Translational Bioinformatics: Past, Present, and Future.

Though a relatively young discipline, translational bioinformatics (TBI) has become a key component of biomedical research in the era of precision medicine. Development of high-throughput technologies and electronic health records has caused a paradigm shift in both healthcare and biomedical research. Novel tools and methods are required to convert increasingly voluminous datasets into information and actionable knowledge. This review provides a definition and contextualization of the term TBI, describes the discipline's brief history and past accomplishments, as well as current foci, and concludes with predictions of future directions in the field.

Introduction

Though a relatively young field, translational bioinformatics has become an important discipline in the era of personalized and precision medicine. Advances in biological methods and technologies have opened up a new realm of possible observations. The invention of the microscope enabled doctors and researchers to make observations at the cellular level. The advent of the X-ray, and later of magnetic resonance and other imaging technologies, enabled visualization of tissues and organs never before possible. Each of these technological advances necessitates a companion advance in the methods and tools used to analyze and interpret the results. With the increasingly common use of technologies like DNA and RNA sequencing, DNA microarrays, and high-throughput proteomics and metabolomics, comes the need for novel methods to turn these new types of data into new information and that new information into new knowledge. That new knowledge, in turn, gives rise to action, providing insights regarding how to treat disease and ideally how to prevent it in the first place.

Translational bioinformatics

Defining translational bioinformatics

According to the American Medical Informatics Association (AMIA), translational bioinformatics (hereafter “TBI”) is “the development of storage, analytic, and interpretive methods to optimize the transformation of increasingly voluminous biomedical data, and genomic data, into proactive, predictive, preventive, and participatory health” (http://www.amia.org/applications-informatics/translational-bioinformatics). Put more simply, it is the development of methods to transform massive amounts of data into health. Dr. Russ Altman from Stanford University delivers a year-in-review talk at AMIA’s summit on TBI. In his 2014 presentation he provided the following definition for TBI: “informatics methods that link biological entities (genes, proteins, and small molecules) to clinical entities (diseases, symptoms, and drugs)—or vice versa” (https://dl.dropboxusercontent.com/u/2734365/amia-tbi-14-final.pdf). Figure 1 gives a visual depiction of the way in which TBI fits within the bigger picture of biomedical informatics and transforming data into knowledge [1]. Along the X axis is the translational spectrum of bench-to-bedside, while the Y axis from top to bottom represents the central dogma of informatics, transforming data to information and information to knowledge. Toward the discovery end of the spectrum (the bench) is bioinformatics, which includes storage, management, analysis, retrieval, and visualization of biological data, often in model systems. The discovery end of the spectrum has some overlap with computational biology, particularly in the context of systems biology methods. Toward the clinical end of the spectrum (bedside) is health informatics. TBI fits in the middle of this space. On the data-to-knowledge spectrum, data collection and storage are the beginning steps. After that comes data processing, analysis, and then interpretation, thereby transforming the information that has been gleaned from the data into actual knowledge, useful in the context of clinical care, or for further research. In that way the data go from just being “bits”—1’s and 0’s—to new knowledge and actionable insights.

Figure 1

Translational Bioinformatics in context

The Y axis depicts the “central dogma” of informatics, converting data to information and information to knowledge. Along the X axis is the translational spectrum from bench to bedside. Translational bioinformatics spans the data to knowledge spectrum, and bridges the gap between bench research and application to human health. The figure was reproduced from [1] with permission from Springer.

Where do we come from? A relatively short past

TBI as a field has a relatively short history. In the year 2000, the initial drafts of the human genome were released, arguably necessitating this new field of study (http://web.ornl.gov/sci/techresources/Human_Genome/project/clinton1.shtml). In 2002, AMIA held its annual symposium with the name “Bio*medical Informatics: One Discipline”, meant to recognize and emphasize the spectrum of subdisciplines. In 2006, the term itself was actually coined by Atul Butte and Rong Chen at the AMIA annual symposium in a paper entitled “Finding disease-related genomic experiments within an international repository: first steps in translational bioinformatics” [2]. In 2008, AMIA had its first annual AMIA summit on TBI, chaired by Dr. Butte. Year 2011 saw the first annual TBI Conference in Asia, held in Seoul, Korea. Finally, an online textbook on TBI was published in 2012 by PLoS Computational Biology, edited by Maricel Kann. Initially intended to be a traditional print textbook, this resource was published using an open source model, making it freely available on the Internet (http://collections.plos.org/translational-bioinformatics).

