CrosstalkNet: A Visualization Tool for Differential Co-expression Networks and Communities Free

Co-expression networks are a representation of putative gene–gene relationships, built by investigating the similarity of transcriptomic profiles across samples. During the last decade, co-expression networks have grown in popularity as they enable the integration of large-scale transcriptomic datasets, allowing identification and exploration of multiple genes with similar expression patterns across conditions (1–6). Given the relevance of these co-expression networks in cancer research, we developed CrosstalkNet, a web-based tool that enable users to explore and mine co-expression networks. We present here an application of CrosstalkNet in the context of tumor epithelial-stromal cross-talk, although the tool is widely applicable to the analysis of bipartite networks in general.

One of the most important signaling networks in tumor biology is the communication between tumor epithelial (epi) cells and neighboring stromal cells (7), which is mediated through regulatory loops between compartments. A promising approach to shed light on such a cellular cross-talk lies in using the transcriptional profiles of epithelial and stromal sample pairs obtained from laser capture micro-dissection (Supplementary Fig. S1) to construct co-expression networks (3). Given the computational and biological complexity of the co-expression network, it is challenging to explore and comprehend these large-scale networks. Epi–stromal co-expression networks can be represented as a bipartite network where each node represents a gene and each edge denotes a co-expression relationship between the epithelium and stroma (Fig. 1A). Differential networks have recently been used to characterize the rewiring of condition-specific interactions (8). We present here the first application of this methodology in the context of epi–stroma cross-talk to uncover the differential interactions gained and lost by tumor (Supplementary Fig. S2).

Most common visualization tool is Cytoscape, but, BiLayout is the only layout in Cytoscape that allows for visualizing bipartite networks. However, it requires manually selecting the nodes of interest as well as the neighbors to be plotted. In addition, it is difficult to visualize second neighbors. CrosstalkNet overcome the limitations the existing web-based visualization tools (Supplementary Information) to better explore large-scale bipartite networks Supplementary Video S1). We believe that this web-based application will assist researchers across disciplines in mining bipartite biological co-expression networks, as exemplified by our study of the epithelial–stromal cross-talk that occurs in breast tumors.

The datasets used in this study are presented in the Supplementary Table S1. The genome-wide tumor epi–stroma network was constructed by estimating pairwise co-expression using Pearson's correlation coefficient between the tumor epithelium and tumor stroma, whereas the normal network was constructed using the same approach based on normal epithelium and normal stroma. Each node represents a gene and co-expressions between gene_i in the epi cells with gene_i, gene_j, gene_k in the stromal cells are represented by edges, as seen in Fig. 1A and Supplementary Fig. S3. This approach is extended by statistically comparing the interactions inferred from tumor and matched normal samples (8) and are applied here for the first time to infer epi–stroma networks (Supplementary Information).

CrosstalkNet uses a node.js server that interacts with Rscript to execute user queries. Rscripts are responsible for interacting with the bipartite network files and determining neighbors and paths in those networks. Communication between node.js and Rscript happens via JavaScript Object Notation (JSON). Once the network files have been accessed, the results are sent back to the user and displayed via cytoscape.js, which is embedded in an AngularJS and Angular Material front-end.

CrosstalkNet is divided into five major panels, namely, Main Graph, Interaction Explorer, Path Explorer, Degree Explorer, Community Explorer (Fig. 1B). Each panel has its own modular workflow, and is shown in Fig. 1C. The functionalities of the panels are described below in detail. Case studies will be further presented to exemplify the use of each of the panels in the Supplementary Material.

Users have the ability to upload their own adjacency matrix files to the server. Uploaded files must have an Rdata extension and must contain a single dgCMatrix (sparse matrix) from the Matrix package saved using the saveRDS function. Furthermore, row names and column names have to match, and the file must not contain missing (NA) values. In addition, row and column names and either be gene symbols or entrez gene identifiers. Uploaded files will be private to the user that uploaded them and cannot be accessed by other users, and can be deleted at anytime.

The Main Graph is used to view the first and second neighbors of multiple genes of interest at the same time (See Supplementary Video S2). To prevent graphs from getting unmanageably large in size, a filtering option is provided in order to allow filtering of both first and second neighbors individually. This makes it possible to focus on the more relevant interactions. There is a table view showing the interactions between the genes in the corresponding graph, and the table is downloadable as a comma-separated value (CSV) file.

The interaction explorer enables this functionality by providing the ability to view arbitrarily many levels of neighbors (See Supplementary Video S3).

The path explorer makes it possible to obtain all paths with a maximum of one hop between two genes of interest (See Supplementary Video S4).

The degree explorer displays either the top genes based on their degree, or by showing genes having a degree greater or equal to a specified number, or using both the options (See Supplementary Video S5). It is capable of showing an arbitrarily large list of genes, which can be downloaded in a simple CSV format.

Communities of interactors hold the potential to reveal functional processes of a biological system. The community explorer functionality allows the user to upload communities of the co-expression network using their favorite community detection algorithms,visualize them as seen in Fig. 1D, and can be downloaded in a CSV format. See Supplementary Video S6 for a detailed tutorial.

