Biomarker development in the precision medicine era: lung cancer as a case study

Abstract

Precision medicine relies on validated biomarkers with which to better classify patients by their probable disease risk, prognosis and/or response to treatment. Although affordable ‘omics’-based technology has enabled faster identification of putative biomarkers, the validation of biomarkers is still stymied by low statistical power and poor reproducibility of results. This Review summarizes the successes and challenges of using different types of molecule as biomarkers, using lung cancer as a key illustrative example. Efforts at the national level of several countries to tie molecular measurement of samples to patient data via electronic medical records are the future of precision medicine research.

As announced by the US President Barack Obama during the 2015 State of the Union Address¹, with further details provided by leaders at the US National Institutes of Health^2,3, the Precision Medicine Initiative promises to improve human health by combining clinical data and biomarker measurements on a massive scale. The goal of these precision medicine efforts is to use multiple types of data to classify patients into precise groups that will benefit from a given treatment approach. Similar efforts have already begun in the United Kingdom⁴, in Denmark⁵ and in Germany⁵, where universal health care has made data collection easier. The term ‘precision medicine’ gained momentum with the publication of the 2011 US Institute of Medicine’s National Research Council report Toward Precision Medicine: Building a Knowledge Network for Biomedical Research and a New Taxonomy of Disease⁶. This report summarizes a research pathway to redefine and unite the taxonomic systems by which the medical and scientific communities classify diseases. Patients with different biomarkers present with different risks of developing a disease, different disease prognoses or different responses to treatment; therefore, new biomarkers will be added to the current standards of phenotypic features (symptoms and histology) and medical history to revise the definition of a disease to include a new subtype (taxa) (FIG. 1). New standards of care will then be developed for these newly defined disease taxa. For example, epidermal growth factor receptor (EGFR)-positive non-small cell lung adenocarcinoma is currently its own taxon as opposed to being included in the general lung adenocarcinoma taxon and is treated with different chemotherapy from that of non-EGFR-driven adenocarcinomas⁷. The term precision medicine is new and arguably more accurate than its predecessor ‘personalized medicine’, but the overarching approach to developing improved biomarkers has not appreciably changed (BOX 1). Rather, the primary difference between the precision medicine research approach and traditional biomarker development is the magnitude of data collected and the speed at which data from different sources is (usually simultaneously) analysed. The Institute of Medicine’s report serves as a unified reference document that includes standard terminology that researchers of multiple disciplines can use when designing precision medicine studies.

Figure 1|. Classifying patients into new, specific taxa.

Patients with the same signs and symptoms of cancer often have different outcomes. The precision medicine approach provides a research strategy to develop biomarkers that can be used to classify patients with the same cancer into finer taxa (subclass 1 versus subclass 2) by biomarkers that predict prognoses derived from the synthesis of large amounts of data to identify discriminating biomarkers. For example, patients in subclass 1 who have a worse prognosis (that is, have biomarkers that are associated with poor survival) may be given a more aggressive treatment (treatment X) versus those in subclass 1 who have a better prognosis (that is, have biomarkers that are associated with good outcome) and require a less aggressive therapy (treatment Y). Additionally, the converse may be true where individuals with a worse prognosis are provided less aggressive therapy if no benefit from aggressive treatment has been observed for this subclass.

Box 1|. A brief history of the term ‘precision medicine’.

Hippocrates gave the field of medicine the sound advice that “It is far more important to know what sort of person the disease has than what sort of disease the person has.” Thus, the notion that individuals presenting with the same signs and symptoms may require different treatment is not new. Precision medicine has a long list of predecessor terms with similar meaning, including personalized medicine, P4 (predictive, preventive, participatory and personalized) medicine, genomics medicine, predictive medicine and individualized medicine. Regardless of the name of the approach, the goal is to use molecular data in addition to more traditional clinical information (for example, symptoms, personal history and histology) to tailor medical care to provide the most benefit while minimizing risk. The application of precision medicine is anticipated to improve all areas of medicine, including predicting an individual’s risk of disease, disease prognosis and risk of side effects versus positive response to disease treatment approaches. Thus far, the greatest advances of precision medicine have been achieved in the prediction of response to a drug therapy using companion diagnostics (that is, biomarkers that can predict response to a specific drug treatment).

The precision with which a disease can be classified into subtypes (or taxa) rests heavily on the success of the research framework outlined in the Institute of Medicine’s report⁶ and on supportive mechanistic research. Specifically, this report proposes that an ‘Information Commons’ will serve as a giant reservoir of medical history, demographic, molecular measurement and disease outcome data (FIG. 2). The data mined and analysed from this Information Commons should then be integrated with other published biomedical literature on mechanisms of action to generate an educated Knowledge Network, which will identify biomarkers that classify patients by differential risk, diagnosis, response or outcome. Insights gained from the Knowledge Network will then be used to revise subtype definitions and standard of care for each disease subtype. This process of Information Commons-Knowledge Network-redefining taxon definitions is an iterative one that is ideally evaluated in real time to improve the precision with which patients can be diagnosed with a disease subtype. That is, the entire process is actively provided updated patient outcome data or relevant mechanistic data as soon as they are available to continually improve taxonomy, apply the new taxonomy and collect more data on patient outcomes using the new taxonomy. The real-time update of patient and biomarker data is a goal of the US Precision Medicine Initiative for its national cohort of 1,000,000 citizens; this Initiative will therefore be a true test of the limits of the precision medicine approach to define new disease taxonomy³.

Figure 2 |. A precision medicine research strategy.

