scientificscientificreport report Structure homology and interaction redundancy for discovering virus–host protein interactions Benoıˆt de Chassey 1,Laure`ne Meyniel-Schicklin 2,3, Anne Aublin-Gex 2,3, Vincent Navratil 2,3,w, Thibaut Chantier 2,3, Patrice Andre´1,2,3 &VincentLotteau2,3+ 1Hospices Civils de Lyon, Hoˆpital de la Croix-Rousse, Laboratoire de virologie, Lyon , 2Inserm, U1111, Lyon , and 3Universite´ de Lyon, Lyon, France Virus–host interactomes are instrumental to understand global purification coupled to mass spectrometry, led to the publication perturbations of cellular functions induced by infection and of the first virus–host interactomes [3–4]. Although incorporation discover new therapies. The construction of such interactomes is, of data from different methods or variation of the same method [5] however, technically challenging and time consuming. Here we has improved the quality of the data sets, diversification of describe an original method for the prediction of high-confidence methods is still clearly needed to generate high-quality interactions between viral and human proteins through comprehensive virus–host interactomes. In addition, regarding a combination of structure and high-quality interactome data. the size of host genome and the huge diversity of viruses, millions Validation was performed for the NS1 protein of the influenza of binary interactions remain to be tested. Therefore, accurate and virus, which led to the identification of new host factors that rapid methods for the identification of cellular interactors control viral replication. controlling viral replication is a major issue and a crucial step Keywords: interactome; prediction; protein interaction; towards the selection of original therapeutic targets and drug structure; virus development. This could benefit from predictive methods EMBO reports advance online publication 6 September 2013; doi:10.1038/embor.2013.130 preselecting subsets of putative interactions. Computational methods are essentially dedicated to predict intraspecies interac- tions [6] but applications to pathogen–host protein inter- actions are now emerging [7]. These methods are generally INTRODUCTION evaluated by the overlap between predicted and published Viruses are obligate parasites that rely on cellular functions to interaction data sets. Here, we present and experimentally replicate. Identifying the cellular functions that a virus must use, validate an original method for the prediction of virus–host counteract or interfere with to assure its replication is therefore a interactions combining protein structure homology and major issue for both basic knowledge and drug discovery. interaction redundancy. Biological processes being largely dependent on protein–protein interactions, cellular functions involved in viral replication can be identified by mapping virus–host protein–protein interactions and RESULTS AND DISCUSSION by constructing virus–host interactomes. Until 2007, viral protein– Principle of the method host protein interactions have essentially been identified in Several methods have proposed to use structural information to low-throughput experiments. Several databases are integrating predict protein–protein interactions [8–10]. The method described these data spread in the literature [1–2]. The construction of here relies on the assumption that when two proteins are physical viral ORFeomes and the development of high-throughput structurally homologous, they are more likely to have interactors technologies, such as yeast two-hybrid (Y2H) and tandem affinity in common (Fig 1A). Using the Structural Classification of Proteins from SCOP database [11] and the high-quality data sets of 1Hospices Civils de Lyon, Hoˆpital de la Croix-Rousse, Laboratoire de virologie, 103 grande rue de la Croix-Rousse, Lyon, F-69004, protein–protein interactions from the VirHostNet database [2], we 2Inserm, U1111, 21 Avenue Tony Garnier, Lyon, F-69007, showed that this assumption is a general feature of human 3Universite´ de Lyon, Lyon, France and viral proteins (supplementary information online). Briefly, wPresent address: Laboratoire de Biome´trie et Biologie E´volutive, CNRS, Universite´ Claude Bernard Lyon 1, 43 boulevard du 11 Novembre 1918, 69622 Villeurbanne for a viral protein having a solved structure, structurally Cedex, France. homologous human and viral proteins are first selected. Inter- +Corresponding author. Tel: þ 33 437262971; Fax: þ 33 437282341; actors of these homologous proteins are then identified. These E-mail: [email protected] proteins are considered as putative interactors and ranked Received 14 February 2013; revised 29 July 2013; accepted 31 July 2013; according to a score that favors proteins independently identified published