While phenotypically diverse, many CVDs share pathophysiologic intermediaries. For example, atherosclerosis is a chronic inflammatory condition characterized by plaque buildup in the intima of blood vessels. This dominant cause is shared by various vascular diseases through mechanisms, such as (i) thickening of arterial walls in CAD and PAD, (ii) plaque rupture and thrombosis in its onset of acute complications (e.g., MI), and (iii) changes to aortic media causing aneurysms. However, components of the immune system underlie atherosclerosis development and aggravation. Dysregulation of innate immune systems may accelerate atherosclerosis through mechanisms such as impaired efferocytosis, sustained macrophage activation, and activation of the NLRP3 inflammasome (11, 12). Similarly, adaptive immune mechanisms may promote atherosclerosis through interferon-γ, tumor necrosis factor-α, and interleukin-17 (IL-17) (13, 14). While inflammation appears to promote several atherosclerotic vascular diseases, including CAD, accumulating evidence suggests various CVDs share inflammation as a common pathology (Figure 1). Such shared pathologies motivate a systems-based, holistic view of CVD research, along with independent studies of different diseases.

Despite shared pathologies, individual CVDs that affect different organs and/or tissues naturally display unique sets of symptoms and outcomes. Within each CVD, increasing evidence suggests heterogeneity in clinical manifestations and underlying mechanisms. Diverse clinical presentations of CAD, including asymptotic stenosis due to atheroma, angina pectoris due to thrombosis, and ST-segment elevation MI (STEMI) vs. non-STEMI (15), are some examples of phenotypic heterogeneity in disease expression. Heterogeneity in hypertrophic cardiomyopathy (HCM) in patients with diverse sarcomere-independent morphological features (16) is another example.

Disease subtyping or phenotyping, as per nosology, derives from clinical presentations or other observable characteristics without suggesting underlying mechanisms. It is done by clustering subjects with similar observable clinical characteristics. Such phenotypic assessments (i.e., identifying disease attributes that describe clinical differences in presentation of diseases) helps refine diagnoses and determine treatment strategy. Akin to diverse clinical presentation of the same disease in different patients, heterogeneity also manifests in patients’ drug response, such as seen in HF with either preserved (HFpEF) or reduced (HFrEF) ejection fraction. Traditionally viewed as different stages of the same disease, evidence now indicates that these conditions may be different disorders, with different etiologies and responses to therapy. While they share many common risk factors (e.g., obesity, hypertension, and diabetes) (17), many drug trials failed to demonstrate a similar efficacy in HFpEF to what can be achieved in HFrEF. This discrepancy highlights the importance of considering pathophysiology, not just clinical presentation, in disease classification and the subsequent design of clinical trials (18).

However, disease endotyping is defined by distinct functions and pathobiological mechanisms as well as altered molecular pathways. Endotyping improves patient stratification, leading to more accurate diagnoses and tailored therapeutic strategies. Proper endotyping requires a framework that can integrate pathways and mechanisms with phenotypic features, molecular measurements, and demographic data (19–21).

Only a few CVDs are monogenic, including hypercholesterolemia, either familial (mutant gene: LDLR) or autosomal recessive (mutant gene: ARH); sitosterolemia (mutant genes: ABCG5/8) (22); and single mutation disorders such as Marfan syndrome, Loeys-Dietz syndrome, vascular Ehlers-Danlos syndrome, to name but a few (23). The pathogenesis of most CVDs is multifactorial and involves many complex genetical risk factors, including multiple pathogenic genes, hundreds of single nucleotide polymorphisms (SNPs), copy number variations, and genetic loci (24). Moreover, the combinatorial effect of multiple SNPs imparting a strong heritable component of CVDs indicates the presence of multiple overlapping pathological mechanisms that remain unidentified. In the context of CVDs (25, 26), genome-wide association studies (GWAS) indicate that identifying missing heritability may require understanding and functional characterization of risk factors and other intermediate traits leading to disease development and progression (27).

Cardiovascular diseases involve not only genetic, but also various environmental and lifestyle factors. These causal factors determine individuals’ CAD risk and cause multilayer heterogeneity in CVD patients. Along with genetics, data from other omics modalities (e.g., transcriptomics, proteomics, metabolomics, and epigenetics) provide a way to examine the disruptions and dysregulation caused by disease. While each type of omics data provides helpful information about changes in molecular composition or chromatin states, integrating them into unified models can give valuable insights into mechanisms underpinning disease pathology (28). The success of precision medicine in CVDs requires (i) consideration of complex overlaps of phenotypes and shared pathologies between different CVDs; (ii) the presence of endotypes within the same disease; (iii) accounting for heritability through intermediate risk factors; and (iv) integration of different omics data. Such an approach should also be able to handle big biomedical data, incorporate interdisciplinary research methodologies, and provide a framework to integrate multidimensional information about the disease.

