Modeling the CD8+ T cell response to solid tumors: A multi-scale immunoinformatics approach to prognosticating disease outcome

Abstract

Robust CD8 T cell activity is crucial for anti-tumor immune response. CD8 T cells keep tumors in check using direct effector mechanisms that cause tumor death. The key to this activity is the recognition of tumor cells as being "foreign" or "non-self." Tumor peptide antigens known to elicit CD8 T cell response can be split into tumor-specific antigens (TSA) and tumor-associated antigens (TAA) based on whether tumors uniquely present them. Neo-antigens are the most intuitively understood form of tumor-specific antigens derived from mutated proteins in the tumor. However, barring a countable number of point mutations, indels, and structural variants, the tumor proteome resembles that of non-tumors, and the odds of any specific mutation resulting in a tumor-specific neoantigen are vanishingly small, especially in tumors that are not widely mutated. Alternatively, tumor-specific responses can be targeted against proteins solely or differentially expressed in tumors, although it is not apparent which TAAs can elicit them. Accurately modeling the CD8 T cell response to tumors requires characterizing the specific antigens expressed by the tumor.

CD8 T cells scan through a sea of peptides presented by HLA-I molecules on the cell surface and enact their effector activity if their cognate T-cell receptor (TCR) can recognize any specific pHLA-I complex. In a physiologically healthy situation, the TCR repertoire should be depleted of receptors that recognize self-antigens by the negative selection during thymic education. However, this process is not absolute, and self-reactive T cells are seen in the periphery. A causal link has been established between the absence of protein expression in the thymus and the presence of these circulating T cells. Self-antigens for the negative selection process are primarily by a subset of medullary thymic epithelial cells (termed mTEChi), and while they cover an impressive amount of the proteome, it is not complete, and some antigens are produced much more than others. A key aspect in understanding anti-tumor response is thus also contingent on understanding tumor antigens poorly represented during thymic education, but accurately quantifying the self/non-self divide remains a lacunae that hinders this understanding.

We address this gap in this work and model different steps in the spectrum of molecular events required for initiating and effecting cytotoxic T-cell responses against tumors. Towards this, we first build a ‘Thymic Education’ model that has provided predictivity on the recognizability of tumors by factoring in the repertoire of peptides presented in the medullary thymic epithelial cells. We devise a score termed as the "thymus-likeness score" (TLS) and reason that tumors with lower TLS scores are more likely to present poorly represented proteins during the negative selection process and should be easier to recognize compared to tumors with higher TLS. We then developed DeepHLAffy, a deep learning model that accurately predicts peptide-HLA complex formation. The algorithm utilizes large-scale sequence and structural information of HLA alleles and peptides and is based on deep learning approaches and has an added advantage of being ‘interpretable’. We also develop DIPPred, a comprehensive algorithm for modeling the immunopeptidome in individual patients, that caters to precision modeling of the antigen processivity, presentation and recognition to HLA alleles in an unbiased manner and outputs a high likelihood set of pHLA-I complexes. Next, we integrate these with further steps of generating a T-cell response into an end-to-end computational pipeline iTRPred, which is able to take in a tumor transcriptome as an input, read in the individual HLA genotype in that patient, predict the set of possible T-cell antigens encompassing both self and non-self based on the previous modules and precisely quantifies CD8-T cell response in that individual patient. iTRPred combines DeepHLAffy, DIPPred and ERGO-II, a previously reported deep-learning based algorithm for predicting pHLA-TCR complex formation. The individual modules in the pipeline utilize exome sequences and transcriptomes of individual samples and are designed for performing with high speed to enable genome-wide scanning of a given tumour sample in about 3 hours. Put together, the pipeline enables rapid identification of the entire repertoire of tumor antigens and quantifies the resultant CD8 T-cell responses.

We apply the individual modules and the whole pipeline to four different solid tumor cohorts from TCGA, skin melanoma, lung adenocarcinoma, breast carcinoma and head and neck squamous carcinoma. From the thymic education model, we observe that the selfness is largely defined by similarity in expression values to the medically thymic epithelial cells, which can be utilized to compute a ‘selfness score’ for each patient sample. We observe that the tumors exhibit an increased selfness as compared to their adjacent non-tumours in the same individuals in 8 different cancer cohorts. We show that the score is also indicative of patient survival in many of these cohorts, indicating that gene expression modulation in the tumors could be a potential immune sculpting mechanism.

Next, we observe DeepHLAffy and DIPPred are able to correctly predict the known tumor antigens. Their application to studying the solid tumor cohorts resulted in predicting the immunopeptidomes in each patient in each of the cohorts, which includes several new antigens that are self aberrantly expressed tumor associated antigens, which were previously unknown. Application of the integrated iTRPred pipeline, resulted in quantifying CD8 T-cell responses in individual patients. The CD8 T-cell response scores in different patients spanned a wide range and were correlated with survival. Our response scores therefore are seen to have the prognostic power in predicting disease outcomes in multiple cancer cohorts, with the underlying biology captured being consistent with T-cell activation.