Identification of a gene-expression predictor for diagnosis and personalized stratification of lupus patients

Abstract

Systemic lupus erythematosus (SLE) is an autoimmune disease characterized by a wide spectrum of clinical manifestations and degrees of severity. Few genomic biomarkers for SLE have been validated and employed to inform clinical classifications and decisions. To discover and assess the gene-expression based SLE predictors in published studies, we performed a meta-analysis using our established signature database and a data similarity-driven strategy. From 13 training data sets on SLE gene-expression studies, we identified a SLE meta-signature (SLEmetaSig100) containing 100 concordant genes that are involved in DNA sensors and the IFN signaling pathway. We rigorously examined SLEmetaSig100 with both retrospective and prospective validation in two independent data sets. Using unsupervised clustering, we retrospectively elucidated that SLEmetaSig100 could classify clinical samples into two groups that correlated with SLE disease status and disease activities. More importantly, SLEmetaSig100 enabled personalized stratification demonstrating its ability to prospectively predict SLE disease at the individual patient level. To evaluate the performance of SLEmetaSig100 in predicting SLE, we predicted 1,171 testing samples to be either non-SLE or SLE with positive predictive value (97-99%), specificity (85%-84%), and sensitivity (60-84%). Our study suggests that SLEmetaSig100 has enhanced predictive value to facilitate current SLE clinical classification and provides personalized disease activity monitoring.

Fig 1. Co-expression analysis of the 100 meta-signature genes from the SLE training data sets.

Using EXALT meta-analysis, thirteen SLE signatures in columns with similar phenotypes indicated in S1 Table were displayed in a heat map with 100 genes (SLEmetaSig100) displayed in rows. The colors in the meta-heat map represent the direction of differential gene expression within a given transcriptional profile (red for up, green for down, and black for a missing match). Color intensity reflects the confidence levels of differential expression in the signatures.

Fig 2. Pathway analysis of SLEmetaSig100 genes.

Pathway analyses of SLEmetaSig100 genes identified genes involving DNA sensors and the cytokines constructing an innate immune DNA-sensor model. SLEmetaSig100 genes are marked in white circles or rectangles. DNA sensors include MB21D1(cGAS), multiple TLR genes, TMEM173/STING, and IF16 genes. In the Toll-like receptor signaling pathway, the stimulation of DNA sensor genes by microbe-derived and/or host DNA are positively regulated by MYD88 and TMEM173/STING genes and negatively regulated by TREX1 and TREX2 genes. The downstream cytokine-cytokine receptor interaction genes include NF-kappa B signaling pathway mediated IFNs, inflammatory cytokines (e.g. IL1R1), and STATs mediated chemokines (CXCL and CXCR genes).

Fig 3. Stratification of human samples by clustering the SLEmetaSig100 meta-profiles.

Meta-heat maps from unsupervised hierarchical clustering depict meta-profiles in two test data sets, (A) GSE65391 and (B) GSE11909. In each panel, the gene expression patterns from one given test data set are represented in rows and samples are clustered in columns. The colors in the heat map represent the direction of differential gene expression within a given transcriptional profile (red for up, green for down, and black for a missing match). Color intensity reflects the confidence levels of differential expression. Sample groups in columns are determined by the top hierarchy nodes of dendrograms (yellow bar) into left and right sample groups. The sample phenotype patterns underneath each sample dendrogram panel are indicated by black (SLE) and white (healthy) bars. The classification of healthy samples from total samples (healthy/total) by SLEmetaSig100 profiles between two sample groups (left and right) and statistic tests (Mann-Whitney U test) results are listed in a table (C).

Fig 4. Receiver operating characteristic (ROC) curves for SLEmetaSig100.

Area under receiver operating characteristic curve (AUC) for performance of SLEmetaSig100 were calculated in two testing cohorts, GSE65391(solid line) and GSE11909(dash line), and SLEmetaSig100 significantly outperforms the random prediction of SLE disease (AUC, 0.89 in GSE65391 and 0.85 in GSE11909). The sub-table shows SLEmetaSig100 prediction performance in two test datasets. *Note: SLE prediction by SLEmetaSig100 in two test data sets was examined by Fisher Exact test (P value = 1.48E-36).