Studies on Chemical Data Sets

The influence of the studies summarized above can be seen in the methods subsequently implemented by many other researchers for their applications (see the section on Chemical Applications). One method that was included in the original assessment studies, but not in the later assessments, is k-means. This method did not perform particularly well on the small data sets of the original studies, and the resultant clusters were found to be very dependent on the choice of initial seeds hence it was not included in the subsequent studies. However, k-means is computationally efficient enough to be of use for very large data sets. Indeed, over the last decade k-means and its variants have been studied extensively and developed for use in other disciplines. Because it is being increasingly used for chemical applications, any future comparisons of clustering methods should include k-means. 

---

Hattis et al. (1987) examined the variability in key pharmacokinetic parameters (elimination half-lives (Ty ), area under the curve (AUC), and peak concentration (C ax) in blood) in healthy adults based on 101 data sets for 49 specific chemicals (mostly drugs). For the median chemical, a 10-fold difference in these parameters would correspond to 7-9 standard deviations in populations of normal healthy adults. For one relatively lipophilic chemical, a 10-fold difference would correspond to only about 2.5 standard deviations in the population. The authors remarked that the parameters studied are only components of the overall susceptibility to toxic substances and did not include contributions from variability in exposure- and response-determining parameters. The study also implicitly excluded most human interindividual variability from age and diseases. When these other sources of variability are included, it is likely that a 10-fold difference will correspond to fewer standard deviations in the overall population and thus a greater number of people at risk of toxicity. 

In the chemical safety report, the hazard assessment of a particular substance is based on the data set provided in the technical dossier. This contains substance-specific information on physicochemical properties as well as on toxicological and ecotoxicological hazards. One objective of the hazard assessment is the substance s hazard identification, which comprises the determination of its physicochemical and hazardous properties for the purpose of classification. Concerning human health hazards, both human and nonhuman information is taken into consideration and evaluated with respect to the classification criteria laid down in the Dangerous Substances Directive and in the CLP Regulation, respectively. However, in most cases human data do not exist, so the hazard identification has to be based on data from animal experiments. With respect to teratogenicity, this hazardous property may in principle be detected in the following toxicity studies ... 

There are two general types of aerosol source apportionment methods dispersion models and receptor models. Receptor models are divided into microscopic methods and chemical methods. Chemical mass balance, principal component factor analysis, target transformation factor analysis, etc. are all based on the same mathematical model and simply represent different approaches to solution of the fundamental receptor model equation. All require conservation of mass, as well as source composition information for qualitative analysis and a mass balance for a quantitative analysis. Each interpretive approach to the receptor model yields unique information useful in establishing the credibility of a study s final results. Source apportionment sutdies using the receptor model should include interpretation of the chemical data set by both multivariate methods. 

No case study is exactly like another case. Each case study has its own features, defined by the natural setting and the nature of data obtained. The following case studies are an assortment of studies, heavily based on chemical data. 

Little has been reported on the use of hierarchical divisive methods for processing chemical data sets (other than the inclusion of the minimum-diameter method in some of the comparative studies mentioned above). Recursive partitioning, which is a supervised classification technique very closely related to monothetic divisive clustering, has, however, been used at the GlaxoSmithKline and Organon companies. 

---