The Research Domain Criteria (RDoC) is an initiative developed by the National Institute of Mental Health (NIMH) to guide research and facilitate better communication about mental disorders, and psychopathology in general. Recent advances in neuroscience have not offered significant improvement in treatment modalities and patient care for people afflicted with mental health problems. The RDoC project is an attempt to address the heterogeneity of diagnostic categorizations and the lack of progress in research into the neurobiological foundations of mental disorders. The core of RDoC is based on a Matrix in which functional aspects of behavior, named Constructs, are related to genetic, neurological, and phenotypic research findings, along with the various assays, self-reports and paradigms that generate the data used to make such findings. The RDoC Matrix suffers from several problems, which need to addressed before it can deliver on the NIMH's long-term goals of fostering translational research via the broad sharing of data relevant to psychopathology. One of the most difficult challenges for RDoC is in providing researchers and users of the Matrix a formalized unambiguous way of linking findings in genetics, molecular biology, and neuroscience to the constructs for which they are thought to be associated. The purpose of this paper is to discuss those challenges. We expand our previous analysis of the RDoC matrix and introduce an ontological representation of the Constructs, the RDoC Ontology (RDoCOn), that provides a method for incorporating the RDoC framework with current biomedical ontologies. We demonstrate a way in which particular Elements in the Matrix can be usefully linked to Constructs.
The Research Domain Criteria (RDoC) is an initiative
developed by the National Institute of Mental Health (NIMH)
to guide research and facilitate better communication about
mental disorders, and psychopathology in general. The
NIMH offers RDoC as a framework for conducting
research, one that is based on dimensions of observable
behaviors and neurobiological measures
The clinical categorizations of mental disorders as seen in
the Diagnostic and Statistical Manual of Mental Disorders
(DSM) and the International Classification of Disease (ICD)
repeatedly fail to correlate with valid sets of biomarkers that
could be useful for diagnosis or treatment
The core of RDoC is based on a matrix in which
functional aspects of behavior, named Constructs, are related to
genetic, neurological, and phenotypic research findings,
along with various assays, self-reports and paradigms that
generate the data used to make such findings. Grounding the
Constructs are eight Units of Analysis: Genes, Molecules,
Cells, Circuits, Physiology, Behavior, Self-Report, and
Paradigms. Particular Elements, such as brain-derived
neurotrophic factor (BDNF), Dopamine, Hypothalamus, Fear
Potential Startle, or Drifting Double Bandit, populate the
cells for each of the 41 Constructs (Table 1). Central to the
long-term aims of NIMH’s vison for RDoC is to foster
translational science and the broad sharing of data relevant
to mental disorders. As part of realizing this goal, an RDoC
Database (RDoCdb) has been created
Negative Valence Systems
Acute Threat
(“Fear”) Potential Threat
(“Anxiety”) Sustained Threat
Loss Frustrative
Non-reward
Bsteridanteurcmleiunsaloisf Potentiated startle Hypothalamic nuclei
Dysregulated HPA axis
Avoidance Amygdala neuroimmune
Anhedonia
Stranger Tests
Contextual Threat Physical and relational aggression
Change in attributional style Buss-Durkee and Social dominance
Buss Perry test 5-HTTLPR
GABA
Parasympathetic system
In health informatics, there is a clear and demonstrable
need to provide standardized, extensible and semantically
interoperable solutions for the sharing of data and
markingup of information sources to enhance knowledge discovery.
It is widely accepted that modern healthcare requires
highquality information, which is readily accessible, easily
searchable, reliably encoded, and stored and structured in
such a way as to provide the capability to be automatically
manipulated and reasoned over
Perhaps the most visible, and arguably most successful,
example of the expanding role of ontologies is the use of the
Gene Ontology (GO) in describing the associations of gene
products (proteins and RNAs) with biological processes,
functions, and cellular components
The purpose of this paper is to propose how realismbased methods for developing and utilizing ontologies can be used to support and improve RDoC and its methodology, for example in the linking of GO annotations to research which utilizes the RDoC Matrix. We summarize previous work in which we analyzed the Matrix for terminological and ontological principles. We then illustrate how we have redefined the RDoC constructs as bodily systems which bear functions, and discuss existing resources that can be incorporated, along with their limitations, and what will need to be improved. 2
The rows in the Matrix consist of Constructs, which are grouped together into five broad Domains that attempt to represent our current understanding about key psychological systems. The RDoC Domains are: negative valence systems, positive valence systems, cognitive systems, social process systems, and arousal and regulatory systems. The columns in the Matrix are the Units of Analysis, which are populated by Elements that are thought to be associated with any particular construct. (Table 1).
