When President Obama announced his Precision Medicine Initiative at his State of the Union address in 2015, I wondered whether he had questioned the wording and knew what he was promising.
The National Institutes of Health describe precision medicine as “an emerging approach for disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle for each person. This approach will allow doctors and researchers to predict more accurately which treatment and prevention strategies for a particular disease will work in which groups of people. It is in contrast to a one-size-fits-all approach, in which disease treatment and prevention strategies are developed for the average person, with less consideration for the differences between individuals.”
Precision medicine assumes that data can predict the success of a treatment for individual patients beyond what works best for them as a group. It assumes that doctors know and agree on what is the success of a treatment and that factors can be identified that increase the likelihood the treatment will work for a patient. If data need to support treatment decisions, then the precision medicine approach also assumes that the success of alternative treatments can be predicted as well and that the predictive ability of all algorithms are such that they can identify single best treatment (if not, then one size did fit all). And in all this, side effects, comorbidities, costs, and patient preferences are not yet considered. The bar for making precision medicine a reality is high.
I study how well DNA can predict complex diseases and traits. From our earliest studies, we observed that predictive accuracy remains poor when many genetic risk factors are combined in algorithms except when one or more factors have a major role in their development.
For over a decade, I have illustrated the genetic prediction of complex diseases in my lectures using examples from predicting other complex outcomes: the crashing of planes and capsizing of ferries. I ask audiences whether they expect that all planes and all ferries crash due to the same causes. I know their answer is no. And I know that we agree that disasters happen because of unique combinations of contributing factors and that it cannot be known in advance which factors need to be measured ahead of time to predict and prevent a disaster from happening.
The same holds for complex diseases. Each patient gets their disease through a largely unique and unpredictable combination of contributing factors. Disease prediction remains probabilistic because it cannot be known in advance which risk factors should be measured, when, and how, to prevent the disease from developing.
Science can improve prediction, but the complexity of the origins of complex diseases will limit the predictive power of tests. We should expect the same limits for the prediction of treatment benefits.
Every now and then, someone in the audience thinks I am too fast with my conclusions. They argue that with more DNA data, combined with more other data, in larger studies, with machine learning and artificial intelligence, with other technological and computational advances that we don’t know about yet, there may be better opportunities for prediction. That it is early days.
These people, I believe, see the limits of prediction as a problem of limited data and computational approaches, not as the reality of nature’s complexity.
I needed another analogy.
I am a novice baker. I have a simple recipe that is only four ingredients (flour, water, yeast, salt), a few steps, and waiting. I weigh the ingredients and follow the steps, but, after more than 25 bakes, I still was failing to get a consistent, good loaf. The taste was good — that’s the ingredients — , but the shape and texture varied from bake to bake. I had no clue why.
This summer, I took a class during a family visit to the Netherlands. In my home town is a beautiful industrial complex, a former leather factory, that now houses the bakery and restaurant of the host of the Dutch edition of The Great British Bake Off, Robèrt van Beckhoven. It is the factory where my father and grandfather bought leather for their shoe factory before the Dutch shoe industry imploded in the early ’70s.
The class was hands-on. We baked more than ten loaves in two days but didn’t see a single recipe. For the first two loaves, we had the correct amounts of ingredients on our benches, for the others we received them already mixed. The class was about the dough.
We learned how to knead the dough and find out when the kneading was enough. People vary in how they knead. They knead with big or small hands, strong or weak, warm or cold, with firm or gentle strokes, fast or slow. These factors determine the kneading time, but since the impact of each depends on all others and may vary during the kneading, it cannot be said beforehand how long the kneading time needs to be.
That’s why we learned how to observe the quality of the dough. Does it need more water, more flour, more kneading? And how does the dough handle the rising? Is the temperature right, the humidity? Does it need more time? A second proof? We learned how to recognize over-kneading and over-proofing, which are lessons learned to improve the next bake.
The class was all about observing and reasoning. Learning the many details that, over time, become intuition and experience not a longer recipe.
I now understand why bread recipes often only specify ingredients and a few steps. When there are too many factors that influence the too many processes that are going on in the dough, it’s impossible to specify them in a recipe. That is why you knead until the dough is a smooth ball that holds its shape and let it rise until the dough has doubled in volume. A recipe specifies the ingredients and assumes a certain level of experience to recognize when the dough is where it needs to be.
It is reasonable to assume that the human body is more complex than a ball of dough — that too many factors determine whether a treatment will work, that some of these factors can be known and measured in advance and used for prediction, but that many others will remain unobservable due to the complex nature of disease and treatment processes. Think of body weight, the timing of drug intake, use of other drugs, diet, and, I almost forgot, adherence, which all influence how well a drug will work.
It seems too optimistic to expect that the response to treatments can be predicted, let alone that these can be predicted for multiple treatments with high enough accuracy to decide which will be best for a patient.
Treating patients is choosing the best possible treatment based on a few indicators, observing if it works, and knowing what can and needs to be done if not.
With more data and advances in data sciences, the prediction might become better, but it is the complexity of nature, of how people become ill and get better, that prevents accurate prediction.