Network inference algorithms elucidate Nrf2 regulation of mouse lung oxidative stress

Abstract

A variety of cardiovascular, neurological, and neoplastic conditions have been associated with oxidative stress, i.e., conditions under which levels of reactive oxygen species (ROS) are elevated over significant periods. Nuclear factor erythroid 2-related factor (Nrf2) regulates the transcription of several gene products involved in the protective response to oxidative stress. The transcriptional regulatory and signaling relationships linking gene products involved in the response to oxidative stress are, currently, only partially resolved. Microarray data constitute RNA abundance measures representing gene expression patterns. In some cases, these patterns can identify the molecular interactions of gene products. They can be, in effect, proxies for protein-protein and protein-DNA interactions. Traditional techniques used for clustering coregulated genes on high-throughput gene arrays are rarely capable of distinguishing between direct transcriptional regulatory interactions and indirect ones. In this study, newly developed information-theoretic algorithms that employ the concept of mutual information were used: the Algorithm for the Reconstruction of Accurate Cellular Networks (ARACNE), and Context Likelihood of Relatedness (CLR). These algorithms captured dependencies in the gene expression profiles of the mouse lung, allowing the regulatory effect of Nrf2 in response to oxidative stress to be determined more precisely. In addition, a characterization of promoter sequences of Nrf2 regulatory targets was conducted using a Support Vector Machine classification algorithm to corroborate ARACNE and CLR predictions. Inferred networks were analyzed, compared, and integrated using the Collective Analysis of Biological Interaction Networks (CABIN) plug-in of Cytoscape. Using the two network inference algorithms and one machine learning algorithm, a number of both previously known and novel targets of Nrf2 transcriptional activation were identified. Genes predicted as novel Nrf2 targets include Atf1, Srxn1, Prnp, Sod2, Als2, Nfkbib, and Ppp1r15b. Furthermore, microarray and quantitative RT-PCR experiments following cigarette-smoke-induced oxidative stress in Nrf2(+/+) and Nrf2(-/-) mouse lung affirmed many of the predictions made. Several new potential feed-forward regulatory loops involving Nrf2, Nqo1, Srxn1, Prdx1, Als2, Atf1, Sod1, and Park7 were predicted. This work shows the promise of network inference algorithms operating on high-throughput gene expression data in identifying transcriptional regulatory and other signaling relationships implicated in mammalian disease.

Figure 1. CLR algorithm results showing connections to gene Nfe212 (protein Nrf2).

Regulatory interactions involving Nrf2 as determined using the CLR algorithm. Across 260 microarrays, profiles of genes categorized by the Gene Ontology as participating in the response to oxidative stress were examined. Z-scores were calculated on the basis of the CLR mutual information based values. At a z-score cutoff of 2.0 (two standard deviations above the mean score of all pair-wise CLR calculations), eighteen edges were reported that involved any of the Affymetrix probe sets representing the Nfe2l2 gene. These edges are shown in Figure 1. Thirteen of the eighteen putative edges had z-scores of 2.45 or higher. Some of these edges had the same gene at the other end (duplicate edges from the different Nfe2l2 probe sets), resulting in a total of twelve genes shown connected to Nfe2l2 in Figure 1 and twelve entries reported for CLR in Table 1. The nodes represent genes and the lines (edges) between them represent transcriptional regulatory relationships. Interactions involving Nrf2 (Gene Symbol: Nfe2l2) are depicted in this diagram. Multiple edges between two nodes indicate multiple array probe-sets on the arrays referencing the same gene.

Figure 2. ARACNE algorithm results showing connections to gene Nfe212 (Nrf2).

Regulatory interactions involving Nrf2 as determined using the ARACNE algorithm. Across 260 microarrays, profiles of genes categorized by the Gene Ontology as participating in the response to oxidative stress were examined. The DPI tolerance was set at 0.15; p = 1e-7. The nodes represent genes and the lines (edges) between them represent transcription regulation relationships. Interactions involving Nrf2 (Gene Symbol: Nfe2l2) are depicted in this diagram. Multiple edges between two nodes indicate multiple array probe-sets on the arrays referencing the same gene.

Figure 3. Oxidative stress markers in Nrf2^+/+ and Nrf2^−/− cigarette smoke (CS)-exposed or air-exposed lungs.

Figure 3A - lower induction values of total GSH were observed after CS exposure in Nrf2^−/− (NO) lungs after CS as compared to Nrf2^+/+ (WT) CS exposed lungs. Figure 3B - levels of TBARS (marker of lipid peroxidation) were elevated in NOCS lungs as compared to WTCS lungs. The data is shown as Mean±SD based on three replicates (n = 3) in each of the four conditions.

