TCR-antigen foundation model to empower TCR-based diagnostics and therapeutics - Project Summary Background: The interaction between T cell receptors (TCRs) and antigens presented by MHC proteins is crucial for initiating antigen-specific T cell responses in various biomedical contexts. Understanding these interactions can provide insights into adaptive immunity, antigen discovery, therapeutic TCRs, and the impact of T cells on immune checkpoint inhibitor (ICI) and other immunotherapy treatments. However, traditional experimental methods for studying TCR-antigen interactions are inefficient, costly, and labor-intensive. Solution: We have developed a deep foundation model, pMTnet-omni, which accurately predicts TCR-antigen interactions for both human and mouse, and for class I and II pMHCs. This tool overcomes the caveats of traditional experimental approaches. We will innovate upon this foundation model to (a) discover TCRs for therapeutic use, (b) perform antigen discovery from spatial-TCR-sequencing data, and (c) develop TCR-based biomarkers for predicting immune-related adverse events (irAEs) for patients on ICI treatments. Team: The multidisciplinary team includes bioinformaticians, immunologists, and clinical researchers, ensuring comprehensive development and application of the proposed aims. We have developed AI-based techniques in predicting TCR-antigen interactions and have done pioneering works in multi-instance learning, spatially resolved transcriptomics, T cell engagers, biomarker development, which laid the foundation for this proposal. Aim 1: We will use AI (pMTnet-omni) to discover and optimize TCRs for T cell engagers, against a KRAS neoantigen. AI, coupled with limited experimental validation, will reduce the cost and time spent on engineering TCRs and will result in better TCRs. We will experimentally validate these TCRs’ specificity and efficacy. Aim 2: We will apply multi-instance learning and pMTnet-omni to spatial-TCR-sequencing data to identify novel antigens, based on spatial location, gene expression and TCR sequences. The discovered antigens can serve as druggable targets by therapeutic TCRs, or be incorporated into tumor vaccines. Aim 3: We will develop a biomarker model using pMTnet-omni, TCR-sequencing data generated from patient peripheral blood, and autoantigens of healthy organs to predict irAEs in ICI-treated patients. This biomarker approach is highly generalizable as we can develop this biomarker for any antigen of interest in any disease. Significance: (1) The project aims to create platform technologies and resources, based on the deep foundation model, for broad research and translational uses. (2) We will prove that AI can accelerate the discovery of therapeutic TCRs, which will ultimately benefit patients suffering from various diseases. (3) Significant resources have been spent on inventing spatially resolved transcriptomics (including spatial TCR- seq). We will showcase the significant value that can be derived from such expensive data. (4) Our biomarker method will provide an interpretable and non-invasive approach to extract clinically actionable information from patient TCR repertoire. (5) Our research outcome will be shared publicly through webservers and databases.