We discovered a variety of problems with the matrix as it
is currently formulated
From an ontological perspective, we would like to see clear definitions of what a ‘Gene’ and ‘Molecule’ is in the RDoC system. We would suggest that the RDoC system refer to genes in the modern genetic sense of a DNA sequence encoding a protein or functional RNA, and refer to encoded proteins or RNAs directly when particular experimental data address the nature of those types of entities directly. This is particularly important when considering spliced forms of proteins or proteins with different types of post-translational modifications. ‘Molecule’ might best be reserved for non-polypeptide entities, i.e., specific biochemical or small molecules such as dopamine, norepinephrine, and acetylcholine that play roles in the functioning of the nervous system. For all these entities, appropriate terms should be selected from existing OBO Foundry ontologies.
In addition to there being a lack of clarity in how to determine whether particular Elements belong as Genes or Molecules, there similarly exists overlap between Cells and Circuits, Circuits and Physiology, and Behavior and Paradigms. Furthermore, there is the problem of how to distinguish between all of these Units of Analysis and the Constructs themselves. We contend that this is because RDoC Constructs are not defined well enough to unravel this overlap. For example, ‘Animacy Perception’ is defined as “the ability to appropriately perceive that another entity is an agent’, while the term ‘ability to appropriately attribute animacy to other agents’ is an Element within the Behavior Unit of Analysis. How would this Element be related to the Construct? Via equivalency? It would seem that an Element should not be regarded as an ability, but rather as a process which realizes that ability.
While the developers of RDoC and the Matrix admittedly regard it to be in a nascent state and open to revision, to serve as inspiration for guiding research, if it is to provide the kinds of semantic integration needed for translational research and data sharing, a more consistent method for development should be considered. We contend that the best way to ensure success for future iterations of RDoC and the Matrix is to formalize the Constructs according to established ontological principles, such as those found in the OBO Foundry, ones which are already being used in health informatics solutions.
RDoC Constructs are described as systems responsible for behaviors. Consider the definition for Approach Motivation, within the Positive Valence Systems Domain:
A multi-faceted Construct involving mechanisms/processes that regulate the direction and maintenance of approach behavior influenced by pre-existing tendencies, learning, memory, stimulus characteristics, and deprivation states. Approach behavior can be directed toward innate or acquired cues (i.e., unconditioned vs. learned stimuli), implicit or explicit goals; it can consist of goal-directed or Pavlovian conditioned responses.1 What does it mean for an entity, a system, a mostly neurological system, to be responsible for some behavior? It is responsible in the sense that the system has, as a result of the way in which it is configured, some disposition towards certain kinds of behavior. This behavior is realized when appropriate environmental conditions and stimuli, both external and internal to the organism, are present. Thus, we take the terms used as RDoC Constructs to refer to bodily systems that bear functions, which are realized in particular kinds of behaviors, ones which ultimately can be observed, described, and measured. These behaviors are those used to diagnose mental dysfunctioning, and which are used as a basis for the criteria in making the traditional categorical diagnoses as seen in the DSM or ICD.
The approach we have adopted is to redefine the RDoC
Domains and Constructs as subtypes of ‘bodily system’,
which are defined necessarily by the ‘function’ they bear.
Thus, S is a bodily system for organism O if and only if S is
an element of O, and S bears a critical function for O, and S
is not a part of any other system that has a critical function
for O
The advantage of this approach, which we have
implemented in OWL as the RDoC Ontology (RDoCOn)2, is in clearly
separating the system as a material entity, most likely a
complex aggregate of functionally-related and more
granular entities, from the capacity of that system to contribute in
some way to the realization of mental processes and
behavior. These systems can be composed of circuits, cells,
pathways, molecules, and so on. The system can in turn be
connected to physiology, which, from the RDoC prospective, is
considered as generic biological processes, well-established
measures of which have been validated in assessing
constructs, such as heart rate or galvanic skin response
However, RDoC currently makes no clear distinction between behavioral paradigms (instances of which would exist on the side of the patient) and assay or testing paradigms (instances of which are planned processes that produce information about the patient). They are currently lumped together in both columns of the Matrix, as either Behavior or Paradigms. A realism-based approach would make this distinction unambiguous and explicit. The advantage is that it would prevent incorrect use of the Matrix, potentially making false assertions, and thus producing false inferences from tools using automated reasoning techniques. For example, RDoC Paradigm Attention Blindness is certainly not a testing paradigm, but rather a complex behavior. While there are tests for measuring attention blindness, such as the invisible gorilla test, the use of the term ‘attention blindness’ for a Paradigm is ambiguous and could easily lead to problematic results. At an instance level representation of data, it may result in inferring that a behavior of attentional blindness, when asserted as a type of Paradigm, would somehow produce data.
