Linking Co-regulated Gene Modules with Polycyclic Aromatic Hydrocarbon-related Cancer Risk in the 3D Human Bronchial Epithelium

Abstract

Exposure to polycyclic aromatic hydrocarbons (PAHs) often occur as complex chemical mixtures, which are linked to numerous adverse health outcomes in humans, with cancer as the greatest concern. The cancer risk associated with PAH exposures is commonly evaluated using the relative potency factor (RPF) approach, which estimates PAH mixture carcinogenic potential based on the sum of relative potency estimates of individual PAHs, compared to benzo[a]pyrene (BAP), a reference carcinogen. The present study evaluates molecular mechanisms related to PAH cancer risk through integration of transcriptomic and bioinformatic approaches in a 3D human bronchial epithelial cell model. Genes with significant differential expression from human bronchial epithelium exposed to PAHs were analyzed using a weighted gene coexpression network analysis (WGCNA) two-tiered approach; first to identify gene sets co-modulated to RPF, and second, to link genes to a more comprehensive list of regulatory values, including inhalation-specific risk values. Over 3,000 genes associated with processes of cell cycle regulation, inflammation, DNA damage, and cell adhesion processes were found to be co-modulated with increasing RPF with pathways for cell cycle S phase and cytoskeleton actin identified as the most significantly enriched biological networks correlated to RPF. In addition, co-modulated genes were linked to additional cancer-relevant risk values, including inhalation unit risks, oral cancer slope factors, and cancer hazard classifications from the World Health Organization’s International Agency for Research on Cancer (IARC). These gene sets represent potential biomarkers that could be used to evaluate cancer risk associated with PAH mixtures. Among the values tested, RPF values and IARC categorizations shared the most similar responses in positively and negatively correlated gene modules. Together, we demonstrated a novel manner of integrating gene sets with chemical toxicity equivalence estimates through WGCNA to understand potential mechanisms.

Graphical Abstract

INTRODUCTION

PAHs as global contaminants

Polycyclic aromatic hydrocarbons (PAHs) are a ubiquitous class of environmental contaminants found in ambient air, indoor air, water, and soil. Of hundreds of identified and studied PAHs, currently 16 PAHs are listed on the US Environmental Protection Agency (EPA) high priority pollutant list¹. These priority PAHs were identified based on relative abundance in the environment and toxicity, but there are PAHs beyond the list of 16 that have higher toxicity in complex environmental mixtures². Atmospheric PAHs can serve as an important source of PAH exposure and can participate in long range transport to settle in soil and water^3–7,8. The hazards and risks of PAH exposure may depend on the mixture composition and exposure route, which can include ingestion, inhalation, and dermal contact in occupational and environmental settings^8–12. Adverse human health outcomes can be diverse for the class, but include cancer as the health effect of greatest concern with the US EPA and the European Chemicals restricting the exposure and sale of petroleum products containing at least eight of these carcinogenic PAHs. Despite their importance and prominence in mixtures of concern, accurate human health risk assessment of PAHs remains a challenge.

Cancer risk assessment of PAHs

The current EPA framework for PAH regulation involves estimating cancer risk using a component-based relative potency factor (RPF) approach in which risk for PAH mixtures is estimated based on sum relative potency for component PAHs present in the mixture¹³. In this approach, benzo[a]pyrene (BAP) is used as a reference carcinogen and cancer risk is reported in terms of BAP equivalents (BAPeq). A major limitation of this approach includes its heavy reliance on dose-response assessments of traditional, two-year in vivo rodent carcinogenicity bioassays that are lengthy and resource-intensive. Further, RPF estimates exist for only 27 PAHs limiting applicability of the approach to the broader class of over 1500 polycyclic aromatic compounds, many of which co-occur in mixtures. In addition, RPF values are primarily derived from studies based on intraperitoneal injection or dermal exposure, which may not adequately represent alternate PAH routes of exposure in humans and could fail to capture the true risk of individuals PAHs. These data suggest an urgent need to develop alternative approaches for chemical hazard assessment and regulation¹⁴. In addition, other risk and regulatory values should be considered to adequately evaluate lung-specific PAH cancer hazards. Regulatory values that incorporate additional routes of exposure include inhalation unit risk (IUR), cancer slope factor, reference concentration (RfC), reference dose (RfD), and cancer hazard categorizations are available through the International Agency for Research on Cancer (IARC) and assessments from the US EPA Integrated Risk Information System (IRIS). Currently, there are limited investigations examining possible links between biological mechanisms and such increased cancer risks and toxicity values across different chemicals within a chemical class.

Utility of Transcriptional Signatures in Chemical Hazard Assessment

Transcriptomics has emerged as an informative method for evaluating biological functions and pathways targeted by xenobiotics. Transcriptional profiling of gene signatures has been especially useful in identifying mechanisms linked to adverse outcomes and cancer modes of action^15–17. Additional applications in toxicology include modeling transcriptional information for risk assessment^18–20 and evaluating chemical mixture interactions^21,22. Further applications of bioinformatics and transcriptomics have been used to evaluate biomarkers of chemical exposure at the transcriptional level²³.