As outlined in the 2011 Institute of Medicines National Research Council report entitled Toward Precision Medicine: Building a Knowledge Network for Biomedical Research and a New Taxonomy of Disease⁶, an Information Commons will be analysed to develop a Knowledge Network to inform research and medicine. The Information Commons will serve as a reservoir of data on a group of individuals from multiple sources (clinical data, demographic and epidemiological data, and multiple types of ‘omics’ data). Analyses of the Information Commons will result in the generation of a Knowledge Network that will specify clinical, demographic and ‘omics’ characteristics that predict disease risk, diagnosis, response and prognosis, thus allowing for the reclassification of individuals into subtypes (taxa). These new taxa will require further research and clinical follow-up to validate their existence and to determine the most suitable taxon-specific standards of care. Adapted with permission from REF. 6 by the National Academy of Sciences, Courtesy of the National Academies Press, Washington, DC, USA.

Biomarkers are the foundation of improving diagnostic precision. Biomarkers can be correlational (that is, only associated with disease) and/or functional (that is, they have an identified mechanism of action related to disease). Functional biomarkers can also be used as potential therapeutic targets. Biomarkers can be measured alone or in a group, often called a biomarker panel, to infer risk, diagnosis, prognosis and therapeutic response. DNA, RNA, proteins, metabolites, host cells and microorganisms can all function as biomarkers, and are now often measured using ‘omics’ methodology. Lower throughput, tissue visualization-and imaging-based biomarkers are also commonly used due to the availability of formalin-fixed, paraffin-embedded clinical specimens. Biomarkers can be measured in a variety of biological material (for example, blood, organ tissue, stool, saliva and urine). Analogous to the phases of drug development, Pepe et al.⁸ outlined a set of five phases to serve as an early guide for researchers aiming to bring biomarkers to the clinic. These five phases are as follows: preclinical, exploratory studies; clinical assay development and validation studies; longitudinal, retrospective studies; prospective, screening studies; and studies to determine whether the biomarker reduces morbidity or mortality in the population. The vast amount of data generated by precision medicine research is likely to affect each of these phases.

The development of biomarkers is directed by regulatory guidelines. In this regard, the US Food and Drug Administration (FDA) is recognized as a leader in the regulation of precision medicine because it was the first to set regulatory guidelines in this field⁹. For example, the FDA provides a guided pathway for the development of precision medicine biomarkers that are paired with a companion therapeutic agent. These paired biomarkers are termed companion diagnostics¹⁰. The European Union’s European Medicines Agency (EMA)^11,12 and Japans Pharmaceutical and Medical Device Agency (PMDA)¹³ have also released guidance for companion diagnostics, and harmonization efforts are ongoing within each country and between countries to improve regulatory pathways^14–16. These regulatory processes have resulted in the approval of companion diagnostics such as EGFR biomarkers that pair with the EGFR inhibitor afatinib for lung cancer¹⁷. Now, international guidelines exist for the use of EGFR and other non-small cell lung cancer companion diagnostics¹⁸. The pathways for FDA approval of other biomarkers (that is, biomarkers of risk, diagnosis, response and prognosis not related to a companion therapeutic), designated as Laboratory Developed Tests, have fallen under the FDA purview only recently, and a regulatory framework for these biomarkers is still being developed¹⁹. These other types of biomarker have not yet been specifically addressed by the European Union or Japan. With respect to technical validity and reproducibility, the United States requires complex biomarkers to be tested in a Clinical Laboratory Improvement Amendments of 1988 (CLIA)-certified laboratory²⁰, whereas the European Union and Japan do not currently require the use of standardized laboratories for biomarker measurement. The development of regulatory guidelines in even well-developed countries to ensure the safety and utility of biomarkers is an immense effort; thus, the regulation and use of complex biomarkers in less developed countries are likely to continue to lag behind.

The aim of precision medicine and its regulation is to move molecular measurements through validation and ultimately to patient populations in need of improved diagnostic precision. Although the cornerstone of precision medicine will be the large national cohort studies, currently, the precision medicine research approach is being used on existing cohorts and in clinical trials with a much smaller number of individuals. In these studies, large amounts of data from a variety of sources (for example, histology, DNA, protein and RNA) are analysed to determine which pieces of data are the best predictors of disease risk, treatment response and/or prognosis, and thus should be moved forward for validation as biomarkers in future studies. Oncology is a field ripe for using diagnostic tests to subdivide patient populations into risk and treatment categories (that is, new disease subtypes or taxa)². Lung cancer, in particular, causes more deaths worldwide than the other top three cancers combined²¹ and therefore presents with a great need for improved diagnostic precision. This Review discusses the biological and statistical strengths and weaknesses of using different types of molecule as biomarkers. Specific examples will be taken from lung cancer to illustrate the potential and pitfalls of each. We then discuss the grouping of different molecule types together to form biomarker panels. Finally, we propose future pathways for precision medicine biomarkers and discuss the potential of biomarkers to stratify patients with lung cancer into different treatment groups to enable precision medicine.

Single-molecule types of biomarker

The measurement of genomics, transcriptomics, pro-teomics, metabolomics, microbiomics and other omics’ methods has now reached a point of reasonable return on investment. The giant reservoirs of molecular data linked to continuously updated electronic medical records envisioned by the Institute of Medicine are now within reach in the foreseeable future. The general precision medicine approach to analysing data is conceptually no different from what has been performed in the past to identify biomarkers in cohort studies. However, it is the proposed magnitude of the amount of data that will need to be collected, stored and analysed to identify the most precise biomarkers possible that is challenging this field (BOX 2).

Box 2 |. The challenges of developing a national precision medicine cohort.

