Mapping disease course across the mood disorder spectrum through a Research Domain Criteria (RDoC) framework

Abstract

Background:

The National Institute of Mental Health Research Domain Criteria (RDoC) initiative aims to establish a neurobiologically valid framework for classifying mental illness. Here, we examined whether the RDoC construct of Reward Learning and three aspects of its underlying neurocircuitry predicted symptom trajectories in individuals with mood pathology.

Methods:

Aligning with the RDoC approach, we recruited individuals [n=80 with mood disorders (58 unipolar, 22 bipolar) and n=32 controls; 63.4% female] based on their performance on a laboratory-based reward learning task, rather than clinical diagnosis. We then assessed (1) anterior cingulate cortex prediction errors using electroencephalography, (2) striatal reward prediction errors using functional magnetic resonance imaging, and (3) medial prefrontal cortex glutamatergic function (mPFC Gln/Glu) using ¹H magnetic resonance spectroscopy. Severity of anhedonia, (hypo)mania and impulsivity were measured at baseline, 3 months and 6 months.

Results:

Greater homogeneity in aspects of brain function (mPFC Gln/Glu) was observed when individuals were classified according to reward learning ability rather than diagnosis. Furthermore, mPFC Gln/Glu levels predicted more severe (hypo)manic symptoms cross-sectionally, worsening (hypo)manic symptoms longitudinally, and explained greater variance in future (hypo)manic symptoms than diagnostic information. However, rather than being transdiagnostic, this effect was specific to individuals with bipolar disorder. Prediction error indices were unrelated to symptom severity.

Conclusions:

Although findings are preliminary and require replication, they suggest that heightened mPFC Gln/Glu warrants further consideration as a predictor of future (hypo)mania. Importantly, this work highlights the value of an RDoC approach that works in tandem with, rather than independent of, traditional diagnostic frameworks.

Introduction

The Diagnostic and Statistical Manual of Mental Disorders (DSM) (1) and International Classification of Diseases (2) classify Major Depressive Disorder (MDD) and Bipolar Disorder (BD) as separate conditions distinguishable by a history of (hypo)mania, with evidence supporting a disease-specific treatment approach (3, 4). Although these nosological systems provide a useful common language for clinicians and researchers, their value for understanding mood disorder pathophysiology remains limited. Accordingly, the Research Domain Criteria (RDoC) (5, 6) was proposed as a strategic change in scientific inquiry, and seeks to classify psychiatric disorders according to measurable variability within and across different domains of functioning. Subsequently, the Positive Valence Systems domain – in particular, the subdomain of ‘Reward Learning’ – has emerged as an especially promising target for understanding the mechanisms underpinning mood symptoms.

Reward learning refers to the ability to adaptively modulate behavior as a function of positive reinforcement. Abnormalities in reward learning and underlying neurocircuitry have been strongly implicated in mood disorders (7, 8). For example, performance on behavioral reward learning paradigms has been shown to (i) differentiate patients with MDD or BD from controls during symptomatic and asymptomatic states (9, 10), (ii) predict anhedonia severity and treatment outcome (11), (iii) change following pharmacological dopaminergic manipulations (12, 13), (iv) be linked to striatal dopamine transporter function and frontostriatal functional connectivity (14), and (v) be heritable (15). Decades of research in laboratory animals has identified the neurobiological processes underpinning reward learning (16). Therefore, examining how these processes vary across the mood disorder spectrum represents a fruitful avenue for identifying the neurobiological basis underpinning mood disorder heterogeneity.

Imaging and computational studies suggest that the brain employs distinct hierarchical systems to support learning (17, 18) and to date, the neural circuitry involved in learning from positive reinforcement has been especially well characterized (19–21). Importantly, individuals with MDD and BD have been found to exhibit dysregulation in three key aspects of this neurocircuitry. First, a fundamental mechanism that supports reward learning is the reward prediction error (RPE), which is a striatal dopamine-based signal that encodes violations of reward expectancies (22). Individuals with MDD have been found to have blunted striatal RPE signals during learning (23–25) and this blunting has been linked to a more recurrent depressive illness course (23). Similar abnormalities have been observed in individuals with BD, although the direction of effects often diverges from those observed in studies of unipolar MDD. Relative to healthy controls, euthymic individuals with BD or individuals with subthreshold hypomania have been found to have elevated striatal activation during reward anticipation (26) and reward outcome (27). Similarly, manic individuals with BD show striatal responses that fail to differentiate between receipt and omission of rewards, suggestive of abnormal RPE signaling (28).

