Development of a blood-based gene expression algorithm for assessment of obstructive coronary artery disease in non-diabetic patients

Abstract

Background: Alterations in gene expression in peripheral blood cells have been shown to be sensitive to the presence and extent of coronary artery disease (CAD). A non-invasive blood test that could reliably assess obstructive CAD likelihood would have diagnostic utility.

Results: Microarray analysis of RNA samples from a 195 patient Duke CATHGEN registry case:control cohort yielded 2,438 genes with significant CAD association (p < 0.05), and identified the clinical/demographic factors with the largest effects on gene expression as age, sex, and diabetic status. RT-PCR analysis of 88 CAD classifier genes confirmed that diabetic status was the largest clinical factor affecting CAD associated gene expression changes. A second microarray cohort analysis limited to non-diabetics from the multi-center PREDICT study (198 patients; 99 case: control pairs matched for age and sex) evaluated gene expression, clinical, and cell population predictors of CAD and yielded 5,935 CAD genes (p < 0.05) with an intersection of 655 genes with the CATHGEN results. Biological pathway (gene ontology and literature) and statistical analyses (hierarchical clustering and logistic regression) were used in combination to select 113 genes for RT-PCR analysis including CAD classifiers, cell-type specific markers, and normalization genes.RT-PCR analysis of these 113 genes in a PREDICT cohort of 640 non-diabetic subject samples was used for algorithm development. Gene expression correlations identified clusters of CAD classifier genes which were reduced to meta-genes using LASSO. The final classifier for assessment of obstructive CAD was derived by Ridge Regression and contained sex-specific age functions and 6 meta-gene terms, comprising 23 genes. This algorithm showed a cross-validated estimated AUC = 0.77 (95% CI 0.73-0.81) in ROC analysis.

Conclusions: We have developed a whole blood classifier based on gene expression, age and sex for the assessment of obstructive CAD in non-diabetic patients from a combination of microarray and RT-PCR data derived from studies of patients clinically indicated for invasive angiography.

Clinical trial registration information: PREDICT, Personalized Risk Evaluation and Diagnosis in the Coronary Tree, http://www.clinicaltrials.gov, NCT00500617.

Figures

Figure 1

Gene discovery and algorithm development patient and logic flow schematic. Initial gene discovery (CATHGEN repository) included both diabetic and non-diabetic patients. Gene discovery from PREDICT involved non-diabetic patients in a paired microarray analysis, that yielded 655 significant genes in common with those from the CATHGEN arrays. For RT-PCR 113 genes were selected and tested on 640 PREDICT patient samples, from which the final algorithm was derived and locked.

Figure 2

RT-PCR analysis of diabetic status impact on significant genes from CATHGEN microarray analysis. Significance of individual genes selected from the CATHGEN microarray cohort in non-diabetic (ND) and diabetic (D) patients is shown. The sex/age adjusted p values from a CAD logistic regression analysis in each subset are plotted (log scale). Significant p values (<0.05) are indicated in red with gene symbols, non-significant ones in black.

Figure 3

Venn diagram of microarray, RT-PCR, and algorithm gene sources. A total of 7718 genes were identified, 2438 and 5935, respectively, from the CATHGEN and PREDICT microarray analyses, with an intersection of 655 genes. For the 113 RT-PCR genes, 52 were from PREDICT, 22 from CATHGEN, and 29 from both; 10 were either normalization genes or from previous studies [13]. The final algorithm contained 20 informative genes: 10 from both microarray studies, 8 PREDICT alone, and 2 CATHGEN alone.

Figure 4

Gene ontology analysis of 655 CAD genes identified from microarray studies. The 655 CAD genes identified were analyzed using the BINGO algorithm to ascertain significant biological processes. Significant processes (p < 0.01 after FDR correction) are colored with the gradient of p values reflected in the colors as indicated, and the biological process annotated. A total of 55 processes were significant in this analysis at p < 0.05.

Figure 5

Heat-Map representation of Hierarchical Clustering Results on 113 RT-PCR Genes. Clusters were generated by hierarchical clustering yielding 20 groups of correlated. Clusters were annotated as to cell type expression using BioGPS (http://www.biogps.gnf.org). Extent of correlation is indicated by color as shown in the bar.

Figure 6

Schematic of the final algorithm structure and genes. The algorithm consists of overlapping gene expression functions for men and women with a sex-specific linear age function for the former and a non-linear age function for the latter. The genes in each term and their weights are shown. For the gene expression components, 16/23 genes in 4 terms are gender independent: Term 1 - neutrophil activation and apoptosis, Term 3 - NK cell activation to T cell ratio, Term 4, B to T cell ratio, and Term 5 -AF289562 expression normalized to TFCP2 and HNRPF. In addition, Term 2 consists of 3 sex-independent neutrophil/innate immunity genes (S100A8, S100A12, CLEC4E) normalized to overall neutrophil gene expression (AQP9, NCF4) for women and to RPL28 (lymphocytes) for men. The final male specific term is the normalized expression of TSPAN16. Algorithm score is calculated as described (Additional file 1).

Figure 7

ROC analysis of final algorithm. The cross-validated ROC curve for the final algorithm in the algorithm development cohort is shown. The AUC is 0.77 ± 0.04.