Exploring the prediction of emotional valence and pharmacologic effect across fMRI studies of antidepressants

Abstract

Background: Clinically approved antidepressants modulate the brain's emotional valence circuits, suggesting that the response of these circuits could serve as a biomarker for screening candidate antidepressant drugs. However, it is necessary that these modulations can be reliably detected. Here, we apply a cross-validated predictive model to classify emotional valence and pharmacologic effect across eleven task-based fMRI datasets (n = 306) exploring the effect of antidepressant administration on emotional face processing.

Methods: We created subject-level contrast of parameter estimates of the emotional faces task and used the Shen whole-brain parcellation scheme to define 268 subject-level features that trained a cross-validated gradient-boosting machine protocol to classify emotional valence (fearful vs happy face visual conditions) and pharmacologic effect (drug vs placebo administration) within and across studies.

Results: We found patterns of brain activity that classify emotional valence with a statistically significant level of accuracy (70% across-all-subjects; range from 50 to 87% across-study). Our classifier failed to consistently discriminate drug from placebo. Subject population (healthy or unhealthy), treatment group (drug or placebo), and drug administration protocol (dose and duration) affected this accuracy with similar populations better predicting one another.

Conclusions: We found limited evidence that antidepressants modulated brain response in a consistent manner, however found a consistent signature for emotional valence. Variable functional patterns across studies suggest that predictive modeling can inform biomarker development in mental health and in pharmacotherapy development. Our results suggest that case-controlled designs and more standardized protocols are required for functional imaging to provide robust biomarkers for drug development.

Keywords: Antidepressant; Drug development; Emotional valence; Machine learning; Predictive analysis; Task-based fMRI.

Fig. 1

Protocol summary. Primary-data analysis (A) was performed at the subject level to model task effects. Study and group-level analyses took place in MNI152 space and served as a QA step (B, see Methods). Feature reduction (C) took place in native subject space to maximize registration accuracy. The contrast of parameter estimates (COPE, see Methods) were used as features in the machine learning protocol (D).

Fig. 2

Accuracies for the emotional valence (left) and pharmacologic effect (right) classification. Studies are organized on a clinical spectrum, from healthy (H), to low neurotic (LN), to high neurotic (HN), to dysphoric (DYS), to major depressive disorder (MDD). Green lines indicate significance at respective level: (i) within study classification: no correction for multiple comparisons; (ii) across-study: p < (0.05/10) Bonferroni correction for multiple comparisons; (iii) across all-subjects: no correction for multiple comparisons; (iv) across-all-studies p < (0.05/10) Bonferroni correction for multiple comparisons. Accuracies based on a bimodal distribution test, numerical p-values are shown in the Supplementary Materials. Yellow lines are illustrate groups with higher shared accuracies. Shown below are results for the within-study classification (diagonal) and across-study classification (off-diagonal). Results for the across-all-subjects classification may be referenced in Table 2 (final row) and Supplementary Fig. 2. Results for the across-all-studies classification may be referenced in Supplementary Fig. 2. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)