Second, event-related potential (ERP) studies highlight the reward positivity (RewP) as another important reward circuit component linked to mood pathology (29). The RewP is a frontocentral electroencephalographic (EEG) deflection that is elicited by RPEs, and is thought to originate from the anterior cingulate cortex (ACC) and striatum (30). Smaller RewP amplitudes, as well as weaker RewP-related ACC activation, predict poorer reward learning (31, 32). Furthermore, abnormal RewP amplitudes have been observed in individuals with hypomania (33) and those with MDD (34), and have been found to predict future depression onset in healthy individuals (35). Critically, the source of these RewP signals is believed to be distinct from that of striatal dopaminergic RPEs (36), hence, they offer complementary information to fMRI-based RPE studies in terms of understanding the biological basis of reward learning dysfunction.

Finally, while the reward learning literature has historically emphasized the role of dopamine, the hedonic effects of dopamine are thought to be partially mediated by its interactions with glutamatergic signals originating in the medial prefrontal cortex (mPFC) (37). In line with this notion, in animal studies, disrupted glutamate signaling between mPFC and striatal regions impairs reward motivation (38), and in psychiatrically healthy humans, mPFC glutamate levels (measured using magnetic resonance spectroscopy; MRS) predict reward-based decision-making (39). Human MRS studies often focus on the glutamine/glutamate ratio (Gln/Glu) because glutamate is released into the synaptic cleft, taken up by glial cells, converted into glutamine, and cycled back into neurons (40), making mPFC Gln/Glu a proxy measure of the integrity of the glutamatergic synapse. Of note, meta-analyses of MRS studies have highlighted mPFC glutamate abnormalities in MDD and BD, albeit in opposite directions, with glutamatergic transmission being reduced in MDD (41), but elevated in BD (42) across manic (43), depressive (44) and euthymic (45) mood states.

Taken together, these studies suggest that striatal and ACC-mediated PE signals, along with mPFC Gln/Glu, are promising biomarkers of reward learning that may be implicated in mood pathology. Therefore, the aim of this study was to determine whether variation in reward learning neurocircuit function predicts variability in symptom trajectories in individuals with mood disorders. In line with the grant mechanism supporting this study (RFA-MH-14-050, Dimensional Approaches to Research Classification in Psychiatric Disorders), we recruited individuals based on their performance on a well-validated behavioral reward learning task, rather on the basis of specific DSM diagnoses. We then examined whether neurobiological indices of reward learning predicted cross-sectional and longitudinal variation in three reward-relevant symptom domains, namely anhedonia, (hypo)mania and impulsivity. We predicted that potentiated striatal and ACC-mediated PEs, and elevated mPFC Gln/Glu, would predict worsening (hypo)mania and impulsivity. In contrast, we predicted that blunted striatal and ACC-mediated PEs, and reduced mPFC Gln/Glu, would predict worsening anhedonia. Importantly, we assessed whether these reward learning biomarkers provided superior predictive validity in determining symptom trajectories, relative to clinical diagnostic information alone.

Materials and Methods

Participants

The mood pathology group were required to have depressive, mixed or hypomanic symptoms severe enough to cause distress/impairment. Participants could pursue treatment but were excluded from further testing if they initiated one of the exclusionary treatments (Supplemental Methods). Psychotropic medication load was quantified using previously established procedures (see Supplemental Methods). The control group had no lifetime psychiatric disorders or psychotropic medication use. This study was approved by the Partners Human Research Committee. Participants provided written, informed consent prior to participating.

Study design and recruitment

Fig. 1A shows the study design. Recruitment occurred as follows: healthy controls and treatment-seeking individuals with mood disorders were screened on a Probabilistic Reward Task (PRT) (10, 46). Screening continued until two conditions were met: (i) until a sample of 32 healthy controls with valid PRT data, and who met study eligibility criteria, was recruited, and (ii) until a sample of 80 individuals with mood pathology whose PRT performance spanned the full range of a normative distribution, and who met study eligibility criteria, was recruited. For the 80 individuals with mood pathology, we focused on equally populating quintiles of reward learning (i.e., n=~16 per quintile; Fig. 2) that were defined using cut-offs derived from a prior normative sample of 572 controls who had performed the PRT in prior studies. In total, n=272 individuals had to be screened on the PRT to reach these two criteria (see Fig. S1 for study flow diagram).

