Exploring mechanisms underlying heterogeneity of response in personalized cancer immunotherapy by using machine-learning techniques

T cell recognition of a cognate peptide-MHC (pMHC) complex presented by infected/malignant antigen-presenting cells are of utmost importance for mediating a robust and long-term immune response. The recognition is mediated by specific molecular interactions between heterodimeric T Cell Receptors (TCRs) and pMHC ligands and instructs the nature of ensuing adaptive immune response. A better understanding of TCR:pMHC interaction would allow further harnessing of the adaptive T cell immunity and may lead to the development of  vaccines and therapeutics  both in the context of personalized cancer immunotherapies and infectious diseases such as COVID19. The research interests in the Koohy group are focused on the development of machine-learning and Bayesian statistical models to help us better understand two key components of this interaction: A) architecture of the immune repertoire and its dynamics upon exposure to antigens, B) processing and presentation of antigens by MHC molecules to their cognate T cells.

Cancer is usually characterized by accumulation of genetic alterations. Tumour-specific somatic mutations may generate small mutated proteins known as neoantigens that are presented on the surface of cancer cell as ‘cancerous flags’ in association with class I and II HLA molecules.  Neoantigens can be recognized by autologous T cells as foreign and therefore are considered as targets for improved cancer vaccines and adaptive T cell therapies. Almost similar mechanisms are applied to infectious diseases with the difference that the immunogenic epitopes on the surface of infected cells originate from invasive pathogens. Prediction of both immunogenic viral epitopes and cancer neoantigens has been at the centre of extensive research around the globe over the past couple of decades but remains unsolved.   We have been developing various statistical models such as Bayesian Hidden Markov Models to predict immunogenic epitopes that can be used as targets for vaccines for personalized cancer immunotherapy as well as infectious diseases such as COVID1,2.

Over the past decade we have witnessed unprecedented achievements on various cancer immunotherapies in which patients’ own immune system is modulated to find and kill cancer cells. This is evident by the 2018 Nobel Prize for development of Immune Checkpoint Blocked ICB that has greatly improved patients care. However, not all patients respond the same way, besides, some patients develop immune related adverse events such as checkpoint colitis. 
Multiple factors affect immune response to treatment including mutation burden rate, cytotoxic T cell infiltration, antigen processing and presentation defects, mutation-driven clonal signature and the composition of intestinal microbiota. Owing to advances in high throughput sequencing technologies, in particular recent single cell advancements, these features can now be measured from patients’ samples at single cell level at multiple time points including before, during and after the treatment.  We take readouts of these experiments in the form of high throughput sequencing data including genomics, transcriptomics, T cell receptor repertoire, and epigenomics data to train  statistical and machine learning models to study the mechanisms underlying heterogeneity of the response.