To improve the precision of QSAR modeling, a novel nonlinear combinatorial forecast method based on support vector regression (SVR) and K-nearest neighbor (KNN) was proposed. Kernels and descriptors optimization based on SVR were evaluated by the rule of minimum MSE value, and "multi-round enforcement optimization" was taken to illuminate the effect of retained descriptors on forecasting precision. The heterogeneity of sample set was characterized by different KNN and multiple sub-models were assembled, then combinational forecast was car...