Quantitative structure-activity relationship analysis has been carried out for 74 diaryl ureas including aminobenzoisoxazole ureas, aminoindazole ureas, aminopyrazolopyridine ureas against vascular endothelial growth factor receptor-2 kinase using both linear and non-linear models. Considering simplicity and predictivity, multivariate linear regression was first employed in combination with various variable selection methods, including forward selection, genetic algorithm and enhanced replacement method based on descriptors generated by e-dragon software. Another model using support vector regression has also been constructed and compared. Performances of these models are rigorously validated by leave-one-out cross-validation, fivefold cross-validation and external validation. The enhanced replacement method model significantly outperforms the others with R(2) = 0.813 and R(2)(pred) = 0.809. Robustness and predictive ability of this model is prudently evaluated. Moreover, to find out the most significant features associated with the difference between highly active compounds and moderate ones, two classification models using linear discriminant analysis and support vector machine were further developed. The performance of support vector machine significantly outperforms linear discriminant analysis, with leave-one-out cross-validation and external validation prediction accuracy reaching 0.838 and 0.857, respectively. The resulting models could act as an efficient strategy for estimating the vascular endothelial growth factor receptor-2 inhibiting activity of novel diaryl ureas and provide some insights into the structural features related to the biological activity of these compounds.