神经抗原引导的TCR发现流程以提高特异性并降低AML中的脱靶毒性

Identifying and engineering neoantigen-specific T cell receptors (TCRs) remains a major barrier to advancing adoptive immunotherapy for acute myeloid leukemia (AML), a malignancy characterized by a low mutational burden and, consequently, limited neoantigen availability. Recent studies have highlighted recurrent AML-associated mutations as promising sources of therapeutic neoantigens, arising from chromosomal alterations, motivating the development of new strategies to systematically discover and evaluate TCRs targeting antigens while minimizing off-target toxicity. We developed a structure-guided TCR discovery pipeline that designs neoantigen-specific TCRs, determines their presence within AML patient repertoires, and evaluates binding specificity and predicted safety. From the literature, we selected recurrent AML neoantigens restricted by HLA-A*02:01, including the TP53 Y220C peptide (VVPCEPPEV) and the NPM1 mutant peptide (AIQDLCVAV), both of which are among the most common TP53 hot-spot and the most frequent molecular alterations in AML. A melanoma-derived TCR-pHLA structure (PDB: 2BNQ) was used as an unbiased scaffold for inserting each AML neoantigen. ProteinMPNN, a deep learning-based protein sequence design, was then applied to redesign residues within CDR1-3, generating 50,000 candidate TCR sequences predicted to optimize interface complementarity and binding. Single-cell TCR-sequencing (scTCRseq) data were analyzed to extract paired alpha/beta patient TCRs, with high-confidence receptors identified based on barcode redundancy and immune phenotype. Clustering of patient and designed TCRs using GLIPH2 revealed a dominant cluster comprising >90% of TP53-associated patient TCRs and two designed TCRs sharing a CDR3 motif, whereas no convergence was observed for NPM1 sequences. Three designed TCRs (two motif-convergent and one top-ranked by structural confidence) and the 62 patient TCRs from the convergent cluster were structurally modeled using TCRmodel2. All modeled TCRs are now being evaluated with STAG-LLM to predict binding specificity and interface similarity to the TP53-HLA complex. The top candidates will progress to molecular dynamics simulations to characterize contact fingerprints and evaluate potential off-target toxicity using CrossDome, which evaluates TCR cross-reactivity based on biochemical similarity between peptide-HLA ligands and predicts the off-target toxicity risk of T-cell-based immunotherapies. Initial results suggest structural and repertoire-based convergence toward recognition of TP53, supporting its relevance as an immunologic target for AML. By integrating rational TCR design, patient repertoire interrogation, and computational safety screening, this pipeline provides a scalable, reproducible framework for discovering neoantigen-specific, potentially low-toxicity TCR candidates for AML immunotherapy.