The proposed hybrid model is basically composed of two parts. For the first part, a series of experiments were performed to find the best prediction model for each single feature we extracted. For the feature one hot encoding, AAC, PAAC, AAindex, and PSSM profile, we trained them by CNN, RF, and SVM separately. Moreover, every model was trained with peptides ranged from window size 15 to 35 to reach for the proper length of features. The default CNN model we designed included two rounds of the convolutional layers and max pooling layers following by three fully connected layers. The default RF model we designed was setting up the numbers of decision trees to 1000 with other parameters suggested by the RF package. Radial basis function kernel was chosen to be the default kernel function of our SVM model. In order to find the optimal model with the suitable window size, the one with the highest AUC score was selected. At this stage, we have 5 single-feature models for mammalian proteins and 5 single-feature models for plant proteins. Tensorflow and sklearn learn package in Python programming language was utilized.
The purpose of the second part was to aggregate the separated models trained by different features to make a final predicted decision. There were three methodologies for aggregation: another neural network, SVM, and majority votes. For the mammalian group, we found out that seeing the CNN models as a feature extractor and combing together to feed into another deep learning neural network returned the best AUC. We also noticed that using feature one hot encoding, AAindex, and PSSM profile three features has a better AUC than using all five features. As for the plant group, feeding the positive probability of each feature to a SVM model reached the best performance. The cutoff threshold is set to 0.046 for mammalian model, 0.195 for plant model.