Evaluation of cross-platform compatibility of a DNA methylation-based glucocorticoid response biomarker

Abstract

Background: Identifying blood-based DNA methylation patterns is a minimally invasive way to detect biomarkers in predicting age, characteristics of certain diseases and conditions, as well as responses to immunotherapies. As microarray platforms continue to evolve and increase the scope of CpGs measured, new discoveries based on the most recent platform version and how they compare to available data from the previous versions of the platform are unknown. The neutrophil dexamethasone methylation index (NDMI 850) is a blood-based DNA methylation biomarker built on the Illumina MethylationEPIC (850K) array that measures epigenetic responses to dexamethasone (DEX), a synthetic glucocorticoid often administered for inflammation. Here, we compare the NDMI 850 to one we built using data from the Illumina Methylation 450K (NDMI 450).

Results: The NDMI 450 consisted of 22 loci, 15 of which were present on the NDMI 850. In adult whole blood samples, the linear composite scores from NDMI 450 and NDMI 850 were highly correlated and had equivalent predictive accuracy for detecting DEX exposure among adult glioma patients and non-glioma adult controls. However, the NDMI 450 scores of newborn cord blood were significantly lower than NDMI 850 in samples measured with both assays.

Conclusions: We developed an algorithm that reproduces the DNA methylation glucocorticoid response score using 450K data, increasing the accessibility for researchers to assess this biomarker in archived or publicly available datasets that use the 450K version of the Illumina BeadChip array. However, the NDMI850 and NDMI450 do not give similar results in cord blood, and due to data availability limitations, results from sample types of newborn cord blood should be interpreted with care.

Keywords: 450K versus 850K; Algorithmic biomarker; Cord blood; DNA methylation; Dexamethasone; Glucocorticoid; Whole blood.

Fig. 1

Workflow. A Venn diagram demonstrating the neutrophil-specific CpG probes present on 450K and 850K platforms, the output of elastic net regression, and calculation of the linear composite score for NDMI 450 and NDMI 850. Of the 2621 neutrophil-specific probes of interest on the 850K array, 897 probes were available in the 450K. While for NDMI 850, elastic net chose 28 CpG probes as discriminative of DEX use, 22 were chosen for NDMI 450. 15 of these probes were shared between the two algorithms. The coefficients of the respective CpG probes were used to calculate the linear composite score (NDMI score) for each patient

Fig. 2

High correlation between NDMI 450 and NDMI 850 scores in the IPS and AGS samples. Those taking dexamethasone at the time of blood draw are in blue, and those who are not are in red. A In the AGS pilot cases comparing the same set of patients run on both arrays, the Pearson correlation was r = 0.97 (p < 0.0001). B In IPS training data, correlation was r = 0.99 (p < 0.0001). C In AGS glioma cases, those with IDH/1p19q/TERT-WT WHO 2016 classification of gliomas denoted by circles had r = 0.97 (p < 0.0001), and other AGS gliomas denoted by triangles had r = 0.98 (p < 0.0001). D In the adult controls, the correlation was r = 0.97 (p < 0.0001)

Fig. 3

Comparable classification accuracy of DEX exposure by NDMI 450 in IPS and AGS patients. A In the training set (IPS pre-surgery samples), the AUROC of the NDMI 450 (blue) is compared to that of the NDMI 850 (the gold standard, light blue), which was 99.4% (95% CI 98.7%, 100%). B In the test set of 552 AGS cases and controls used to evaluate NDMI 850, the AUROC was 92.3% (95% CI 88.5%, 96%) for the NDMI 450 (red) and 91.9% (95% CI 88.3%, 95.6%) for the NDMI 850 (light orange)

Fig. 4

On the 850K array, NDMI 450 and NDMI 850 scores were consistent in whole blood. In cord blood on the 850K array, NDMI scores between the two algorithms were very different. NDMI 450 scores were a lot lower than that of AGS controls, but NDMI 850 scores in cord blood were relatively similar to that of AGS controls. A The first four rows denote the adult whole blood samples from the AGS and IPS studies that were run on the 850K array. The NDMI 450 and NDMI 850 score distributions are similar regardless of platform. In the next three rows, with the first one being the 850K cord blood datasets combined for the purpose of analyzing phenotypes of NGT and GDM, NDMI 450 and NDMI 850 score distributions are significantly different from each other, but NDMI 850 scores from cord blood are similar to that of the adult whole blood controls. The difference between the two score distributions is less extreme when the 2 850K datasets are combined. B Scatterplot of NDMI 450 and NDMI 850 in the combined 850K cord blood dataset with moderate correlation (r = 0.77, p < 0.0001), but NDMI 450 scores are more negative. Normal glucose tolerance (NGT) samples are denoted in red, and gestational diabetes mellitus (GDM) in blue. C The absolute differences in methylation values of the 22 CpGs in NDMI 450 between the AGS controls (whole blood) and 850K cord blood. The 15 shared CpG probes with NDMI 850 are shown in green, and the 7 unshared probes are in blue. The color of the points denotes the degree of absolute differences

Fig. 5

Comparison of NDMI 450 and immune profiles in predicting maternal risk factors from cord blood. Three logistic regressions were run on the GEO cord blood datasets, with the independent variable set as the phenotype, and NDMI 450 score (purple), immune cell proportions (green), or NDMI 450 and immune cell proportions (blue) as predictors. Significance is denoted by the shading and shape of the symbol for the model estimate. A GSE152380: 450K, full-term (ref.) versus preterm. NDMI 450 score was significant in the univariate model, and in the model with just cell proportions, NK cell and B cell proportions were significant in distinguishing full-term versus preterm. CD4T cell proportion was a significant predictor in the third model, but NDMI score was not. B GSE104376: 450K, newborns born to low (ref.) versus high anxiety individuals. C GSE153219: 450K, normal for glucose tolerance (ref.) versus gestational diabetes mellitus. In the second model, CD8T cell proportions were a significant predictor. D GSE122288, GSE122086: 850K, normal for glucose tolerance (ref.) versus gestational diabetes mellitus