patient_123 participant_in attentionBlindness_123 attentionBlindness_123 has_output dataItem_123 The correct interpretation is where some assay measures the behavior that the patient is participating in, or hypothesized to be. Thus, if this were to be encoded in some knowledge base, the proper sequence would look something like: patient_123 participant_in attentionBlindness_123 attentionBlindnessAssay_123 has_output dataItem_123 dataItem_123 is_about attentionBlindness_123
Of course, these are patient-level statements, and much of the data relevant to RDoC, especially in regard to genetics, will be about populations of patients that all share some properties in common. We believe this distinction can be adequately addressed, whether through the use of aggregates of patients and processes, or as canonical class-level assertions, restricted to some evidential context, in much in the same way as GO annotations are considered. 4
There exist many mid-level ontologies relevant to RDoC: The Gene Ontology (GO), Chemical Entities of Biological Interest (ChEBI), Protein Ontology (PRO), Cell Ontology (CL), Human Phenotype Ontology (HP), Relations Ontology (RO), Neurological Disease Ontology (ND), Neuropsychological Testing Ontology (NPT), Cognitive Paradigm Ontology (COGPO), Evidence Ontology (ECO), Mental Disease Ontology (MDO), Mental Functioning Ontology (MFO), Emotion Ontology (MFOEM), among others.3 3 For sake of space, we do not include individual citations for these ontologies, but refer to: http://www.obofoundry.org However, gaps exist. Currently no ontology represents the level of granularity for behavioral processes necessary to express all of RDoC’s current content. Many of these ontologies are not currently being developed, nor are users documented or curation issues being addressed. Part of our goal in developing RDoCOn is to revitalize, and eventually support the integration of these ontologies. To illustrate this, consider how the results of a study in the genetics related to impulsivity and psychopathology could be represented using ontologies to link the research findings to the RDoC framework, and thus provide the kind of automated discovery and interoperability the NIMH hopes for.
In a recent study (Sanchez-Roige et al., 2017) did a genome-wide association study of delayed discounting, in which the Money Choice Questionnaire was used as a measure of the behavioral paradigm. They found significant association in the intron of GPM6B (Neuronal Membrane Glycoprotein M6B), which has been previously associated with serotonin transport and impulsivity behavior. Looking up GO annotations for GPM6B, we find 85 annotations, 14 of which are for Homo sapiens. GPM6B is associated with a range of GO biological processes, such as protein transport, nervous system development, positive regulation of bone mineralization, and negative regulation of serotonin uptake. The last is of interest since previous findings have indicated that lower levels of serotonin are associated with an increase of delayed reward discounting behavior (Schweighofer et al., 2008).
Delayed discounting is the tendency to favor immediate rewards over a (potentially) more valuable distant reward. It is considered an important feature of impulse control, and can be increased, or exaggerated, in people who are diagnosed with mental disorders, such as ADHD, addiction, and depression (Sanchez-Roige et al., 2017). Delayed Discounting in RDoC is listed as an Element under Paradigm and relevant to the Reward Valuation Construct. Our definition of the function associated with Reward Valuation: ‘reward valuation function’ =def An approach motivation function that, when realized, is realized in some mental process that assesses the benefits of a prospective outcome in choosing some reward.
This aligns with RDoC’s attempt to link the paradigm as a measure of behavior to the Construct that is responsible for realizing the behavior. However, as noted above, the current version of RDoC often confuses behaviors with the testing paradigms (assays) that measure behaviors. As a behavioral “paradigm”, delayed discounting behavior is part of a broader process that realizes a reward valuation function. As an assay “paradigm”, a delayed discounting assay process is a planned process of assaying a subject’s delayed discounting behavior in a controlled circumstance using some validated instrument, such as a questionnaire, the goal of which is to produce data about the subject’s behavior.
