The DIKW of transcriptomics in ecotoxicology: extracting information, knowledge, and wisdom from big data

In the early 2000s, transcriptomics was emerging as a new science with seemingly limitless potential. Originally coined by Charles Auffray (McGettigan, 2013), the term transcriptomics refers to the measurement of levels of all expressed genes across the genome (Figure 1). To researchers in our field in the early aughts, transcriptomics promised to uncover the global molecular response to environmental contaminants. We imagined replacing the limited view provided by individual biomarkers such as ethoxyresorufin-O-deethylase (EROD) or vitellogenin, with a detailed picture of all the changes occurring in the cell and the pathways impacted. The idea that each individual chemical or mixture would have a unique signature that could be linked to organismal-level outcomes was prominent. Over 20 years later, many of us who are using transcriptomics in our research are still baffled by how to analyze these big data sets, and how to apply the results to salient questions in the field of ecotoxicology. The paradox is that we have exponentially more information, but a limited capacity to interpret what it all means. The objective of this Focus Article is to explore the promise of transcriptomics for ecotoxicology and to break down current realities in the field. We hope to convey a realistic view of both the challenges and opportunities associated with this scientific approach. Our article is geared toward researchers, government regulators, and students who are interested in transcriptomics but have limited knowledge or experience in this area. We have organized our article around the Data, Information, Knowledge, Wisdom (DIKW) framework (Figure 2), and an illustrative case study (Textbox 1). The DIKW framework comes from the field of information science, and can be used to illustrate the process of moving from raw data to valuable insights. Steps in a typical transcriptomics experiment. Methodological details from an illustrative case study are provided. The case study involved exposing double-crested cormorant (Nannopterum auritum) embryos to ethinylestradiol via egg injection. Detailed methods and results are presented elsewhere (Jeon et al., 2023). One transcriptomics experiment can produce hundreds of gigabytes (GBs) of raw data. Next generation sequencing allows us to generate transcriptomics data for any species. With the price of sequencing dropping precipitously, transcriptomics experiments are becoming more affordable. The current leading technology for generating transcriptomics data is RNA sequencing (RNA-Seq), with a cost in the range of $100 USD/sample. New multiplex approaches promise to drive the price/sample down further still. An experiment like our illustrative case study (Textbox 1) can generate GBs of data for less than $2000 USD. An RNA-Seq experiment can easily be performed in a basic molecular biology laboratory by students or technicians with minimal experience. One of the great advantages of RNA-Seq (as opposed to, for example, microarray technologies) is that it is species agnostic; the experimental steps to generate data are the same regardless of the species used. The raw data from an RNA-Seq experiment come in the form of "sequencing reads," which are fragments of nucleotide sequences that are approximately 100 letters (100 base pairs [bp]) long. The length of an expressed gene is typically >1000 bp, so overlapping reads need to be matched to the gene they correspond to before they can be counted. With the cost of sequencing decreasing, the ease of generating transcriptomics data is increasing. The bigger challenge comes when you need to figure out how to extract information from the 100s of GB of data that you receive from the sequencing laboratory. Extracting information from sequencing data is not trivial but it is getting easier with new tools. The approach differs for species with or without established reference genomes. Statistical uncertainty means that there is no single "right" analysis: different bioinformatics pipelines can produce different results. To extract information from sequencing reads, the overlapping fragments need to be mapped to known genes and then counted. This is fairly straightforward for species with publicly available reference genomes. For nonmodel species whose genomes are not yet sequenced (e.g., the double-crested cormorant Nannopterum auritum [Textbox 1]), the entire constellation of reads received from the sequencing laboratory can be assembled to create a de novo transcriptome, which is used in place of a reference genome. De novo transcriptome assembly is not trivial. In addition to piecing together all the sequencing reads like a puzzle, the researcher needs to identify and annotate (i.e., assign a function to) the genes within the assembled transcriptome. A new online tool allows researchers to skip the computationally intensive steps associated with de novo assembly (Ewald et al., 2023; Liu et al., 2023). The Seq2Fun algorithm (accessible via ExpressAnalyst, 2024) aligns raw sequencing data from any species to functional groups of genes (orthologs) compiled from the genomes of over 600 species. This produces a table of expression values for approximately 12,000 to 16,000 gene groups, the majority of which have known function. Compared with de novo assembly and annotation, some information is lost using this approach, but there are also many advantages. For one, it is often easier to interpret dysregulation of a smaller number of known genes rather than hundreds of thousands