online 6 September 2013 from multiple structural homologues. &2013 EUROPEAN MOLECULAR BIOLOGY ORGANIZATION EMBO reports 1 Prediction of virus–host protein interactions scientificreport B. de Chassey et al A Protein of interest Viral and human Human Best predicted New picture of (full-length 3D structural interactors interactors the interactome or domain) homologs Scoring Protein-protein interaction db (VirHostNet) B 16,187 HH ppi VirHostNet high–quality dataset 72,339 HH ppi Chain 22,052 PDB 14,422 H protein chains (3,163 Splitting 344 homologous 1,377 predicted structures H proteins) H proteins H interactors 1/all Influenza NS1 full- pairwise 2< 5 % Experimental length structure structure z-score 69 predicted Scoring benchmarking PDB (3F5TA) alignments <20 best H interactors (Dali) Score 4,933 PDB 3,130 V protein Chain 42 homologous 86 predicted >3 chains (440 V structures Splitting V proteins H interactors proteins) 26 predicted VirHostNet 2,539 VH ppi high- H interactors 4,458 VH ppi quality data set Fig 1 | Description of the method. (A) Principle of the method. (B) Detailed steps of the method applied to influenza A virus NS1 protein. The score 2 2 favors independent predictions provided by interaction redundancies. Score (putative interactor) ¼ [(nH /kH) þ (nV /kV)] Â a. With nH and nV number of distinct human interactors and viral interactors, respectively; kH and kV degree of the protein in the human protein–protein interaction data set and in the virus–human protein–protein interaction data set, respectively; a ¼ 3 for a prediction from both human and virus–human data sets, a ¼ 2 from only virus–human data set, 1 else. H, human; V, virus. Prediction of influenza A virus NS1 protein interactions of protein–protein interactions from the VirHostNet database [2], As a proof of concept, we applied this method to identify new a total of 1,384 human proteins were predicted to interact with partners of the influenza A NS1 protein whose full-length structure NS1- 1,377 interacting with human proteins that are homologous has been solved by x-ray crystallography [12], (Fig 1B). A total of to NS1 and 86 interacting with viral proteins that are homologous 22,052 human protein structures and 4,933 viral protein structures to NS1. These 1,384 proteins were then ranked according to respectively corresponding to 3,163 human proteins and 440 viral the number and origin of the NS1 homologues weighted by proteins were extracted from the Protein Data Bank (PDB). The their degree. The 5% top-ranked candidates, corresponding to standalone version 3.1 of the Dali programme was used to perform 69 proteins, were selected for further validation (Fig 1B, one-against-all pairwise alignments between NS1 and these large supplementary Table S1 online). In silico mapping of the predicted data sets of viral and human protein structures [13]. NS1 structural interactions can also be performed when the structure of domains homologues with a Dali z-score between 2 and 20 were retained or fragments are available. For NS1, an RNA-binding domain to include weakly homologous proteins and exclude fully and an effector domain have been structurally defined (PDB homologous NS1 proteins [14]. All homologues that are full- identifiers: 2Z0A and 3D6R, supplementary Table S2 online). Fifty length or part of NS1 proteins of influenza A viruses were also eight of the 69 predicted interactors for the full-length NS1 protein excluded. Hence, known NS1–human interactors were not favored. were predicted to interact with the RNA-binding domain while Influenza NS1 protein was found structurally homologous to 344 none of them were predicted to interact with the effector domain human proteins and 42 viral proteins. Using high-quality data sets (supplementary Table S1 online). 2 EMBO reports &2013 EUROPEAN MOLECULAR BIOLOGY ORGANIZATION Prediction of virus–host protein interactions B. de Chassey et al scientificreport C 22.0 AB20.0 Final predicted Interaction siRNA 18.0 cellular interactor validation assay 16.0 Official GST pull Neuraminidase 14.0 Ensembl ID Y2H PFU name down activity 12.0 ENSG00000002330 BAD × × 10.0 ENSG00000153094 BCL2L11 × × ×× 8.0 ENSG00000107223 EDF1 × × 6.0 ENSG00000110324 IL10RA × × ×× ENSG00000169047 IRS1 × × 4.0 ENSG00000143393 PIK4B × × Replication fold increase 2.0 ENSG00000132155 RAF1 × × ××GST GST–NS1 GST–NS1–RBD GST–NS1–effector 0.0 ENSG00000160691 SHC1 × × ENSG00000136560 TANK × × IB: myc BID CTRL TAF1 ENSG00000175564 UCP3 × × ENSG00000168646 AXIN2 × ×× D BCL2L11 ENSG00000087088 BAX × ×× Pull 1.2 ENSG00000015475 BID × ×× down 1 ENSG00000179151 EDC3 × ×× IB: GST ENSG00000030110 BAK1 ENSG00000157764 BRAF 0.8 ENSG00000101224 CDC25B 0.6 ENSG00000189229 CDK11B ENSG00000107566 CHUK 0.4 ENSG00000143466 IKBKE IB: myc Input ENSG00000185950