From Genomes to Societies: A Holistic View of Determinants of Human Health

Abstract

Both biological and social sciences have identified contributing factors to human health. However, health outcomes are unlikely to equal a simple sum of these identified factors. This article makes an attempt to put together the information, methods, and technologies that relate to health outcomes from biological, behavioral, and social disciplines. Much of this information was obtained by controlling for the variations of the factors in “other” disciplines. For example, genetic factors were controlled for in identifying the behavioral determinants of health. Looking forward, better understandings of health outcomes may require exploiting the interactions of health determinants that were identified from different disciplines. We propose the concept of “systems health” studies, which take health outcomes as the outputs of a system, where the inputs and their interactions from multiple disciplines are considered.

INTRODUCTION

Human health has been studied by inadvertently isolated disciplines, including biology, behavioral and social sciences. This reductionist approach effectively narrowed down the parameters to consider, facilitated technology developments, and led to the identification of important contributing factors in each discipline to human health. These achievements prompted this abridged overview of what is known and unknown about health. From prevention and intervention perspectives, it would be good to know what parameters in biological, behavioral, and social domains can be adjusted for promoting health outcomes. Towards this goal, a systems approach that considers multiple domains of information and their interactions is likely to be instrumental.

BIOLOGICAL DETERMINANTS OF HEALTH

Genomic factors associated with physiological traits and health outcomes

Early evidence of genetic contribution of health outcomes came from familial diseases, including cystic fibrosis [1–3] and others. These diseases are associated with and often caused by polymorphisms in the genome, sometimes a single nucleotide polymorphism (SNP). Not only the missense SNPs that changed protein coding sequences, but also the SNPs that changed cis-regulatory sequences such as enhancers can alter health outcomes [4]. These Mendelian genomic loci represent the best known genomic determinants of health outcomes [5,6]. Nevertheless, a larger variety of physiological traits are associated with combinations of alleles at multiple genomic loci, and thus considered complex traits [7]. Genome-wide association studies (GWAS) were carried out to identify risk loci of the genome to a variety of common diseases, including cardiovascular, mental, and autoimmune diseases [8–11]. Common physiological traits have also been mapped, including height, body mass index (BMI), HDL cholesterol, and others [11]. A recently added dimension of research is mapping the variability in drug responses to genomic loci and alleles [12]. SNP microarray has been the primary technical platform used in GWAS studies [13,14], partially due to its cost advantage to genome sequencing.

Personal genome sequencing allowed for identification of rare and common genomic variants associated with a disease. Sequencing based analyses were carried out around two major experimental designs. The first is the comparison of genomes within an individual. Two early studies on a skin cancer patient [15] and a lung cancer patient [16] sequenced and compared the “normal” genomes of their lymphoblast cells and the “abnormal” genomes of metastasized tumors. Comparing the normal and abnormal genomes of the same patient allowed for identification of cancer associated somatic mutations. Single-cell sequencing technology enabled sequencing the genomes or partial genomes of dozens and even hundreds of single cells within a tumor or a normal tissue [17,18]. Using the frequencies of a mutation or a copy number variation (CNV), multiple teams traced the earlier somatic genomic changes in the cancer cells [19–21]. These early changes were reasoned to be “driver” changes that have causal relationships with tumorigenesis [22]. Surprisingly, normal neurons in the brain possess mosaic CNVs within the same person, posing a challenge to redefine the concept of a “normal” genome [18].

The second major experimental design is to sequence individual genomes from a patient cohort, and then summarize the shared genomic variations among these patients. Examples include but not restricted to sequencing Autism cohorts [23–25] and cancer cohorts [26][27][28][29]. A major lesson learned from these studies is that there are much fewer shared mutations [23] or CNVs [24] among patients of similar diagnoses than previously expected. Considering the hallmarks of cancer are perturbations of molecular pathways, a bioinformatics approach was developed to aggregate mutations on genes of the same pathway and then cluster patients based on these aggregated impacts of mutations and molecular pathways [30]. This approach stratified prostate cancer patients into subgroups using genomic sequences alone, nevertheless these subgroups exhibited different survival time.

