AlphaFold 3 scores T cell receptor specificity from sequence alone, reaching a median AUC of 0.92 on held-out class I antigens
Synopsis
Whether a T cell receptor recognizes a given antigen has usually been settled by binding assays or tetramer staining, one receptor at a time. Researchers instead used AlphaFold 3 to predict the peptide-MHC-receptor triad and scored the predicted alignment error at the receptor-peptide-MHC interface. On held-out sets excluded from training, the antigen-centric area under the ROC curve had a median of 0.92 for class I and 0.81 for class II, 0.24 to 0.33 AUC units above the earlier AF2-multimer model.
Interpretation
Given a receptor sequence and the peptide-MHC antigen it faces, a structure prediction model can now supply a comparable recognition confidence before any binding experiment is run. The prior route was experimental measurement, or a specialized AlphaFold-multimer pipeline for peptide-MHC-receptor complexes, AF-TCRdock, which brings curated chain-template information into inference. Across 284 solved ternary structures, AF3's median CDR RMSD against the crystal structures was 1.5 and 1.8 angstroms for pre-cutoff class I and II triads and 3.2 and 2.4 angstroms for post-cutoff triads, significantly better than AF-TCRdock in all four cases by a median of 0.8 to 2.9 angstroms.
The score uses a single interface confidence value taken from the predicted structure, so no known structure is needed as a reference. Structural deviation in RMSD can only be computed against a real crystal structure, which is unavailable for candidate receptors. On 1024 class I and 248 class II antigen-study combinations curated from IEDB, 46 of 47 class I features and 44 of 48 class II features departed significantly from chance; the best single feature was the inverted mean predicted alignment error at the receptor-peptide-MHC interface, termed AF3-PTI-PAE.
The score still separates cognate from noncognate triads on data excluded from training, and its performance does not fall with distance from the training data. The earlier AF2-multimer model AF-TCRdock showed no performance decrease between triads released before and after the AF3 training cutoff (P = 0.51 for class I, P = 0.21 for class II), while Boltz-2 dropped on post-training triads relative to pre-training triads. The class I held-out set was 16 triads deposited in the PDB after the AF3 training cutoff; the class II set was 205 triads assigned by the single-cell CRESTA assay across eight structurally uncharacterized Mycobacterium tuberculosis antigens. Antigen-centric median AUCs were 0.92 and 0.81, receptor-centric medians 0.86 and 0.94, with no correlation to distance from training data.
Perspective
The demonstrated scope is human class I and class II MHC-restricted peptide-MHC-receptor triads: the structural benchmark comes from solved PDB complexes, and the classification evaluation from IEDB-curated data plus two held-out sets, the class II one being Mycobacterium tuberculosis antigens assigned by CRESTA.
Classification performance varies markedly with the assay type used to identify the triads: binding assays gave median AUCs of 0.93 and 0.94, while multimer/tetramer assays gave 0.61 and 0.52, yet the latter contributed more than 80% of class I and more than 60% of class II triads. The score correlates with dissociation constant, but the correlation is weaker for natural receptors alone (P = 0.044), and only four class II triads have measured dissociation constants.