Network Medicine: Complex Systems in Human Disease and Therapeutics

by Joseph Loscalzo

In a time when little was known about HIV, this study that described a social network was pivotal in understanding the spread of the virus and the risk factors for the disease.

Social processes and structures emerge from the friendship, kinship, contact, resource-exchange, and information-exchange ties between people.

Berkman and colleagues (2000) have provided one of the more useful and detailed conceptual models that shows how the social and physical relate to one another to affect human health and disease.

Berkman’s model has at least three important features. First, the model is ecological and clearly positions social networks between the broader macrosocial environment and the narrower micro-level of psychosocial processes.

Disease transmission may be slowed in a community that is characterized by isolated network components with low reachability, as compared with a tightly connected community network with high reachability. Second, the model outlines plausible causal mechanisms whereby specific network characteristics may influence psychosocial processes.

High density and reciprocity make it easier for social support to flow through the network. Another example is that networks with more multiplexity (different types of coexisting network ties) may make it more likely that health information flows through both word of mouth and professional communication networks.

Consider an adolescent network that exhibits disassortative mixing regarding smoking status, which means that many teens who smoke are connected to many teens who do not smoke.

Disassortative mating (also known as negative assortative mating or heterogamy) means that individuals with dissimilar genotypes or phenotypes mate with one another more frequently than would be expected under random mating. Disassortative mating reduces the genetic similarities within the family

The implication is that not only static network characteristics may be important influences on disease, but also dynamic processes such as tie formation and network growth may be important predictors of disease.

It is important to understand that even though the model is ecological, it is not linear. In particular, disease processes and health outcomes influence earlier parts of the model (as reflected by the feedback loops).

First, we can design interventions that work directly on the networks themselves to prevent the occurrence of disease or reduce its impact. Second, we can use network information to inform or enhance more traditional disease interventions.

Traditional epidemic models of infectious disease such as the SIR (Susceptible-Infected-Recovered) model and others function well when there are relatively clear mechanisms of transmission, infectivity, and outcomes

However, most of the research using epidemiological models with three categories: susceptible, infected, and recovered (SIR models) assumes random mixing within a population; that is, social structure is ignored.

By adding social-network variables and random graph modeling, improved models can be created that optimize control of epidemics

It is no longer a radical idea to understand infectious diseases as being driven by social and relational processes, and network analysis is a primary tool for understanding those relationships.

Social networks are comprised of interpersonal ties that form a pipeline to structure contagious flows (Borgatti and Halgin 2011).

Until treatments can be devised and vaccines can be developed, social isolation and distancing are often the most effective methods of preventing mass infection.

After extensive interviewing, network techniques showed that those with syphilis had higher scores of betweenness centrality and relationship microstructures, with network cliques playing a large role in transmission.

Historically, one of the most consistent network findings in chronic disease has been the association of social isolation and integration with mortality.

Although we still do not know much about the precise mechanisms that link social network characteristics to the risks for developing cancer and heart disease, research over the past 20 years has started to shed light on two particular pathways: the link between social isolation and chronic disease survival, and the relationship between network characteristics and risk-reduction strategies such as cancer screening.

Clearly, more studies are needed that move beyond simple network size counts as predictors of chronic disease survival. The second prominent line of research has focused on the role of social networks in risk-reduction strategies for chronic diseases such as cancer screening, genetic testing, and health communications.

Network analysis has been used primarily to address two broad sets of tobacco-control questions: How do social networks influence individual tobacco use? How are community, state, national, and international tobacco-control systems structured?

Almost all people who start smoking do so before they are 18, and when they start they typically get their first cigarettes from a friend or family member

Another study confirmed that there is an accumulation of risk takers in isolated positions and that individuals in isolated positions drifted toward risk-taking groups, while clique members shifted from non–risk-taking behaviors to risk-taking behaviors over time (Pearson and West 2003).

Specifically, they found that popular students (defined in network terms) were more likely to smoke in schools with high smoking prevalence and less likely to smoke in schools with low smoking prevalence.

Essentially, teens who were more centrally positioned in their friendship networks were less likely to be involved in smoking even when they had more friends who smoked. They proposed the interpretation that teens with higher social status (i.e., greater centrality) may have more resources to help withstand the influence of friends who smoke.

They found that friendship ties were significantly more likely to occur between teenagers who were of the same weight, as compared with those of different weights. Using more sophisticated exponential random graph model (ERGM) analysis of the much larger National Longitudinal Study of Adolescent Health, Schaefer and Simpkins (2014) found that nonoverweight adolescents (as measured by body-mass index [BMI]) tended to avoid forming friendship ties with overweight peers.

