Application of Text Mining in Risk Assessment of Chemical Mixtures: A Case Study of Polycyclic Aromatic Hydrocarbons (PAHs)

Hazard identification is the first step in cancer risk assessment and requires a thorough review of the literature for generating information on the carcinogenic potential of chemicals. An efficient and reliable literature review is crucial to properly assess impact on human health (Goodman and Lynch 2017). However, owing to the exponential growth and complexity of the scientific literature, manual evaluation has become extremely challenging and time consuming (Korhonen et al. 2012). While considering human exposures to multitude chemical carcinogens and the pace of current risk assessment practices, novel approaches need to be introduced to evaluate cancer risks associated with these chemicals. Therefore, we have developed text-mining (TM) tools for literature review and to identify mechanistic information for cancer risk assessment of chemicals (Korhonen et al. 2012; Silins et al. 2012). Although prior research has produced some tools to support practical tasks such as literature curation and development of semantic databases in biomedicine, none of these are focused on chemical cancer risk assessment (Harmston et al. 2010; Rebholz-Schuhmann et al. 2012; Zhu et al. 2013).

Based on TM technology, we developed an open access Cancer Risk Assessment using Biomedical literature (CRAB) tool that automatically organizes and classifies literature according to carcinogenic modes of action (MOAs) (Ali et al. 2016; Korhonen et al. 2009, 2012; Silins et al. 2012). Our evaluations have shown that the automatic classification results generated by these tools are very accurate, often between 95% and 99% (Korhonen et al. 2012). In addition, the CRAB tool can be used to reveal similarities and differences between chemicals by, for example, identifying similar MOAs shared by chemicals in mixtures. We have employed the tool in practice, including studies of MOA analysis for groups of chemicals. For example, in a study of polychlorinated biphenyls (PCBs), we analyzed literature profiles of dioxin-like- and non-dioxin–like PCBs and compared them with those of their indicator PCBs (Ali et al. 2016). A comparison of carcinogenic MOAs identified by the CRAB tool with a manual evaluation of PCBs found similar evidence on key characteristics (Smith et al. 2016), associated with carcinogenicity of these compounds. We also investigated carcinogenic MOAs of pesticides in fruits and identified that CRAB-generated MOA profiles of many pesticides were similar, containing evidence for both genotoxic and nongenotoxic MOAs (Silins et al. 2014). Together, these results suggest that the CRAB tool can be of great use in effectively identifying groups of chemicals with similar MOAs and to evaluate the relevancy of a reference compound for a group of chemicals.

In addition to the CRAB tool, we have developed a Cancer Hallmarks Analytics Tool (CHAT) (Baker et al. 2017), primarily intended to assist cancer researchers. CHAT provides detailed analyses of literature by automatically organizing and classifying literature on the basis of the evidence on cancer hallmarks and associated processes. CHAT consists of 37 distinct categories, which specify the biological processes underlying each of the original eight cancer hallmarks and two enabling characteristics (Hanahan and Weinberg 2000, 2011). CHAT analyses provide mechanistic information on the biological processes involved in cancer development.

Exposures to chemicals occur in mixtures, and mixture risk assessment is challenging (Gwinn et al. 2017). One such group of chemicals comprises the polycyclic aromatic hydrocarbons (PAHs), a family of more than 1,500 compounds with diverse structural features, that is, ring configurations and substitutions (NTP 2012). Based on the evidence on their carcinogenicity, the International Agency for Research on Cancer (IARC) has classified the PAH benzo[a]pyrene (B[a]P) as a human carcinogen and several other PAHs as possible or probable human carcinogens (IARC Working Group on the Evaluation of Carcinogenic Risks to Humans 2010). Although the evaluation of carcinogenicity of PAHs has been made based on their mutagenic and tumor-initiating activities, data are limited on their tumor-promoting activity (Asada et al. 2005; Sakai et al. 2010). Recently it was suggested that some of the PAHs with relatively low carcinogenic potential exhibit strong tumor-promoting activity (Misaki et al. 2016). Importantly, many complex mixtures containing PAHs have been classified as human carcinogens—including coal tar, diesel exhaust, and coke oven emission—and are of serious health concern, not only for the general population, but also for those in occupational settings (IARC 2016). Although sufficient data are available to support the carcinogenicity of some PAH mixtures in humans, there is a general lack of epidemiological data for individual PAHs, for example, B[a]P (Baan et al. 2009).

