Network Medicine: Complex Systems in Human Disease and Therapeutics

by Joseph Loscalzo

The proximity and degree of overlap of two network neighborhoods has been found to be highly predictive of the pathobiological similarity of the corresponding diseases

A negative sAB indicates topological overlap of the two node sets, whereas a positive sAB indicates topological separation of the two node sets.

An important property of networks is, therefore, their robustness, or resilience, against the breakdown of nodes or links that may break such paths.

Networks in which only a fraction of nodes and/or links are present have been studied extensively in the framework of percolation theory

Generally, as long as a certain critical fraction of all N nodes (or L links) is present, the network remains globally connected

Below this critical fraction, the giant component disappears and the network breaks into small disconnected components.

In probability theory and statistics, the binomial distribution with parameters n and p is the discrete probability distribution of the number of successes in a sequence of n independent experiments, each asking a yes–no question, and each with its own boolean-valued outcome: success/yes/true/one (with probability p) or failure/no/false/zero (with probability q = 1 − p). A Boolean-valued function (sometimes called a predicate or a proposition) is a function of the type f : X → B, where X is an arbitrary set and where B is a Boolean domain, i.e. a generic two-element set, (for example B = {0, 1}), whose elements are interpreted as logical values, for example, 0 = false and 1 = true, i.e., a single bit of information.

A scale-free network is a network whose degree distribution follows a power law, at least asymptotically

Since the vast majority of nodes in scale-free networks have only a few connections, random failure mostly affects such low-degree nodes, which, in turn, have little impact on the overall integrity of the network. Such networks are, therefore, remarkably tolerant against random node removal.

This robustness against random failure has also a down side: the networks are particularly vulnerable to a targeted attack that systematically removes the hubs, that is, the nodes in the network with the highest degrees (Albert, Jeong, et al. 2000). The precise fraction of removed hubs under which the network breaks down depends on the details of the degree distribution. For the interactome, we find that removing ~30% of the nodes is sufficient to destroy the network completely

In the same way that a whole network can fall apart under random node removal or attack, subgraphs inside a network can become disconnected if network incompleteness exceeds some threshold.

The percolation threshold is inversely proportional to m; that is, smaller subgraphs require a higher network completeness in order to have a giant component.

Machine-learning approaches, such as neural networks, support vector machines, or Bayesian networks, typically combine protein-interaction data with other sources of information, such as protein sequence and structure, pathway membership, gene expression, or the genome-wide association study (GWAS) p values

A number of methods aim to identify possible new disease gene candidates relying solely on the position of known

To maintain the degree distribution of a network, we randomize the interaction partners of the nodes while preserving each node’s degree.

Randomization of the network topology is primarily used to identify the impact of the network topology on the system’s behavior.

One could argue, for example, that the high number of connections among disease proteins is a result of their relatively high degree.

The interdependence between different layers of the network can give rise to cascading failure, where the breakdown of a node in one layer propagates throughout all other layers, leading to a global breakdown.

The rapidly evolving field of temporal networks aims to incorporate these dynamic aspects of networks and to explore the impact of this temporality on its structural and dynamic characteristics

In systems of biochemical reactions, for example, it has been found that by monitoring a few selected nodes one can infer the complete state of the entire system

These results could have immediate application in the rational design of biomarkers for disease states, as well as in rational drug target(s) selection. The ultimate goal is to control these systems, that is, to drive a cell from a disease state to a healthy state

To accomplish the biological processes needed for organisms to survive, cells act through dynamic complex systems formed by interacting macromolecules.

The fundamental challenge is to identify and catalog the protein–protein interactions that can take place among all members of a proteome. From that information, the next challenge is to understand when and where these interactions take place in vivo.

(1) curation of protein-interaction data from the existing scientific literature, (2) computational predictions of protein interactions based on available orthogonal information, and (3) systematic experimental mapping at proteome scale to identify: (a) co-complex associations or (b) binary interactions.

Such information can be mined by either direct annotation by trained curators or by computational text-mining methods (Mosca, Pons, et al. 2013), stored in publicly accessible databases (Orchard 2012), and used to produce large interactome-network maps.

With few exceptions, negative results are not reported in the primary literature and hence are not available in curated protein-interaction databases

Homology is the state of having the same or similar relation, relative position, or structure

The two main strategies for experimental interactome mapping. In yeast two-hybrid (Y2H) screening, physical interaction between two hybrid proteins within the nucleus of a yeast cell reconstitutes a transcription factor that activates expression of a selectable reporter gene. Y2H returns binary interactions. In affinity purification followed by mass spectrometry (AP–MS), a protein complex between a tagged bait protein and associated cellular proteins is purified by affinity purification, and the constituents of the complex are identified by mass spectrometry. AP–MS returns co-complex associations.

