Developing and validating an approach for diagnosing and prognosticating cancer from biochemical motifs in T-cell receptors.

Abstract

e15260 Background: Immune repertoire deep sequencing allows profiling T-cell populations and enables novel approaches to diagnose and prognosticate cancer by identifying T-cell receptor sequence patterns associated with clinical phenotypes and outcomes. Methods: Our goal is to develop a method to diagnose and prognosticate cancer using sequenced T-cell receptors. To determine how to profile the specificity of a T-cell receptor, we analyze 3D X-ray crystallographic structures of T-cell receptors bound to antigen. We observe a contiguous strip typically 4 amino acid residues in length from the complimentary determining region 3 (CDR3) lying in direct contact with the antigen. Based on this observation, we extract 4 residue long snippets from every receptor’s CDR3 and represent each snippet using biochemical features encoded by its amino acid sequence. The biochemical features are combined with information about the abundance of the snippet or the receptor and scored using a machine learning based approach. Each predictive model is fitted and validated under the requirement that at least one positively labelled snippet appears per tumor and no positively labelled snippets appear in healthy tissue. Results: Using a patient-holdout cross-validation, we fit predictive models to distinguish: 1. colorectal tumors from healthy tissue matched controls with 93% accuracy, 2. breast tumors from healthy tissue matched controls with 94% accuracy, 3. ovarian tumors from non-cancer patient ovarian tissue with 95% accuracy (80% accuracy on a blinded follow-up cohort) 4. and regression of preneoplastic cervical lesions over 1 year in advance with 96% accuracy. Conclusions: Immune repertoires can be used to diagnose and prognosticate cancer.