Graphical modeling of gene expression in monocytes suggests molecular mechanisms explaining increased atherosclerosis in smokers

Abstract

Smoking is a risk factor for atherosclerosis with reported widespread effects on gene expression in circulating blood cells. We hypothesized that a molecular signature mediating the relation between smoking and atherosclerosis may be found in the transcriptome of circulating monocytes. Genome-wide expression profiles and counts of atherosclerotic plaques in carotid arteries were collected in 248 smokers and 688 non-smokers from the general population. Patterns of co-expressed genes were identified by Independent Component Analysis (ICA) and network structure of the pattern-specific gene modules was inferred by the PC-algorithm. A likelihood-based causality test was implemented to select patterns that fit models containing a path "smoking→gene expression→plaques". Robustness of the causal inference was assessed by bootstrapping. At a FDR ≤0.10, 3,368 genes were associated to smoking or plaques, of which 93% were associated to smoking only. SASH1 showed the strongest association to smoking and PPARG the strongest association to plaques. Twenty-nine gene patterns were identified by ICA. Modules containing SASH1 and PPARG did not show evidence for the "smoking→gene expression→plaques" causality model. Conversely, three modules had good support for causal effects and exhibited a network topology consistent with gene expression mediating the relation between smoking and plaques. The network with the strongest support for causal effects was connected to plaques through SLC39A8, a gene with known association to HDL-cholesterol and cellular uptake of cadmium from tobacco, while smoking was directly connected to GAS6, a gene reported to have anti-inflammatory effects in atherosclerosis and to be up-regulated in the placenta of women smoking during pregnancy. Our analysis of the transcriptome of monocytes recovered genes relevant for association to smoking and atherosclerosis, and connected genes that before, were only studied in separate contexts. Inspection of correlation structure revealed candidates that would be missed by expression-phenotype association analysis alone.

Figure 1. Analysis workflow.

Microarray expression data were analyzed at two levels, a probe level (top) and a gene cluster level (bottom). Of the 35,358 probes with a “good” or “perfect” score according to ReMOAT, 23,214 were detected in monocytes of 936 subjects. Of these, 3,960 probes were associated to smoking or atherosclerotic plaques at an FDR ≤0.1, corresponding to 3,368 unique genes that were further clustered in 29 expression patterns by ICA. Causality testing revealed 4 patterns that were compatible with expression mediating the relationship between smoking and plaques. The skeleton of the network connecting genes, smoking, risk factors and plaques was then inferred using the PC algorithm.

Figure 2. Graphical models for equivalence classes tested among smoking, gene expression and atherosclerotic plaques.

Variables are represented by squares and causal associations are indicated by directed edges among nodes. Undirected edges indicate bidirected edges. The two classes colored in brown represent the causal models of interest where gene expression (G) mediates the association between smoking (S) and plaques (P). In class (f), the covariation between S and P is entirely explained by G, whereas in class (k), there is residual covariation between S and P after conditioning on G.

Figure 3. Probability of selection of the different causality models for each ICA expression pattern.

Probability was estimated from 1000 bootstraps of the data. The bottom of the graph shows the probabilities of the 11 models described in Figure 2. The top of the graph shows the sum of the probabilities for models (f) and (k) representing the causal classes.

Figure 4. Subnetwork of PC skeleton for Module 21.

This graph represents a consensus network from 1000 bootstraps. Edges among variables are drawn if detected in at least 60% of bootstrapped samples. The recovery percentages are indicated to the right of the medial section of each edge. Line thickness is proportional to the edge's partial correlation. Black edges denote positive and pink edges negative partial correlations. Plaques and risk factors are in blue. Genes directly connected to smoking are in green and those directly connected to plaques are in orange. Other genes are in gray. Only genes that are involved in the shortest paths connecting smoking to plaques are shown. The full network for this and other patterns are found in Text S2.