Clustering/dimensionality reduction-based approaches have the capacity to transform different data types into a common data space, thus facilitating downstream integration. This can be achieved through graph or kernel-based methods followed by grouping data features into a smaller number of variables. These approaches are the most straightforward methods to define biomarkers of disease or disease subtypes, thereby facilitating diagnosis and prognosis. The advantages of clustering/dimension reduction include the ability to retain within-data type properties and the robustness to different units of measurement. The drawback, however, is that the transformation of different data types may alter the underlying interaction between data types, even if within-data properties are retained (6).

Clustering-based approaches typically include hierarchical clustering (53), biclustering (54), and k-means clustering (21), which are used to find disease subpopulations (21, 55), refine disease characteristics, and help identify markers (56). Various methods such as iCluster (21), ICM (22), TMD (23), and others have been developed to use clustering for multidimensional integration (Table 1). For example, iCluster models the associations between different data types and the structure within each data type to bring the data onto the same feature space allowing for k-means clustering. This workflow has been applied on breast and lung cancer data sets to identify novel disease subtypes, which cannot be resolved using a single data type (21). We did not identify specific applications of multi-omics clustering in CVD research, although this type of approach has been applied based on individual data types (57–59). Future applications of such approach engaging multidimensional data will facilitate more accurate patient stratification based on multi-omics patterns and help identify unique biomarkers of CVD subtypes.

Dimensionality reduction can be achieved either intrinsically, which scales the dataset of interest using an analytical method, or extrinsically, which uses information outside of the dataset. Intrinsic approaches are the most widely used for dimensionality reduction of genomics data. Standard techniques include principle component analysis, factor analysis, multidimensional scaling, and others, which have been covered in a review of feature selection and extraction methods by Hira and Gillies (60). Tools utilizing dimensionality reduction techniques for multidimensional integration include CIA/MCIA (26), FALDA (27), and others (Table 1). Multifactorial dimensionality reduction has been applied by Badaruddoza et al. to identify environmental and genetic interactions in type 2 diabetes and CVD (61).

Predictive modeling is another powerful data-driven approach that is primarily utilized for the discovery of composite biomarkers in a multi-omics, big data landscape. In broad terms, it comprises a set of algorithms capable of learning from data to make predictions, which theoretically become more accurate with increasing amount of data. A series of machine learning techniques are commonly implemented, including logistic regression, support vector machines, random forest, neural nets, Bayesian models, and boosting (62) to select the most predictive features. This is typically done through weighting, where the most predictive features contribute more weight to the final model.

Among the various predictive modeling approaches used for multidimensional data integration (Table 1), an example is Causal Modelling with Expression Linkage for cOmplex Traits (Camelot) (29). Camelot implements elastic net regression to select the most significant features and uses bootstrapping to reduce the set of features to potential causal genes (29). There has also been widespread usage of machine learning methods in CVD-related fields to identify CVD risk variants (34) and estimate cardiometabolic risks (63). Specifically, Chen et al. trained an ensemble classifier to prioritize non-coding risk variants using multi-omics data and found that the variants associated with repressed chromatin were often the most informative (34). Kupusinac et al. leveraged artificial neural networks to predict cardiometabolic risk using easy to obtain, non-invasive primary risk factors and achieved comparable performance to predictions based on more invasive secondary risk factors (63).

As discussed previously, there are intrinsic biological relationships between data dimensions that can inform on mechanisms, and quantitatively assessing the association between the omics domains can help capture such relationships in a data-driven manner. Pairwise omics data integration is therefore an intuitive and commonly used approach that characterizes interactions between two omics domains. This type of integration comes in two broad categories based on whether genetic information is under consideration (Figure 1). The first category is genetics of intermediate traits analysis, in which DNA variants are tested for association with downstream omics markers. The second category is correlation analysis between two non-genetic omics data types (e.g., between metabolites and microbiome).

For genetics of intermediate trait analysis, expression quantitative trait loci (eQTLs) are the most well-known pairwise integration where genetic variations are linked to transcriptomic alterations, achieved through an association test between variants and gene expression levels (64). There are numerous methods available to conduct eQTL analyses such as GEMMA (38) and Matrix eQTL (40), which have been discussed in detail elsewhere (65). Genetic loci can also be associated with omics data types other than transcriptomics, such as methylation quantitative trait loci (66), microRNA QTLs (miR-eQTLs) (67, 68), protein quantitative trait loci (69–71), metabolite quantitative trait loci (72–74), and microbiome quantitative trait loci (75). Correlations between downstream omics data are also informative, although it may be difficult to infer a causal relationship. For example, expression quantitative trait methylation has been defined as the correlation of CpG methylation levels to gene expression (66). This type of analysis can be extended to the other omics data types (e.g., between microbiome and metabolome).

