Are randomized trials obsolete or more important than ever in the genomic era?

Th e genomic era has raised the possibility of major changes in the design, conduct, and even the existence of randomized trials as we know them [1­3]. Randomized trials are often seen as a slow, laborious, expensive, and diffi cult step in the translational process and are associated with a high attrition rate for drugs. Indeed, most tests that are in use for the screening, diagnosis, prognosis, monitoring or management of patients have never been scrutinized by a randomized trial. Th is has largely been due to a failure to realize that tests can do as much harm and as much good as drugs or devices; thus, a rigorous appraisal of their clinical utility, including both the possible benefi ts and the possible harms, is necessary. Moreover, numerous new omics­based tests are continu­ ally being proposed, especially in the context of targeted preventive or therapeutic interventions. Given rapid develop ment of these new biomarkers, can we make ran­ dom ized trials more adaptable to a changing land scape? Furthermore, do we still need randomized trials at all? Our answers to these two questions are: yes, to some extent; and yes, defi nitely. We will explain our reasoning in this article. Revolutionizing randomized trials in the genomic era Genomic applications in medicine are highly diverse. Th ey range from diagnostic and carrier tests for many rare Mendelian conditions to complex biomarkers that inform preventive or therapeutic interventions. One relatively new area is the explicit linkage of a treatment and a companion diagnostic test (a molecular assay to indicate the likelihood of a patient responding to this specifi c treatment or reacting adversely). With this in mind, at least three types of potentially informative


Revolutionizing randomized trials in the genomic era
Genomic applications in medicine are highly diverse. Th ey range from diagnostic and carrier tests for many rare Mendelian conditions to complex biomarkers that inform preventive or therapeutic interventions. One relatively new area is the explicit linkage of a treatment and a companion diagnostic test (a molecular assay to indicate the likelihood of a patient responding to this specifi c treatment or reacting adversely). With this in mind, at least three types of potentially informative designs are available for the conduct of omicsinformed clinical trials. Th e fi rst is the use of enrichment strategies, in which participants are selected for enrollment on the basis of a validated predictive tool to ensure the selection of those who are likely to have the best response rates (this approach is used to test biologically targeted thera pies) [1,2]. Th e second design makes use of surrogate markers (for example, a protein, metabolic, epigenetic, or other marker), especially validated markers, to conduct trials with a smaller sample size and a shorter followup period than traditional trials require. Th e third design incorporates a range of adaptive designs so that the trial can be optimized and modifi ed as it progresses. Various aspects can be altered, such as the inclusion and exclusion criteria and sample size, the data collection processes, and even the defi nitions of the endpoints and types of analyses. Th is fl exible approach would address the expectation that new information or discoveries will accumulate at such a rapid pace that external information could aff ect the conduct of trials midstream and could thus eff ectively guide their modifi cation. Trials could then be completed faster and might yield larger treatment eff ects and more defi nitive answers.
Despite these expectations of how genomic discoveries will alter clinical trials, omics research to date has not had the major clinical potential that was heralded. For example, despite decades of pharmacogenetics research, only a limited number of genetic variants have been robustly documented (at a genomewide signifi cance level) as being associated with drugrelated outcomes. Of the several thousand genetic associations that have been validated by genomewide association studies (GWAS), few pertain to pharmacogenetics, and far fewer have the large eff ect sizes that would make them readily action able. In the 50 most recent GWAS that were indexed in the National Human Genome Research Institute GWAS catalog as of 28 February 2013 [4], only 3 of the 290 well validated variants (with P < 5 × 10 8 ) pertain to pharmaco genetics, and all 3 have small eff ect sizes that are unlikely to be useful in clinical practice.
Th e US Food and Drug Administration records 119 pharmaco genetic associations that are listed on drug labels [5]. Only four, however, pertain to a clear require ment or a strong recommendation for genetic testing (for example, cetuximab (EGFR), trastuzumab (HER2), mara viroc (CCR5), and dasatinib (BCR-ABL)). For the other associations, few randomized trials have proved the clinical utility of routine genetic tests; an exception is HLA-B*5701 testing to assess the suitability of treatment with abacavir [6]. In several additional cases, there is some supporting evidence from randomized trials, but this is based on surrogate outcomes: for example, CYP2C19 testing for patients to be treated with clopido grel, in which the main outcome assessed was platelet reactivity rather than major bleeding [7]. Most pharma co genetic associations listed on drug labels have not had their utility tested in clinical trials.
Many other markers, in diverse omics fields, lack even proper validation at the association level. It makes little sense to pursue and incorporate nonvalidated markers or nonvalidated molecular signatures in clinical trials. Standards for reproducible research need to be strength ened [8]. As to adaptive confirmatory trials, they remain mostly theoretical constructs. We lack examples of newly arising genomic information leading to an adaptation of the design of already ongoing large trials with major impact on their eventual success. Unjustified adaptations can end up being indistinguishable from spurious manipulations of the data collection and analysis plans.

Getting rid of randomized trials in the genomic era?
A common complaint is that clinical trials are too large and expensive to be practical and, instead, one could wish that omics research would make these trials more efficient or even replace them. The reality is different. Even for blockbuster drugs, randomized trials are rarely large enough to demonstrate conclusively benefits for major clinical outcomes and death [9]. Usually, informa tion is fragmented across many small or modestsized trials, with selective publication and reporting. Rational i z ing the agenda of traditional randomized trials mostly requires common sense, not fancy omics. For example, designing several large, pivotal trials with reliable measurements and clinically relevant outcomes may yield more reliable information and at a lower cost than traditional randomized trials or trials incorporating expensive omics tests. Conversely, the most expensive part of many grant applications that propose to run clinical trials is the exploratory analysis with the latest crop of nonvalidated biomarkers.
One may argue that perhaps randomized trials could be skipped, especially when it comes to appraising all of the new omics tools and technologies. There are so many new tests that it is expensive to assess them with randomized trials. Moreover, these tools are constantly evolving; thus, if one genomic risk score is tested now, this score will alter or evolve while the trial is being conducted. Alternatively, perhaps other types of designs, such as studies with just one participant (nof1 studies) [3] are more commensurate with personalized medicine. These arguments can only go so far. The prostatespecific antigen (PSA) test, for example, was widely overused, and it took decades of debate and millions of misdiagnosed and mistreated patients before extirpating its un warranted routine use from clinical practice guidelines. Genomic tools can yield numerous biomarkers, all of which could fail in clinical use much like PSA. We agree that the targets for evaluation in randomized trials need to be chosen wisely and, indeed, some tests will evolve; there fore, the trials that are conducted will primarily be proof ofconcept experiments. Nof1 designs have value [3], but it is difficult to envisage that they could substitute for traditional randomized trials to answer most questions.
Most of the emerging genomic information that is meandering its way toward health applications is still either nonvalidated noise [10] or true signals with validated small effects, which are not suitable for applying to clinical practice. For the relatively few discoveries that represent more than noise or mere curiosities, random ized trials are indispensable to find out what they can really achieve. Novel trial designs are worth exploring the use of some types of omics information, and modeling, observational, and other nonrandomized comparative effectiveness research might be useful in some cases. Nonetheless, such approaches will not eliminate the need for solid evidence from traditional randomized trials of the type that have been known for over 60 years but have rarely been performed and reported properly.