Despite the acceptance of transcriptomics, standard methods for global gene expression analyses, which typically filter data based on differential expression criteria for pathway enrichment, can restrict the number of possible toxicogenomic biomarkers identified and limit downstream interpretation. We previously analyzed transcriptional gene signatures exposed to a range of carcinogenic PAHs and identified pathway gene sets for successful development of a classification model to accurately predict chemical carcinogenic potency^24,25. However, a limitation of this gene classifier identification was the reliance on pathway annotation. Novel bioinformatics approaches can help overcome these limitations. One such method, known as the weighted gene co-expression network analysis (WGCNA), provides a systems biology approach to explore genes regulated together without a priori knowledge^26,27. Researchers have noted the benefit of co-expression approaches where gene signature identification would be independent of pathway annotations²⁸. In addition, co-expression network analysis allows for identification of genes associated with a specific parameter rather than focusing on chemical-specific gene regulation. Groups of genes are identified as modules through an initial correlation analysis; then gene modules are related to traits or parameters, which can span dose, time, chemical traits, and biological phenotypes^26,29. This co-expression approach has been applied successfully in the field of toxicology to identify novel transcription factors associated with bisphenol A exposure³⁰ and genes and proteins linked to adverse health effects associated with increased inorganic arsenic exposure³¹.

While the cancer-related health effects of PAHs and PAH mixtures have been evaluated in numerous cell-based and animal studies, there remains a lack of understanding of the genes and pathways affected by different carcinogenic PAHs in an inhalation-relevant model. There also exist data gaps in understanding genes and signaling events implicated in chemical carcinogenesis. WGCNA can be used to correlate transcriptional changes with increasing cancer risk (RPF), thereby allowing comparisons across chemicals and mixtures with different relative potencies. Therefore, co-expression analysis serves as a novel approach to investigate PAH-induced system-wide biological changes in an organotypic tissue model and may contribute MOA information to future cancer hazard and risk assessment. The goal of the present study is to identify gene sets co-modulated in relation to estimated human cancer risk and regulatory values and to better understand biological signaling important for chemically-induced disease. Network analysis of these co-modulated molecules was carried out to understand the biological functions and signaling networks of genes most significantly co-modulated in relation to cancer risk.

METHODS

Chemicals and Reagents

Cell culture media and phosphate buffered saline (PBS) were provided by MatTek Corporation (Ashland, MA). Benzo[a]pyrene (BAP) CAS# 50-32-8 and dibenzo[def,p]chrysene (DBC) CAS# 189-64-0 were purchased from MRIGlobal (Kansas City, MO). Benz[a]anthracene (BAN) CAS# 56-55-3, phenanthrene (PHE) CAS# 85-01-8, and pyrene (PYR) CAS# 19-00-0 were acquired from Sigma-Aldrich Chemical (St. Louis, MO.). PAHs used in the simulated air mixture (AM) were purchased from Sigma-Aldrich Chemical (St.Louis, MO.) Coal Tar Extract (CTE) standard reference material (SRM 1597a) was purchased from the National Institute of Standards & Technology (Gaithersburg, MD.) DNase I, TRIzol^® reagent, Superscript® III First Strand Synthesis System, qPCR primers, and Pierce^™ LDH Cytotoxicity Assay Kit were from Thermo Fisher Scientific (Waltham, MA). 2X SsoAdvanced^™ Universal SYBER^®Green Supermix was purchased from BioRad Laboratories, Inc. (Hercules, CA.) A simulated AM designed from published results of air samples collected in Beijing, China and analyzed for parent, nitrated, and oxygenated PAHs (Wang et. al., 2011) was produced as follows. Five parent PAHs were selected based on their abundance and biological relevance. Three additional methylated PAHs were selected for addition of non-carcinogenic components. Benzo[b]fluoranthene CAS# 205-99-2, benz[e]pyrene CAS# 192-97-2, benzo[g,h,i]perylene CAS# 191-24-2, benzo[a]pyrene, dibenzo[def,p]chrysene, 2-methylnaphthalene CAS# 91-57-6, 1-methylnaphthalene CAS# 90-12-0, and 1,3-dimethylnaphthalene CAS# 575-41-7 were added in the reported proportions relative to the sum of the eight chosen PAHs (Supplemental File S1). The simulated mixture was solubilized in acetone vehicle.

