Ioannidis JPA. Invited Commentary—Genetic Prediction for Common Diseases. Arch Intern Med. 2012;172(9):744-746. doi:10.1001/archinternmed.2012.931
Author Affiliations: Departments of Medicine and Health Research and Policy, Stanford Prevention Research Center, Stanford University School of Medicine, and Department of Statistics, Stanford University School of Humanities and Sciences, Stanford, California.
A major promise of human genetics has been the use of genetic information to predict the risk of common diseases in order to prevent and treat these conditions more effectively. Most common diseases have a complex etiology, and genes are expected to explain much of their risk. However, even though PubMed already retrieves more than 2 million articles with “gene OR genetic” (n = 2 015 109 as of February 10, 2011) and half (n = 1 040 434) are tagged as “Human,” there are formidable difficulties in materializing this promise.1
Genome-wide association studies have now successfully identified thousands of common genetic variants that influence the risk of complex diseases. Large-scale evidence, agnostic testing with stringent statistical criteria, and rigorous replication standards guarantee that this literature has high credibility. Nevertheless, the discovered gene variants do not markedly expand our predictive ability compared with what can be achieved by using only information from long-known traditional risk factors. In this issue of the Archives, Smith et al2 add another example for in which genetic information does not materially improve prediction. Genotyping for 3 robustly replicated gene variants for atrial fibrillation enhanced the area under the curve (AUC) by only 0.009 and 0.005 for prevalent and incident atrial fibrillation, respectively.
This failure is not surprising. The risks conferred by the identified common gene variants are almost ubiquitously small; therefore, they are unlikely to change predictive discrimination perceptibly. Simulations show that new risk factors with small odds ratios cannot change the AUC by more than 1% when there are already other traditional risk factors available that can already achieve modest predictive discrimination.3 Much larger effect sizes and/or a very large number of genetic variants are needed to start making substantial improvements in predictive discrimination. One disease for which we have been closest to such a success is probably age-related macular degeneration. With close to 20 gene variants identified to date, and with some of them having an odds ratio of greater than 2 per risk allele, consideration of traditional plus genetic risk factors for age-related macular degeneration can generate an AUC above 0.90.4 Such predictive discrimination has not yet been achieved for other common diseases, even when we have discovered many dozens of genetic variants with small effects. Apparently, for some common diseases such as type 2 diabetes mellitus or coronary heart disease, the genetic variants that are implicated are probably in the range of many hundreds. Knowing several dozens or even 100 of them is still too few.
In theory, even minor changes in discrimination may have clinical importance if they achieve marked improvements in reclassification.5 Reclassification metrics, such as the net reclassification index,5 examine the extent to which the addition of information from new risk factors allows patients to be recategorized into more (rather than less) appropriate risk categories. Optimal management and best-indicated interventions differ in these categories. Several highly visible articles have claimed highly promising reclassification metrics for some common diseases, eg, 1 article claimed that information on just 9 cholesterol-related variants can markedly improve the categorization of patients with intermediate predicted risk of cardiovascular disease.6 However, simulations show that the net reclassification index estimate is extremely volatile and highly influenced by minor changes in the selection of thresholds to separate the risk categories.7 For almost all diseases, we have no consensus or good evidence on what the ideal cutoffs of risk are that should define categories with different management plans. Even when we do, eg, the categories proposed by the Framingham Risk Score for cardiovascular disease, in practice the selection of cutoffs frequently deviates from the proposed thresholds.8 Therefore, the most extravagant net reclassification index results may simply be examples of how a metric that is susceptible to the choice of assumptions and selective reporting can fuel spurious expectations.
Eventually, we have to prove with pragmatic patient outcomes that genetic or other predictive information works in real life. While discrimination and reclassification exercises provide some initial screening, the proof of principle requires demonstration of benefits in randomized controlled trials. Predictive discrimination or reclassification metrics are unable to fully foretell the results of such trials. For example, an ongoing randomized trial is allocating participants who consider initiation of statins to 2 strategies: either learn their 10-year risk estimate for coronary heart disease with information added from their genomic profile on 19 relevant gene variants or get this genomic information after the end of the trial (clinicaltrials.gov Identifier: NCT01406808). A similar concept is also being tested in another randomized trial using genomic information on diabetic risk.9 Studies on Mendelian disorders, eg, familial hyperlipidemia, suggest that individuals may be responsive to genetic information and may increase adherence to treatment.10 By extrapolation, genetic profiling of common variants may also motivate individuals to modify their behavior, lifestyle choices, and adherence to drug treatment. Preliminary observational data offer some reassuring evidence that provision of such genetic information does not increase the levels of anxiety in individuals11 but also show no major impact on exercise or dietary habits.11 The results of randomized trials will give us some more rigorous evidence on what we can expect to achieve with genetic risk information.
In designing this generation of personal genomics randomized trials, there are hundreds of diseases and genetic variants that one can choose to study. It is probable that priority should be given to common diseases with a high burden of disease as well as to those diseases for which (1) risk information seems most predictive, (2) effective interventions are available, and (3) intervention is likely to be effective and/or cost-effective only when prescribed to patients above a given risk threshold. There is probably a window of opportunity for performing such studies now that a wide segment of the general public is excited, fascinated, or at least intrigued by the prospects of personal genomics. In the next few years, we should be able to find out whether genetic information offers only the joy of learning about the complexity of our genes (an otherwise splendid educational or recreational activity) or whether we can also make some real use of this information for the betterment of health and health care.
Correspondence: Dr Ioannidis, Departments of Medicine and Health Research and Policy, Stanford Prevention Research Center, 1265 Welch Rd, MSOB X306, Stanford, CA 94305 (firstname.lastname@example.org).
Financial Disclosure: None reported.