Network Medicine: Complex Systems in Human Disease and Therapeutics

by Joseph Loscalzo

Modern genomic techniques are emerging as useful tools in resolving current relatively crude clinical syndromes, but until we are able to define preclinical or even prepathological endophenotypes, the predictive goals of modern medicine are unlikely to be realized. This will require much more systematic exploration of phenotypic space and similar approaches to the definition of a vast range of stimuli that are currently unmeasured.

For any given quantitative trait, while the effects of individual alleles may contribute evenly to the normal distribution across that trait, the presence or absence of specific alleles may be discernible through distinctive effects on other orthogonal traits between alleles.

The inescapable conclusion is that many genetic determinants of phenotypes are undetected by modern investigation, not as a result of any limits of genotyping, but rather as a result of the limitations of phenotyping (MacRae and Vasan 2011).

Without the relevant phenotypic assay or without the necessary stimulus to generate a particular dynamic phenotypic response, the major alleles may simply be silent. These phenotypic phenomena likely account for a significant proportion of the “hidden” heritability inferred in many modern human genetic studies

Deconvoluting interactions on such a scale, and in the setting of high rates of de novo variation, is currently inconceivable without a holistic probabilistic framework such as network medicine.

More sophisticated models that incorporate genes of differing effect size, environmental factors, and stochastic components will have to be developed.

If nothing else, these data force us to reconsider the deterministic implications of specific sequence variants, and to reevaluate the particular phenotypes we have associated with these variants (Seidman and Seidman 2001).

Human disease characterization is, for many biological processes, the highest-resolution phenotyping that exists. Much of the technological innovation that has taken place in biology has been driven by the incentives of clinical medicine, in which we often imagine that the resolution of our diagnoses can barely be improved.

Pathophenotypes, however, are likely more heterogeneous than the basal physiological traits, simply as a result of the range of mechanisms by which normal biology may be perturbed.

Even if the network components or their connectivity has not been disrupted, pathophenotypes or endopathophenotypes may operate in parts of the dynamic range discrete from the physiological state. Here again, the central role of a defined stimulus to elicit (and organize) a measured response can be appreciated. Without a series of calibrated stimulus–response pairs that traverse the different network states from organogenesis to adulthood or from a timescale of microseconds to decades, it may be difficult even to detect an abnormal disease network.

For each and every phenotype there are so-called phenocopies or “indistinguishable” biological outcomes that do not share the same underlying genotype

The concept of increased dimensionality in phenotyping that is incorporated in network approaches may improve the resolution of any trait and thereby reduce the likelihood of a mechanistically unrelated phenocopy.

The silence of many pharmacologic response alleles without drug challenge is an excellent example of how many traits may simply be undetectable without the correct assay or without the correct conditioning variable

The same technologies that are being exploited to improve the collection of phenotypic data must be adapted to begin to record the exposome at a similar level of resolution

Ultimately, the collection and analysis of “big data” on metrics such as nutritional or other purchase data, personal habits, travel, and other exposures, will be a core component of network approaches to understanding human biology

An important goal for network medicine will be to redefine the tools for capturing variation in the phenome in a way in which the information content can be maximized.

No other organism is capable of being phenotyped with the granularity that is feasible in humans. Many phenotypes that incorporate higher cognitive faculties are inconceivable or not present in other organisms. Self-report alone defines a phenotypic universe that cannot be accessed in any other species and often exceeds the sensitivity of current objective technologies.

This complex set of definitions is built on the general strategy of breaking an individual disorder down into a series of nonoverlapping dimensions. These traits themselves have been refined iteratively on their performance in studies of natural history, genetics, and therapies, and the DSM is now in its fifth iteration.

The term diathesis derives from the Greek term (διάθεσις) for a predisposition, or sensibility. A diathesis can take the form of genetic, psychological, biological, or situational factors. A large range of differences exists among individuals' vulnerabilities to the development of a disorder.

Parallel work in the common variant genetics of psychiatric disease has revealed a large number of loci that are shared across apparently very different phenotypes

Ultimately, these insights have led the psychiatric community to drive toward more global approaches to phenotype firmly based in quantitative objective metrics. It is clear that as biology enters a more quantitative and holistic era, a rigorous reappraisal of all phenotypes will be required.

Together these observations on the origins, evolution, and limitations of the phenotype concept argue for a much more comprehensive approach, one that documents the fundamental genetic molecular and cellular architecture of physiological and pathophysiological traits, not only under baseline conditions, but also in response to a range of empirically tested perturbations.

Systematic approaches to detection and quantitation of the exposome will be required to make any inferences regarding the relevant biological networks.

In many instances, simply changing the resolution of a phenotype might dramatically improve our understanding of the pathophysiology of disease.

For example, elegant metabolomic data from population cohorts demonstrate abnormalities of amino acid metabolism in those who will go on to develop overt diabetes over a decade prior to any detectable abnormality of glucose handling (Wang, Larson, et al. 2011).

Together these emerging examples highlight a profound discordance between the limited dimensionality of even the most modern clinical phenotypes and the biologic heterogeneity of the very diseases for which these current phenotypes are the effective gold standard.