Documentation along with a video tutorial for the web-application are provided within the tool (http://crosstalknet.pmgenomics.ca/app/#/documentation). The examples used in the web application, namely, normal, tumor and differential epi-stroma networks can be downloaded directly from the tool. These examples are provided to guest users as well as registered users. The methodology used to construct differential network as well as the web-based tool can be used to mine networks based on pan cancer transcriptomic profiles. The code is open-source and freely available from GitHub (github.com/bhklab/CrosstalkNet/).

In this study, we show how the CrosstalkNet web-application can be used to uncover the epi–stroma co-dependencies in ER⁺ breast cancer patients. The details of the network topology are provided in the Supplementary Text. The degree distributions of epi, stroma genes in the tumor network is presented in Supplementary Fig. S4A and S4B, respectively. The number of interactions with respect to their significance (P value) is provided in Supplementary Fig. S5. We used a false discovery rate (FDR) of 15% to select the set of significant interactions. Concordant with our previous work (3), we found that the proportion of self-loops to be significantly higher in tumor compared to normal (Supplementary Fig. S6). CrosstalkNet enables (i) efficient exploration of a shared neighborhood (first neighbors) of genes of interest (Supplementary Fig. S7A); (ii) exploration of paracrine interactions (Supplementary Fig. S7B); (iii) identification of paths between any functionally similar genes (Supplementary Table S2); and (iv) identification of hubs as potential drug targets (Supplementary Table S3). We show hereafter how CrosstalkNet enables efficient exploration of gene interactions through two case studies describing how to mine a differential network, and extract the community structure from the tumor network.

To identify the epi–stroma interactions that are gained or lost in ER⁺ breast tumors, we used the Delta option in the main graph panel to explore a differential network. With the tumor-specific epi–stroma (differential) network selected, we explored the interactions involving S100A7, a protein in the S100 family known to be a key contributor to the onset of aggressive and invasive tumors with the help of the tumor stroma (9). We identified a large change in the co-expression of S100A7-epi and S100A7-stroma between normal and tumor networks (Δ = 0.99; Supplementary Fig. S7C), indicating that there is a strong gain of interaction in the tumor. This supports the fact that S100A7 is one of the key genes that is differentially co-expressed between normal epi-stroma and tumor epi–stroma, and contributes to tumorigenesis (9).

Biological networks are known to have specific topology and structure, consisting of nodes that interact with each other and govern specific cellular processes. These groups of nodes are termed a community in the biological network. We used the CONDOR method (10) to obtain communities in the tumor epi–stroma interaction network, resulting in 62 communities, which can be displayed using the Community Explorer (Supplementary Fig. S8A). Each community consists of both epi–genes and stroma–genes (Supplementary Fig. S8B). We then performed a pathway analysis on all the communities that have ≥5 using all the genes with Gene Ontology terms (11). A total of 705 pathways are overrepresented with an FDR < 5% (Supplementary Fig. S8C). Broadly, the enriched GO terms belong to cell cycle, DNA synthesis and integrity, metabolism, etc., which are known to regulate tumor progression, resistance in ER⁺ breast cancer patients (12–14).

The availability of high-throughput data along with the advancement of web-based technologies has led to the development of network visualization tools to mine large-scale interaction graphs. We developed CrosstalkNet, a user-friendly web-application that can be used to explore and mine interactions. The application gives users the ability to upload their customized files and, to search and highlight a gene of interest in a dense network, along with multiple layout options. Users can also explore their inferred bipartite networks by focusing on a specific subgraph, filtering the network based on interaction strength, investigating the possible paths between two genes, identifying the hubs along with identification of communities in the network.

In the present study, we have described the use of CrosstalkNet with biologically relevant case studies reporting relevant epi–stroma interactions in ER⁺ breast cancer. We showed how to mine interactions in differential co-expression networks and how to extract the community structure from the epi–stroma interaction networks. Overall, our case studies highlight the importance of differential networks and community structure in a biologically meaningful way. Moreover, these case studies (including the ones in Supplementary Information) reiterate the relevance and versatility of the CrosstalkNet web-application in cancer research.

In conclusion, we developed CrosstalkNet, a web-application enabling exploratory network analysis for researchers without requiring computational skills. CrosstalkNet will assist researchers across various disciplines along with the clinicians in mining complex networks to decipher novel biological processes in the tumor epithelial–stroma cross-talk, as well as in other studies of intercompartmental interactions.

No potential conflicts of interest were disclosed.

We thank Gary Bader and Max Franz for their help in using the cytoscape.js framework that has been used to develop the CrosstalkNet web-application. V.S.K. Manem was supported by the US National Institutes of Health (subaward #236136) and the Cancer Research Society (grant #21363). G.A. Adam was supported by the Canadian Institutes of Health Research. B. Haibe-Kains was supported by the Gattuso Slaight Personalized Cancer Medicine Fund at Princess Margaret Cancer Center, the Canadian Institutes of Health Research, the Natural Sciences and Engineering Research Council of Canada, the Ministry of Economic Development and Innovation/Ministry of Research & Innovation of Ontario, and a Stand Up To Cancer Canada—Canadian Cancer Society Breast Cancer Dream Team Research Funding, with supplementary support of the Ontario Institute for Cancer Research through funding provided by the Government of Ontario (Funding Award Number: SU2C-AACR-DT-18-15). Stand Up To Cancer Canada is a program of the Entertainment Industry Foundation Canada. Research funding is administered by the American Association for Cancer Research International—Canada, the Scientific Partner of SU2C Canada.