Two Maps of the Mind: Why Psychiatry’s Biggest Frameworks Are Finally Merging

Mental health research is having a moment of reckoning. For decades, the field has been organized around diagnostic categories — depression, anxiety, schizophrenia — that clinicians can use and insurers can bill, but that increasingly look like shaky foundations for science. Two ambitious reform movements have emerged to fix this, each from a different direction. One, the Hierarchical Taxonomy of Psychopathology (HiTOP), rebuilds the map of mental illness from the bottom up using patterns of symptom covariation. The other, computational psychiatry, tries to explain mental disorders in terms of formal models of the brain’s neurocognitive machinery. A new Perspective published in Nature Reviews Psychology, led by Timothy A. Allen of the University of Pittsburgh and Michael N. Hallquist of the University of North Carolina at Chapel Hill, together with a large multidisciplinary workgroup, argues that these two movements are natural allies — and that their integration could be the key to finally connecting symptoms to brains.

The case begins with a problem that has haunted psychiatry for generations: the diagnostic categories themselves. The DSM and ICD systems carve mental distress into discrete boxes, yet decades of evidence show that patients rarely fit neatly into one box. Comorbidity is the norm rather than the exception; the National Comorbidity Survey Replication documented pervasive overlap among common disorders, and longitudinal work from the Dunedin Birth Cohort showed that mental health problems over four decades are better described by a single general dimension of vulnerability than by separate, stable diagnoses. Taxometric studies, meta-analyzed by Haslam and colleagues, consistently favor dimensional over categorical models of psychopathology. In other words, the boundaries between disorders appear to be conventions, not discoveries.

HiTOP is the most systematic attempt to replace those conventions with something empirical. Rather than starting from clinical tradition, HiTOP applies factor-analytic methods — the same statistical machinery behind principal component analysis — to the way symptoms actually co-occur in large samples. The result is a hierarchy: at the top sits a general factor of psychopathology, sometimes called the p factor, which captures overall severity; below it are broad superspectra such as internalizing, externalizing, and psychosis; below those are spectra, syndromes, and finally individual symptoms and traits. A 2023 meta-analysis by Ringwald, Forbes, and Wright found substantial structural evidence for this arrangement, and subsequent work has extended it to neurodevelopmental conditions and to a data-driven reconstruction of DSM-5 symptoms published in 2025. Crucially, HiTOP dimensions have shown real-world predictive value: higher-order internalizing dimensions predict treatment response in partial hospitalization programs, and the p factor forecasts long-term psychiatric and functional outcomes in anxious youth.

But HiTOP, by design, is descriptive. It tells us how psychopathology is structured, not why. Its dimensions are latent variables — statistical constructs inferred from covariation — and critics, including the authors themselves in earlier papers, have noted that latent variables alone do not constitute a theory of mechanism. This is precisely where computational psychiatry enters the picture. The field, crystallized in a landmark 2012 Trends in Cognitive Sciences paper by Montague, Dolan, Friston, and Dayan, aims to explain psychiatric phenomena using formal mathematical models of cognitive and neural processes: reinforcement learning models that quantify how people learn from reward and punishment, drift diffusion models that decompose decision-making into speed and caution, Bayesian models of belief updating, and dynamic causal models of brain circuits. Instead of asking whether a patient has depression, a computational psychiatrist asks whether that patient’s learning rate, reward prediction errors, or evidence-accumulation efficiency deviate from normative functioning — and by how much.

The Perspective distinguishes two cultures within computational psychiatry, echoing a well-known 2019 commentary by Bennett, Silverstein, and Niv. One is data-driven: machine learning approaches that mine neuroimaging, genetic, and behavioral data to predict diagnoses, treatment outcomes, or disease trajectories. The other is theory-driven: mechanistic modeling that begins with an explicit hypothesis about a cognitive process and fits it to behavior. Both cultures, the authors argue, are hampered by the same weakness — the messiness of the diagnostic labels they are asked to predict or explain. Machine learning models trained on DSM categories inherit the heterogeneity and label noise of those categories, which may partly explain why clinical prediction models have struggled to generalize. A 2024 commentary in JAMA Psychiatry asked pointedly whether psychiatric nosologies are limiting the success of clinical prediction models. Meanwhile, theory-driven studies that compare a single diagnostic group with healthy controls risk conflating distinct processes that happen to co-occur in one disorder.

This is where the two frameworks begin to fit together like puzzle pieces. HiTOP offers computational psychiatry something it badly needs: reliable, specific, empirically grounded phenotypes. Consider the growing literature on reinforcement learning in mood and anxiety disorders. Meta-analyses by Pike and Robinson and by Halahakoon and colleagues found that reward-processing deficits in depression are real but modest and heterogeneous when patients are defined by diagnosis. Studies that instead relate learning parameters to transdiagnostic symptom dimensions — such as anhedonia or compulsivity — tend to find cleaner signals. Claire Gillan’s work showed that deficits in goal-directed control relate to a self-reported compulsivity dimension more strongly than to an obsessive-compulsive disorder diagnosis per se. Similarly, Wise and Dolan linked aversive learning parameters to transdiagnostic symptoms in a general population sample, and a computational factor modeling approach by Wise, Robinson, and Gillan demonstrated that mechanisms can be identified across diagnoses when the phenotype is dimensional.

