Support Vector Machine Classification of the Carcinogenic Activity of Polycyclic Aromatic Hydrocarbons
Ovidiu Ivanciuc, BioChem Press, Ovidiu Ivanciuc · 2002
Motivation. Structure–activity relationships (SAR) can be efficiently used to predict the carcinogenic hazard of new chemicals, before producing them on a large scale or even before synthesizing them. SAR models that detect potential carcinogens can also supplement short–term tests of genotoxicity, long–term tests of carcinogenicity in rodents, or epidemiological evidence in humans. Method. Support vector machine (SVM) is an efficient classification algorithm that can provide highly predictive SAR models for the carcinogenic hazard. We have applied the SVM model to identify the carcinogenic activity of 46 methylated and 32 non–methylated polycyclic aromatic hydrocarbons (PAH). The PAH chemical structure was encoded by four theoretical descriptors computed with PM3, namely the energy of the highest occupied molecular orbital EHOMO, the energy of the lowest unoccupied molecular orbital ELUMO, the hardness HD, and the difference between EHOMO and EHOMO–1. Results. A wide range of SVM experiments were performed using the dot, polynomial, radial basis function, neural, and anova kernels. The results obtained for the classification of PAH carcinogenicity demonstrate that the performances of SVM depend strongly on the kernel type and various parameters that control the kernel shape. The best prediction results were obtained with the radial basis function kernel with J = 0.5, the anova kernel with J = 0.5 and d = 1, and the anova kernel with J = 0.5 and d = 2. In the first case, from 34 carcinogenic compounds, 28 were correctly classified, while from 44 non–carcinogenic compounds, 40 were correctly classified. Conclusions. SAR models for predicting the carcinogenic hazard can benefit from the use of support vector machines, which determine a maximum separating hyperplane between carcinogenic and non–carcinogenic compounds. The solution of the SVM model is a unique hyperplane which can be computed very fast, but the classification results heavily depend on the kernel type and structural descriptors. Extensive cross–validation tests should be made to find the kernel with the optimum predictive power.