Applying deep learning to predict T cell receptor binding specificity of neoantigens and response to checkpoint inhibitors

Project Summary Background: In-depth study of neoantigens will promote our knowledge of the fundamental mechanisms of basic immunology and immune-related disease processes, such as response to cancer immunotherapy. Neoantigens play a key role in the recognition of tumor cells by T cells and are increasingly shown to be targets of checkpoint inhibitor-induced immune response. However, several missing links exist in neoantigen research. (1) Only a small proportion of neoantigens can elicit T cell responses. It is even less clear which neoantigens will be recognized by which specific T cell receptor (TCR). (2) Although neoantigens are important during the course of action of immunotherapies, how neoantigen repertoire data can be used to predict patient response is only poorly understood. (3) The lack of standardized analysis pipelines and limited sharing of neoantigen data have hindered efficient and consistent research in the tumor immunogenomics field. Aim 1: Build a transfer learning-based model to predict immunogenicity of neoantigens. So far, only a very limited number of reports have created predictive models determining whether a neoantigen/MHC complex can elicit any T cell response. Even fewer of them are capable of predicting the TCR-binding specificity of neoantigens. However, the capability to predict the overall immunogenicity and the TCR-binding specificity of neoantigens is critical for improving the benefit of immunotherapy. Aim 1 addresses this challenge with advanced transfer learning algorithms, followed by benchmarking and laboratory validations. Aim 2: Predict response to checkpoint inhibitors by integration of the immunogenicity and other properties of all neoantigens in a patient, through a Bayesian multi-instance learning model. To date, most studies have focused on the neoantigen/mutation load approach in correlation with response of patients to immunotherapy administration. This simplistic approach misses the rich information contained in the whole repertoire of neoantigens per patient and has been successful in only a few studies, but not others. Aim 2 addresses this important inadequacy by creating a Bayesian multi-instance learning model that fully considers various quality features, including immunogenicity, of all neoantigens in a patient for prediction of treatment response. Aim 3: Create a web portal to provide neoantigen-related computational services and to share neoantigen data. The PI will establish a public webserver providing cloud-based standardized services, including prediction of neoantigens and the advanced analysis methods developed in Aim 1 and 2. The webserver will openly share neoantigen/TCR and patient phenotype data, in accordance with IRB and HIPAA regulations. Expected impact: (1) This project will predict the immunogenicity of neoantigens, which could inform neoantigen vaccine development. (2) This project will predict response to checkpoint inhibitors and other forms of immunotherapy based on patient neoantigen profiles. (3) The neoantigen database will propel research and also lead to clinical applications for cancers and other immune-related diseases, such as COVID-19.