Sarcoidosis blood transcriptome reflects lung inflammation and overlaps with tuberculosis

Abstract

Rationale: Sarcoidosis is a granulomatous disease of unknown etiology, although M. tuberculosis may play a role in the pathogenesis. The traditional view holds that inflammation in sarcoidosis is compartmentalized to involved organs.

Objectives: To determine whether whole blood gene expression signatures reflect inflammatory pathways in the lung in sarcoidosis and whether these signatures overlap with tuberculosis.

Methods: We analyzed transcriptomic data from blood and lung biopsies in sarcoidosis and compared these profiles with blood transcriptomic data from tuberculosis and other diseases.

Measurements and main results: Applying machine learning algorithms to blood gene expression data, we built a classifier that distinguished sarcoidosis from health in derivation and validation cohorts (92% sensitivity, 92% specificity). The most discriminative genes were confirmed by quantitative PCR and correlated with disease severity. Transcript profiles significantly induced in blood overlapped with those in lung biopsies and identified shared dominant inflammatory pathways (e.g., Type-I/II interferons). Sarcoidosis and tuberculosis shared more overlap in blood gene expression compared with other diseases using the 86-gene signature reported to be specific for tuberculosis and the sarcoidosis signature presented herein, although reapplication of machine learning algorithms could identify genes specific for sarcoidosis.

Conclusions: These data indicate that blood transcriptome analysis provides a noninvasive method for identifying inflammatory pathways in sarcoidosis, that these pathways may be leveraged to complement more invasive procedures for diagnosis or assessment of disease severity, and that sarcoidosis and tuberculosis share overlap in gene regulation of specific inflammatory pathways.

Figure 1.

Flow diagram of bioinformatic analyses presented in the study. (A) Overview of random forest classifier development using derivation and validation blood gene expression datasets. (B) Overview of the analyses used to compare gene expression in blood versus lung biopsies. (C) Overview of blood gene expression datasets used to qualitatively compare gene expression between cohorts of sarcoidosis, tuberculosis, hypersensitivity pneumonitis, and other infectious and Th1-type inflammatory diseases.

Figure 2.

Receiver operator characteristic curve analysis of subject classification derived using the random forest algorithm on blood gene expression data. (A) University of California, San Francisco (UCSF) Derivation (training) set performance for all patients with sarcoidosis (n = 38) versus control subjects (n = 20). (B) Performance of the model using an external validation test set (Oregon, 12 patients with sarcoidosis and 12 control subjects). Derivation set performance is estimated using the out-of-bag error rate that is returned by the random forest algorithm, whereas the validation set performance is a comparison of the predicted versus the actual class for each subject.

Figure 3.

Principal component analysis (PCA) using (A) the 100 genes from blood most important in the classification algorithm based on their value in discriminating groups (defined in random forests as the Gini coefficient). Patients with sarcoidosis who were taking corticosteroids were intermixed with patients who were not, and both groups were well distinguished from healthy control subjects. PC1 and PC2 denote the first and second principal components. (B) PCA using the 10 most important genes in the classification algorithm suggesting that smaller subsets of genes have the potential to retain discriminative capacity. (C and D) Similar findings were derived using PCA of the Oregon dataset.

Figure 4.

Quantitative PCR confirmation of array-based blood gene expression measurement from 10 of the most important classifier genes (as ranked by the Gini coefficient). The correlation of array to quantitative PCR data was statistically significant for all genes.

Figure 5.

Expression levels of the 10 most discriminative genes from blood in classifying patients with sarcoidosis. Relative gene expression levels of microarray data are compared between the internal derivation dataset (University of California, San Francisco [UCSF]: 20 healthy control subjects and 38 patients with sarcoidosis) and the external validation dataset (Oregon: 12 healthy control subjects and 12 patients with sarcoidosis). For the UCSF dataset, all 10 genes were statistically significantly different when comparing sarcoidosis groups with the control group. *P < 0.05. ^†P < 0.05 when comparing patients with sarcoidosis with normal lung function to those with low lung function.

Figure 6.

Overlap in differentially expressed genes in blood and lung tissue in sarcoidosis. Dots represent the mean log fold change in microarray gene expression of individual genes measured in lung biopsies (OHIO dataset; n = 6 patients with sarcoidosis and 6 control subjects) versus whole blood (University of California, San Francisco [UCSF] dataset; n = 38 patients with sarcoidosis and 20 control subjects). As described in the online supplement, we identified significant gene expression overlap (red dots) as those genes with an FDR q value of less than 0.1 in both datasets and a greater than 2-fold change in either dataset. Specific genes, such as those that are interferon-inducible or related to the interferon signaling pathway and IL-15, are highlighted.

Figure 7.

The distribution of blood gene expression patterns between sarcoidosis and active pulmonary tuberculosis (PTB). Relative gene expression was assessed by generation of heat maps (left) and principal component analysis (right) as described in Materials and Methods. Other infectious and Th1-type inflammatory diseases are included for comparison. (A) Analyses using the most discriminative genes identified by the random forest classifier built to distinguish patients with sarcoidosis from control subjects. (B) Analyses using the PTB-specific gene signature described by Berry and colleagues (29). (C) Analyses using the top 50 most discriminative genes identified by the random forest classifier built to distinguish PTB from sarcoidosis. For heat map figures, genes are listed from top to bottom, and disease cohorts are listed below heat map. From left to right: SARC = sarcoidosis (Oregon, n = 12; University of California, San Francisco [UCSF], n = 38); PTB = pulmonary TB (UK, n = 13; UK, n = 21); LTB = latent TB (UK, n = 17; UK, n = 21); PSLE = pediatric systemic lupus erythematosis (n = 49); ASLE = adult systemic lupus erythematosis (n = 28); HP = hypersensitivity pneumonitis (UCSF, n = 6); STAPH = Staphylococcus aureus infection (n = 40); STREP = group A Streptococcus infection (n = 12).