With innovation in this arena, functional abnormalities might be detectable at a stage where downstream final common pathways are not yet activated and when more robust etiologic discrimination is feasible.

Nosology (from Ancient Greek νόσος (nosos), meaning 'disease', and -λογία (-logia), meaning 'study of-') is the branch of medical science that deals with the classification of diseases.

The ideal phenotypes for network medicine are beginning to be imagined, not only in the ivory towers of academic medicine, but also in laboratories where digital devices are being developed, in online data warehouses, and within lay organizations such as the “quantified self” movement.

The most successful next-generation phenotypes will likely emphasize several core attributes, while reevaluating, extending, or complementing existing technologies. A suite of digital clinical measurements that reproducibly captures information in multiple dimensions, across different spatial and temporal scales will be vital for network science.

Ontology is a set of concepts and categories in a subject area or domain that shows their properties and the relations between them. "what's new about our ontology is that it is created automatically from large datasets

Ultimately, as innovative personal and ambient technologies emerge, many of these strategies will move beyond the domain of disease classification to the active documentation and maintenance of wellness.

Reassessment of existing technologies outside the constraints of current nosology using techniques such as machine vision and machine learning will offer early insights. Innovative phenotyping tools, standardized perturbations, and vast electronic health records integrating institutional and personal data will allow the generation and validation of in silico network models capable of predictive utility

For rigorous analytics that might facilitate the early detection of deviations from health, the prediction of disease, and the enabling of real-time monitoring, we will need to collect a standardized multidimensional dataset on vast numbers of individuals across their lifespan.

To undertake this analysis at the necessary scale will require the definition of a novel minimal phenotyping dataset for modern medicine, paralleling the standard “history and physical” of the past two centuries

This dataset must incorporate new smart ontologies for self-reported data, definitive imaging series at high resolution (with no exposure risk), functional genomics, novel digital inpatient or outpatient technologies, and personal data on a scale that has to date on...

Medicine will need to train a generation of generalists, capable of reconciling the deterministic approaches of recent decades with the holistic integrative strategies of network medicine.

Perhaps the most daunting hurdle to be overcome in this transformation is the change in scale of the data that clinicians must manage, understand, and act on. These challenges will mandate that physicians have access to the requisite skills to develop network analytics, as well as innovative methods of data display, knowledge management, and real-time education to allow all providers to practice network medicine.

Network medicine is a rapidly evolving, new field that strives to apply network science and systems biology principles to disease mechanism and to pharmacotherapeutics

Biomedical research into disease was ever in search of the cause (or a limited number of causes) of pathogenic processes as complicated as myocardial infarction, heart failure, or cirrhosis.

At this stage of any disease process, the drivers of disease expression are not specific to the disease, but represent later-stage mechanisms that underlie all disease—the intermediate (patho)phenotypes of inflammation and immune response, thrombosis or hemorrhage, fibrosis, apoptosis, and cell proliferation.

Similar reasoning leads to the conclusion that many so-called specific disease therapies actually focus on these common intermediate pathobiological mechanisms rather than on the specific underlying disease determinants, such as the use of antithrombotics in acute myocardial infarction, or the use of anti–tumor necrosis factor therapies for rheumatoid arthritis.

Network medicine has as its basic operating tenet the identification of disease networks or disease modules within networks responsible for specific pathophenotypes.

One can argue, however, that these loci may well have more powerful effects if their function were explored in the network within which they operate. One can, therefore, view genome-wide association studies as providing the “parts list” of potentially important functional variants; however, the assembly diagram is required to appreciate how these variants may affect function or cause dysfunction and disease.

Network modeling can be approached either statically or dynamically. Static modeling focuses only on network architecture or topology, generally without regard for the directionality of a link between nodes, the strength of association, or the changes in node expression or activity as a function of network perturbation.

Dynamic modeling in the extreme requires knowledge of the formal kinetics and kinetic constants, as well as reaction stoichiometry, for a deterministic approach using systems of nonlinear ordinary differential equations of the form: where dS(t)/dt refers to the time derivative of the concentration vector for all species in the pathway or network, N is the stoichiometry matrix for all reactions in the pathway or network, and ν(S,k)is the flux vector defined by the rate constants (k) governing all reactions in the pathway or network.

Simulation of network dynamics can provide, for example, an approach for defining stable operating ranges of the network with regard to substrate availability (using ordinary differential equations) or transcription factor activity (using stochastic differential equations given the limited [at least two] number of potential cis-binding sites for any given transcription factors).

This network structure suggests that perturbations of a single node can evoke wide-ranging effects through multiple modules, leading to significant consequences for functional (patho)phenotypes.

The identification of the disease module(s) within a network requires first that one recognize that genes governing chronic, complex diseases are typically nonessential and generally weakly linked within the interactome-derived network

In addition, it requires that one can distinguish among three types of network modules: topological, or modules that are defined purely by proximity of nodes to one another; functional, or modules that comprise nodes with common functional consequences, often acting through a common pathway; and disease, or modules that reflect altered functional consequences owing to changes in the behavior of a (mutant) node, the loss of a links between normally functioni...