Dissertations and Theses (Open Access)
Author ORCID Identifier
0000-0001-7899-7961
Date of Graduation
8-2026
Document Type
Dissertation (PhD)
Program Affiliation
Quantitative Sciences
Degree Name
Doctor of Philosophy (PhD)
Advisor/Committee Chair
Ken Chen
Committee Member
Alexandre Reuben
Committee Member
Khaled Sanber
Committee Member
Jeffery Chang
Committee Member
Francesca Cole
Abstract
Recognition of antigen by a T cell begins with a single physical event, the engagement of the T cell receptor (TCR) with a peptide-major histocompatibility complex (pMHC), and this interaction is encoded in the geometry with which the receptor docks onto the composite pMHC surface. Whether that docking geometry is biologically organized, and whether the structures now produced at scale by prediction preserve recognition or only reproduce that organization, are the two questions this thesis addresses. Two gaps stand in the way. The first is a gap of measurement, because docking geometry has never been made quantitatively comparable across many complexes. The second is a gap of inference, because a predicted structure can be confident and plausible without being evidence that two molecules bind.
To close the measurement gap, this thesis develops FramePose, a coordinate framework that parameterizes the position, direction, and rotation of the receptor relative to a pMHC reference frame and resolves both whole-receptor and CDR3-local geometry within one system. Applied to experimentally solved complexes, FramePose shows that docking geometry is not arbitrary but organized, a germline-encoded scaffold set by V-region framework and tuned locally by MHC allele and peptide, chiefly through rotation about the groove-normal axis. This organized geometry explains interface size and burial, but it does not predict affinity from static pose, marking a boundary that a residue-level extension, RotamerPose, shows persists even at side-chain resolution.
To address the inference gap, the thesis applies a falsification standard throughout, crediting a structural feature as evidence of recognition only if it adds transferable information beyond sequence, germline, chemistry-matched counterfactuals, and familiarity, under genuinely unseen contexts. Judged this way, predicted confidence, pose, and interface burial separate binders from non-binders within a peptide context but do not transfer across peptides, and their apparent signal has different origins. Predicted confidence is a layered modelability readout, shaped by peptide and repertoire representation, template familiarity, and register-fit rather than by binding, and predicted burial largely follows the same axis. Predicted pose carries real within-context structural information beyond confidence, but its dominant signal lies in germline-loop and register-organized docking rather than in a transferable CDR3-local binding geometry. Pan-specific performance is therefore context interpolation rather than a general recognition rule, and the plausibility or confidence of a predicted complex does not confirm recognition. The thesis contributes the FramePose framework, evidence that docking geometry is organized, the boundary that separates organization from affinity, and a falsification standard for judging when predicted or generated structures carry transferable mechanistic information.
Recommended Citation
Kim, Kun Hee, "From Geometry to Recognition: A Quantitative Framework for TCR-pMHC Structural Analysis in the Age of Predicted Structures" (2026). Dissertations and Theses (Open Access). 1588.
https://digitalcommons.library.tmc.edu/utgsbs_dissertations/1588
Keywords
TCR-pMHC recognition, Structural immunology, Protein structure prediction, TCR docking geometry, Machine learning