Tissue Culture and Chemical Exposures

Tissue culture and chemical exposures were described previously³². Briefly, primary human bronchial epithelial cells (HBEC) differentiated on transwell inserts (EpiAirway^™ 100, Mattek, Ashland, MA) equilibrated for 24 hours at 37°C, 5% CO₂ followed by a change of fresh medium. Single PAHs and mixtures were solubilized in acetone and applied (0.01 ml/insert) to the apical surface of tissues (n=4 per treatment) for up to 48 hrs, BAP (1–500 ug/ml or 0.02–8.33 ug/cm²), DBC (1–50 ug/ml or 0.02–0.83 ug/cm²), BAN (10–500 ug/ml or 0.17–8.33 ug/cm²), PYR (10–250 ug/ml or 0.17–8.34 ug/cm²), PHE (10–250 ug/ml, or 0.17–4.17 ug/cm²), AM (125–2000 ug/ml or 2.08–33.3 ug/cm²) and CTE (250 −1500 ug/ml or 4.17–25 ug/cm²). Individual PAHs were chosen to span carcinogenic potency, ranging from non-carcinogenic to those with lower/higher potency than BAP¹³. Forty eight hours post treatment, PBS (0.20ml) was added to the apical surface of each insert and 48 hr transepithelial electrical resistance (TEER) recorded as previously reported^25,32. Apical washes and basal media were transferred to clean, sterile tubes and stored at −80°C. At the end of each exposure regimen, lysis buffer was added to each insert, collected, and stored at −80°C until extraction using RNeasy Mini kit (Qiagen, Venlo, Netherlands).

Transcriptomic Analyses

Total RNA samples were isolated from HBEC (n=4 per treatment) after treatment with 10, 100, and 500 ug/ml BAP, 10 ug/ml DBC, 100 ug/ml PHE, 500 ug/ml PYR, 500 ug/ml BAN, 1500 ug/ml CTE and 250 ug/ml AM. Dosing was chosen based on BAP equivalents at concentrations similar to those used in prior animal studies as previously reported. Briefly, RNA was isolated using RNeasy Mini Kit (Qiagen, Venlo, Netherlands) and was quantitated on a SYNERGY/HTX plate reader equipped with a Take3 module, then evaluated for quality with a Bioanalyzer 2100, Agilent (Santa Clara, CA). Samples were sequenced by Illumina HiSeq3000 through the Oregon State University’s Center for Genome Research and Biocomputing Core facilities yielding 22–25 million reads per sample as described previously^25,32. Sequenced reads were processed with Cutadapt (version 1.8.1) to trim adapter sequences from the paired-end reads. The human genome assembly GRCh38.84 was indexed using botwie2-build (version 2.2.3) while the transcriptome was indexed using Tophat (version 2.1.1.)³³. TopHat was used to align the trimmed reads to indexed transcriptome and genome³⁴. FeatureCounts, from the Subread package (version 1.6.0) was used to generate count tables.

Statistical Evaluation of Transcriptional Changes Associated with PAHs

Sequence count data were used to identify genes with expression levels associated with each PAH exposure. A comparative analysis of RNAseq data was determined in DESeq2 (version 1.26.0)³⁵. The non-normalized count table output from FeatureCounts was input into DESeq2 to estimate differential gene expression as logarithmic fold change (log2FC) compared to vehicle control. Genes included for further analysis in WGCNA met the statistical criteria of q<0.05 for at least one chemical exposure group with no undetected values in the remaining exposure groups.

Organizing Cancer Risk Values across Databases

We first used RPFs from the U.S. Environmental Protection Agency IRIS 2010 recommendation of an RPF approach for the carcinogenic risk assessment of PAH mixtures. The RPF approach assigns a numerical value to each chemical based on experimentally derived carcinogenic potency, scaled as greater or less than BAP (RPF of 1). The dose-specific cancer risk of PAH mixtures were calculated by multiplying available RPFs to the proportions (mg/kg) of the individual PAH present in the mixtures and reported as weighted averages (Supplemental File S2). Chemicals without RPF carcinogenic estimates were assigned values of 0 for the purpose of calculating a sum mixture cancer risk. Individual PAHs were calculated by multiplying dose with RPF.

Additional cancer-relevant regulatory values, such as oral slope factor (OSF), IUR, RFC and RFD were downloaded from databases, prioritized in order of availability from the EPA CompTox Chemistry Dashboard³⁶, then predicted from Conditional Toxicity Value Predictor (CTV)³⁷. While many of these health risk values were derived or predicted with variable confidence, it was important to consider these with the goal of including inhalation risk values, to serve as comparisons against RPF and OSF, which are derived through oral and dermal routes of exposure. IRIS and IARC categories also were converted to a numerical scale, with 0 representing least and 4 representing greatest hazard, or group 1. Chemicals with no IRIS or IARC information were grouped as the least hazard.

