Nature Human Behaviour

Functional neuroimaging aims to uncover brain processes underlying behavior and disease, yet studies are often underpowered to detect these effects. How this literature has shaped our understanding of brain function remains unknown, and little guidance exists for planning better powered studies. An underappreciated barrier is that commonly reported effect sizes across the brain are inflated, biasing study planning. Here, we introduce a correction for this inflation bias and show how more accurate studies can be planned using corrected effect size benchmarks from a mega-analysis of 63 typical studies across seven large datasets (52,979 participants). We find that common methods of planning studies based on uncorrected effects lead to roughly half the expected detections at typical sample sizes, with limited spatial overlap with original findings. These missed effects collectively explain meaningful additional variance in the desired outcome. We show how to recover missed effects by planning not only for power but also for a target number of detections via corrected benchmarks, or by taking a whole-brain approach with multivariate effects that individual research groups can detect (n < 50 compared to n > 1,000 for a typical univariate effect). These findings lay the groundwork for more informed study planning and a richer understanding of the widespread nature of brain effects, with implications for shared challenges (and solutions) across biomedicine.

Stimulant use disorder (StimUD) is a significant public health problem, but genetic studies have been limited by small sample sizes. We conducted genome-wide association studies (GWAS) of StimUD in the Million Veteran Program (MVP) and All of Us (AOU), followed by meta-analysis with FinnGen and 10 additional datasets, for a total of 709,369 individuals (Ncases=33,977, Ncontrols=675,392) in four broad ancestry groups: European (EUR) (Ncases=22,564, Ncontrols=624,672), African (AFR) (Ncases=7,574, Ncontrols=34,189), Admixed American (AMR) (Ncases=3,657, Ncontrols=15,698), and East Asian (EAS) (Ncases=182, Ncontrols=833). Population-specific SNP heritability was 6.1% in EUR and 2.4% in AFR. We discovered a total of 19 genome-wide-significant loci, six in EUR, including DRD2*rs5794864, P=7.32E-10, one in AFR, five in a multi-ancestry meta-analysis, including CHRNA5*rs55781567, P=3.27E-9, two in a male-only meta-analysis, including FTO*rs8057044, P=9.50E10-9, and five in a meta-analysis of sex-stratified results. In a hold-out AOU subsample (NEUR=18,841, NAFR=12,263, NAMR=9,739), ancestry-specific polygenic risk scores were significantly associated with StimUD in EUR (OR=3.28, 95% confidence interval (CI)=2.89-3.71) and AMR (OR=2.01, 95% CI=1.71-2.37). Transcriptome-wide association studies, fine-mapping, and colocalization analyses prioritized additional genes (e.g., GPX1, BSN). Genetic correlation, Mendelian randomization, and causal mixture analyses revealed relationships with other substance use and use disorder phenotypes, including cannabis use disorder (rg=0.94, P=5.43E-237) and opioid use disorder (rg=1.01, P=4.40E-107), and other psychiatric traits, including anxiety, depression, neuroticism, and attention-deficit/hyperactivity disorder. This is the first well-powered GWAS of StimUD, and it offers significant insights into disease biology.

Adaptive learning requires distinguishing volatility, changes in the latent state of the environment, from moment-to-moment stochasticity of observations. The two demand opposite adjustments to the learning rate: volatility calls for faster updating, stochasticity for slower. Disentangling them is computationally difficult because both inflate experienced variance, leaving the inference prone to systematic individual differences with potential consequences for psychopathology. Three computational phenotypes capture this variation: intact learners; stochasticity-blind learners, who over-update by treating noise as change; and volatility-blind learners, who under-update by treating change as noise. In two large online samples and across three tasks, we found a double dissociation between these phenotypes and transdiagnostic psychiatric dimensions: stochasticity-blind learners scored higher on Internalizing (anxiety, depression), volatility-blind learners on Externalizing (behavioral addiction, compulsivity). Distinct symptom dimensions thus correspond to distinct failures of inference about uncertainty, supporting a selective rather than generalized account of learning-under-uncertainty deficits in psychopathology.

