Abstract
We predicted the biodegradability of 554 chemicals by using a nonlinear partial least squares (PLS) method, called kernel PLS (KPLS), and compared the predictive performance of KPLS and that of linear PLS, which is widely used for modeling structure-activity relationships. Moreover, prediction using support vector machine (SVM) was performed to confirm the utility of KPLS. The KPLS models correctly categorized 429 (77.4%), 443 (80.0%) and 454 (81.9%) chemicals out of 554, whereas the PLS models were correct for 419 (75.6%), 434 (78.3%) and 439 (79.2%) in cases of using 6, 50 and 89 descriptors, respectively, based on the chemical structures of chemicals analyzed in this study. By properly tuning the necessary parameters, KPLS showed better predictive performance for the biodegradability of chemicals than PLS and SVM did, which showed 79.6% accuracy with 89 descriptors.
