CORAL: The prediction of biodegradation of organic compounds with optimal SMILES-based descriptors

Abstract

CORAL software (http:/www.insilico.eu/coral) has been used to build up quantitative structure-biodegradation relationships (QSPR). The normalized degradation percentage has been used as the measure of biodegradation (for diverse organic compounds, n=445). Six random splits into sub-training, calibration, and test sets were examined. For each split the QSPR one-variable linear regression model based on the SMILES-based optimal descriptors has been built up. The average values of numbers of compounds and the correlation coefficients (r 2) between experimental and calculated biodegradability values of these six models for the test sets are n=88.2±11.7 and r 2=0.728±0.05. These six models were further tested against a set of chemicals (n=285) for which only categorical values (biodegradable or not) were available. Thus we also evaluated the use of the model as a classifier. The average values of the sensitivity, specificity, and accuracy were 0.811±0.019, 0.795±0.024, and 0.803±0.008, respectively. © Versita Sp. z o.o.