January 18, 2011

Personalized Medicine is not just about therapeutics, it is also about personalized diagnostics

By: Douglas S. McNair, MD, PhD

Most people think of “personalized medicine” as choosing the best (most-effective, safest, etc.) treatment according to the details of the attributes of the person, including genetic and genomic attributes. What's not generally recognized is that personalized medicine also involves personalizing the decision-levels of diagnostic tests for the individual--taking the numerical limits of the reference ranges (“normal ranges”) of laboratory tests and vital signs and other diagnostic and monitoring procedures, and tailoring those ranges to reflect what is “normal” for persons with those specific attributes.

A simple but important example was published last month by Julius Gudmundsson and coworkers at deCODE in Iceland (see link below), involving genomic effects on the range of “normal” prostate-specific antigen (PSA) levels in men who are being screened for prostate cancer. They find that, depending on your genotype at a series of commonly-occurring variations (single-nucleotide polymorphisms; SNPs) in several genes, age-adjusted PSA “normal” values may be lower or higher than overall population “normals” by up to -40% or +56%.

Most of Gudmundsson's attention is directed to those variations (combinations of genotypes, of two or more SNPs at the same time) that produce elevated “normal” values, which in turn lead to large numbers of prostate biopsies being performed in men who do not have cancer, because the elevated “normal” PSA values are normal for them.

In other words, the SNP alleles associated with [higher-than-population-average “normal” baseline-] PSA levels are significantly less common among men with a positive biopsy than in men who have a biopsy that comes back negative for cancer. The results for these variants show that the alleles associated with “increased” PSA levels increase the probability that a normal prostate will be biopsied [biopsied needlessly and more frequently than serves any useful health purpose, which leads to excess morbidity and risk, excess cost, and excess anxiety].

But here is the reason for this blog post: there is something that Gudmundsson and coworkers do not mention--something that is more worrisome than needless biopsies that find only normal tissue: the failure to biopsy and the failure to find high-grade aggressive cancer early when a difference in outcome might be achievable. The alleles associated with lower-than-population-average baseline PSA levels are significantly less frequent among men with a negative biopsy than in men with a positive biopsy. That means that, for men with those alleles, their low-percentile PSA levels will tend to cause false reassurance, 'stand-down' decisions on biopsy (due to the still-“normal” PSA level), diminished cancer detection opportunities when undergoing PSA-based screening, and consequent failure to detect prostate cancer early.

"An error common throughout the history of prostate cancer screening might be described as self-fulfilling prophecy: If you don’t look for it, you won’t find it."
--Ian Thompson and Donna Ankerst, 2007

log_PSA_distributions_by_haplotype_class

The conventional set of un-personalized PSA diagnostic decision-levels ignores the fact that genomic variations cause 63% of men with prostate cancer to have PSA levels lower than 3 ng/mL, a commonly-used numerical cut-point above which biopsy is offered.

To solve that problem, you need to increase the statistical specificity and sensitivity of the PSA test's decision-levels by working out a statistical model that defines the “normal” PSA value for each man. Such a model would have to take into account age and race, because the prevalence of both benign and malignant conditions that can cause serum PSA elevation increases with age and because of other factors that are associated with the person's race.

Sequence variants are also important when determining personal normal PSA levels, because it has been shown that expression of a large percentage of genes is under the genetic control of both cis- and trans- components (see Emilsson 2008). Using the genotype information for the SNPs above, it is possible to figure out whether your PSA level of 1.8 ng/mL is, say, at the 77th percentile for your age and race (in which case you will probably want to proceed with a biopsy, even though your PSA is still far below the un-personalized conventional decision-level, because the biopsy is likely to catch an aggressive cancer and catch it early).

Alternatively, for your haplotype at these 6 SNPs it is possible that your 4.9 ng/mL PSA places you in the 12th percentile for your age and race (in which case you can safely not go forward with biopsy right now, even though the PSA is above the un-personalized conventional decision-level, because a biopsy would likely only find normal, noncancerous prostate tissue).

The number of haplotype combinations of 6 SNPs, each of which can have 3 genotype values, is 36 = 729. But, epidemiologically, some of those haplotype combinations occur with some SNPs' genotype values frequently together, and other of the haplotype combinations give rise to just about the same PSA concentration distributions by age and race as other haplotype combinations. So, to develop a personalized model, we don't need to do a study that includes hundreds of thousands of men, to capture a large sample of each of the 729 haplotype combinations that are possible. Instead, it is practical to sort the haplotypes in to, say, <a href="http://en.wikipedia.org/wiki/Quantile" target="_blank">quintile</a> categories according to the logarithm of PSA concentration (as we did to create the graphic above). That way, it is possible to create age- and race-adjusted genomics-based personalized PSA diagnostic decision-levels with a study that has only about 50,000 men in it, still a large number but far more practical than the alternative.

• rs2735839 (20% lower baseline PSA per copy of 'A')
• rs4430796 (homozygous 'GG' 9% lower PSA)
• rs10993994 (homozygous 'CC' 9% lower PSA)
• rs10788160 (homozygous 'GG' 8% lower PSA)
• rs11067228 (homozygous 'GG' 8% lower PSA)
• rs401681 (homozygous 'TT' 7% lower PSA)

These SNPs are included in your raw results from direct-to-consumer (DTC) genomics companies like deCODEme and 23andme, who use Illumina microarray chips. The companies do not provide detailed statistical guidance like what we are discussing above. But you can look at your results and discuss them with your doctor.

One last thing: Gudmundsson and his coauthors don't discuss the relevance of these same SNPs for decision-making about prostate cancer time-to-recurrence, PSA doubling-time after recurrence, androgen ablation, etc. But it is highly likely that the same inherited polymorphisms also underlie population variations in those PSA-centric diagnostics too, not just variations in baseline PSA.

Even without genomic markers data, the Gudmundsson paper tends to support the position of some leading urologists who favor using 'relative' (percentage) PSA increases from post-radical prostatectomy PSA nadir concentration, or the PSA doubling-time metric, instead of transgression of an 'absolute' PSA threshold such as has been conventional [un-personalized] practice in the U.S. for many years.

Deantoni E. Age-specific reference ranges for PSA in the detection of prostate cancer. Oncology 1997;11:475-89.

Emilsson V, et al. Genetics of gene expression and its effect on disease. Nature 2008;452:423–8