The integration also runs in the other direction: computational psychiatry gives HiTOP a path from description to explanation. The Perspective highlights examples where model-based measures behave like computationally defined traits. Efficiency of evidence accumulation, estimated with sequential sampling models, emerges as a task-general marker of cognitive efficiency that is more reliable than many traditional executive function measures and predicts substance use in emerging adulthood. Drift diffusion modeling has disentangled which components of externalizing psychopathology — antagonism rather than disinhibition — are tied to poor cognitive control. Neurocomputational studies of social exchange have illuminated mechanisms behind callousness, and model-based planning deficits in compulsivity have been traced to faulty neural representations of task structure. These are the kinds of findings that could eventually populate HiTOP’s dimensions with explanatory content: not just ‘this set of symptoms co-occurs,’ but ‘this set co-occurs because it reflects a common deviation in a specific neurocognitive process.’

The neurobiological evidence reinforces the case. Transdiagnostic neuroimaging studies have repeatedly found that brain signatures cut across DSM categories: common structural and functional disruptions span mood, psychotic, and anxiety disorders, and a transdiagnostic network for psychiatric illness has been derived from atrophy and lesion data. Normative modeling approaches, which quantify each individual’s deviation from expected brain development, show that transdiagnostic dimensions of psychopathology explain unique deviations in brain structure. Genetic studies point the same way: psychiatric disorders share substantial polygenic risk, and the genetic architecture aligns better with hierarchical dimensional models than with diagnostic categories. If biology does not respect the DSM’s boundaries, the authors argue, then a framework that organizes phenotypes dimensionally — and pairs them with computational measures of process — offers the most promising route to a genuinely mechanistic clinical neuroscience.

Yet the authors are candid about the barriers. Cross-field collaboration remains limited; HiTOP grew out of clinical psychology and quantitative psychometrics, while computational psychiatry has roots in cognitive neuroscience, machine learning, and even economics, and the two communities publish in different journals and attend different meetings. Multimodal data pose technical challenges: combining self-report, behavior, neuroimaging, and genomics requires measurement models that most studies lack. Measurement reliability is a persistent headache — many classic cognitive tasks show poor test–retest reliability at the individual level, and landmark analyses by Elliott and colleagues and by Marek and colleagues showed that reproducible brain–behavior associations often require thousands of participants. There are also unresolved conceptual questions about whether computational and biological measures should be allowed to revise the HiTOP structure itself, or whether the taxonomy should remain purely phenomenological. A 2025 meta-analysis asking whether cognitive functions belong in the HiTOP model illustrates how live this debate is.

The stakes of getting this right are enormous. The authors point to early evidence that computational measures can mediate treatment effects: reinforcement learning disruptions in depression predict sensitivity to cognitive behavioral therapy, neural correlates of reward prediction errors classify therapy response, and different components of cognitive-behavioral therapy have been shown to affect specific cognitive mechanisms. If HiTOP provides the reliable descriptive scaffolding and computational psychiatry supplies the mechanistic engine, the combination could transform how researchers stratify patients, predict outcomes, and design interventions — moving the field from a nosology of named boxes toward a quantitative science of mental processes. The Perspective is, in effect, an invitation: two of psychiatry’s most promising reform movements have been building toward the same destination from opposite ends, and the shortest path forward may be the one that connects them.

Subject of Research: Integration of the Hierarchical Taxonomy of Psychopathology with computational psychiatry to connect dimensional symptom models with neurocognitive mechanisms

Article Title: Integrating the Hierarchical Taxonomy of Psychopathology and computational psychiatry

Article References: Allen, T. A., Schreiber, A. M., Blain, S. D., Carlisi, C., DeYoung, C. G., Fornito, A., Gillan, C. M., Goghari, V. M., Hall, N. T., Homan, P., Kaczkurkin, A. N., Kotov, R., Krueger, R. F., Latzman, R. D., Okan, A., Tackett, J. L., The HiTOP Neurobiological Foundations Workgroup, & Hallquist, M. N. (2026). Integrating the Hierarchical Taxonomy of Psychopathology and computational psychiatry. Nature Reviews Psychology. https://doi.org/10.1038/s44159-026-00621-7

Image Credits: AI Generated

DOI: 10.1038/s44159-026-00621-7

Keywords: HiTOP, computational psychiatry, psychopathology, dimensional models, reinforcement learning, drift diffusion modeling, machine learning, neuroimaging, DSM, transdiagnostic, p factor, mental health research