Fire risk evaluation of heat-not-burn products and combustible cigarettes: Thermal exposure of different materials under ambient conditions.

Major depressive disorder (MDD), which afflicts more than 332 million of the world population (), is often difficult to treat once fully established () and is highly prone to relapse (). Consequently, the prevention and early detection have emerged as global health priorities (). While association-based studies of the human brain have cataloged numerous neural correlates (), the field suffers from a crucial bottleneck: the absence of mechanistically interpretable biomarkers that can predict individual risk (). Bridging the gap between descriptive associations and computational mechanisms is essential to transform psychiatric nosology and enable preemptive prevention. Deficient emotional processing is a hallmark feature of depression () and a promising candidate marker for illness vulnerability (). However, progress has been limited due to the lack of mechanistic understanding of these deficits. Leading theoretical frameworks, such as the theory of constructed emotion (), posit that conceptual knowledge from memory systems shapes sensory coding in the visual hierarchy (). Neuroimaging studies support this view, showing that emotional categories can be decoded from visual cortex activity (–), and its dysfunction during negative facial emotion processing is linked to depression (, ). We focused specifically on angry facial expressions as they serve as potent signals of social threat and rejection. Altered processing of such social cues is central to the interpersonal difficulties and negative processing biases often observed in depression (). More recently, the brain’s responses to angry faces have been decomposed into two networks, an orbitofrontal-related network and an occipital-related network, both of which are linked to the resilience to developing emotional disorders following childhood adversity (). Nevertheless, these observations remain fundamentally correlative, leaving a knowledge gap: How are emotion-specific visual representations computationally formed through learning, and how does the disruption of this process confer vulnerability to disease? To resolve this gap, we introduce a perturbation-based neurocomputational framework. We instantiate a dual-pathway deep neural network (DNN) () in which a “conceptual pathway” encodes abstract emotional concepts that then regulate learning in a “perceptual network.” Conceptually, “regularization” here refers to the process where the brain uses abstract knowledge to guide and constrain how it processes raw visual information. Ideally, this helps us recognize facial emotions quickly. However, if this guidance becomes too rigid (i.e., “overregularized”), perception may become biased by negative concepts rather than accurately reflecting the visual reality. This architecture provides a computational testbed: By systematically perturbing the strength of the conceptual regularization in the model, we can quantitatively simulate the emergence of both normative and pathological emotion representations. We then directly test this model against human neuroimaging data, evaluating the alignment between DNN-derived representations and brain activity during the perception of angry faces (). This approach moves beyond traditional correlational neuroscience to propose a theoretical link between a defined computational mechanism and depression risk, thereby suggesting a previously unrecognized class of candidate mechanistic markers for preemptive psychiatry.