Network medicine provides a framework that considers different components and constituents in the system and integrates them in an architecture that unravels the disease etiology, facilitates drug target discovery, and enables proper disease endotyping (21, 29).

The advent of high throughput technologies, the decreasing cost of sequencing, and willingness in public to allow the use of biomedical data for research through biobanks and consortia has yielded a vast amount of data from different omics platforms as well as clinical and phenotypic information. Network medicine effectively integrates information from multiple omics data types and extracts meaningful information and mechanistic understanding from different omics layers (e.g., transcriptomics, proteomics, genetics, and phenomics) (Figure 3). Several challenges remain in leveraging these different omics datatypes, especially during integration (e.g., data harmonization, differences in scaling and normalization, matching genes to proteins and metabolites, and different batch effects in different omics layers). Krassowski et al. discuss major considerations during data integration (36). Joshi et al. provide a detailed discussion about each omics type in the context of CVD (37). Use of networks to analyze biomedical data from different omics platforms can be classified in two ways. One approach involves obtaining information from different omics layer separately and then assembling it. Because different biological components work together, information from each omics layers can provide a facet of the system under study. In this approach, separate networks are constructed for each omics layer using a variety of different algorithms, depending upon the omics type, to establish a relationship among the features. After analyzing the networks to identify important genes, proteins, or epigenetic marks relevant to the disease in each omics modality, the overlapping entities can be used to identify important biomolecules relevant to the underlying biology (Figure 3A). We can also identify relationships between different omics layers using this approach (38).

Another approach to network medicine-based multiomics integration involves simultaneously analyzing different components from different omics layers by incorporating them into unified models. Networks constructed with this approach may have different types of nodes and the nature of edges between different node types may differ. For example, genes, proteins, or epigenetic marks may be connected based on biological principals, signaling pathways, and central dogma, along with correlated abundance levels. The edges may be present among same or different node types such as protein–protein, protein–DNA, or mRNA–protein. Some network methods use mathematical inferences as an additional step to give weights to the interactions (39). This approach is akin to a complex electric circuit, where different components exchange information to construct a more coherent output that informs the system under study, in this case underlying biology of the system (Figure 3B). It also reflects the complexity of multiomics integration, where the function of a complete circuit cannot be inferred from its individual components. However, unlike an electric circuit, this approach does not need information from all omics sources; available omics data can be built into coherent models to probe the system. Examples of this approach include Passing Attributes between Networks for Data Assimilation (PANDA) (40), Similarity Network Fusion (SNF) (41), and PAthway Recognition Algorithm using Data Integration on Genomic Models (PARADIGM) (42), among others (39). The concept of multiplex or interconnected networks also falls within this category (43). Mishra et al. provide a detailed list of variety of methods and tools for integrative omics (44).

Network medicine, artificial intelligence (AI), and/or machine learning are at the core of systems biology, which aims to identify the molecular determinants of disease. Using various classifiers, machine learning—supervised or unsupervised—is a major tool for identifying data patterns. Advanced models of machine learning include deep learning models that allow construction of various architectures of neural networks and innovations in loss functions, consequently performing different tasks in a variety of applications. These learning-based approaches are valuable when identifying diagnostic, predictive, prognostic, or therapeutic signatures of disease, including biomarkers. Despite many parallels (e.g., unsupervised machine learning as modularity detection in networks), significant differences exist between networks- and learning-based approaches; primary among them is their black-box nature and lack of interpretability of the latter. In contrast to linear regression and decision trees, which are easier to understand and have fewer parameters, deep learning algorithms base their decision on intricate neural networks with a huge number of weight parameters. Even knowing all the weights does not allow comprehension of a model’s behavior, leading to its black-box nature.

One of the most lucrative outcomes of biomedical big data analysis is the use of multiomics approach to identify new molecular or clinical biomarkers of disease. Such biological signatures could include disease-associated proteins, mutations, deletions, and copy number variation of genes identified by various omics platforms. Biomarker discovery has benefitted greatly from the advent of AI, especially its ability to handle large multivariate and unstructured clinical data (e.g., electronic health records and data from imaging modalities such as chest computer tomography and X-rays). These types of data are not easily amenable for network medicine approaches due to their size, complexity, heterogeneity, incompleteness, and unstructured nature. However, the strength of network medicine lies in leveraging omics data to find important biomolecules that serve as “network biomarkers” (45). Such biomarkers are often identified using genes responsible for network topology changes due to disease. For example, biomarkers for major adverse cardiac events were identified using protein interactions and signaling pathways (46). Other advances include using network-based models to identify blood-based and circulating biomarkers (47–49). The following sections discuss some examples in which network medicine approaches identify CVD biomarkers. Similarly, AI-based biomarker discovery using deep phenotyping with multiomics and analysis of digital electrocardiogram and data from wearable devices shows promising results in HF (50), left ventricular systolic dysfunction (51, 52), and arrhythmia (53). Although reviewing specific applications of AI in cardiovascular contexts exceeds the scope of this review, several interesting review articles address this area (54–58). Giordano and colleagues discuss the efficacy and application of machine learning and AI in clinical decision making when developing personalized models of patient care (59).