What are we? TBI today

A 2014 review article [3] categorized recent themes in the field of TBI into four major categorizations: (1) clinical “big data”, or the use of electronic health record (EHR) data for discovery (genomic and otherwise); (2) genomics and pharmacogenomics in routine clinical care; (3) omics for drug discovery and repurposing; and (4) personal genomic testing, including a number of ethical, legal, and social issues that arise from such services.

Big data and biomedicine

As technology enables us to take an increasingly comprehensive look across the genome, transcriptome, proteome, etc., the resulting datasets are increasingly high-dimensional. This in turn requires a larger number of samples in order to achieve the statistical power needed to detect the true signal. The past decade or so has seen an increasing number of large-scale bio-repositories intended for clinical and translational research all over the world. These projects comprise both information and biospecimens from individual patients, enabling researchers to reclassify diseases based on underlying molecular pathways, instead of the macroscopic symptoms that have been relied on for centuries in defining disease. Examples are listed in Table 1. These various projects involve different models of participation, ranging from explicit informed consent to use of de-identified biospecimens and their associated clinical information from EHRs (also de-identified). The informed consent is the most ethically rigorous model, but also the most expensive. The use of de-identified specimens and data is more scalable, and financially feasible. However, as complete genomic data are increasingly used, it is impossible to truly de-identify these data [4]. This raises ethical issues regarding patient privacy and data sharing. In the United States, legislation known as the “Common Rule” addresses these issues. In 2015, a notice of proposed rule-making (NPRM, http://www.hhs.gov/ohrp/humansubjects/regulations/nprmhome.html) was released to solicit feedback on some major revisions to the law, which was originally passed in 1991. Much has changed within biomedical research in the intervening years [5].

Table 1

Large-scale research initiatives integrating human specimens with clinical annotation

Name	Link	Description
Million Veteran Program	http://www.research.va.gov/mvp/veterans.cfm	US Veterans Affairs (VA)-sponsored research program to partner with veterans to study how genes affect health
Personal Genome Project	http://www.personalgenomes.org/	Based at Harvard University with an emphasis on open access sharing of genomic, environmental, and human trait data
MURDOCK Study	http://www.murdock-study.com	A community-based registry and biorepository aimed at reclassifying disease based on molecular biomarkers
UK Biobank	http://www.ukbiobank.ac.uk	UK-based national health resource aimed at improving the prevention, diagnosis, and treatment of disease
Genomics England	http://www.genomicsengland.co.uk/	Company formed to sequence samples in the UK-based 100,000 Genomes Project, focused on rare diseases, cancer, and infectious disease
Framingham Heart Study	https://www.framinghamheartstudy.org/	A long-term, ongoing study started in 1948, based in Framingham, Massachusetts. The study is now on its third generation of participants
China Kadoorie Biobank	http://www.ckbiobank.org/site/	Focused on genetic and environmental causes of common chronic diseases in the Chinese population
Kaiser Permanente/UC San Francisco Research Program on Genes, Environment, and Health	https://www.dor.kaiser.org/external/DORExternal/rpgeh/index.aspx	A collaborative resource that will link electronic medical records, behavioral and environmental data, and biospecimens to examine the genetic and environmental factors that influence common diseases
Google Baseline	http://www.wsj.com/articles/google-to-collect-data-to-define-healthy-human-1406246214∗	Designed to collect numerous different types of clinical and molecular data to help define what a “healthy” individual looks like
US Precision Medicine Cohort	http://www.nih.gov/precision-medicine-initiative-cohort-program	A United States population-based research cohort that aims to engage a million or more volunteers over many years to improve health outcomes, fuel new disease treatments, and catalyze precision medicine
The eMERGE Consortium	https://emerge.mc.vanderbilt.edu	An NIH-funded national research network that combines DNA biorepositories with EHRs for large-scale, high-throughput genetic research to enable genomic medicine
National Biobank of Korea	http://www.nih.go.kr/NIH/eng/contents/NihEngContentView.jsp?cid=17881	A collection of well-annotated, high quality human biospecimens for distribution to Korean scientists, and to facilitate international cooperation toward personalized medicine
Estonian Biobank	http://www.geenivaramu.ee/en/access-biobank	An Estonian population-based cohort recruited at random by physicians. Significant data beyond medical information is collected: places of birth and living, family history spanning four generations, educational and occupational history, physical activity, dietary habits, smoking, and alcohol consumption, among others

Note: ∗a database weblink for Google Baseline is not available; the link to a news report about the project is provided instead. eMERGE, electronic medical records and genomics; EHR, electronic health record.