Identifying Genes with Co-Modulated Expression Patterns through WGCNA

To identify co-modulated gene sets within the transcriptomic dataset of PAH and PAH mixtures treated HBEC, WGCNA was employed to calculate integrated measures from co-modulated genes. WGCNA is a quantitative, systems biology approach that describes correlation patterns between groups of genes²⁶. All genes that were identified as significantly regulated across at least one treatment (13644 genes in total) were analyzed collectively using the WGCNA R package²⁷. WGCNA was used to identify clusters (modules) of highly interconnected genes across both experimentally and database-derived traits and regulatory values, all correlated with each other, and thus considered co-modulated. The steps have been previously summarized^{26–28,30,38}. Overall network and module statistics were performed following methods described by Langfelder et al.²⁷. First, Pearson correlation coefficients were calculated for all pairwise comparisons of genes. The resulting Pearson correlation matrix was transformed into an adjacency matrix resulting in a weighted network describing connection strengths between genes. To allow genes with both negative and positive correlations to be grouped together, an unsigned network was used. A power of 16 and minimum gene module size of 50 was used to construct co-expression gene networks that were weighted to prioritize genes with high correlations as previously described³¹. Modules were then identified as groups of interconnected genes in the weighted network analysis with high topological overlap, measured based on an average linkage hierarchical clustering. The resulting modules represent clusters of highly interconnected genes with high positive or negative correlations. With these defined modules, the first principal component of the modules were calculated and referred to as eigenmodules, then correlated with toxicity values where statistical significance was evaluated²⁶.

Network Analysis of Co-Modulated Genes associated with Cancer Risk Values

Network analysis was carried out to understand systems-level responses that occur within a co-modulated set of mRNAs identified through WGCNA. RPF-associated genes from prioritized module eigengene were further analyzed for pathway enrichment analysis using the Process Network knowledgebase in Metacore (Clarivate Analytics, Philadelphia, PA). In addition, protein-protein interactions were evaluated in Metacore to identify over-connected genes or hubs associated with the dataset. Statistical significance of over-connected interactions was calculated using a hypergeometric distribution, where the p value represents the probability of a particular mapping arising by chance for experimental data compared to the background³⁹. Average FDRs for network categories were calculated for further analysis. Network visualizations were prepared in Cytoscape (v3.8.0)⁴⁰ using gene and protein interactions provided from the Metacore Process Network knowledgebase. The functional networks were analyzed through the Cytoscape NetworkAnalyzer tool to calculate betweenness centrality for each gene based on interactions (undirected) described from Metacore as a way to prioritize the role of genes within the network⁴¹. Betweenness centrality provides a measure of how important the node is to the flow of information through a network and may represent important regulators in a biological network^42–45.

RESULTS

Organizing Cancer Risk Values across Databases

To relate human cancer risk to our PAH treatments, we first used RPFs from the U.S. Environmental Protection Agency Integrated Risk Information System’s (IRIS) 2010 recommendation of a component-based RPF approach for the carcinogenic risk assessment of PAH mixtures, in which RPFs for component PAHs are summed after scaling for concentration in the mixture¹³. Cancer potency factors (referred to as BAP equivalents, or BAPeq) were calculated using the RPF approach for PAHs with RPF estimates available. For the PAH mixtures, RPFs were missing for as many as 57 out of the 73 PAHs in Coal tar extract mixture (78%), and 4 out of the 8 PAHs in the air mixture sample (50%). Calculations by mass determined that 75% of the components in CTE were unaccounted for by RPF, while 58% of the PAHs by mass in AM were unaccounted for by RPF, which is a primary limitation of relying on RPF for estimating potency of mixtures. Additional cancer-relevant regulatory values, such as RFC, RFD, OSF and IUR were downloaded from databases, prioritized in order of availability from EPA IRIS, and then missing values were predicted using the CTV in silico prediction approach to fill most missing values (Supplemental File S2). Cancer-relevant risk and classification values from IRIS, and IARC cancer hazard classifications were included for completeness. While the applicability domain was high for almost all predicted values, indicating low confidence in those predictions, those values were included in the analysis to provide inhalation-relevant risk values for comparison.

Identifying Co-Modulated Genes associated with Cancer RPF

Overall, this study evaluates molecular mechanisms related to PAH cancer risk through coexpression analysis of global transcriptomic data collected in vitro with regulatory values relevant for PAH cancer risk. The effects of PAH and PAH mixture exposure in 3D HBEC were investigated using a transcriptomic profiling approach through RNAseq analysis. HBEC (n=4 per treatment) were treated with both individual PAHs and PAH mixtures, including 10, 100, and 500 ug/ml BAP, 10 ug/ml DBC, 100 ug/ml PHE, 500 ug/ml PYR, 500 ug/ml BAN, 1500 ug/ml CTE and 250 ug/ml AM, as proof-of-concept for utilizing coexpression analysis to identify genes co-modulated with cancer risk and regulatory values for PAH chemicals. Raw and normalized sequencing files are available online at NCBI Gene Expression Omnibus (GSE128471 and GSE156147). Over 60,000 mRNA transcripts were mapped and evaluated, and 13,644 genes were identified as statistically significant (q <0.05). The number of genes that showed significant differential expression varied according to each exposure condition, with numbers ranging from 3 (pyrene) to over 9000 (air mixture). Because the number of significantly regulated genes varied so highly between exposure groups, it was advantageous to employ an approach that evaluated co-modulated expression profiles regardless of significant differential expression.