One of the most consistent characteristics of social networks is that of homophily—the tendency for individuals (or other social actors) to be connected to similar others. Thus comes the saying “birds of a feather flock together” (McPherson et al. 2001).

The challenge has been to disentangle two different social processes that may explain the homophily pattern: (1) social selection, in which network members select partners who are similar to themselves; and (2) social influence, in which the behavior of network partners exerts influence on an actor and the behaviors become more similar over time

Valente and colleagues (2004) frame this as a “chicken or the egg” argument; that is, does friendship grouping precede or follow substance use?

Using these new modeling techniques (ERGMs), Mercken and colleagues (2010) have shown that both selection and influence are important processes that shape adolescent smoking behavior. Similarly, Schaefer and colleagues have identified selection and influence processes operating on obesity-related behavior (physical activity and BMI) among friendship networks (Simpkins et al. 2012).

Interventions can take advantage of this effect, either by working directly to influence network characteristics or by using the social network itself as the setting for the treatment.

This network-contact map reveals an important behavioral fact that should be considered when developing an intervention (e.g., to reduce infectious transmission rates)—viz., that nurses play a more critical interactional role than other staff, including physicians. Network characteristics can reveal important information about social and organizational systems that can directly inform development of disease treatments (Luke and Stamatakis 2012).

Network analysis studies of human disease have gone through three stages. The earliest stage, which accounts for most of the research, is made up of studies that have collected simple data on basic network characteristics such as overall size

The second stage of network disease studies is characterized by research that focuses on whole networks (as opposed to egocentric networks) and uses more sophisticated network visualization and descriptive methods.

Finally, we are just now entering the third stage of network disease research. Here, studies move beyond viewing networks as static structures and use modern social network modeling and computational modeling techniques to explore network dynamics and how networks are part of larger, complex systems (Luke and Stamatakis 2012).

Our ability to understand, treat, and prevent human disease is based on a broad and deep evidence base constructed by geneticists, physiologists, clinicians, epidemiologists, and scores of other basic science, medical, and social science disciplines.

Disease has long been classified as a series of phenomenological clusters (symptoms, signs and objective data) associated with discrete outcomes (Jones 1868).

There is an ongoing tension between increasing study size and decreasing resolution of observation that sets limits on the biological effects that are detectable.

There is a recognition that specific tradeoffs in the interest of increasing study size have resulted in the aggregation of large numbers of etiologically distinct entities (on the basis of superficial resemblance) in both epidemiological investigation and clinical trials, in turn leading to the dilution of relationships between cause (or intervention) and effect.

While there have been numerous efforts to define rigorous mathematical models of very specific biological events, the completion of the human genome project (and parallel projects for other genomes) has highlighted advantages of global descriptions in biological frameworks

Network science approaches the problems of medicine through comprehensive descriptions of the components of an entire system, their quantitative relationships, and their combined response to specific perturbations.

At the core of the tensions between reductionism and holism in medicine lies the biological complexity of the phenomena themselves and the relatively rudimentary tools that we use for observation and modeling.

In practical terms, realization of the vision of network medicine demands observational outputs that can describe precisely the network state under a broad range of conditions and across a broad range of scales

Once genome sequencing had defined the comprehensive static genetic code for an organism, all downstream consequences were readily aggregated in the definition of phenotype.

Only in recent years, as the dynamism of the individual genome during development and beyond has been recognized (Poduri, Evrony, et al. 2013) and organismal phenotypes have been discovered that represent the integrated output of multiple genomes (e.g., via host–microbiome interactions), have these distinctions again become blurred (Wang, Klipfell, et al. 2011; Brestoff and Artis 2013).

Indeed, many of the major problems with etiologic heterogeneity might be resolved if phenotype definitions incorporated dynamic responses to standardized challenges or objective transmission likelihoods.

The limited predictive utility of major effect-size genes suggests that the observed phenotypes are strongly modified by other genes or by nongenetic factors.

Diathesis is a hereditary or constitutional predisposition to a disease or other disorder.

A central argument for changing the approach to phenotyping is our inability to rigorously attribute the factors determining penetrance or pleiotropy without systematic studies of the heritability of each component of the phenotype and measurement of the relevant environmental contributions.

Among the benefits of network science is the feasibility of employing specific network modules and module states as endophenotypes.