The ECS network presented here was constructed from a manually curated component set grounded in the study by Bernabò et al. (2013) and subsequently expanded using established databases. The network constructed in this study incorporates protein–protein and protein-chemical interactions retrieved from STRING and ChEMBL databases, respectively, building on the curated ECS components from Bernabò et al. (2013). While this approach provides biological relevance and conceptual consistency, it necessarily reflects the current state of ECS research. As with any literature-informed curation strategy, the resulting network may be influenced by publication bias, uneven experimental characterization, and database coverage (Wodak et al. 2013; Snider et al. 2015; Schaefer et al. 2015; Gillis et al. 2014; Prakash et al. 2007; Schnoes et al. 2013). Proteins and chemicals that are more extensively characterized may exhibit higher connectivity within the network, reflecting both their biological importance and the depth of experimental investigation available in current literature and databases (Wodak et al. 2013; Snider et al. 2015; Schaefer et al. 2015; Gillis et al. 2014; Schnoes et al. 2013). Accordingly, this network should not be considered a complete or exhaustive representation of the ECS, but rather a structured approximation based on currently available data. Nevertheless, the integrated dataset provides a sufficiently robust foundation for examining the organizational and topological properties of the ECS at a systems level.

Protein–protein interactions were identified among the ECS proteins based on the components from Bernabò et al. (2013) and expanded with STRING interactions filtered by a specified confidence threshold to prioritize higher-confidence associations. STRING aggregates evidence from curated experimental data, database annotations, co-expression analyses, text mining, and computational inference (Szklarczyk et al. 2023). Although applying a confidence cutoff reduces the inclusion of lower-support interactions, the resulting PPI layer may still contain indirect or functionally inferred associations that do not necessarily correspond to direct physical binding events (Wodak et al. 2013; Snider et al. 2015; Gillis et al. 2014). Thus, the PPI component of the network represents high-confidence literature-supported and computationally integrated relationships within the curated ECS protein set, rather than an exclusively structure-resolved interactome (Wodak et al. 2013; Snider et al. 2015; Gillis et al. 2014; Schnoes et al. 2013). Emerging large-scale and structure-informed interaction datasets, including 3D interaction mapping and deep learning–based PPI prediction frameworks (Baranwal et al. 2022; Wu et al. 2023a; Patel et al. 2017; Xiong et al. 2025; Fang et al. 2021), may further refine ECS network resolution in future work by distinguishing direct molecular interfaces and expanding interaction coverage.

Protein–chemical interactions in the present network reflect curated literature-supported associations rather than quantitative pharmacological parameters (Bernabò et al. 2013). Because ChEMBL integrates diverse assay types, including binding and functional studies conducted across varying experimental contexts, individual interactions may differ substantially in affinity, efficacy, and physiological relevance (Zdrazil et al. 2024). For example, ChEMBL captures interactions derived from assays such as ligand binding affinity measurements (e.g., Ki or Kd), functional activity assays (e.g., EC50 or IC50) and cellular response studies conducted under various experimental conditions. By representing these heterogeneous measurements as binary edges, the current framework emphasizes structural connectivity over biochemical precision. While this abstraction enables global topological analysis, it does not distinguish between high-affinity canonical interactions and weaker or context-dependent associations. Incorporating affinity-weighted edges or multilayer representations would allow future studies to more directly model pharmacological strength and signaling specificity (Zheng et al. 2019; Berenstein et al. 2016; Zhang et al. 2016; Jiang et al. 2022; Cheng et al. 2012).

In addition, all interactions were modeled as unweighted and non-directional edges. This abstraction enables global topological analyses but does not capture interaction strength, binding affinity, temporal dynamics, or signaling directionality (Bosque et al. 2014; Martínez et al. 2023; Gursoy et al. 2008; Wuchty 2006). Consequently, measures such as degree or betweenness centrality reflect structural prominence within the curated interaction framework rather than quantitative biochemical influence or causal regulatory flow. This simplification of interactions in the network limits the extent to which signaling dynamics, interaction strength, and directional effects can be inferred, and highlights the importance of integrating additional data sources to develop more detailed and context-specific ECS models. Future extensions incorporating weighted edges, context-specific interaction strengths, or directed signaling relationships could provide a more mechanistic representation of ECS dynamics (Berenstein et al. 2016; Zhang et al. 2016; Jiang et al. 2022; Cheng et al. 2012; Wuchty 2006; Nasser et al. 2011; Rush and Repsilber 2018; Chen et al. 2021; Li et al. 2024b). Despite these limitations, the current network offers a systematic and reproducible framework for interrogating ECS organization at a systems level.

