Controlling for Confounding Factors when Diagnosing Cancer from T Cell Receptor Repertoires

Abstract

Abstract Our publications have shown that T cell receptor (TCR) repertoires sequenced from tissue biopsies can be used to distinguish tumor from control tissue with > 90% classification accuracies, suggesting that TCR repertoires can be used to diagnose cancer. TCR repertoires sequenced from blood samples of cancer patients are now publicly available, and it is tempting to use these existing samples to develop blood-based assays for diagnosing cancer. Already, research groups including us obtained promising results. However, these results are obtained by combining cancer and control samples from mismatching studies. Without controlling for age, sex, race, and even the reduced amount of blood volume collected from cancer patients relative to controls, these exciting results could merely be an artifact of confounding factors. In this study, we control for possible confounding factors and observe that our ability to diagnose cancer from TCR repertoires sequenced from blood does not vanish. On patient-holdout cross-validations, we achieve diagnostic accuracies of 90% for colorectal cancer (controlling for depth coverage), 84% for breast cancer (controlling for depth coverage and age), and 83% for lung cancer (controlling for age, sex, race, and depth coverage). While it is impossible to control for all potential confounding factors like variations in geographic locations, these results are encouraging. We propose future studies that (i) collect matching controls and cases and (ii) confirm TCR repertoire sequencing provides a distinct signal for diagnosing cancer independent of other blood-based assays and therefore can improve the diagnostic performance of these other blood-based assays.