The Future of RDoC
What are you going to do with RDoC?
This was by far the most common question I got last summer after it was announced that I would take over as director of NIMH. There is intense interest in the NIMH’s foray into categorizing and measuring human behavior, from supporters and critics alike. When I arrived last September I told everyone that I was thinking about this issue, but that I would take my time learning and listening. Well, the academic year is winding to a close, my son came home from his first year of college, and swimming pools are opening up all over the DC area. Over the past nine months, I’ve spent a lot of time learning about the Research Domain Criteria (RDoC) project.i I’ve come to three conclusions. First, the utility of the RDoC approach has yet to be determined. Second, a bottom-up, data-driven approach to defining RDoC domains would be valuable. And third, computational approaches will be useful in defining these domains, connecting them with neurobiology, and testing their clinical utility. In this message I will cover the first two points; the third will be addressed in a follow-up message to come in the next few weeks.
The Short Answer: The Great RDoC Hypothesis
On the face of it, the RDoC project is an excellent idea. Take behavior, break it down into its component parts, and study the psychobiological factors relating to these parts, from genes and molecules to neural circuits and behavior. Indeed, this is a natural extension of what cognitive psychologists, behavioral neuroscientists, and others have been doing for decades. The basic building blocks of behavior should be linked to their neurobiological roots, and therefore more amenable to scientific interrogation than the multifaceted and heterogeneous clinical diagnoses that have proved to be unreliable guides to understanding cause. Couple this with RDoC’s dimensional approach, where within each behavioral domain, traits and behavior are measured along a continuum from subtle to strong, rather than with Boolean (yes or no) symptom lists on which current diagnosis is based, and we can hopefully resolve psychobiological questions that have long evaded resolution.
Key word: hopefully. Because the idea that RDoC will facilitate rapid, robust and reproducible neurobiological explanations for psychopathology (as observed within and across DSM disorders) represents a broad hypothesis. A compelling hypothesis worth exploring, but a hypothesis in the end. And what do we scientists do with hypotheses? We test them. So this is what we must do. We must do our very best to construct these behavioral domains carefully; to conduct well-controlled, sufficiently powered studies asking the appropriate questions about the interface between behavior/cognition and neurobiology that RDoC calls for; and evaluate the results we get. If such studies reveal interesting and reproducible results, we can then move forward to accelerating the applications to clinical trials and prevention. If not, we evaluate why such a seemingly promising framework did not provide sufficient results, and move on to new approaches that can incorporate brain science and our understanding of the symptoms and dysfunction associated with psychopathology.
I would argue that we have not yet reached a position to say whether RDoC has succeeded. So the short answer to the question posed above is: We are going to continue with the RDoC experiment.
The Longer Answer: Big Data
How might we get the kind of data that would actually allow us to better evaluate the RDoC hypothesis? This is a big question, with a big answer: big data.
But to get to this answer, we have to go backward a bit. I noted above that the first step to testing the RDoC hypothesis is to construct these behavioral domains carefully. Have we really done so? I would argue that we have not done so as systematically as we might have. We constructed the RDoC domains using a similar method to what we used to construct DSM diagnoses. We took a bunch of scientists and clinicians and locked them in a room together. Only instead of asking them, “What are the various kinds of psychiatric illnesses?” we asked them, “What are the various kinds of behaviors, and the neural systems that implement them?” Being experts in their field, they came up with answers based on the most well-supported current data. But our understanding of the complex relationships between brain and behavior is incomplete and rapidly changing. By necessity, then, even the consensus decisions of the most knowledgeable experts are bound to be incomplete.
We call the approach used for DSM and RDoC a top-down, expert-driven approach. But for questions as complex and daunting as these, most would prefer a bottom-up, data-driven approach. The challenge with such an approach towards RDoC is the sheer size of the endeavor that would be required. So far we have defined five RDoC domains, with more under consideration. Within these domains, we have defined 23 constructs and 25 sub-constructs (see the RDoC matrix), each representing a facet of behavior which would have to be assayed by multiple behavioral tests, which in turn would have multiple measurable variables. We could imagine measuring all these variables—likely hundreds and possibly thousands—in a group of subjects, and using computational techniques (clustering analyses or dimensional reduction techniques) to examine the underlying structure of behavioral performance across these tasks. Such an analysis would reveal which tests tend to correlate with each other, and which tests tend to give independent results. That is, such an analysis would have the potential to reveal the underlying structure of human behavior.
To be reliable, however, multidimensional analyses such as these typically require large datasets—the number of test subjects should, in general, exceed the number of measured variables by at least an order of magnitude, preferably even more. So if we have thousands of variables to measure, we need tens of thousands, or even hundreds of thousands of subjects. To gather such an immense number of individuals, each characterized comprehensively across a wide range of behavioral tests, will require tremendous resources—or one very carefully constructed resource: the All of Us Research Program. All of Us is a key element of the Precision Medicine Initiative. Currently in its initial beta testing phase, the program will enroll and gather information from a diverse cohort of 1 million or more volunteers from across the United States. Participants will complete surveys, donate biosamples, and agree to make their de-identified electronic medical records available for research purposes. Most crucial for the purposes of RDoC, they will be continuously engaged in the cohort via a web-based portal that will enable their participation in various research endeavors. Our task is thus simplified. Develop a panel of web-based behavioral measures, each covering an aspect of the RDoC universe, and roll these out to a subset of these volunteers—on the order of 100,000 of them—enriched for those with psychiatric illnesses. Gather the data, make it freely available for researchers, and fund those researchers to use data-driven approaches to parse behavior into its basic building blocks.
We are still some years away from being able to implement such a plan, but the NIMH is serious about the effort. We are currently identifying and preparing a panel of web-based behavioral tests and surveys that span the RDoC domains, so that when the All of Us program and similar cohorts currently under development are ready, we will be too. And while the initial use of these big data approaches will be to revise the RDoC domains in a more bottom-up, data-driven way, there are even more powerful potential applications. For more on that, check out my next message.
i For readers unfamiliar with RDoC, here is the key concept behind this initiative: at its base, RDoC is an effort to understand the connection between behavior (and disorders) and biology, taking into account the processes of development (fetus to adult), and the impact of environment and experience on a living individual. One impetus for RDoC is that knowledge of the underlying causes of disorders may provide a deeper understanding of why they occur and, as a result, much better targets for the development of new treatments. Additionally, there is hope that focusing on graded readouts of behavior rather than diagnostic categories may lead to novel approaches to treatment and better measurements of treatment efficacy.