In order to accrue the numbers of samples required for the “big data” discipline that biomedical research is becoming, the ability to use patient data and samples in research would be of significant benefit. One major point addressed in the aforementioned NPRM is the ability for patients to give broad consent for future use of data and samples, without knowing the specifics of research studies ahead of time. As we move toward the learning healthcare system (LHS) model [6] in which every encounter is an additional data point, explicit research registries will become less relevant. They will be too expensive to maintain, and larger numbers of patients/participants will be available through federated initiatives that allow a researcher to query across institutions regionally, nationally, and even internationally. The National Patient-Centered Clinical Research Network (PCORnet) takes this approach, enabling clinical outcome research through federated pragmatic clinical trials. Importantly, this initiative emphasizes partnership with patients and their advocates, so that they are empowered as collaborators, with a say in what research questions matter most [7].

LHS is about moving from evidence-based practice, i.e., clinical care decisions based on conscientious use of current best evidence, to practice-based evidence, i.e., the generation of evidence through collection of data in the real-world as opposed to the artificially-controlled environment of randomized clinical trials [8]. In recent decades, the biomedical enterprise has strived to practice medicine in a way that is supported by the best possible evidence from randomized clinical trials. But clinical trials have their own issues. They are expensive, and they tend to be very different from real life scenarios [9]. Criteria for inclusion in a trial often include the absence of common comorbidities or use of common medications. Compliance tends to be high, but the cohort being studied is often not representative of the target population for the treatment in question. In LHS, translation becomes bi-directional. Research is used to inform practice, whereas data that are generated in the course of clinical care can in turn be used for both hypothesis generation and validation through pragmatic trials. Data derived from clinical care can thus inform clinical guidelines and future practice.

Secondary use

Secondary use of data refers to data that are created or collected through clinical care. In addition to use in caring for the patient, these data may also be crucial for operations, quality improvement, and comparative effectiveness research. Some assert that the term “secondary use” should give way to the term “continuous use.” They argue against the notion that data collected at the point of care are solely for clinical use, and everything else is secondary. We should be maximally leveraging this valuable information. Nonetheless, there is a legitimate concern about data quality. Data in the EHR are often sparse, incomplete, even inaccurate [10]. This makes these data wholly unsuitable for certain purposes, but still sufficient for others. For instance, Frankovich et al. described a case in which an adolescent lupus patient was admitted with a number of complicating factors that put her at risk for thrombosis [11]. The medical team considered anti-coagulation, but were concerned about the patient’s risk of bleeding. No guideline was available for this specific case, and a survey of colleagues was inconclusive. Through the institution’s electronic medical record data warehouse, Frankovich and colleagues were able to look at an “electronic cohort” of pediatric lupus patients who had been seen over a 5-year period. Of the 98 patients in the cohort, 10 patients had developed clots, with higher prevalence in patients with similar complications as the patient in question. Using this real-time analysis based on evidence generated in the course of clinical care, Frankovich and colleagues were able to make an evidence-based decision to administer anti-coagulants [11]. Subsequently, researchers at Stanford University have proposed a “Green Button” approach to formalize this model of real-time decision support derived from aggregate patient data and data capture to help inform future research and clinical decisions [12].

TBI tends to focus on molecules, newly accessible in high dimensions based on novel high-throughput technologies. Phenotyping is a closely-related challenge, more complex than it might seem. Disease is not binary: even within a very specific type of cancer, a tumor’s genomic profile may be quite different among the precise sampling locations and sizes [13]. There are a number of groups focusing on this problem: the Electronic Medical Records and Genomics (eMERGE) Network (http://www.genome.gov/27540473), the NIH Collaboratory (https://www.nihcollaboratory.org/), PCORnet (http://www.pcornet.org/), and the MURDOCK Study (http://murdock-study.com/), among others [14], [15], [16], [17]. The Phenotype KnowledgeBase website (https://phekb.org/) is a knowledge base of phenotypes, offering a collaborative environment to build and validate phenotype definitions. The phenotypes are not (yet) computable, but it serves as a resource for defining patient cohorts in specific disease areas [18]. Richesson et al. looked at type 2 diabetes, a phenotype that one might expect to be fairly straightforward [19]. But defining type 2 diabetes mellitus (T2DM) using the International Classification of Disease version 9 (ICD9) codes, diabetes-related medications, the presence of abnormal labs, or a combination of those factors resulted in very different counts for the number of people diagnosed with T2DM in Duke’s data warehouse [19]. Using only ICD9 codes gave 18,980 patients, while using medications yielded 11,800. Using ICD9 codes, medications, and labs all together yielded 9441 patients. Note that the issue is not just a matter of semantics and terminology, where if everyone could agree to a single definition and use the same code, then the terms would become uniform. For different purposes, different definitions of diabetes may be needed, depending on whether the use case involves cohort identification or retrospective analysis. In different cases, one might care more about minimizing false positives (e.g., retrospective analysis) or maximizing true positives (e.g., surveillance or prospective recruitment).