More data collected on more individuals are required to generate validated biomarkers that can improve the precision to which we can categorize an individual’s risk of disease or response to therapy. However, the collection of large amounts of personal data presents complex challenges to researchers, the medical community and individuals whose data are being collected. The challenges facing countries that are developing national cohorts include the following:

• Collecting, handling, storing and transporting millions of biospecimens and then analysing these data using multiple different molecular measurement techniques

• Collecting electronic medical record data, merging data from different types of medical records and questionnaires, and then storing large amounts of these data

• Analysing data from different sources (for example, questionnaires, molecular measurements and electronic medical records) while respecting the strengths and limitations of each type of data

• Combining expertise from multiple different disciplines, including clinicians, laboratory researchers, bioinformaticians, biostatisticians and lawyers

• Dissemination of these data for researchers to use while ensuring that legal, ethical and privacy concerns of all participants are addressed

The feasibility of national cohorts for the purpose of precision medicine is restricted to countries and regions with the resources to meet all of the above challenges. This requirement will limit the ability of precision medicine to rapidly move to under-resourced regions of the world and its applicability to different races and ethnicities.

This section of the Review focuses on studies that have applied a simplified version of the precision medicine approach, wherein only a single type of molecule (for example, metabolite or protein levels, but not both) in a single biospecimen type is measured in a single cohort. Once identified, putative biomarkers should be validated in other cohorts and a mechanism of action should be elucidated. Once these steps have been accomplished, if the biomarkers were originally measured in tissues that can only be obtained through invasive methods (for example, solid tumour tissue), then blood, urine, saliva and other routes of less invasive specimen extraction should be tested and validated⁸. Once a feasible platform for measurement is identified (for example, genomic biomarkers may be discovered using next-generation sequencing but often PCR is the preferred platform in the clinic), these biomarkers can then serve to better classify patients in the clinic.

Genomic biomarkers in precision medicine.

Studies of genomic biomarkers in tumour tissue have advanced our understanding of lung adenocarcinoma disease sub-types and taxonomy, particularly at the diagnosis stage. Historically, lung cancer has been grouped into small cell carcinoma, non-small cell squamous cell carcinoma, non-small cell adenocarcinoma and large cell carcinoma subtypes. In the late 1980s and the mid-2000s the research community began to recognize that lung adenocarcinoma could further be subdivided beyond histology into cancers that were driven by KRAS²² and/or EGFR²³ gene mutations (FIG. 3). Finally, the development of next-generation DNA sequencing technologies facilitated the comprehensive characterization of the lung cancer genome. Specifically, The Cancer Genome Atlas (TCGA) network has, to date, sequenced approximately 1,025 lung cancer exomes of various histologies and identified at least 15 unique candidate genes that can drive oncogenesis in lung adenocarcinoma when a somatic mutation occurs^7,24. New mutations continue to be identified in lung adenocarcinomas such as the recent identification of mutations in protein phosphatase 3 catalytic subunit-α (PPP3CA), histone H3K79 methyl-transferase DOT1L and FtsJ methyltransferase domain containing 1 (PTSJD1; also known as CMTR2)²⁴. Owing to a lack of companion therapeutics and mechanistic studies linkizng newly identified subtypes to disease outcomes, only EGFR-driven or anaplastic lymphoma kinase (ALK)-translocation-driven tumours are treated as different disease subtypes in current clinical practice^25,26. Multiple reviews have described the ongoing and promising efforts in the development of genetic biomarkers that are also targets for therapy in lung cancer^27–29.

Figure 3|. Knowledge of non-small cell lung adenocarcinoma has evolved in recent decades.

Traditionally, lung cancer was grouped by histology into small cell lung cancer and non-small cell squamous cell carcinoma or adenocarcinoma. In 1987, a KRAS mutation was identified in ~25% of all non-small cell lung cancers, and 50% of lung adenocarcinomas²². In 2004, epidermal growth factor receptor (EGFR) mutations were identified as an additional mutation in lung adenocarcinomas²³. The Cancer Genome Atlas (TCGA) Network’s next-generation sequencing of lung adenocarcinoma in 2014 led to the identification of more than 15 different gene events that could be exploited for treatment and/or used for subclassifying patients into new taxa⁷. ALK, anaplastic lymphoma kinase; amp, amplification; ex, exon; RIT1, Ras like without CAAX1. Data in the left panel were abstracted from Rodenhuis et al.²². Data in the middle panel were abstracted from Paez et al²³ and Riley et al.¹³². The right panel of the figure is from REF. 7, Nature Publishing Group.

The measurement of DNA mutations and translocations as biomarkers has paved the way for further subdivision of lung adenocarcinomas into subtypes that are associated with different outcomes and with different responses to treatment^30,31. The best example of these biomarkers currently used in the clinic for patients with lung cancer is the PCR-based companion diagnostic test for EGFR mutations in tumour tissue. This companion diagnostic is used to determine whether the tumour is of the EGFR-positive subtype³¹; if the patient is EGFR positive, then they can be prescribed an EMA-or FDA-approved EGFR-inhibiting drug (for example, afatinib^17,32). Similarly, ALK translocations identified by fluorescent in situ hybridization (FISH) leads to patients being eligible for an EMA-or FDA-approved ALK-inhibiting drug³⁰. Despite sequencing efforts on squamous cell lung cancers³³, the identification of new biomarkers and targeted therapy for this type of lung cancer remains limited. However, immunotherapy that targets programmed cell death protein 1 (PD1; also known as PDCD1) is an example of a rare success for this histological subtype³⁴. Many studies have provided evidence that alterations in the non-coding regions of DNA and gene polymorphisms are also associated with disease risk, response or prognosis^35–41. The targeting of non-coding regions in particular may be a new avenue of exploration for the prevention and treatment of lung cancer.

