Polygenic risk-informed white matter integrity improves deep learning-based prediction of youth depression

Early detection of youth depression is crucial, given its rising prevalence and long-term consequences. Although genetic factors contribute significantly to youth depression, their integration with neuroimaging remains limited. This study aims to bridge the gap by predicting youth depression through a combined genetic and neuroimaging framework. We present a deep learning framework using polygenic risk scores to pretrain a 3D convolutional neural network on track-weighted fractional anisotropy. This approach captures gene and brain associations from a multi-ethnic cohort of 4741 youth in the Adolescent Brain Cognitive Development Study (age range: 9-10 years old). We fine-tune the model on separate held-out datasets for cross-sectional and 2-year follow-up prediction, respectively. Here we show that the model improves cross-sectional (266 participants) and two-year predictions of depression and suicidality, with area under the curve values of 0.61 to 0.66. It outperforms unimodal models, increasing accuracy over genetics-only and brain-only models. Explainable artificial intelligence identifies key white matter tracts, including the superior longitudinal fasciculus, cingulum, and corpus callosum, as primary predictive features. The model effectively generalizes to an independent Korean youth sample (age range: 9-17 years old), achieving an area under the curve of 0.67. This establishes the cross-ethnic scalability of integrating genetics with brain imaging. These findings highlight the promise of multimodal deep learning for precision psychiatry and early clinical intervention. Early detection of youth depression is crucial given its rising prevalence. We aimed to predict this risk by combining genetic and brain data. We developed an artificial intelligence model, first pretraining it to learn associations between 9-to-10-year-olds’ brain scans and depression-related genetic risk. Then, we taught the model to predict if each youth has or will develop depression within two years. Pretraining with genetic information significantly improved the model’s ability to predict depression and suicidal behavior in two years. Our approach outperformed standard comparison models and performed well for an independent group of Korean youth. Using genetics to guide artificial intelligence in reading brain scans can detect at-risk youth earlier, enabling personalized support. ED SUM: Min et al. develop a deep learning framework pretrained on genetic and white matter associations, then fine-tuned to predict youth depression. This pretraining strategy predicts risk up to two years in advance, pinpoints key white matter tracts, and generalizes to an independent cohort.