BackgroundQuantifying the alignment between scientific production and population-level indicators remains a persistent methodological challenge in health research evaluation. While longitudinal ecological models have been increasingly used to explore associations between research output and societal outcomes, their feasibility, interpretability, and structural limitations have not been systematically examined. MethodsWe conducted a longitudinal ecological meta-research analysis integrating global bibliometric data on mental health publications with country-level indicators of mental disorders, mental health infrastructure, and subjective well-being. Analyses were stratified by World Bank income groups and implemented using a three-step framework comprising income specific linear regression models, random-effects meta-analyses, and meta-regressions to assess association patterns, heterogeneity, and potential moderators. ResultsScientific production was highly concentrated in high-income countries. Income-stratified regression models revealed divergent association patterns across contexts, with inverse associations observed in higher income groups and predominantly positive coefficients in low-income countries. Meta-analyses showed extreme between-group heterogeneity for most indicators, yielding largely attenuated pooled estimates. Only one subjective well-being indicator retained a significant pooled association. ConclusionsLongitudinal ecological models linking scientific production to population-level indicators can identify broad association patterns and structural asymmetries but are strongly constrained by contextual heterogeneity and data availability.

Adaptive learning requires distinguishing environmental volatility from observation stochasticity, two sources of uncertainty that demand opposite adjustments to the learning rate but inflate experienced variance similarly. Disentangling them is computationally difficult with no tractable closed-form solution. Particle-filter methods are the natural tool for this kind of joint inference, but their stochastic likelihoods and non-differentiable objectives force derivative-free fitting protocols and discourage the individual-difference analyses central to cognitive modeling, where small effect sizes leave little room for additional estimator noise. We introduce the Categorical Bayes Filter (CBF), a deterministic alternative that preserves the conditional structure of recent particle-filter accounts but replaces the stochastic outer layer with a categorical distribution on a quantile grid parameterized through differentiable Beta quantile functions. The procedure performs evidence maximization with an exact, deterministic marginal likelihood that is fully differentiable in the grid parameters. In a volatility-stochasticity task with N = 643 participants, fitted CBF dispersion parameters reveal a cross-over phenotyping pattern between volatility-blind and stochasticity-blind subjects that is not recoverable from particle-filter parameters fit to the same data under a state-of-the-art protocol. The deterministic structure also yields a trial-by-trial ambiguity signal that predicts response times not used in fitting. More broadly, the approach opens individual-level analyses in cognitive modeling and computational psychiatry that stochastic methods have effectively foreclosed.

How do humans store sequences that far exceed working memory capacity? Using visuo-spatial and binary auditory sequences, we previously showed that a Language of Thought (LoT) architecture -- in which simple primitives are recursively combined into hierarchical programs -- enables efficient storage of structured sequences. Here we ask whether this principle extends to purely ordinal structure: sequences defined by how items repeat and in what order, as in AABBCCAABBCC, independently of their spatial content. Across three experiments, participants reproduced 12-item sequences of spatial locations with various ordinal structures. The minimal description length derived from the LoT model predicted recall accuracy with remarkable precision (r = .96), substantially outperforming Shannon entropy, Lempel-Ziv complexity, chunking models and subjective complexity ratings. Critically, fine-grained analyses of participants inter-click intervals during reproduction revealed systematic slowdowns at the hierarchical boundaries predicted by the LoT programs, providing a behavioral signature of the underlying mental syntax. These results identify a compact vocabulary of mental primitives -- repetition, mirroring, and interleaving -- whose composition accounts for the symbolic compression of ordinal structures. For ordinal regularities, human sequence memory operates as a form of program induction, leveraging a domain-general capacity for hierarchical compression to encode complex structured information. Author SummaryHuman short-term memory is heavily limited, holding no more than a few items at once. Yet humans routinely memorize complex sequences that far exceed this capacity. How is this possible? We propose that the brain acts like a programmer: rather than storing each element individually, it compresses sequences into short mental "programs." Just as a programmer writes "repeat ABC four times" instead of typing ABCABCABCABC, the brain leverages regularities such as repetitions (ABC-ABC) or mirror patterns (ABC-CBA) to encode sequences efficiently. We tested this idea across three experiments: two in which participants memorized and reproduced sequences of spatial positions on a screen, one where they only rated their perceived complexity. Sequences described by shorter programs were remembered far better and judged as simpler -- even when they were the same length as less structured sequences. When reproducing sequences, participants paused longer at structural boundaries, revealing the internal organization of their mental programs. Strikingly, program length predicted memory performance better than participants own complexity ratings, suggesting that these mental representations are not fully accessible to conscious awareness. Finally, we identified key new patterns -- including temporal inversion and interleaving -- that extend the Language of Thought framework. Together, these findings suggest that a compositional Language of Thought is a fundamental aspect of how the human brain efficiently store and represent structured information.

