23andMe & Navigenics’ Open Letter to Nature

The following is a joint letter addressing the Opinion piece by Pauline C. Ng, Sarah S. Murray, Samuel Levy and J. Craig Venter that appeared in the October 8, 2009 issue of Nature (coverage here, here, here, and here).  Unfortunately, Nature could not publish the letter because of space restrictions, so 23andMe and Navigenics decided to publish the letter to our respective sites.

Dear Editor:

We read with interest the Opinion piece entitled “An agenda for personalized medicine” in the October 8, 2009 edition of Nature. Our two companies, though commercially distinct with differentiated products, would like to respond to this piece jointly to show our commitment to working together in an open, transparent fashion.

Our companies agree with most of the recommendations Ng and colleagues made.  Without doubt, genotype-based risk prediction for common, multifactorial diseases is still in its infancy.  More work must be done to standardize markers used; to better explain the contribution of genetics to common, complex diseases; and to incorporate common genetic variants into clinical practice.  Each company, however, has a few points of disagreement and/or explanation it feels important to articulate.  These points from each company follow.

Response by Navigenics:

With regard to the specific recommendations, Navigenics agrees with most of the suggestions.  For example, we agree with the authors that results showing less than average risk should not be a primary point of focus, a viewpoint that has been incorporated into our service offering in a variety of ways.  Ng et al. recommend “that DTC companies report the proportion of the genetic contribution of a disease that can be attributed to the markers used in their test…” There are many metrics that can be used to describe the completeness/accuracy of these types of tests. However, each of them has advantages and disadvantages, and all can be misinterpreted by experts and laypersons alike.  Furthermore, the call for such information must be put in context with currently implemented non-genetic risk communication. For example, does your doctor know/communicate what percentage of the total risk of cardiac disease is contributed by your cholesterol level?  Or your family history?  Clearly, risk communication has, and will continue to be, an important area of research for the community.

We agree that associations must be replicated in other ethnicities; that prospective studies will be helpful in further assessing the validity of predictions; and that sequencing should be used when the technology becomes more affordable.  The monitoring of behavioral outcomes is another important avenue for future research, and to this end, Navigenics is collaborating with the Scripps Translational Science Institute on a 20-year longitudinal outcomes study.  Pharmacogenomic markers may also offer immediate value to individuals.

We also agree on the importance of including the same strong-effect markers, and with a few exceptions, our companies are consistent.  A standard set of markers would be valuable to the industry and personalized medicine in general, and it may be most practical for a third party to assess clinical validity.  The catalog of genome-wide association loci sponsored by the National Human Genome Research Institute is an example of such a resource.  Further public-private efforts could be placed into grading the cumulative evidence supporting various marker-disease associations using, for example, the Venice criteria.

Regarding the use of surrogate risk markers, Navigenics initially used markers in linkage disequilibrium with published SNPs (with a requirement of r2 = 1) to tag SNPs that were not on its genotyping platform. However, Navigenics now directly targets published SNPs, except for a few loci in the HLA region.

Response by 23andMe:

We would like to discuss two technical points about the article.  The authors presented these points in a relatively balanced manner but the subtleties have led to misinterpretations in subsequent media coverage.

The first point is that in the comparison of risk estimates, the thresholds for what Ng and colleagues consider to be “average risk” (0.95 ≤ relative risk ≤ 1.05) are somewhat restrictive.  Using this definition would be akin to telling an individual that her risk of a condition was “increased” because it was 5% more than average—technically true, but unlikely to be meaningful.  While there is no scientific consensus on what magnitude of risk equates with “meaningful” increased risk, one typically does not find relative risks less than 1.5 used today in clinical practice (e.g. family history, non-genetic biomarkers, environmental risk factors).  Thus, it would have been more appropriate to perform the consistency comparison using a wider window for “average risk”.

The second point concerns measures of genetic contribution to disease.  Ng and colleagues correctly note that for some diseases, less than half of the genetic contribution can be explained by known markers, which may lead to false negatives and false positives.  However, the metrics Ng and colleagues use to assess genetic contribution―heritability and proportion of genetic contribution explained―are population-level measures that should not be applied to individual-level risk estimates, as illustrated by two examples.

Estimates of the genetic component (heritability) of Parkinson’s disease range from 0% to 27%, meaning that genetics explains at most 27% of the variance in disease risk.  However, the G2019S mutation in the LRRK2 gene confers a 40% to 60% lifetime risk of developing Parkinson’s disease, compared to 1% to 2% on average.  The low heritability cannot—and should not—be interpreted to mean that the risk estimate of 40% to 60% is inaccurate or irrelevant, or that relative risk due to genetics must be small.  (Visscher et al. address common misconceptions about heritability in a 2009 review.)