Figure 1.

Study methods overview

Panel A shows a summary of the study flow. Participants were screened on a Probabilistic Reward Task (PRT) and the patient group was recruited so that their scores on the PRT spanned the entire range of possible scores on a pre-existing normative distribution. If eligible, a clinical assessment was conducted, then participants returned for two baseline neuroimaging visits (EEG and fMRI/MRS sessions), as well as 3- and 6-month follow-up assessments. Panel B displays source localization analyses demonstrating that scalp-recorded RewP amplitude correlated with current source density in the dorsal ACC (p<0.005 uncorrected; x=−3), validating RewP amplitude as a marker of ACC-mediated activation. Panel C shows the bilateral NAc region-of-interest (y=10) from which striatal RPEs were extracted. Panel D shows the 2 × 2 × 2cm voxel placed in the mPFC (x=0), from which Gln/Glu metabolites were extracted.

Figure 2.

Recruitment based on behavioral reward learning

Panel A shows the number of participants with mood pathology whose PRT performance fell in each quintile of reward learning performance according to the normative distribution (the normative distribution was based on a separate existing sample of N=572 healthy controls). The dotted line indicates the a priori target of n=16 per quintile that was set to ensure that we recruited individuals who spanned the entire range of reward learning performance. This target was met in all but the lowest quintile, however, this quintile was still adequately represented with a sample of n=13. Panel B shows frequency histograms of reward learning performance across the control, unipolar and bipolar groups.

For participants who were screened on the PRT and had valid data, study eligibility criteria and clinical diagnoses were further evaluated via a Structured Clinical Interview for DSM-IV (47) (conducted by Masters- or PhD-level interviewers). Participants were also screened with the Young Mania Rating Scale (YMRS) (48) to ensure that at least one third of the mood pathology sample exhibited (hypo)manic symptoms. Eligible participants completed five study visits: (i) behavioral testing and clinical assessment; (ii) a baseline EEG/ERP recording; (iii) a baseline MRI scan; (iv) a 3-month follow-up clinical assessment; and (v) a 6-month follow-up clinical assessment. Participants received $15/hr in compensation plus earnings on the behavioral and imaging tasks.

Primary outcomes

Anhedonia was measured using the Anhedonic Depression subscale of the 62-item Mood and Anxiety Symptom Questionnaire (MASQ-AD) (49) and impulsivity was assessed using the Barratt Impulsiveness Scale (BIS) (50). (Hypo)mania was measured using the Bipolar Inventory of Symptoms Scale Mania subscale (BISS-mania), which was chosen over the YMRS as it measures an extended range of (hypo)manic symptoms (51) and showed greater variance across both unipolar and bipolar groups. These measures were completed at baseline, and again at 3- and 6-month follow-up assessments. All three scales demonstrated good internal consistency (Supplementary Methods).

Probabilistic Reward Task (PRT) – quantifying reward learning

Reward learning was assessed using a well-validated computer-based PRT (46). On each trial, a fixation cross (500ms) was followed by a schematic, mouthless face (500ms). Next, a short (11.5mm) or a long (13mm) mouth appeared (100ms). Participants indicated whether the mouth was long or short. There were 3 blocks of 100 trials, and for each block, 40 correct trials were rewarded (“Correct!! You won 20 cents”). Although long and short mouths were presented at equal frequency, unbeknownst to participants, correct identification of one mouth (the ‘rich stimulus’) was rewarded three times more than the other (the ‘lean stimulus’).

Following quality control (Supplemental Methods), we used signal detection analysis (52) to compute response bias (the tendency to bias responding to the rich stimulus), using the formula:

$$\text{log}b=\frac{1}{2}\text{log}\left(\frac{\mathit{\text{ Ric}}{\text{h }}_{\mathit{\text{correct}}}*\mathit{\text{Lean}}{}_{\mathit{\text{incorrect}}}}{\mathit{\text{ Ric}}{\text{h }}_{\mathit{\text{incorrect}}}*\mathit{\text{Lean}}{}_{\mathit{\text{correct}}}}\right)$$

To allow calculation of response bias for cases that included a zero in the formula, 0.5 was added to each cell of the matrix (53). Reward learning was defined as the increase in response bias from block 1 to block 3.

