APPINetwork: an R package for building and computational analysis of protein–protein interaction networks

Introduction

Protein–protein interactions (PPIs) are central to many cellular processes. PPIs are identified and characterized experimentally by different methods which determine whether two proteins make physical contact or if they belong to a transient or permanent complex. There are advantages and limitations to any method of identifying and measuring PPIs, reviewed by Snider et al., 2015. Well-known drawbacks of some methods are the identification of proteins that interact in the experimental conditions but not in a biological context (false positives) or failing to identify known or probable interactions that are biologically significant (false negatives). To fully appreciate the range of PPIs that are possible within the predicted proteomes of several model organisms (Tran, Hamp & Rost, 2018), it is of interest to supplement the information on experimentally identified PPIs with PPI predictions (Humphreys et al., 2021).

PPI data is stored in repositories of various formats. The experimental results or computing methods used to identify or predict PPIs are diverse. In addition, the IDs and descriptions are not comparable from one database to another. To ensure easy access to the data and reliable outputs, the Human Proteome Organization (HUPO) initiative (Orchard & Hermjakob, 2008) and the International Molecular Consortium (IMEx) (Porras et al., 2020) have defined guidelines including accepted terminology and standardized data formats that should be used by authors reporting PPIs. Many curators have already adopted these principles for handling the data which is greatly facilitating exchanges and comparisons, although some disharmony still exists.

The Universal Protein Resource UniProt (Apweiler et al., 2004; The UniProt Consortium, 2018) is a collection of sequences with functional annotations and diverse information about each protein. The nomenclature and vocabulary are standardized, and various formats are available. This rich and user-friendly resource provides the reference proteome of species and several feature viewers that summarize and give access to data on localization, interactions, and molecular structures.

There is a large variety of biomolecular interaction databases. Some are specific to a particular type of interaction, others focus on a given organism type (fly, yeast, bacteria), or disease (Miryala, Anbarasu & Ramaiah, 2018). In this article, we will only consider some of the most frequently used PPI databases. The BioGRID database of physical, genetic, and chemical interactions reported in various organisms is updated monthly (Oughtred et al., 2019; Oughtred et al., 2021). The data can be downloaded in multiple formats, and more tools and resources are provided for analysis. The iRefIndex database (Razick, Magklaras & Donaldson, 2008) is a secondary database that collates non-redundant data on interactions from freely available sources. A confidence score is calculated for each accession. The open-source IntAct database provides interaction data derived from literature curation or direct submission as well as interactomes from different species or datasets. APID (Alonso-López et al., 2019) provides curated interactomes of 400 organisms based on PPI information from six primary databases of molecular interactions and experimentally resolved 3D structures. APID also includes a data visualization tool. APID’s user-friendly and intuitive interface can be used to look for physical, genetic, or predicted interactions alongside expression or localization data from an input set of genes of interest. The Proteomics Standard Initiative Common QUery InterfaCe (PSICQUIC) (Aranda et al., 2011) aggregates molecular interaction data from 23 servers. In the first PSICQUIC version, each PPI was described by 15 fields corresponding to PSI-MITAB2.5 format. Since this first version, other file formats give more information about the reported interactions and more facilities to the users. STRING, one of the most popular tools for representing PPI networks (Szklarczyk et al., 2019), aggregates details of experimental or predicted physical and functional interactions from other databases. In total, STRING provides protein interaction data with associated confidence scores from 5090 organisms. The network resulting from the user’s request can be easily exported in different formats of text and image files.

The analysis of protein interaction networks (PIN) is of great interest when studying biological activities, pathways, or drug targeting and is the reason why many web tools or plugins for visualization and analysis of protein interaction networks have been developed. For example, Pathguide (Bader, Cary & Sander, 2006) provides a list and brief description of 702 pathways and molecular interaction resources.

Cytoscape (Shannon et al., 2003) plugins have been developed to export PPI data and visualize PPI networks (Martin et al., 2010; Doncheva et al., 2019; Holmås et al., 2019; Legeay et al., 2020). Here, we will only present three of them because their functionalities are close to those of the package we developed. GeneMANIA (Warde-Farley et al., 2010) imports an interaction network from a list of genes with their annotations and putative functions. The interactions of the network are associations, i.e., the most closely related genes to a query gene set are identified using guilt-by-association. With this approach, new members of a pathway or a complex are found and weights are assigned to the interactions. From a combination of the most trusted datasets from UniProt, Intact, and other curated sources, BioGateway (Antezana et al., 2009) provides a network of interactions of different types, among which are PPIs annotated with GO terms. The interactome browser mentha (Calderone, Castagnoli & Cesareni, 2013) provides interactomes of eight model species based on PPI data from databases set up by the IMEx consortium.

