Proteins interact in the cell forming a complex ordered functional mesh. Some of these interactions are necessary for maintaining cell organization whereas others support the cell response to internal and external stimuli, and are often transient (Acuner Ozbabacan et al., 2011). A variety of approaches suitable for high throughput analysis have been used to reveal the physical contacts between proteins without informing on the dynamic of signal propagation (Xing et al., 2016). More than 400 K “physical interactions” between human proteins have been reported in the literature by using these methods and for 85% of the proteins in the human proteome we know at least one physical partner in public databases (Orchard et al., 2014; Oughtred et al., 2021). Physical interactions are symmetrical by nature and, having no directionality, are represented as “undirected” graphs (Figure 1A). Transient signaling interactions, on the other hand, are often short lived and as such may not be revealed by the methods developed for physical interactions. They are often causal as one of the partners, the regulator, causes a functionally relevant modification of the target protein. These latter types of interactions may be modeled in two ways that are often referred to as “process descriptions” (PD) and “activity-flow” (AF) (Figure 1) (Le Novère, 2015; Türei et al., 2016; Touré et al., 2020).

Let us consider, as an example, the experimental observation that the phosphatase PTPRJ (DEP1) binds to MAPK1 (ERK2) and inactivates it by removing a phosphate (Sacco et al., 2009). As shown in Figure 1B, an “undirected PPI” model represents this statement as a link between PTPRJ and MAPK1 that has no direction. The “activity-flow” (AF) model, on the other hand, renders this information as a binary interaction where the two proteins are connected by an edge that has direction from PTPRJ to MAPK1 and a sign which is graphically symbolized with a specific edge-form or color. This representation captures the evidence that PTPRJ is the regulator and MAPK1 the target and that this interaction has the consequence of inactivating MAPK1. AF models offer the advantage of being represented as a set of binary interactions in a signed directed graph which is more informative than an undirected graph. Finally, the PD model captures additional mechanistic details. In this representation the target entity, MAPK1 in our example, is split into two nodes representing the phosphorylated and unphosphorylated forms of the protein. The two forms are connected by a directed edge symbolizing the transition from one form to the other. The activity of the regulatory protein PTPRJ is represented as an edge promoting the removal of the phosphate from MAPK1. A limitation of this latter model is that the impact of the phosphorylation on the activation status of MAPK1 cannot be directly derived; it is only implicit as it can only be inferred from the reconstruction of the downstream chains of reactions.

The different representations serve different purposes and answer different questions. For instance, analysis of highly connected regions of an undirected protein interaction network may reveal the formation of macromolecular complexes (Wang et al., 2009; Havugimana et al., 2012). Similarly, the function of a protein that is trapped in a subnetwork formed by proteins that are annotated to a specific biological process may provide hints on its function (Oliver, 2000). On the other hand, process description and activity-flow networks are appropriate to sketch the information flow from a receptor sensing a stimulus to activation of a transcription factor driving phenotype modulation.

Another major difference between physical and causal interaction datasets is proteome coverage as the latter have significantly lower coverage. This is partly due to incomplete curation of reported experimental evidence and partly to the lack of appropriate high throughput experimental approaches to reveal causal interactions on a large scale. In addition, many resources annotating PPI have, in recent years, joined their efforts forming a consortium (Orchard et al., 2014; Porras et al., 2020) for distributing curation investment and using common standards and curation rules, whereas causal resources have not reached such an agreement yet. For these reasons, many of the network approaches presently rely on networks based on physical protein interactions (PPI) (Zhang and Itan, 2019).

Approaches based on networks assembled by using information on causal relationships, however, are gaining momentum as they provide information that can be relatively easily converted into Boolean or ordinary differential equation models thus enabling users to compute the behavior of a system in different conditions (Le Novère, 2015).

Although there is no strict separation between the experimental evidence that can be captured by the different models, it is crucial to understand the data structure adopted by each resource as analyses built on information extracted from distinct databases may lead to different biological conclusions (Mubeen et al., 2019).