Figure 4. Oxidative stress-mediated induction in Sod1, Nqo1 and Als2 mRNA.

Increases in expression of Sod, Nqo1, and Als2 mRNA only in Nrf2^+/+ (WT) CD-1 mice but not Nrf2^−/− (NO) mice following cigarette smoke (CS) exposure. This figure depicts mean (n = 3) mRNA expression from the microarrays on Nrf2^+/+ air-exposed (WTAir), Nrf2^+/+ CS-exposed (WTCS), Nrf2^−/− air-exposed (NOAir) and Nrf2^−/− CS-exposed (NOCS). Results shown suggest regulation of these genes by Nrf2.

Figure 5. Oxidative stress-mediated induction of numerous predicted Nrf2 associated genes.

Nqo1, Sod1, Ercc6, Prdx6, Als2, Txnrd2, Park7, Srxn1 and Epas1 mRNA were induced selectively more in Nrf2^+/+ mouse lungs after CS exposure. Some of the Nrf2-associated predicted genes as Sod2, Ppp1r15b, Fos and Jun either show no differential induction or an inverse relation with Nrf2 gene. Nrf2 apparently exerts a negative regulatory influence on the expression of Sod2, Ppp1r15b and Fos. The results are plotted as relative fold changes (RFC) with WT air (WTAir) as the baseline for three replicates (n = 3).

Figure 6. ARACNE algorithm results showing connections to genes Nqo1 and Nfe212 (Nrf2).

Results of ARACNE runs on microarray data showing extension of the Nfe212 (Nrf2) network from Nqo1, one of its targets. Transcriptional regulatory interactions involving Nrf2 and Nqo1 as determined using the ARACNE algorithm. Across 260 microarrays, profiles of genes categorized by the Gene Ontology as participating in the response to oxidative stress were examined. The DPI tolerance was set at 0.15; p = 1e-7. The nodes represent genes and the lines (edges) between them represent transcription regulation relationships. Interactions involving Nrf2 (Gene Symbol: Nfe2l2) and Nqo1 (one of its regulatory targets) are depicted. Multiple edges between two nodes indicate multiple array probe-sets on the arrays referencing the same gene.

Figure 7. Alternate view of the ARACNE algorithm results, here focused on the subset of connections that directly involve Nqo1.

Figure 8. A subset view of three-party direct dependencies involving Nrf2 (Nfe2l2), Nqo1 and gene “X” from the ARACNE algorithm results.

Gene X is a placeholder for any of these genes: Sod1, Srxn1, Txnrd2, Prdx1, Prdx2, Prdx6, Atf1, Park7 and Als2. Nrf2 transcriptionally regulates Nqo1 expression; this defines one of the three edges.

Figure 9. ARACNE algorithm results and a possible feed-forward loop.

A depiction of a possible feed-forward loop involving Nrf2, Sod1 and Nqo1 captured in the networks generated using ARACNE on microarray data. In order to assign directionality to the edges of the generated subgraph, there is a need for biological context: Nrf2 transcriptionally regulates both Sod1 and Nqo1. In lung epithelial cells (A549-S), inhibition of Nqo1 gives the same effect on the generation of hydrogen peroxide by low dose quinones as the introduction of exogenous Sod1 . Inference: Nqo1 is a repressor of Sod1.

Figure 10. Predicted feedforward loops involving Nrf2, Nqo1, Srxn1, Prdx1, Atf1, and Als2.

On the basis of LibSVM predictions (Tables 1 and 3), Nrf2, Nqo1, Srxn1, Prdx1, Atf1 and Als2 are all regulatory targets of Nrf2 transcriptional activity. ARACNE runs on 260 microarray data indicate direct dependencies between Nqo1 and Srxn1, Prdx1, Atf1, Als2 and Nrf2. In addition, there is evidence Atf1 acts as a transcriptional repressor on the anti-oxidant response element of another promoter . This figure captures these relationships. Transcriptional regulatory relationships are depicted by arrows and less well defined relationships are depicted with hidden detail (dotted lines).

Figure 11. A depiction of analyses across networks.

Use of the CABIN tool to conduct exploratory analysis for comparison and integration of interactions evidence obtained from the ARACNE and CLR algorithms along with the promoter region analysis using LibSVM and interaction evidence obtained using the Agilent Literature Search tools. The interactions involving Nrf2 are selected and highlighted in blue.