From Tumor Immunogenomics to Deep-Learning for Immunogenicity Prediction

Date of Award

Spring 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Genetics

First Advisor

Krishnaswamy, Smita

Abstract

Tumor immunobiology research has critically guided the design of a variety immunotherapeutic strategies over recent years, but a key challenge remains in sustaining high immunogenic response rates in the context of immunotherapy. An overt immune response within a tumor is broadly determined by its immunometabolic state, the processing and presentation of mutant peptides on the cell surface, and the TCR-based recognition of peptide-MHC complexes. Diving deeper into each of these individual components by integrating different data types may reveal further complexities that are highly specific to certain tumor subtypes. Accordingly, this PhD dissertations begins with multi-modal analyses of two distinct immunometabolic landscapes in lower-grade glioma and lung carcinoma, surfacing tumor-specific immunometabolic states that associate with clinical features such as overall survival. To move beyond examining these patterns of immunometabolic crosstalk and towards therapeutic angles for precise immune modulation, I developed a new deep-learning model, ImmunoStruct, to improve epitope-based vaccine design. Traditional immunogenicity prediction tools predominantly train only on protein sequence data and lack structural and biochemical context when predicting epitope immunogenicity. To address this, I built ImmunoStruct as a multi-modal deep learning model that integrates peptide-MHC sequence, peptide-MHC structure, and biochemical properties to achieve state-of-the-art class-I pMHC immunogenicity prediction. By leveraging wide-scale AlphaFold2-generated pMHC structures, ImmunoStruct can capture conformational patterns and physicochemical determinants that improve epitope selection for cancer neoantigens and infectious disease antigens. Unlike sequence-based models, it also provides an added degree of biostructural interpretability. Furthermore, it highlights the benefit of integrating in-silico generated structural data for complex property prediction tasks, which may also extend to other domains of biology. This work transitions from first surfacing immunometabolomic snapshots in tumors, to active immune engineering of antibodies and epitope-based vaccines. Together, it becomes apparent how multi-modal deep-learning and structural immunology can refine and improve computational vaccine design and epitope-based immunotherapy.

This document is currently not available here.

Share

COinS