Biochemical and brain heterogeneity characterizes psychiatric and non-psychiatric illness Authors: Maria A. Di Biase, William R. Reay, Hadis Jameei, Yuanzhe Liu, Ye E. Tian, Elysha Ringin, Susan Rossell, James A. Karantonis, Tamsyn E. Van Rheenen, Junhao Wen, Christos Pantelis, Andrew Zalesky Published: 2026 ABSTRACT Biological heterogeneity presents a formidable challenge in understanding psychiatric illness, yet it remains underexplored at scale or in a pan-disease setting. We characterized biochemical and brain heterogeneity in 445,374 individuals, considering 9 psychiatric disorders and 30 non-psychiatric medical conditions. Psychiatric and non-psychiatric illnesses were characterized by greater biological heterogeneity than predicted by healthy variation. Notably, heterogeneity in psychiatric disorders was comparable to non-psychiatric illnesses and often diverged from diagnostic classification. Circulating glucose and glycated hemoglobin levels were the most heterogeneous traits in 65% of diagnoses, suggesting that glucose dysregulation is pervasive yet variable across diseases. Diagnosis-specific patterns also surfaced—e.g., in schizophrenia, developmental brain markers ranked among the most heterogeneous traits. These characterizations of biochemical and brain heterogeneity lay the groundwork for biological focal points to aid precision psychiatry research. To support this effort, we provide an interactive online resource for exploring and downloading illness-specific heterogeneity measures. Introduction Biological heterogeneity is a defining but poorly understood feature of psychiatric disorders1. Individuals diagnosed with the same condition can differ markedly in brain structure2,3, biochemical profiles4, and underlying etiology5,6—challenging treatment and our basic understanding of psychopathology7. Traditional case-control studies have emphasized mean group differences in biological traits; however, these rarely yield clinically penetrant biomarkers. Limited translational success has shifted attention toward characterizing variability within diagnostic categories to expose biological patterns of clinical and/or biological relevance—e.g., diagnostic and prognostic subgroups1,8,9. Analogous patterns in other chronic disease settings offer important insights. For instance, in chronic kidney disease, serum creatinine clearly differentiates cases from controls. However, cystatin C—despite greater variability among patients—better predicts outcomes10, highlighting the importance of characterizing biological heterogeneity. To determine whether observed heterogeneity in psychiatric disorders reflects illness-related variation, comparisons with chronic non-psychiatric conditions (where pathological mechanisms are more well established) as well as healthy individuals, offer useful benchmarks to anchor expectations for biological variability. We present a comprehensive quantification of biochemical and brain heterogeneity across psychiatric and chronic non-psychiatric illnesses using large-scale biochemical and neuroimaging datasets. We compare the degree of heterogeneity across illness categories alongside the magnitude of mean group differences. We examine whether genetic factors, as captured through polygenic scores11 for each biological trait, help explain these illness-specific heterogeneity profiles. We further test whether biological heterogeneity conforms to hierarchical models of psychopathology—for instance, whether broader psychiatric categories exhibit greater biological heterogeneity compared to specific disorders or subtypes within a disorder. Characterizing biological heterogeneity is critical as the field endeavors to advance toward precision medicine. Results Illness-related heterogeneity exceeds normative variation Figure 1 provides an overview of our study design. We utilized neuroimaging and urine- and blood-derived phenotypes (Fig. 1a) acquired from 445,374 individuals (mean[SD]; range age, 56.94[8.1]; 37–74 years). Individuals were grouped into 40 categories: 9 psychiatric disorders, 30 non-psychiatric chronic medical conditions, and a healthy comparison group (Supplementary Tables 1–4). Across these groups, the mean sample size for biochemical phenotypes was 23,207 ± 24,756 (range = 773–98,279), and for brain phenotypes was 1872 ± 2420 (range = 88–11,170).Fig. 1Study overview.a A range of phenotypes were examined, including 60 biochemical phenotypes grouped into 11 panels and 5 brain magnetic resonance imaging (MRI)-derived phenotypes. b For each phenotype, normative ranges of variation as a function of age and sex were established in a healthy group (N = 59,105 for biochemical and N = 11,170 for brain phenotypes). Deviations from the normative median were computed for each individual diagnosed with one or more of the illnesses studied. Heterogeneity was computed for each illness category, defined as the standard deviation of individual deviation scores (see the section “Sensitivity analyses” for an alternative heterogeneity measure). c Biological heterogeneity was examined after adjusting for the genetic prediction of biological phenotypes (single nucleotide polymorphism [SNP]-based polygenic scores); and d across different resolutions of psychiatric nosology. This figure was created with BioRender.com (https://BioRender.com/6by4qui). Study overview. a A range of phenotypes were examined, including 60 biochemical phenotypes grouped into 11 panels and 5 brain magnetic resonance imaging (MRI)-derived phenotypes. b For each phenotype, normative ranges of variation as a function of age and sex were established in a healthy group (N = 59,105 for biochemical and N = 11,170 for brain phenotypes). Deviations from the normative median were computed for each individual diagnosed with one or more of the illnesses studied. Heterogeneity was computed for each illness category, defined as the standard deviation of individual deviation scores (see the section “Sensitivity analyses” for an alternative heterogeneity measure). c Biological heterogeneity was examined after adjusting for the genetic prediction of biological phenotypes (single nucleotide polymorphism [SNP]-based polygenic scores); and d across different resolutions of psychiatric nosology. This figure was created with BioRender.com (https://BioRender.com/6by4qui). Heterogeneity was quantified for 60 biochemical phenotypes grouped into 11 biomarker panels (see Supplementary Table 5 for trait-to-panel annotation) and five brain phenotypes (Fig. 1a). Normative models12 were trained on the healthy individuals to establish age- and sex-specific reference charts (Fig. 1b). These models were then applied to all individuals, yielding z-scores that reflect individual deviation from normative phenotypic values. For each diagnostic group, we calculated the standard deviation of these z-scores to index heterogeneity (Fig. 1b; see the section “Sensitivity analyses” for an alternative heterogeneity measure). Our goal was to identify diagnostic categories and biomarkers with variability exceeding normative levels. Illness-specific heterogeneity estimates are presented in Supplementary Tables 6 and 7 and can be visualized through an interactive online resource, which allows users to filter results by diagnoses and biomarker, and to explore individual z-score distributions within each illness group. Secondary analyses examined (1) the extent to which polygenic biomarker scores account for empirically measured biomarker variability (Fig. 1c) and heterogeneity levels across