Scalp-recorded RewP amplitude – quantifying ACC PEs

The RewP was computed from 128-channel scalp-recorded EEG, acquired while participants performed a counterbalanced version of the PRT. After preprocessing, temporo-spatial principal components analysis was used to decompose the time-domain ERP (54). Temporal variance in the averaged ERP waveforms was examined using temporal PCA and Infomax rotation. Based on the scree plot used to determine the factors to retain in a PCA analysis, 12 temporal factors were retained for rotation. The spatial distribution of these temporal factors was then examined using spatial PCA and Infomax rotation, with a spatial PCA being conducted for each temporal factor. Eight spatial factors were retained for each temporal factor. Analyses focused on the PCA component with a timing and topography most consistent with the RewP (TF8/SF2; Supplemental Methods). Furthermore, source localization (55) confirmed that the RewP had a source in the dorsal ACC (Fig. 1B). Our primary variable of interest was the difference in RewP amplitude following feedback on lean versus rich trials (ΔRewP), which captures the degree to which the ACC tracks reward probability across different contexts.

fMRI-based learning task – quantifying striatal RPEs

Striatal RPE signals were assessed using a well-validated explicit reinforcement learning paradigm (19, 56) that required participants to learn reward contingencies through trial-and-error. On each trial, participants were asked to choose between two symbols and each symbol in the pair was associated with an 80%/20% probability of a given outcome (gain: $1/$0; loss: $0/−$1; neutral: grey square/’nothing’). We used Q-learning to calculate the RPE (19) from participants’ behavioral data, and imaging analyses focused on a parametric modulation contrast for RPE signals (Supplementary Methods).

Anatomically-defined regions-of-interest (ROIs) in the left and right NAc were selected from prior research showing links between dopamine transporter function and reward learning (14) (Fig. 1C). Beta weights from RPE contrasts were extracted from these ROIs. A one-sample t-test confirmed that the RPE in both ROIs was >0 [left: t(106)=3.07, p=0.003; right: t(106)=4.12, p<0.001], so beta values were averaged to create a single NAc RPE beta weight that was used for subsequent analyses. A positive RPE beta value signified higher activation for unexpected reward and lower activation for unexpected omission of rewards during gain trials.

Magnetic resonance spectroscopy (MRS) – quantifying mPFC glutamate

¹H-MRS was used to assess mPFC Gln/Glu. A 2 × 2 × 2cm voxel was placed in the mPFC, midsagittally, anterior to the genu of the corpus collosum (Fig. 1D). The voxel was automatically shimmed, with further manual shimming performed as needed. A modified J-resolved protocol (2D-JPRESS) (57) was used to resolve glutamatergic metabolites. This sequence involved the collection of 22 TE-stepped spectra with an echo time ranging from 35ms to 250ms in 15ms increments (TR=2s, f1 acquisition bandwidth=67Hz, spectral bandwidth=2kHz, readout duration=512ms, NEX=16/TE-step, approximate scan duration=12min).

To quantify glutamate and glutamine with the JPRESS data, the 22 TE-stepped free-induction decay series (FIDS) was zero-filled out to 64 points, Gaussian-filtered, and Fourier-Transformed using GAMMA-simulated J-resolved basis sets modeled for 2.89 T. Every J-resolved spectral extraction within a bandwidth of 67Hz was fit with the spectral-fitting package LCModel and its theoretically-correct template. The integrated area under the entire 2D surface for each metabolite was calculated by summing the raw peak areas across all 64 J-resolved extractions (Supplemental Methods). Our primary outcome was the glutamine to glutamate ratio (Gln/Glu).

Statistical analysis

Multivariable regression analyses examined whether ΔRewP, NAc RPE or mPFC Gln/Glu predicted anhedonia, (hypo)mania or impulsivity in the clinical sample cross-sectionally and longitudinally. Separate regression models were run for each outcome (MASQ-AD, BISS-mania, BIS). Models included covariates (age, sex, medication load), mood polarity/diagnosis (Group: dummy-coded 0=unipolar, 1=bipolar), the three neural predictors (ΔRewP, NAc RPE, mPFC Gln/Glu), and a Group × Predictor interaction term for each neural predictor. Models predicting follow-up symptom severity also controlled for baseline symptom severity.