Our network-based analysis demonstrates how centrality measures can systematically identify nodes with high topological prominence within the ECS interaction network. Distinct patterns emerged across various metrics. Betweenness and degree centrality highlighted both canonical and non-canonical proteins, including CB₁, TRPV1, PPARα, COX-2, FAAH, and DAGLα, indicating their roles as hubs or connectors within the network. In contrast, closeness and eigenvector centrality highlighted chemicals such as CBD, THCV, and CBDV, demonstrating their extensive network reach and connectivity (see Fig. 2b and Table S13). The consistent top rankings of CB₁, CB₂, and GPR55 across multiple centrality metrics reflect their central positions in the curated ECS framework rather than pre-assigned biological weighting. Collectively, these findings illustrate that ECS organization arises from a combination of highly connected receptors, metabolic enzymes, and ligands that integrate receptor-mediated, metabolic, and signaling pathways. By highlighting both canonical and non-canonical nodes, these centrality-based insights provide a framework for prioritizing experimental studies to understand network-level relationships, structural dependencies, and potential targets for pharmacological exploration.

Clustering analysis revealed a modular yet interconnected architecture of the ECS network (Fig. 2c, Figures S3–S4, Tables S6–S11). The network’s connectivity alone was sufficient to uncover biologically coherent functional groupings. Rather than being randomly organized, the network is divided into communities that correspond to key areas such as receptor signaling, endocannabinoid synthesis and degradation, transcriptional regulation, non-canonical signaling mediated by TRP channels, and inflammatory pathways derived from arachidonic acid. This organization highlights that the ECS is structured around functional modules, which remain interconnected through shared ligands, enzymes, and signaling intermediates (Lu and Mackie 2021; Sikyta et al. 2024; Barrie and Manolios 2019; Campolongo and Trezza 2012; Dörnyei et al. 2023; Bernabò et al. 2013; Pistis and Melis 2010; Rezende et al. 2023; Șerban et al. 2025). By delineating distinct yet interacting communities, this framework offers a structured basis for hypothesis-driven investigations and can serve as a guide for future studies focused on specific clusters and their constituent nodes when examining receptor regulation, metabolic flux, transcriptional control, or inflammatory crosstalk. Integrating contextual layers into future network models, such as tissue-specific expression patterns or condition-dependent interactions (Zheng et al. 2019; Berenstein et al. 2016; Zhang et al. 2016; Nasser et al. 2011; Rush and Repsilber 2018; Chen et al. 2021; Li et al. 2024b), could further refine these modules and clarify how ECS connectivity may vary across physiological and pathological states.

Our perturbation analysis indicates that removing highly connected nodes alters the topological structure of the ECS interaction network. Consistent with previous research describing the ECS as a scale-free network (Bernabò et al. 2013), a small subset of highly connected nodes exerts a disproportionate influence on overall connectivity. The targeted removal of these nodes redistributes centrality across the remaining network, increasing the relative importance of some components while decreasing that of others. For instance, after the removal of CB₁, receptors such as CB₂ and GPR55 showed increased centrality, whereas certain enzymes like DAGLβ and endocannabinoid-like lipids like LEA showed reduced positional influence. These changes reflect a reconfiguration of the shortest paths and connectivity patterns within the static interaction map, rather than a functional compensation.

