Network Medicine: Complex Systems in Human Disease and Therapeutics

by Joseph Loscalzo

These approaches include strategies that identify novel disease genes through their direct interaction with known disease genes (linkage-based), by their presence in a disease module (disease module–based), or by their presence in modules or pathways that are in the proximate neighborhood of a known disease module (diffusion-based). By identifying the network location of novel disease genes, the disease network is enriched and potential novel links identified through which disease modules can be characterized more thoroughly.

The higher the degree of localization, the greater the biological similarity of the disease genes, supporting the notion of well-localized disease modules in the interactome.

Topology, then, clearly recapitulates functionality within a disease. In addition and importantly, however, is the more generalizable finding that the network-based, interactome-defined location of a disease also determines its pathobiological relationship to other diseases. Topologically overlapping diseases show significant gene co-expression, symptoms similarity, and comorbidity, while topologically distinct and separated diseases do not

Despite increased data sets of potential drug targets and their structural characterization, drug discovery has been waning for over a decade (Loscalzo 2012). In part, this limitation is a consequence of the reductionist approach to discovery, which seeks a single molecular cause for each disease that can be targeted with an Ehrlichian magic bullet.

For these reasons, it is no surprise that drug discovery strategies that focus on phenotype screening are almost twice as successful as those that focus on target-based screening, even in the current era (Swinney and Anthony 2011).

In Figure 6–9, drugs are viewed as network perturbants that can affect the robustness (or ability to withstand perturbations—in this case by disease agents) of a (cellular) system in response to a disease agent.

As shown in the figure, ideally, a drug affects the pathophenotype landscape to restore homeostasis when diseased or to maintain homeostasis in the face of potential disease stress, preventing its expression. In this paradigm, drugs can also yield system dysfunction and cell death in the worst case (Csermely, Korcsmaros, et al. 2013).

Armed with this integrative map of molecular interactions, network medicine strategies facilitate the unbiased identification of novel disease pathways and modules, and provide support for rational approaches to therapeutics, including rational polypharmacy.

However, for most complex diseases, a large percentage of heritability—the fraction of the total phenotypic variation attributed to genetic causes—remains unexplained

Epistasis is the phenomenon wherein the effect of one gene (locus) is dependent on the presence of one or more 'modifier genes', i.e. the genetic background

many functional variants are likely regulatory variants of moderate effect

gene transcriptional regulation is complicated by tissue specificity as well as dynamic temporal and spatial controls

the impact of genetic variation on gene expression may escape detection with current methods in available tissue samples

The development of new bioinformatic approaches to prioritize potentially functional genetic variants, such as the Bayesian approach using dense genotyping data along a region of interest developed by Fahr and colleagues (Farh, Marson, et al. 2015), will likely assist in the iden...

We used chromosome conformation capture (3C) assays, an approach that cross-links DNA to characterize chromatin structure, in both lung epithelial (Beas-2B) and lung fibroblast (MRC5) cell lines to identify a 7-kb region approximately 85 kb upstream from the HHIP gene, which showed a long-range interaction with the HHIP promoter.

Since many genetic determinants of complex diseases appear to influence gene-regulatory networks, approaches like 3C that assess long-range chromatin interactions will likely be essential in defining the relevant genes involved in a particular disease-related network.

Network-based approaches can assist in this process. Erlich and colleagues used disease-network analysis to provide another layer of filtering of genetic variants identified by whole exome sequencing and to implicate KIF1A as the causative gene in a small consanguineous pedigree with hereditary spastic paraparesis (HSP)

One of these methods was Endeavor, which includes multiple data types such as protein–protein interactions to rate potential candidate genes.

O’Roak and colleagues sequenced whole exomes in parent-child trios in which the child was affected by autism to find de novo mutations, then created a protein-protein interaction network of genes with truncating or severe missense mutations using GeneMania

A high percentage of the mutated genes mapped to a particular interconnected network neighborhood, which may represent a disease network module for autism.

To identify genes regulated by HHIP and potentially related to COPD pathogenesis, my research group performed gene expression microarray analysis in a human bronchial epithelial cell line (Beas-2B) after RNA interference; the Beas-2B cells were stably infected with short hairpin RNAs (shRNAs) targeting the HHIP mRNA

Using the Predictive Networks Web application, we found that the extracellular matrix and cell proliferation genes influenced by HHIP tended to be interconnected more than expected by chance

Baranzini and colleagues developed a protein interaction, network-based, pathway analysis approach and applied this method to two GWAS of multiple sclerosis (Baranzini, Galwey, et al. 2009).

In addition to the known relationships of multiple sclerosis to immunological defects, they identified a neural-based network, which had not been implicated in many previous genetic association studies of multiple sclerosis.

Applying this approach to a breast cancer GWAS from the Nurses’ Health Study generated 9212 modules. They focused on the top 1% of modules for further analysis, which included 166 genes in a merged breast cancer subnetwork. As expected, some of these genes were highly associated with breast cancer, while others were weakly associated genes that were linked to strongly associated genes in the network.

They used two schizophrenia GWAS datasets to build and validate their genetic-network modules, and they implemented new approaches to assess the statistical significance of disease modules to overcome the biases related to gene length, SNP density, and baseline degree distribution. They found a disease module composed of 205 genes for schizophrenia; 76 of those genes showed nominal evidence for association in a third schizophrenia GWAS.