different resolutions of psychiatric diagnoses as defined by the International Classification of Diseases (ICD-10) classification system (Fig. 1d). Biochemical phenotypes showed consistently elevated heterogeneity across all illness groups compared to healthy controls (mean[SD] t = 8.04[2.7]; Cohen’s d = 1.5[.50], FDRp < 0.05), with variability increases ranging from 3% to 25% (SD = 0.86–1.66; Fig. 2b). Brain phenotypes revealed a similar pattern, with 70% of illness groups showing increased heterogeneity and 40% reaching significance (mean[SD] t = 4.59[1.21]; mean[SD] Cohen’s d = 2.9[.76], FDRp < 0.05). Heterogeneity in brain measures was 2–20% higher than norms (SD = 0.87–1.35; Fig. 2b). Biochemical traits exhibited greater heterogeneity compared to brain traits across most illness categories. However, in neurobiological disorders such as multiple sclerosis, schizophrenia, and bipolar disorder, brain traits displayed greater variability than biochemical traits (Fig. 2b and c).Fig. 2Illness-specific heterogeneity in common biochemical and brain phenotypes.a Normative ranges of variation, used to define heterogeneity (see Fig. 1), are shown for an example biochemical and brain trait. b Diagnostic categories are ordered by mean heterogeneity estimates, averaged across 11 biochemical panels and 5 brain phenotypes. Heterogeneity equals one in the healthy comparison group (dashed line). Filled circles indicate diagnostic groups with a significant difference in heterogeneity relative to the healthy group after FDR correction. c Heterogeneity estimates for select diagnoses (intended solely to aid visualization) are shown, for example, biochemical panels and brain traits (Supplementary Tables 6 and 7 contain complete heterogeneity data). d Violin plots display the distribution of average heterogeneity estimates across psychiatric and non-psychiatric diagnostic groups for biochemical (teal; n = 11) and brain (brown; n = 5) phenotypes. Boxplots overlay the median (central line), interquartile range (box edges) and data within 1.5 times the interquartile range (whiskers). Mean heterogeneity was compared between psychiatric disorders and each non-psychiatric category using GLMs. Analysis was performed separately for each biological domain (biochemical/brain) using phenotype-level heterogeneity values averaged within diagnoses as observations. P-values were adjusted for multiple comparisons across the six psychiatric-versus-category contrasts using the Benjamini–Hochberg FDR procedure (q = 0.05). CKD chronic kidney disease, GAD generalized anxiety disorder, GORD gastroesophageal reflux disease, OCD obsessive compulsive disorder, IBD inflammatory bowel disease. Icons created with BioRender.com (https://BioRender.com/4mya26j). Illness-specific heterogeneity in common biochemical and brain phenotypes. a Normative ranges of variation, used to define heterogeneity (see Fig. 1), are shown for an example biochemical and brain trait. b Diagnostic categories are ordered by mean heterogeneity estimates, averaged across 11 biochemical panels and 5 brain phenotypes. Heterogeneity equals one in the healthy comparison group (dashed line). Filled circles indicate diagnostic groups with a significant difference in heterogeneity relative to the healthy group after FDR correction. c Heterogeneity estimates for select diagnoses (intended solely to aid visualization) are shown, for example, biochemical panels and brain traits (Supplementary Tables 6 and 7 contain complete heterogeneity data). d Violin plots display the distribution of average heterogeneity estimates across psychiatric and non-psychiatric diagnostic groups for biochemical (teal; n = 11) and brain (brown; n = 5) phenotypes. Boxplots overlay the median (central line), interquartile range (box edges) and data within 1.5 times the interquartile range (whiskers). Mean heterogeneity was compared between psychiatric disorders and each non-psychiatric category using GLMs. Analysis was performed separately for each biological domain (biochemical/brain) using phenotype-level heterogeneity values averaged within diagnoses as observations. P-values were adjusted for multiple comparisons across the six psychiatric-versus-category contrasts using the Benjamini–Hochberg FDR procedure (q = 0.05). CKD chronic kidney disease, GAD generalized anxiety disorder, GORD gastroesophageal reflux disease, OCD obsessive compulsive disorder, IBD inflammatory bowel disease. Icons created with BioRender.com (https://BioRender.com/4mya26j). To assess whether these patterns differed by sex, heterogeneity levels were examined separately in males and females using self-reported sex. Biochemical heterogeneity showed a strong correspondence between sexes (r = 0.85, p < 0001), whereas brain heterogeneity showed more sex-specific divergence (r = 0.68, p < 0001), with females exhibiting slightly higher variability for several neurological and metabolic conditions. Sex-stratified results, including ranked heterogeneity profiles (Supplementary Fig. 1), are provided in the Supplementary Materials. Of the psychiatric disorders, schizophrenia exhibited the greatest variance across both biochemical and brain phenotypes, with biochemical heterogeneity 15% higher, on average, than the normative reference (SD: 1.06–1.22, where SD = 1 reflects normal variation, SD > 1 indicates variation that exceeds norms) and brain heterogeneity averaging 20% higher (SD: 1.02–1.36). Bipolar disorder followed closely across both measurement types (biochemical SD: 1.02–1.2; brain SD: 0.92–1.20). All other psychiatric disorders showed lower but still elevated heterogeneity in the biochemical measures, including anorexia nervosa (SD: 1.05–1.12), obsessive-compulsive disorder (OCD; SD: 1–1.11), depression (biochemical SD: 1–1.08), social anxiety/phobia (SD: 0.99–1.08), generalized anxiety disorder (GAD) (SD: 0.99–1.07), panic attacks (SD: 0.98–1.06), and bulimia nervosa (SD: 0.92–1.13). Contrary to our expectation, heterogeneity was not significantly greater within psychiatric disorders relative to all non-psychiatric illnesses for either biochemical (t = −1.28, p > 0.05) or brain phenotypes (t = -0.24, p > 0.05) or to broad illness categories, including neurological, cardiovascular, gastrointestinal, musculoskeletal, and respiratory diseases (refer to Fig. 2d for statistics and Supplementary Table 8 for illness-to-category annotations). Heterogeneity is distinct from group-mean differences Across both biochemical and brain domains, heterogeneity profiles provided distinct insights beyond group-mean differences (Fig. 3a) and remained largely preserved after adjusting for case-control effects (biochemical: r = 0.94, p < 0.0001; brain: r = 0.98, p < 0.0001). For example, the inflammation panel (serum C-reactive protein, CRP) showed the largest effect sizes for case-control differences in 38 out of 40 illness categories (95%), including all psychiatric diagnoses, whereas the glycemic panel, comprising glucose and glycated hemoglobin (HbA1C), was the most heterogeneous panel in 26 out of 40 (65%) illness categories, including 3 out of 9 psychiatric diagnoses (33%).Fig. 3Biomarker heterogeneity is distinct from group-mean differences and varies by disease mechanism.a Biochemical panels (left) and brain traits (right) exhibiting the greatest heterogeneity and mean group difference for each diagnosis. Line thickness represents the degree of heterogeneity or effect size for mean differences, categorized into low, medium, and high. Bold text indicates psychiatric