Rhythmic temporal structure improves working memory, but how this benefit emerges from recurrent dynamics remains unclear. Here, we trained excitatory-inhibitory recurrent neural networks with short-term synaptic plasticity to perform a sequential delayed match-to-sample task with either regular or jittered sample timing. Rhythmic input produced a small but reliable improvement in task accuracy and was associated with more differentiated population trajectories during encoding. This behavioral advantage was accompanied by an organization of population dynamics around the dominant input frequency: temporal regularity progressively brought stimulus arrivals closer to preferred encoding phases, modulated phase advancement during stimulus presentation, and reduced the deviation of inter-stimulus phase-progression frequency from the dominant input rhythm. As a result, internal oscillations increasingly tracked the temporal structure of the input across the sequence, providing a phase-based scaffold for encoding ordered information. This scaffold preferentially supported temporal-order representations rather than uniformly enhancing all stimulus features. Decoding analyses further showed that stronger temporal regularity increased the fidelity and persistence of stimulus information in both neuronal activity and synaptic efficacy, whereas perturbing synaptic efficacy produced the largest impairment during delay-period maintenance. These findings suggest that rhythmic input supports sequential working memory by imposing a reliable temporal structure on recurrent dynamics and stabilizing synaptic-state representations.

Working out whether others care for us is crucial in personal relationships and when seeking professional help. It is often most difficult for those most in need, e.g. following interpersonal traumata. Here, we introduce a simple Caring Attributions task and elucidate key computational mechanisms involved and, using EEG, the cortical activity representing the degree of belief that another is beneficent or maleficent. We find evidence for a new type of neurocomputational processing: valence-partitioned temporal-difference inference (TD-Bayes). This employs primary processing about latent causes, but also separate channels to represent these different valenced attributions, inspired by value-partitioned associative learning (VPAL). TD-Bayes uses slow propagation of beliefs using temporal-difference updating. These models gave a very good account of behaviour, slightly better than VPAL, but crucially, their partitioned representations have stronger, distinct representations in ERP signals. They provide a promising inroad into the understanding of how people may jump to atttibutions about caring vs. uncaring others.

Human perception is shaped by both sensory input and prior knowledge or expectations. But how does prior contextual information influence rapid visual processing? Here, we combined eye tracking with feature-based encoding models across two experiments to predict detection latencies in a core visual task: rapid face detection in natural scenes (N = 38 per experiment). In the first experiment, we manipulated the presence of faceless scene previews. In the second experiment, we additionally restricted peripheral visual input using a moving-window paradigm, thereby increasing reliance on prior information. Across both experiments, prior context facilitated face detection, particularly for challenging images. This facilitation was already evident in the very first eye movement, suggesting that previews shape perceptual strategies from the outset. To quantify what information guided behavior, we modeled detection latencies using a set of image-based predictors capturing (i) sensory information and (ii) a scene-derived spatial prior: the expected face location. Both predictor classes explained latency variation across images. Among sensory predictors, the difference in deep neural network responses induced by the presence of the face provided the strongest out-of-sample prediction of detection latency. Critically, when scene previews were available, the contribution of the spatial prior increased, while reliance on sensory-driven features was generally reduced. Together, these findings indicate that prior scene context shifts the balance of information used for rapid face detection from sensory-driven to expectation-based spatial guidance.

