The process to discover and develop molecular biomarkers for cancer diagnosis (or prognosis) is a work in progress and is evolving. "Discovery" research -searching for markers by use of highthroughput technology without an a priori hypothesis -has produced few useful biomarkers over the last 10 years despite numerous claims in research publications and news reports about "95% sensitivity and 95% specificity" or better. The degree of disconnect between claims and products should make us ask why

more »

... rogress is so slow: Is it the normal stop-and-start of science? Or is there some systemic problem with the process that we currently use to discover and develop markers? After all, it took decades to evolve the process to discover and develop drug therapies. So, where does the evolution of the discovery and development of biomarkers stand and how can we improve the process? Although the process is complicated, there are two major questions to address in research. First, can a marker discriminate well? That is, can it discriminate cleanly and reliably (eg, reproducibly and not due to artifact or bias) between persons with early-stage cancer vs those without cancer (or good vs bad prognosis)? Second, does that discrimination, when coupled with an intervention such as surgery or chemotherapy, lead to improved outcome? Answering the second question requires a randomized clinical trial, a design that is powerful and well understood but also expensive and time consuming. A randomized clinical trial is not even considered until a marker clearly shows reliable discrimination that warrants such an effort. Failures in current marker research occur not at the randomized clinical trial stage but earlier, when results that appear to have promising discrimination in early discovery turn out not even to be reproducible. Pepe et al. ( 1 ) proposed a formal structure in 2001 "to guide the process of biomarker development" consisting of fi ve "phases [that] are generally ordered according to the strength of evidence that each provides in favor of the biomarker, from weakest to strongest [and] the results of earlier phases are generally necessary to design later phases." The phase structure has been widely adopted for use in various research projects, including, as noted by Feng et al. ( 2 ), "by the EDRN [the National Cancer Institute's (NCI) Early Detection Research Network], a number of Specialized Program of Research Excellence (SPORE) consortiums, and other biomarker research projects." With this background, the article in this issue of the Journal (3) from key authors of the phase proposal provides insight about the evolution of this process and about how to improve it. The article proposes a specifi c study design for the step just before a randomized clinical trial is done to demonstrate reliable discrimination in the kinds of persons who might then be studied in a randomized clinical trial. Reliability is achieved in design by prospective-specimen collection (by which the authors mean before the diagnosis is known) and retrospective-blinded-evaluation (PRoBE) ( 3 ). The goal of the design is to avoid biases that compromise so much biomarker research: "Common biases that pervade the biomarker research literature would be eliminated if these rigorous standards were followed" ( 3 ). The fact that a proposal is being made now about study design provides lessons about the state of evolution of marker research. Although using a design intended to avoid bias may seem natural and even taken for granted, it may surprise laboratory investigators unfamiliar with clinical research design that avoiding bias is not achieved at all by using phases. Phases provide a list of research questions and a rough idea of the kinds of subjects to study ( 4 ); they do not specify any details of research design that protect against the biases ( 4 ) that, as stated by Pepe et al. ( 3 ) "pervade the biomarker research literature." In the 6000-word article about phases ( 1 ), the word bias appears once; in the proposal for PRoBE ( 3 ), bias is the main focus. Design (and avoiding bias) and phases are simply different topics in the larger process of marker discovery and development. The PRoBE design is already being used in a few settings and