of de novo assembled transcripts, many of which are unannotated and identified with low confidence. Mapping to a common ortholog database with species-agnostic identifiers encourages research on species with unannotated genomes and helps facilitate cross-species comparisons. The seq2Fun approach is a practical starting point for researchers who are studying nonmodel organisms and who are newcomers to the world of bioinformatics. The output of Seq2Fun or other bioinformatics approaches will tell you which genes were expressed in your sample, and their relative abundance. Most researchers will then want to compare these "counts" of expressed genes across experimental conditions. What emerges from this analysis is a list of differentially expressed genes (DEGs), that is, genes that are expressed at different levels in control and treated samples (Figure 1). This information can confirm that previously identified modes of action of a chemical are active, or suggest new hypotheses. Different visualization methods can be used to holistically explore the data, as we illustrate using results from our case study. Common approaches include principal components analysis (PCA) plots (which illustrate the relative similarity between control and treated samples and help to identify potential outliers), Venn diagrams (which compare the number of up- or downregulated genes under different experimental conditions), heat maps (which highlight common patterns of gene expression between samples), and volcano plots (which highlight the p value and directionality of differences in gene expression). Some of these visualization methods are illustrated in Figures 3 and 4. One challenge with differential expression analysis for transcriptomics experiments is that the uncertainty inherent in any statistical approach is magnified when one is working with tens of thousands of expressed genes, high levels of biological variability, and small sample sizes. Because toxicogenomics studies are expensive and often have a secondary goal of reducing animal use, they tend to use a lower number of replicates than traditional toxicity studies. For example, the Organisation for Economic Co-operation and Development (OECD) test methods measuring endpoints related to growth, reproduction, and survival in rats typically require sample sizes of at least 10 animals/experimental group (e.g., OECD, 1998, 2008), whereas sample sizes of three to five are typical for toxicogenomics experiments. Differential expression analysis employs thousands of t tests to differentiate between levels of expression of every single gene in a treated sample compared with a control sample. The t tests make a variety of assumptions that a sample size of only three to five cannot reliably provide across the thousands of genes in the data set. Bioinformaticians use a variety of approaches to mitigate this problem, and different algorithms can return different numbers of DEGs for the same data set even though each algorithm is well established and considered to follow "best practices." The end result is that differential expression analysis often has low statistical power, and there can be differences in the final list of DEGs produced by two different scientists analyzing the same raw data. For example, the number of up- and downregulated genes identified in our illustrative case study varied depending on which statistical approach (Limma vs. EdgeR) and fold change cutoffs (Log2FC 0 vs. Log2FC 1) were used (Figure 4). The availability of varied and evolving approaches for analysis of transcriptomics data can create the impression that the data analysis step is never "done" as well as a lack of confidence if different methods yield different results. In our opinion, this lack of confidence in the "final" list of DEGs can, in some ways, be well justified. Researchers must shift their understanding to encompass the uncertainty implicit in an analysis of big data and look for large overall patterns in the transcriptomics data rather than one definitive answer. It can be helpful to think of differential expression analysis as a tool for prioritizing features for further investigation instead of producing a final list of set-in-stone results. Alternatively, researchers who are interested in a more stringent statistical analysis can use multiple statistical packages to analyze their data and publish a consensus list of DEGs. Regardless of the bioinformatics approach used, unlocking the full potential of transcriptomics and deriving new knowledge often requires further analysis. Generating new knowledge from transcriptomics data requires complementary information on the genome, molecular pathways, and organismal level response. New approaches such as transcriptomic dose–response analysis (TDRA) hold promise for making comparisons between transcriptomic and organismal level responses. Extracting knowledge from information requires biological understanding of the model system, the experimental design, and the underlying scientific question. An initial scan through a list of DEGs will allow you to identify groups of genes that are dysregulated in your experiment. Clustering of gene expression profiles can organize perturbed genes into groups that have similar patterns across samples. Prior knowledge about the function of these genes can lead to inferences about the mechanism of action of the chemical. For example, in our illustrative case study, dysregulation of classical biomarkers of estrogen exposure, vitellogenin and apovitellenin 1, is apparent, as would be expected for individuals exposed to ethinylestradiol (Figure 3B). This type of gene-by-gene approach can build confidence in the results, but tends to focus our attention on genes that we are already familiar with and away from potential new insights. Unlocking the full potential of transcriptomics data requires a broader understanding of the global transcriptional response and the molecular pathways that may be involved. Pathway analysis, a common next step in the RNA-Seq bioinformatics pipeline, attempts to do this. As with differential expression analysis, a variety of ever-evolving statistical packages can be used for pathway analysis, with the potential for disparate results with different packages. Two currently popular approaches are gene set enrichment analysis (GSEA) and overrepresentation analysis (ORA). For the nonmodel species that are often used in ecotoxicology, pathway analysis is particularly challenging. The databases that are currently used to group genes into pathways (e.g., Kyoto Encyclopedia of Genes and Genomes, and Gene Ontology) are largely based on biomedical models (e.g., humans, rodents) and are therefore often missing the appropriate context for analyzing results from nonmammalian species. Our case study with double-crested cormorant illustrates this problem when it highlights pathways related to Parkinson's, Alzheimer's, and Huntington's diseases that do not seem to be relevant to our experimental model (Figure 5). However, identifying pathways associated with human disease in nonhuman organisms should not necessarily be interpreted as nonsensical. For example, Parkinson's and Huntington's disease pathways include many genes involved in mitochondrial respiration (e.g., nicotinamide adenine dinucleotide:ubiquinone oxidoreductase and cytochrome C oxidase isoforms), whose dysregulation may be interpreted differently in birds. In other cases, key pathways that are very relevant to our animal model may be missing. The classic biomarker of estrogen exposure, vitellogenin, is not expressed in mammals. This well-conserved response in oviparous animals will be missed if the statistical analysis relies on databases built from rodent and human genomes. More research is needed in this area to develop databases that are tailored to ecotoxicological models. Another approach to generating knowledge from transcriptomics data is the use of TDRA. This approach is rapidly gaining attention because it provides a way to compare overall transcriptomic responses with apical outcomes of regulatory concern (Johnson et al., 2022). Specifically, TDRA fits individual curves to each of the expressed genes in a data set and calculates a gene-level benchmark dose for significant curve fits. A transcriptomic point of departure is then calculated from a distribution of the gene-level benchmark doses and can be used to provide estimates of chronic toxicity at an organismal level (Pagé-Larivière et al, 2019). Transcriptomic dose–response analysis has the advantage of being able to capture the entire transcriptomic response without requiring an understanding of the function of all of the genes, or what pathways they belong to, which is an especially attractive advantage for research involving nonmodel species. These new approaches are promising, but building transcriptomic knowledge into the wisdom needed to make sound regulatory decisions remains an elusive goal. Transcriptomics data can contribute to the assembly of knowledge that is required to make informed decisions about chemicals in the environment. Global efforts to reduce animal use are stimulating interest in using transcriptomics data in the context of new approach methods (NAMs). More research needs to be done to develop, standardize, and validate transcriptomics methods before they can be incorporated into regulatory frameworks. Adopter-centric approaches prioritizing the user experience, fit for purpose, and implementation will be critical to the success of these efforts. In the context of our field, wisdom could be defined as an assembly of knowledge with the potential to make a real-world impact on chemical contamination of the environment. Wisdom requires integration of knowledge from different sources and fields of study. Insights derived from transcriptomics studies are only one piece of the puzzle. In a regulatory context, wisdom is critical for sound decision-making. Regulators may consider transcriptomics data in their assessments, but decisions have traditionally been based on organismal-level endpoints, that is, effects on growth, reproduction, or survival derived from standardized tests using animal models. Although these types of data provide a direct link to outcomes of regulatory concern, they may not detect more subtle effects that may nonetheless be indicative of long-term harm, or help to elucidate mechanisms of action. Additionally, the chronic animal tests that are currently critical to regulatory decision-making are widely understood to be inefficient, prohibitively expensive, and ethically concerning (Pain et al., 2020). Given the shortcomings of the status quo, organizations such as the US Environmental Protection Agency (USEPA) and Health Canada are now promoting the use of NAMs that reduce, refine, or replace animal use. This could include in vitro or embryo-based assays that integrate high-throughput transcriptomics and can be used to screen and prioritize large numbers of chemicals for regulatory purposes or commercial use. However, integration of knowledge derived from such transcriptomics studies into regulatory frameworks is not easy. Regulators are interested in using new approaches, but they require