This first tier of the analysis identified and prioritized module eigengenes, or gene modules, linked to RPF. Twenty-three gene modules were identified through WGCNA (Fig. 1A), representing groups of genes that are co-modulated across PAH and PAH mixtures treatments. Seven modules were significantly correlated to RPF cancer risk (Table 1). The top most significant positively and negatively correlated modules associated with RPF were prioritized for further analysis to identify which genes, cellular networks and processes were most co-modulated and involved in the processes of cellular response to carcinogenic PAH treatment. One module (“MEturquoise”) was identified as the most significantly positively correlated with RPF (p=0.01), while another module (“MEred”) was identified as the most significantly negatively correlated with RPF (p=0.0007). The MEturquoise module identified 2078 total co-modulated transcripts and included 7 miRNAs and 303 lncRNAs (Supplemental File S3). This module contained co-modulated genes related to cell cycle and inflammation, including breast cancer type 1 susceptibility gene (BRCA1), breast cancer type 2 susceptibility gene (BRCA2), nuclear receptor subfamily 2 group F member 2 (NR2F2), tumor protein p53 regulated apoptosis inducing protein 1 (TP53AIP1), JunB proto-oncogene, AP-1 transcription factor subunit (JUNB), and heme oxygenase 1 (HMOX1) (Fig. 1B). In the MEred gene module, 1021 transcripts were found to be significantly negatively co-modulated with RPF (Supplemental File S4). Many of these genes are downregulated in response to RPF and include 2 miRNAs and 126 lncRNAs (Fig. 1C). These include genes important for cell motility junction organization, including formin 1 (FMN1), dynein cytoplasmic 2 light intermediate chain 1 (DYNC2LI1), dynein axonemal light chain 1 (DNAL1), and claudin 8 (CLDN8).

Figure 1. Module-trait relationships and genes of interest (A) Module-trait relationships matrix. WGCNA identification of gene modules correlated to PAH cancer risk, calculated through RPF.

Seven module eigengenes (MEs) were highly correlated with RPF. Degree of correlation is colored, with red as positive and blue as negative, and p-values in parentheses. (B) MEturquoise genes of Interest. The most positively correlated module eigengene (“MEturquoise”) was prioritized for evaluation; genes of interest that had expression levels correlated to MEturquoise are shown. Expression levels are Z-score normalized. (C) MEred genes of interest. The most significantly negatively correlated module eigengene (“MEred”) was priotized for evaluation; genes of interest shown, ranked by RPF, with expression levels Z-score normalized.

Table 1.

Gene modules significantly (p< 0.05) associated with RPF.

Module	Correlation	p-value
Turquoise	0.79	0.01
Magenta	0.73	0.03
Lightgreen	−0.65	0.05
Black	−0.68	0.05
Grey60	−0.69	0.04
Royalblue	−0.72	0.03
Red	−0.82	0.007

Network Analysis of Co-modulated Genes in Prioritized Modules

To further assess biological relevance of the genes identified in each prioritized module, a pathway enrichment analysis of MEturquoise and MEred was conducted (Fig 2). Network analysis of genes was used to identify biological processes that were significantly enriched (FDR<0.05) within both prioritized modules, as well as processes uniquely enriched by one module and not the other. These biological processes were organized into broad biological categories for direct comparison between modules. Genes within the positively correlated MEturquoise module were significantly enriched (p<0.05) for biological processes related to cell cycle regulation, DNA damage, signal transduction, and cardiac development (Fig 2). These biological process categories were uniquely enriched by the MEturquoise module and MEred did not share any significant enrichment in these functional categories. In contrast, genes within the MEred module were significantly enriched (p<0.05) for processes related to cell adhesion, cytoskeleton, protein folding, transcription, and translation. Genes within both modules were enriched for processes related to inflammation (9 specific processes enriched by MEturquoise, and 15 specific processes enriched by MEred).

Figure 2. Network category table of gene modules MEturquoise and MEred that are most significantly correlated to PAH RPF.

Pathway enrichment analysis was conducted, and a broader network category-level significant enrichment score was calculated. Numbers indicate number of pathways enriched under each network category, and ave FDR (average false discovery rate) shown in parentheses. Distinct biological networks were enriched by MEturquoise and MEred, as well as several network categories commonly enriched, notably Inflammation.

To study network regulation of the genes in each prioritized pathways, process networks were visualized by comparing module membership of network genes. The Cell Cycle S phase network was the most significantly enriched network (FDR=1.16E-09) unique to MEturquoise (Fig 3A), while the Cytoskeleton-Actin filaments network was the most significantly enriched network (FDR=4.24E-06) unique to MEred (Fig 3B). The networks were analyzed through Cytoscape (v3.8.0) with the size of the nodes in the network mapped to calculated betweenness centrality for each gene and edges mapped to annotated interactions (undirected) from Metacore. The betweenness centrality parameter represents the extent to which a vertex is in the shortest path between other vertices. Genes with high betweenness centrality measures likely represent important regulators in biological processes. In the Cell Cycle S phase gene network, the top gene nodes with the highest centrality measures from MEturquoise were cyclin dependent kinase 1 (CDK1), minichromosome maintenance complex component 2 (MCM2), and proliferating cell nuclear antigen (PCNA). In the Cytoskeleton-Actin filaments network, the gene nodes with the highest centrality measures from MEred included actin gamma 1 (ACTG1), SRC proto-oncogene, non-receptor tyrosine kinase (SRC), actin beta (ACTB), and cortactin (CTTN). All of these genes were also identified as significantly over-connected hubs within the networks (p<0.005) based on the number of shared edges with network members.