Thousands of papers have been published describing genome-wide association studies (GWAS), in which researchers look across the entire genome to find SNPs that are statistically enriched for a given phenotype (usually a disease) compared with healthy controls [20]. Researchers at Vanderbilt University turned this approach on its head, developing a method known as phenome-wide association studies (PheWAS, https://phewas.mc.vanderbilt.edu/). Instead of looking at the entire genome, PheWAS evaluates the association between a set of genetic variants and a wide and diverse range of phenotypes, diagnoses, traits, and/or outcomes [21]. This analytic approach asks, for a given variant, do we see an enrichment of a specific genotype in any of these phenotypes? Figure 2 illustrates results using this approach [22]. In standard GWAS analyses, the different color bands at the bottom represent the different chromosomes. In the case of PheWAS, they are different disease areas, e.g., neurologic, cardiovascular, digestive, and skin. Pendergrass et al. [23] used a PheWAS approach for the detection of pleiotropic effects, where one gene affects multiple different phenotypes. They were able to replicate 52 known associations and 26 closely-related ones. They also found 33 potentially-novel genotype−phenotype associations with pleiotropic effects, for example the GALNT2 SNP that had previously been associated with HDL levels among European Americans. Here they detected an association between GALNT2 and hypertension phenotypes in African Americans, as well as serum calcium levels and coronary heart disease phenotypes in European Americans.

Figure 2

A PheWAS Manhattan plot for a given SNP

This plot shows the significance of association between SNP rs965513 and 866 different phenotypes. Along the X axis different disease groups are shown in different colors. This is in contrast to an analogous plot for GWAS in which the X axis would represent the different chromosomes. The Y axis reflects the P value for each phenotype. Blue and red horizontal lines represent P value of 0.05 and Bonferroni corrected P value of 5.8 × 10−5, respectively. PheWAS, phenome-wide association studies; GWAS, genome-wide association studies. The figure was reproduced from [22] with permission from Elsevier.

Another aspect of big data in biomedicine is the use of non-traditional data sources. These were well illustrated, both literally and figuratively, in a 2012 paper by Eric Schadt [24]. A complex and detailed figure (Figure 3) showed various data types that could be mined for their effects on human health: weather, air traffic, security, cell phones, and social media among others. But strikingly to those reading the paper just a few years later, the list did not include personal activity trackers, e.g., FitBit, Jawbone, or even the Apple watch. This omission of such a popular technology today is indicative of what a fast-moving field this is.

Figure 3

Heterogeneous and non-traditional sources of big data

Technological advances have enabled the collection and storage of big data beyond biomedicine, including everything from credit card transactions to security cameras to weather. Notably absent from this 2012 figure are smart watches and fitness tracking devices, which became pervasive in the years that followed. The figure was reproduced from [24] under Creative Commons Attribution license.

Genomics in clinical care

One sees a number of examples of how genomic data are used in clinical care in the context of pharmacogenomics [25]. But molecular data, and genomic data derived from next-generation sequencing (NGS) in particular, have been used in a number of other contexts as well. One example took place at Stanford’s Lucile Packard Children’s Hospital, where a newborn presented with a condition known as long QT syndrome. (http://scopeblog.stanford.edu/2014/06/30/when-ten-days-a-lifetime-rapid-whole-genome-sequencing-helps-critically-ill-newborn/). In this specific case, the manifestation was unusually severe—the baby’s heart stopped multiple times in the hours after its birth. Long QT syndrome can be caused by mutations in a number of different genes [26]. It is necessary to know which gene harbors the mutation in order to know how to treat the condition [26]. In this case, a whole-genome sequencing (WGS) was performed enabling identification of a previously-studied mutation, as well as a novel copy number variation in the TTN gene that would not otherwise have been detectable through targeted genotyping alone. Moreover, NGS enabled the answer to be obtained in a matter of hours to days instead of weeks.