Tissue-level DNA measurements are not without their limitations. First, collection of tissue is usually invasive and changes in the DNA are not always functional. Patients with cancer routinely undergo biopsy and tumour removal, which makes this approach feasible despite its invasiveness for initial treatment decisions. However, if subsequent biopsies are needed to make future treatment decisions, then this approach is more challenging. Second, biopsies may not be representative of the whole tumour due to the heterogeneity of multiple malignant cellular clones within a tumour⁴². Thus, there is an impetus to find less invasive markers of tumour DNA status, for example, measurement of circulating cell-free DNA in lung cancer to identify an increase in EGFR mutations⁴³. However, the low sensitivity of these blood-based measurements will probably limit their use to monitoring, as opposed to diagnosing, lung cancer⁴⁴. Third, treatment may select for certain cancer clones, as evidenced by the rise in the ratio of EGFR mutations observed in patients with lung adenocarcinoma who had developed resistance to tyrosine kinase inhibitors⁴⁵. The TRACERx (TRAcking non-small cell lung Cancer Evolution through therapy (Rx)) study⁴⁶ aims to monitor the impact of tumour heterogeneity on therapeutic outcomes and will address many questions about the response of tumours to treatment over time. The findings from this study are particularly relevant to precision medicine because the behaviour of tumours over time may explain changes in patient outcome, which will be continuously collected from electronic medical records. The invasiveness of biopsies, challenges in identifying functional versus non-functional changes, temporal variation in biomarker values and tumour heterogeneity are not unique to genomic biomarkers. Rather, these factors are challenges to all biomarker development regardless of molecular type.

Transcriptomic biomarkers in precision medicine.

The global measurement of mRNA expression, termed transcriptomics, has provided an understanding of cancer subtypes but, in contrast to DNA, is tissue specific. Evaluation of the transcriptome of non-small cell lung cancer to identify mRNA expression biomarkers began with microarray technology^47–50. Using microarray technologies, many studies have identified panels of mRNA expression biomarkers that classify lung cancer into more precise subtypes based on associations with disease outcomes^51–55. In contrast to microarrays, which are limited to preselected mRNA probes, mRNA sequencing enables the sequencing of all mRNA present in a sample then, similar to DNA sequencing, maps sequences back to a reference library⁵⁶. Putative mRNA biomarkers in non-small cell lung cancer are being identified using this newer, more comprehensive approach to searching for biomarkers (for example, Janus kinase (JAK)-signal transduction and activator of transcription (STAT) pathway mRNA⁵⁷ and mRNA of tumour-educated platelets⁵⁸) but it is still in preliminary stages.

Epigenomic biomarkers in precision medicine.

DNA methylation, histone protein modifications (for example, methylation and acetylation), microRNA (miRNA) and long non-coding RNA (IncRNA) are all measureable epigenomic biomarkers that function principally to regulate RNAs^59,60. miRNA and IncRNA can also have other functions, such as serving as ligands for receptors⁶¹, and are measured using the same general methods as mRNA. DNA methylation and histone modifications usually require immunoprécipitation of the epigenomic mark of interest⁶² or, for DNA methylation, bisulfite conversion⁶³ or restriction enzyme use⁶⁴ before microarray or sequencing analyses. Epigenomics has been used largely to explore the aetiology of lung cancer, and epigenomic biomarkers are candidate mechanistic biomarkers for classifying individuals based on disease risk. However, as Lilogou et al.⁵⁹ have recently reviewed, epigenetic biomarkers also have the potential to serve as biomarkers for identifying subclasses of patients with lung cancer. For example, the promoter methylation status of five genes was recently identified as a classifier of non-small cell lung cancer prognosis⁶⁵. Global methylation patterns, such as CpG island methylator phenotype, have also been associated with prognosis in adenocarcinoma⁶⁶. Interest in this area continues to grow and there are ongoing clinical trials to move these biomarkers into the clinic.

Proteomic biomarkers in precision medicine.

Immuno-histochemical staining of proteins in formalin-fixed, paraffin-embedded lung tissue samples has been recommended by international experts for use in the clinic to classify tumours⁶⁷. Studies that used tissue microarrays and existing immunohistochemical protein stains in a high-throughput manner have identified new, putative lung tissue biomarkers⁶⁸. Advances in mass spectrophotometry that, analogous to next-generation sequencing, enable mapping of a multitude of mass spectrophoto-metric peaks to reference libraries to identify proteins has facilitated the global assessment of the non-small cell lung cancer proteome^69,70. Using this new technology, 17 different circulating proteins were recently identified and validated as putative biomarkers for non-small cell lung cancer⁷¹. Circulating proteins have also been explored as less invasive biomarkers in lung cancer. Specifically, due to their role in carcinogenesis⁷², circulating inflammatory proteins have demonstrated clinical utility in lung cancer prognosis^73–79. However, moving proteomic biomarkers from the exploratory mass spectrophotometry-based analyses phase into the clinic, which would require more stable measurement platforms, remains challenging. Füzéry et al.⁸⁰ have thoroughly reviewed such challenges in the framework of the existing FDA approval pathways.