Network medicine has grown tremendously over the years, producing many successful applications across various disease areas and pathological landscapes. This growth spurred several excellent review articles and books that discuss its different aspects (30, 60–62). Notable among these is a series of articles by Loscalzo and colleagues and other torchbearers for network medicine (29, 63, 64) who have discussed different facets of network medicine approaches that focus on specific disease areas, including pulmonology (60), pathobiology (65), cardiology (66, 67), and coronavirus disease 2019 (68–70). These reviews discuss various concepts, applications, applicability to specific problems, practices, pitfalls, and above all, the promises of network medicine.

Many other interesting reviews and books discuss various aspects of network medicine in general and in the context of CVD. While some focused on discussing different omics technologies and how network methods can be leveraged to analyze them, others discussed individual network projects in detail or the philosophical underpinnings of the field, providing a context for understanding the contributions of network medicine to biomedical advances. The present review article is at the cross-roads of these ideas. We do not provide a chart of network methods, but rather try to inculcate an understanding that will empower researchers to select a specific network medicine approach, depending on data type, disease context, and outcome expectation. For this purpose, we rely on our previous review article (71), which discusses network medicine in the context of analyzing biomedical big data from different omics platforms. We also classified network medicine research into three different paradigms: (i) an interactome of protein–protein interactions (PPIs) with other omics data to elucidate disease context, (ii) pattern analysis in the co-abundance of measured analytes in disease, and (iii) deciphering gene regulation principles through phenotype-specific networks (Figure 4). Here, we use that classification template to discuss the application of network medicine and review several examples and applications in CVD contexts. While these paradigms provide a bird’s eye view of the field, we will cover CVD applications where PPIs and co-expression networks are used together.

Large-scale efforts to collect data have yielded far-reaching effects on biomedical research. For example, the Human Genome Project mapped the human genome and the Encyclopedia of DNA Elements (ENCODE) (72) provided functional elements of the genome. Consortia such as the Human Protein Atlas (HPA) and Genotype-Tissue Expression (GTEx) provided detailed atlases of human proteome and transcriptome in different tissues, respectively. Data from these efforts provided a comprehensive “parts-list” of cellular machinery that approaches like network medicine can assemble. Similarly, high-throughput mapping of protein—protein interactions (PPIs) is a significant milestone in biomedical research (72–77). PPIs are networks with proteins as nodes and interactions between them as connections. PPIs are a collection of interactions between proteins measured in various ways, including a yeast-two-hybrid system (78, 79), co-immunoprecipitation followed by mass spectrometry (80, 81), literature curation (82), or collection of various evidences (83–87). Primarily, PPIs are undirected because the direction of interactions between proteins is undefined. Previously, paradigm 1 of our earlier review article discussed different features of PPIs in detail (Figure 4A). Here, we recap some important features and mention the studies that have implemented them in the cardiovascular context. We do not intend to provide an exhaustive compendium of all the studies, but rather intend to discuss a few exemplar research studies to elucidate the efficacy of the approach.

