Abstract
A recent report in the New England Journal of Medicine on prenatal screening for Down's syndrome (Malone et al 2005) received much attention in the non-scientific press, such as Time (Wallis 2005) and Forbes (Gordon 2005) magazines. This paper describes the results from a multicenter study that showed that, when an integrated, multi-marker approach was combined with a specific algorithm, it was possible to detect 90–95% of fetuses affected with trisomy 21 relatively early in a pregnancy with high accuracy, giving the prospective parents time to make appropriate choices. This study represents the culmination of over 25 years of research that has gradually increased the utility of prenatal screening for chromosomal defects. This effort started in the 1970s with the observation that there was an increased risk for chromosomal abnormalities, primarily trisomy 21, in mothers of advanced age at the time of pregnancy (Hook 1976a, b). In the mid 1980s it was shown that combining age with the results from AFP (alpha-fetoprotein) and beta-hCG (human chorionic gonadotropin) measurements in the maternal serum could increase the detection rate from below 40% to around 50–55% (Bogart et al 1987; Cuckle et al 1987). By 1990, the so-called triple test was developed, which takes into account the maternal age in combination with the levels of AFP, beta-hCG, and uE3 (unconjugated estriol) in the maternal serum during the second trimester of pregnancy, thereby achieving a detection rate of up to 70% for trisomy 21 (Cuckle et al 1988; Wald et al 1989). This detection rate was further improved a few years later with the addition of dimeric inhibin A measurement to form the quadruple test, which increased the detection rate to almost 80% (Wald et al 1994a; Aitken et al 1996). Today, either the triple or quadruple test is the standard of prenatal care for expectant mothers and their unborn child. At the same time that the quadruple test was introduced, various groups started to develop tests that could be used earlier in a pregnancy, i.e., during the first trimester, which led to the combination of the levels for the biochemical serum markers PAPP-A (pregnancy-associated plasma protein A) and free beta-hCG with the physical marker of nuchal translucency determined by ultrasound (Wald et al 1994b; Wald et al 1995; Wald and Hackshaw 1997). These developments finally culminated in two landmark studies, one in England (SURUSS) (Wald et al 2003), the other in the USA (FASTER) (Dugoff et al 2004; Malone et al 2005), that demonstrated that a combination or integration of second trimester screening with first trimester screening gives the highest sensitivity for the prenatal detection of chromosomal defects, while at the same time keeping the false-positive rate low.
So, is there a lesson in this story for the early detection of cancer by screening? We believe so. It clearly shows that a combination of multiple markers, each of which by itself has limited sensitivity and/or specificity, can lead to a powerful screening test. In cancer screening, the traditional approach has been to seek the one and only “silver bullet” type biomarker that detects all cancers early without too many false positive results. The latter characteristic is especially important because of the relatively low incidence of individual cancers among the general population, and the potentially multiple invasive, complex and expensive follow-up tests triggered by a positive screening result. Unfortunately, this magic marker has proven elusive, the occasional claim to the contrary notwithstanding. In contrast, the scientific literature is full of reports on initially promising new biomarkers that, when further investigated, turn out not to be that good after all. But are they really not good, or is it just a matter of the paradigm that they are expected to fit? Perhaps, if a similar strategy to that successfully used in prenatal screening were adopted, some of these markers could be useful after all (Li et al 2005). Why not develop a multi-marker screening approach for cancer? Indeed, some form of multi-marker testing is already in usage, and more are in development. For example, in the USA screening for prostate cancer in men over 50 years of age now consists of a digital rectal exam (DRE; a physical marker) and a PSA (prostate-specific antigen) test in serum (a biochemical marker), that together have a substantially higher detection rate than does either marker alone (Catalona et al 1994; Schröder et al 1998; Schröder et al 2001; Vis et al 2001). Furthermore, in cases where the results put the patient into the elevated risk, yet diagnostically ambiguous grey area – a PSA value between 4 and 10 ng/ml and a negative DRE – a second biochemical marker, namely free PSA, can be added, that helps to distinguish benign from malignant conditions (Catalona et al 1995; Raaijmakers et al 2004). While far from perfect, and while certainly not yet as sophisticated as the integrated screening for fetal defects, this triple-marker approach for prostate cancer has proven capable of detecting a substantial number of early cancers, although the effect of early detection on overall mortality from prostate cancer is still unclear (Postma and Schröder 2005). Other multi-marker test protocols, the ProstAsure index (Babaian et al 1998) and ProstataClass (Stephan et al 2002a, b), use neural network-derived, non-linear algorithms to calculate the risk for prostate cancer from values for three serum markers and the patient's age (ProstAsure), or free and total PSA, DRE, patient age and prostate volume (ProstataClass). The latter algorithm introduces another physical marker, prostate volume, that is derived by an imaging technique, namely transrectal ultrasound or TRUS (
