Combining gene expression data from inflamed tissue and machine learning for blood biomarker discovery

Abstract

Abstract Inflammation is common in virtually all diseases. Since immune cells travel through the blood to inflamed tissues, peripheral blood reflects inflammation at distant sites. However, blood biomarkers are hard to detect in global gene expression data since blood is not the site of perturbation and statistical corrections for multiple testing reduce sensitivity. We hypothesized that combining transcriptional signatures from inflamed tissue and machine learning would be useful to derive blood biomarkers. We first tested this in the mouse model of influenza infection. We identified a set of early genes in the lungs associated with lethality after infection. We then tested their differential abundance early in the blood of mice infected with a range of influenza doses. We applied elastic net to build a biomarker able to predict disease outcome (4 classes: early lethal, late lethal, non-lethal vs. non-infected; accuracy on training and validation sets: 91% and 79%). We next confirmed our hypothesis in lung cancer and breast cancer in humans. We identified sets of genes associated with tumors but not with tumor-free tissues and tested their differential abundance in the blood of patients compared to healthy controls. Using elastic net, we built two blood biomarkers to classify subjects according to disease status (case vs. control; accuracy on training and validation sets in breast cancer: 92% and 67%; in lung cancer: 89% and 78%). In conclusion, combining inflamed tissue signatures and machine learning, we were able to derive blood biomarkers for three different diseases in two different species. We expect that this novel strategy can be applied in other diseases as well. This work was supported by the Intramural Research Program of NIAID, NIH.