Network Medicine: Complex Systems in Human Disease and Therapeutics

by Joseph Loscalzo

Charting the path from genotype to phenotype in human disease is not straightforward. In addition to genetic and environmental modifiers, differences in gene variants can affect how disease genes are expressed and function, and how a disease phenotype manifests.

The situation is more intricate for complex traits, for which the disease states appear to result from multiple perturbations of molecular machines and pathways

Many proteins have been found to interact with p53, a central player in the regulation of cell proliferation and apoptosis frequently found mutated in cancers

What these telling examples relate is that human disorders do not necessarily reflect the absence of the incriminated gene product, as per conventional models of human disease (Botstein and Risch 2003). Often, the disease results from subtler molecular defects that impinge on specific protein–protein interactions (Chakravarti, Clark, et al. 2013).

In network models of the interactome, these truncating mutations can be thought of as the removal of one node along with all of its edges—a node removal

While investigation of the precise interaction defects associated with point mutations is of course not new, the term edgetic promotes a subtle yet meaningful archetype shift from conventional gene-centric approaches toward interaction-centric approaches.

Edge perturbing network models, which emphasize consideration of which specific edges are affected by a mutation, complement and extend classic gene-centric models, which ascertain only whether a gene product is present or not present and neglect less overt alterations of a given gene or gene product

Two complementary strategies can be pursued to investigate edgetic perturbation models: forward edgetics and reverse edgetics

Direct tests of the edgetics model reported so far are beginning to resolve particular issues of confounding genetic heterogeneity in human disease

Modeling network perturbations induced by mutations in alleles associated with human Mendelian disorders led to an estimate that up to half of the disease-associated mutations examined could potentially encode proteins with edgetic interaction defects

Reverse engineering presents a potent approach to delineate genotype–phenotype relationships. Y2H assays are an especially potent tool to delineate mutations that perturb specific protein–protein interactions and even to demarcate protein–protein binding sites

Reverse engineering can also generate proteins with point mutations that do not lose an interaction but that actually have increased affinity for a specific interactor, a gain-of-interaction

Many gain-of-interaction point mutations in PCNA actually exhibit more deleterious phenotypes than mutations that lose the binding altogether. Such subtle edgetic perturbations can produce more pronounced effects than conventional genetic perturbations that remove a protein node altogether

Despite impressive progress, the wealth of information provided by the genomic revolution has not yet been translated into actionable knowledge about disease pathogenesis

It is now undeniable that the complexity of genotype–phenotype relationships will not be understood without considering the underlying system. Part of the resolution will hopefully come from the emergence of edgetic perturbation models.

The C. elegans apoptosis ced-9 gene, the homolog of mammalian BCL2, was originally identified by the isolation of a dominant allele, ced-9(n1950), corresponding to a G169E mutation. This mutation prevents embryonic cells from undergoing programmed cell death, or apoptosis, by precluding the EGL-1-induced dissociation of the CED-9/CED-4 apoptosis effector complex

Hengartner and Horvitz 1994). The isolation of this allele, which is now understood to be edgetic

The impact of all of these four alleles can now be understood to represent a node removal, depleting CED-9 in the cell

From a network medicine perspective, the diseases associated with viral infection can be attributed to the interaction between viral proteins and host proteins

AP–MS technologies have witnessed a similar progression from small-scale (mapping the host interactors of hepatitis B virus multifunctional protein HBx [Zhang, Xie, et al. 2013]) to systematic virus–host co-complex mapping.

These systematic proteome-scale screens provide unbiased lists of viral-targeted proteins through which to identify cellular pathways and functions crucial to the viral life cycle.

In molecular biology, an interactome is the whole set of molecular interactions in a particular cell. The term specifically refers to physical interactions among molecules (such as those among proteins, also known as protein–protein interactions, PPIs; or between small molecules and proteins) but can also describe sets of indirect interactions among genes (genetic interactions). The interactomes based on PPIs should be associated to the proteome of the corresponding species in order to provide a global view ("omic") of all the possible molecular interactions that a protein can present.

Host proteins directly targeted by HIV proteins predominantly localized in groups of human proteins strongly interconnected in the underlying network

Interacting with highly connected cellular proteins enables a viral protein to manipulate diverse cellular pathways simultaneously and efficiently.

