Optimization of structure-guided models of peptide presentation and T-cell receptor specificity

Abstract public domain (NIH RePORTER)

PROJECT SUMMARY Motivation: Accurate prediction of T-cell receptor (TCR) specificity to peptide-Major Histocompatibility Complex (pMHC) ligands is a critical challenge—the "holy grail" of computational immunology—as current immunotherapies fail in most patients due to the inability to precisely match TCRs to cancer neoantigens. Existing computational methods rely heavily on sequence similarity, leading to poor generalization to novel antigens and a failure to reliably discriminate between epitopes that differ by a single amino acid, the resolution needed for mutation-derived neoantigen prediction. Additionally, predicting which peptides are presented by MHC (pMHC prediction) remains a key, incompletely solved challenge for developing effective precision therapies. Approach and Results: We developed enFoldX (Ensemble of Folded compleXes), a high-throughput, highly parallelized algorithm that leverages protein structure prediction models (such as AlphaFold3) to jointly fold the TCR:pMHC complex. By generating an ensemble of predicted structures, enFoldX extracts customized biophysical features and confidence metrics that overcome the high false-positive rates and failure modes of previous structural predictors. enFoldX is the first model to robustly generalize across species (human to mouse data) and demonstrate meaningful predictive power in discriminating neoantigen-like epitopes based on single amino acid substitutions, which is crucial for the cancer setting. Moreover, our preliminary data suggests that this ensemble-based approach can improve predictive performance on pMHC binding and can identify non-binding decoy peptides missed by gold-standard tools like NetMHCPan. Impact: We propose to iteratively refine enFoldX and develop enFoldX-MHC to specifically improve Class I and Class II peptide presentation prediction, optimizing both tools through a virtuous cycle of fine-tuning and validation using extensive experimental data, including TCR targets and immunopeptidomes from NSCLC and PDAC immunotherapy cohorts. The resulting highly optimized, fast, and open-source tools will accurately predict TCR:pMHC specificity and MHC presentation of peptides, removing a critical bottleneck and providing a new foundation for designing personalized precision immunotherapies, such as engineered cellular TCR therapies (TCR-Ts), and advancing quantitative immune monitoring.

Public health relevance public domain (NIH RePORTER)

PROJECT NARRATIVE Current cancer immunotherapies, such as checkpoint blockade, succeed in treating only a minority of patients, revealing a critical public health problem rooted in the inability to precisely match T-cell receptors (TCRs) to the specific mutation-derived neoantigens they recognize. The Major Histocompatibility Complex (MHC) is central to this process, as it displays processed peptide antigens, making them visible to the T cells, and thus accurate prediction of both MHC presentation and subsequent TCR binding is the key bottleneck to creating precise therapies. This project addresses this challenge by developing highly accurate AI models to reliably predict these necessary TCR-antigen pairs and MHC presentation, enabling the design of personalized cancer immunotherapies and improved disease monitoring.