1340

A.SABLJIC AND W.PEIJNENBURG

(biophores), while the second file was searched for fragments explaining nonbiodegradability in the test, respectively (biophobes). The program located 48 biophores that could explain all 269 biodegrad-able compounds, as well as 10 biophobes. Finally, a multiple linear regression (MLR) relationship was built between the 58 selected fragments and the biodegradation data measured according to the MITI-I test protocol [12]. The model was capable of correctly classifying 92.5% of the data in the training set.

4.1.4 Biodegradation rules developed by the inductive machine learning method

The artificial intelligence technique (i.e., the inductive machine learning method [7–9]) and biodegra-dation data measured according to the MITI-I test protocol were used to develop structural rules for fast ultimate biodegradation [59]. The following seven rules have been derived for fast biodegradation (i.e., the chemical will biodegrade fast if any of the following seven rules applies):

?esters, amides, or anhydrides with a larger number of ester groups than rings

?all chemicals with at least one acyclic C–O bond and molecular weight below 129 ?chemicals built of C, H, N, and O atoms and with larger number of esters groups than rings but without nitro group

?organic acids with molecular weight below 173 and with more acid groups than halogen atoms ?chemicals built of C, H, N, and O atoms with weight below 129 having equal number of aromat-ic amino groups and acid groups but without nitro group

?esters, amides, or anhydrides with molecular weight below 173 and at least one acyclic C–O bond ?chemicals built of C, H, N, and O atoms with molecular weight below 173 and at least one acyclic C–O bond, equal number of aromatic amino groups and acid groups, but without a nitro group This set of 7 rules is based only on 11 structural descriptors, selected from a pool of 17 [8]. It was possible to correctly classify 84.3% of chemicals from the MITI database with a well-balanced classi-fication of easily and poorly biodegradable chemicals.

An analysis of structural descriptors present in biodegradation rules was performed to extract infor-mation on requirements for either fast or slow biodegradation. The low molecular weight (below 173); presence of only C, H, N, and O atoms in a chemical; presence of C–O bonds; acyclic structures; as well as acid, ester, amide, and anhydride functional groups all seem to be stimulating features for biodegradation. On the other hand, the presence of rings, aromatic amines, halogen atoms, or a nitro group seems to retard the biodegradation process. These findings are also consistent with the general experience.

4.2 Validation of biodegradation models

4.2.1 BIODEG model

To assess the real value of the predictive power of biodegradation models, it is most useful to focus on the external validation of models. The linear BIODEG model based on evaluated biodegradation data [18,55] has been evaluated on a large set of consistent biodegradation data (i.e., 733 compounds test-ed with the MITI-I test). We expect those evaluation results to be realistic, as well as a solid indicator of the model’s future performance in predicting biodegradability of new compounds, since they are based on a large dataset of structurally diverse chemicals. The results of this validation [12] are given in Table 1.

This test showed that the overall performance of the BIODEG model is only 61.1% of correct pre-dictions. This is a relatively poor result. However, in the case of “ready” biodegradable chemicals its prediction score is high, namely, 91.1% of the predictions are correct. However, the BIODEG model was not modeled to predict the outcome of the MITI-I test, which was found to be a more strict meas-ure of biodegradability than the evaluated biodegradation data from the BIODEG dataset. Thus, it was logical to find out that the BIODEG model predicts a significant number of chemicals to be “ready”biodegradable although they were evaluated by MITI-I test as “not ready” biodegradable chemicals. It

?2001 IUPAC, Pure and Applied Chemistry73, 1331–1348