Results

Sample characteristics

The sample was 63.4% female (n=71), with a mean age of 28.6 (SD=9.1, range 18–60). Of the patient group, 72.5% (n=58) had unipolar mood pathology (MDD/dysthymia, or MDD in partial remission), 27.5% (n=22) had bipolar mood pathology (BD-I/II, depressed, mixed or hypomanic), and 40% (n=32) took medication (see Table 1 and Supplemental Methods for details). Sample sizes for each of the analyses varied when a participant had missing data on one or more of the neural indices and/or follow-up measures. Accordingly, sample sizes ranged from 25–32 for the control group, 38–58 for the unipolar group, and 12–22 for the bipolar group (sample sizes for each analysis are specified below).

Table 1.

Demographic and clinical characteristics of sample

	HC (n=32)	Unipolar (n=58)	Bipolar (n=22)	Test	P value
Age, M (SD)	28.4 (7.7)	28.0 (8.6)	30.5 (12.1)	F=0.59	0.56
Female, N (%)	17 (53.1)	41 (71.7)	13 (59.1)	χ2=2.96	0.23
Years education, M (SD)	17.0 (3.2)	16.0 (2.8)	15.6 (3.1)	F=1.77	0.18
White, N (%)	21 (65.6)	40 (69.0)	19 (86.4)	χ2=10.02	0.26
Hispanic, N (%)	2 (6.3)	6 (10.3)	2 (9.1)	χ2=0.43	0.81
Clinical diagnoses
Current MDD, N (%)	- -	49 (84.5)	- -	-	-
Current dysthymia, N (%)	- -	1 (1.7)	- -	-	-
MDD in partial remission, N (%)	- -	8 (13.8)	- -	-	-
BD-I depressed, N (%)	- -	- -	7 (31.8)	-	-
BD-I mixed, N (%)	- -	- -	0 (0.0)	-	-
BD-I hypomanic, N (%)	- -	- -	2 (9.1)	-	-
BD-II depressed, N (%)	- -	- -	9 (40.9)	-	-
BD-II mixed, N (%)	- -	- -	1 (4.6)	-	-
BD-II hypomanic, N (%)	- -	- -	3 (13.6)	-	-
Comorbidities
Alcohol abuse, N (%)	- -	0 (0.0)	2 (9.1)	χ2=5.41	0.02
EDNOS or BED, N (%)	- -	2 (3.4)	2 (9.1)	χ2=1.07	0.30
GAD, N (%)	- -	3 (5.2)	2 (9.1)	χ2=0.42	0.52
Panic disorder, N (%)	- -	1 (1.7)	0 (0.0)	χ2=0.38	0.54
PTSD, N (%)	- -	3 (5.2)	2 (9.1)	χ2=0.42	0.52
Social Phobia, N (%)	- -	8 (13.8)	3 (13.6)	χ2=0.00	0.99
Specific Phobia, N (%)	- -	3 (5.2)	2 (9.1)	χ2=0.42	0.52
Medication
Antidepressants, N (%)	- -	19 (32.8)	4 (18.2)	χ2=1.65	0.20
Mood stabilizer or anticonvulsant, N (%)	- -	1 (1.7)	7 (31.8)	χ2=16.05	<0.001
Anticonvulsants, N (%)	- -	0 (0.0)	1 (4.5)	χ2=2.67	0.10

Note. HC=healthy control; MDD=major depressive disorder; BD=bipolar disorder; EDNOS=eating disorder not otherwise specified; BED=binge eating disorder; GAD=generalized anxiety disorder; PTSD=post-traumatic stress disorder. All tests are two-tailed.

Correlations among units

Pearson correlations were used determine the degree to which the three neural indices mapped onto behavioral reward learning (see Tables S1 & S2; differences in units of analysis between diagnostic groups are reported in the Supplementary Results and Figure S2). Across the sample, higher mPFC Gln/Glu correlated with better reward learning (r=0.27, p=0.007, n=102; Fig. S3A). This was consistent with the linear trend shown in Fig. S3A, where mPFC Gln/Glu values increased across the learning quintiles. Furthermore, the quintiles explained a greater proportion of the variance in mPFC Gln/Glu relative to diagnosis (5% vs. 2%; R² change=0.05, F change=5.25, p=0.02).

