Transfer learning via multi-scale convolutional neural layers for human-virus protein-protein interaction prediction.

MOTIVATION: To complement experimental efforts, machine learning-based computational methods are playing an increasingly important role to predict human-virus protein-protein interactions (PPIs). Furthermore, transfer learning can effectively apply prior knowledge obtained from a large source dataset/task to a small target dataset/task, improving prediction performance. RESULTS: To predict interactions between human and viral proteins, we combine evolutionary sequence profile features with a Siamese convolutional neural network (CNN) architecture and a multi-layer perceptron. Our architecture outperforms various feature encodings-based machine learning and state-of-the-art prediction methods. As our main contribution, we introduce two transfer learning methods (i.e. 'frozen' type and 'fine-tuning' type) that reliably predict interactions in a target human-virus domain based on training in a source human-virus domain, by retraining CNN layers. Finally, we utilize the 'frozen' type transfer learning approach to predict human-SARS-CoV-2 PPIs, indicating that our predictions are topologically and functionally similar to experimentally known interactions. AVAILABILITY AND IMPLEMENTATION: The source codes and datasets are available at https://github.com/XiaodiYangCAU/TransPPI/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Current status and future perspectives of computational studies on human-virus protein-protein interactions.

The protein-protein interactions (PPIs) between human and viruses mediate viral infection and host immunity processes. Therefore, the study of human-virus PPIs can help us understand the principles of human-virus relationships and can thus guide the development of highly effective drugs to break the transmission of viral infectious diseases. Recent years have witnessed the rapid accumulation of experimentally identified human-virus PPI data, which provides an unprecedented opportunity for bioinformatics studies revolving around human-virus PPIs. In this article, we provide a comprehensive overview of computational studies on human-virus PPIs, especially focusing on the method development for human-virus PPI predictions. We briefly introduce the experimental detection methods and existing database resources of human-virus PPIs, and then discuss the research progress in the development of computational prediction methods. In particular, we elaborate the machine learning-based prediction methods and highlight the need to embrace state-of-the-art deep-learning algorithms and new feature engineering techniques (e.g. the protein embedding technique derived from natural language processing). To further advance the understanding in this research topic, we also outline the practical applications of the human-virus interactome in fundamental biological discovery and new antiviral therapy development.

HVIDB: a comprehensive database for human-virus protein-protein interactions.

While leading to millions of people's deaths every year the treatment of viral infectious diseases remains a huge public health challenge.Therefore, an in-depth understanding of human-virus protein-protein interactions (PPIs) as the molecular interface between a virus and its host cell is of paramount importance to obtain new insights into the pathogenesis of viral infections and development of antiviral therapeutic treatments. However, current human-virus PPI database resources are incomplete, lack annotation and usually do not provide the opportunity to computationally predict human-virus PPIs. Here, we present the Human-Virus Interaction DataBase (HVIDB, http://zzdlab.com/hvidb/) that provides comprehensively annotated human-virus PPI data as well as seamlessly integrates online PPI prediction tools. Currently, HVIDB highlights 48 643 experimentally verified human-virus PPIs covering 35 virus families, 6633 virally targeted host complexes, 3572 host dependency/restriction factors as well as 911 experimentally verified/predicted 3D complex structures of human-virus PPIs. Furthermore, our database resource provides tissue-specific expression profiles of 6790 human genes that are targeted by viruses and 129 Gene Expression Omnibus series of differentially expressed genes post-viral infections. Based on these multifaceted and annotated data, our database allows the users to easily obtain reliable information about PPIs of various human viruses and conduct an in-depth analysis of their inherent biological significance. In particular, HVIDB also integrates well-performing machine learning models to predict interactions between the human host and viral proteins that are based on (i) sequence embedding techniques, (ii) interolog mapping and (iii) domain-domain interaction inference. We anticipate that HVIDB will serve as a one-stop knowledge base to further guide hypothesis-driven experimental efforts to investigate human-virus relationships.

Human Gene Functional Network-Informed Prediction of HIV-1 Host Dependency Factors.

Human immunodeficiency virus type 1 (HIV-1) depends on a class of host proteins called host dependency factors (HDFs) to facilitate its infection. So far experimental efforts have detected a certain number of HDFs, but the gene inventory of HIV-1 HDFs remains incomplete. Here, we implemented an existing network-based gene discovery strategy to predict HIV-1 HDFs. First, an encoding scheme based on a publicly available human tissue-specific gene functional network (GIANT; http://giant.princeton.edu/) was designed to convert each human gene into a 25,825-dimensional feature vector. Then, a random forest-based predictive model was trained on a data set containing 868 known HDFs and 1,736 non-HDFs. Through 5-fold cross-validation, an independent test, and comparison with one existing method, the proposed prediction method consistently revealed accurate and competitive performance. The highlight of our method should be ascribed to the introduction of the GIANT encoding scheme, which contains rich information regarding gene interactions. By merging known HDFs and genome-wide HDF prediction results, network analysis was conducted to catch the common patterns of HDFs in the context of the GIANT network. Interestingly, HDFs reveal significantly lower betweenness than HIV-1-interacting human proteins (i.e., HIV targets). In the meantime, the functional roles of HDFs were also examined by mapping all the HDF candidates into human protein complexes. Especially, we observed the frequent co-occurrence of HDFs and HIV targets at the protein complex level. Collectively, we hope the proposed prediction method not only can accelerate the HDF identification and antiviral drug target discovery, but also can provide some mechanistic insights into human-virus relationships.IMPORTANCE Identification of HIV-1 HDFs remains a crucial step to understand the complicated relationships between human and HIV-1. To complement the experimental identification of HDFs, we have implemented an existing network-based gene discovery strategy to predict HDFs from the human genome. The core idea of the proposed method is that the rich information deposited in host gene functional networks can be effectively utilized to infer the potential HDFs. We hope the proposed prediction method could further guide hypothesis-driven experimental efforts to interrogate human-HIV-1 relationships and provide new hints for the development of antiviral drugs to combat HIV-1 infection.