Other related tools have been developed in the R programming language (Ihaka & Gentleman, 1996) to display the shortest paths of functional interaction between proteins and are provided by Bioconductor (Gentleman et al., 2004). The package Path2PPI (Philipp, Osiewacz & Koch, 2016) helps researchers find proteins and interactions of pathways or biological processes in fully sequenced organisms for which virtually no PPI is known. With the cisPath package (Wang et al., 2015), cloud users can integrate downloaded functional information on PPI from different online databases or private data to construct, visualize, manage, and share functional protein interaction networks.

In developing these tools, the respective authors carefully considered how to benefit the most from existing data when seeking to answer different biological questions. Depending on the topic, one tool may be better than another. In their article from 2016, Pan et al. reviewed the computational approaches to analyze PINs. While the above mentioned tools successfully integrate information on first-order neighbors in the network, they do not readily deal with all the experimental second-order PPIs involved in the biological process of interest. However, to find clusters in a PPI network, it is necessary to account for the second-order PPIs.

In the present article, we describe APPINetwork, an R package for constructing PPI networks to search for (1) sets of proteins involved in the same biological process and (2) proteins or protein sub-complexes that play a role in the assembly of a protein complex (Glatigny et al., 2017). Starting from an input set of proteins, APPINetwork builds the PPI networks of the first or second-order, by using PPIs derived from all the available PPI databases and potentially any privately held data. APPINetwork thus provides the most exhaustive possible network of PPIs, whether experimental or predicted. The fact that this network integrates public and private data, makes the package particularly useful for all research teams who have identified new PPIs, and for proteomics platform groups that have accumulated large datasets of unpublished PPIs. Through a user-friendly interface, APPINetwork allows users (1) to choose the specie the user is studying from among hundred species currently included, (2) to select the queried PPI databases including any proprietary data files, (3) to select the order of network desired (first or second order), and (4) select the analysis to perform. We first present how the package is implemented and then discuss the advantages of similar tools.

Materials and Methods

Implementation

All functions of the package APPINetwork are implemented in R except for the script to search for all proteins involved in a biological process (see Fig. 1) which was developed in C (Gambette & Guénoche, 2011), and the scripts that format the BioGRID data file (see Fig. 1) and generate the ID correspondence file (see Fig. 1) which were implemented in Python. Indeed, the use of Python is optimal for writing functions for text mining large text files such as BioGRID files (Oughtred et al., 2021).

Results

To start with APPINetwork

The graphical interface of the APPINetwork menu offers the choice between five actions: (i) construct a correspondence file between different IDs; (ii) format iRefIndex, IntAct, or BioGRID PPI files; (iii) build a network; (iv) identify proteins involved in a biological process, and (v) identify the assembly intermediates of a protein complex (see Fig. 2 and the user guide). When using APPINetwork for the first time, actions (i), (ii), and (iii) must be executed successively. Indeed, formatted PPI files are required to build the network for which a correspondence file with the different identifiers is necessary. On the second use, if the user is continuing to work on the same organism, it is not necessary to execute steps 1 and 2 again. Another network can be built from another input list or from the same input list but with different parameters. In the same way, any network built with another tool, if formatted as described in the user guide, can be analyzed with APPINetwork using actions (iv) or (v).

Building of the network

To build a network (see Fig. 3), the user must select the “input list” file that has been prepared in advance (see section “required files” in the section “Material and Methods”). Then the formatted PPI files are selected by clicking the “select database” button. The user must also select an ID correspondence file by clicking on the “ID correspondence file” button and choosing the organism. Finally, the user must decide whether the network should contain (a) PPIs determined experimentally by physical or genetic methods or both, (b) PPIs of the first or second order, (c) proteins interacting (whether first or second order PPI) with a single protein of the input list (termed “unique link”), (d) interactions of a protein with itself (see Fig. 2 and the user guide), and if necessary (e) indicate a maximum number of first-order protein partners.

The script looks for all the PPIs involving proteins of the input list (first-order PPI) inside the formatted files of PPIs. In case of discrepancy between gene names or protein IDs, the program sends a warning to the user, who can then manually correct them. It removes redundant PPIs and records the IDs of publications mentioning each PPI. It calculates the number of these publications and records it in the file because it is an index of the reliability of the interaction. Other information related to each PPI, such as the type of experiment, the biological function of each network protein the name of the organism is also stored in the file. The network of order one or two is saved as a flat file with 13 columns containing all PPIs of order one or two and related information (see Materials and Methods), that can be exported to other analysis tools.