disorders. b Radial plots depict heterogeneity profiles for selected diagnoses. CKD chronic kidney disease, GAD generalized anxiety disorder, O CDobsessive-compulsive disorder, FA fractional anisotropy. Biomarker heterogeneity is distinct from group-mean differences and varies by disease mechanism. a Biochemical panels (left) and brain traits (right) exhibiting the greatest heterogeneity and mean group difference for each diagnosis. Line thickness represents the degree of heterogeneity or effect size for mean differences, categorized into low, medium, and high. Bold text indicates psychiatric disorders. b Radial plots depict heterogeneity profiles for selected diagnoses. CKD chronic kidney disease, GAD generalized anxiety disorder, O CDobsessive-compulsive disorder, FA fractional anisotropy. For non-psychiatric illnesses, although different markers contributed to mean differences and heterogeneity estimates, the most pronounced effects were observed in systems and biomarkers directly linked to their underlying disease processes (examples shown in Fig. 3b). For instance, in multiple sclerosis, subcortical volume emerged as the most heterogeneous trait, aligning with its established role in predicting disability progression13,14. In contrast, fractional anisotropy measured in white matter—the primary tissue affected in multiple sclerosis—showed the largest case-control difference but relatively less variability across patients. This distinction illustrates how case-control differences can point to central disease mechanisms, whereas heterogeneity captures inter-individual variation within patient groups, often along dimensions with prognostic or stratification value. Together, these complementary metrics underscore the importance of considering both mean effects and individual variability in biomarker-driven disease characterization. Heterogeneity-case-control discrepancies were also observed among psychiatric illnesses. For example, traits showing the greatest case-control differences included blood CRP and cortical thickness and volume in schizophrenia and bipolar disorder, respectively—findings consistent with previous case-control studies15,16. However, the most heterogeneous traits differed from those showing the largest mean differences, with glycemic and kidney biochemical markers, as well as the developmental brain marker surface area, emerging as the most variable in both schizophrenia and bipolar disorder (Fig. 3b). Thus, while group-level mean differences highlight established markers of disease burden, inter-individual variability is more evident in metabolic and neurodevelopmental markers. Group means align across biological domains; heterogeneity does not To test whether illness-related patterns of heterogeneity and case-control differences are conserved across biological domains, we examined the correspondence between biochemical and brain trait profiles across diagnostic categories. Each data point represented a single illness, defined by both its level of heterogeneity and the magnitude of case-control differences across traits. After controlling for heterogeneity, case-control differences in biochemical traits were significantly correlated with case-control differences in brain traits (r = 0.49, p = 0.002), suggesting that group-level differences across peripheral and brain markers may reflect shared aspects of disease severity. In contrast, heterogeneity estimates in biochemical traits were not significantly correlated with those in brain traits after adjusting for case-control differences (r = 0.24, p > 0.05). Thus, heterogeneity in these traits captures distinct, domain-specific patterns of biological variation. Sensitivity analyses Heterogeneity profiles across the biochemical traits were used to probe the influence of several covariates on our observed findings: illness comorbidities (psychiatric and non-psychiatric), variation in sample size among diagnostic groups, the choice of heterogeneity metric, demographic (age and sex) characteristics and several biochemical confounds (BMI, smoking status, medication exposure and fasting duration; refer to Supplementary Materials for details). Due to limited sample sizes (n < 100) after excluding cases with comorbid conditions, six diagnostic groups were omitted from the comorbidity-related sensitivity analysis: Crohn’s disease, inflammatory bowel disease, asthma, obsessive compulsive disorder, social anxiety/phobia and bulimia nervosa. Across the remaining 34 groups, diagnosis-level heterogeneity estimates (one value per illness) were consistent before and after removal of individuals with comorbidities (r = 0.80, p < 0.0001). In addition, the relative ranking of heterogeneity across illnesses was largely preserved under multiple sensitivity analyses, including: (i) restricting each diagnostic group to 1000 randomly selected individuals (r = 0.99, p < 0.0001; bulimia nervosa excluded due to insufficient sample size), (ii) estimating heterogeneity using the interquartile range instead of the standard deviation (r = 0.98, p < 0.0001), (iii) regressing out mean age and sex ratio across diagnostic groups (r = 0.85, p < 0.0001), and (iv) adjusting for body mass index (BMI), smoking status, medication exposure, and fasting duration (r = 0.67, p < 0.0001). In the latter analysis, concordance was lower compared to the other sensitivity analyses, an effect driven predominantly by BMI, as heterogeneity profiles remained highly concordant after adjusting for smoking status, medication exposure and fasting duration alone (r = 0.98). Diagnoses showing the largest BMI-related reductions in heterogeneity were diabetes (−9.9%), gout (−8.6%), glomerular disease (−6.8%), coronary artery disease (−6.4%), and sleep apnea (−6.4%). Notably, these illnesses are linked to metabolic, or obesity-related pathophysiology, indicating that BMI may reflect core disease biology rather than extraneous confounding. Consequently, controlling for BMI in these conditions may remove substantive illness-related biological variability, not merely background noise. Modest contribution of polygenic scores to heterogeneity In all the above analyses, individuals in a disease category were benchmarked to the variability seen in healthy comparison individuals of the same age and sex. However, factors beyond age and sex may account for biomarker variability, including environmental and genetic effects. To explore the genetic contribution to heterogeneity, we computed polygenic scores (PGS) for each biochemical and brain phenotype. We regressed each subject’s z-score (i.e., deviations from the normative median for each phenotype) on polygenic scores and recalculated heterogeneity using the resulting residuals, allowing us to assess the extent to which genetic scores contribute to heterogeneity (Fig. 4).Fig. 4Illness-specific heterogeneity profiles remain largely unaffected by polygenic architecture.For each subject with available data, PGSs were computed for each biochemical/brain phenotype. A general linear model was used to regress out the effect of PGS from each subject’s z-score (deviations from a normative age- and sex-adjusted median). a Distribution of heterogeneity (measured as SD) across biochemical (n = 11) and brain phenotypes (n = 2), before and after PGS adjustment. Significant reductions in heterogeneity following adjustment were observed in the healthy control (p = 0.0002) and migraine group (p = 0.002). b Mean percentage change in heterogeneity across diagnoses (n = 40) following PGS adjustment, shown