Atom-Level Structure-Involved TCR-Peptide-MHC Specificity and Affinity Prediction Model

We introduce introduce AtomTCR, an atom-level structure-based computational framework that enables simultaneous prediction of TCR-pMHC binding specificity and affinity for both known and novel peptides. AtomTCR employs a pre-trained structure prediction module to generate TCR-pMHC complexes for arbitrary peptides, followed by extraction and encoding of key binding regions—specifically the TCR CDR3 loops and peptide-MHC interfaces—into atomic-level embeddings. To enhance generalization across diverse peptides, we design a dual-expert architecture that synergistically combines amino acid sequence features with structural data, enabling accurate predictions even for unseen epitopes.