From Geometry to Recognition: A Quantitative Framework for TCR-pMHC Structural Analysis in the Age of Predicted Structures

Kun Hee Kim, MS (Advisor: Ken Chen, PhD)

Adoptive and engineered T cell therapies depend on antigen-specific recognition, yet T cell activation is shaped by many factors, including cellular state and external signaling. Underlying these, recognition is initiated by a physical molecular event: engagement of the T cell receptor (TCR) with peptide-MHC (pMHC). This interaction is structurally encoded by docking geometry, interface contacts, and energetic complementarity, and is therefore informative of binding. Historically, TCR-pMHC structural analysis has relied on qualitative, case-by-case interpretation of a limited set of experimentally solved complexes. Recent advances in protein structure prediction now enable high-throughput modeling of TCR-pMHC complexes, creating an opportunity for quantitative structure-to-binding analysis while raising a risk, that the prediction yields geometrically plausible structures even where no binding occurs.

This thesis develops TCR-FramePose, a quantitative framework for representing and comparing TCR-pMHC geometry across global and local scales. FramePose parameterizes the relative position, direction, and rotation of the receptor with respect to a pMHC reference frame, converting docking-pose variation into interpretable structural descriptors. Applied to experimentally solved complexes, FramePose reveals that docking geometry is biologically organized and that CDR3-local pose features contain interface and binding-associated information not captured by conventional descriptors.

The thesis then evaluates whether predicted structural features can discriminate binding from non-binding contexts. Predicted confidence, pose, and interface burial separate binders from non-binders within individual peptide contexts, but these signals do not transfer across peptides. Further analysis shows that these apparent structural signals have different origins. Predicted confidence is a layered modelability readout, shaped by peptide/repertoire representation, receptor-template familiarity, and structural register-fit rather than by binding itself. Predicted burial largely follows this confidence-modelability axis. Predicted pose contains real within-peptide structural information beyond confidence, but the dominant signal lies in germline-loop and register-organized docking rather than in a transferable CDR3-local binding geometry. Predicted structures are therefore valuable for studying docking organization, public-register structure, and germline-repertoire effects, but the plausibility or confidence of a predicted complex does not confirm recognition. Together, this work establishes a framework for quantitative structural immunology and defines the promise and limits of predicted TCR-pMHC structures for binding interpretation.

Advisory Committee:

• Ken Chen, PhD, Chair
• Jeffery Chang, PhD
• Francesca Cole, PhD
• Alexandre Reuben, PhD
• Khaled Sanber, PhD

Join via Zoom (Plese contact Ms. Kim for her Zoom meeting info.)