Novel Computational Representations of TCRs and pMHCs Incorporating Structural Information for Translational Applications

Event start: 2026-07-03T14:00:00Z Event end: 2026-07-03T15:00:00Z Location: Duncan Hall   Speaker: Jared Slone Doctoral Candidate Thesis Defense Department: Computer Science Location: Duncan Hall 1049 The ability to accurately predict TCR-pHLA binding specificity would have major impacts on infectious disease research, autoimmunity, and personalized cancer immunotherapy. Experimental characterization of TCR-pHLA binding specificity is labor-intensive and costly, however, and, despite significant advances in machine learning, computationally predicting TCR binding specificity remains difficult. This difficulty arises from both the immense combinatorial space of possible TCR-pHLA pairings and the fact that the proteins in question possess high degrees of freedom, making it difficult to model their physical interactions. This thesis advances the computational characterization of TCR-pHLA interactions through the development of novel computational methods for modeling, representing, and predicting the behavior of TCR-pHLA pairs.   Central to this thesis is the use of three-dimensional structural information when predicting TCR-pHLA binding specificity. First, this thesis introduces STAG, a graph convolutional neural network for predicting TCR-pHLA binding specificity from modeled protein structures. STAG incorporates geometric and physicochemical information into a novel graph representation of TCR-pHLA complexes. The results from this work demonstrate that using holistic structural representations of the proteins as input outperforms prior structure-based approaches for binding prediction which were based on interprotein contacts. This thesis next presents STAG-LLM, a multimodal machine learning architecture that integrates protein language models with structure-based geometric learning. By combining amino acid sequence-derived embeddings with graph representations of three-dimensional protein structure, STAG-LLM achieves improved predictive performance and increased data efficiency compared to approaches that use sequence or structure alone. Recognizing that TCR-pHLA interactions are inherently dynamic, this thesis further introduces STEGG, a computational framework for generating structural ensembles of TCR-pHLA complexes. Through domain-specific conformational sampling and optimization techniques, STEGG efficiently explores the conformational landscape of TCR-pHLA interactions and generates diverse conformational ensembles without the computational expense of molecular dynamics simulations. Collectively, these contributions establish new computational paradigms for computationally representing TCR-pMHC structures, modeling their flexibility, and predicting their interactions. More broadly, this thesis demonstrates advances in structural bioinformatics that improve the computational characterization of TCR-pHLA interactions and aid in the development of personalized immunotherapies. (Department : Computer Science)