| Protein function prediction is a fundamental problem in the field of bioinformatics and plays a crucial role in understanding cellular processes and molecular interactions.Accurate identification of submitochondrial protein localization and protein-protein interaction(PPI)binding sites is essential to unravel complex biological networks and guide drug discovery efforts.Current computational models for predicting protein submitochondrial localization and PPI binding sites rely primarily on biometric or evolutionary information,which severely limits the expression of sequence information and the performance of computational models.Moreover,the potential of deep learning for these tasks has not been fully exploited,and there is a growing need for more efficient and accurate models.Based on this,this thesis innovatively proposes two independent computational frameworks that fuse dynamic word embedding coding and deep learning techniques to address the challenge of protein function prediction.First,this thesis proposes a new computational method,called i Deep Sub Mito,to predict the location of mitochondrial proteins in mitochondrial compartments.The method employs the Protein ELMo algorithm based on dynamic word vector encoding to model the probability distribution of protein sequences and represents protein sequences as continuous vectors.Then a bidirectional LSTM-based convolutional neural network architecture with a self-attentive mechanism is implemented to efficiently explore contextual information and semantic features of protein sequences.To demonstrate the effectiveness of i Deep Sub Mito,this thesis performs crossvalidation on two datasets,which contain 424 proteins and 570 proteins,respectively,consisting of four different mitochondrial regions(matrix,inner membrane,outer membrane and intermembrane regions).The experimental results showed that i Deep Sub Mito outperformed other computational methods.In addition,the performance of i Deep Sub Mito was tested on the datasets of M187,M983 and Mito Carta3.0 to further validate its effectiveness.Finally,Motif analysis and interpretability analysis are conducted to reveal new insights of i Deep Sub Mito on the subcellular biological functions of mitochondrial proteins.Secondly,this paper also proposes a computational framework for identifying PPI binding sites,the Ensemble Deep Learning Model(EDLM)-based Protein-Protein Interaction(PPI)site identification method(EDLMPPI).EDLMPPI is implemented through a Dynamic word embedding model(Prot T5)based on Transformer architecture to extract potential associations between protein level structures,capturing their functional and structural properties only from readily available protein sequence data.After that,EDLMPPI is based on Bi LSTM to fully learn the contextual associations between features and retain the contextual information through capsule network to further improve the generalization performance.Experimental results show that EDLMPPI can successfully capture protein-protein interactions and identify binding residues involved in the interactions.the accuracy of EDLMPPI is nearly 10% higher than the state-of-the-art model,significantly improving the efficiency and accuracy of PPI binding site prediction.In addition,the analysis of biological and interpretable dimensions provides new perspectives on the identification and characterization mechanisms of protein binding sites from different perspectives. |
