T3 PM Aquatic Ecotoxicology (Part 2)
Tuesday, 15 November 2005: 1:50 PM - 5:30 PM in Ballroom 3

(MEN-1117-788379) Predicting toxicity of Alcohol Ethoxylates homologues by an artificial neural network.

Meng, YB¹, Lin, B-L¹, Tokai, A¹, Tominaga, M¹, Nakanishi, J¹, ¹ Research Center for Chemical Risk Management, National Institute of Advanced Industrial Science and Technology, Tsukuba City, Ibaraki Prefecture, Japan

ABSTRACT- The toxicity of alcohol ethoxylate (AE), a widely used non-ionic surfactant depends on its alkyl chain, polyoxyethylene (EO) chain and other conditions. The distribution of AE homologues in environment, termed as fingerprint, is usually different from that of AE products under use or upon disposal, which entails prediction of the untested homologues' toxicity from toxicity data of the tested homologues. We constructed a feed-forward artificial neural network (ANN) consisting of one hidden layer to predict the toxicity. The input of the ANN consisted of alkyl chain length, branching extent in alkyl chain, EO number, test condition, test endpoint, exposure time and species taxon. Existing toxicity data from experiments of lab and mesocosm were collected and broken down into 545 toxicity data for training of the ANN. Using early-stopping technique, six-neuron in the hidden layer was selected for good generalization. A leave-one-out cross validation process indicated that the 95% confidential interval (CI) of predicted toxicity data could include the 'true' (untrained) toxicity data with about 90% probability and the 99% CI with more than 95% probability. Simulating the ANN for AE homologues revealed that those with shorter alkyl chain, more EO, or branched alkyl chain were less toxic, but toxicity of long alkyl chain (C>14) homologues tended not to increase upon alkyl length. The predicted no-observed-effect-concentrations, NOECs, for various species, were fit to a log-normal species sensitivity distribution to infer a hazardous concentration, HC5.

Key words: acohol ethoxylate, artificial neural network, hazard concentration, species sensitivity distribution