for biochemical panels (top row: heatmap) and brain measures (bottom row: boxplot). Each cell in the heatmap represents the average percent reduction; bubble size overlays indicate the spread of the pairwise distribution (SD across diagnoses, range = 0–0.67). For all boxplots (a and b, bottom row), data are presented as median values (center line); box edges represent the interquartile range (IQR; 25th and 75th percentiles); and whiskers extend to 1.5 times the interquartile range (minima/maxima). For brain phenotypes in a, whiskers are absent as n = 2. Asterisks indicate significant differences in mean heterogeneity or percentage reduction based on FDR-corrected comparisons: p < 0.01 (**) and p < 0.001 (***). GAD generalized anxiety disorder, OCD obsessive compulsive disorder, GORD gastroesophageal reflux disease, IBD inflammatory bowel disease, CKD chronic kidney disease. Icons created with BioRender.com (https://BioRender.com/4mya26j). Illness-specific heterogeneity profiles remain largely unaffected by polygenic architecture. For each subject with available data, PGSs were computed for each biochemical/brain phenotype. A general linear model was used to regress out the effect of PGS from each subject’s z-score (deviations from a normative age- and sex-adjusted median). a Distribution of heterogeneity (measured as SD) across biochemical (n = 11) and brain phenotypes (n = 2), before and after PGS adjustment. Significant reductions in heterogeneity following adjustment were observed in the healthy control (p = 0.0002) and migraine group (p = 0.002). b Mean percentage change in heterogeneity across diagnoses (n = 40) following PGS adjustment, shown for biochemical panels (top row: heatmap) and brain measures (bottom row: boxplot). Each cell in the heatmap represents the average percent reduction; bubble size overlays indicate the spread of the pairwise distribution (SD across diagnoses, range = 0–0.67). For all boxplots (a and b, bottom row), data are presented as median values (center line); box edges represent the interquartile range (IQR; 25th and 75th percentiles); and whiskers extend to 1.5 times the interquartile range (minima/maxima). For brain phenotypes in a, whiskers are absent as n = 2. Asterisks indicate significant differences in mean heterogeneity or percentage reduction based on FDR-corrected comparisons: p < 0.01 (**) and p < 0.001 (***). GAD generalized anxiety disorder, OCD obsessive compulsive disorder, GORD gastroesophageal reflux disease, IBD inflammatory bowel disease, CKD chronic kidney disease. Icons created with BioRender.com (https://BioRender.com/4mya26j). The diagnostic ranking of heterogeneity remained consistent after adjusting for polygenic effects. Across all diagnoses and brain/biochemical phenotypes, heterogeneity decreased modestly following the removal of PGS-related variability (Fig. 4a), with percentage reductions ranging from 1.33% to 2.89% for biochemical phenotypes and 0.09–0.64% for brain phenotypes. A significant difference in heterogeneity pre- and post-PGS adjustment was observed only in the healthy control group (t = –4.50, Cohen’s d = 1.90, p < 0.001) and, to a lesser extent, the migraine group (t = 3.55, Cohen’s d = 1.50, p < 0.01). This may reflect that, in most illnesses, non-genetic sources of variability dilute the observable effects of polygenic risk. Despite these small effects, significant differences in mean percentage change following PGS adjustment were observed (Fig. 4b) between biochemical panels (mean[SD] across significant effects: t = 19.22[10.33], Cohen’s d = 4.3[2.3], p < 0.001) and between brain traits (t = –5.09, Cohen’s d = 1.14, p < 0.0001). These findings indicate that the polygenic scores used in this study capture only a small, but variable, fraction of heterogeneity across diagnoses and traits. This likely reflects limited PGS power rather than a lack of genetic influence on these measures. The remaining variability likely reflects contributions from additional genetic, environmental, and disease-specific factors, among others. Psychiatric heterogeneity defies ICD-10 diagnostic hierarchies Biological characteristics often misalign with psychiatric diagnostic categories17. Here, we explored the extent to which heterogeneity within psychiatric populations conforms to diagnostic structures prescribed by the ICD-10 (refer to Supplementary Materials for details)18. We hypothesized that a broad psychiatric population would exhibit greater biological variability compared to specific groups, such as individuals diagnosed with a particular disorder (e.g., bipolar disorder) or a subtype of that disorder (e.g., bipolar disorder with psychotic features). In general, we found that heterogeneity across hierarchical taxonomies of psychopathology deviates from diagnostic structures (Fig. 5a and Supplementary Fig. 2). In contrast to ICD-10 demarcations, a set of all individuals with any psychiatric illness often displayed less heterogeneity compared to discrete subtypes at lower levels of the hierarchy, both in biochemical and brain domains. This was evidenced by broader distributions in subtypes, relative to an integrated psychiatric population. In terms of diagnostic subtypes, greater heterogeneity often aligned more closely with symptomatic complexity rather than with diagnostic categories. For example, bipolar disorder showed greater heterogeneity than depression and bipolar disorder with psychotic features displayed greater heterogeneity relative to those without psychotic features across both biochemical (Fig. 5a) and brain (Supplementary Fig. 1) traits. In contrast, heterogeneity was comparable between disorders such as schizophrenia and schizoaffective disorder, consistent with previous findings suggesting that certain diagnostic boundaries are less biologically and genetically distinct than others19.Fig. 5Heterogeneity at multiple diagnostic resolutions.Heterogeneity estimates averaged across biochemical traits for psychiatric (a) and non-psychiatric (b) illnesses. Heterogeneity was estimated based on individual deviations from a healthy control normative range and defined as the SD within each category. Teal nodes represent observed mean heterogeneity (measured as the SD across n = 11 biochemical phenotypes), while gray nodes represent the null expectation. Shaded error bars (teal) indicate 95% confidence intervals derived via bootstrap resampling (1000 iterations) within each category. Null conditions (light gray) and their corresponding uncertainty intervals (5th–95th percentile) were generated by re-estimating heterogeneity in equally sized subpopulations drawn from higher-order diagnostic structures. Diagnostic Categories are based on ICD-10 classifications. Lower-order groups are subsets of higher-order groups but do not necessarily sum to the total; for instance, of 2646 individuals with bipolar disorder (BPAD), only those with sufficient clinical data are subclassified into BPAD with (n = 216) or without (n = 207) psychosis. Not all individuals were eligible for subclassification due to insufficient clinical data or failure to meet specific subcategory criteria. Dep depression, BPAD bipolar disorder, w with, w/o without, comp. complications. Heterogeneity at multiple diagnostic resolutions. Heterogeneity estimates averaged across biochemical traits for psychiatric (a) and non-psychiatric (b) illnesses. Heterogeneity was estimated based on individual deviations from a healthy control normative range and defined as the SD within each category. Teal nodes