Cell physiology is governed by a large connected network of physical and causal interactions. Nevertheless, biologists sometimes prefer to consider the cell model as an ensemble of unconnected pathways that, in a first approximation, function in isolation and do not crosstalk. However, this approximation neglects the effects of the cell network as a whole that may significantly affect the behavior of the pathway subnetworks. Although networks are useful abstractions, their functional integration into a cell model remains an important challenge. Capturing the experimental information for the assembly of protein interaction networks from primary literature data is an intimidating task. To assist scientists, over the past 20 years, a number of resources have set out to annotate an excerpt of the experimental facts related to protein interactions in structured formats in public repositories. However, different databases have been developed to serve different purposes, they adopt different curation policies and describe the same biological fact at different levels of abstraction and granularity.

We here focus on resources that capture causality, hereafter referred to as “causality resources” (Figure 2). Considering the chosen representation model, databases can be grouped into two broad classes (Figure 2, Supplementary Table 1): “interaction databases,” where relationships are integrated in a global network and “pathway databases,” where interactions are curated and displayed in the context of the pathway they participate in.

The three most popular pathway resources are KEGG, Reactome and WikiPathways (Kanehisa and Goto, 2000; Slenter et al., 2018; Jassal et al., 2020). Among the pathway databases those adopting AF as interaction model are KEGG, SPIKE (Paz et al., 2011) and CBN (Boué et al., 2015) (Supplementary Table 1). The signaling information annotated in these databases is reviewed by domain experts and covers more than 50% of the human proteome. Aside from their descriptive value in the representation of cell physiology, they have proven useful in the analysis and interpretation of -omics data when coupled with algorithmic approaches such as gene set enrichment analysis GSEA and signaling pathway impact analysis (SPIA) (Subramanian et al., 2005; Tarca et al., 2009; Sprent, 2011). However, they do not provide an integrated picture of cell functioning as interactions are accessible only in the context of pathways and miss to offer a holistic view. Another class of resources, including Cell Collective (Helikar et al., 2012), Biomodels (Malik-Sheriff et al., 2020), the GINsim repository (Naldi et al., 2009) and the PyBoolNet repository (Klarner et al., 2017) collect assembled logical models. Briefly, these resources store models curated and tested by different groups for specific projects. Users can download the models and adapt them to different purposes. However, the models do not necessarily follow a common annotation standard. As a consequence, the integration into larger models is often not straightforward.

A third class of resources, such as SIGNOR (Licata et al., 2020), SignaLink (Csabai et al., 2018), OmniPath (Türei et al., 2016; Ceccarelli et al., 2020) or PhosphoSitePlus (Hornbeck et al., 2019) annotate interactions without necessarily listing them as members of a pathway. We will refer to these with the generic term “causal interaction databases” (Figure 2). This organization of the interaction data, which is not pathway centric, allows users to assemble an integrated cell network where all pathways are connected, thereby allowing to monitor pathway crosstalk.

The conclusion of the analysis in the previous section is that, in order to increase coverage, users should consider collating datasets from different resources. However, in large curation efforts, in some instances, the same experimental evidence can lead different curators to different interpretations. In addition, experimental reports addressing the same biological question reach sometimes contrasting conclusions. Thus, it is not surprising to observe that a causal relation between protein A and protein B is annotated as activating by one database and inactivating by another. However, this represents a problem in the assembly of AF networks from an integrated dataset. To investigate how serious this issue was, we next assessed the fraction of causal relationships that are inconsistently annotated.

Ninety five percent of the edges that are curated by more than one database are consistently associated with either up- or down- regulation (Supplementary Table 2). About 3,200 interactions are annotated with the same consensus effect in at least three resources, thereby accounting for a high-confidence subset of causal interactions. Conversely, 5% of the pairs are associated with both up- and down- regulation in different databases. Besides trivial curation errors, some discrepancies might reflect differences in the annotation policies of the different primary resources. Alternatively, it could be the consequence of conflicting literature reports or complex effects of an interaction leading to clashing consequences on the target protein function. For instance, GSK3-mediated MAF phosphorylation leads both to transcriptional activation and to degradation of the target (Rocques et al., 2007).

