Table 1.

Summary of machine learning classification methods for protein–protein interaction hot spot prediction.

Classification Methods	Description	References
Nearest neighbor	The model consists of 83 classifiers using the IBk algorithm, where instances are encoded by sequence properties.	Hu et al. [58]
	Training the IBk classifier through the training dataset to obtain several better random projections and then applying them to the test dataset.	Jiang et al. [16]
Support vector machine	The decision tree is used to perform feature selection and the SVM is applied to create a predictive model.	Cho et al. [30]
	F-score is used to remove redundant and irrelevant features, and SVM is used to train the model.	Xia et al. [28]
	Proposed two new models of KFC through SVM training	Darnell et al. [31]
	The two-step feature selection method is used to select 38 optimal features, and then the SVM method is used to establish the prediction model.	Deng et al. [11]
	The random forest algorithm is used to select the optimal 58 features, and then the SVM algorithm is used to train the model.	Ye et al. [59]
	Use the two-step selection method to select the two best features, and then use the SVM algorithm to build the classifier.	Xia et al. [3]
	When the interface area is unknown, it is also very effective to use this method.	Qian et al. [48]
Decision trees	Formed by a combination of two decision tree models, K-FADE and K-CON.	Darnell et al. [31]
Bayesian networks	Can handle some of the missing protein data, as well as unreliable conditions.	Assi et al. [65]
Neural networks	Does not need to know the interacting partner.	Ofran and Rost [66]
Ensemble learning	The mRMR algorithm is used to select features, SMOTE is used to handle the unbalanced data, and finally AdaBoost is used to make prediction.	Huang and Zhang [72]
	Random forest (RF) is used to effectively integrate hybrid features.	Wang et al. [71]
	Bootstrap resampling approaches and decision fusion techniques are used to train and integrate sub-classifiers.	Deng et al. [11]