represent observed mean heterogeneity (measured as the SD across n = 11 biochemical phenotypes), while gray nodes represent the null expectation. Shaded error bars (teal) indicate 95% confidence intervals derived via bootstrap resampling (1000 iterations) within each category. Null conditions (light gray) and their corresponding uncertainty intervals (5th–95th percentile) were generated by re-estimating heterogeneity in equally sized subpopulations drawn from higher-order diagnostic structures. Diagnostic Categories are based on ICD-10 classifications. Lower-order groups are subsets of higher-order groups but do not necessarily sum to the total; for instance, of 2646 individuals with bipolar disorder (BPAD), only those with sufficient clinical data are subclassified into BPAD with (n = 216) or without (n = 207) psychosis. Not all individuals were eligible for subclassification due to insufficient clinical data or failure to meet specific subcategory criteria. Dep depression, BPAD bipolar disorder, w with, w/o without, comp. complications. To contextualize these findings, we examined heterogeneity levels in a non-psychiatric illness, namely, diabetes mellitus (Fig. 5b), where biochemical markers are used in clinical practice20. Here, we characterized heterogeneity in markers relevant to diagnosing and monitoring diabetes, including HbA1c and glucose levels. Similar to psychiatric populations, a broad category (diabetes) was less heterogeneous than a specific subtype (Type 1 diabetes). As shown in Fig. 5b, biochemical heterogeneity was more pronounced in Type 1 than in Type 2 diabetes, possibly due to greater variability in insulin administration and glucose regulation21. This was particularly evident in subgroups with severe complications, such as ketoacidosis and coma, where fluctuations in blood glucose levels may contribute to heightened variability. These findings suggest that variability is shaped by a combination of phenomenological complexity, disease severity, and treatment-related factors—and less by where a condition sits within a diagnostic hierarchy, across both psychiatric and non-psychiatric illnesses. Discussion Our findings reveal widespread inter-individual variability in biochemical and brain traits across psychiatric and non-psychiatric illnesses. The pattern and structure of this heterogeneity vary systematically across diagnoses and traits and often align with disease-relevant biological systems. Phenotypic heterogeneity does not systematically differ between psychiatric and non-psychiatric conditions, which challenges the assumption that psychiatric disorders are uniquely heterogeneous. The ranking of heterogeneity across illnesses remains largely stable after accounting for key demographic, statistical, illness-related and genetic factors. However, the comparatively lower concordance observed after BMI adjustment suggests that metabolic burden contributes to illness-associated variability. Together, these results underscore the need to move beyond case-control frameworks toward heterogeneity-focused models that prioritize highly variable phenotypes as targets for precision-medicine research to inform diagnostic, prognostic, and therapeutic strategies. In non-psychiatric conditions, heterogeneity tracked with known pathophysiology—for example, AST levels in cirrhosis22. In psychiatric disorders, our findings point to potential leads of interest: biochemical panels for distinguishing patient subgroups or illness states. For example, pronounced variability was seen within glycemic and kidney panels for schizophrenia and bipolar disorder, supporting evidence for metabolic dysfunction23–25 and elevated risk of diabetes and kidney disease26. These markers may not distinguish psychiatric patients from healthy controls, but their heterogeneity may be relevant for identifying clinical subtypes23,24,27 or treatment responses28,29. Serum marker, CRP, showed the largest mean difference across nearly all psychiatric and non-psychiatric conditions, while glycemic markers showed the highest within-group heterogeneity. This may reflect distinct physiological dynamics. Though traditionally considered an inflammatory marker, CRP also plays roles in pathogen response and, in its pentameric form, exerts anti-inflammatory effects30. Considering these varied functions, it follows that dysregulated CRP is a well-documented observation in many illness manifestations—ranging from infectious and chronic inflammatory diseases to cardiovascular diseases, metabolic syndrome and psychiatric illness31–33. In contrast, glycemic markers showed greater heterogeneity, in part due to their rapid oscillations—occurring within minutes to hours—in response to various physiological demands34, compared to slower changes in CRP levels35. Longitudinal and continuous-glucose-monitoring studies also show substantial short-term variability in glycemic measures36, supporting this interpretation. Although our data are cross-sectional and cannot capture these fluctuations directly, elevated glycemic variability may reflect glucose homeostasis linked to systemic inflammation, stress, medication effects, poor nutrition, reduced activity, or hormonal changes37–41. Importantly, disentangling cause from consequence remains challenging: glycemic instability may reflect core pathophysiological processes, downstream systemic effects of illness, or bidirectional interactions between metabolic, physiological and environmental states. The divergence between mean shifts and variability highlights how different traits reflect illness—either through sustained level changes or instability in regulation. While group-level mean differences were similar across biochemical and brain domains, heterogeneity patterns diverged. Interpreted at the level of biological systems, these findings point to a general “mean severity” factor embedded within diagnoses, whereby illness manifestations are evident across both peripheral and brain systems, regardless of the specific affliction or whether it primarily involves the central nervous system (e.g., psychiatric disorders) or peripheral organs (e.g., gout and cirrhosis). This convergence may reflect bi-directional brain–body interactions and/or shared pathological mechanisms influencing both biochemical and brain profiles, including genetic, epigenetic, neurotransmitter, metabolic, neuroplasticity, and environmental factors42. In contrast, heterogeneity levels differed across illnesses between brain and peripheral systems, potentially arising from differential sensitivity of biochemical and neuroanatomical measures to disease stage, progression, and individual responses. Biological heterogeneity is a ubiquitous theme with many layers of complexity, shaped by numerous biomedical and genetic factors as well as by how heterogeneity is operationalized. In the analysis and interpretation of heterogeneity, acknowledging its inherent dependency on context is essential43. Here, our reference point reflected a group of individuals without a diagnosis of a psychiatric and/or chronic medical condition. Alternative reference points may have exposed different pathological patterns of interest that potentially cross-cut traditional diagnostic boundaries. When examining heterogeneity across levels of diagnostic hierarchy, we found that the magnitude of biological variability did not map neatly onto ICD-10 classifications. Broader categories (e.g., “any psychiatric illness”) often exhibited less heterogeneity than more specific disorders or subtypes. A