Another example of DNA sequencing in clinical care involved diagnosis of infectious disease. A 14-year-old boy with severe combined immunodeficiency (SCID) had been admitted to the hospital repeatedly. He had headache, fever, weakness, nausea, and vomiting. His condition continued to decline to the point where he was put into a medically-induced coma. A normal diagnostic work up was “unrevealing” and the doctors were unable to determine the etiology of his condition. The patient was enrolled in a research study for pathogen detection and discovery in hospitalized patients. The protocol for this study involved performing NGS on the subject’s spinal fluid. The results included detection of 475 reads corresponding to leptospira infection. Of note, the normal test for leptospirosis involves detection of the patient’s antibody response to the infection. In this case, the patient’s SCID status prevents such a response, so the infection was not detectable through standard means [27].

Cancer has been one of the most active areas for the translation of genomic discoveries into changes in clinical care. One of the biggest players in this space is Foundation Medicine, which makes the FoundationOne test, a targeted panel that uses NGS to test all genes known to harbor mutations in solid tumors. Johnson et al. [28] looked retrospectively at approximately 100 patients who had undergone FoundationOne, and found that 83% of them had potentially actionable results from that test, of which 21% received genotype-directed treatment. Explanations for why the indicated treatment was not given in 79% of those cases included the decision to use standard therapy and clinical deterioration. These results indicate that application of genomic technology has transcended the research domain. In many cases, the findings are clinically actionable. Mirroring this fact, it is worth noting that some medical insurance companies in the United States started to offer coverage of the FoundationOne test in late 2014 (http://www.clinicalinformaticsnews.com/2014/10/17/foundation-medicine-wins-insurance-coverage-priority-health.html).

In addition to targeted panels, deep cancer sequencing can help shed light on drug sensitivity and resistance. Wagle et al. [29] describe a case of a 57-year-old woman with anaplastic thyroid cancer. Her specific tumor was initially sensitive to everolimus, a mammalian target of rapamycin (mTOR) inhibitor. Her doctors were able to sequence the tumor before it became resistant, revealing a mutation in TSE2, which encodes a negative regulator of mTOR. Normally mTOR is down-regulated by TSE2, but the mutation caused TSE2 not to down-regulate mTOR to a sufficient level. Therefore, everolimus, which inhibits mTOR, was effective in treating this specific cancer. Later this patient’s cancer became drug-resistant, whereupon the newly drug-resistant part of the tumor was sequenced again. It was discovered that an mTOR mutation had cause mTOR not to be inhibited by allosteric inhibitors like everolimus. An allosteric inhibitor binds to the protein in question somewhere other than on the active site of the molecule. Knowing the specific cause for the newly-acquired resistance leads to other treatment options. They were able to switch to mTOR kinase inhibitors, in order to down-regulate the pathway through other mechanisms.

Despite these compelling cases, it is worth noting that at this point, tumor sequencing is the exception in clinical oncology, and not one of the routine procedures. Only after “first line” treatment has failed are tumors sequenced, and even that is largely confined to large academic medical centers. It will be interesting to see if, when, and how that changes.

While blood clotting and cancer are the areas where most actionable pharmacogenomics findings have been made, a notable exception is work by Tang and his colleagues [30], in which they describe the identification of a genotype-based treatment for T2DM with an alpha2A-adrenergic receptor antagonist. A specific genetic variant causes over-expression of the alpha2A-adrenergic receptor, and impaired insulin secretion. They hypothesized that if they could block the over-expressed receptor, they could increase insulin secretion. The authors looked at patients with and without the mutation in question. Those with the mutation showed a dose response to the drug as measured through levels of insulin. Participants without the mutation showed no such response. This research is especially interesting because it goes beyond cancer and blood thinners, into a chronic disease that is affecting an increasing portion of the world’s population.

One of the earliest and best known DNA sequencing success stories was that of Nic Volker, who at age 2 developed severe gastro-intestinal issues resembling Crohn’s disease, for which a diagnosis could not be determined. His story was covered in the Milwaukee Journal Sentinel in 2011 in a piece that was later awarded a Pulitzer Prize for explanatory reporting (http://www.jsonline.com/news/health/111641209.html). After multiple life-threatening situations and a protracted diagnostic odyssey, Volker’s genome was sequenced to look for a causal mutation. Doing so enabled the discovery of a mutation in the gene X-linked inhibitor of apoptosis (XIAP). Equally important, there existed a known therapeutic intervention for disorders caused by XIAP mutations. A progenitor cell transplant was performed and the patient’s condition improved. Though Nic is not without ongoing health challenges, he celebrated his 11th birthday in October, 2015.