Figure 3. Gene networks showing most significantly enriched networks associated with prioritized gene modules significantly correlated to PAH RPF. (A) MEturquoise significant enrichment of the Cell cycle S phase gene network, and (B) MEred significant enrichment of the Cytoskeleton actin filament gene network.

Node sizes denote betweennenss centrality measures calculated in each network. A larger node, or “bottleneck gene” thus represents a gene or biomolecule that stands in between the shortest paths connecting the maximum number of neighboring nodes to each other.

Identifying Co-Modulated Genes associated with Additional Cancer Regulatory Values

To evaluate relationships between co-modulated genes and regulatory values and hazard classifications, we performed a second analysis to identify co-modulated genes associated with additional database-derived values, particularly inhalation-relevant values such as IUR, RfC, OSF, RfD, and IRIS cancer risk values and IARC cancer hazard classifications, and compared to RPF (Fig. 4). The gene modules that correlated most significantly with RPF were MEfive, at p = 0.02 (positive correlation) and MEfourteen, at p = 0.02 (negative correlation) (Figure 4 and Table 2). A pathway enrichment analysis on the transcripts from MEfive and MEfourteen identified processes related to transcription, translation, and cell cycle regulation as among the top most significantly (p<0.05) enriched. Certain regulatory and hazard classifications’ correlations were found to be more similar to RPF’s module correlations than other regulatory values. Specifically, RPF and RfD did not share any significantly correlated module eigengenes; however RPF shared two significantly correlated modules with OSF. When the regulatory value groups’ overall correlations were evaluated through unsupervised hierarchical clustering, we found that RPF’s correlations to the module eigengenes were the most similar to IARC hazard classifications, followed by IRIS, then OSF, IUR, RfD, and RfC (Fig. 4). Through this analysis, we observed that select module eigengenes were significantly correlated with RPF, yet different module eigengenes were significantly correlated with different risk value types. MEthirteen was highly significantly correlated (p=3E-4) with OSF, while IUR was significantly linked to MEten. These uniquely correlated modules suggest that a combined approach of multiple regulatory values can be used to identify potential biomarkers that may have greater applications in chemical safety and hazard assessment.

Figure 4. Module-trait relationship correlation matrix linking module eigengenes with risk values.

WGCNA identification of gene modules significantly correlated to risk values and cancer categorizations (RfC, RfD, inhalation unit risk, oral slope factor, RPF, and IRIS and IARC carcinogen classes). Correlation matrix shows that several module eigengenes (MEs) were highly correlated with PAH cancer risk values (RPF), as well as correlated in common between RPF, OSF, and IARC classes. Unsupervised hierarchical clustering was performed on the correlation values and the dendrogram identifies RPF and IARC as having the most similar module correlation patterns. Degree of correlation is colored, with red as positive and blue as negative, and p-values in parentheses.

Table 2.

Gene modules significantly (p< 0.05) associated with cancer risk values.

Module	Correlation (max)	p-value (min)	Risk value(s)*
Two	0.82	0.02	RPF
Five	0.85	0.02	OSF, RPF
Fourteen	−0.84	0.02	OSF, RPF, IARC
Seventeen	−0.77	0.05	RPF

^*Note that IARC is a cancer classification value (as opposed to risk)

DISCUSSION

As the toxicology community shifts towards in vitro testing and computational biology, novel approaches in transcriptional data analyses must be developed and refined to fill unmet needs in chemical hazard assessment⁴⁶. The ultimate goal of toxicity testing is to generate and understand data to protect the public and environment from harmful chemicals. However, traditional in vivo toxicity testing is limited in its throughput and capacity to provide mechanistic information linking apical endpoints to the underlying processes involved in chemical toxicity⁴⁷. Current “-omics” level approaches evaluating PAH carcinogenicity have focused on identifying chemical-specific mechanisms and been largely successful contributing towards development of predictive approaches of chemical MOA^24,48–51. However, past studies have not focused on identifying broader patterns in affected biomolecules across different parameters of PAHs, such as cancer risk. Here, we evaluated the transcriptional signatures associated with a range of carcinogenic PAHs tested in a 3D in vitro airway epithelium and utilized a WGCNA approach to identify gene modules significantly correlated to increasing cancer risk calculated through RPF.

