Considering social risk alongside genetic risk for bipolar disorder in the All of Us Research Program

Summary

Polygenic risk scores (PRSs) have shown increasing utility for risk stratification across complex diseases, but for psychiatric disorders such as bipolar disorder (BD), current PRSs explain only a fraction of disorder liability (∼1%–9%), with predictive performance further diminished in non-European populations and electronic health record (EHR)-linked biobanks. To explore the potential of integrating social and environmental risk factors alongside genetic liability to improve risk characterization, we evaluated the relationship between a PRS for BD (PRS_BD) and six social risk measures—perceived stress, discrimination in medical settings, neighborhood social cohesion, perceived neighborhood disorder, cost-related medication nonadherence, and adverse childhood experiences (ACEs)—to BD case status in 115,275 participants (7,000 affected individuals and 108,275 control subjects) from the All of Us Research Program. PRS_BD was associated with BD case status across ancestry groups, though liability-scale variance explained was attenuated relative to what has been reported for curated research cohorts (R² = 1.86% in European, 0.60% in African, and 1.65% in Latino/Admixed American ancestries). Each social risk factor tested exhibited a larger effect size than PRS_BD, with perceived stress (odds ratio [OR] = 2.05 per SD) and ACEs (OR = 2.68 for ≥4 ACEs) demonstrating the strongest associations. Individuals in the lowest genetic risk decile with high social burden exhibited BD prevalence comparable to or exceeding that in the highest genetic risk decile with low social burden. These findings demonstrate the substantial explanatory power of social risk factors and support the development of integrated genetic-social risk frameworks for more accurate and equitable characterization of psychiatric risk.

Polygenic risk scores capture only a fraction of bipolar disorder liability. This multi-ancestry study of 115,275 All of Us participants demonstrates that social risk factors confer substantial explanatory power beyond genetic liability alone, emphasizing the necessity of incorporating social context into precision psychiatry models.

Introduction

Neuropsychiatric conditions, including major depression, anxiety, and bipolar disorder (BD; MIM: 125480), affect an estimated 1 in 5 adults in the United States (US) each year and are leading contributors to the global burden of disease.^1,2 In recent decades, the psychiatric genetics community has made substantial progress identifying the genetic contributions of these disorders.^3,4,5,6 Yet for many of these conditions, common genetic factors explain only a fraction of disorder risk, motivating efforts to improve risk prediction by integrating genetic information with environmental and social contexts.^7,8,9

Polygenic risk scores (PRSs) aggregate estimated effects of common trait-associated variants across the genome into a single value, representing the additive component of an individual’s genetic liability for the trait. Although PRSs have shown increasing utility in risk stratification across a number of diseases,^10,11 current PRSs capture only a fraction of total heritability for psychiatric disorders. For BD, SNP-based heritability estimates reach ∼22%, well below twin-based estimates of ∼60%–80%, and PRSs explain only ∼9% of phenotypic variance on the liability scale.^{12,13,14,15,16} The gap between twin and SNP heritability may in part reflect contributions from genetic factors not captured by PRSs, including rare coding and structural variants that may be particularly relevant for some psychiatric disorders.^17,18 Moreover, PRS derived from predominantly European-ancestry discovery samples show substantially reduced predictive performance in non-European-ancestry populations due to differences in linkage disequilibrium (LD) structure and allele frequencies, a limitation that remains a major barrier to equitable clinical implementation.¹⁹

The likelihood of developing a polygenic disorder such as BD is not solely a function of genetic factors but is also associated with social, environmental, and economic risk factors, collectively referred to as social determinants of health (SDOHs).²⁰ Several such factors have been associated with BD onset and severity. Childhood maltreatment or adversity has been consistently linked to earlier onset and disorder severity, though whether this relationship is moderated by polygenic burden remains unclear, with some studies reporting gene-environment interactions^21,22,23 and others finding primarily additive effects.²⁴ Stressful life events have been associated with mood episode onset in genetically at-risk individuals, and socioeconomic deprivation has been robustly associated with the prevalence and severity of mood disorders.^25,26 Additionally, experiences of discrimination in medical settings have been linked to diagnostic disparities, with evidence suggesting that Black Americans with BD are disproportionately misdiagnosed with schizophrenia, potentially reflecting both clinician bias and differential access to psychiatric care.^27,28 Notably, while several of these social factors have been studied as antecedents to psychiatric disorders, the temporal relationship between social risk and BD is often ambiguous; factors such as perceived stress and cost-related medication nonadherence may also arise as consequences of living with a chronic psychiatric condition, and it is important to note that cross-sectional study designs cannot fully disentangle these directions of association.

