Outliers toxicity

Lipnick, R.L. (1991). Outliers Their origin and use in the classification of molecular mechanisms of toxicity. Science of the Total Environment 109/110, 131-153. 

Toxicology has long recognized that no population, animal or human, is completely uniform in its response to any particular toxicant. Rather, a population is composed of a (presumably normal) distribution of individuals some resistant to intoxication (hyporesponders), the bulk that respond close to a central value (such as an LD50), and some that are very sensitive to intoxication (hyperresponders). This population distribution can, in fact, result in additional statistical techniques. The sensitivity of techniques such as ANOVA is reduced markedly by the occurrence of outliers (extreme high or low values, including hyper- and hyporesponders), which,... 

Occasionally a study in toxicology can reverse the direction of information flow. In a set of anilines which seemed to be more toxic as their log P was increased, there was one outlier, the 2,6-di-isopropyl aniline shown in Figure 5.9. It was much less toxic than predicted... 

To investigate the influence of the change of the algorithm for Am jx calculation on the coefficients in the model, only the compounds, considered in Schultz (1999) for training (n = 200) were used in the analysis (see Table 12.1). The compounds, being not toxic at saturation as well as those detected as outliers in Schultz (1999), were excluded prior to the modeling. The resulting equation ... 

Evaluation of all 215 compounds with measurable toxicity, reported by Schultz (1999), with the newly calculated A showed that only one compound (4-chloro-3,5-dinitrobenzonitrile, Res. = 1.16, St. res. = 3.24) is a statistically significant outlier and this was excluded from the data set. The resulting model is shown below ... 

Ehtylbenzene and 3-xylene (log Kow values of 3.15 and 3.20, respectively) were removed because they were outliers (R2 = 0.73 when included in the analysis) (data not shown). Their removal can be justified by the fact that for the highest concentrations tested, these compounds were found to be toxic to a luminescent metabolic biosensor of P. putida FI, and it may be assumed that they were toxic to P. putida TVA8 accordingly. As a consequence of the removal of these data, there is a significant gap of chemicals for octanol-water partition coefficients between 2.73 and 4.90. 

The presence of outliers is not necessarily bad because they can indicate aberrant behavior where, for example, the compound may undergo a different mechanism than the other compounds in the data set. For example, aldehydes did not fit an otherwise good correlation equation for a toxicity index presumably because of Schiff base formation at membrane surfaces by the aldehyde group with amino groups. Outliers also can be indicative of an... 

---