There are many studies have been devoted to solve the all-atom protein back- bone reconstruction problem (PBRP), such as Adcock¡¦s method, MaxSprout, SAB- BAC and Chang¡¦s method. In the previous work, Wang et al. tried to solve this problem by homology modeling. Then, Chang et al. improved Wang¡¦s result by refining the positions of oxygen based on the AMBER force field. We compare the results in CASP7 and 8 from Chang et al. and SABBAC v1.2 and find that some proteins get better predicting results by Chang¡¦s method and others do better in SABBAC. Based on SVM, we propose a tool preference classification method for determining which tool is potentially the better one for predicting the structure of a target protein. We design a series of steps to select the better feature sets for SVM. Our method is tested on the proteins with standard amino acids in CASP7 and 8 dataset, which contains 30 and 24 protein sequences, respectively. The experimen- tal results show that our method has 7.39% and 2.94% RMSD improvement against Chang¡¦s result in CASP7 and 8, respectively. Our method can also be applied to other effective prediction methods, even if they will be developed in the future.