Cancer risk assessment of PAHs is principally based on one of two approaches. One is based on B[a]P as a surrogate marker for all PAHs and thus solely considers the environmental levels of B[a]P to assess risks. This approach is used for assessing airborne PAHs within the European Union (WHO 2000). The other approach considers the contribution from other PAHs by applying potency factors that describe the relative carcinogenic potency to B[a]P (RPFs) and is used by, for example, the U.S. Environmental Protection Agency (EPA) for assessing environmental PAH pollution (U.S. EPA 1993). Although the latter approach includes multiple PAHs, only a limited number of PAHs have been assigned a RPF, and an extended list is currently under evaluation (U.S. EPA 2010a). A potentially serious limitation of both approaches is that they do not allow for taking interaction effects into account. A number of studies have shown that PAH mixtures can induce nonadditive effects, resulting in antagonistic or synergistic genotoxic or carcinogenic effects (Jarvis et al. 2014). Therefore, the development of new methodology that will optimize the possibility of taking advantage of published scientific databases is highly motivated.

In this article, we present a case study of 22 U.S. EPA and EU priority PAHs and complex PAH mixtures that are of concern for the general population and in occupational settings. We used the CRAB3 tool (http://crab3.lionproject.net/)—the latest iteration of the CRAB tool—and CHAT (http://chat.lionproject.net/) to analyze PAH literature and to evaluate the usefulness of these tools for hazard identification. CHAT analyses complement those of the CRAB3 tool, and to the best of our knowledge, for the first time they were used together.

By using a combination of our TM tools CRAB3 and CHAT, we present a case study of PAHs based on the PubMed literature (March 2019) and show that our tools can identify similarities and differences in carcinogenic MOAs. First, we analyzed the literature concerning the 22 priority PAHs and tested whether the information generated by the CRAB3 tool can be useful to classify PAHs into subgroups. To evaluate the classification, we employed RPF values assigned to these PAHs and found a correlation between CRAB3- and CHAT-generated profiles and RPFs and that the CRAB3 tool could group PAHs according to their potency. We also found that differences in the carcinogenic potential of PAHs and PAH derivatives, namely, methylated- or nitrated-PAHs, and of other chemicals, for example, crystalline silica, can be identified. An unexpected finding was that the tools revealed major mechanistic differences in the characterization of naturally occurring PAH mixtures. These data show that CRAB3 and CHAT analyses gave similar answers to our questions and had the capacity to evaluate a large amount of literature concerning the complex group of PAHs. Moreover, comparison of CRAB3 analyses with the IARC evaluation in identifying key characteristics on carcinogens showed similar evidence and correlated with traditionally conducted assessments.

The classification of literature by each of these tools is based on different taxonomies and TM techniques. Nevertheless, the output data from CRAB3 and CHAT complemented each other in this study. CRAB3 classification of literature is based on the evidence for carcinogenic MOAs, and the tool was developed primarily for cancer risk assessment purposes (Korhonen et al. 2012). CHAT analysis was developed primarily for cancer researchers and for classifying cancer literature according to cancer hallmarks and associated processes (Baker et al. 2017). CHAT analysis has the capacity to identify more detailed end points not covered by the CRAB taxonomy. Because CHAT analysis is built on concepts employed in basic research, using CHAT in combination with the CRAB3 tool might be particularly helpful for basic researchers interested in risk assessment issues.

Our data suggest that both TM tools can identify literature profiles that reflect potency and can, for example, support grouping chemicals with similar relative potency. Comparing MOA profiles for structurally related single chemicals within the PAH group and with strongly differing potency, such as chrysene and 5-methylchrysene, complemented the analysis based on RPF values by showing that both tools could differentiate literature on structurally similar chemicals but with widely differing potency. Even analysis of B[a]P and DB[al]P, which are structurally related but have largely differing sizes of literature bases, revealed expected results with selectively more data on DNA damage-related effects for the latter. These results were based on PAHs, for which RPFs previously have been developed, but suggest that these tools also could be used for potency grouping of chemicals that lack potency factors.