Delayed diagnosis and poor antiretroviral therapy (ART) adherence remain primary drivers of HIV-related morbidity in low-resource settings, yet real-world AI validation at scale is lacking. We conducted a retrospective validation study using two publicly available, de-identified datasets: a Quality of Care cohort of 27,288 HIV-positive patients on ART across multiple healthcare facilities, and the CEPHIA multi-country assay database comprising 165,444 specimen records from six countries. Four machine learning classifiers were evaluated using 10-fold stratified cross-validation with SMOTE applied strictly to training folds. Explicit data leakage prevention, ablation analysis, calibration assessment, and bootstrap confidence intervals were applied. Economic projections used one-way sensitivity analysis. This study adheres to TRIPOD reporting guidelines. Random Forest achieved AUC-ROC of 0.9753 (95% CI: 0.970-0.975), sensitivity 87.3% (95% CI: 86.4-88.2%), specificity 95.7% (95% CI: 95.2-96.2%), and Brier score 0.079. Ablation testing confirmed robustness (AUC 0.963 without the primary predictor). Temporal validation on held-out future patients yielded AUC 0.772 (95% CI: 0.744-0.802), confirming generalisation across time. Real-world analysis revealed median diagnosis-to-ART delay of 74 days, with 47.3% of patients exceeding 90 days and 36.7% presenting with CD4 below 200 cells per microlitre. Multi-country CEPHIA analysis identified 18.6% HIV recency within the 130-day early-intervention window. Decision curve analysis confirmed net clinical benefit across threshold probabilities 0.03-0.45. Subgroup analysis demonstrated consistent AUC across sex, age, CD4 strata, and WHO staging (max difference 0.051). Economic modelling projected base-case savings of USD 415 per patient (USD 2.07 million per 5,000-patient cohort). These findings provide large-scale empirical evidence that AI-driven informatics can predict ART adherence failure and quantify systemic care gaps, offering a scalable framework for equitable HIV care delivery in resource-limited settings. Prospective external validation is required before clinical deployment.

We study how confidence in perceptual decisions depends on whether it is communicated verbally (e.g., "very likely") or numerically (e.g., "80% certainty"). We find that verbal expressions more reliably distinguish correct from incorrect choices than numerical reports, challenging the common assumption that numerical probabilities provide more precise representations of uncertainty. Additionally, in a dyadic decision-making task in which participants can revise their initial reports based on a partners choice and expressed confidence, verbal and numerical reports are equally effective in supporting accurate revisions of initial judgments. Together, these results underscore the effectiveness of verbal expressions as a means of conveying decision confidence.

Anhedonia - the diminished capacity to experience or anticipate pleasure - is among the most common consequences of early-life unpredictability, yet how these co-occurring conditions jointly shape real-world decision-making remains unknown. Here, we use a sequential foraging-under-threat task to probe motivational conflict decisions in 357 individuals varying in early-life unpredictability and anhedonia symptoms. We find that unpredictability and anhedonia exert opposing influences on choice: unpredictability shifts behavior away from the survival-optimal policy in a sex-dependent manner, while anhedonia promotes adherence to it, partly through heightened sensitivity to unexpected threatening outcomes. A mediation analysis reveals that anhedonia partially buffers the deleterious effects of unpredictability on decision quality. These results demonstrate that co-occurring conditions can mask one anothers behavioral signatures and suggest that the heterogeneous expression of transdiagnostic constructs like anhedonia may reflect context-dependent adaptations to distinct underlying etiologies.

Skilled action requires expressing motor memories as appropriate for the current context, but context is often uncertain. Theoretical models make conflicting proposals about memory expression under contextual uncertainty, predicting belief-weighted combination of memories versus the selection of the most probable memory. We tested these predictions by training human participants to reach in two opposing force fields cued by the direction of a random dot motion stimulus whose coherence varied. When participants moved before reporting dot direction, adaptation scaled with coherence: low-reliability cues produced partial expression of both memories. Fitting Bayesian observer models to behavior favored belief-weighted memory combination. In contrast, when participants reported their choice before moving, adaptation was independent of coherence and model fits favored categorical memory selection. Thus, sensorimotor memories are expressed as either a probabilistic combination or categorical selection, depending on whether participants contextual inference remains implicit or is made explicit at the time of memory expression.

