Artificial Avidity Optimization of T-cell Receptors by Integrating Deep Learning Model and Mammalian Cell Surface Display

Weak binding avidity restricts the application of natural T-cell receptors (TCRs) as therapeutic agents. Currently, display technologies have been extensively utilized for avidity optimization of TCRs, yet they tend to be resource-intensive and random. Deep learning models have emerged as promising tools capable of surmounting the constraints of experimental screening. However, the lack of training data has hampered their application in avidity optimization of TCRs. In this study, we profiled peptide-human leukocyte antigen (pHLA) binding fitness landscapes, surveying each amino acid in complementarity-determining region 3 of an influenza matrix peptide (58-66)-HLA-A*02:01 targeting TCR, and performed deep sequencing on the site-directed mutant libraries displayed in human embryonic kidney 293 T cells after pHLA screening, to generate unique training data for deep learning models. The fully trained model was subsequently employed to screen a computationally generated multi-site variant library and predict the antigen-specific optimized TCR candidates. Finally, twenty variants with antigen specificity and three variants with enhanced avidity were identified. The binding affinity calculated by free energy perturbation demonstrated consistent results. The integration of deep learning models and mammalian cell display allows for investigation of any residue within the TCR-antigen interface and a larger accessible sequences of TCRs, leading to an efficient avidity optimization of TCRs.

The online version contains supplementary material available at https://doi.org/10.1007/s43657-025-00295-w.