As discussed previously, phenotypic context does not come naturally to the interactome or PPI because they are a static map of protein interactions. Disease context must be embedded onto them through seed genes or proteins. By design, co-expression-based networks are built by using measured analytes in the disease or phenotypic context. In our previous review, paradigm 2 discussed different ways to leverage co-abundance information to construct co-expression networks that use a variety of network methods and correlation measures (Figure 4B). Identifying biomolecules expressed at similar levels can provide insight on their cooperative function or how they co-regulate each other or other constituents of the biological process (134–136). Among many methods that calculate co-expression networks, weighted gene correlation network analysis (WGCNA) remains among the most widely used methods in cardiovascular medicine. WGCNA builds co-expression networks by calculating the correlation between genes and organizes the network based on identified highly correlated clusters of genes. After identifying the “eigengene” (i.e., the first principal component of the expression matrix of the corresponding module) of each cluster as representative of that cluster, WGCNA determines its associations with disease or other clinical phenotypes, thus revealing which cluster is most informative of the given phenotype. Numerous applications of this method have been applied in a variety of CVD areas. For cardiac hypertrophy and failure, co-expression module analysis identified ZIC2 as a regulator of failing myocardium (137). Another study investigated cardiac hypertrophy by using co-expression analysis and micro-RNA (miRNA) data to identify the involvement of PPARα in sex differences (138). Regarding septic cardiomyopathy, WGCNA identified the critical role of various genes (139) in obesity-related cardiomyopathy (140), hypoplastic left heart syndrome (141), viral myocarditis (142), and ischemia-reperfusion in patients undergoing coronary artery bypass graft surgery (143). Some studies of pulmonary arterial hypertension imparted critical understanding of its pathogenesis by using gene co-expression networks (144–153). Analysis of transcriptomics networks in nearly 3,600 individuals with blood pressure measurement identified SH2B3 as a key regulator of blood pressure (154). In a major investigation, Bertero and colleagues used network analysis to implicate miRNA-130/301, which targets PPARγ in different cell types, in the development of pulmonary hypertension (155). In the context of hypertension, a comparison of co-expression hub genes and miRNA revealed important associations in the interaction between gene co-expression and miRNA (144). Others applied WGCNA analysis to diseases such as non-syndromic thoracic aortic aneurysm (156) and acute aortic dissection (157) using module trait association analysis. Genotyping and transcriptomics data from the MAGnet consortium helped Cordero and colleagues determine that PPP1R3A regulates HF pathology (158). Furthermore, a co-expression network investigation revealed central transcriptional regulators in macrophage activation under different stimuli, which is an important research problem in atherosclerosis development and progression (159).

Because co-expression networks themselves are symmetrical (i.e., do not include directionality between interactions), directionality pertaining to regulator and effector nodes require inclusion by design. Gene regulatory networks (GRN) comprise interactions between effector nodes, such as TFs, other signaling proteins, and receiver nodes that are genes (Figure 4C). While there is no consensus on the proper definition of GRN, most networks that include co-expression between genes use this broad terminology. Gene regulation studies primarily investigate the regulatory potential of TFs in cellular, tissue, or disease contexts. Many algorithms have been proposed to leverage different omics datatypes, primarily transcriptomics and transcription factor binding, to infer gene regulatory networks. To establish directionality between gene–gene interactions, suggested strategies include model-based approaches like Boolean networks, Bayesian network models; similarity-based approaches like linear and non-linear co-expression; and conditional mutual-information network models. Most apply to bulk data such as the Algorithm for the Reconstruction of Accurate Cellular Networks (ARACNE) (190) and context likelihood and relatedness (191), but some methods include single-cell data such as the SCENIC method (192), the GENIE3 algorithm, and the SCODE method (193). Methods, including PANDA (40), that explicitly model the role of transcription factor to infer their regulatory gene targets using message passing between co-expression and the interactome are useful across many diseases (194–197), sex-specificity (198), and the regulatory architecture of healthy human tissues (199). This message-passing approach now includes a variety of multiomics data types such as miRNA in PANDA using MicroRNA Associations (PUMA) (200), chromatin accessibility in Seeding PANDA Interactions to Derive Epigenetic Regulation (SPIDER) (201), and genotypes in Estimating the Genetic Regulatory Effect on TFs (EGRET) (202). In the context of CVD, an investigation of mitral valve disease that used PANDA to explore the role of serotonin 5-HT dysregulation on mitral valve interstitial cells revealed increased expression of 5-HTR2B and 5-HT receptor signaling (203).

Regulatory networks centered on transcription factors have played an important role in CVD research. Macrophage activation has an important implication in atherosclerosis. Network analysis of macrophage activation in response to bacterial lipopolysaccharide (LPS) revealed the role of activating transcription factor 3 and NRF-2 during inflammatory response (204). Using GRNs to integrate gene expression, histone modifications in promoters and enhancers, and TF motifs in different human tissues, Schmidt and colleagues showed that expression levels of relevant genes correlated with chromatin accessibility in different tissues (205). Another study used GRNs to explore the role of Toll-like receptors in transcriptional control of macrophage activation (206). A similar approach was considered for studying epigenetic architecture, histone modifications, and gene expression in pulmonary arterial hypertension (207). GRN analysis has aided the study of atherosclerotic plaques (208, 209) and heart failure (158). Moreover, differential network analyses are useful tools when describing the sources of dysregulation between two phenotypes (31, 210) (Figure 4C).

An emerging target of CVD therapeutics is non-coding RNAs, especially due to their stability in plasma and confinement in extracellular vesicles. Additionally, miRNAs and long non-coding RNAs (lncRNAs) serve as biomarkers of CVD (211–213). Particularly, lncRNAs regulate macrophage functions and thus play a significant role in CVD pathophysiology (214). Some studies determined a regulatory relationship between different mRNAs and non-coding RNAs (215, 216). Network medicine can investigate such an interplay of non-coding regulatory network mechanisms (217).