| Predicting protein function is one of the most important things in Bioinformatics to help understand the disease’s behaviors.However,there are still many challenges to predict protein function.Recently,developed methods have been done and give good results using Convolutional Neural Networks(CNNs).The existing methods use three main protein features;protein sequences,protein interaction networks,and protein domains as predictive criteria.However,the available PPI data is still so less to obtain.So,the sequences similarity-based prediction is proposed in this research using CNNs as well.Furthermore,a comparison will be done on combining other features in the extended CNNs step for achieving the best state-of-art.In this experiment,the protein function prediction model is designed with two main input domains;protein sequence and protein structure.The protein sequence domain uses 3grams method as data extraction to make a scalable vector which can be calculated in CNNs architecture.And the protein structure domain uses PSSM matrix with three different handlings;the plain dataset which means using the original PSSM matrix,the weighted sum dataset which uses addition in normalized PSSM matrix,and the weighted product dataset which applying multiplication as the normalized PSSM matrix.The CNNs model is made up with 1-dimensional convolutional layer,Re LU layer as the activation function and max pooling layer.After that,the fully connected layer is classified by using Gene Ontology(GO)classifier which annotates the protein function in the GO terms.As the result,the evaluation performance metrics calculates its score to find whether the proposed model is good enough and better than the existing method.Fmax and AUC evaluation metrics is being used in this experiment. |
