Modeling and Predicting Protein-Protein Interactions of Type 2 Diabetes Mellitus Using Feedforward Neural Networks

Abstract

Data of protein-protein interactions (PPIs) are still limited. More data of PPIs are required so one can find significant proteins representing a disease more accurately. Computational approach which can predict PPIs is one of alternatives to reduce time and cost that generally required by experimental work. This research focused on predicting PPIs of Type 2 Diabetes mellitus using feedforward neural network (FNN). Impact of different activation functions, number of units per hidden layers and number of hidden layers themselves to estimation error were observed. Rectifier activation function, seven hidden layers and 36 units per hidden layers gave smallest MSE separately. The model with those configurations predicted a PPI with predicted combined score of 0.922. FNN model had better prediction accuracy than random forest and support vector regression models.