The ImmunoGlobe immune interaction network codifies immune interactions described in Janeway’s Immunobiology
To construct a comprehensive immune interaction network, we manually curated the 2799 immune interactions (edges) published in Janeway’s Immunobiology [18], widely regarded as an essential and comprehensive immunology text [17]. The data in this textbook is derived directly from the research literature, and focuses on physiologic functioning of the immune system rather than rare or atypical phenomena that may result from some experimental setups.
Detailed information about 253 immune system components (nodes) and the nature of each directional interaction was recorded into a network table (Table S1). Nodes are general representations of each immune component and do not represent particular samples. For each interaction (edge), we extracted the names of the source and target nodes, the direction and type of the interaction, and the source of the data in the textbook (Fig. 1a). Additional information, such as the receptors involved, the activation states of the source and target nodes, and the immune process in which a given edge participates were recorded if available. This codification of the textbook was repeated twice and verified by an independent panel of reviewers.
A table (Table S2) designating node attributes was also generated to provide functional detail about each individual node. Each node was categorized into one of five types reflecting its identity: cell, cytokine, antibody, effector molecule, or antigen. A subtype was further assigned to reflect the function of each node. Of the 2799 interactions extracted (Table S1), 1112 were unique (Table S3). These interactions linked 253 nodes.
An example of the type of information used for construction of the network is presented in Fig. 1b. Analysis of this sentence reveals seven individual edges (interactions) between six distinct nodes (immune system components) (Fig. 1c), which were used to generate a graphical network (Fig. 1d). Although the amount of information provided by the sentence and the graphical network is identical, the graphical network formalizes the mechanistic relationships between the nodes, and enables the application of graph theory and network analysis principles to immunology.
The edge list and node attributes table were used to generate ImmunoGlobe, a graphical immune interaction network model (Fig. 2a). ImmunoGlobe was manually organized to group nodes according to function, with node type indicated by shape (Fig. 2b). Immune cells are at the top, organized according to the differentiation tree from a common hematopoietic stem cell [19]. Innate immune cells are on the left, and adaptive immune cells are on the right. Non-immune cells that interact with the immune system are collected in a column on the left. Cytokines are grouped together, separated into subgroups of interleukins, chemokines, and other cytokines. Immune effector molecules are grouped together and further clustered by subtype (e.g. Complement, reactive oxygen species). Antigens (foreign or pathogenic molecules that can stimulate an immune response) are shown at the bottom of the network. Antibody isotypes are shown on the right. Different edge types are represented by lines of different colors and styles, detailed in Fig. 2c. Edge types that are considered positive interactions (i.e., activate, recruit, or promote survival) are in green. Negative interactions (i.e., inhibit, kill) are in red. Secrete is in purple. Other edges (differentiate, polarize) are in grey. Definitions of the edge types can be found in Note S2. ImmunoGlobe thus provides a visual catalog of directional interactions between immune components and is available as an interactive network for download (File S1) and online at www.immunoglobe.org.
The immune network model recapitulates known features of the immune system
A high-level analysis of the ImmunoGlobe network confirms known features of the human immune system, providing confidence that the topology and characteristics of this network accurately reflect our prior knowledge of immune system functioning. Most of the nodes in the network are cytokines (n = 109), followed by cells (n = 51), effector molecules (n = 59), antigens of various types (n = 30), and antibodies (n = 4) (Fig. 2e). The immune interaction network is large with 253 components (nodes) and 1112 interactions (edges) but has a low density of 0.02, meaning that only 2% of all possible edges in the network actually exist (Fig. 2d). This low density reflects specificity in the action of immune components, as a single node with excessively high connectivity could lead to pathologic immune responses if it were to become dysfunctional [20]. The network diameter of 7 indicates that the longest path between any two nodes is 7 steps. The average path length of the network is 3.25: It takes on average 3.25 steps along existing directional edges (interactions) to connect any two randomly selected nodes. This is shorter than would be expected by a random graph (Fig. S3), indicating that the network structure allows the rapid dissemination of information across its components [21], which is critical in the timely initiation of immune responses [22] (Fig. 2d).
The most common edges in the immune network describe the effects of cytokines on cells. The second most frequent edge type is cells secreting cytokines, followed by direct cell to cell interactions. The final category captures all edges involving antibodies, effector molecules, and antigens (Fig. 2f). The “Other” category in Fig. 2f groups together interactions between immune cells and effector molecules, antigens, and antibodies. A visualization of the interactions between all node types shows that cells are involved in over half of the total edges (Fig. 2g).