Although the ΔRewP and NAc RPE were not correlated with our a priori-defined learning measure (block 3 minus block 1 response bias), they were correlated with the total overall response bias. Specifically, heightened NAc RPE (r=0.37, p=0.04, n=32; Fig. S3B) and ΔRewP (r=0.41, p=0.04, n=25; Fig. S3C) correlated with greater overall response bias in controls but not in patients (p>0.10, n=75). Furthermore, across the whole sample, heightened NAc RPE was associated with faster learning in block 1 (r=0.23, p=0.02, n=107).

Elevated mPFC Gln/Glu correlates with more severe (hypo)manic symptoms cross-sectionally

Standardized values for each outcome measure across the reward learning quintiles are shown in Figure S4 (patients only). Multimodal regression models assessed whether the three reward circuit markers were associated with symptom severity cross-sectionally.

A significant Group × mPFC Gln/Glu interaction (β=0.28, p=0.04, n=57) emerged from the model predicting baseline (hypo)mania severity (BISS-mania), indicating that the effect of mPFC Gln/Glu on baseline (hypo)mania severity differed across the unipolar and bipolar groups (Table 2). To unpack this interaction, we examined the correlation between mPFC Gln/Glu and baseline BISS-mania scores (both residualized for other variables in the model) in each group. mPFC Gln/Glu was associated with higher BISS-mania scores in the bipolar (r=0.56, p=0.045, n=13) but not the unipolar group (r=−0.24, p=0.12, n=45).

Table 2.

Models predicting (hypo)manic symptom severity on the BISS-mania scale

Dependent variable: Baseline (hypo)manic symptom severity
	B	SE	β	t	p
(constant)	9.14	2.33		3.92	<0.001
Age	−0.09	0.08	−0.13	−1.18	0.24
Sex	−2.77	1.52	−0.20	−1.82	0.08
Medication load	−0.71	0.42	−0.19	−1.71	0.09
Group	10.46	1.60	0.69	6.52	<0.001
ΔRewP	−0.70	1.24	−0.07	−0.56	0.58
NAc RPE	0.03	0.57	0.01	0.05	0.96
mPFC Gln/Glu	−9.33	16.39	−0.07	−0.57	0.57
Group × ΔRewP	−1.48	2.27	−0.09	−0.65	0.52
Group × NAc RPE	2.11	1.90	0.13	1.11	0.27
Group × mPFC Gln/Glu	68.06	31.51	0.28	2.16	0.04
Dependent variable: 3-month (hypo)manic symptom severity
	B	SE	β	t	p
(constant)	3.49	1.93		1.81	0.08
Age	0.06	0.06	0.15	0.99	0.33
Sex	−2.14	1.19	−0.27	−1.80	0.08
Medication load	−0.29	0.32	−0.12	−0.90	0.37
Baseline BISS-mania	0.16	0.11	0.28	1.54	0.13
Group	−1.15	1.61	−0.13	−0.71	0.48
ΔRewP	−1.02	0.99	−0.16	−1.03	0.31
NAc RPE	−0.04	0.43	−0.01	−0.10	0.92
mPFC Gln/Glu	−38.43	14.30	−0.46	−2.69	0.01
Group × ΔRewP	1.03	1.93	0.08	0.53	0.60
Group × NAc RPE	−0.07	1.35	−0.01	−0.05	0.96
Group × mPFC Gln/Glu	96.92	24.90	0.70	3.89	<0.001

Note. Group was dummy-coded 0=unipolar, 1=bipolar; BISS-Mania=Bipolar Inventory of Symptoms Scale mania subscale score; RewP=reward positivity; NAc RPE=nucleus accumbens reward prediction error; mPFC Gln/Glu=medial prefrontal cortex ratio of glutamine to glutamate.

In contrast, none of the neural indices predicted anhedonia severity (MASQ-AD), or impulsivity (BIS; all ps>0.05) cross-sectionally.

