The Algorithmic Agent Perspective and Computational Neuropsychiatry

Depression is a stuck attractor, not a chemical imbalance — and this paper builds the formal machinery to treat it that way.

The core idea is disarmingly simple: a brain is an agent that builds models of the world, evaluates how good its current situation is (the "valence" signal), and plans actions to improve things. Depression, in this framework, is what happens when that valence signal gets persistently stuck low — not because the world is necessarily bad, but because one or more of the three computational modules driving that signal has broken down. The Modeling Engine builds distorted or ruminative world models. The Objective Function miscalibrates reward and threat. The Planning Engine loses the ability to identify escape routes. Each failure mode maps onto specific brain circuits and produces a recognizable clinical picture: anhedonia, rumination, cognitive dyscontrol, threat dysregulation.

The dynamical systems framing is where this gets mechanistically interesting. Think of the brain's state as a ball rolling around a landscape, where altitude represents how bad you feel. A healthy brain has a varied landscape — the ball can roll toward higher-valence regions. A depressed brain has eroded into a deep pit: trauma carves the trough, maladaptive plasticity deepens it, and the system gets trapped. This isn't just metaphor — the paper formalizes it using Lyapunov attractor theory, which gives you precise language for why recovery requires more than removing the original stressor, why relapse is asymmetric, and why some interventions work by brute-force dislodging (ECT, psychedelics) while others work by slowly reshaping the landscape (SSRIs, CBT). Psychedelics, for instance, are interpreted as temporarily flattening the entire landscape — increasing brain entropy so the system can explore states it was previously locked out of — followed by a plasticity window where a new, healthier attractor can form.

The clinical payoff is a biotype stratification: rather than treating "depression" as one thing, the framework predicts that patients whose primary dysfunction is anhedonia (broken reward circuits, hypoactive nucleus accumbens) need different targeting than patients whose primary dysfunction is cognitive dyscontrol (hypoactive dorsolateral prefrontal cortex) or threat dysregulation (hyperactive amygdala). This isn't just taxonomic tidiness — it directly implies which brain region to stimulate, which therapy to prioritize, and eventually which personalized whole-brain model to simulate before touching the patient at all.

The paper is honest about what remains aspirational. The circuit-to-module mappings are first drafts. The biotypes are suggestive, not validated. No biophysically realistic whole-brain model of MDD yet exists. But the framework is explicit enough to generate falsifiable predictions — brain complexity measures should correlate with depression severity, psychedelics should measurably flatten the valence landscape, biotype-matched interventions should outperform generic ones — which is more than most psychiatric theories can claim. For a field that has historically oscillated between "serotonin is low" and "childhood trauma," having a unified computational account that explains why drugs, therapy, stimulation, and psychedelics all work, and when each is appropriate, is a genuine step forward.