Diagnosing Diamond-Blackfan anaemia: 'Vorsprung durch Technik'.

Abstract

Especially since the turn of the millennium, the growth of precision medicine1 has promised a change in our approach to patient care. The prospect now exists of an objective, accurate and detailed diagnosis, leading to the selection of the appropriate therapy at the appropriate dose for the needs of the patient. This marks a sea-change in the approach to diagnosis and treatment over the period of publication of this Journal, as can be seen from the very first paper published – even the title points to acute leukaemia in general in a way that soon became obsolete.2 However, some conditions persist even today in being diagnosed by exclusion, or in Justice Potter Stewart’s famous definition of pornography, using the principle of ‘I know it when I see it’.3 With, as is made clear by van Dooijeweert et al.4 in this issue of the Journal, with no validated functional screening test, Diamond–Blackfan anaemia (DBA) falls into this category. The results of the study presented here showing a metabolic signature for DBA therefore have great potential interest. The promise of being able to characterise a condition using genomics or metabolomics is seductive. Gone is all subjectivity, replaced by a single test that gives an accurate answer. Even better, the signature offers the lure of the predictive marker: identifying therapeutic targets for tailored therapy and tracking response to that therapy. The success or failure of a treatment can be detected early, while simultaneously the chances of a treatment being ineffective in a particular patient are reduced. However, the derivation of such signatures, especially in rarer conditions is not without challenges. The number of different biomarkers that can be analysed in the laboratory means that one is presented with more potentially explanatory variables than patients on which to test them. If sufficiently many markers are available (with data on many genes or many proteins), good discrimination can be seen even if these markers are chosen entirely at random from the set available. For example, on a set of 59 patients with chronic myeloid leukaemia it was shown empirically that it is possible for 30 randomly selected gene expression probes to predict response correctly, in that dataset, on average 90% of the time.5 This is a manifestation of a familiar high school mathematics problem: with n unknowns and n independent simultaneous equations there exists a unique solution. The high degree of accuracy in discriminating responders from non-responders is because the variables chosen fit the data, as opposed to the disease. The other challenge comes in the design of the experiment: in terms of diagnosis one is typically aiming to distinguish cases from controls. Often these controls will be healthy, or use a convenience sample of material from unrelated procedures. Laying aside the issue that one may in fact be identifying a factor associated with the selection of the controls,6 do the controls chosen actually represent the population in which one is looking for the correct diagnosis? For example: if one were looking to identify patients with acute promyelocytic leukaemia (APL) in one’s clinic, then it would be important to distinguish between APL and acute myeloid leukaemia (AML). A test that was derived based only on APL cases and healthy controls might identify factors common to AML and APL that are not present in healthy controls – rendering it all but useless in the clinic. In other words the test needs to distinguish between those with the condition and those who might reasonably be suspected to have the condition but do not.6 So how does the work presented by van Dooijeweert et al.4 fare when considered on these two points? There are two particular features here that provide grounds for optimism. First, although small, there is a validation cohort that shows similar detection rates to the original training set. And second, and indeed rarer, is a comparison between DBA and congenital dyserythropoietic anaemia (CDA) demonstrating a different signature for DBA than for CDA. Both of these features are heartening: that this signature fits the condition, as opposed to being a chance ‘best fit’ to the dataset provided; and second, that the signature can distinguish DBA from other similar conditions – the test that a diagnostic would need to pass in a clinical setting. As the authors themselves admit, there is more work to be done. Larger independent validation in a real-life clinical setting will reinforce the findings. And at present there is no indication of how the signature can work to identify therapeutic options, or as a serial marker of disease activity. But the findings here provide proof of principle for an approach that is less invasive than current techniques, and may lead to an objective set of diagnostic criteria, and a gateway to choosing therapy and monitoring its effect.