Ng and colleagues also note that the contribution of known genetic markers to disease can be measured using percent variance explained, another population-level measure.  This statistic, too, can be misleading if applied at an individual level.  As an example, risk data for BRCA1/2 mutations and breast cancer in Ashkenazi Jews can be entered into a liability threshold model, which has been used to estimate percent variance explained by markers discovered in genome-wide association studies.  According to the model, the three BRCA1/2 mutations most commonly seen in Ashkenazi Jewish populations only account for about 2.5% of the variance in breast cancer risk in that population (Table 1), primarily because of the rarity of the mutations.

It would be misleading to advise a woman receiving a positive result for the 5382insC mutation in BRCA1 not to take the 81% lifetime risk of breast cancer seriously just because that mutation explains only 1.1% of the variance in breast cancer risk.  Though we agree that showing the percent variance explained by reported markers could indicate the state of genetic research on a phenotype, a low value does not necessarily mean that individual-level estimates of risk are unreliable or should be disregarded.

Mutation

Relative risk (RR)

Breast cancer risk

(RR x 12.5% average risk)

Carrier frequency

Percent variance explained

BRCA1 185delAG

2.90

36%

0.0092

0.6%

BRCA1 5382insC

6.44

81%

0.0026

1.1%

BRCA2 6174delT

2.90

36%

0.0120

0.8%

All 3 mutations

2.5%

[Table 1.  Percent variance of breast cancer risk explained by BRCA1/2 mutations in Ashkenazi Jews, calculated using model in Raychaudhuri et al. (2008).  All data (except percent variance explained) are from Fodor et al. (1998).]

In closing, both of our companies thank Ng and colleagues for their serious consideration of genomics and personalized medicine.  We welcome further dialogue on how best to improve our offerings to the public.

Sincerely,

23andMe, Inc.

Navigenics, Inc.






  • Pauline Ng

    First, I would like to thank Navigenics and 23andMe for participating in this study, as we were unable to properly acknowledge this in the opinion piece due to space restrictions. These companies donated their tests so that we could do such a study, and our recommendations would not have been possible without their participation.

    We would like to address the two main points. The first point, raised by 23andMe, was our use of restrictive thresholds for “average risk” (0.95  relative risk  1.05). We used these thresholds because we wanted to pinpoint differences in the companies, if any. When wider thresholds were tried, there were many more “average” risk predictions, thus preventing a more detailed analysis.
    23andMe uses the thresholds of 0.8 <= relative risk <= 1.2 for "typical" risk. When we applied this new threshold, approximately 1/3 of the predictions do not agree, which is a number similar to what we stated in our opinion piece.

    We could expand the threshold to 0.5 <= RR < = 1.5, but as less restrictive thresholds are used, there are more "average" risk predictions. This opens up the question if consumers would find these tests attractive if most predictions state "average risk." For example, using 23andMe's threshold, all individuals would have an average risk of heart attack, as seen in the table above.

    The second point we would like to address, raised by both Navigenics and 23andMe, relate to reporting known genetic contribution to disease. Most of the markers tested by DTC companies are variants common in the population, with minor effects on an individual's health. These markers largely were identified by genome-wide association studies. These studies are ongoing, and we, along with many in the genetics community are concerned about the missing heritability in these studies (see a paper by the Director of National Institutes of Health on this issue http://www.ncbi.nlm.nih.gov/pubmed/19812666). A consumer could be misled into thinking he or she is in a low risk category because the currently known markers point to an average or below-risk of disease. When in actuality, the consumer may have a disease variant that the genetics field has not yet discovered. It is this disclosure of how much is unknown that we advocate. An analogy would be consuming a "fruit drink", would you like to know if your drink has 20% juice or 5% juice?

    23andMe cited mutations in breast cancer and Parkinson's disease as examples, but these mutations are not representative of the markers typically used by both 23andMe and Navigenics for disease prediction. The variants referred to in 23andMe’s examples are actually mutations with strong effects but that are rarely seen in the human population. We agree that rare variants can have a huge impact on health, but people that carry such rare strong-effect variants typically have a family history for this disease. In this case it may be more appropriate consider a diagnostic company that specializes in the particular disease diagnosis (i.e., Myriad Genetics for BRCA1/2 testing) than employ a genome-wide scan.

    Best Regards,
    Pauline Ng
    (an author on the Nature piece)

Return to top