Abstract 3559: A validated bioinformatics tool-set for predicting TCR specificity

Abstract Background: The T-cell receptor (TCR) is a transmembrane receptor on T cells, which is responsible for recognizing any foreign antigen derived peptide, presented on infected or abnormal cells by the MHC-I complex. Predicting TCR specificity computationally is a long-standing problem. Many methods leverage large libraries of TCR sequences and cognate epitopes as input for machine learning and predict specificity from sequence alone. Most methods generally do not perform very well, especially in the case of unseen peptides, i. e. for cases where peptides were not part of the input library. Methods: Here we describe three independent but interacting in-silico modules for TCR discovery. 1) The first module predicts specificity from deep phenotypic T cell profiles. We produce these profiles using a high-throughput pipeline, which combines multiplexed CyTOF with subsequent VDJ CITE-Seq. The generated multiomics data is aggregated into a database called TCAPS, which is the basis for machine learning models. These models can predict T cell specificity from the deep phenotype alone, i. e. without TCR sequence information. The resolution however is limited to the target class and hence the actual epitope has to be determined through subsequent deorphanization. This deorphanization step is augmented by the other two modules also described here. 2) The second module enables ultrafast sequence similarity search, which is used to find similar TCRs with known epitope specificity 3) The third module ranks TCR candidates for prioritisation for in-vitro validation. This is achieved by running the Alphafold-based structure prediction program TCRDock. Results: 1) We demonstrate how TCR specificity can be correctly predicted from multiomics deep T-cell profiles and specifically highlight two in-vitro validated predictions, for which the corresponding epitope/antigen were not part of the input panel. 2) Using a recently published TCR clustering benchmark TCRScapes, we demonstrate the superior performance of our ultrafast sequence similarity search. 3) We furthermore showcase TCRdock-directed experiments on TCRs specific for SF3B1mut-induced splice variants. TCRdock correctly ranked TCR candidates, including artificial combinations, whose functionality we prove through in-vitro validation. Conclusions: We describe three modules for TCR discovery and demonstrate their utility by benchmarking and in-vitro validation of multiple predictions, including two examples against unseen epitopes/antigens. Citation Format: Andreas Wilm, Florian Schmidt, Melissa Wirawan, Manisha Cooray, Kan Xing Wu, Katja Fink, Dan MacLeod. A validated bioinformatics tool-set for predicting TCR specificity abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts) ; 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84 (6Suppl): Abstract nr 3559.