The combination of genetics-based and non-genetic correlative analyses can help infer causality. This concept has been widely used in CVD research to infer candidate causal genes (12, 76–80). Schadt et al. (81) were among the first to develop a formal procedure to incorporate eQTLs, genetic disease association, and gene–trait correlation to infer disease causal genes. Yang et al. (76) applied this approach to identify tissue-specific causal genes for atherosclerotic lesions. Laurila et al. applied a combined approach using both eQTLs and pathway analysis to link genomics, adipose transcriptomics, and lipidomic profiling, highlighting a shift toward inflammatory HDLs in individuals with low HDL (82). Huan et al. (83) combined eQTLs, miRNA-eQTLs, correlative analysis between gene expression and microRNAs, and GWAS to identify microRNA–gene pairs that are putatively causal for CVD. In another effort toward this direction, Zhu et al. proposed a summary data-based Mendelian randomization method that integrates diverse types of QTLs with GWAS to infer candidate genes for complex traits (41).

Network approaches have emerged as another powerful platform for multidimensional data integration. Networks depict omics markers as nodes and connections between markers as edges that reflect correlations, regulatory relations, or physical interactions. There are many types of network inference approaches, including regression, mutual information, correlation, and Bayesian networks (44) (Table 1). Among the widely used network methodologies, particularly in the CVD field, are correlation-based methods such as the weighted gene coexpression network analysis (42). These approaches primarily focus on gene expression data and use correlation patterns to group functionally related genes into modules, which significantly reduce the complexity of overlaying other types of omics data onto transcriptomics. It is also feasible to apply these coexpression network approaches to other types of omics data (e.g., DNA methylation data).

In network-based applications, different data types are typically mapped to features (e.g., genes) that can be projected onto networks. For example, Huan et al. integrated coexpression networks with genetic variants to identify causal functional modules for coronary heart disease (78). Yao et al. built an eQTL coexpression network to reveal CVD-related modules (84). Shang et al. inferred a transcription factor regulatory network from blood macrophages transcriptomics profiles and identified a key driver, LIM domain binding 2, for atherogenesis (85). Public network depositories such as protein–protein interaction (86) and BioGRID (87) have also been used to identify novel candidate CVD genes from diverse datasets (88, 89). Recently, the Björkegren group integrated Bayesian networks with CAD genetics and transcriptomics data from CVD relevant tissue types and identified CVD-causal subnetworks and key drivers (80).

Many of the available tools and methods applied to better understand the etiology of a complex disease like CVD utilize combinations of the various principles discussed above (Table 1). The integration of the various methods and data types is typically done in a sequential manner where a common overlapping feature (e.g., genes) is used to convert the output of one part of the analysis to be a compatible input for the next step. One example is the Analysis Tool for Heritable and Environmental Network Associations (50, 51), which utilizes neural nets and has been previously applied to predict HDL cholesterol (90). Specifically, this method generates a separate neural net model for each individual data type, and the features with the top predictive power from each model are combined in an integrative model, which possesses higher predictive power than any of the individual models (91). An alternative approach is to employ a majority voting scheme from each of the independent models from the individual omics types, thereby avoiding the additional step of merging multiple models but still leveraging information from multiple data types to predict a clinical outcome (92). As another example, Inouye et al. constructed metabolic networks where metabolites were identified to be associated with the genes identified in the eQTL analysis, thereby layering an additional data modality. The expression levels of the prioritized candidate genes were found to be associated with the phenotypes of the disease, demonstrating the effectiveness of this integrative method (15). Our lab has recently developed a highly generalizable analytical framework, named Mergeomics (3, 52), to more effectively incorporate multidimensional data and various integration strategies. Mergeomics can reveal pathogenic processes underlying diseases by interrogating enrichment patterns from diverse omics association data, and then leverage tissue-specific networks to identify key perturbation points of the significant processes. With this approach, we have prioritized novel regulatory genes and therapeutic targets for CAD and hypertension from diverse genomics, transcriptomics, and molecular network resources (12, 79, 93).