It feels like every other day that another putative biomarker is identified that will predict presence or extent of some disease or another, usually with an absurdly low p value.  So, if these biomarkers are so common, why is the subsequent commercialisation and clinical use of these potential diagnostics so difficult to achieve?

I believe that the biggest of the problems is in the design of the studies carried out, particularly in the early stages of biomarker research.

The first stage in the development of a biomarker is the study in which it is identified.  This initial study is usually designed to maximise the difference in phenotype between your control subjects and patients with disease.  This is usually thought of as the way of best identifying a biomarker for the disease in question.  However, it should always be remembered that the result of your experiment will always be dictated by its design.  Assuming that the scientific aspects of the study are carried out rigorously, the best outcome of a biomarker identification study can only be a biomarker that best distinguishes your two study groups.  However, the groups of subjects studied in the first identification of a biomarker are rarely those that a clinician will want to discriminate between.  A clinician is seldom faced with the need to determine whether a sample is indicative of a patient with disease or a healthy individual.  More typically their problem is in distinguishing different underlying pathologies with similar symptoms or, in the context of screening, in determining which of two apparently healthy subjects have underlying asymptomatic disease.

Once this initial mistake in study design has been made, it leads to something we all see often – gradually reducing diagnostic power the more you work with a diagnostic.  After the first study has been carried out, you try and repeat your preliminary work, usually with greater numbers of patients.  Another clinician is brought on board, or an additional clinical site.  The study is repeated, and the ability to distinguish subjects with disease from those without disease is markedly weaker than your first study.  This should be entirely expected.  Your biomarker identification study looked for the difference between healthy subjects and those with your target disease.  Often your follow up studies are not actually testing the same thing.  Now you are looking at distinguishing subjects with disease from subjects with similar symptoms, but who may have completely different underlying pathologies.  It should come as no surprise that your sensitivity and specificity has dropped.

At this stage consternation sets in for you and your investors.  You feel that there must be some improvement that can be made in the measurement of your biomarker.  Protocols are tightened up, samples re-assayed, statisticians called in.  But they are all in vain.  Your initial choice of biomarker was flawed, and it is all too late.

So what are the answers to this chain of events?  The answer is actually quite simple.  When you are designing your early biomarker studies, make absolutely sure that you have done all of your homework.  Understand exactly what problems clinicians have when trying to identify particular pathologies, and target your FIRST clinical studies accordingly.  You might be less likely to find that wonderful biomarker in the very first study, but you will quickly find yourself in one of two situations:  either you will fail early (and cheaply, which your investors will thank you for) or you will find a biomarker that is much more likely to stand the test of time.

David Mosedale