Overall, we identified groups of genes with similar expression patterns that were significantly correlated to PAH cancer RPF estimates. Through our co-expression analysis, we identified multiple co-modulated gene clusters significantly correlated with RPF. Genes within the two most significantly correlated modules were enriched for processes of cell cycle regulation and cytoskeleton, suggesting the importance of these processes for cancer risk associated with RPF. We also evaluated additional risk values both specific and non-specific to inhalation. Our expanded co-expression analysis identified shared gene modules among RPF, IARC hazard classes, and cancer slope factor indicating similarity in these values for PAH cancer risk.

Carcinogenic PAH exposure is associated with alterations in cell cycle S phase gene regulation

This study identified cell cycle S phase genes as an important and highly enriched pathway associated with a module of high interest associated with cancer RPF values, MEturquoise. MEturquoise genes were significantly positively correlated with increased RPF estimates. Our results are consistent with other findings that PAH exposure is linked to alterations in cell cycle processes. Previous research has identified dysregulation of cell proliferation as a potential mechanism of BAP and BAN activity⁵². The p53 signaling pathway is also activated in response to stress, which includes DNA damage⁵³. Several genes identified in MEturquoise (Fig 1B) (i.e. TP53BP2 and MDM2) are involved in p53 signaling^54–56. Cyclin-dependent kinase (CDK1) was also identified as a gene of interest in MEturquoise, and normally functions in cell cycle progression via S phase initiation⁵⁷. Dysregulation of these signaling pathways may be important in early cellular responses to PAH exposure. A well-understood mechanism of BAP carcinogenesis is its genotoxic activity forming diolepoxide DNA adducts^58–61. Our findings support several reports that exposure to BAP can cause S phase arrest through DNA damage signaling^52,62,63.

Bottleneck genes identified in the network (CDK1, MCM2, PCNA) have roles central to controlling cell cycle progression, including M-phase transition, and responding to DNA damage (Fig. 3A). However, their involvement with environmental carcinogens have not been clearly established. Our network analysis identified several highly connected bottleneck genes central to the network, which include the MCM2, a regulator of DNA synthesis, and CDK1, which controls progression into S phase and M-phase transition of the cell cycle and induces senescence in response to DNA damage^64–67. Malfunction or dysregulation of these proteins have been linked to the development of cancers in multiple tissues, including lung and breast^67–71. While BAP exposure has previously been linked to alterations in CDK1 and MDM2 activity, the findings of this network analysis support S phase gene regulation as a major cellular network perturbed by carcinogenic PAH exposure^62,63,72.

Carcinogenic PAH exposure dysregulates actin cytoskeleton gene networks

In our analysis of prioritized MEred module genes significantly correlated to RPF, we identified the cytoskeleton actin filament network as a top negatively correlated gene network to PAH cancer risk. Several genes (ACTG1, ACTB, SRC) were also identified as bottleneck genes with high betweenness centrality measures (Fig. 3B). These findings suggest that genes integral to actin cytoskeleton structure, maintenance, and regulation, are negatively associated with increased PAH cancer risk. Similar mechanistic studies also found that short-term exposure to PAHs or PAH-containing mixtures result in the down-regulation of several cell adhesion, motility, and cytoskeletal-involved processes, including epithelial-mesenchymal transition^32,73–76. These processes can be indicators of cancer promotion and progression, and can co-regulate in complex signaling networks during cancer progression. Actin filament reorganization is directly responsible for driving cancer cell motility, morphological changes, and alterations in extracellular matrix and adhesion^77–79. Additionally, actin-based adhesion protein complexes are integral in linking cells via cell-cell adhesion and to the extracellular matrices. Taken together, these findings support the role of actin cytoskeletal networks as early biomarkers and indicators of cellular changes affected by carcinogenic PAH exposure.

Comparisons of regulatory values and associated gene modules

PAH cancer risk assessment has traditionally relied on the use of RPF estimates available only for a subset of PAHs¹³. In this study, we adapted WGCNA in a novel manner to identify genes similarly associated to RPF and other risk values and hazard classifications. Notably, we identified overlap in significant gene module membership between RPF and at least one other risk value in two modules. We also found that RPF correlation patterns were the most similar to that of IARC cancer classes. These results suggest that the RPF approach currently has more similar hazard designations towards chemicals, compared to IARC classifications, which classifies chemicals as carcinogens based on weight of evidence^80–82. Inversely, many chemicals, including heterocyclic PAHs, present in PAH mixtures evaluated did not have RPF estimates nor IARC classifications; yet had OSF and IUR values. While direct comparisons between differing regulatory and risk values is not feasible, this analysis focused on identifying shared gene clusters and exploring correlations across risk values. There were also several challenges and limitations involved with using RPF and risk values. The RPF estimates available to calculate BAPeqs were derived from murine and rodent models from studies where animals were exposed orally and through intraperitoneal injection, which do not represent typical routes of PAH exposures for human. Most importantly, inhalation was not considered in deriving RPF estimates. The IURs we included from CTV completed our trait table, however a majority of the chemical values sourced for chemical mixtures were predicted from CTV, rather than experimentally derived. CTV was a useful tool in supplementing data-poor chemicals with predicted health regulatory values. However, this data was frequently low confidence, from CTV’s high percent of applicability domain. Additionally, the second tier of our analysis did not allow taking chemical dose into consideration; only one dataset per chemical was able to be associated to database-derived regulatory and hazard values.