Elevated mPFC Gln/Glu correlates with more severe (hypo)manic symptoms longitudinally

Next, we examined whether, after controlling for baseline (hypo)manic severity, the reward circuit markers were associated with 3- and 6-month follow-up symptom severity (Figure S5 shows mean symptom severity across time). A Group × mPFC Gln/Glu interaction (β=0.70, p<0.001, n=49) emerged for the model predicting 3-month BISS-mania scores (Table 2). To unpack this interaction, we again examined the correlation between mPFC Gln/Glu and 3-month BISS-mania scores (residualized for other variables in the model) in each group. Increased mPFC Gln/Glu was associated with less severe hypomanic symptoms in the unipolar group (r=−0.35, p=0.03, n=38) but more severe hypomanic symptoms in the bipolar group (r=0.85, p<0.001, n=12; Fig. 3) at 3 months.

Figure 3.

Group × mPFC Gln/Glu interaction for longitudinal (hypo)manic symptom severity

Residualized scatter plots showing the relationship between mPFC Gln/Glu and (hypo)manic symptom severity (BISS-mania subscale scores) at baseline (A, B) and at the 3-month follow-up assessment (C, D) in the unipolar and bipolar mood disorder groups. Residualized values on each axis control for the other variables in the model, which were: age, sex, baseline BISS-mania subscale scores, ΔRewP amplitude, NAc RPE beta weights.

In contrast, the reward learning markers did not predict 6-month follow-up BISS-mania scores, or 3- or 6-month MASQ-AD or BIS scores (all ps>0.05; see Supplemental Results for exploratory unimodal analyses).

Predictive value of mPFC Gln/Glu

Next, we compared a simple model containing covariates (age, sex, medication load, baseline BISS-mania) and diagnostic information (Group), to a model containing terms for mPFC Gln/Glu and Group × mPFC Gln/Glu. The simple model explained 15.8% of the variance in 3-month (hypo)manic symptom severity, F(5,44)=1.65, p=0.17. However, adding the mPFC Gln/Glu terms explained an additional 24.3% of the variance in 3-month hypomanic symptom severity, F(7,42)=4.01, p=0.002, and this change in R² was significant (F change=8.49, R² change=0.24, p=0.001). This indicates that mPFC Gln/Glu explained greater variance in future hypomanic symptom severity relative to baseline diagnosis alone. Furthermore, we confirmed that mPFC Gln/Glu explained a greater proportion of the variance in 3-month (hypo)manic symptom severity relative to behavioral reward learning alone (F change=3.91, R² change=0.09, p=0.03; Table S3), indicating that adding this biomarker enhanced predictive power over and above behavioral data.

Discussion

Using a novel recruitment method, a transdiagnostic sample, and a multimodal longitudinal design, we examined whether variation along the RDoC Positive Valence Systems domain of Reward Learning and the underlying neurocircuitry predicted variability in three reward-related mood symptoms – anhedonia, (hypo)mania and impulsivity. In doing so, we focused on three components of reward learning neurocircuitry linked to mood disorder pathophysiology that span distinct units of analysis across physiology (ACC-mediated PEs), circuits (striatal RPEs), and molecules (mPFC Gln/Glu).

As predicted, the three neural components correlated with aspects of behavioral reward learning on the PRT. In terms of symptoms, elevated mPFC Gln/Glu predicted more severe cross-sectional and longitudinal (hypo)manic symptoms in those with bipolar pathology. Importantly, baseline mPFC Gln/Glu levels explained a greater proportion of the variance in (hypo)manic symptoms at 3 months relative to diagnosis alone. These findings extend prior case-control MRS studies (41, 42) by showing that elevated mPFC Gln/Glu is also associated with (hypo)mania severity dimensionally.

We replicated prior findings linking blunted ΔRewP amplitude with greater anhedonia in exploratory unimodal analyses (see Supplement), however, neither ΔRewP nor NAc RPE signals were associated with symptom severity when entered into a multimodal model with mPFC Gln/Glu. Although the lack of a relationship between NAc RPE and anhedonia in our unimodal analyses contrasts with recent findings showing that striatal RPEs predicted improvement in anhedonic symptoms (58), we used a more complex instrumental fMRI learning paradigm designed to assess striatal RPEs in the context of learning, as opposed to a more traditional guessing-type paradigm (which maximizes the RPE signal yet involves minimal learning).

