From Experimental Approaches to Computational Techniques: A Review on the Prediction of Protein-Protein Interactions

Table 4

A summary of machine and statistical learning approaches applied to the predictive task of inferring PPI. Advantages and limitations for each approach are presented along with a reference to the studies where they have been applied.

Classifier	Description	Reference
NB	Ability to integrate diverse heterogeneous data. Can handle missing data. Assumes conditional independence between datasets. Performance of NB deteriorates when dependencies between features exist.	[10, 11]
KNN	Classification method which has been considered “simple but powerful” providing competitive performance compared with other classifiers [76]. Classifier performance may deteriorate if many variables are used or if the GS is not balanced.	[17, 74]
SVM	Can handle non-linearly separable datasets. Can incorporate prior information.	[77]
RF	Can handle missing values. Can integrate diverse heterogeneous data	[78, 79]
ANN	Ability to recognise complex patterns	[80–82]