Accurate and Interpretable Prediction of Antidepressant Treatment Response from Receptor-informed Neuroimaging

Abstract

Conventional antidepressants show moderate efficacy in treating major depressive disorder. Psychedelic-assisted therapy holds promise, yet individual responses vary, underscoring the need for predictive tools to guide treatment selection. Here, we present graphTRIP (graph-based Treatment Response Interpretability and Prediction) - a geometric deep learning architecture that enables three advances: 1) accurate prediction of post-treatment depression severity using only pretreatment clinical and neuroimaging data; 2) identification of robust biomarkers; and 3) causal analysis of treatment effects and underlying mechanisms. Trained on data from a clinical trial comparing psilocybin and escitalopram (NCT03429075), graphTRIP achieves strong predictive accuracy (r = 0.72, p = 6.8 x10-8), and shows clear generalization to both an independent dataset and across brain atlases. The model identifies stronger functional connectivity within sensory networks as a robust predictor of poorer response across both treatments. In contrast, causal analysis implicates frontoparietal and default mode networks as key moderators of differential response, with stronger 5-HT1A- and 5-HT2A-related signalling in the frontoparietal network predicting escitalopram response but psilocybin resistance. Overall, this work advances precision medicine and biomarker discovery in depression.