Antigenic proteins expressed on lung cancer and immune cell surfaces are attractive targets for the development of immunotherapeutic antibodies, and have been reviewed extensively^81,82. Briefly, intravenous administration of PD1 and PD1 ligand 1 (PDL1) IgG antibodies demonstrate efficacy in the treatment of non-small cell lung cancer. Indeed, two anti-PDl therapies — pembrolizumab^83,84 and nivolumab^85,86 — have been approved by the EMA and FDA for use in nonsmall cell lung cancer. Recently, PDL1 measurement via immunohistochemistry was approved as a companion diagnostic for pembrolizumab by the FDA⁸⁷, but the path to approval for this biomarker was fraught with specificity and immunohistochemical challenges^88,89. Other attractive immune biomarkers include CD8⁺ lymphocytes identified by quantitative fluorescence in tumours, which have been associated with better prognosis⁹⁰. However, tissue imaging-based biomarker identification is hampered by limited throughput and, even with tumour microarrays, often requires a large sample input for a small amount of data output compared with proteomics technology. Nonetheless, the mechanistic relationship between other antigen biomarkers, as companion diagnostics, and antigen-targeting therapies continues to incentivize more research in this rapidly advancing area.

Metabolomic biomarkers in precision medicine.

Similar to proteomics, metabolomics (also known as metabo-nomics) can be assessed in a targeted or unbiased manner, and mass spectrophotometry is used to identify chromatogram peaks as specific metabolites^91,92. Metabolomics is particularly promising for biomarker development because altered metabolism is considered a hallmark of cancer⁷². Moreover, metabolites are frequently exported to the blood for transport or removal from the body via urine or faeces; therefore, these metabolites could serve as non-invasive biomarkers that accurately reflect the metabolic activity of tumour tissues. Tissue^93,94, blood^95–98 and urinary^99,100 metabolomics analyses have yielded putative biomarkers that classify patients into subtypes of lung cancer; however, these findings have yet to be sufficiently validated in more than one cohort while also using positive controls to ensure accurate identification of the purported metabolite. Validation of metabolomic biomarkers requires not only analyses in other cohorts but also a known standard to confirm the identity of the putative metabolite peak. Improved libraries of synthesized standards to authenticate peak identity are an area of need to move this research forward and to build more reliable platforms for metabolite analyses that can progress to the dinic¹⁰¹.

Microbiomic biomarkers in precision medicine.

Using modified extraction procedures, microbial DNA is generally measured in the same manner as human-derived DNA, with the popular exception of 16S rRNA-specific gene sequencing to identify bacteria predominantly at the genus level^102,103. The normal lung was thought to be generally sterile until the advancement of culture-independent microbial DNA sequencing techniques for microorganism identification in the lung^104,105. As a result of such techniques, we know that cigarette tobacco contains bacteria¹⁰⁶ and that cigarette smoke can disrupt the respiratory tract mucosal barrier¹⁰⁷ to allow microbiota migration into the lung. These findings have led to the hypothesis that the lung microbiome may play a part in carcinogenesis. To our knowledge, the global analysis of the lung cancer tissue microbiome remains in progress. However, one study has suggested that bacillus species in lung sputum could serve as a non-invasive biomarker of increased lung cancer risk¹⁰⁸. The recent identification of Fusobacterium nucleatum as a functional microbial biomarker in colon cancer^109–112 provides strong support for the continued interrogation of the microbiome as a functional biomarker with which patients with lung cancer could be classified into risk, response or prognostic subtypes. Antibiotic, probiotic or prebiotic treatment could then be prescribed for different diagnostic sub-types to modify their risk and/or response to therapy¹¹³. Standards for faecal microbiome research approaches are emerging^114,115, but standards for other biosample types are needed to accelerate the development of non-faecal microbiome biomarkers.

The exposome in precision medicine.

The term ‘exposome’ was first coined by Christopher Wild in 2005 (REF. 116) and refers to all types of molecules and events from the environment to which humans can be exposed; for example, drugs, diet or the microbiome (FIG. 4], Aspects of the exposome are commonly measured by questionnaires, which are administered to patients in the clinic. For lung cancer in particular, information about cigarette smoke and asbestos exposure is requested in addition to information on age and sex to generate a panel of information that is used to stratify people into subtypes of risk of lung cancer^117–119 (also see Further information). However, questionnaire responses are biased and error prone. For example, self-reported tobacco exposure does not always correlate with measured tobacco carcinogen exposures¹²⁰. Thus, development of molecular biomarkers of tobacco smoke^121,122 and other exposures that could be used in the clinic to more accurately reflect a patient’s expo-some, and therefore lung cancer risk classification, are ongoing.

Figure 4|. The lung exposome.

The exposome of the lung comprises a diverse array of molecules and events (including carcinogens from tobacco, asbestos and radon) that come from the external and internal lung environment. These external and internal influences interact with each other and host’omes’to alter the lung cell environment (including inflammation and the microbiome) and may promote or protect against the development of the hallmarks of cancer⁷². Smoking is estimated to cause 90% of lung cancers. Occupational exposures to carcinogens and radon exposure are estimated to cause 9–15% and 10% of lung cancer cases, respectively¹⁷⁴. Measurement of the exposome, in addition to other host ‘omics’, has led to the development of precise biomarkers of risk, diagnosis, treatment response and prognosis by which patients can be classified into new taxa. These new taxa then require different standards of care for cancer screening, diagnosis, prevention and therapy.

Interestingly, the exposome alters the effects of other molecular measurements (for example, inflammation) and the effects of the exposome may be altered by changes in other molecules such that it is imperative that the exposome be considered when developing any type of biomarker. Specifically, among patients with lung adenocarcinoma, smokers have a higher mutation frequency overall than non-smokers¹²³, and it is known that exposure to smoke can cause gain-of-function TP53 mutations^124,125, which drive lung carcinogenesis. Although questionnaire-derived data have been analysed predominantly by epidemiologists and molecular data by basic scientists, these data are more commonly being incorporated together in research projects, and collaboration across disciplines is essential to ensure these data are analysed to their fullest potential. The combination of the exposome and molecular biomarkers for improved prediction of disease subclassifications is discussed in more detail below.