similar pattern emerged in diabetes mellitus. Rather than diagnostic breadth, greater heterogeneity was observed within diagnoses associated with more severe, multifaceted, or fluctuating symptom profiles. The recurrence of this pattern across distinct disease domains underscores the importance of interpreting heterogeneity in clinical context—where it may better index illness complexity—rather than within phenomenological diagnostic hierarchies. Our analysis focused on the standard deviation (and the interquartile range in a sensitivity analysis) to characterize biological heterogeneity—other distributional measures such as skewness, bimodality, and entropy may provide complementary insights into illness-specific variability profiles. Another limitation is that the UK Biobank imaging cohort is, on average, older than most clinically ascertained psychiatric samples. Although age-related effects are explicitly modeled and removed within the normative framework—and sensitivity analyses confirmed that heterogeneity profiles were robust to age residualization—this demographic skew may nevertheless limit direct generalization of absolute heterogeneity estimates to younger clinical populations. In terms of the genetic analyses, because the sample sizes of individuals with non-European genetic ancestry participants are limited in the UK Biobank, analyses were necessarily restricted to individuals with European genetic ancestry to ensure sufficient power, limiting generalizability to other genetic ancestries. Furthermore, estimates of genetic levels (Fig. 4) were constrained to brain traits where only a small portion of phenotypic variability is currently explained by PGS-predicted scores (~2–3% of out-of-sample variance explained, as reported in Grasby et al.44). This is despite relatively high SNP heritability (h2; ~25–35% for global cortical measures and ~5–35% for regional cortical measures). Increases in sample size for future GWAS will likely help to better reconcile individual-level predictive accuracy and population-level heritability estimates45. Finally, although polygenic scores influenced overall trait variability (see Supplementary Tables 9 and 10), their contribution to illness-specific heterogeneity was modest, suggesting that factors beyond PGSs—genetic and/or environmental—play a significant role in shaping illness-specific variability, which were not examined here. Despite these limitations, our characterization of biological heterogeneity in a pan-disease setting indicates that different illnesses can be distinguished by their degree of variability in biochemical and brain phenotypes. Our findings highlight specific biological features as promising targets for research aimed at predicting illness states and identifying biological subgroups with diagnostic, prognostic, or therapeutic relevance. This need is especially acute in psychiatric disorders, where clinical presentations are highly heterogeneous and robust biological markers remain scarce. A critical next step will be to apply heterogeneity-informed models to deeply phenotyped clinical cohorts with longitudinal measures of symptoms, cognition, and treatment response. Doing so would allow population-level heterogeneity signatures to be translated into clinically grounded biotyping frameworks and to test whether individuals with marked deviations from normative biological ranges cluster into reproducible subgroups with distinct symptom profiles, illness trajectories, or treatment responses. To facilitate this work, we provide an empirically grounded online resource summarizing the broad range of implicated biochemical panels and brain traits across diverse illness groups. Methods Participants Biochemical and genotype data were sourced from the UK Biobank (UKB; n = 445,374). Brain imaging data were sourced primarily from the UKB. To enhance the sample size of diagnostic-specific brain imaging analyses, UKB data were integrated with data from the Australian Schizophrenia Research Bank (ASRB; n = 270)46 and the Cognition Study in Bipolar Disorder (COGSBD; n = 61)47, respectively. Biochemical assessments were sourced from 445,374 individuals (mean [SD]; range age, 56.94 [8.1]; 37–74 years) participating in the UKB. T1-weighted magnetic resonance imaging (MRI)-derived brain phenotypes were derived from 39,042 individuals aged 18–95 years, and diffusion-weighted MRI brain images were available for 39,406 individuals (mean [SD]; range age, (64.2 [8.25]; 18–85 years) from all cohorts. Refer to Supplementary Tables 1–4 for consortia-specific sample sizes and demographic characteristics. Details pertaining to each cohort, as well as data integration procedures, are described in Supplementary Materials. Ethical approval was obtained as follows: for UKB, from the North West Multi-centre Research Ethics Committee; ASRB, Melbourne Health Human Research Committee (Project ID: 2010.250); COGSBD, the Alfred Human Research Ethics Committee (Project 3-16) and Swinburne University (2016/017). Written informed consent was obtained from all participants, according to the ethical standards of the institutional and/or national research committee of each site and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline. Clinical groups To investigate illness-specific signatures of biological heterogeneity, individuals were grouped into 40 categories (see Supplementary Tables 1–4). These comprised 9 psychiatric disorders—schizophrenia, bipolar disorder, depression, anorexia nervosa, bulimia, generalized anxiety disorder, obsessive compulsive disorder, social anxiety/phobia and panic attacks—30 non-psychiatric conditions, and a healthy control group. ICD and MHQ codes corresponding to each of the 39 clinical groups are listed in Data S1. A reference point of biological heterogeneity was defined by a healthy group, operationalized as individuals who had no lifetime history of any of the 39 diagnostic categories or any chronic medical condition before their biochemical/imaging assessment date, as ascertained through hospital records, primary care data, and self-report. Refer to Supplement Materials for detailed information regarding data sources and diagnostic grouping procedures. Biochemical and brain phenotypes Blood and urine sample assays were available for UKB participants. We utilized 60 biochemical traits, which passed quality control measures (Field IDs listed in Supplementary Table 5). Assessment details and procedures for processing biological samples are described elsewhere48. Biochemical traits were grouped into biomarker panels based on shared physiological roles and established clinical groupings to aid biological interpretation. Refer to Supplementary Materials for panel grouping procedures and Supplementary Table 5 for biochemical trait to panel annotation. Consistency across traits within biochemical panels was confirmed with the intraclass correlation coefficient (ICC), indexing the degree of absolute agreement among measurements (ICC range across biochemical panels = 0.76–0.94, p < 0.0001)49. The following brain gray matter measures were selected from T1-weighted MRI data: gray matter volume, cortical thickness, surface area and subcortical gray matter volume. Brain white matter microstructure was profiled using tract-specific measures of fractional anisotropy. Refer to Supplementary Materials for details regarding imaging processing, quality control, and phenotype extraction. Characterizing heterogeneity Generalized Additive Models for Location, Scale and Shape (GAMLSS) frameworks12 stratified by sex (and site for brain phenotypes) were applied to healthy individuals to establish reference curves as a function of age, using GAMLSS implemented in R v2021.09.1 (Build 372)50,51. The GAMLSS framework is a semi-parametric normative modeling framework that accounts for heteroscedasticity, non-Gaussian distributions, and nonlinear trajectories. GAMLSS fitting procedures are described in Supplementary Materials. A separate GAMLSS model was fitted to each phenotype (see GitHub for code and implementation details), as shown in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y={{SHASH}}(\mu,\sigma,\nu,\tau )$$\end{document}y=SHASH(μ,σ,ν,τ)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu={f}_{\mu }^{1}\left({{\rm{age}}}\times {{\rm{female}}}\right)+{f}_{\mu }^{2}\left({{\rm{age}}}\times {{\rm{male}}}\right)+{\beta }_{\mu }{z}_{{{\rm{site}}}\,({{\rm{for}}}\; {{\rm{brain}}}\; {{\rm{phenotypes}}}\; {{\rm{only}}})}$$\end{document}μ=fμ1age×female+fμ2age×male+βμzsite(forbrainphenotypesonly)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma={f}_{\sigma }\left({{\rm{age}}}\right)+{\beta }_{\sigma }{z}_{{{\rm{site}}}}({{\rm{for}}}\; {{\rm{brain}}}\; {{\rm{phenotypes}}}\; {{\rm{only}}}),\,\nu={\beta }_{\nu },\tau={\beta }_{\tau }$$\end{document}σ=fσage+βσzsite(forbrainphenotypesonly),ν=βν,τ=βτwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{SHASH}}}$$\end{document}SHASH denotes the Sinh-Arcsinh distribution—a flexible four-parameter distribution that generalizes the normal distribution by allowing for skewness and heavy or light tails52. The distribution was parameterized with four parameters: median/location (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document}μ), centile-based coefficient of variation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}σ), skewness (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\nu$$\end{document}ν) and kurtosis (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau$$\end{document}τ). For brain phenotypes, the location parameter also modeled the effect of several neuroimaging covariates (not shown above for brevity), including positions of center-of-gravity and table/coil, intensity scaling, and signal-to-noise and contrast-to-noise ratios (Supplementary Table 11 presents a complete list of neuroimaging confounds). For IDPs with more than three neuroimaging confounds, principal component analysis (PCA) was used to reduce dimensionality into three components that captured the most variance, as described in previous work53. Each individual with available phenotype data was positioned on these reference charts and a continuous measure of deviation (Δ) from the established normative range was expressed as a z-score for each individual, reflecting the extent of deviation from the median (50th centile). To characterize dispersion in the differences from the reference curve, heterogeneity was defined within diagnostic categories as the standard deviation (SD) of individual deviation scores, rather than raw phenotypic values. The interquartile range was used in sensitivity analyses. To ensure robustness, heterogeneity estimates in healthy controls were also evaluated using a 10-fold cross-validation framework. In each fold, z-scores were computed for the 10% of healthy controls that were left out of the GAMLSS model training. Importantly, heterogeneity estimates derived from cross-validation were identical to those obtained using the full healthy control sample, across five randomly selected biological phenotypes. This confirms that the normative modeling procedure yields stable and consistent estimates. Finally, normative models were trained separately within each cohort (UKB, ASRB, and COGSBD). Cohort-specific variances were pooled using sample-size weighting, and the square root of the pooled variance was taken to derive a combined standard deviation to minimize scanner-specific bias while preserving cohort-level structure. Polygenic scores for biochemical and brain phenotypes Quality control Quality control (QC) involved the selection of single-nucleotide polymorphisms (SNPs) and samples. For SNP QC, autosome variants were retained while SNPs failing any of the following criteria were excluded: (1) genotype call rate <0.95, (2) minor allele frequency (MAF) < 0.01, (3) Hardy–Weinberg equilibrium (HWE) test p-value < 10−7, (4) imputation information score below 0.8, (5) duplication, or (6) missingness. For sample QC, participants were retained if they were included in the genotype principal component (PC) analysis and were of white British ancestry. In addition, participants were excluded if they had (1) a genotype missing rate exceeding 0.05, (2) more than 10 third-degree relatives based on the kinship table provided by the UKB, (3) high heterozygosity rates, (4) sex chromosome aneuploidy, (5) a mismatch between reported sex and sex chromosome determination, or (6) withdrawn consent. All UKB genomic coordinates were available in NCBI Build GRCh37/Hg19. Genome-wide association study (GWAS) for biochemical traits GWAS analyses were conducted on cohorts comprising individuals who met genetic quality control criteria (see above) and were free from a lifetime diagnosis of the 39 specified clinical conditions and a cancer diagnosis prior to the biochemical assessment. For each biochemical trait, between 54,595 and 69,215 individuals met these criteria; however, only 34,216 participants had complete and overlapping biochemical data across all 60 markers. To standardize GWAS discovery sample sizes across traits and to retain a large, fully independent cohort for downstream polygenic scoring, we randomly selected a fixed sample of 40,000 individuals per biochemical trait. This ensured that (i) all traits were analyzed using the same discovery N, (ii) no individual contributed to both GWAS discovery and PGS estimation, and (iii) a sufficiently large non-overlapping target sample remained available for PGS analyses. Replicate GWASs using independent random draws yielded highly concordant SNP effect estimates (r > 0.98), indicating that this subsampling strategy did not distort genetic signal or bias effect estimates. GWAS was performed using PLINK version 2, which fitted a linear regression for marker-SNP associate testing (--glm) to obtain SNP summary statistics for each biochemical marker. We ran GWASs using two phenotype-specification strategies: (i) the raw biochemical trait as the outcome with covariate adjustment in the linear model, and (ii) a rank-based inverse normal transformation (INT) of residuals after regressing the trait on covariates. We implemented and compared both to balance robustness to non-normal trait distributions (via INT) against interpretability and scale preservation (raw trait), noting that our large sample size should, by the central limit theorem, mitigate modest violations of normality in OLS. In both strategies, we adjusted for sex, genotyping batch (UKB Data-Field 22000), and the first 15 genetic principal components (Data-Field 22009). We then characterized genetic architecture and polygenicity with an empirical-Bayes mixture approach that estimates the genome-wide proportion of non-zero SNP effects, as described in refs. 54,55. Polygenic scoring The biochemical summary statistics produced as described above were used to derive polygenic profiles for biochemical markers in all individuals who were not included in the discovery GWAS (n = 337,113). The sample size of individuals with both MRI and SNP data was too small to