APPINetwork thus gives a lot of information on PPIs and offers to build a particular second-order network, which other software does not offer. The downside is that it takes time. As an example of the time required, on a laptop computer with 32 GB of memory and an Intel Core I7, the computation time was about 7 min for a network built from a database of 683,389 PPIs, with a threshold of 300 for the maximum number of partners of each protein.

Discussion

APPINetwork is more than a PPI network building tool since it also offers two clustering methods to analyze the resulting PIN. It can therefore be used to search for proteins involved in a biological process of interest or to model the assembly of protein complexes by looking for clusters in PPI networks centered on the studied process. Second-order networks can be built and analyzed as well as first-order networks. APPINetwork provides information on PPIs in a large number of species or strains while other tools or databases are focused on a limited number of model species. Biologists working on lesser studied species or strains will therefore gain from using APPINetwork.

To remove protein data that may bias the clustering of a network, APPINetwork provides filtering options that are not offered by existing tools for building and analysis of PPI networks. Indeed, many proteins that are not specific to the biological process of interest are represented in second-order networks, while they interact with only one protein in the first-order network. If such proteins are not eliminated, the analysis tends to erroneously cluster proteins that have no biological relationship. Similarly, when looking for assembly intermediates of a protein complex, it can be useful to remove self-loops, because they are penalizing for the Jaccard index of dimeric proteins, which leads to assembly models where the monomeric proteins are assembled first.

APPINetwork removes inter species PPIs, which differentiates it from APID, PSICQUIC and mentha. This can be illustrated using the Elongator (Elp) complex of S. cerevisiae as an example. In the first-order networks built from the six subunits of the Elp complex with APID, PSICQUIC or mentha, one PPI is identified involving a human protein, namely the interaction between the protein ELP3 and the human histone H3.3. Notably, APPINetwork does not take interactions between a protein and another macromolecule into account. For example, according to the PSICQUIC network, the proteins ELP3 and IKI3 interact with the tRNA Glu UUC, but APPINetwork discounts these interactions.

APPINetwork merges all provided PPIs present in public and private PPI databases, so it builds a more complete network than other available tools. For example, by querying the BioGRID database, APPINetwork built a first-order PPI network of the ELP complex with more proteins related to the studied biological process than did APID, PSICQUIC or mentha. The additional identified proteins belong to transcription complexes TFIID, TFIIIB, SPT4-SPT5 and Facilitator of Chromatin Transcription (FACT) complex as well as the ribosome. None of these proteins are represented in the very small network obtained with STRING. Even when the STRING network is extended, it includes additional proteins with functions that do not seem to be coherent with those of the ELP complex proteins.

Using APPINetwork to integrate laboratory PPI data and PPI from public databases into the same network is particularly useful for analyzing the numerous PPIs identified by interatomic platforms. It will result in a more comprehensive network. An additional feature of APPINetwork is that the output contains information on the interactions of the network and the associated publication(s) describing them. This file is convenient for users because all known information on PPIs involving the proteins of interest is easily accessible.

An advantage of APPINetwork is that the user can build a PPI network with particular second-order PPIs, excluding more proteins that are unrelated to the biological process of interest. Moreover, as some proteins have a very many partners (several hundred), there is an option to filter out these partners which tend to strongly bias the clustering of the graph. The resulting clusters will thus be more relevant than when starting from a classical second-order network.

Finally, the first and second-order networks obtained with APPINetwork are provided in files that can be easily exported in other software like Cytoscape (Shannon et al., 2003) allowing them to be visualized and analyzed through other applications.

Conclusion

The APPINetwork package is a tool for PPI network building and analysis involving proteins from numerous biological processes in numerous species or strains. It offers users the choice of using public (experimental or predicted) PPI databases to build the PPI network and to add unpublished PPI data.

It has a user-friendly graphical interface allowing access to the different options for building a network suited to the type of analysis to be carried out. For example, a network built with genetic or predicted interactions, as well as unpublished interactions, could identify more PPIs involved in the studied biological process. A first-order network without self-loops could improve the likelihood of identifying assembly intermediates of a protein complex while a second-order network would identify sets of proteins involved in the same biological process. Other options of the interface allow to choose between the two types of analysis and modify their parameters.

APPINetwork provides the PPI network as a flat file containing the list of PPIs with various information about the interaction and the interacting proteins (PubMed IDs, experimental methods, all identifiers of involved proteins) that can be a very useful resource for biologists. It also provides a text file containing all proteins of each cluster identified by TFit and additional files containing results of the hierarchical clustering modeling the assembly of a complex.

Finally, the APPINetwork package can be freely downloaded from the GitHub repository and comes with a user guide and examples that facilitate its use.