standardization, validation, and evidence that benchmark doses derived from transcriptomic approaches are not overprotective. Social sciences research reveals that the design of new toxicogenomic methods must prioritize simplicity and compatibility with regulatory routines and workflows (i.e., an adopter-centric view) versus focusing on novel functionality (i.e., an innovation-centric view; Pain et al., 2020). Toward this end, the introduction of the USEPA Transcriptomic Assessment Product in 2024 as an approach to yield transcriptomic-based reference values represents an important step forward. In a research context, wisdom stimulates further scientific exploration and innovation. For example, knowledge derived from transcriptomics studies may be used to design new materials that are less hazardous, meeting the green chemistry principle of "benign-by-design." Knowledge may also be used to construct new test methods built on gene expression that pay careful attention not only to scientific and technical validity but also to user experience, fit for purpose, and pathways to implementation. This research is critically important given a regulatory desire for NAMs that reduce animal use in favor of in vitro or embryo-based approaches. For over 20 years, ecotoxicogenomics research has been met with both excitement and skepticism about our capacity to link transcriptional changes to ecologically meaningful outcomes. As our ability to generate and harness the power of big data improves, optimism grows. Our reflection on the DIKW of transcriptomics leaves us with several concluding thoughts about the path forward. As is common for new technologies, transcriptomics has sometimes been perceived to over-promise and under-deliver. In this regard, interpretation and communication of transcriptomics data may benefit from an honest reckoning about where the analysis fits within the DIKW framework. For some experiments, particularly those using nonmodel species, it may be appropriate to stop at the information stage and form hypotheses based on lists of DEGs rather than struggle to perform meaningful pathway analysis. This is an important first step that will facilitate further research and innovation. Generating knowledge about pathways affected can currently be more easily accomplished in model species with well-annotated genomes. Finally, new approaches driven by regulatory adopters such as TDRA may bring us closer to the desired goal of harnessing data and information from transcriptomics studies to support knowledge-based decisions regarding the safety of environmental chemicals. Regardless of where you are in the DIKW hierarchy, it is critical to have a realistic understanding of the strengths and limitations of transcriptomic technologies and be aware of the statistical uncertainty that is implicit in bioinformatic analysis. A transcriptomics data analysis pipeline has multiple acceptable options for each step. In a research context, spending effort on making the analysis transparent and reproducible, and publicly sharing the data (i.e., adhering to the Findability, Accesibility, Interoperability, and Reusability [FAIR] principles) may be more beneficial than agonizing over the specific methods and parameters to use. In a regulatory context, standardization of transcriptomics data analysis approaches may be required as exemplified by the Transcriptomics Reporting Framework initiative from the OECD's Extended Advisory Group on Molecular Screening and Toxicogenomics (Harrill et al., 2021). The challenges we have discussed are not unique to transcriptomics data. We would not expect a single traditional toxicity test with a sample size of three to five to be conclusive in either a regulatory or a research context. The power of omics data lies in the ability to capture thousands of endpoints simultaneously, rather than a few predetermined ones (e.g., mortality or a single biomarker response), but it does not replace requirements for adequate sample size, consideration of toxicokinetics, or replicated studies. With new regulations about restricting animal use, and thousands of chemicals waiting for evaluation, the days of single biomarkers and "kill 'em and count 'em" toxicity testing are receding into the past. It will be up to the current generation of trainees and researchers to use adopter-centric thinking to modernize toxicity testing strategies. We must be realistic in the handling and interpretation of big data to harness its power to efficiently and wisely support decisions that protect us all. This article was adapted from a plenary presentation delivered by J. H. at the Canadian Ecotoxicology Workshop (CEW) annual meeting in October 2021 (Halifax, NS). The authors would like to thank the CEW organizers and B. de Jourdan for the invitation. We acknowledge that support over the years from Genome Canada and Genome Quebec (via the following programs: 2016 Large Scale Applied Research Program; 2018 Bioinformatics and Computational Biology Program; 2023 Genomics Applications Partnership Program) have helped to develop our thinking about the role of transcriptomics in ecotoxicology. Jessica A. Head: Conceptualization; Data curation; Formal analysis; Visualization; Writing—original draft. Jessica D. Ewald: Data curation; Formal analysis; Writing—review & editing. Niladri Basu: Conceptualization; Writing—review & editing. The raw data for the illustrative case study are available in the Gene Expression Omnibus of the National Center for Biotechnology Information under accession number GSE214620.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.