Multimodal deep representation learning for protein interaction identification and protein family classification

摘要

Protein-protein interaction (PPI) networks are becoming increasingly crucial for analyzing biomedical functions, retrospecting species evolution and analyzing different compounds that cause diseases. Moreover, comprehending the intrinsic patterns behind PPI networks facilitates the understanding of cancer-related protein-protein interfaces and the topological structures of the cancer networks. Normally, two groups of research methods can be formulated when analyzing PPI networks: computational biology methods and high-throughput experimental methods. Given a PPI network, computational biology methods calculate the distances between proteins according to network theory metrics (e.g. betweenness, centrality, average degree) or machine learning algorithms[ – ]. High-throughput techniques, on the contrary, including yeast two-hybrid screens (Y2Hs)[ ], mass spectrometry protein complex identification (MS-PCI) [ ] and Nuclear Magnetic Resonance (NMR)[ ], etc. pro- duce large amounts of data for constructing primary protein databases. These databases provide primary and rich sources for developing molecular and functional networks. Nevertheless, these genome-based techniques demand expensive wet-lab investment and exhaustive lab work. Also, because of the equipment biases in the experimental environment, the results generated by these genome-based methods are subjected to inevitable inaccuracy. Moreover, compared with the significant amount of protein sequence data, the functional units that have been discovered are comparatively restricted. Previously, traditional machine learning algorithms such as decision trees (DT), naive bayes (NB) and nearest neighbor (NN)[ ] have been utilized efficiently in lots of data mining tasks. Yet, these traditional machine learning techniques lack the capacity of discovering hidden associations and extracting discriminant features from the input complex data. Lately, accompanied with the advancement of AI techniques, deep learning methodologies[ ] extracting non-linear and high dimensional features from the protein sequences [ , ] have emerged as a new tendency. These deep learning techniques and frameworks have been recently applied in tremendous biomedical research fields, biological network analysis, and medical image examination. However, since natural and real-world data distributions are highly complex and multimodal, it is essential to incorporate different modalities and patterns from the data to attain satisfactory performance. Additionally, discovering biological pattern from the graph topology of these protein networks is fundamental in comprehending the functions of the cells and their constitutional proteins. When applying deep learning techniques to biological network analysis, these modalities include topological similarities such as 1 -order similarity, 2 -order similarity, and homology features extracted from protein sequences. Additionally, next-generation sequencing technologies also generate large amounts of DNA/RNA sequences which are then translated into protein peptides in the form of stacked amino acid residues. These protein sequences consist of fundamental molecules which perform biological functions for various species [ – ]. Thus, the functionality of a protein is encoded in the amino acid residues. To recognize the protein functionalities, researchers categorize proteins into various families such that proteins within the same family share similar functions or become the parts on the same pathway. In this paper, we propose a advanced multi-modal deep representation learning framework preserving different modalities to harvest both protein sequence similarity and topological proximity. This framework leverages both relational and physicochemical information from proteins and successfully integrates them using a late feature fusion technique. These concatenated features are provided to the interaction identifier and protein family classifier for the training and testing tasks.