Binary and co-complex screens interrogate vastly different parts of the interactome

Rather, third-party users have made attempts to distinguish binary from indirect interactions by defining the methods that have been used to map protein interactions as primarily binary or primarily indirect

It is not that one technology provides better or lesser maps of the interactome. The two principal experimental technologies provide complementary, almost orthogonal, views of the protein interactome

As the biases of different approaches—literature, computational, or experimental—in mapping protein interactions become appreciated (Futschik, Chaurasia, et al. 2007; Gillis, Ballouz, et al. 2014), strategies toward reaching a complete human interactome map can be suitably adapted.

All these efforts to rigorously control the quality and reproducibility of systematic protein-interaction maps have greatly increased confidence in their quality.

Much as the Human Genome Project could reach completion only once efforts transitioned from small-scale sequencing efforts of uneven and inestimable quality to sequencing consortiums of definable quality and rigor

And just as with the Human Genome Project, proteome-scale interactome mapping projects do not preclude small-scale focused investigations of biological mechanisms, but rather, stimulate and augment them

The problems posed by inspection bias extend beyond mapping of protein interactions to the development of pharmacological agents and other aspects of modern biomedicine

With continued advancement of systematic protein-interaction mapping efforts, the expectation is that interactome “deserts,” the zones of the interactome space where biomedical researchers simply do not look for interactions owing to the lack of prior knowledge, might eventually become more populated

An effective way to identify novel disease genes is to examine the interaction partners of proteins encoded by known disease genes

Focused protein–protein interaction mapping efforts have identified novel interactions among proteins encoded by known disease genes and have also predicted new disease-susceptibility genes.

Usher syndrome is an inherited, genetic condition. The main symptoms are hearing loss and sight loss, due to a condition called retinitis pigmentosa (RP) and, in some cases, balance difficulties.

The common finding among these disease-centric interactome models is the discovery of unexpected relationships between disease genes that initially appeared unrelated. Accordingly, building and analyzing disease-centered networks is a critical step toward the fundamental understanding of underlying disease mechanisms and may also lead to improved predictions of disease survival.

Interactome maps have highlighted new candidate disease genes, disease pathways, and disease-modifier genes, but the task of investigating the impact of causal variants on interactome networks is in its infancy.

A disease module is defined as a group of interacting network nodes that ordinarily contribute to a common cellular function and that, when disrupted, lead to disease

Given these findings, fuller maps of the human interactome would lead to more accurate maps of modules, which in turn would lead to novel or more complete hypotheses regarding disease-gene candidates

Bardet–Biedl syndrome (BBS) is a ciliopathic human genetic disorder that produces many effects and affects many body systems. It is characterized principally by obesity, retinitis pigmentosa, polydactyly, hypogonadism, and kidney failure in some cases. Historically, slower mental processing has also been considered a principal symptom but is now not regarded as such.

Meckel syndrome (also known as Meckel–Gruber syndrome, Gruber syndrome, dysencephalia splanchnocystica) is a rare, lethal, ciliopathic, genetic disorder, characterized by renal cystic dysplasia, central nervous system malformations (occipital encephalocele), polydactyly (post axial), hepatic developmental defects, and pulmonary hypoplasia due to oligohydramnios.

Alström syndrome (AS) is a rare autosomal recessive disease characterized by multiorgan dysfunction. The key features are childhood obesity, blindness due to congenital retinal dystrophy and sensorineural hearing loss. Associated endocrinologic features include hyperinsulinemia, early-onset type 2 diabetes and hypertriglyceridemia. Thus, AS shares several features with the common metabolic syndrome, namely obesity, hyperinsulinemia and hypertriglyceridemia. Mutations in the ALMS1 gene have been found to be causative for AS with a total of 79 disease-causing mutations having been described.

Joubert syndrome is a rare autosomal recessive genetic disorder that affects the cerebellum, an area of the brain that controls balance and coordination.

Kaufman syndrome: The condition is named for Dr. Robert L. Kaufman and Victor McKusick. It is sometimes known by the abbreviation MKS. In infancy it can be difficult to distinguish between MKS and the related Bardet–Biedl syndrome, as the more severe symptoms of the latter condition rarely materialise before adulthood. McKusick-Kaufman syndrome affects 1 in 10,000 people in the Old Order Amish population. Research has not identified cases outside of this population. Clinically, McKusick–Kaufman syndrome is characterized by a combination of three features: postaxial polydactyly, heart defects, and genital abnormalities: - Vaginal atresia with hydrometrocolpos - Double vagina and/or uterus. - Hypospadias, chordee (a downward-curving penis), and undescended testes (cryptorchidism). - Ureter stenosis or ureteric atresia

A finding often replicated is that cancer-gene products tend to be highly connected nodes, or hubs, in human interactome-network models