LAUSR

W the past 20 years, metabolomics has moved from an exciting innovation within the environmental sciences to something that is almost routine. It can be considered as a means to generate metabolite biomarkers, although it is also important to note the cogent criticisms of the environmental biomarker approach that have been made within ecotoxicology: briefly, that biomarkers are surrogates for macro phenotypes (e.g., survival, reproduction, and behavior) that have population-level effects and that it is generally more straightforward and meaningful to measure these end points directly. Some studies have emphasized instead the ability to gain potentially relevant mechanistic information, even for nonmodel organisms, especially when used as part of a multiomic approach. An improved biological understanding is often implicitly or explicitly part of the justification of including metabolomics in a study. So far, so good, but there is a problem: there is no simple, universally accepted way of reverse engineering mechanistic understanding from metabolomic data, even for model organisms, and the problem is even more complicated for nonmodel species. The closest thing to a standard approach is pathway analysis (PA), i.e., making use of existing biochemical knowledge. There are multiple approaches to PA, but we will focus on just one, over-representation analysis (ORA). (NB that the term ORA is often not used, and many authors refer generically to “pathway enrichment” methods.) It should clearly be understood, though, that ORA is certainly not the only approach to analyzing metabolomics data. It is beyond the scope of this work to review the options available, but we direct the interested reader to recent reviews. ORA uses the intuitive approach of identifying metabolite biomarker “hits” and comparing them to the numbers of metabolites in specific pathways, to determine if there are either more or fewer hits than one would expect by chance. It therefore has the twin advantages of being simple to calculate and simple to understand. It does, though, have disadvantages. One potential limitation is shared with all methods that rely on predetermined pathway definitions: traditional pathways are, generally, subjective and heuristic approaches to imposing order on a biochemical network. While this is an important point, we will simply note it here and pass on, and bear in mind that “pathways” are, at least to some extent, arbitrary definitions. The problem is exacerbated for nonmodel organisms, in that accurate metabolic pathway definitions may not be available. It should also be noted that metabolites may contribute to many different pathways: for example, glucose is present in 23 of 263 pathways (KEGG, human), and ATP is present in 880 of 1669 pathways (Reactome, human). Just because a metabolite may be part of a particular pathway, then, does not mean that changes in that metabolite necessarily mean changes in that pathway. Particular care must be taken with environmental organisms not to misinterpret changes with respect to examples from human medicine. A second obvious limitation of ORA is that the criteria for defining significant metabolites are also arbitrary, usually, but not necessarily, based on selecting a threshold for P values from null hypothesis significance testing. It is also possible to draw incorrect conclusions from ORA. For instance, the online Metaboanalyst web server provides a suite of tools for metabolomic analysis, including, but not limited to, ORA. These have become justly popular, as they are free to use, available online, integrated with data processing and biostatistical modules, and updated to ensure they remain current. They also provide some opportunities to set parameters that affect the results, opening up the possibility of inadvertently misusing the tools. (NB that this is not an implicit criticism of the team behind Metaboanalyst: individual researchers should take responsibility for their own results, including interpretation.) We recently published a study of the sensitivity of ORA of metabolomics data to some of the different parameters that can be chosen. A wide range of different factors affect the results (Figure 1). First, and