Advantages of identifying co-modulated gene sets across cancer risk values

Data generation from chemical transcriptional profiling has yielded a wealth of chemical-gene signatures; however researchers lack adequate measures to understand broader patterns across chemical groups. The results from this study demonstrate a novel and effective method to identify gene associations across multiple chemicals and parameters. There were several advantages to using RPF as the primary chemical risk value for identification of co-modulated gene associations. RPF estimates are derived through a weight of evidence approach, and so RPF values are substantiated by in vivo tumorigenic evidence¹³. Typically, multiple chronic rodent cancer bioassays are evaluated towards the development of an RPF estimate. We identified cellular networks positively and negative associated with cancer risk in the 3D HBEC model, and our findings supported several reports in the literature, indicating that short-term gene signatures can be particularly useful in identifying genes linked to cancer. A major benefit of a co-expression approach is that subtle gene changes and biological interactions may be captured²⁸, allowing identification of similarly regulated gene networks. Previous research using WGCNA has found that gene modules identified in drug-induced hepatocyte transcriptional changes are correlated to adverse outcomes and toxicity⁸³; however some of the top ranked gene modules identified had little biological annotation, indicating a unique capability of this approach⁸⁴. Investigation of these poorly annotated genes and lesser understood signaling networks may prove useful in developing novel biomarkers to identify potential drivers of chemical toxicity. In this study, we identified 9 miRNAs and over 300 lncRNAs as significantly correlated with PAH cancer risk. Thus there is high potential for identifying novel or understudied biomarkers related to PAH-specific cancer risk through applications of WGCNA. In addition, expansion of this analysis to include more PAHs of varying carcinogenic potential could help strengthen gene associations and identify broader patterns across carcinogenic potency.

CONCLUSIONS

This study examined relationships between co-modulated gene sets and chemical risk values relevant to human cancer risk, with the goal of identifying key pathways, and biomarkers linked to cancer risk of PAH and PAH mixtures. For the first time, PAH-responsive genes were linked to RPF through a quantitative systems biology approach. We prioritized positively and negatively associated modules for further study, and these genes significantly enriched distinct biological functions including cell cycle, cytoskeleton, and inflammation. This study also identified module eigengenes significantly associated with additional regulatory values and hazard classifications including IARC classifications, oral cancer slope factors, and inhalation unit risks. Some regulatory values (IARC, IRIS, OSF) shared greater numbers of commonly correlated modules with RPF than other regulatory values (IUR, RfC, RfD), suggesting potential value in using them when RPF is unavailable. This study shows the potential for using transcriptional signatures in response to short-term chemical exposures to identify biomarkers and associations between regulatory values and biological activity. Such an approach could be used to relate organotypic, human in vitro-derived toxicogenomics information towards regulatory applications in the near future.

Supplementary Material

Supplemental File S1

Supplemental File S1. Chemical composition of the artificial air mixture.

Supplemental File S2

Supplemental File S2. Summary of risk regulatory values (RPF, RfC, RfD, OSF, IUR, IARC, IRIS).

Supplemental File S3

Supplemental File S3. List of genes in the MEturquoise module.

Supplemental File S4

Supplemental File S4. List of genes in the MEred module.

Abbreviations

ACTB

actin beta

ACTG1

actin gamma 1

AM

simulated air mixture

BAN

benz[a]anthracene

BAP

benzo[a]pyrene

BAPeq

benzo[a]pyrene equivalents

BRCA1

breast cancer type 1 susceptibility gene

BRCA2

breast cancer type 2 susceptibility gene

CDK1

cyclin-dependent kinase 1

CLDN8

claudin 8

CTE

coal tar extract

CTTN

cortactin

CTV

Conditional Toxicity Value Predictor

DBC

dibenzo[def,p]chrysene

DNAL1

dynein axonemal light chain 1

DYNC2LI1

dynein cytoplasmic 2 light intermediate chain

FMN1

formin 1

HBEC

human bronchial epithelial cells

HMOX1

heme oxygenase 1

IUR

inhalation unit risk

MCM2

minichromosome maintenance complex component 2

ME

module eigengene

JUNB

JunB proto-oncogene, AP-1 transcription factor subunit

NR2F2

nuclear receptor subfamily 2 group F member 2

OSF

oral slope factor

PAH

polycyclic aromatic hydrocarbon

PBS

phosphate buffered saline

PCNA

proliferating cell nuclear antigen

PHE

phenanthrene

RfC

reference concentration

RfD

reference dose

RPF

relative potency factor

SRC

SRC proto-oncogene, non-receptor tyrosine kinase

U.S. EPA

United States Environmental Protection Agency

TEER

transepithelial electrical resistance

TP53AIP1

tumor protein p53 regulated apoptosis inducing protein 1

WGCNA

weighted gene co-expression analysis