Meta-Analysis of in vitro-Differentiated Macrophages Identifies Transcriptomic Signatures That Classify Disease Macrophages in vivo

Abstract

Macrophages are heterogeneous leukocytes regulated in a tissue- and disease-specific context. While in vitro macrophage models have been used to study diseases empirically, a systematic analysis of the transcriptome thereof is lacking. Here, we acquired gene expression data from eight commonly-used in vitro macrophage models to perform a meta-analysis. Specifically, we obtained gene expression data from unstimulated macrophages (M0) and macrophages stimulated with lipopolysaccharides (LPS) for 2-4 h (M-LPS_early), LPS for 24 h (M-LPS_late), LPS and interferon-γ (M-LPS+IFNγ), IFNγ (M-IFNγ), interleukin-4 (M-IL4), interleukin-10 (M-IL10), and dexamethasone (M-dex). Our meta-analysis identified consistently differentially expressed genes that have been implicated in inflammatory and metabolic processes. In addition, we built macIDR, a robust classifier capable of distinguishing macrophage activation states with high accuracy (>0.95). We classified in vivo macrophages with macIDR to define their tissue- and disease-specific characteristics. We demonstrate that alveolar macrophages display high resemblance to IL10 activation, but show a drop in IFNγ signature in chronic obstructive pulmonary disease patients. Adipose tissue-derived macrophages were classified as unstimulated macrophages, but acquired LPS-activation features in diabetic-obese patients. Rheumatoid arthritis synovial macrophages exhibit characteristics of IL10- or IFNγ-stimulation. Altogether, we defined consensus transcriptional profiles for the eight in vitro macrophage activation states, built a classification model, and demonstrated the utility of the latter for in vivo macrophages.

Keywords: adipose tissue macrophages (ATMs); alveolar macrophages (AMs); elastic net classification; macrophage identifier (macIDR); macrophages; meta-analysis; synovial macrophages (SMs).

Figure 1

Overview study design. Transcriptome datasets were found on the Gene Expression Omnibus (GEO) or ArrayExpress (AE) and screened according to the inclusion criteria yielding 18 datasets. A separate meta-analysis and classification analysis was performed and results thereof were subjected to functional pathway analyses.

Figure 2

Summary meta-analysis. (A) Heatmap of the Cohen d pairwise Spearman correlation coefficients. (B) Principal component analysis of the Z-values obtained from the meta-analysis. (C) Protein-protein association network as obtained from the top 100 cDEGs using the STRING database. Node colors represent the unbiased estimator of the effect size (mu), whereas the edge colors and thickness represent the source of the cataloged association and the weight of the evidence. (D) Heatmap of the canonical pathways with the intensity representing the activation z score. Two most defining clusters have been enlarged and annotated on the right.

Figure 3

Summary classification analysis. (A) Heatmap of the median-stabilized log odds ratios per macrophage activation state for each of the 97 predictor genes. (B) Confusion matrix representing the number of correctly classified samples (entries on the diagonal) vs. the misclassified samples (entries on the off-diagonal). Classes on the y-axis represents the reported class while classes on the x-axis represent the predicted class. Colors represent the predicted classes with purple, red, yellow, and blue representing M-dex, M-LPS+IFNγ, M-IFNγ, and M-IL10, respectively. (C) Bar plots of the misclassified samples depicting the classification signal on a scale of 0 to 1 where the class with the largest signal represents the predicted class. Blue bars represent the incorrectly predicted class and orange bars represent the reported class. Border colors represent the predicted classes with purple, red, yellow, and blue representing M-dex, M-LPS+IFNγ, M-IFNγ, and M-IL10, respectively. (D) Boxplots representing the classification signal on a scale of 0 to 1 where classes with the largest signal represents the predictions. Colors represent GM-CSF differentiated macrophages (GM-MDMs), monocyte-derived dendritic cells (MoDCs), fibroblast-like synoviocytes (FLS), B lymphocytes (B), T lymphocytes (T), natural killer cells (NK), and neutrophils (NP). (E) GM-MDMs and (F) MoDCs colored by stimulation.

Figure 4

Classification of in vivo macrophages. Summarized classification results per dataset with cross-bars representing the mean and the standard errors of the log odds colored by the macrophage in vivo type. Dots above represent the log odds ratio [log(OR)] relative to the sum of the log odds ratios if all predictor genes were measured. (A) Alveolar macrophages obtained from smoking individuals, chronic obstructive pulmonary disease (COPD), asthma patients, as well as healthy controls. (B) Adipose tissue macrophages obtained from diabetic obese and non-diabetic obese patients. (C) Synovial macrophages obtained from rheumatoid arthritis (RA) patients and MDMs from healthy controls (HCs).

Figure 5

Analysis of the rheumatoid arthritis-derived synovial cells. (A) t-distributed stochastic neighbor embedding (tSNE) visualization of the synovial biopsy-derived cells as obtained through Louvain clustering. (B) Pie chart depicting the frequency of each macrophage activation model as predicted by macIDR.