Another cancer for which a multi-marker approach has been attempted is ovarian cancer, for which currently no single good early detection marker exists. The most widely used biochemical marker for ovarian cancer is CA125 (Bast et al 1998). However, while most advanced-stage cancers show elevated CA125 serum levels, the marker's sensitivity for early-stage disease is less than 50% (Skates et al 2004). In order to improve early detection sensitivity without sacrificing specificity, one or more additional markers, including, among others, CA72–4, CA15–3, OVX1, lipid-associated sialic acid (LASA,) and macrophage colony-stimulating factor (M-CSF), were added (Terry et al 2004; Bast et al 2005). The various combinations of markers were then evaluated with a number of prediction models and methods, including artificial neural networks, logistic regression, classification trees and mixture discriminant analysis (Berek and Bast 1995; Woolas et al 1995; Zhang et al 2000; Skates et al 2004). As for prostate cancer detection, there was a clear performance improvement when using multi-marker analyses, compared to measurement of CA125 alone. Furthermore, a combination of new biomarkers identified through proteomics strategies also exhibited increased sensitivity and specificity for the early detection of ovarian cancer (Zhang et al 2004). Finally, a novel recursive approach has shown that, instead of measuring multiple individual analytes, it is feasible to analyze many markers simultaneously, using mass spectroscopy to determine patterns of markers, whereby the pattern itself becomes the analytical entity. Thus, the pattern is what is diagnostic for the presence or absence of cancer (Petricoin et al 2002). Although the initial study has generated considerable controversy, intensive efforts are now underway to further develop and validate this strategy.
Together, these examples clearly show that multi-marker screening can have its place in early cancer detection. So what, if anything, is holding us back? A priori, there seem to exist no scientific reasons why such an approach should not work for cancer screening. However, a number of issues need to be addressed before any multi-marker test can become the standard of care. As for any new clinical procedure, such a test has to be validated in large, well designed prospective trials.
One of the difficulties is that cancer is a rather complex disease, and therefore the confirmation of whether cancer is indeed present or not after a positive screening result is rarely as easy and unequivocal as a karyotype for trisomy 21. Furthermore, whether the cancer will develop and cause clinical symptoms may not be known for many years, whereas after fetal defect marker testing the ultimate outcome will be known within several months of the analysis. In addition, cancer screening is not just a one-time procedure, in contrast to screening for fetal defects, which has to be conducted only once during a pregnancy. Rather, screening for cancer needs to be periodically repeated during a person's life, thereby raising a number of other issues. These include cost/benefit and the appropriate interval between tests repetitions.
Finally, what is unique about the triple and quadruple test for prenatal fetal defect screening is that neither uses complex models; from an informatics standpoint, their implementation amounts to simply counting positively ruling-in markers. Much of the modeling that is done today in the discovery phase of cancer biomarker studies explores highly computationally intensive multivariate prediction models (logistic regression, CART, PCA, support vector machines, neural networks) using dozens to hundreds, if not thousands of individual markers. We wonder if, and how, these types of prediction models will generalize for use in the clinic, for several reasons. First, there is the technical challenge to accurately and reproducibly measure a large number of markers. Second, the current models seek and use all markers in a panel in all patients in a fixed manner. Third, the output, while resulting in a specific class prediction, can be difficult for clinicians to interpret. This is in contrast to the triple/quadruple test, in which markers that ‘miss’ a positive clinical diagnosis for Trisomy-21 are provided ‘back-up’ by other markers that provide evidence for a positive diagnosis. Finally, the relative performance of models generated using the various alternatives have not been sufficiently explored.
While we think that complex fixed-marker models should continue to be explored, simpler models may be possible and ultimately needed to avoid some of the real and/or perceived limitations that lead to resistance in the adoption of the complex models. Researchers involved in cancer biomarker model development should become aware of and consider simple current clinical diagnostics solutions, such as the triple/quadruple test, as a guide for their own diagnostic model formulation.
These and other key issues obviously need to be addressed before any new test will have a significant impact on early cancer detection. Unless new and better individual cancer-specific markers are discovered, a multi-marker testing approach may hold the greatest promise for the improved early detection of cancer.
Footnotes
Acknowledgements
We would like to thank Dr. James Lyons-Weiler for suggesting this contribution, and for engaging in a discussion on simple vs. complex disease prediction model paradigms for cancer biomarker development.