Beyond genomic sequences

Omics, the study of a total collection of a molecular species, revealed substantial amounts of health-related information beyond the genomes. The transcriptome, epigenome, proteome, and metabolome were four major data strata that exhibited strong associations with diseases. Before the accomplishment of the human genome project, genome-wide gene expression data were used to classify two types of leukemia (AML, ALL) [31]. Most strikingly the physicians changed their diagnosis on one child in this study, after seeing the transcriptome based results. With this proof of principle, gene expression differences have been extensively utilized to narrow down disease candidate genes.

The epigenome, the chemical modifications on histones and DNA, appears to strongly correlate with cellular behaviors and thus with health outcomes. The genome-wide distributions of as few as 1–3 histone modifications were effective in predicting prognosis of multiple cancer types [32]. Almost every cancer being analyzed exhibited hyper- and hypo- modified regions in the chromosome of the tumor cells [32–34]. Perhaps most strikingly, DNA methylation levels at a defined set of genomic locations are strongly predictive of human aging [35,36]. Mechanistically, the epigenome interacts with the genome to modulate the personal specific transcription factor binding, and thus directly contributes to personal variation of gene expression [37]. Moreover, the temporal changes of the epigenomic modifications (e.g. 5-hmC) during a biological process are predictive of gene expression changes [38]. Because the epigenome is jointly determined by the genomic sequence and the environmental signals, and the epigenome is likely to be less sensitive than the transcriptome to transient environmental changes, the epigenome may serve as a preferred molecular layer for quantifying personal responses to interventions.

Protein levels often correlate well with physiological outcomes, and sometimes with the subtypes of a disease. Breast cancer is a case in point. The protein levels of estrogen receptor, progesterone receptor, and Her2 are currently used to stratify patients for prognosis and treatment purposes. Thus, high-throughput protein identification technologies, especially mass spectrometry (mass-spec) hold the potentials for becoming diagnostic tools [39]. Serum proteomic profiling is a heavily pursued approach for identifying disease biomarkers, catalyzed by increasing sensitivity and specificity of mass spec.

Microbiome, our very close neighbors

“Within the body of a healthy adult, microbial cells are estimated to outnumber human cells ten to one” [40]. The diversity and the social structure of the microbial community in humans could hold the key to unexplained parts of health outcomes. The metagenomics methods that utilize deep genomic sequencing to identify the microbial species as well as their relative population sizes have started to reveal the personal variations of oral, nasal, skin, gastrointestinal, and urogenital microbiomes. Emerging data have suggested associations between the compositions of gut microbiome with obesity [41], inflammatory bowel diseases (IBDs) [42], colon cancer [43], sepsis [44], and other diseases [45]. Metagenomic analyses may offer a new approach to test outstanding hypotheses of disease etiology, including but not restricted to the hygiene hypothesis of asthma [46] and the intestine-toxin hypothesis of fatty liver diseases [47].

Environmental and behavioral impacts to biological determinants

Behaviors can impact health outcomes through modulating every layer of hitherto mentioned biological factors, including the genome, epigenome, transriptome, protein interactions and signal transduction, and gut microbiome (Figure 1). Sun tanning and smoking can stimulate somatic genomic mutations which are found in skin and lung cancers [15,16]. Smoking also induces DNA methylation at specific genomic loci [48]. Physical activities correlate with breast cancer survival at least partially by modulating the epigenome [49]. Early experience of children correlate with gene expression in the brain [50]. Dietary restriction may affect insulin receptor signaling and is reproducibly correlated with longevity [51]. Alcohol overconsumption results in the secretion of pro-inflammatory cytokines and alters gut microbiome, which may be causally linked to liver diseases [47]. We proceed to summarize the framework and methods for studying the behavioral and social determinants of health.

Figure 1.

Examples of behavior-related health outcomes that were mediated by biological factors.

