Author: David Harry
It was nice to take a few days off during the holidays—a chance to catch up on some reading and head outdoors.
Among my readings, an intriguing article about genetic testing was published by Kira Peikoff in the New York Times (http://www.nytimes.com/2013/12/31/science/i-had-my-dna-picture-taken-with-varying-results.html). Ms. Peikoff’s article described her personal experience after submitting her own DNA for testing by three companies. The article seemed particularly timely since one of the companies had recently received a stern warning from the US FDA
regarding its direct-to-consumer testing program (http://www.nytimes.com/2013/11/26/business/fda-demands-a-halt-to-a-dna-test-kits-marketing.html). At first glance, Ms. Peikoff’s results appeared surprising because they suggested conflicting interpretations regarding her genetic predisposition for certain afflictions.
While I found all this interesting, I was on holiday, the day was still young, and I had a chance to get outside. I called up my favorite smartphone weather app, which provides access to forecasts wherever I happen to travel. Lo and behold, two sources provided strikingly different forecasts for the day! One called for partly sunny skies and little chance of rain; the other called for clouds and an 80% chance of rain. What to do? I grabbed my weather gear and headed outside.
Luckily for me, the weather that day turned out to be very pleasant. My mind was free to wander, yet my curiosity had been piqued. We readily accept discrepancies in weather forecasts as they happen all the time. Why, then, should we be so surprised that genetic test results from different providers would yield different interpretations of genetic risks? To elucidate this, let’s take a step back and think about what forecasts entail.
Probability and Risk
In the case of weather, meteorologists have amassed an impressive arsenal of instrumentation, computational power, atmospheric mechanics, and historical observations which are all combined using sophisticated probabilistic forecasting models. Local forecasts reflect a greatly simplified interpretation of these complicated models, providing the likelihood of alternative weather scenarios (e.g. chance of rain or whatever) in the hours or days ahead. In my experience, weather forecasts have improved substantially in recent decades, but even so, I tend to give more credence to short-term (next few days) than I do to long-range forecasts. Why? Different forecasting models predict different outcomes, and differences among the models are more pronounced for longer-range forecasts. Forgive the cliché, but it doesn’t take a rocket scientist to figure this out. Simplistically, weather forecasts help us evaluate our individual circumstances so that we may balance risks and plan activities in the future.
At one level, weather forecasts and predictive genetic testing share common themes. Both provide insights into the likelihood of future events based on appropriate and available predictive models. Yet whereas meteorologists have benefitted from decades of modern climatological observations and experience (not to mention hundreds of years of anecdotal history), the science of predictive personal genomics is still in its infancy. The first draft sequence of the human genome became available only in 2000, and most gene-based risk assessments have been developed since then. Moreover, gene-based predictions typically evaluate different genetic markers (e.g. DNA signposts), a fundamental problem which has been likened to determining a novel’s plot after reading random letters drawn from a small percentage of its words. Different readers, with access to different letters and words, are likely to derive different interpretations. Another fundamental difference is timeframe. As discussed, weather forecasts for the immediate future are often given more credence than are long-range forecasts. Predictive genetic tests attempt to assess the likelihood of outcomes that, should they occur, may not become obvious for many years. Even for those relatively few genes for which strong disease relationships have been established (e.g. BRCA genes with cancer risk), many other non-genetic factors modulate the likelihood of disease. The vast majority of tested genes are at best only loosely tied to a genetic disease, and for these, the intricate complexities of “nature vs. nurture” are even greater. Does this imply that genetic diagnostics are meaningless? Perhaps not if the results are interpreted in the proper context. But how can any one individual interpret their personal results?
Statistical Forecasts and Individual Consequences
Both weather models and genetic models predict alternative outcomes with varying degrees of likelihood. But if weather models can (and often do) produce incongruous forecasts, why not genetic models? Forecasts attempt to provide information to evaluate possibilities—they do not claim to tell us what will happen. Over the years, we have realized that weather forecasts are often helpful, and yet they are far from perfect. Genetic testing, especially of the sort provided by personal genomic service companies (as in Ms. Peikoff’s example), has a long way to go to improve reliability and consistency.
To evaluate the overall merits of genetic testing, two additional issues come to mind. The first concerns methods to evaluate the accuracy or precision of the estimated likelihoods. Political pollsters, for example, routinely evaluate the precision of their estimates. A polling result of, e.g. 46% +/-2% conveys a sense that the actual likelihood is somewhere between 44-48%, whereas a result of, say, 46% +/- 25% conveys a sense that the results are unreliable. Might such estimates of precision become available for predictive genetic tests? Time may tell.
Another factor, which in my opinion is both more subtle and even more important, concerns the meaning of probabilistic interpretations as applied to populations vs. individuals. Statistical models (i.e. of genetic predispositions) may be meaningful when applied to larger samples, but individuals want to know what such results mean for themselves. Even though a given genetic predisposition may indicate a low overall risk of disease (at a population level), such likelihoods are meaningless to those individuals who do end up suffering the affliction. For people, evaluating genetic predispositions must also keep individual consequences clearly in mind.
Application to Breeding
Time and again, practices in agricultural genetics and breeding are informed by advances in human genetics. As with previous blog posts, I often conclude by evaluating my musings in light of agricultural applications. Relative to human medical genetics, an obvious advantage in breeding is that, at least for the most part, we need not overly concern ourselves with individual plants or animals. We are certainly interested in changing (i.e. “improving”) the overall average performance of a population (as a collection of individuals), but the individuals themselves are typically dispensable. Hence, provided genetic models are “reasonably” accurate, they can be used to our advantage. Because predictive models deal with probabilities, certain individuals with above-average genetic merit will sometimes be discarded, and likewise, others with below-average genetic merit will be retained. The basic principles of building and refining genetic models are similar for both medical geneticists and agricultural geneticists, yet the take-home message is that utilizing genetic testing is more straightforward for agricultural geneticists than it is for medical geneticists. Yes agricultural geneticists have much to learn (and gain) by applying genetic predictions, and are lucky in that ethical considerations are not nearly as consequential as in disciplines such as medical genetics.
David Harry, Ph.D., is TerViva’s R&D Director