conduct a robust GWAS and still retain independent samples for analyzing genetic effects in brain phenotypes. Therefore, for brain phenotypes, including regional surface area and cortical thickness measures, GWAS summary statistics were obtained from the Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) consortium genetics working group, excluding UKB samples56. GWAS summary statistics for the remaining three MRI phenotypes used in the phenotypic analyses were not available from independent cohorts and were thus excluded in the PGS-adjusted tests. For each biochemical and brain phenotype, we applied a flexible approach for tuning polygenic scores that does not require an independent tuning sample and relies on relatively simple assumptions, making it well-suited for the diverse genetic architectures of the traits, as described elsewhere54. Briefly, this method employs probabilistic fine-mapping within independent linkage disequilibrium (LD) blocks across the genome, allowing for the estimation of variant-wise posterior probabilities of causality. This serves as a shrinkage approach to account for correlated effect sizes induced by LD but does not inherently require an independent genotyped sample for tuning. Approximate Bayes Factors (ABFs) were calculated for each LD block using a prior variance of 0.15, a standard assumption for quantitative traits, in R v2021.09.1. Genome-wide significant loci for each biochemical and brain phenotype were fine-mapped using ABFs to identify credible causal variants under the assumption of a single causal variant, also using a prior variance of 0.1557. Final polygenic profiles were generated using PLINK version 2 (--score), incorporating fine-mapped SNP weights for each phenotype. Polygenic scores were then adjusted for sex, batch number (Data-Field 22000), and the first 15 principal components of population structure (Data-Field 22009) before correcting for genetic effects in illness-specific heterogeneity profiles. Statistical analyses All statistical analyses were conducted in MATLAB R2024b. Comparing heterogeneity with healthy controls (Fig. 2) General linear models (GLMs) compared heterogeneity—defined as the SD of normative-model z-scores—between each diagnostic group and healthy controls. Analyses were run separately within each biological domain, using phenotype-specific heterogeneity values as observations (biochemical: 11 panels; brain: 5 measures). Two-sided tests were used. For multiple testing, each domain was treated as a separate multiplicity family: (i) 39 diagnosis–control contrasts for biochemical phenotypes and (ii) 39 diagnosis–control contrasts for brain phenotypes. Within each family, p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure55, controlling the FDR at q = 0.05. Comparing heterogeneity between psychiatric and non-psychiatric illnesses (Fig. 2) Non-psychiatric conditions were grouped into higher-order categories (Supplementary Table 8): all non-psychiatric, neurological, cardiovascular, gastrointestinal, musculoskeletal, and respiratory diseases. GLMs compared the mean heterogeneity between psychiatric disorders and each category, run separately for each biological domain (biochemical and brain). Observations were phenotype-level heterogeneity values averaged within the relevant diagnoses (11 biochemical panels; 5 brain measures). For each domain separately, we defined the multiplicity family as the six psychiatric–versus–non-psychiatric category contrasts, and p-values were adjusted within that family using the Benjamini–Hochberg FDR procedure (q = 0.05). Heterogeneity and mean group differences (Fig. 3) Independent t-tests (two-sided) assessed case–control differences in z-scores (derived from normative modeling; age/sex adjusted) between each diagnostic group and healthy controls, run separately for each biochemical panel and brain phenotype (sample sizes shown in Supplementary Tables 1–4). Cohen’s d quantified effect sizes. Heterogeneity was recomputed after adjusting for the case–control d effect per phenotype, and pre- and post-adjusted heterogeneity was correlated across diagnoses (n = 40 data points) using two-sided Spearman’s r. Correlations were computed at the domain level using, for each diagnosis, the average heterogeneity across biochemical panels and, separately, across brain phenotypes. To assess cross-system alignment, two-sided Spearman partial correlations were performed across diagnoses (n = 40 datapoints). First, we correlated (i) biochemical and brain case–control effect sizes, controlling for biochemical and brain heterogeneity, and (ii) biochemical and brain heterogeneity (mean SD of deviation scores), controlling for biochemical and brain case-control effect sizes. Genetic effects on illness-specific heterogeneity profiles (Fig. 4) Deviation scores (normative-model z-scores) were residualized on each participant’s PGS, and heterogeneity was recomputed as the SD of the residualized scores. For biochemical phenotypes, within-diagnosis comparisons of pre- vs. post-PGS heterogeneity were performed using a GLM with heterogeneity as the outcome and adjustment status (pre/post-PGS adjustment) as the predictor; the 11 biochemical panels served as observations/data points. No inferential test was conducted for brain phenotypes (only two measures). A second set of GLMs examined whether the average percentage change in heterogeneity (pre vs. post PGS adjustment) differed systematically between biochemical panels and brain traits, respectively. In these analyses, data points reflect the diagnoses (n = 40), the dependent variable was the percentage change in heterogeneity, and the independent variable was phenotype (biochemical panel or brain trait). Comparisons—across diagnoses (39 contrasts) and among biochemical phenotypes (55 unique pairwise correlations)—were FDR-corrected (q = 0.05); no correction was applied in the brain domain because only a single pairwise contrast (between the 2 brain phenotypes) was tested. Heterogeneity within diagnostic constructs (Fig. 5) To analyze heterogeneity across hierarchical frameworks for classifying mental disorders (e.g., any psychiatric disorder > mood disorders > depression or bipolar disorder), heterogeneity was recomputed across the various resolutions of psychiatric diagnoses. To assess whether observed heterogeneity within each diagnostic subgroup exceeded what would be expected by chance, null conditions were constructed by re-estimating heterogeneity in equally sized patient subpopulations (5000 permutations) drawn from the higher-order diagnostic group to which each subtype belonged. For each permutation, individuals were randomly sampled without replacement from the broader diagnostic category (e.g., all individuals with a psychiatric disorder), preserving the sample size of the original subtype (e.g., mood disorders). Heterogeneity (SD of z-scores) was then calculated for each resampled group. This procedure yielded a null distribution representing the degree of heterogeneity expected if no additional biological differentiation were present within subtypes beyond what is observed in their parent diagnostic group. In parallel, we quantified uncertainty around each subtype’s heterogeneity estimate by bootstrap resampling with replacement within the subtype (5000 permutations), repeatedly recalculating the (normalized) SD and deriving 95% confidence intervals from the 2.5th and 97.5th percentiles of the resulting bootstrap distribution. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file