Looking at existing biomedical knowledge resources, we see MESH defines ‘Delayed discounting’ as:
The ability to resist the temptation for an immediate reward and wait for a later reward. The tendency to devalue an outcome as a function of its temporal delay or probability of achievement. It can be evaluated in a psychological paradigm that involves the choice between receiving a smaller immediate reward or a larger delayed reward.4 It is defined as an “ability to resist”, which puts it within the specifically dependent branch of a BFO hierarchy, most likely as a realizable entity. However, it is asserted as a subclass of ‘Choice Behavior’, sibling to ‘Career Choice, and part of a subsumption hierarchy that does not align with its definition as an ability.
A much better example is found in the Cognitive
Paradigm Ontology (COGPO), where ‘Delay Discounting Task
Paradigm’ is defined as “a Behavioral Experimental
Paradigm in which, Subjects perform a type of reward task
(correct performance is associated with reward, often monetary
reward) in which they choose between earning a small
reward immediately or a larger reward at a later time”
When representing the study described above (SanchezRoige et al., 2017), a class for money choice questionnaire assay process would need to be created since none exists in any ontology. It would be asserted as a subclass of COGPO ‘delayed discounting testing paradigm’. However, COGPO’s labeling is potentially misleading and could be more explicit, i.e., ‘delayed discounting assay process’. The assay process produces data which provides some measure of the behavioral paradigm. (Figure 1).
COGPO does not define the behavioral processes that these behavioral experimental paradigms produce information about. We believe this is appropriate, as terms for representing behavior and mental processes should be maintained in separate ontologies from one that contains classes for assays. This aligns with the OBO Foundry principle of Scope to delineate content and maintain orthogonality with other ontologies.6 Currently no ontology adequately represents these more granular processes and behaviors, although some appear in the GO, or the Neuro-behavior Ontology. These intermediate processes can be used as way of linking Elements together and ultimately to the Constructs that RDoC intends for consideration as dimensional axes for understanding psychopathology. 5 http://www.cogpo.org/ontologies/CogPOver1.owl#COGPO_00049 6 http://www.obofoundry.org/principles/fp-005-delineated-content.html
The initial stage in the creation of RDoCOn is complete. We have redefined the Constructs and created OWL classes for representing each of the 41 Constructs. They are grouped into five high-level and seven mid-level classes, which mirror the current RDoC taxonomy. We have taken some liberties in our representation of the Constructs. Our goal is to find a balance between the need for ontological rigor, realism-based analysis and development, along with the implementation and eventual use of the ontology. It is important at this stage to maintain close alignment with the current state of RDoC and the Matrix. We are offering RDoCOn as an application ontology to support data integration, query writing, and knowledge discovery in general. It shall serve as a tool to demonstrate the kind of semantic integration that is attainable for research aligned with the RDoC framework by using existing ontologies, especially via the myriad of GO annotations. We consider this a beginning to clearing up the ontological confusion surrounding the RDoC framework. Deeper consideration of the validity of RDoC Constructs and to what extent a strictly realism-based approach can be faithful to the RDoC framework shall continue.
We hope development and use of RDoCOn will promote further development of ontologies related to the mental health domain, such as MFO and COGPO. We would like to see NIMH take note of our efforts and attempt to resolve these issues in the Matrix and underlying RDoC methodology, most notably the lack of clarity and consistency in how the Matrix is constructed. Revision will be needed as RDoC grows and adapts to vetting by the scientific community, potentially even radically altering its organization even as bottom-up data driven approaches are now being considered for reconfiguring the Constructs7. RDoCOn, and especially its use and integration with other ontologies, will need formal review by domain specialists as well as by the biomedical ontology community at large.
RDoC is currently under revision and the subject of a numerous articles that both support and criticize the project, its methods and content. There exists a tension amongst the top-down construction of the matrix as it stands, and using bottom-up statistical methods to look for alternate ways of developing functional constructs. Our goal here is not to addressing why or how RDoC was developed Although we have reviewed the content herein and previously, we are not recommending changes to that content, but rather promoting better attention the terminological component of RDoC. We are offering an ontological representation that will facilitate data integration and analysis using RDoC, regardless of how its developers alter the content.