Deep learning reveals a neurocomputational mechanism predicting depression risk in adolescents

Abstract

Early detection and prevention of psychiatric disorders, particularly depression, remain as major global health challenges, yet reliable tools for identifying individuals before symptom onset are lacking. Here, we combine functional neuroimaging with computational modeling to identify a mechanistic biomarker of depression risk. In a population-based adolescent cohort (IMAGEN, N = 1332), we found that weakened neural representations of emotional signals were linked to depressive symptoms. Perturbation experiments in a brain-aligned deep learning model showed that this deficit reflects overregularized emotion perception, producing a negative perceptual bias. A neurocomputational signature of this mechanism predicted depression symptom onset up to 4 years later at the IMAGEN follow-up (N = 725), was associated with both a genetic-risk variant and polygenic risk for depression, and improved depression classification in a patient cohort (STRATIFY, N = 411). These findings suggest a possible mechanism linking genetic vulnerability to altered emotion perception and future depression, and propose a predictive computational marker with potential for early detection and prevention.