Importantly, such perturbation-driven changes identify nodes whose structural roles are contingent upon the presence of dominant hubs and highlight components that assume greater topological prominence under altered configurations. While these results do not imply biological resilience or adaptive signaling responses, they provide a systematic framework for prioritizing hypothesis-driven investigations. Nodes that increase in centrality following hub removal, such as CB₂ and GPR55, may represent underexplored components whose influence is revealed under perturbed conditions. Conversely, nodes that exhibit substantial decreases in centrality, like DAGLβ and LEA, may reflect interaction patterns highly dependent on specific connectors. Together, these observations offer a network-based strategy for selecting candidate proteins or ligands for experimental validation, mechanistic study, or therapeutic exploration within the broader context of ECS-related signaling and cross-talk (Zheng et al. 2019; Nada et al. 2024; Carels et al. 2023; Ruffner et al. 2007; Tiwari and Dwivedi 2019).

GPR55 consistently ranked among the top nodes across multiple centrality measures, indicating strong structural prominence within the ECS interaction network. In the combined network comprising 80 nodes, GPR55 ranked among the top five nodes for betweenness (5th), closeness (1st), degree (5th), and eigenvector centrality (5th), exceeding the rankings of other non-canonical GPCRs such as GPR18 and GPR119, which were placed substantially lower across these measures (Table S14). These patterns were even more pronounced in the protein–protein network (40 nodes), where GPR55 ranked first in both closeness and eigenvector centrality and third in degree centrality (Table S14). Its normalized eigenvector centrality score approached 1.0, reflecting strong connectivity to other highly ranked nodes within the defined ECS network. In this context, high eigenvector centrality indicates that GPR55 is embedded within densely connected regions of the network, emphasizing its topological prominence.

Consistent with these structural observations, perturbation analysis showed that removal of CB₁ increased GPR55’s relative centrality ranking, indicating that its positional prominence becomes more pronounced under altered network configurations. These shifts reflect the redistribution of centrality within the static interaction map rather than evidence of biological compensation. Together, these findings identify GPR55 as a structurally central component within the curated ECS framework. Independent experimental studies have implicated GPR55 in cannabinoid-responsive signaling and diverse physiological processes such as synaptic signaling and inflammatory responses (Ryberg et al. 2007; Lauckner et al. 2008; Kargl et al. 2012). While the present analysis does not assign functional significance, the consistent topological prominence of GPR55 across network configurations supports its prioritization for further hypothesis-driven experimental investigation within ECS-related signaling contexts.

Although we constructed a systems-level network of the ECS by integrating experimentally validated interactions between proteins and chemical compounds, this network has inherent limitations that restrict functional interpretation. The network captures structural connectivity and organizational patterns, illustrating how key nodes and interactions shape network topology (Fig. 2a and b), how molecular components cluster into coherent communities (Fig. 2c), and how perturbation of specific nodes redistributes centrality and connectivity (Table 1, Figs. 3, 4, 5, 6). However, it does not incorporate interaction directionality, signaling dynamics, or context-dependent regulation, limiting insights to structural topology rather than functional activity.

Looking forward, this framework lays the groundwork for several avenues of future investigation and addresses key gaps in ECS research. Incorporating interaction directionality, experimentally measured interaction strengths, signaling dynamics, and tissue- or context-specific interactions (Zheng et al. 2019; Berenstein et al. 2016; Zhang et al. 2016; Jiang et al. 2022; Cheng et al. 2012; Wuchty 2006; Nasser et al. 2011; Rush and Repsilber 2018) would enable predictive modeling of ECS responses to endogenous ligands, phytocannabinoids, and synthetic compounds, capabilities that remain largely unexplored in current studies (Bernabò et al. 2013; Bernabò et al. 2014; Liu et al. 2024). Complementary network approaches, such as network proximity analysis, could quantify distances between ECS nodes and disease-associated or drug-target modules, revealing potential cross-talk with other signaling pathways and informing therapeutic strategies (Wu et al. 2023b; Han et al. 2024; Brohée et al. 2008; Cheng et al. 2018). Moreover, perturbation- and centrality-driven analyses can prioritize understudied or structurally sensitive nodes, providing a systematic framework for hypothesis-driven experiments to test their roles in ECS connectivity, signaling, and pharmacological modulation. By combining these approaches, future studies could substantially expand both the conceptual and practical impact of ECS network mapping, uncovering non-obvious regulatory components, identifying candidate targets for experimental validation, and guiding the design of interventions that leverage network topology for translational or therapeutic benefit.