To quantify this lack of consistency, we compared the datasets from the four primary repositories. For this analysis, we first filtered out from each dataset those pairs that in each database are annotated with both a positive and a negative effect as in the GSK3-MAF example mentioned earlier (internal “inconsistencies”). As shown in Figure 3C, the percentage of incongruent pairs between DBs is relatively small, SignaLink and SIGNOR are the two repositories showing the highest number (and percentage) of contradicting interaction annotation. This subset of conflicting pairs has already been discussed (Perfetto et al., 2016) and can be explained by the differences in annotation granularity adopted by the two resources. For instance, SIGNOR annotates the mechanisms (such as ubiquitination, phosphorylation, etc.) involved in the interaction and, when provided, also the modified residues, while SignaLink only provides information about the causal effect.

We next asked whether the combined causal information captured by the different primary resources is sufficiently complete to be used to assemble informative disease networks linking most of the genes that are found mutated in patients. We used the expert curated information of the Cancer Gene Census (Sondka et al., 2018) that annotates 389 cancer types with lists of genes observed to be significantly mutated in cancers. Similarly, we used the information collected by the DisGeNET resource (Piñero et al., 2020) to download lists of gene-disease associations (GDAs) for 4,713 polygenic diseases (DisGeNET score > 0.5). These lists have different sizes ranging from one up to 83 genes in the case of “Malignant neoplasm of breast.” We filtered the lists by selecting diseases with at least two genes annotated, 163 and 823 diseases in Cancer Gene Census and DisGeNET, respectively (Figure 4). These disease-gene lists were used to query the AF resources for interactions linking the disease genes. We also included in the network the proteins that by forming a bridge between the query proteins, allow to connect them. The rationale for inclusion of “bridge proteins” is further discussed in the next paragraph.

The results of the approach are shown in Figure 4 as violin plots illustrating the distribution in the number of edges in the networks assembled by this automatic procedure. As proteome coverage is far from being complete, not all disease gene lists could be connected to form a network in the different resources. Above each violin we have indicated the number of diseases for which it was possible to assemble a network by interrogating each of the resources together with the average network size (average number of edges). As a larger coverage corresponds to a higher number of connections, retrieving interactions from SIGNOR allowed the assembly of a higher number of disease networks (446 and 107 from the DisGeNET and Cancer Gene Census lists, respectively) in comparison with the other primary resources. Similarly, SIGNOR-derived networks tend to be larger, in terms of number of connections. OmniPath that integrates all the primary databases allows an even higher coverage (both in terms of number of diseases and in average network size). However, as already noted, by integrating the data of the four primary resources and OmniPath an even higher number of disease genes could be assembled into connected networks.

It is finally to note that among the resources compared here, only SIGNOR and OmniPath have implemented a web tool to extract connections between a list of input proteins and to return the results either in graph or table format. To apply a similar procedure to the dataset offered by the other databases dataset-manipulation and/or parsing is necessary.

As an example of the results that one obtains by the procedure detailed in the previous section we will describe in more detail the networks retrieved in the case of the Gray Platelet syndrome (GPS). GPS is a rare recessive autoimmune disorder characterized by a variety of symptoms including the absence of platelet alpha-granules, bleeding disorders and bone marrow fibrosis (Gunay-Aygun et al., 2010). NBEAL2 is the most frequently mutated gene in patients affected by this condition. However, due to the rarity of GPS, the molecular mechanisms underlying the disease are still poorly understood (Gunay-Aygun et al., 2011). We first assembled a list of 36 GPS associated genes and used this list to interrogate the different primary datasets and OmniPath (Figure 5A). As shown in Figure 5, in network assembly we also included “bridge proteins,” nodes that link two “disease proteins.” One advantage of using “bridge proteins” is that they allow for the expansion of the search space and for the retrieval of a graph connecting most of the disease proteins. By applying the aforementioned method, we succeeded in retrieving networks with a relevant (>2) number of interactions only by interrogating SIGNOR and OmniPath (Figures 5B,C).