Despite the successes described above, there is also reason for caution regarding use of NGS sequencing in clinical care. Dewey et al. [31] performed Illumina-based WGS on 12 different participants. Confirmatory sequencing was formed on 9 of those participants by Complete Genomics Inc. Their findings included the fact that 10%–19% of inherited disease genes were not covered to accepted standards for SNP discovery. For the genotypes that were called, concordance between the two technologies for previously-described single nucleotide variants was 99%–100%. However, for insertion and deletion variants, the concordance rate was only 53%–59%. Approximately 15% were discordant, and approximately 30% of the variants could not be called by one technology or the other. In addition, inter-rater agreement for whether findings should be followed up clinically was only 0.24 (“fair”) by Fleiss’ kappa metric. Rater agreement was even worse for cardio-metabolic diseases, with the rate at which the two experts agreed on the need for clinical follow-up worse than random in those cases. Notably, the estimated median cost for sequencing and variant interpretation was about US $15,000, plus the price of the computing infrastructure and data storage. This means the cost of interpretation is significantly more than the proverbial US $1000 genome goal, but is also significantly less than the US $100,000 or $1 million some had feared [32].

Omics for drug discovery and re-purposing

Much has been said about the protracted process involved in getting a drug through the FDA approval pipeline. Estimates are that the process can take on average 12 years between lead identification and FDA approval. This makes the prospect of drug repurposing an appealing one. Drug repurposing refers to taking an existing, FDA-approved compound and using it to treat a disease or condition other than the one for which it was originally intended [33]. In the past, inspiration for this type of “off label use” has been largely serendipitous. For example, Viagra was initially aimed at treating heart disease, and turned out to be useful for erectile dysfunction [34]. By using a pre-approved compound, early phase clinical trials can be avoided, which can save significant time and money.

Computational approaches to drug repurposing may take a number of different forms as described in two recent reviews [33], [35]. One is to look for molecular signatures in disease and compare those to signatures observed in cells, animal models, or people who have been treated with different drugs. If anti-correlated signatures can be identified between diseases and drugs, administration of that drug for that disease may help cure the condition, or at least to alleviate the symptoms. One of the prominent early examples of a computational approach to drug repositioning was the Broad Institute’s Connectivity Map (CMap) [36]. The authors identified gene expression signatures for disease states and perturbation by small molecules and then compared those signatures. They made the data available as a resource intended to enable the identification of functional connections between drugs, genes, and diseases. Another example is work by Jahchan et al. [37], in which they identified anti-depressants as potential inhibitors of lung cancer. The authors looked at numerous disease and drug profiles and found an anti-correlation between gene expression seen after administration of anti-depressants and the pattern of expression observed in small cell lung cancer. They next transplanted these types of tumors into mice and found that with the drug, those tumors either shrunk or didn’t even grow. They also used indigenous tumors in the mouse model for this type of cancer and found that the anti-depressant had promising results on those tumors and for other endocrine tumors as well. They were able to start a clinical trial, much faster than would have been possible by chance or by various other traditional methods, though unfortunately that trial was ultimately terminated for lack of efficacy. A similar approach was used by Sirota et al. to identify the anti-ulcer drug cimetidine as a candidate agent to treat lung adenocarcinoma and validate this off-label usage in vivo using an animal model of the lung cancer [38].

An alternative to the largely computational methods described above is an experimental approach to drug repositioning. For example, Nygren et al. screened 1600 known compounds against 2 different colon cancer cell lines [39]. They used Connectivity Map data to further evaluate their findings, and identified mebendazole (MBZ) as having potential therapeutic effect in colon cancer. Finally, Zhu et al. mined data from PharmGKB [40] and leveraged the web ontology language (OWL) to perform semantic inference. They were able to identify potential novel uses and adverse effects of approved breast cancer drugs [41].

Personalized genomic testing

The year 2008 saw the founding of several companies that offered direct-to-consumer (DTC) genetic testing, reporting on a variety of genes for both health and recreational purposes. As of 2016, 23andMe was the last major player standing in the United States, with other companies having been acquired and/or changing their business models.

DTC genetic testing raises a number of interesting ethical, legal, and social issues. For several years, there was an open question as to whether or not these tests should be subject to government regulation. In November 2013, the US FDA ordered 23andMe to stop advertising and offering their health-related information services. The FDA considered these tests to be “medical devices” and as such to require formal testing and FDA approval for each test. In February 2015, it was announced that the FDA had approved 23andMe’s application for a test for Bloom syndrome (http://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/UCM435003), and in October 2015 it was announced that the company would once again be offering health information in the form of carrier status for 36 genes [42]. Note that a 23andMe customer is able to download his or her raw genomic data and to use information from other websites to interpret the results, including Promethease (http://www.snpedia.com/index.php/Promethease), Geneticgenie (http://geneticgenie.org/), openSNP (https://opensnp.org/), and Interpretome (http://interpretome.com/) for health-related associations.

