Researchers developed a system combining high-throughput yeast display and fine-tuned protein language models to create deep peptide recognition profiles for TCRs. This approach details binding against millions of peptides and outperforms models like AlphaFold3 in predicting T cell activation. The…
Predicting T cell receptor specificity from sequence proves difficult. TCRs with similar sequences often bind different antigens, while those with distinct sequences can target the same ones. This creates a core issue in understanding adaptive immunity responses to pathogens, cancer, and autoimmunity.
TCRs interact with a surface formed by antigenic peptides and major histocompatibility complex molecules. The complexity of this interface offers chances for biotech advances but also poses hurdles. Sequence clustering tools like GLIPH and TCRdist identify general trends in epitope specificity, yet they fall short in pinpointing precise differences among related TCRs.
Existing methods to connect sequence and function have clear limits. High-throughput yeast or mammalian display screens map peptide interactions on a large scale, but turning that data into strong predictive models remains tough. Computational tools relying on sequence similarity or structure predictions, such as AlphaFold3 and tFold-TCR, use interface predicted template modeling scores to rank bindings, but they often fail at accurate de novo predictions, especially for new epitopes and varied TCRs.
Synthetic peptide libraries in display systems also miss ties to relevant native proteomes. These gaps slow efforts to find autoantigens in HLA-linked diseases like ankylosing spondylitis and acute anterior uveitis, both tied to HLA-B 27.
Scientists introduced an experimental-computational platform that produces deep peptide recognition profiles for TCRs with high-resolution views of their pMHC recognition. It pairs high-throughput yeast display, testing single TCRs against millions of peptides, with protein language models fine-tuned on the binding results.
In a set of HLA-B 27:05-restricted TCRs from people with ankylosing spondylitis and acute anterior uveitis, these profiles show binding focused almost solely through CDR3β. Crystal structures confirm this strong emphasis of CDR3β on peptide contacts in the group. The profiles allow detailed mapping of structure-activity links.
Protein language models trained on these deep peptide recognition profiles exceed AlphaFold3 and tFold-TCR in forecasting T cell activation. This edge led to finding and confirming new candidate autoantigens for ankylosing spondylitis and acute anterior uveitis.
One example is a peptide from PSG5, checked in patient-specific T cells. The approach applies to TCR families visualized at high detail, where CDR3β drives peptide recognition.
Model performance on new TCRs links to functional distance, measured as peptide recognition profile divergence, instead of sequence likeness. Researchers added a built-in uncertainty measure to gauge prediction reliability.
This combination of deep profiling and machine learning targets tight, disease-related TCR groups. It uncovers specificity patterns missed by sequence methods alone.
Related reading: Peptide-Guided Method Activates PAK1 Kinase, Eli Lilly Commits Up to $300M for CrossBridge Bio Buy, CSBio's 2-Hour Mini-Course on SPPS Fundamentals, Ved Srivastava on Peptide Science from Early Days to AI Era.