The degree counts, which measure the number of edges a node has, recapitulate prior knowledge as well. The degree distribution of the immune network skews right (Fig. 2h), showing that most nodes have relatively low degree, although there are a number of highly connected nodes. We examined the degrees of cytokine nodes by plotting the number of connections in versus the number of connections out for each individual cytokine (Fig. 2i). The number of connections in, or the “in” degree, reflects how many cell types secrete that cytokine, and “out” degree reflect the nodes that the cytokine influences. Some cytokines have low degrees and thus are highly specific: These cytokines are either secreted by or affect few cell types, whereas others with high degrees are secreted by or act upon many types of cells. The cytokines with the highest degrees are those related to inflammation (e.g. IFNγ, TNFα) and immunosuppression (e.g. TGFβ, IL10), which are relatively nonspecific processes that require broad activity across multiple modules of the immune system [23]. These processes are both initiated by many cell types and affect many immune cell types.
We next examined the degrees of the cell nodes (Fig. 2j). Cells have the highest degree of all node types because their functions are versatile, and cells can have different (and sometimes even opposing) responses depending on their physiologic context [24]. Cells carry out these varying functions by interfacing with and producing different components of the immune system. Antigen-presenting cells (APCs; here referring to dendritic cells, as described in Note S1) both sense a wide range of inputs and express or secrete numerous immune cell effectors [25]. Myeloid cells (including granulocytes), whose primary responsibility is to sense and respond rapidly to threats from the environment, have high “in” degrees but lower “out” degrees, reflecting their limited effector mechanisms [26]. Lymphocytes, the main effectors of the adaptive immune system, have lower degrees than other immune cells, reflecting their specialized and antigen-specific functions [27]. Immune cell precursors have low “in” degrees and slightly higher “out” degrees, reflecting their sensing of specialized growth and differentiation signals and their subsequent differentiation into mature immune cell subsets [19].
ImmunoGlobe accurately represents multi-step immunologic mechanisms
One potential value of the ImmunoGlobe network lies in its capacity to uncover novel multi-step immune pathways. To test this, we performed two case studies of multi-step pathways assembled from individual network interactions to determine if there was evidence for them in the literature. Iwamoto et al. [28] reported that activation of monocyte-derived dendritic cells by TNFα and GMCSF influences their capacity to induce differentiation of CD4+ T cells into Th1 and Th17 cells (Fig. 3a). Although this particular pathway was not described in the textbook it exists in the network because the eleven cell types and cytokines involved exist as nodes in ImmunoGlobe, and 13 of the 14 interactions comprising it were described in other contexts in the textbook. Only one of the 14 interactions reported by these authors was absent in ImmunoGlobe (secretion of IL23 by monocytes). ImmunoGlobe also identifies several additional interactions between these nodes not reported in the Iwamoto paper. In the second study, Daftarian et al. [29] reported that IL10 secretion is enhanced in CD4+ T cells by the cytokines IL6 and IL12, and in monocytes by TNFα (Fig. 3b). In the ImmunoGlobe network, all edges described in the paper are present, along with additional interactions between the nodes not described in the paper. The abstracts for both papers are included in Note S3. Thus, ImmunoGlobe links interactions reported individually in the textbook into more extensive pathways supported by experimental evidence but not explicitly described in the source text. This illustrates the comprehensiveness of the network despite its being based on a single source text, and suggests that the network can be mined for previously unknown or unaccounted for interactions and pathways of interest.
Mouse and human immune systems differ largely in the properties of their respective immune system components
Next we used ImmunoGlobe to investigate whether differences between mouse and human immune systems are reflected in the immune network structure. Each mention of a difference between mouse and human immune components (including cells, proteins, or molecules) described in Janeway’s Immunobiology was classified into one of four categories (Table S4) and annotated with the nodes and immune processes affected. We classified differences in node properties into four categories (Fig. 4a). Category 1 differences are those in which the component is the same between mouse and human, but form, function, or copy number differs. Category 2 are different components that perform equivalent functions. Category 3 differences are those in which the components are identical, but their levels or expression patterns differ. Category 4 are components that have no equivalent in one of the species. The most common differences between mouse and human immune components were those in Category 1 (Fig. 4b), with Category 4 being the least common. This predominance of subtle differences between the species highlights the common origin of their immune systems [30]. Indeed, the Category 4 differences (CCL6, CCL9, CCL12, SAP, and dendritic epidermal T cells are found only in mice, Granulysin and MIC molecules are found only in humans) all affect innate immune functions such as inflammation and barrier immunity, likely reflecting the different evolutionary pressures encountered by each species since their divergence [31].