Aside from specific associations with BD, promising results indicate improved risk prediction and phenotypic discrimination when modifiable risk factors are included in models alongside genetic risk scores,^29,30,31 and researchers have increasingly called for the inclusion of SDOHs in risk prediction models across traits.^32,33,34 Methodologically, omission of strong social covariates from these models can bias risk estimates regardless of whether the omitted factors correlate with the PRS or modify its association strength.³⁵ Yet how genetic and social risk factors combine, or potentially interact, in their associations with complex disorder risk remains poorly understood.³⁶ Particularly in neuropsychiatry, rigorous evaluation of how social and environmental factors potentially modulate disorder susceptibility is required before PRSs can be adopted for clinical use.^7,37,38

Here, we evaluate the associations of social risk and genetic risk with BD in a diverse US population. We leverage data from the All of Us (AoU) Research Program to assess the predictive power of an externally developed PRS for BD (PRS_BD) derived from the latest Psychiatric Genomics Consortium (PGC) meta-analysis, alongside several social risk measures derived from AoU surveys. We model the effects of genetic and social risk factors and test for potential interactions to quantify their individual and combined associations with BD case status. By integrating genetic liability with social context, we seek to improve risk prediction strategies for complex psychiatric traits and underscore the value of incorporating social determinants into genetic risk frameworks for a holistic approach to psychiatric epidemiology.

Subjects and methods

AoU cohort

We utilized data from the AoU Research Program, a diverse cohort aiming to recruit one million participants across the US, a detailed report of which, including objectives, data availability, and limitations, has been previously published.³⁹ This study utilized data from 414,830 AoU participants with short-read whole-genome sequencing (srWGS) data enrolled by October 1, 2023 (data release v.8) and was conducted under the AoU Data Use Agreement.

Of the 414,830 participants with linked srWGS data, 115,275 participants were ultimately included in statistical analyses following genomic quality control (QC), electronic health record (EHR) filtering, and application of our phenotyping schema (Figure 1). Included participants had linked EHR data and had completed the relevant AoU survey modules for the investigated social risk variables (at missingness thresholds defined below). To ensure sufficient health record density for subsequent analyses, we restricted the cohort to participants with an EHR length of 3 years or more and a record depth of 3 or more distinct clinical visits. Finally, we restricted the final cohort to individuals with predicted genetic ancestry of either European, African, admixed American, or East Asian descent (matching the genome-wide association study [GWAS] source ancestries), who were aged 20–69 years at their last recorded event. This age range was selected to capture the period following typical BD onset while minimizing confounding from age-related multimorbidity.^40,41,42 South Asian and Middle Eastern predicted ancestry groups were excluded from all analyses due to case counts below 20 in each group.

Figure 1.

Flowchart detailing derivation of analysis cohort

Flowchart illustrating sample quality control, participant inclusion criteria, and EHR phenotyping steps taken to produce the final analysis cohort.

WGS data

We utilized the centrally processed v.8 genotype data provided by the AoU Researcher Workbench. Specifically, we used the allele count/allele frequency (ACAF)-thresholded PLINK-formatted files derived from srWGS data, which retain variants exceeding a population-specific allele count of 100 or a population-specific allele frequency of 1%.

We removed samples flagged by AoU as having low-quality WGS data during the joint callset QC process. To account for cryptic relatedness, we utilized the pre-computed kinship scores provided by AoU and excluded the AoU-recommended list of samples to generate a maximally unrelated set of participants.⁴³

Demographic and ancestry variables

We defined a three-level sex/gender variable based on concordance between sex assigned at birth and self-reported gender identity. Participants were classified as cis-male (assigned male at birth, identified gender as male), cis-female (assigned female at birth, identified gender as female), or sexual and gender minority (SGM) (discordant sex and gender indications, intersex, or non-binary identities). Participant age was calculated as the interval between the date of birth and the date of the last recorded event in their EHR data. EHR length and record depth were derived from available EHR data, with EHR length calculated as the interval between the first and last recorded events and record depth as the number of distinct recorded events. Finally, genetic ancestry, as well as genetic ancestral principal components (PCs), was defined using the pre-computed ancestry predictions provided by AoU as previously described.⁴³

