For decades, psychiatrists have prescribed antidepressants largely by trial and error — try one, wait six to eight weeks, switch if it doesn’t work, try again. In June 2024, a Stanford team led by Leanne Williams and Leonardo Tozzi published something that could change that pattern: using functional MRI scans of 801 people with depression and anxiety, they identified six statistically distinct “biotypes” of the illness, each defined by a different pattern of dysfunction across specific brain circuits — and, critically, each responding differently to medication versus talk therapy. It’s one of the clearest pieces of evidence yet for an idea computational psychiatry has been building toward for years: that “depression” isn’t one disease with one best treatment, but a label covering several distinct brain-level conditions that have been treated as interchangeable mostly because psychiatry lacked the tools to tell them apart.
The Scientific Foundation
The Stanford study used a standardized imaging protocol probing six brain circuits previously implicated in depression — including the default mode circuit (active during rest and self-referential thought), the salience circuit, and the cognitive control circuit — measuring both resting brain activity and activity evoked by specific emotional and cognitive tasks, an approach the authors compared to how cardiologists use both resting and stress-test measurements to fully characterize heart function. Clustering patients by these brain circuit measures produced six biotypes with meaningfully different clinical profiles: one subtype, marked by reduced activity in a circuit governing attention, responded significantly worse to standard behavioral talk therapy than the other groups, while a different subtype with elevated activity in circuits tied to problem-solving showed better symptom relief from that same therapy. This built on earlier work from the same research group, including a 2023 JAMA Network Open study identifying a specific “cognitive biotype” of depression from a randomized clinical trial, that similarly found brain-circuit measures predicted differential treatment response.
This fits into a broader shift the field calls computational psychiatry, which treats psychiatric symptoms not as a checklist of behaviors but as the observable output of measurable, quantifiable brain dynamics. A 2025 Viewpoint in Nature Computational Science, authored by researchers including Thomas Wolfers, described the current push toward integrating “normative modeling” — statistical models of how brain measures vary across large healthy populations, against which an individual patient’s brain can be compared to quantify exactly how and where their neurobiology deviates — as a central path toward genuinely precision psychiatric care.
The Cross-Domain Connection
What makes this a real cross-domain story is that it borrows its central toolkit from dynamical systems theory and control engineering, fields with no traditional connection to psychiatry at all. A striking December 2025 paper in Neuropsychopharmacology, “Re-engineering the disordered mind,” proposes treating a person’s mental state as a point moving through a mathematical space called a neural manifold, and building individualized AI models — trained on dense, repeated measurements from a single patient over time — that can simulate counterfactual interventions before they’re tried on the actual person. The authors describe a closed-loop, “N-of-1” experimental paradigm: rather than testing one treatment across many patients and hoping the average result applies to any given individual, an AI surrogate model is trained on one person’s own longitudinal data, used to simulate how that specific person’s brain state would likely respond to different treatment perturbations, and then used to select an intervention optimized for them specifically.
This is genuinely novel because it applies engineering concepts developed for controlling physical systems — feedback loops, system identification, simulated interventions before real-world deployment — to something as historically qualitative as a person’s mood and cognition. It also draws on network theory borrowed from sociology and physics, treating mental disorders as patterns of interacting symptoms and processes rather than singular underlying causes, an approach formalized in psychologist Denny Borsboom’s influential 2017 network theory of mental disorders.
What Remains Undemonstrated
The gap between this framework and everyday clinical psychiatry remains wide. The Stanford biotype study’s own commentary was explicit that these findings are not yet ready for standard clinical use, and represent an initial proof of concept rather than a validated diagnostic tool — replication in larger, more diverse patient populations is still needed before biotype-guided treatment selection could become routine care. The N-of-1 AI surrogate model approach described in the 2025 Neuropsychopharmacology paper is, by the authors’ own framing, a proposed clinical experimentation paradigm rather than an already-validated method; no large-scale trial has yet demonstrated that AI-simulated counterfactual interventions reliably predict real-world treatment outcomes in psychiatric patients. There are also documented equity concerns actively being studied: a 2025 analysis found significant variation in how AI systems including several large language models assessed psychiatric symptoms depending on racial cues embedded in prompts, a reminder that computational tools trained on historical clinical data can inherit and amplify the same biases and access gaps that already exist in mental healthcare.
Why It Matters
The World Health Organization projects mental disorders will become the leading cause of global disease burden by 2030, and a 2023 cross-national analysis in The Lancet Psychiatry estimated that roughly half the world’s population will experience one or more mental disorders by age 75. Against that backdrop, the current trial-and-error model of psychiatric treatment, where patients often cycle through multiple medications over months before finding one that helps, carries a real human and economic cost. Biotype-guided treatment selection or individualized dynamical models, if they mature, could meaningfully shrink that time-to-effective-treatment window — directing a patient toward talk therapy, a specific medication class, or another intervention based on their actual brain dynamics rather than population averages that may not describe them well at all.
The Human Dimension
There’s something genuinely hopeful in the idea that a diagnosis long treated as a single, monolithic label might actually be several different conditions wearing the same name — because it means the frustrating, demoralizing experience so many depression patients describe, of trying medication after medication that simply doesn’t work, may not be a personal failure or a mysteriously treatment-resistant case, but simply evidence that they were never matched to the right intervention for their particular brain in the first place.
Sources:
1. Tozzi, Zhang, Pines et al., “Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety,” Nature Medicine, 2024 — https://www.nature.com/articles/s41591-024-03057-9
2. Hack, Tozzi, Zenteno et al., “A Cognitive Biotype of Depression and Symptoms, Behavior Measures, Neural Circuits, and Differential Treatment Outcomes,” JAMA Network Open, 2023 — referenced via ClinicalTrials.gov, NCT00693849
3. Kheirkhah, Shariatpanahi, Hahn et al., “Re-engineering the disordered mind: clinical experimentation, dynamical systems, and AI for personalized psychiatry,” Neuropsychopharmacology, 2025/2026 — https://www.nature.com/articles/s41386-025-02303-z
4. Akiki, Williams, Wolfers et al., “Transforming psychiatry with computational and brain-based methods,” Nature Computational Science, 2025 — https://www.nature.com/articles/s43588-025-00884-9
5. “Harnessing artificial intelligence for mental health care,” eBioMedicine, 2025 — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11784658/
6. Okesanya et al., “Artificial intelligence in psychiatry: transforming diagnosis, personalized care, and future directions,” Exploration of Digital Health Technologies, 2025 — https://www.explorationpub.com/Journals/edht/Article/101174
7. “Six distinct types of depression identified in study combining brain imaging with machine learning,” Medical Xpress coverage of Tozzi et al., 2024 — https://medicalxpress.com/news/2024-06-distinct-depression-combining-brain-imaging.html
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 4.6. Published at artificialideas.org.
Artificial Ideas 