BEHAVIORAL AND SOCIAL DETERMINANTS OF HEALTH

Conceptual frameworks

Humans are social beings. A typical social science approach to identifying health determinants is to analyze the correlation between the variations of social environment and that of the health outcomes. It is suggested that the “conditions in which people are born, grow, live, work and age” correlated to health outcomes, and “these circumstances are shaped by the distribution of money, power, and resources at global, national and local levels” [52]. Dahlgren and Whitehead in 1991 proposed a conceptual framework for summarizing social determinants of health [53], including four highly interconnected categories: 1) age, gender, and constitutional factors, 2) individual lifestyles, 3) social interactions and communities, and 4) socioeconomic, cultural and environmental conditions.

Social and economic development has produced physical products that can modulate behaviors, and in turn affect health outcomes. The man-made context for human activities has been formalized into the notion of built environment [54], which may include local facilities, infrastructures, and food environment. By incorporating the built environment, Barton, Grant and Guise categorized the determinants of health into seven groups [55].

Adding together the proposed factors by Dahlgren and Whitehead [53] and Barton et al. [55] would provide a relatively comprehensive list of behavioral and social factors that should be considered in health studies. However, from a prevention or intervention perspective, we recognize that not all factors are equal in their potentials to be changed for promoting better health. Some factors can be changed by personal determinations, whereas some may require policy and economic reforms. For this reason we propose a three-tier system to summarize and categorize behavioral and social factors that may contribute to health outcomes (Figure 2). The first tier includes the factors directly associated with the individuals. The second tier includes the local environment where individuals live and work. The third tier includes macro environmental factors. The factors within each tier can interact with each other and may independently or collectively interact with the factors of other tiers.

Figure 2.

Three layers of behavioral and social determinants of health. Adapted from Dahlgren & Whitehead 1991 and Barton et al. 2005.

The ongoing research of health-related behavioral and social studies can fit nicely into the framework as described above. An example is the studies of obesity, the prevalence of which has increased considerably during the past several decades [56]. Some studies assessed individual and local characteristics associated with obesity [57–60], and suggested that individual socioeconomic and lifestyle profile, local food access, and physical activity environment have significant impacts. Some other studies looked at macro level determinants (third tier), and reported that, price and taxation policies on food and other relevant products accounted for the variation in obesity in addition to individual level factors, and the level of urbanization mattered as well [61–64].

Analysis methods

In contrast to the flexibility of generating experimental data from model organisms, ethical and practical constraints limit the capacity of experimentations that involve human subjects. The ethical and practical constraints often require a correlation of the outcome and the factors of interest as preliminary evidence for trials. Moreover, such correlations would be likely to increase the cost effectiveness of the designed trials. Observational data, in which human subject are observed in natural circumstances that cannot be controlled, is therefore often used to detect correlations of health and explanatory factors, understand interactions of factors, and provide support for hypothesis testing. Examples of observational data include surveys that sample individual units from a population and ask a number of questions to the respondents at one time point (cross-sectional survey) or repeated time points (longitudinal survey); behavioral data extracted by electronic technologies such as pedometers, smartphones, or internet; government administrative or surveillance data; and data collected by public or private entities such as medical claims. Data from different sources can be merged to enable a more thorough look of the determinants at multiple levels of a given health outcome. Statistical power and representativeness of the data are essential, and they can be achieved by large sample size and careful sampling of the study units.

A well-known problem of observational data is the presence of confounding factors, which simultaneously contribute to the health outcome with the factors of interest [65]. To isolate the effects of the factors of interest, statistical and epidemiological approaches are often employed on the observational data to make the confounding factors “under control”. For example, multivariate regression analysis is often used to determine the function that describes how a vector of factors responds to the changes in others and how they collectively affect the health outcome of interest. Factor analysis and principle components analysis allow researchers to create a new set of synthetic variables and investigates the contribution of each set to the outcome [66]. Cluster analysis groups population into clusters so that individuals within cluster share more similar characteristics than individuals in other clusters. Multilevel models are appropriate when observational data are organized at multiple levels, such as individuals nested in local communities, which are further nested in states and countries [67]. Case control studies are used to compare two groups that differ in health outcomes but are otherwise similar, and cohort studies track a cohort of population over time to observe the development of health outcomes [65].