PRS calculation

We computed PRS_BD in our cohort using publicly available summary statistics from the largest multi-ancestry GWAS of BD to date, performed by the PGC.¹² Specifically, we utilized summary statistics from the meta-analysis multi-ancestry analysis excluding self-reported affected individuals (n_{affected individuals} = 67,948; n_{control subjects} = 867,710). We selected the multi-ancestry GWAS without self-report because O’Connell et al. demonstrated that self-reported affected individuals have a distinct genetic architecture with substantially lower SNP-heritability (h²_SNP = 0.08) compared to clinically ascertained affected individuals (h²_SNP = 0.22) and that PRSs derived from this sub-cohort showed better predictive performance in clinical cohorts (weighted liability R² = 0.090) compared to models including self-reported data (R² = 0.058).¹²

Following the approach used in O’Connell et al., we computed PRSs for our cohort using the Bayesian regression framework PRS-CS-auto to estimate posterior SNP effect sizes.⁴⁴ While the source GWAS was a multi-ancestry meta-analysis, a majority of participants (∼82%) were of European descent. Consequently, we elected to use the 1000 Genomes Project Phase 3 European Reference panel provided by the PRS-CS developers.⁴⁵ PRS-CS-auto was run on each chromosome with 10,000 Markov chain Monte Carlo (MCMC) iterations and 5,000 burn-in steps, as recommended by the developers.⁴⁴

Posterior effect sizes were lifted over from GRCh37 to GRCh38 using Hail (v.0.2.130).⁴⁶ We performed allele harmonization by removing strand-ambiguous variants and aligning effect alleles to the target dataset, flipping the sign of the effect size where the effect allele matched the reference. Variants that failed liftover or had unresolved allele mismatches were excluded. Individual-level PRSs were calculated using PLINK v.1.9⁴⁷ with the --score function, summing the dosage of the effect allele weighted by the posterior effect size estimated by PRS-CS-auto. PRS-CS-auto estimated posterior effect sizes for 1,061,483 variants. After liftover to GRCh38 (414 variants failed), removal of strand-ambiguous variants, and allele harmonization with the AoU target dataset, 1,045,187 variants were retained for score calculation.

To address potential effects of population stratification, raw PRS values were adjusted for the top 16 genetic ancestry PCs.⁴⁸ The resulting residuals were standardized (mean = 0, SD = 1) within the full cohort to generate the final global PRS used for analysis. We confirmed that this residualization of the PRS on a continuous measure of ancestry largely removed large population-scale differences in the PRS distributions (Figure S1).

As a sensitivity analysis, we also computed PRS using PRS-CSx, a multi-ancestry extension of PRS-CS-auto that jointly models multiple ancestry-specific GWAS summary statistics with matched ancestry-specific LD reference panels to estimate posterior SNP effect sizes.⁴⁹ We applied PRS-CSx to the BD GWAS summary statistics for European, African, and East Asian individuals; Latino/admixed American summary statistics were not available in the source GWAS. All other parameters and downstream processing were identical to the PRS-CS-auto procedure described above. Results from this comparison are reported in Table S1.

BD phenotyping

BD case and control status was determined using EHR data within the AoU Researcher Workbench. Affected individuals were defined as participants meeting one of the following criteria: (1) presence of at least 2 diagnostic codes (ICD-9/10 or SNOMED) for BD more than 30 days apart, (2) presence of at least 1 BD diagnostic code and a prescription record for a BD-indicated medication within 1 year of each other, or (3) presence of at least 1 BD diagnostic code and a high volume of BD-indicated medications (≥30 total records or ≥3 unique medications). BD-indicated medications included lithium, valproate, lamotrigine, and carbamazepine (for a detailed list, see the supplemental information). For participants meeting at least one of the above criteria, we next excluded those with codes for a conflicting psychiatric condition (e.g., schizophrenia) that outnumbered the participant’s BD code count, unless the BD codes were the most recent. This case definition strategy was adapted from a recent clinician-informed EHR-based phenotyping approach for early-onset BD.⁵⁰ Requiring convergent evidence from repeated diagnostic codes or concordant medication records provides a more stringent case definition than single-code approaches, which have been shown to capture heritable variation unrelated to the target disorder.⁵¹

Control subjects were defined as participants with no BD diagnostic codes and no exclusionary psychiatric diagnoses. We note that 3.7% of control subjects had at least one prescription record for a BD-indicated medication (predominantly lamotrigine, valproate, or carbamazepine), most likely reflecting use for non-psychiatric indications such as epilepsy; lithium prescriptions were rare among control subjects (0.22%) (Figure S2). These individuals were retained as control subjects given the absence of BD diagnostic codes in their records. For full code and medication lists used to construct the cohort, see the supplemental information.