Formal writing is evolutionarily recent, yet the brains of literate adults contain regions - the OTS-words subregions - that respond more strongly to written text than other stimuli. We tested a novel solution to this multi-disciplinary paradox: Does symbolic language processing, which emerged early in human history, lay the neural foundation for reading? In a longitudinal fMRI study, we followed 17 children through their first year of literacy training and related neural responses to text, symbolic language processing, and emerging reading skills over time. We found that middle OTS-words is engaged in symbolic language processing before children learn to read, and that this early engagement predicts later text selectivity and reading ability. These findings suggest that literacy builds on a pre-existing neural scaffold linking vision and language.

Context is critical for both human and artificial speech comprehension systems. While the role of preceding context in speech processing has been well documented, the neural mechanisms supporting the integration of subsequent input -- phonemes and words that occur in the future -- remain poorly understood. Here, we leverage advances in artificial speech systems to model the contribution of different sources of context on the neural encoding of speech in the human brain. For neural encoding, context-informed but not context-uninformed speech model embeddings explain unique variance in human neural activity beyond acoustics, including in early speech processing regions. In particular, model embeddings informed by past, future, and surrounding context explain activity in distinct intracranial electrodes. These electrodes are left-lateralised, and spatially intermixed in the temporal lobe. We find that beyond-word context is crucial for the representational quality of speech model embeddings, and in particular for the encoding of abstract linguistic information. Our finding that spatially neighboring yet distinct neural populations in the temporal lobe encode representations shaped by different contextual sources (past, future, and surrounding input) provides key insight into the neural circuitry that integrates multiple forms of contextual information. Furthermore, our results may inform the downstream use of self-supervised speech representations in language technology tasks, and in models of speech comprehension in the human brain.

Human behavioral and mental health outcomes arise from interactions among genetic, environmental, and neurobiological systems. Existing frameworks often model these components jointly, but many treat variables independently or use static representations. This limits their ability to capture system-level dynamics and changes over time. To address this, we developed DynoSys, a unified framework that integrates these signals using three layers: predictive models, relationship exploration models, and mechanism-oriented explanation models. Building on this framework, we introduce DynoSys 2.0, a graph-based temporal modeling approach inspired by the free-energy principle by Karl Friston. In this framework, each individual is represented as a dynamic graph that evolves over time. We hypothesize that healthy development and adverse mental health outcomes correspond to different system states and trajectories. Using longitudinal data from the Adolescent Brain Cognitive Development (ABCD) Study, we construct time-indexed graphs that integrate polygenic risk scores (PRS), multi-domain environmental features, and neuroimaging-derived representations. We study six phenotypes: externalizing behavior, internalizing behavior, and sub-stance use initiation (alcohol, nicotine, cannabis, and any substance). In these graphs, nodes represent domain-level features, and edges capture relationships derived from data-driven feature selection and temporal dependencies. We model graph evolution using recurrent neural networks and graph-temporal learning methods. We also define system-level measures, including graph energy and state transitions, to quantify dynamic patterns. Our results show that DynoSys 2.0 can model behavioral development using longitudinal multi-domain data. The framework achieved meaningful prediction for both continuous behavioral symptoms and substance-use initiation outcomes, but performance differed by outcome type. Externalizing behavior was predicted more accurately than internalizing behavior, and alcohol and any substance initiation showed stronger prediction than cannabis and nicotine initiation. Graph-derived energy measures showed clearer separation for high-versus low-symptom externalizing and internalizing groups, suggesting that continuous behavioral symptoms may be linked to different latent system states over time. Overall, DynoSys 2.0 provides a flexible framework for studying behavioral risk as a dynamic developmental process, while rare-event prediction and detailed graph-level interpretation require further work.

