Fig. 6

Development of an OA cartilage gene signature. a Receiver operator characteristic (ROC) curves for gene classifiers arising from 10 randomized test and training cohorts using data from the RAAK study, n = 40.⁴⁹ Plot shows the true-positive rate (sensitivity, y-axis) against the false-positive rate (1-specificity, x-axis). Curves have been staggered at the origin (0,0) for clarity. The area under the receiver operator curve (AUROC), the capacity for a test to distinguish between two groups, is shown in the figure legend for each group of genes (Classifier). The ROC curve of the random classifier (AUROC = 0.5) is shown as a broken line bisecting the plot through the origin. In all cases, the AUROC for selected genes from the H4 module indicated that the probability that these classifiers would correctly rank healthy cartilage samples over OA samples was greater than the area under the curve of the random classifier. b Principal component analysis of gene expression data from two discriminatory genes (ARHGDIB, RGS5) derived from the human H4 module to exemplify the division of samples. Data arise from healthy (n = 7) and osteoarthritic (n = 33) cartilage samples (class) in young (16–18), mid-aged (>60), and aged (>70) individuals (see key). The first two principal components (PC1, PC2) described most of the variation between the young and mid-aged/aged samples. This combination of two genes had a representative low classification error (0.033) using 10-fold cross-validation as a robust estimate. Other combinations of genes were defined for each test, but in all cases ARHGDIB was the top-scoring gene (Supplementary Data SD42). c Expression of ARHGDIB in three age groups (top panel) was significantly different (p = 3.9e−05, Kruskall–Wallis test) with cartilage from young donors found to be lower. Expression of the cartilage hallmark gene collagen type II, COL2A1, (lower panel) was more variable across age groups (p = 3.7e−03)