A variety of other computational methods have been developed to analyze GWAS results within a network context, including Network Interface Miner for Multigenic Interactions (NIMMI) (Akula, Baranova, et al. 2011), Disease Association Protein-Protein Link Evaluator (DAPPLE) (Rossin, Lage, et al. 2011), and NetworkMiner (Garcia-Alonso, Alonso, et al. 2012). These genetic-network approaches assume that genetic association studies and protein–protein interaction networks offer complementary information regarding a complex disease. These methods have provided strong evidence that protein products of genes associated with a complex disease are likely to have a physical interaction.

Epistasis refers to the interaction between different genes. Epistasis was originally defined in biological terms, to denote an interfering effect of one genetic variant on the phenotypic impact of another variant (Cordell 2002).

Testing for statistical evidence of epistasis may not capture a relevant biological interaction, such as binding between two proteins. Given the likely great importance of gene–gene interactions in complex diseases, it has been surprisingly difficult to demonstrate epistasis in human complex-disease research.

Their model including epistatic effects had 18.3% better predictive power than the model without epistatic effects. They found that intragenic interaction effects were generally greater than intergenic effects in HIV, and the strongest interactions were between nearby variants within the same protein structural domain.

They performed genome-wide association analysis in both a set of inbred Drosophila lines and a laboratory-derived outbred population created by crossing a subset of those inbred lines.

They point out that epistatic effects can contribute to the additive genetic variance as well as the interaction genetic variance, especially when minor allele frequencies are low.

One of the limitations of efforts to identify epistasis in human complex diseases has been the statistical approach of searching for a nonadditive contribution of pairwise SNP effects, typically by testing for significant cross-product interaction terms between allele counts for these SNPs in regression models. These approaches require pairwise tests of millions of SNPs, and the multiple statistical testing penalty has likely been a limiting factor in epistasis detection.

Hu, Moore and colleagues developed a method to use information theory measures to assess the main and interaction effects between SNPs within a statistical epistasis network. They applied this method to a candidate SNP panel of 1422 SNPs genotyped in patients with bladder cancer as compared with control subjects.

They assessed the statistical significance of their interaction network by permutation of disease status and found evidence for increased numbers of edges and nodes, as well as a large connected component, within the bladder cancer network.

However, even with these more elegant network-based statistical approaches, the assessment of statistical epistasis may not capture the biological epistasis (e.g., physical interaction between proteins), which is likely of greater relevance for disease pathogenesis.

Genetic studies have also been performed using many of the -omics data types that are described elsewhere in this book, including quantitative levels of messenger RNA (mRNA), proteins, and metabolites. The genetic architecture of these “intermediate phenotypes” or “endophenotypes,” which are biologically more proximal to the genetic variant, can be used to gain insight into the biological networks relevant for complex diseases.

Genetic determinants of gene expression levels, known as expression quantitative trait loci, or eQTLs, can be located near the coding gene (cis effects) or at a great distance on the same chromosome or on different chromosomes from the coding gene (trans effects).

Integrative genomics studies can provide functional insights into complex disease genetic associations; cis-eQTLs associated with disease can implicate a specific gene in disease pathogenesis and suggest that disease susceptibility is influenced by regulation of that gene, while trans-eQTLs can provide insight into the network of interacting genes and proteins involved in a complex disease

Their results suggested that network analysis could implicate key genes in disease-related phenotypes that would not have been identified using traditional genetic methods.

The application of statistical methods to test network relationships between genetic variants and gene expression levels will be helpful in rigorously delineating genetic networks.

Most of the reported pQTL studies have focused on a small number of proteins of interest. More recently, genetic-association analysis of a more comprehensive set of proteomics data was reported by Wu and colleagues

Of interest, including both a pQTL SNP (in the AGER gene) and the sRAGE protein biomarker level together provided substantially stronger associations to emphysema than either the SNP or protein biomarker level alone.

Mendelian randomization is a statistical approach that uses instrumental variable analysis to assess whether a biomarker is causally related to disease risk (Lawlor, Harbord, et al. 2008).

Mendelian randomization approaches can assist in determining whether a significant genetic determinant of a protein level is also related to disease risk, thus addressing one of the key challenges of second-generation genetic studies.

Metabolomic studies can provide complementary information to transcriptomics and proteomics, but because affordable, high-throughput metabolomic assay systems have been developed only recently, efforts to identify genetic determinants of large panels of metabolites have been more limited.

They analyzed genome-wide SNP genotyping data and all individual metabolites and ratios of metabolites in a screening stage with log-transformed metabolite values. They subsequently performed meta-analysis of both study populations and found 37 loci that were significant genome-wide after Bonferroni correction for multiple statistical testing.

They measured N-acetylornithine levels and found that they were correlated with renal function in their study populations, suggesting a potential role for this metabolite in chronic renal insufficiency. Of interest, six of their 37 genome-wide significant associations to metabolites have previously been related to adverse drug effects; further study of these metabolites and their genetic determinants could provide insight into the mechanisms for drug toxicity.

Network information can be used to interpret SNP associations with complex diseases, thus placing new genetic findings, which may have never previously been considered to be involved in disease pathogenesis, into biological context.

Epistasis networks hold great promise to identify the gene–gene interactions that seem likely to be involved in complex disease pathogenesis but that have been quite difficult to identify.