Summary.

Nearly all omics’ analyses are still limited by factors that have slowed the progression of biomarkers from discovery to deployment in the clinic. The following are the most important factors impeding progress. First, there are technical reproducibility issues of ‘omics’ platforms and variability between laboratories¹²⁶, because the vast majority of biomarker studies are not conducted in CLIA-certified or other regulated laboratories. Second, there are limitations in the quality and size of the reference library used to identify molecules. Third, false positives due to the vast amount of potential biomarkers analysed in global omics’ studies^127,128. Fourth, statistical reproducibility issues arise owing to biases in the original sample or validation cohort, or owing to false discoveries^8,129. Fifth, lack of longitudinal cohorts in which to validate biomarkers over time. Sixth, the need for functional studies of putative biomarkers. Finally, heterogeneity within and between samples leads to inconsistent measurements on the same sample or measurements that are out of the dynamic range of a test, respectively.

These limitations have led to a plethora of putatively identified biomarkers in the literature that lack validation, mechanistic evidence and/or follow-up studies. Moreover, many biomarkers are identified at the tissue level. Although tissue-level studies are often crucial in identifying a mechanistic link between a biomarker and carcinogenesis, biomarkers requiring biopsy are not practical for assessing cancer risk and for monitoring response to treatment in the dinic, even in well-developed health-care systems. Perhaps the expansion of regulations regarding laboratory developed tests in the United States¹⁹, and potentially the European Union and Japan, will incentivize the biomedical community to move putatively identified biomarkers towards less invasive, validated biomarkers not only for companion diagnostics but also for biomarkers of disease prognosis and risk.

The ability of any biomarker to correctly differentiate two subgroups of patients with lung cancer in a statistically significant and clinically meaningful manner is dependent on the relative number of people in each group and the magnitude of the difference in the value and variance of the biomarker in each group (that is, power)^127–129. Furthermore, all analyses comparing two lung cancer subgroups assumes that the groups are identical to each other in all other ways except for the biomarker or biomarkers that differentiate the lung cancer subtype¹³⁰. Thus, the rarer a disease subtype is the less likely it will be that even large cohorts will have the power (number of participants and measured differences between subgroups) to meet the assumptions to identify these subgroups. Nonetheless, innovative clinical trial design approaches (for example, n = 1 studies¹³¹) are leading the way for the approval of companion diagnostics and are helping to minimize the challenge of not observing enough participants in each subgroup²⁷. In addition, the more studies that are conducted and the more sensitive or specific molecular measurement platforms and procedures become the more statistical power we will have to identify and validate new biomarkers. These new approaches, paired with the precision medicine approach⁶ to continuously re-evaluate the outcomes for patients classified to a disease subtype by biomarkers in a longitudinal manner, will change the way biomarkers are developed and will enable the most precise classification of patients in the future.

Multi-molecule-types of biomarker panel

The Institute of Medicine’s seminal report envisioned an Information Commons that contains multiple omics’ approaches such that biomarker panels containing different molecule types, exposome data and/or demographic data could be developed to classify diseases into more precise subtypes⁶. Although single-molecule biomarkers (for example, EGFR) have progressed into the clinic³¹, biomarker panels are still in the discovery stage. Biomarker panels containing various types of molecule are attractive because genes, proteins, RNAs and metabolites all work in concert to prevent or promote the development of the hallmarks of cancer⁷². However, the development of biomarker panels is still limited by two main factors. First, by any weaknesses associated with each individual omics’ technique, molecule type and tissue type included in the panel, and second, by amplification of the factors identified as challenges common to the identification of any biomarker mentioned above^8,127–129. Integrating different types of data leads to more potential biomarker combinations and, consequently, more functional studies and statistical tests that need to be run. These additional statistical tests require more power to detect significant differences while still trying to avoid false positives^127–129. Two conceptual approaches to developing biomarker panels have been used. The first relies on adding new biomarkers to existing biomarkers or biomarker panels to improve the sensitivity or specificity of the panel because of either an interaction or an independent effect of the new biomarker. The second approach employs de novo analysis and integration of multiple sources of molecular data to identify the best combination of putative biomarkers.

Adding new biomarkers to existing biomarker panels in precision medicine.

The addition of new biomarkers to existing, validated biomarker, demographic and exposome information (for example, smoking status, age, race and sex) panels decreases the number of possible biomarker combinations. Thus, the statistical power needed to identify additional biomarkers compared with de novo analysis (discussed below) is also decreased as compared with de novo analyses. An early, seminal example of adding a biomarker to existing predictive information in lung cancer was the discovery that smoking (exposome) can cause TP53 mutations and thus alters lung cancer risk¹²⁴. However, treatments targeting p53 or KRAS pathways (KRAS is also frequently mutated⁷ and associated with smoking in lung cancer¹³²) remain in clinical trial stages. Interestingly, most of the work in this area was conducted before advanced next-genera-tion sequencing and used exhaustive mechanistic studies to identify this relationship. Conversely, the majority of newer biomarker studies pair epidemiological evidence with global ‘omics’-based technology, and then, if validated, few have identified a mechanism of action that explains the relationship between these biomarkers and disease. Identifying a mechanism of action is required for the development of viable companion therapeutics.