Figure 4c shows the distribution of species-specific differences across the immune network, with the specific nodes and immune processes affected detailed in Fig. 4d. The differences between human and mouse affect both the innate and adaptive arms of the immune system, as well as some effector molecules (defensins, granulysin, acute phase molecule SAP) and chemokines (CCL12, CCL8, and CCL9). There are several differences in components involved in antigen presentation, including in the sequences and structures of MHC/HLA molecules, T cell receptors, the structures of antibodies, and the ratios of antibody isotypes. The ratios of circulating immune cells as well as the specific surface markers of various immune cell types differ as well. Innate immune recognition differs in the Toll-like receptors, antimicrobial molecules and enzymes that exist in each species, as well as activation control of B and NK cells. The nodes with the largest number of species-specific differences are those that represent B cells and NK cells. For B cells, these differences include differences in the positioning and sequences of the genes encoding HLA molecules, the structures of the HLA molecules, the effect of cytokines such as IL7 and TSLP on developing B cells, the surface markers that differentiate B cells, the process of recombination of the B cell receptor, and the expression of Toll-like receptors on naïve B cells. For NK cells, the differences impact their role in innate immunity, particularly in antigen recognition and cytotoxicity.
We expected that there would be differences in network structures between mice and humans based on the difficulty in translating immunomodulatory therapies between the species, but instead found that the 59 differences related instead to properties of the nodes themselves, largely in what activates the different immune components and how they are activated. The edges between the nodes do not appear to differ. For example, while TLR expression can be found in B cells of both species, they are expressed in naïve B cells constitutively in mice but only after BCR stimulation in humans [32], and the MIC and KIR genes involved in NK activation in humans are not found in mice [18]. These changes affect the reactivity of immune components rather than their interactions with other parts of the immune system.
Immune network structure can be used to examine the network effects of immune stimuli
To demonstrate the potential application of ImmunoGlobe in helping to interpret experimental data, we performed a mass cytometry experiment to see whether we could use the immune network structure to identify a relationship between network characteristics and the strength of immune cell activation in response to stimuli. Briefly, spleens were harvested from 4 wild-type B6 mice, and whole splenocytes were incubated with LPS, TNFα, or IFNγ for 8 h, after which they were stained with a panel of antibodies that recognize phenotypic markers of major immune cell types as well as several markers known to shift in expression with activation (Fig. S1). We calculated a composite activation score for each combination of cell type and stimulus by finding the difference in average expression of each activation marker between stimulated and unstimulated, then summing across all activation markers for each cell type.
We hypothesized that activation scores would be highest for cell types directly activated by a given stimulus, with a decrease as the number of intermediates between the stimulus and cell type increased. Our findings broadly support this hypothesis (Fig. 5a). One notable exception is the low activation score of T cell subsets, which is likely because no antigen-specific stimuli or costimulatory signals were included in the experimental conditions.
With the exception of cells directly activated by a given stimulus, the distance (defined as the number of steps comprising the shortest path) between stimulus and cell was not correlated with activation score (Fig. S2). Rather, we found that for cells not directly activated by a stimulus, the number of shortest paths between a stimulus and cell type showed a positive correlation with that cell type’s activation score (Fig. 5b), with a Pearson’s correlation coefficient of 0.55 (p-value 0.007). To quantify how likely one was to observe a correlation coefficient of 0.55 or stronger at random, we performed a permutation test which gave an empirical p-value of 0.009. Eosinophils (dark green) and neutrophils (dark orange) are the best examples (Fig. 5b), with the strongest relationships between the number of shortest paths and activation score. Cell types directly activated by a stimulus did not follow this correlation as they were more strongly activated, which is expected given the direct nature of the interaction. These data therefore suggest that the strength of a cell’s response to a stimulus is dependent not just on its direct responsiveness to the stimulus, but also on the number of paths that exist between the stimulus and the cell. This finding held true for all three stimuli tested in this experiment (TNFα, LPS, and IFNγ) and demonstrate that the prediction of how strongly a given immune cell will respond to a stimulus can be informed by knowledge of its place in the immune network structure.