Abstract: 18 new features such as residue contact number and the proportion of relative change of accessible surface area et al. were derived based on the analysis of protein-protein interaction energy hot spots. Two recursion feature elimination methods were used to select discriminative feature subsets and two corresponding prediction models were proposed, noted as SVM- RFE and F-Score-RFE. The experimental  results showed that the prediction model F-Score-RFE could improve 6.25％ in the value of F1 compared with the best existing method on the same independent test dataset, which  indicated that new features defined were significant to improve the performance of prediction.

Key words: prediction, recursion eliminate, protein-protein interaction, energy hot spots, feature selection