Another important question raised by DTC genetic testing include whether the consumers are ready for this information. Traditionally, patients receive troubling health-related information in a face-to-face conversation with their doctor. There is some concern that patients are not competent or well-equipped to receive potentially distressing news through an Internet link [43]. To help mitigate this concern, 23andMe “locks” certain results, making them accessible only if the user clicks through an additional link, indicating they truly want to know.

What about the healthcare providers? Are they ready to incorporate genomic data, patient-supplied or otherwise, into treatment decisions? In a case described in 2012, a 35-year-old woman informed her fertility care provider that her 23andMe results revealed a relatively common (1 in 100) blood clotting mutation [44]. She was surprised when her provider responded by saying that, were she to become pregnant, she would need to be put on an injectable anti-coagulant throughout her pregnancy. With no family history of blood clotting disorders, nor personal history of recurrent miscarriage, this mutation would have gone untested, had it not been for the DTC results. However, when this patient did become pregnant and consulted with a specialist whose expertise was in blood clotting disorders in pregnant women, the anti-coagulant was indeed prescribed. Note that the guidelines have since changed, and the prophylactic treatment for clotting would not be prescribed today. It is also worth noting that in the United States, the Genetic Information Nondiscrimination Act (GINA) prevents employers and health insurers from discriminating against anyone based on their genetic information. It does not, however, cover long-term care, disability, or life insurance. Therefore when this woman applied for life insurance after her twins were born, the rate she was offered was more than twice what it would have been had she not known about the blood clotting mutation and been treated prophylactically for the increased risk it conferred [Tenenbaum and colleagues, unpublished data].

A more positive example of where genetic testing is helping patients is a case presented at the American Neurological Association conference in 2014. A patient had a history of Alzheimer’s disease on her mother’s side of the family. She did not know if she was a carrier, nor did she want to know. But she wanted to ensure that she did not pass that mutation to her future children. Preimplantation genetic diagnosis (PGD) testing enabled her doctors to select embryos that did not have that Alzheimer’s disease gene mutation. The patient herself was never tested, nor was she informed how many (if any) of the embryos contained the mutation. (http://www.wsj.com/articles/genetic-testing-for-alzheimerswithout-revealing-the-results-1413221509).

Privacy

A 2013 Science paper from the Erlich lab at Massachusetts Institute of Technology (MIT) generated much controversy when the authors demonstrated the ability to re-identify a number of individuals using publicly-available genealogy databases and genetic data [45]. The researchers used the short tandem repeat (STR) data from the Y chromosome, year of birth, and state of residence, combined with information from public genealogy websites to identify individuals. They did this by starting with publicly-available STRs and entering them on genealogy websites to identify matches. Note that their accuracy was not 100%. Though they were able to identify Craig Venter based on his genomic data, they failed to identify several other individuals, particularly those with more common last names. Overall, they reported a 12% success rate in recovering surnames of US males. They were also able to reconstruct Utah family pedigrees based on 1000 Genomes Project data and other publicly-available sources. Due to a number of cultural and historical factors, families in Utah tend to be large and genealogically well-documented, and an unusually high proportion of the Utah population has participated in scientific studies involving genomic data. It is worth noting that the researchers did not release any names that were not previously public, nor did they use the information for any nefarious purposes. Their primary interest was to demonstrate that such re-identification was possible. Interestingly, and somewhat surprisingly, the terms of use for the datasets did not prohibit re-identification.

Where are we going? The road ahead

Though we cannot know what the future holds, we can make some informed guesses based on events to date. The author believes that in the not-too-distant future, newborns will be sequenced at birth, just as we currently test for a more limited number of genetic issues. With the cost of sequencing a genome still at or above US $1000, such widespread sequencing is not yet realistic. But researchers are already performing pilot studies in this area, to better understand and anticipate the issues that are likely to arise. As an example, the MedSeq project at Harvard University is a study designed both to integrate WGS into clinical care and to assess the impact of doing so [46]. In addition, the Geisinger Health System has partnered with Regeneron on a project known as MyCode, which aims to sequence the exomes of 250,000 patients in the Geisinger system. In late 2015, the project began returning results to patients for 76 genes (https://www.genomeweb.com/sequencing-technology/geisinger-begins-returning-clinically-actionable-exome-sequencing-results).

