A new neurocomputational model, supported by brain imaging, suggests that overreliance on prior expectations when interpreting facial expressions can signal a heightened risk of developing depression. By analysing how the brain processes angry faces, researchers uncovered a pattern of visual distortion that may foretell symptoms up to four years before they emerge. The study was published in Science Advances and led by Han Lu of Shenzhen University and Qiang Luo of Fudan University, using data from the IMAGEN cohort in Europe.
The team combined functional magnetic resonance imaging (fMRI) with a deep learning framework to explore how sensory input and stored emotional knowledge are integrated in the brain. They built a two-pathway network that mimics the human brain: one pathway processes raw visual information, while a second holds abstract knowledge about different emotions.
Analyzing fMRI data from 1,332 nineteen-year-olds in the IMAGEN population study, participants watched short clips of faces shifting from neutral to angry expressions. Using a machine learning approach, researchers extracted a mathematical representation of how each participant’s visual cortex encoded angry faces, creating what they termed a visual code. They then divided participants into those with typical, highly efficient visual encoding and those with inefficient, noisier processing.
Adolescents in the less efficient group reported higher levels of emotional symptoms—such as anxiety and low mood—on clinical questionnaires. The researchers found that poorer visual coding in the brain was directly related to the presence of emotional distress, suggesting a link between how faces are visually represented and emotional well-being.
To probe the computational mechanism, the scientists trained a deep neural network with the dual processing pathways. The conceptual knowledge pathway was designed to regulate the visual pathway as the model learned to identify facial expressions, mirroring how abstract concepts can guide perception in the human brain. The researchers observed a strong alignment between the model’s internal code and the participants’ brain scans, indicating the simulation captured key aspects of perceptual processing.
The team then manipulated the strength of the conceptual pathway’s regulatory influence over the visual pathway. With an intermediate level of regulation, the network achieved optimal emotion recognition. However, when regulation was forced to an excessive level, the model began to misclassify neutral faces as angry or sad, producing a computational negative bias that paralleled the human data from the more symptomatic group.
The exaggerated regulation in the artificial model produced a brain-like pattern of inefficiency, closely matching the inefficient visual processing observed in adolescents with higher emotional symptoms. This finding points to overregularisation—where preexisting beliefs override sensory input—as a mechanism driving negative emotional biases.
Building on these results, the researchers created a computational marker that quantified each nineteen-year-old’s degree of visual overregularisation from their brain scans. They then followed 725 participants four years later to see if the marker could forecast the onset of mental health issues.
Among young adults with no clinical symptoms at nineteen, the marker successfully predicted those who would develop high levels of emotional symptoms by age twenty-three. The predictive power surpassed traditional baseline measures that rely on demographics and basic brain activation, suggesting that the overregulation mechanism may precede the appearance of emotional symptoms.
The team also explored genetic associations. They found the visual overregularisation marker linked to a genetic variant associated with major depressive disorder, located in a gene involved in brain development. The marker also correlated with the polygenic risk score for depression, while showing no association with genetic risks for alcohol use disorder or eating disorders, indicating a degree of specificity to depression.
In a separate clinical dataset comprising 134 patients with major depressive disorder and 62 healthy controls, the marker distinguished those with depression from controls more effectively than standard brain-activation measures. The marker did not reveal group differences in individuals with alcohol use disorder, anorexia nervosa, or bulimia nervosa, in line with the notion of specificity to depressive states.
The study notes that participants in both population-based and clinical cohorts were predominantly white and drawn from middle-to-high socioeconomic backgrounds. It calls for validation of these computational patterns across more diverse ethnic, racial and socioeconomic groups. While the visual-distortion marker shows promise, the authors caution that its current diagnostic power in the general population is modest and would require broader testing before clinical use.
Looking ahead, researchers see potential for tracking these markers in younger adolescents to determine whether the visual distortion emerges earlier in development and whether targeted interventions to retrain visual processing could mitigate depression risk. They emphasise that understanding whether the marker reflects a stable trait or a temporary state will be crucial for designing effective therapies.
The study, titled Deep learning reveals a neurocomputational mechanism predicting depression risk in adolescents, credited Han Lu and Qiang Luo along with a large collaborative team spanning multiple institutions and consortia, including STRATIFY/ESTRA and IMAGEN.
