Protein structure prediction (AlphaFold)
Predicting a protein's 3-D shape from sequence (AlphaFold being the breakthrough) — applied here to model pMHC complexes and, hardest of all, TCR recognition.
AlphaFold showed that sequence → 3-D structure is learnable at near-experimental accuracy, and the same family of methods is now turned on immunology: modeling how a peptide sits in an HLA groove, and whether a given T-cell receptor will physically dock onto that pMHC surface.
Structure adds information sequence alone can't: binding and especially T-cell recognition are about complementary shapes and contacts. Coupling structure prediction with sequence models is a promising route to the field's hardest problem — predicting actual immunogenicity (real T-cell recognition) rather than mere presentation.