Personalization in practice

See allHide authors and affiliations

Science  16 Oct 2015:
Vol. 350, Issue 6258, pp. 282-283
DOI: 10.1126/science.aad5204

Last month, an advisory committee released recommendations for recruiting at least 1 million individuals to participate in the U.S. National Institutes of Health's Precision Medicine Initiative. This bold approach to disease treatment and prevention seeks to account for an individual's genes, environment, and lifestyle to improve health outcomes. The ability to collect, integrate, analyze, and model relevant data streams is central to this effort. Moving beyond “just” massive data collection will require structured convergence among various disciplines. So, how should data be gathered? Here, computational modeling can be a useful guide. Modeling at the molecular, cellular, tissue, and organismal level will be essential to identify the molecular interactions that underlie progressive diseases and to generate a comprehensive and dynamic picture of the individual.

In 2011, precision medicine was described by the U.S. National Research Council as resting on a “new taxonomy for human disease based on molecular biology” (1), but implicit in this notion is the assumption that defining noncommunicable diseases on the basis of an individual's genomic and epigenomic determinants alone will enable the personalization of therapy. This has led to conflation of the terms “personalized” and “precision.” The overlap is reasonable when the dominant driver of a disease is largely genomic, as in most cancers. However, for many other progressive conditions such as type 2 diabetes, psychiatric diseases, and heart failure, it is not clear whether genomic status is the major driver. All progressive diseases have genomic underpinnings, but it is the impact of diverse environmental influences—mostly unrecognized—on individual genomes that determines interindividual variation in disease progression and drug response (2). In short, we need to know the dynamics of an individual's physiology and pathophysiology.

Empowering the Precision Medicine Initiative requires a formalism to describe relationships between scales of organization and different time domains. This involves a convergence of measurements—from human cell culture experiments to studies in model organisms and clinical measurements in patients—and modeling to reflect unique and general aspects of each system and its relationship to human health and disease throughout the lifetime of an individual. There are several emerging powerful experimental and modeling technologies to do this. For example, induced pluripotent stem cells (iPSCs) enable cell type-specific measurements and provide the opportunity for in vitro experimentation with tissues at the level of an individual. Systems biology provides modeling formalisms to match key features of the molecular, cellular, tissue, and whole-organ physiologies for simulations. Here, Bayesian integration of heterogeneous data (3) can be a good starting point. Graph theory helps build networks that describe the local and regional geography of cells, organs, and organ systems. Dynamical modeling describes how this biological geography changes with environment, lifestyle, and age. Some of the dynamic modeling approaches are monomorphic (e.g., differential equation-based models), whereas some are more modular (linked simulations with different formalisms for different subsystems). Irrespective of the approaches used, modeling disease dynamics must start early, with incomplete data. Simulations can then drive the design of large-scale studies that are both clinical and laboratory-based.

Why do we need to consider dynamics at multiple levels? Empirical observations indicate that genomic and molecular diversity often leads to phenotypic convergence, resulting in a limited number of phenotypes. Dynamic processes at multiple levels are likely to drive phenotypic convergence. Such convergence can offer opportunities for shared treatments for maladies that have different underlying molecular and cellular architectures. For instance, a single kind of arrhythmia such as long-QT syndrome can arise from mutations in multiple genes encoding ion channels (4). As there is little research relating mutations to drug targets and therapy, knowledge of which genes are mutated offers no particular advantage when deciding therapeutic strategies for arrythmias. Another example is hypertension, which can be treated with drugs that are molecularly diverse (e.g., diuretics that act on renal tubules; ACE inhibitors that regulate vessel wall function; beta blockers that act predominantly on the heart). This reflects the molecular and tissue-based diversity underlying this phenotype. Currently, treatment for hypertension, although broadly effective, remains empirical, wherein both the type and timing of drug(s) used are arrived at by trial and error.

Data and dynamics.

Empowering precision medicine requires an iterative process of gathering data in a manner that is driven by integrative computational models. New data can then lead to model refinement, experimental follow-up, and further data, thereby capturing the dynamics of a biological process.


It may be possible to make therapy predictable by integrating several types of models—stochastic and statistical models that relate molecular and environmental determinants to disease origin and progression, and dynamical models of the pathophysiological process and effects of drug treatment. Clinical data such as brain imaging or circulating amounts of a peptide hormone or blood sugar would constrain such models. Modeling should then allow one to frame hypotheses for additional cell-based experiments and even clinical data gathering. These experiments include using human iPSC-derived cells that produce organs on engineered chips for high throughput analysis (5) or using humanized animal models wherein the cellular circuitries have been reengineered to reflect the human genomic and epigenomic determinants. Data from cell-level and humanized model system experiments could, in turn, enable better selection criteria for clinical data gathering. iPSC-derived cell types from patients are showing promise in modeling disease processes (6), and the combination of cell-level experiments with clinical data could enumerate the canonical molecular mechanisms underlying disease and explain how variations in such mechanisms drive disease progression (or lack thereof) in individual patients. In addition, models need to be well coupled to individual pharmacokinetic-pharmacodynamic profiles. Pharmacodynamic analysis is largely conducted early in studies of tolerability after the first introduction of a drug into humans. Given the potential role of the microbiome in controlling disease origination and progression as well as drug disposition and efficacy (and its dependence on diet and time of day), integration of such data may be necessary for pharmacological modeling of precision therapy.

Currently, data gathering, even for big data sets, is largely empirical and is not driven by computational models that can enable predictive simulations of dynamics. Modelers generally only provide input about experimental design relating to considerations of sample size to enable sound statistical correlations. However, what is needed is to build initial integrative models based on current knowledge of genomics and epigenomics and the relevant biochemistry and cellular-tissue physiology, to predict how the to obtain data from these very large cohorts (see the figure). Such predictions could specify which clinical parameters to measure and at what intervals. The clinical measurements could include readouts of environmental influences, such as metabolomics and proteomics, and quantitative and integrative measures of whole-body physiology (e.g., blood pressure, heart rate) and noninvasive imaging of organ functions, such as with functional magnetic resonance imaging. The results of model-directed experiments and clinical readouts would in turn be used to update the initial models, yielding better predictions of molecular, cellular, and organismal pathophysiology and providing directions for further experimental follow-up. This iterative approach to modeling should accelerate convergence to more accurate models by directing experiments to areas where additional information is most necessary.

It is likely that initial integrated models that drive further data gathering will only be partially correct, as our biochemical and cell physiological knowledge of health and disease is incomplete. However, initial poorly performing models can indeed converge to high-performance models through multiple rounds of experimentation and model refinement. A recent example of this comes from animal model studies of kidney disease progression, where better model predictions “treated” a kidney disease by directing drug therapy that reversed one cellular basis for the disease phenotype (7).

The million-person cohort envisaged by the Precision Medicine Initiative should provide an unprecedented wealth and breadth of data. Gathering these data should be driven by modeling analyses that capture the dynamics of disease evolution at multiple scales of organization. These data and models should enable predictions of efficacious therapies over time for an individual, making the aspiration of precision medicine a reality.


View Abstract

Stay Connected to Science

Navigate This Article