Speaker
Description
Predicting T cell receptor (TCR) recognition of peptide–MHC (pMHC) complexes remains a major challenge in immunology. Existing computational approaches are limited by the lack of dense, locally informative binding data. Synthetic co-evolution via yeast display has recently emerged as a powerful technique to probe protein–protein interactions at scale. Applying these methods to TCR–peptide interactions enables datasets where binding affinity across millions of TCR–peptide combinations can be inferred from rounds of in vitro library-on-library selection.
I present a computational pipeline to tackle this inference problem. First, I show how maximum entropy (Potts) models learn patterns of sequence restriction and co-evolutionary couplings, accurately predicting selection dynamics on held-out data. Second, I show how these models disentangle selection pressures from binding affinity versus confounding factors such as differential expression efficiency. Finally, I address what local library scans reveal about the global binding landscape: I simulate how varying complexity of global models shapes the reduced model, then invert this relation to bound the complexity of rules governing TCR–pMHC recognition.
Our work provides a foundation for inferring binding interactions at scale from library-on-library experiments, offering a path towards the throughput needed to guide next-generation models of immune recognition.