The precision medicine research approach has been used in lung cancer to interrogate whether the addition of new molecular data improves an existing biomarker panel. A recent publication using a TCGA dataset has provided evidence that the addition of single-molecule type biomarkers (copy number alterations, protein and miRNA measurements) to existing exposome-only predictors improves prognostic accuracy in lung squamous cell carcinoma¹³³. With respect to early-stage lung adenocarcinoma, mir-21 methylation¹³⁴, homeobox A9 (HOXA9) methylation¹³⁵ and the expression levels of a panel of four genes⁵⁵ have been validated in multiple cohorts to independently predict lung cancer outcomes. However, the addition of mir-21 (REF. 51) and HOXA9 methylation status¹³⁵ to the gene expression biomarker panel improves the predictive accuracy above any of these biomarkers alone⁵⁵. Analyses of multiple international cohorts have provided evidence that the risk of lung cancer from asbestos exposure is increased by the presence of certain genetic variants¹³⁶. Although not multi-omic’, there are ongoing clinical trials that use traditional histological subtyping and then use genomic analysis in patients with lung cancer using a variety of platforms. Examples of such trials include the BATTLE-2 trial¹³⁷ and a US National Cancer Institute trial¹³⁸. These results are then used to make treatment decisions and to identify de novo genetic variations associated with disease response and/or prognosis as candidate biomarkers. These studies are the pinnacle of precision medicine research as it currently exists and, if successful, will change clinical practice within the next decade. This relatively straightforward approach of adding new biomarkers requires a thoroughly validated existing biomarker or panel of biomarkers and could possibly miss unique combinations of biomarkers that may serve as better predictors. With the recent improvements in computational technologies, there has also been a movement towards de novo analysis of multiple omics’ datasets to truly integrate the data and identify novel biomarker panels.

De novo analysis combining global datasets to generate biomarker panels in precision medicine.

Integrating multiple datasets, often derived from global omics’ analyses of different types of molecule, for biomarker panel development has largely been stymied by a lack of methodological approaches that are suitable for combining different data sources. Recent advances in network analysis and other mathematical modelling approaches are swiftly moving this field forward^139–143; however, no preferred approach has emerged. Using TCGA data, Li et al,¹⁴⁴ integrated genomic, transcriptomic and proteomic information to classify patients with non-small cell lung cancer by prognosis. Kim et al.¹⁴⁵ integrated DNA, mRNA, miRNA and methylation sequencing data to identify putative biomarkers that classify female patients with non-smoking-associated lung adenocarcinoma into distinct subtypes. However, identifying mechanisms of action and finding cohorts with sufficient sample size (with meaningful racial, ethnic and geographic diversity) in which to validate these, and other, de novo-assembled biomarker panels remain major challenges. The Precision Medicine Initiative cohort^1,2, the UK Biobank⁴, The International Cancer Genome Consortium¹⁴⁶ and other large studies with molecular measurement data will provide an unparalleled opportunity for validating biomarker panels owing to the projected collection of multiple ‘omics’ data that is anticipated to be publicly available for researchers.

Summary.

Putative biomarker panels in lung cancer are just beginning to accrue^{55,133,144,145}, but their path to the clinic will probably be even longer than for singlemolecule biomarkers owing to exacerbation of typical challenges associated with biomarker development by combining ‘omics’ approaches. Specifically, combining biomarkers leads to multiple potential combinations that require the statistical power to be tested. Using different types of molecule in one panel also leads to logistical challenges on how to measure different molecule types (for example, proteins and DNA) on the same platform when moving towards regulatory approval for clinical use. Nonetheless, biomarker panels are hypothesized to provide a more realistic picture of aberrant regulation in complex diseases such as cancer because molecules do not function in isolation to generate a phenotype. Small shifts in the relative amounts of RNA, protein, epigenetic modifications, metabolites and microorganisms over time may also be useful for early prediction of disease risk or outcomes. For example, a budding area of research in precision medicine is network medicine. Network medicine involves the use of emerging network approaches to integrate omics’ measurements^139–143 to look for small changes over time in the relationship between ‘omics’ that are associated with disease^{143,147–149}. Such dynamic analysis is in contrast to the current use of static panels of biomarkers. This continual monitoring approach is consistent with the greater vision of precision medicine; however, it should rely on only minimally invasive biomarkers and therefore lends itself well to the use of microfluidics¹⁵⁰.

Summary and conclusions

The reduced cost and increased reproducibility of new ‘omics’ technologies, new methodological approaches for integrating different types of molecular data^139–143, and the number of publicly available datasets with molecular measurements have all been pivotal in achieving the substantial leaps forward in biomarker development that we have witnessed over the past decade (BOX 3). The precision medicine research approach is simply a faster paced, larger scale, integrated version of the traditional, single measurement-based biomarker development approach that has only been made possible due to these advances. Currently, there are large prospective, international cohort studies (for example, the European Prospective Investigation into Cancer and nutrition (EPIC)¹⁵¹ and the Women’s Health Initiative¹⁵²) and electronic medical record-based datasets (for example, electronic MEdical Records and GEnomics (eMERGE)¹⁵³ and the UK Biobank⁴) with biobanked samples that allow for ongoing biomarker discovery. The number of biomarkers moving from discovery to clinical trials is worryingly small. Furthermore, the vast majority of registered clinical trials test biomarkers as companion diagnostics to treatment because there are existing regulatory frameworks for companion diagnostics whereas biomarker panels for risk of disease and prognosis (which are unrelated to a specific treatment) have murkier regulatory pathways.

Box 3 |. Promising biomarkers in lung cancer.

A summary of the promising biomarkers in lung cancer is provided below. For more detail, please refer to Lung Cancer and Personalized Medicine: Novel Therapies and Clinical Management as part of the Advances in Medicine and Biology series (2016)¹⁷³.

