Predicting Protein-Protein Interaction Sites by Rotation Forests with Evolutionary Information

摘要

In this paper, according to evolutionary information and physicochemical properties, we selected eight features, combined with Rotation Forest (RotF) to predict interaction sites. We built two models on both balanced datasets and imbalanced datasets, named balanced-RotF and unbalanced-RotF, respectively. The values of accuracy, F-Measure, precision, recall and CC of balanced-RotF were 0.8133, 0.8064, 0.8375, 0.7775 and 0.6283 respectively. The values of accuracy, precision and CC of unbalanced-RotF increased by 0.0679, 0.0122 and 0.0361 over balanced-RotF. Precision values of unbalanced-RotF on our four selected testing sets were 0.907, 0.875, 0.878 and, 0.889, respectively. Moreover, experiment only using two physicochemical features showed evolutionary information has effective effects for classification.