To assess the consistency of our EHR-based case definitions, we examined the concordance with self-reported personal and family history of BD. Specifically, we calculated PRS risk curves across subgroups of participants who self-reported (1) having at least one parent with BD, (2) having BD themselves, (3) currently receiving medication or treatment for BD, and (4) currently seeing a doctor for BD. We additionally compared rates of self-reported BD history between EHR-defined affected individuals and control subjects. While self-report measures are subject to recall bias and respondents may be unable to reliably discriminate between psychiatric diagnoses,⁵² the concordance observed between EHR-based and self-report classifications provides additional support for the reliability of our phenotyping strategy.

SDOH measures

We selected six social risk measures spanning domains with prior evidence of association with BD or related psychiatric outcomes. Perceived stress has been associated with mood episode onset in genetically at-risk individuals^53,54; experiences of discrimination in medical settings may contribute to diagnostic disparities and case misclassification in BD²⁷; measures of cost-related medication non-adherence capture economic barriers to treatment continuity,^55,56 a factor linked to poor psychiatric outcomes; adverse childhood experiences have been consistently associated with earlier BD onset and greater illness severity^22,57,58; and neighborhood social cohesion and perceived neighborhood disorder capture community-level social and physical environment exposures associated with psychiatric morbidity.^59,60,61

Measures were derived from participant-provided responses to “The Basics,” “Social Factors of Health,” “Health Care Access and Utilization Survey,” “Personal/Family Health History,” and “Emotional Health History and Well-being” surveys. From these, individual survey responses were extracted and scored according to validated guidelines for each instrument. These six surveys were perceived stress (PSS; from the PSS-10)^62,63; discrimination in medical settings (DMS),⁶⁴ adapted from the Everyday Discrimination Scale⁶⁵; cost-related medication non-adherence (CRMN)⁶⁶; neighborhood social cohesion (SCNS)⁶⁷; perceived neighborhood disorder (PNDS)⁶⁸; and adverse childhood experiences (ACE).⁶⁹

The following analyses were performed separately for each survey. First, participants who completed less than 70% of the items for a specific survey were excluded from the analysis for that survey. Patterns of item-level missingness within each survey were evaluated using the naniar R package.⁷⁰ The overlap of participants completing the various combinations of these six surveys is shown in Figure S3. For the remaining participants, missing item-level values were imputed using a k-nearest neighbors approach (k = 5)⁷¹ implemented using the caret R package prior to scoring.^72,73 Finally, distributional accuracy of score imputation was assessed by comparing imputed distributions to complete-case distributions (Figure S4).

Prior to association analysis, survey scores were transformed for comparison. Inter-scale correlations among the six social risk measures are shown in Figure S5. Continuous scores (PSS, DMS, SCNS, and PNDS) were standardized to Z scores (mean = 0, SD = 1). Directionality was standardized such that higher scores reflect higher risk. CRMN and ACE scores were analyzed as binary variables, with a CRMN value of 1 indicating any behavior of CRMN and an ACE value of 1 indicating 4 or more adverse experiences.⁶⁹

Due to varying completion rates across surveys and ancestry groups, analytic sample sizes differed for each social risk variable (range: n = 17,098–64,009). Full sample sizes by survey and ancestry group are reported in Table S2, and completion rates by case/control status and ancestry are shown in Figure S6. Descriptive statistics of social risk factors by case status are provided in Table S3.

