MEIC evaluation of acute systemic toxicity - Part VIII. Multivariate partial least squares evaluation, including the selection of a battery of cell line tests with a good prediction of human acute lethal peak blood concentrations for 50 chemicals

The Multicenter Evaluation of In vitro Cytotoxicity (MEIC) programme was se t up to evaluate the relevance for human acute toxicity of in vitro cytotox icity tests. A total of 61 assays were used to test all 50 reference chemic als. The results of all the tests and the human database were presented in the first five papers of this series. An evaluation of the relevance for hu man acute toxicity of all submitted test results with use of hard linear re gression modelling was presented in the next two papers, and demonstrated a high relevance of in vitro tests, notably tests involving human cell lines . In the present study, multivariate partial least square (PLS) modelling w ith latent variables analysis has been used to reach two objectives. The fi rst objective was to study the prediction of human acute toxicity by the 61 assays. The second objective was to select a practical battery from the 61 assays, with an optimal prediction of lethal blood concentrations from hum an acute poisonings of the chemicals. A two-component PLS model of all 61 a ssays predicted three sets of lethal blood concentrations (clinical, forens ic and peak concentrations) very well (R-2 = 0.77, 0.76 and 0.83, Q(2) = 0. 74, 0.72 and 0.81, respectively), providing correlative evidence for a high relevance for human acute toxicity of most of the assays. The assays with human cells were highly predictive, whereas assays with Very short incubati on times and non-fish ecotoxicological assays were least predictive. These findings confirm the previous results from linear regression analysis. To s elect an optimal battery, 24 successive PLS models of in vitro data were co mpared with lethal peak concentrations. The battery selection was based on 38 chemicals with reliable and relevant lethal peak concentrations. An init ial PLS model of all 61 assays was used to select the 15 most predictive an d most distinct assays. Subsequent PLS models were used to measure the decr ease in prediction when assays were deleted from the 15-test battery, as we ll as the increase in prediction when some extra-predictive assays (as iden tified by the deletion process) were added later to an optimal two-test bat tery. The most predictive three-test battery (R-2 = 0.79 and Q(2) = 0.78 fo r all 50 chemicals) included two circumstantial assays. The most predictive and most cost-effective battery consisted of three human cell line assays, with four endpoints and two exposure times, i.e. protein content (24 hours ), ATP content (24 hours), inhibition of elongation of cells (24 hours), an d pH-change (7 days). This 1, 5, 9, 16 battery exclusively measures basal c ytotoxicity, and is highly predictive (R-2 = 0.77 and Q(2) = 0.76 for 50 ch emicals) of the actual lethal peak blood concentrations from acute poisonin gs in humans. The battery prediction compares favourably with the predictio n of human lethal dose by a PLS model of rat and mouse 50% lethal dose (LD5 0) values for the 50 chemicals (R-2 = 0.65 and Q(2) = 0.64). The three assa ys of the battery and other promising MEIC assays should be formally valida ted as soon as possible. The battery can be used immediately for several no n-regulatory purposes, including the high-throughput screening of potential pharmaceuticals.