Humans have a remarkable ability to judge causal relationships from a limited number of un-reliable observations. Past work on causal cognition has largely focused on normative accounts of human behavior, leaving unknown how biologically plausible neural systems could learn causal relationships from observations and update their representations of causal structure with additional evidence. Here, we leverage task-optimized recurrent neural networks to discover candidate implementation-level neural mechanisms of causal judgment. We propose a novel cognitive task in which a subject observes stochastic samples from an unknown causal structure (e.g. among variables A, B, and C with unknown causal relationships), and must judge whether a specific causal relationship is present given a query (e.g. "does A cause C?"). We found that, after training, recurrent neural networks perform the task with high accuracy, adopt strategies that incorporate the behavior of non-queried variables to form their judgments, and, despite being trained only on pairwise queries ("does A cause B?", "does C cause A?", etc), form implicit beliefs about the complete graphical structure underlying the observations. Lastly, we use dynamical systems analysis to identify a set of low-level neural mechanisms that implement causal judgment and representation of causal graphical structure. Together, these findings lay the groundwork for a "bottom-up" approach to causal cognition, providing a potential basis for subsequent experimental study in the brain.

Cultural engagement is associated longitudinally with better mental health and reduced depression incidence, but evidence has largely relied on self-reported symptoms and diagnoses, leaving uncertainty about clinically recorded disorders, and residual confounding remains a concern. Here, we examined whether cultural engagement (including going to cinemas, museums, galleries, exhibitions, theatre, concerts, or opera) predicts hospital-treated mental disorders in 8,274 adults aged 50 years or older from the English Longitudinal Study of Ageing. Participant records were linked to ICD-10 diagnoses in Hospital Episode Statistics and mortality records with follow-up of up to 20 years. In fully adjusted Cox models accounting for sociodemographic, lifestyle, and social factors and multiple testing, frequent cultural engagement was associated with lower risk of any mental disorders (HR 0.71, 95% CI 0.62-0.82, FDR adjusted P value<0.001), dementia (0.71, 0.56-0.89, FDR adjusted P value=0.010), substance misuse (0.75, 0.59-0.95,FDR adjusted P value=0.040), and mood disorders (0.73, 0.56-0.95, FDR adjusted P value=0.044), but not neurotic disorders. Associations persisted after excluding early incident cases and adjusting for baseline depressive symptoms and cognition, and showed robustness to unmeasured confounders. To further probe causality, eye disease, ear disease, and traumatic brain injury, which share similar socio-demographic profiles to mental disorders, were prespecified as negative control outcomes. Cultural engagement was not associated with any negative control outcomes. These findings provide triangulated statistical data to suggest that cultural engagement is associated with reduced risk of several clinically recorded mental disorders and support further testing of cultural engagement as a population mental health strategy.

Suicidality phenotypes, including suicidal ideation (SI), non-fatal suicide attempt (SA), and suicide death (SD), are heritable and exhibit both shared and phenotype-specific genetic influences. Using genomic structural equation modelling, we estimated the shared genetic architecture across GWAS of SI (176,147 cases, 1,010,300 controls), SA (53,919 cases, 1,063,988 controls), and SD (7,584 cases, 652,070 controls) and conducted a multivariate GWAS of a latent suicidality factor capturing their shared liability. This analysis identified 36 genome-wide significant loci, including seven not previously reported in any suicidality GWAS. Follow-up analyses identified residual genetic variance specific to each phenotype, including three SD-specific genomic risk loci. Conditioning suicidality phenotypes on genetic liability to psychiatric disorders revealed significant residual genetic variance across SI, SA, SD, and the suicidality common factor. Together, these results suggest that suicidality reflects both shared genetic liability and phenotype-specific contributions.

Anxiety has been extensively studied in relation to memory, yet its dynamic association with spatial episodic memory in naturalistic clinical settings remains largely unexplored. We developed an anxiety-spatial-memory EMA protocol (asm-EMA) and deployed it in 30 epilepsy patients undergoing inpatient EEG monitoring, delivering combined momentary anxiety ratings and a validated spatial memory task pseudo-randomly every 90-150 minutes across multiple days. Subject-level asm-EMA means and session-to-session variability both correlated significantly with standard neuropsychological assessments, supporting the clinical validity of our design. Elevated within-person STAI-6 was selectively associated with faster retrieval responses, yet spatial memory accuracy was independent of all three anxiety measures, suggesting a shift in response strategy rather than memory impairment. Within-day anxiety showed short-term carryover between consecutive sessions, with little persistence beyond the next session. The asm-EMA protocol provides a feasible, autonomous framework for capturing moment-to-moment anxiety-memory dynamics in naturalistic settings.