Nonmetric Clustering and Association Analysis: Implications for the Evaluation of Multispecies Toxicity Tests and Field Monitoring

Abstract

Many techniques developed by computer scientists in the field of artificial intelligence (AI) are currently being used as standard, state-of-the-art technology. These techniques have proven their value and validity in medicine, geology, agronomy, and astronomy time and again, often beating human experts at their own game. We present here an analysis tool for multispecies data based on nonmetric clustering, an AI technique developed specifically to aid in the interpretation of complex ecological data sets. This technique uses AI search to find an appropriate and meaningful characterization of a multivariate system. After appropriately characterizing the system in this fashion, the relationship between this characterization of the system and the critical environmental variables (pollution, toxicity, etc.) can be quantitatively analyzed to aid in the assessment of the effects of the environment on the system. A priori endpoints or indices are not necessary; the data are allowed to determine the variables that best separate treatment from controls. We have now tested this methodology over a series of multispecies toxicity tests using a variety of Stressors. During the initial blind testing the methodologies could pick treatment groups with high accuracy. When knowledge of treatment group is available, oscillations in the similarity of the treatments to the controls are apparent. Much recent debate in toxicological studies has focussed on appropriate endpoints for multispecies toxicity tests and biomonitoring schemes. We suggest that the search for endpoints appropriate to the entire field of toxicity testing is a fruitless search. We recommend instead an approach that standardizes the common sense approach: different situations, even within a single experiment, call for different endpoints. Typically, the toxicologist, if called upon for an expert opinion, will examine multivariate data, and extract from that data a few critical species. The behavior of these species will give an adequate (though perhaps not complete) picture of the toxic effects. Which species are selected, and whether it is their mortality, behavior, or biomass that is important, will always vary from case to case. We call, therefore, for more research into the automation of the process typically performed by the expert. The selection of species, as well as other parameters, as significant for a particular experiment or field study, can be done automatically by computer algorithms. To be blind to the utility of these tools in the field of toxicology is to work by hand, over and over again, problems which could be solved in a twinkling with their aid.