Machine Learning Identifies Large-Scale Reward-Related Activity Modulated by Dopaminergic Enhancement in Major Depression

Abstract

Background: Theoretical models have emphasized systems-level abnormalities in major depressive disorder (MDD). For unbiased yet rigorous evaluations of pathophysiological mechanisms underlying MDD, it is critically important to develop data-driven approaches that harness whole-brain data to classify MDD and evaluate possible normalizing effects of targeted interventions. Here, using an experimental therapeutics approach coupled with machine learning, we investigated the effect of a pharmacological challenge aiming to enhance dopaminergic signaling on whole-brain response to reward-related stimuli in MDD.

Methods: Using a double-blind, placebo-controlled design, we analyzed functional magnetic resonance imaging data from 31 unmedicated MDD participants receiving a single dose of 50 mg amisulpride (MDDAmisulpride), 26 MDD participants receiving placebo (MDDPlacebo), and 28 healthy control subjects receiving placebo (HCPlacebo) recruited through two independent studies. An importance-guided machine learning technique for model selection was used on whole-brain functional magnetic resonance imaging data probing reward anticipation and consumption to identify features linked to MDD (MDDPlacebo vs. HCPlacebo) and dopaminergic enhancement (MDDAmisulpride vs. MDDPlacebo).

Results: Highly predictive classification models emerged that distinguished MDDPlacebo from HCPlacebo (area under the curve = 0.87) and MDDPlacebo from MDDAmisulpride (area under the curve = 0.89). Although reward-related striatal activation and connectivity were among the most predictive features, the best truncated models based on whole-brain features were significantly better relative to models trained using striatal features only.

Conclusions: Results indicate that in MDD, enhanced dopaminergic signaling restores abnormal activation and connectivity in a widespread network of regions. These findings provide new insights into the pathophysiology of MDD and pharmacological mechanism of antidepressants at the system level in addressing reward processing deficits among depressed individuals.

Trial registration: ClinicalTrials.gov NCT01253421 NCT01701258.

Keywords: Biomarker; Biotypes; Depression; Dopamine; Machine learning; fMRI.

Conflict of interest statement

Competing Interests

Y.L., M.M., and P.A. are current or previous full-time employees at Blackthorn Therapeutics Inc. Over the past 3 years, D.A.P. has received consulting fees from Akili Interactive Labs, BlackThorn Therapeutics, Boehringer Ingelheim, Posit Science, and Takeda Pharmaceuticals and an honorarium from Alkermes for activities unrelated to the current review. All other authors report no biomedical financial interests.

Copyright © 2019 Society of Biological Psychiatry. Published by Elsevier Inc. All rights reserved.

Figures

Figure 1:

a) An illustration of the importance-guided sequential model selection procedure used to find the optimal set of features. First, a full model including all features is trained using logistic regression with elastic net regularization to determine relative importance of individual features. Next, a series of truncated models were trained based on a progressively increasing set of top features rank ordered by the full model. The set of features in the best truncated model on the evaluation set were deemed as the optimal feature set. b) An illustration of the nested cross-validation procedure used to train, validate, and test the models. A grid search procedure with 3-fold cross-validation was implemented on the developmental set to determine the best model parameters. The resulting model was further tested on the evaluation set, which contained an independent set of participants not used in training and validation. The entire procedure was repeated on 100 different random partitioning of the data to allow for stable model performance.

Figure 2:

Comparing classification performance between the data-driven models based on features selected from the whole-brain and the hypothesis-driven models based only on striatal features for a) MDDPlacebo vs. HCPlacebo and b) MDDPlacebo vs. MDDAmisulpride classifications. Asterisks denote significantly different median area under the Receiver Operating Characteristic (ROC) curve measures between the data-driven and hypothesis-driven models as assessed by the Mann-Whitney U test. The black markers denote outliers falling outside the ±1.5 interquartile range. The histogram of the signed sum score from the model-identified most predictive brain regions show high separability between c) MDDPlacebo vs. HCPlacebo and d) MDDPlacebo vs. MDDAmisulpride.

Figure 3:

Weight maps showing the most predictive brain regions for the contrast of the reward minus neutral cue conditions. a) Weight map for the MDDPlacebo vs. MDDAmisulpride model. Positive weights indicate higher BOLD in the MDDPlacebo group relative to the MDDAmisulpride group and negative weights indicate the opposite direction. b) Weight map for the MDDPlacebo vs. HCPlacebo model, with positive weights indicating higher BOLD in the MDDPlacebo group relative to the HCPlacebo group and vice versa. ACC: anterior cingulate cortex; Amyg: amygdala; Cal: calcarine sulcus; Cu: cuneus; dmPFC: dorsomedial prefrontal cortex; Hipp: hippocampus; Ins: insula; lOFC: lateral orbitofrontal cortex; MCC: middle cingulate cortex; OC: occipital cortex; Oper: operculum; Pal: pallidum; PHG: parahippocampal gyrus; PPC: posterior parietal cortex; Precu: precuneus; Put: putamen; SMA: supplementary motor area; TC: temporal cortex; vmPFC: ventromedial prefrontal cortex.

Figure 4:

Weight maps showing the most predictive brain regions/connectivity for the contrast of reward minus no-change outcomes. a) Weight map for the MDDPlacebo vs. MDDAmisulpride model, with positive weights indicating higher BOLD in the MDDPlacebo group relative to the MDDAmisulpride group and vice versa. b) Negative weight assigned to the NAcc-MCC connectivity in the MDDPlacebo vs. MDDAmisulpride model. c) Weight map for the MDDPlacebo vs. HCPlacebo model. Positive weights indicate higher BOLD in the MDDPlacebo group relative to the MDDAmisulpride group and vice versa. d) Negative weights assigned to the Caudate-dACC and NAcc-MCC connectivity features by the MDDPlacebo vs. HCPlacebo model. Abbreviations followed those used in Fig. 2. Cau: caudate; dACC: dorsal anterior cingulate cortex; IFG: inferior frontal gyrus; ITC: inferior temporal cortex; PCC: posterior cingulate cortex; PreCG/PostCG: pre- and post-central gyrus; SFG: superior frontal gyrus; TP: temporal pole.

Figure 5:. Multivariate signatures across groups demonstrated amisulpride-based brain normalization.

The signed BOLD sum scores calculated across the convergent features of MDDPlacebo vs. MDDAmisulpride and MDDPlacebo vs. HCPlacebo models suggest normalization of MDD-related abnormalities following amisulpride administration. MDDPlacebo patients had overall greater multivariate neural signatures compared to HCPlacebo or MDDAmisulpride (p<0.001 for both tests). Equivalence testing demonstrated that HCPlacebo and MDDAmisulpride had statistically equivalent (denoted using “e” in plot) scores (p=0.01).