• Tumour immune and microenvironment biomarkers. For example, programmed cell death protein 1 (PD1), PD1 ligand 1 (PDL1) and vascular endothelial growth factor A (VEGFA)

• Genetic aberration biomarkers. For example, KRAS, HER2 (also known as ERBB2), BRAF, MET, ROSI, RET, fibroblast growth factor receptor 1 (FGFR1), SRY-BOX 2 (SOX2), platelet-derived growth factor receptor-a (PDGFRA), discoidin domain receptor tyrosine kinase 2 (DDR2), PI3K catalytic subunit-a(P/K3CA), PTEN, mixed lineage leukaemia 2 (MLL2; also known as KMT2D)

• Epithelial-to-mesenchymal transition-associated biomarkers. For example, SLUG, forkhead box C2 (FOXC2) and transforming growth factor-β (TGFβ)

• Resistance and susceptibility to treatment biomarkers. For example, ERCCl, ribonucleoside-diphosphate reductase (RRM) and thymidylate synthase (TS)

In addition to changes in regulations, the field of biomarker discovery has innate practical challenges. Specifically, the precision of disease subtyping by biomarkers to predict risk, response or prognosis is limited by a high risk of false positives when seeking to identify a biomarker from the global measurement of thousands of molecules. Moreover, questions of basic statistical power (that is, the number of patients presenting with a disease subtype)^8,127,129, the need for validation and functional studies, and the follow-through to develop non-invasive biomarkers are also limitations. McShane et al.^154–162 have published a series of manuscripts on REporting recommendations for tumour MARKer prognostic studies (REMARK). Consistent adherence to these publishing guidelines will enable improved transparency and allow the scientific community to better judge the quality of the plethora of studies reporting putative biomarker identification. Importantly, the success of these complex studies will require effective collaborative science; a good example is the success demonstrated by the Early Detection Research Network at the US National Cancer Institute¹⁶³. The rate at which new biomarkers enter the clinic will now be benchmarked by the requirements set forth by regulatory agencies^9–16,19. Changes in regulation that result in profitability for biomarkers not tied to a specific drug (for example, biomarker panels to estimate disease risk), will serve to spur the movement of biomarkers from lone studies to the clinic. The speed of discovery and validation of biomarkers could be improved by real-time data collection, which would, theoretically, allow for faster monitoring and revision of new disease taxa as data on patients are collected. However, this benefit must be balanced against the cost of screening a large population multiple times and the increased risk of false positives (over-diagnosis) simply due to the number of measurements being undertaken.

Analyses of existing cohorts and the generation of national cohorts (that is, in the United States^1,2, the United Kingdom⁴, and in Denmark⁵ and Germany⁵) promises to address many of the statistical and logistical concerns in biomarker development, allowing for the progress of truly precise medicine. One can envision that machine learning could be used to mine and continually improve algorithms for calculating patient risk, diagnosis, response and prognosis over time. These algorithms could then be applied to the general population to ensure patients receive appropriate care while measuring only the necessary biomarkers at key time points identified from national cohorts. Furthermore, with respect to the heterogeneous nature of cancers, there may be a day in the future when a tumour is profiled and then a follow-up biomarker panel and drug are developed in real time specifically for a single tumour. Although these possibilities are only likely to be realized decades in the future, currently, the precision medicine approach has led to the identification of new subtypes of non-small cell lung cancers (EGFR, ALK, TP53 and KRAS) and the translation of companion diagnostic biomarkers (for EGFR and ALK) to diagnose and choose appropriate treatment regimens for these new subtypes in the clinic³¹. However, there is an ongoing need for translatable biomarkers of disease risk, particularly for smokers, and prognosis for patients with EGFR-negative and ALK-negative lung adenocarcinomas. There is currently a large screening trial being conducted in the United Kingdom to test the efficacy of low-dose computed tomography (LDCT) before initiating a national lung cancer screening programme¹⁶⁴. LDCT has recently been approved for use in screening individuals at high risk for lung cancer in the United States^165,166, and its implementation will lead to an increased number of individuals diagnosed with early-stage lung cancer. An estimated 8.6 million to 8.8 million people in the United States meet the current criteria for LDCT^167,168 and, if they are screened, this will lead to an overwhelming number of true and false positive findings that will require treatment decisions^165,166,169. This large number is because, despite ongoing efforts to improve the specificity of imaging-based biomarkers¹⁷⁰, the false-positive rate for LDCT is estimated at >90% (REF. 165). Thus, one especially high impact area for the development of precision medicine biomarkers is to further classify early-stage (IA and IB) lung cancers into subtypes of patients at high risk for cancer recurrence to inform treatment decisions¹⁷¹ (FIG. 5). A similar approach would be useful in differentiating findings from mammography, which are also plagued by high false-positive rates¹⁷². This improved specificity would then allow oncologists to treat the patients who would benefit and minimize overtreatment of those who are unlikely to progress. This example is a clear demonstration of the power of precision medicine biomarkers and the future of medicine.

Figure 5 |. Use of precision medicine to classify patients with early-stage lung cancer into subclasses to provide appropriate treatment.

Approximately 25% of patients with stage I lung cancer will have recurrent disease associated with occult metastasis. This figure depicts the classification of early stage (IA and IB) lung cancers by a single biomarker or a panel of biomarkers that predicts risk of recurrence generated using a precision medicine research strategy into’low risk for recurrence’and ‘high risk for recurrence’. Once classified into subclasses (taxa), low-risk patients can be observed post-curative surgery whereas high-risk patients can be provided options for adjuvant therapy post-surgery.