The global landscape of host proteins targeted by proteins from DNA tumor viruses helped to prioritize genes and pathways involved in cancer

Systematic mapping of the interactions and expression perturbations induced by proteins of four classes of tumor viruses, polyomaviruses, papillomaviruses, adenovirus, and Epstein–Barr virus, produced a list of virally perturbed host genes that showed a significant overlap with known genes in the COSMIC Classic list of cancer-associated genes

Genes in close network proximity to host proteins targeted by EBV and HPV16 viral proteins showed significantly shifted expression patterns in diseases associated with the viral infection and are significantly closer to genes associated with virally implicated diseases than to genes associated with other diseases (Gulbahce, Yan, et al. 2012), leading to the Local Impact Hypothesis, similar to the local neighborhood hypothesis for disease modules (Barabási, Gulbahce, et al. 2011).

Exploitation of the local impact hypothesis to viral disease networks prioritized several diseases as candidate virally implicated diseases based on network topology and on population-based clinical associations between candidate diseases and viral infection. One such prioritization highlighted a novel connection between HPV infection and Fanconi anemia (Gulbahce, Yan, et al. 2012).

Alternative splicing vastly increases proteome diversity, and occurs in nearly all eukaryotes

The accumulation of introns, so much more prevalent in mammals (especially in primates) than in metazoans and other eukaryotes, confers an evolutionary advantage by efficiently increasing the number of protein isoforms (Barbosa-Morais, Irimia, et al. 2012) without commensurate increases in the gene count.

Information about which splice isoform is responsible for an interaction is hardly ever presented (Talavera, Robertson, et al. 2013). Incorporating transcript diversity is critical to generation of biologically relevant networks, which would be more informative than currently available interactome datasets

The exact number of different proteins produced by alternatively spliced isoforms from the human genome is not currently determinable

With neuroligin, a protein involved in synaptogenesis, isoforms that bind only β-neurexins stimulate synapse formation, whereas other isoforms that bind both α- and β-neurexins promote synapse expansion

These observations taken together are consistent with removal or inclusion of protein–protein interaction domains being a major role of alternative splicing

Differential protein binding by isoforms could potentially allow for tissue-specific protein-interaction networks without increasing the number of genes in the genome.

There is an urgent need for the evolution of technologies to systematically map proteome-wide the interactions of naturally occurring isoforms.

The complementary technology of RNA-seq can identify thousands of alternatively spliced junctions between exons (Sultan, Schulz, et al. 2008), but it cannot overcome the connectivity problem of determining which alternatively spliced exon is connected linearly to which other exons in any single transcript.

With resources of cloned alternatively spliced ORFs now in hand, systematic isoform-specific interactome maps are appearing

through small regions comprising protein-interaction domains, not through the full protein molecule.

The structures of many domain–motif interaction domains have been solved (Pawson and Nash 2003; Bhattacharyya, Remenyi, et al. 2006; Liu, Engelmann, et al. 2012), providing insight into how cell networks and protein complexes dynamically assemble and disassemble.

Fragmentation for determination of the minimal region of interaction is commonplace because of the ease and convenience of the DNA-based technology underlying Y2H

The interactome networks produced to date are necessarily static interactomes. They describe the interactions that can occur between proteins (biophysical interactions), but within any particular cell or tissue at any particular time, only a subset of the mapped interactions actually functionally occur (biological interactions)

The new technology of AP-SWATH might soon enable reliable quantification of temporal changes of protein interaction networks in response to specific conditions

Condition-specific interactomes cannot currently be generated experimentally at proteome scale by any technology, although two-hybrid approaches are being developed for conditional expression of a third-party protein, such as a methyltransferase or protein kinase, to probe interactions that are dependent on specific protein modifications

Given current limitations of experimental interaction-detection technologies, true condition-specific interactomes mapped systematically, not just computationally, would seem to be a distant but essential goal.

A reference human interactome map, defined as interactions experimentally profiled for at least one isoform for each annotated human gene, now seems attainable.

Just as was the release of the reference human genome, release of a reference human interactome, although a remarkable achievement, would be just the first step toward deciphering the workings of cellular networks.

In a famous illustration from public health, John Snow created a map of cholera cases in Soho, London, demonstrating how the cause of a particular ailment could be ascertained through the detailed investigation of disease and geographic locations.

Snow helped create the modern discipline of epidemiology with his disease-investigation techniques, and he was able to solve a mystery of infection transmission and prevent further outbreaks of cholera.

Just as Snow’s map of disease around a water pump described a physical context for the transmission of cholera, the CDC’s “Patient 0” network map describes a social context for the transmission of HIV.