attentive translator: it looks at a set of receptors and a set of peptides and learns to match them to each other. Unlike previous models that compared pairs one by one, ImmSET considers the entire set at once, like a translator who listens not to individual lines but to the whole conversation.

The scientists tested the model on data of varying sizes and compositions. They found that the more training examples, the more accurate the predictions, like a student who solves more and more problems. With sufficient data, ImmSET even outperformed AlphaFold2 and AlphaFold3, the famous models that predict protein shapes. This means that for the "recognition" task, not only shape but also the text of the sequence itself matters.

This approach is useful not only in immunology. Any task that involves matching sets of biological sequences, such as enzyme interactions or signaling molecules, could benefit from a model that sees the whole picture. And for medicine, this is a step toward tailoring therapy based on a patient's unique "immune passport."