Even clearer is that tumor sequencing will be performed as part of standard of care for cancer. Currently sequencing is performed at certain tertiary and quaternary care facilities, particularly for metastatic tumors. As more is discovered about the various dysregulated pathways in cancer, and about the therapeutic implications for different genetic variations, the blunt mallet that is chemotherapy will be phased out in favor of far more precise and targeted therapies.

The microbiome has seen increasing attention in recent years, a trend that will certainly continue for the foreseeable future. It is not surprising that the make-up of the microbial communities that likely outnumber the cells of the human body [47] can have significant impact on human health, particularly in metabolic and gastro-intestinal disease. The more surprising trend is the connection between the microbiome and other, more unexpected phenotypes, for example, anxiety, depression, and autism [48], [49]. This is likely to continue as this area of research continues to grow.

We will continue to see an increase in analyses of different “omic” types. Genomics has been by far the most popular area of focus to date. As technologies mature, we will continue to see biomarker discovery in proteomics, metabolomics, and other as-yet-unnamed “omic” modalities. We will also see increasingly integrated analysis, taking a systems approach to human biology where to date systems biology has been focused on model organisms, often single cellular ones, in which the system can be methodologically, and ethically, perturbed. Early examples of integrative, multi-modal analysis include the integration of microRNAs and transcription factors to determine regulatory networks underlying coronary artery disease [50], integrative analysis of genomics and transcriptomics to look at cardiovascular disease [51], and the use of metabolomics data with GWAS to elucidate molecular pathways [52].

The coming decade will see more biomarker-based research and insights into mental health disorders. To date, cancer and cardiology have received significant attention, to great advantage. But those disease areas are, by comparison, relatively easy to identify, to distinguish, and even to quantify. This is not the case for neurological and psychiatric disorders. Mental health is an area where diagnosis, and phenotyping more generally, is as much art as science. It is an area that poses enormous burdens on society, both financial and quality-of-life related, and is also ripe for a deeper, more physiologically-based understanding [53]. Even if therapy is still a long way away, having some concrete, quantifiable biomarkers, by which we could classify conditions such as depression, bi-polar disorder, and manic-depressive tendencies, would be a great leap.

Finally, major changes will be required to effectively and efficiently train the workforce of tomorrow. These changes will not simply entail adding a few quantitative courses into medical and graduate level biomedical research training, though that too will be important. The need for more informatics professionals is being addressed to some extent by the significant number of training programs being created at multiple different levels, from certificates to master’s to PhD programs. In addition to the informatics workforce, more genetic counselors are needed. It is estimated that there are only approximately 3500 genetic counselors in the entire United States (http://www.abgc.net/About_ABGC/GeneticCounselors.asp). As genomic sequencing becomes increasingly widespread, this number needs to increase both in the US and around the globe. Genetic counselors today may be compared to pathologists in the early days of the microscope, or radiologists in the early days of X-rays. The respective numbers will never need to be equal—except in the case of cancer, the genome need only be sequenced once in a person’s lifetime. Moreover, pathologists and radiologists detect and describe what is. Even as we learn more about our genomes, they primarily tell us what is more or less likely to be. Still, many genetic findings are already actionable today, and this number will continue to increase.

There are two areas where we need to do better, but I am less optimistic that we will see real progress in the next decade or so. The first is in adoption of data standards. We need better resources for understanding, navigating, and using existing standards [54]. We also need more impactful incentives for adoption, and disincentives for failure to do so. In the current landscape, standards are too difficult to identify and adopt, and the benefit of doing so tends to be realized by people other than those doing the hard work.

Lastly, we need to establish more inter-interdisciplinary coordination and collaboration. Perhaps meta-interdisciplinary is a better term. As biomedical informaticians, we are by definition interdisciplinary, including training and perspectives from medicine, biology, and computer science. But there are a large number of different communities around the world who are working on these problems, talking mostly among themselves. The various professional societies, even the interdisciplinary ones, have their respective meetings. There is some cross pollination, and some overlap in who attends the respective events but still not enough. There could be so much more, and we could make significant progress, reduce redundancy, and increase return on investment for research funding if these groups could be more consciously and proactively in sync.

Conclusion

In summary, we are entering a new era in data-driven health care. Translational bioinformatics methods continue to make an actual difference in patients’ lives. The infrastructure, information technology, policy, and culture need to catch up with some of the technological advances. For researchers working at the cutting edge of translational bioinformatics, opportunities abound, and the future looks bright.

Competing interests

The author has declared no competing interests.