The inferences derived from observational data are often “correlational”, because confounding factors are not able to be observed in the observational data but are correlated with the health outcomes being studied [68]. Biological factors are typical “unobserved heterogeneities” acknowledged in behavioral and social sciences. Other examples are individual preferences, intellectual capabilities, and coping skills. To correct for omitted variable bias and establish “causal” inferences, experimental studies such as randomized trials are still the “gold standard” when applicable. Randomized controlled trials balance the observed and unobserved factors by randomly allocating human subjects to receive one or other of the alternative interventions. Any change between intervention and control group, therefore, is attributable to the intervention alone. The 1971–1982 RAND’s Health Insurance Experiment assigned thousands of people to different health insurance plans. It concluded that the cost sharing reduced medical expenditures, and in general, the reduction in service had no adverse consequences on participants’ health [69–71]. The result of this experiment has encouraged the restructuring of private insurance and the promotion in managed care.

When randomized controlled trials are not considered, quasi-experimental design is often perceived as an alternative [72]. The study design is similar to randomized controlled trials, but lacks the component of random assignment. As a result, the characteristics of groups are not equivalent at baseline and any differences observed after the intervention may not be solely due to the intervention received [72]. A few statistical and econometrical strategies have been developed to address this concern of internal validity. Developed by Paul Rosenbaum and Donald Rubin [73], propensity score matching attempts to statistically mimic the random assignment of the intervention by creating a control sample that has comparable observed characteristics to the intervention sample. Instrumental variable is an observed variable, usually policy changes, that does not affect health outcome itself, but is correlated with the endogenous explanatory variables conditional on other covariates. It allows consistent estimates of the regression relationship when the explanatory variables are correlated with the error terms [74]. In regression discontinuity design [75], a threshold is selected above or below which an intervention is assigned. The causal correlation between the intervention and the health outcome is obtained by comparing people lying closely on two sides of the threshold. Panel analysis utilizes longitudinal observational data which collects repeated measures over time and over the same individuals [68]. The unobserved factors, varying non-stochastically or stochastically over time, can be modeled or differenced out. The health behaviors and health outcomes of siblings or twins are compared assuming that genetic factors are shared and any variation observed would be attributable to environmental contexts.

Key modifiable behavioral and social determinants

Special interests were given to the health determinants that are modifiable by policies or interventions. Six key modifiable determinants have been proposed. These include modifiable health risk behaviors, namely physical inactivity, poor nutrition, tobacco use, and excessive alcohol use, account for much of the burden in morbidity and mortality (CDC). Education is another modifiable health determinant. Quasi-experimental studies have demonstrated causal influences of educational policies on health outcomes and that the improvement of education attainment can lead to improvement in health. Economic stability is causally correlated with health in two ways [76]: through a direct impact on material conditions and through an impact on social participation. Social environment such as social capital, family structure, discrimination and civic participation affects health outcome by modifying the inter-personal relationship, beliefs and perceptions. Health services, including access to care, insurance coverage, and quality of care, are services directly dealing with the diagnosis and treatment of disease, and the promotion, maintenance and restoration of health (WHO). Built environment modifies the physical environment that we live and work. It is associated with health behaviors such as physical activity and alcohol use, and health outcomes such as obesity and depression [77].

FUTURE DIRECTIONS: A SYSTEMS HEALTH APPROACH

A holistic view of “systems health”

An alternative to the reductionist approach of studying biological and social factors of health in separation is a holistic approach where “all things considered”. We take the liberty to propose a name “systems health” for this emerging interdisciplinary research area. Systems health concerns about any biological, behavioral, social factors as well as their interactions that affect health outcomes. Its major goal is to derive mechanistic and predictive models of health outcomes, and major applications are in prevention, intervention, and prognosis.