Cancer risk assessment of a complex group of chemicals, such as PAHs, is challenging because exposure to PAHs occurs mainly in mixtures. Diesel exhaust, coal tar, and coke oven emission are among the most well-studied environmental PAH mixtures, and thus large volumes of literature are available for TM-based analyses. These three natural mixtures of PAHs exhibit quite different literature profiles. For example, the MOA profile of coke oven emission shows a higher than expected co-occurrence with strand breaks and chromosomal changes in the text as compared with B[a]P. Similarly, data on diesel exhaust show a higher than expected co-occurrence with oxidative stress, inflammation, and immunosuppression in the text as compared with B[a]P. Although all these mixtures are classified as human carcinogens (IARC 2016; IARC Working Group on the Evaluation of Carcinogenic Risks to Humans 2010), it is the levels of only B[a]P alone that are used for controlling these emissions and assuring air quality in relation to cancer risk for the general population within the European Union. In addition, there are diesel emission standards managed by, for example, the European Union and the U.S. EPA for particulate matter, nitrogen oxides, and carbon monoxide to limit noncancer health effects (Khalek et al. 2015). The progression of these emission standards has resulted in the development of new-technology diesel engines with reduced emission levels, including PAHs (Khalek et al. 2015). In addition, in 2019, the European Union amended their directive on the protection of workers from exposure to carcinogens to also include diesel exhaust, with a limit value based on elemental carbon (EU 2019). Our findings suggest that we may need to reevaluate the use of B[a]P as a marker for the carcinogenic risk of PAH mixtures and warrant development of new strategies, as has been discussed recently (Andersson and Achten 2015; Jarvis et al. 2014).

In addition, the possibility that PAHs and crystalline silica may interact was suggested by CRAB3 and CHAT analysis, as have been proposed in previous studies (Downward et al. 2014; Vermeulen et al. 2011). The toxicological profiles created by the CRAB3 tool and CHAT respectively for crystalline silica were similar, once again documenting that these two tools can be used for confirming robustness for carcinogens other than PAHs. By employing the CHAT tool and comparing the cancer hallmarks profiles of B[a]P and crystalline silica, we hypothesize that in a co-exposure setting, these chemicals could cause carcinogenic synergies. They may act in accordance with experimental studies employing initiation–promotion protocols (Cohen and Ellwein 1990; Solano et al. 2016). They may also act through different and, presumably, complementing cancer hallmarks, as has been evaluated in the Halifax project (Goodson et al. 2015).

Both CRAB3 and CHAT analysis have many advantages when evaluating large volumes of textual data in abstracts; however, they have some limitations. One of the major limitations of these tools is that they do not have the capacity to distinguish between positive and negative findings or, for example, between down- vs. up-regulation of inflammatory effects. The preferences in the types of research performed, that is, the extent at which a chemical has been studied, may also have an influence on the output. However, it should be noted that the output from these tools is based on available literature and the lack of literature does not necessarily mean the lack of association of a particular MOA or hallmark with the chemical. Furthermore, methodological differences in the study design, that is, dose, route of exposure, and so on, need to be examined manually. Future development of these tools includes features such as identification of positive vs. negative data, consideration of study design, and bibliometric indicators. We plan to upgrade the tools such that chemicals identified by the CRAB3 tool will be assigned a Chemical Abstracts Service number. For classification of compounds with limited amount of human or animal toxicity data, we also plan to integrate structure–activity relationship modeling into our TM tools, as suggested in our previous study (Papamokos and Silins 2016).

In summary, we show that our TM tools CRAB3 and CHAT can be used for literature review in cancer risk assessment of groups of chemicals. Using 22 priority PAHs and complex PAH mixtures, we observed that the similarities and differences between the PAHs identified by our tools can be useful for grouping these chemicals based on their RPF values. Moreover, these tools can facilitate the identification of possible interactions within a group of chemicals such as PAHs and between complex mixtures. A surprising finding was that naturally occurring mixtures that include PAHs exhibited different literature profiles. This case study using PAHs suggests that our tools can be useful not only for risk assessors handling groups of similar chemicals and mixtures but possibly also for cancer researchers addressing cancer risk assessment issues, including gaps in knowledge.

We found that these tools rapidly give literature overviews that complement and enrich the outputs. Used together, they can effectively assist manual reading of the articles. Not only risk assessors, but researchers can also benefit when addressing risk assessment issues. By using these tools, researchers can get a better understanding of how cancer risk assessment might gain from using new and cutting-edge science.

This research was supported by the Medical Research Council-UK and Swedish Research Council for Health, Working Life and Welfare.

The authors declare they have no actual or potential competing financial interests.