The MERS-CoV Receptor DPP4 as a Candidate Binding Target of the SARS-CoV-2 Spike.

The ongoing outbreak of the novel coronavirus pneumonia COVID-19 has caused great number of cases and deaths, but our understanding about the pathogen SARS-CoV-2 remains largely unclear. The attachment of the virus with the cell-surface receptor and a cofactor is the first step for the infection. Here, bioinformatics approaches combining human-virus protein interaction prediction and protein docking based on crystal structures have revealed the high affinity between human dipeptidylpeptidase 4 (DPP4) and the spike (S) receptor-binding domain of SARS-CoV-2. Intriguingly, the crucial binding residues of DPP4 are identical to those that are bound to the MERS-CoV-S. Moreover, E484 insertion and adjacent substitutions should be most essential for this DPP4-binding ability acquirement of SARS-CoV-2-S compared with SARS-CoV-S. This potential utilization of DPP4 as a binding target for SARS-CoV-2 may offer novel insight into the viral pathogenesis and help the surveillance and therapeutics strategy for meeting the challenge of COVID-19.

Understanding Human-Virus Protein-Protein Interactions Using a Human Protein Complex-Based Analysis Framework.

Computational analysis of human-virus protein-protein interaction (PPI) data is an effective way toward systems understanding the molecular mechanism of viral infection. Previous work has mainly focused on characterizing the global properties of viral targets within the entire human PPI network. In comparison, how viruses manipulate host local networks (e.g., human protein complexes) has been rarely addressed from a computational perspective. By mainly integrating information about human-virus PPIs, human protein complexes, and gene expression profiles, we performed a large-scale analysis of virally targeted complexes (VTCs) related to five common human-pathogenic viruses, including influenza A virus subtype H1N1, human immunodeficiency virus type 1, Epstein-Barr virus, human papillomavirus, and hepatitis C virus. We found that viral targets are enriched within human protein complexes. We observed in the context of VTCs that viral targets tended to have a high within-complex degree and to be scaffold and housekeeping proteins. Complexes that are essential for viral propagation were simultaneously targeted by multiple viruses. We characterized the periodic expression patterns of VTCs and provided the corresponding candidates that may be involved in the manipulation of the host cell cycle. As a potential application of the current analysis, we proposed a VTC-based antiviral drug target discovery strategy. Finally, we developed an online VTC-related platform known as VTcomplex (http://zzdlab.com/vtcomplex/index.php or http://systbio.cau.edu.cn/vtcomplex/index.php). We hope that the current analysis can provide new insights into the global landscape of human-virus PPIs at the VTC level and that the developed VTcomplex will become a vital resource for the community. IMPORTANCE Although human protein complexes have been reported to be directly related to viral infection, previous studies have not systematically investigated human-virus PPIs from the perspective of human protein complexes. To the best of our knowledge, we have presented here the most comprehensive and in-depth analysis of human-virus PPIs in the context of VTCs. Our findings confirm that human protein complexes are heavily involved in viral infection. The observed preferences of virally targeted subunits within complexes reflect the mechanisms used by viruses to manipulate host protein complexes. The identified periodic expression patterns of the VTCs and the corresponding candidates could increase our understanding of how viruses manipulate the host cell cycle. Finally, our proposed conceptual application framework of VTCs and the developed VTcomplex could provide new hints to develop antiviral drugs for the clinical treatment of viral infections.

Machine-Learning-Based Predictor of Human-Bacteria Protein-Protein Interactions by Incorporating Comprehensive Host-Network Properties.

The large-scale identification of protein-protein interactions (PPIs) between humans and bacteria remains a crucial step in systematically understanding the underlying molecular mechanisms of bacterial infection. Computational prediction approaches are playing an increasingly important role in accelerating the identification of PPIs. Here, we developed a new machine-learning-based predictor of human- Yersinia pestis PPIs. First, three conventional sequence-based encoding schemes and two host network-property-related encoding schemes (i.e., NetTP and NetSS) were introduced. Motivated by previous human-pathogen PPI network analyses, we designed NetTP to systematically characterize the host proteins' network topology properties and designed NetSS to reflect the molecular mimicry strategy used by pathogen proteins. Subsequently, individual predictive models for each encoding scheme were inferred by Random Forest. Finally, through the noisy-OR algorithm, 5 individual models were integrated into a final powerful model with an AUC value of 0.922 in the 5-fold cross-validation. Stringent benchmark experiments further revealed that our model could achieve a better performance than two state-of-the-art human-bacteria PPI predictors. In addition to the selection of a suitable computational framework, the success of our proposed approach could be largely attributed to the introduction of two comprehensive host network-property-related feature sets. To facilitate the community, a web server implementing our proposed method has been made freely accessible at http://systbio.cau.edu.cn/intersppiv2/ or http://zzdlab.com/intersppiv2/ .