By combining all the interactions, we obtain the most detailed graph incorporating 41 nodes and 91 edges and connecting 19 of the 36 input proteins (Figure 5D). As phenotypes are entities in the SIGNOR dataset, the integrated network also includes the “Platelet alpha granule formation” phenotype. Including phenotype entities improves the readability of the graph and strengthens the biological significance of the derived network.

GPS is a rare and poorly characterized disease. The advantage of this approach is that it compensates the lack of information annotated in the literature about pathways and perturbed molecular events.

To compare the results obtained in the case of GPS to that of a highly characterized disease, we applied a similar strategy to “Malignant neoplasm of breast.” This tumor type has the highest number of GDAs (83) in the DisGeNET resource. Not surprisingly, the retrieved networks are larger than the ones obtained for GPS, including 2,562, 533, 115, and 563 edges for SIGNOR, KEGG, SignaLink and PhosphoSitePlus respectively; 8,430 for OmniPath; and 11,513 for the five resources combined together. Such networks are extremely complex and difficult to interpret and might require stricter search parameters or filtering options that provide contextualization of the network (see next paragraphs).

These observations support the notion that there is no unique strategy to extract a diseases-PKN from AF repositories and the choice of a search method should be guided by quality and amount of information available for that specific pathology.

At the time of our survey only ~28% of the proteome is integrated into a global cell network by the information captured in AF repositories. This represents a severe limitation as for many disease-genes we do not have any clue about the functional consequences of modulating their activities. This can, to some extent, be addressed by increasing the curation effort and perhaps by establishing a collaborative consortium of resources similar to the IMEx consortium in the PPI domain (Porras et al., 2020). However, we also have to accept that for many proteins we have hardly any experimental evidence about their functions, let alone their causal connections with the activity of other proteins in the cell network.

The networks that are derived by the strategy that we have delineated here are highly connected and complex and as such sometimes difficult to understand and model. Some interactions that are not supported by thorough evidence and repeatability or are implausible can be removed after a detailed review of the model connections by a domain expert. However, the development of automatic pruning methods is also desirable. For instance, not all the causal edges are equally supported by experimental evidence. The SIGNOR resource assigns to each causal relationship a score that reflects its experimental support. This can be used to filter the models and delete the connections with little experimental support. However, causal relationships are likely to depend on biological context. Thus, the scoring system should be made context/tissue specific. The increasing availability of tissue specific proteomic and (single cell) transcriptomic data (Fagerberg et al., 2014; Uhlén et al., 2015; Fernandez et al., 2019) should make this possible in a reasonably near future. Computational optimization methods such as CellNetOpt (Terfve et al., 2012), PRUNET (Rodriguez et al., 2015) or MetaReg (Ulitsky et al., 2008) can be used to identify the causal connections that are important to adapt models to context by monitoring their ability to reproduce the response of different cell systems to perturbations.

As briefly discussed in this review, an AF network can be easily converted into simple Boolean models. This conversion process is set back by the observation that proteins in an AF network often receive multiple inputs from upstream proteins and these inputs govern the activity of a node as a function of the activity of the upstream nodes at each cycle of a simulation. To establish the logic functions determining node activity one needs information on how to combine these inputs. For instance, if both the kinase and the phosphatase modulating the phosphorylation state of a substrate site are active, will the substrate be phosphorylated or not? This information cannot be extracted easily from the limited available experimental evidence and approximate approaches are often used. For instance, an inhibitor win approach was often used with some success (Dorier et al., 2016; Palma et al., 2021). Alternatively, once a PKN model has been assembled the connections and the logic gates can be optimized from the ability of different models to reproduce results of perturbation experiments (Terfve et al., 2012). Developments of reasonably high-throughput experimental methods to address this limitation are highly needed.

These considerations underscore the present limits of the approach that we have discussed. Nevertheless, some initial successes in modeling clinically relevant phenotypes, as we have detailed in this review, and the delineation of a strategy to address the current limits provide confidence that cell/disease specific logic models should soon contribute to diagnosis and therapeutic decisions in clinical practice.