It is important to consider what these findings mean for an RDoC approach to mood disorder classification that remains agnostic to DSM diagnoses. On the one hand, mPFC Gln/Glu correlated with reward learning across diagnoses, providing converging evidence that mPFC Gln/Glu is a transdiagnostic marker of this RDoC domain. Additionally, in a heterogeneous sample, more homogeneity in neurobiology (mPFC Gln/Glu levels) was observed within groups when groups were defined on the basis of reward learning than by diagnostic categories. This is consistent with RDoC’s assumption that dimensions of functioning are more proximal to neurobiology than diagnostic categories. Furthermore, dimensional increases in this Reward Learning biomarker (i.e., mPFC Gln/Glu levels) predicted dimensional increases in symptoms characterized by excessive reward responsiveness (i.e., hypomania), rather than membership to a specific diagnostic category. This echoes one of RDoC’s central theses that abnormalities in circuits and associated constructs likely underpin specific features of mental illness, rather than explaining any single disorder in its entirety. Together, these findings partly align with a diagnosis-agnostic approach to mood disorder classification.

However, our results also highlight the considerable value of diagnostic information in predicting symptom trajectories. Specifically, although mPFC Gln/Glu correlated with reward learning transdiagnostically, the link between mPFC Gln/Glu and (hypo)manic symptom severity was disease-specific, and diagnostic information remained an integral component of the final predictive model. If we assume that these findings could inform novel interventions based on neurobiological underpinnings (a key driver of the RDoC approach), then targeting mPFC Gln/Glu may affect reward learning in a similar manner across disorders, but have different effects on symptoms in distinct mood disorder subtypes. The degree to which the RDoC framework predicts purely dimensional variability across disorders versus a blend of transdiagnostic and disorder-specific effects, remains an important topic of debate. Our findings indicate that while information about reward circuit function could improve the prediction of risk for reward-related clinical symptoms (an important finding in its own right), it would do so in tandem with, rather than independent of, existing diagnostic frameworks.

This study has several strengths. By examining neurobiological mechanisms of reward learning across multiple units of analysis, we could probe reward learning circuitry with superior spatiotemporal resolution and at both micro and macro scales, which cannot be achieved with a single unit or modality alone. Furthermore, we tested whether these units of analysis enhanced the ability to predict clinical course, over and above information already used in routine clinical care (diagnosis, baseline symptom severity). Since mPFC Gln/Glu levels can be obtained using MRS in as little as 6 minutes with good test-retest reliability (ICC=0.803) (59), mPFC Gln/Glu warrants further investigation as a potential screening method for individuals at suspected risk for bipolar disorder.

However, some limitations must also be noted. First, mPFC Gln/Glu predicted worse (hypo)manic symptoms specifically in individuals with bipolar mood pathology. As the instance of (hypo)manic symptoms was lower in the unipolar group at follow-up, this may have restricted the variance in symptoms that could be explained by mPFC Gln/Glu. Second, although our three neural indices were selected based on their established association with reward learning, only mPFC Gln/Glu was associated with our a priori reward learning measure, and the three neural indices were not correlated with one another. Although stronger associations may have been evident in a larger sample, the lack of association could also reflect an issue with the construct validity of these units of analysis. For example, it is possible that similar impairments in reward learning may have distinct etiologies (often referred to as ‘equifinality’), particularly when considering individuals with very divergent forms of psychopathology. How equifinality is accounted for remains an important conceptual issue for the RDoC framework. Finally, reductions in sample size for longitudinal analyses (resulting from participant attrition and the need to obtain good quality data across all three neural indices) mean that reduced statistical power is a limitation of our study, and may explain several null findings. The replicability of these results must be interpreted in light of concerns around the generalizability and reproducibility of neuroimaging findings obtained using small samples (60). Accordingly, rather than being definitive, we interpret these findings as novel yet preliminary insights that warrant replication in larger samples.

In sum, we showed that a key component of reward learning neurocircuitry – mPFC Gln/Glu – predicted worse (hypo)manic symptoms. This marker enhanced the ability to predict future (hypo)mania risk over and above diagnostic information alone. Using this marker to improve precision in the diagnosis and treatment of mood pathology therefore represents an important avenue for future research, with a focus on larger, well-powered samples.