A new approach for quantitative structure-activity relationships based on MNA descriptors, fuzzy gradation method and self-consistent regression has been proposed. This approach has been realized in the computer program GUSAR (General Unrestricted Structure Activity Relationships). Our method has been validated on CDK1 inhibitors. Prediction accuracy is comparable with popular methods of 3D QSAR: CoMFA and CoMSIA. However, in contrast to CoMFA and CoMS1A, GUSAR approach does not require information about 3D structure of enzyme and ligand. Application of GUSAR method for heterogeneous training sets has been shown.