The central question of systems health is how gene-environment interactions relate to health. This is a Nature vs. Nurture question [78] with special emphases (Figure 3). First, human health is the outcome of interest. Second, the modifiable social and behavior factors receive special attentions. Third, interventions and preventative procedures can be thought of as feedbacks. One success of this type was made by Caspi and colleagues, who discovered that a polymorphism in the gene encoding for a serotonin transporter protein modulates the impact of life stress to depression [79]. This finding and others led to the idea that the heterogeneities of health outcomes are primarily determined by behavioral and social factors; however personal genomic variation modifies the extent of this correlation [80–82].

Figure 3.

“Systems health” studies gene-environment interactions, with special emphases on the health outcomes and the modifiable input factors.

Challenges in data collection

The major challenges of systems health studies lie in collecting interdisciplinary data and devising new research methods. Ethical and monetary cost issues need to be addressed. On the data end, a few population surveys on health care topics, including the repeated cross-sectional National Health and Nutrition Examination Survey, have incorporated laboratory module to collect respondents’ physiological data. Add Health, a survey that followed approximately 20,000 US adolescents through their adulthood, is currently genotyping 12,000 subjects of its cohort [83].

Electronic equipment including pedometers, personal digital assistants (PDA), and portable personal computers have started to contribute data to behavioral science, opening the possibility of collecting personal data at real time. Electronic data of personal physical, dietary, and social networking activities are among the first to be reported [84–86]. These devices and systems may reduce monetary and time costs of traditional surveys [87], but will certainly bring new challenges in data handling and analyses.

Candidate approaches for “systems health” studies

It is likely too early to speculate the most effective methods, but two general approaches showed promises. The first is to leverage the knowledge of the other disciplines. An example of the success is the process of revealing the mechanistic link between smoking and lung cancer. The prevalence of lung cancer mortality increased dramatically from 5.3% of cancer deaths in men and 2.0% in women in the 1930s to 33.1% in men and 22.8% in women in the 1990s [88], posting a great puzzle to health researchers. Epidemiological studies reported that cigarette consumption, increased from 54 cigarettes per capita per year to 4166 in 1960s, has significantly contributed to the lung cancer epidemic. These evidences led to the first U.S. Surgeon General’s Report in 1964 that “cigarette smoking is causally related to lung cancer in men” [89]. Guided by this knowledge, the genomic mutations resulted from tobacco exposure were analyzed and documented [16], revealing the chain of events from smoking to genomic mutations and then to cancer development.

Biological knowledge may lend a hand to behavioral and social research of health outcomes as well. It has been four decades since the 1973 report [90] on the effects of shift work on worker health [90–94]. The next steps may take advantage of the biological findings on gene expression [95] and metabolic outcomes [96] of the perturbation of circadian rhythms and their mechanistic links to depression and obesity. Furthermore, the data on the role of Melatonin on the regulation of human circadian rhythms and sleep [97] are perhaps worth considering in future contemplation of intervention strategies.

Another approach is stratification, where the information of one discipline is used to stratify the data analysis of another discipline. For example, stratifying the recruits by the polymorphisms of their neural transporter genes may increase the chances of detecting associations in a case control study of stress and depression where the sample size is limited by monetary cost. A drawback of stratification is the potential difficulty of recruiting sufficient respondents in a stratum. This difficulty may have to be addressed by an integrated analysis.

Integrated data analysis may deploy machine learning methods that take both genomic and survey data as inputs to learn information about the variation of health outcomes. These methods are relatively mature as long as the desired data are collected for all subjects. Incomplete data will likely to be a major challenge, in which case some individuals would only have either social or genomic data. The computational methods that aim for unbiased estimates under the missing data scenario can be useful [98].

HIGHLIGHTS.

• “Systems health” is an emerging interdisciplinary field.

• Systems health integrates biological, behavioral, social factors and their interactions.

• Special emphases are given to health outcomes and modifiable social and behavior factors.

• Intervention and preventative procedures can be regarded as feedbacks.