Statistical analyses

All statistical analyses were performed on the AoU Researcher Workbench using R v.4.4.0.⁷⁴ Data manipulation was conducted using tidyverse, and descriptive statistics and cohort characteristics tables were generated using the gtsummary and flextable packages.^75,76,77 We assessed the performance of the PRS using logistic regression models adjusted for age (linear and quadratic terms), sex/gender, EHR length and depth, and 16 genetic PCs. Model discrimination was evaluated using the area under the curve (AUC) calculated using the pROC package.⁷⁸ Variance explained was estimated using liability-scale R² assuming a population prevalence of 2%¹² with 95% confidence intervals (CIs) estimated via bootstrapping (1,000 replicates) implemented using the boot package.^79,80,81

For each social risk variable, we used logistic regression to model BD case status as a function of the continuous, standardized PRS, the social risk variable, and the aforementioned covariates. All PRS associations were estimated using the continuous, standardized PRS. Decile-based analyses are presented for visualization purposes only. Model outputs were processed using the broom package.⁸² To assess potential multiplicative effects, we fitted models including a PRS × social risk interaction term. We tested for multiplicative interaction using a likelihood-ratio test (LRT) comparing the interaction model to a base model without the interaction term. Statistical significance was assessed after correcting for multiple testing using the Benjamini-Hochberg procedure. All PRS performance and social risk association models were fit in the combined “all-ancestries” sample and separately within European, African, and admixed American ancestry groups. East Asian participants were included in the all-ancestries analysis but excluded from stratified models due to insufficient sample size (see demographic and ancestry variables).

Even if the multiplicative interaction term in a logistic model is null, there can still be excess risk due to interactions between variables when risk is assessed on the additive scale.⁸³ To assess interaction on the additive scale, we computed the relative excess risk due to interaction (RERI) for each social risk variable using the delta method applied to the logistic regression coefficients and their covariance matrix via the interactionR package.⁸⁴ RERI > 0 indicates super-additive interaction, whereby individuals with both high genetic and high social risks have a greater risk of BD than would be expected by simply summing the effects of each individual risk factor.⁸⁵ Because standard RERI estimation requires categorical exposures, we dichotomized genetic and social risk prior to testing. High genetic risk was defined as a PRS in the top tertile (>67^th percentile). High social risk was defined using the top tertile of the cohort distribution for the four continuous measures (PSS, DMS, SCNS, and PNDS) and established thresholds for the two count-based indicators (4+ reported ACEs or any CRMN).

To test the robustness of our social risk associations, we conducted two sensitivity analyses. First, to assess potential selection bias from differential survey completion, we applied inverse probability weighting (IPW). Within the full analytic cohort (respondents and non-respondents), we modeled the probability of survey completion using logistic regression on BD case status, age, sex/gender, 16 genetic PCs, EHR length, and EHR depth. Separate completion models were fit within the all-ancestries cohort and separately within each ancestry group (European, African, and Latino/admixed American); ancestry was included as a covariate only in the all-ancestries model. Weights derived from each model were applied to the corresponding all-ancestries or ancestry-stratified outcome regressions. Weights were then calculated as the inverse of the predicted probability and truncated at the 1^st and 99^th percentiles to stabilize variance. Covariate balance was assessed via absolute standardized mean differences using the cobalt R package, and weighted association models were then fit in both the all-ancestries and ancestry-stratified groups using the survey R package.^86,87 Second, to evaluate whether associations varied by diagnostic opportunity, we stratified the analytic cohort into tertiles based on total EHR depth (low, medium, and high utilization) and re-estimated the primary social risk models within each stratum.

Finally, to examine cumulative social risk burden, we constructed a count variable reflecting the number of high-risk social factors present for each participant, with high risk defined for each factor as in the RERI analysis described above. BD prevalence was then examined as a function of this cumulative risk count (range: 0–6).

Results

Study design and cohort characteristics

Out of 414,830 participants in the AoU v.8 cohort with available srWGS data, a total of 115,275 individuals (7,000 affected individuals and 108,275 control subjects) were retained for further analyses after applying QC and cohort selection criteria (Figure 1). The cohort was predominantly cis-female (66%), with a mean age of 50.5 years (SD = 13.2) (Table 1). From here, we stratified further analyses by genetically predicted ancestry group as defined by the centrally processed AoU ancestry predictions. As expected, most participants were of European ancestry (4,054 affected individuals and 63,370 control subjects), followed by participants of African (1,953 affected individuals and 24,014 control subjects) and Latino/admixed American (944 affected individuals and 18,273 control subjects) ancestry. East Asian participants were ultimately excluded from ancestry-stratified analyses due to low sample size (49 affected individuals and 2,618 control subjects) but were retained in the all-ancestries data. For subsequent survey-based analyses, participants were included only if they completed at least 70% of a given social risk survey. The proportion of participants meeting this threshold differed by survey, ancestry group, and case status (Figure S6), with European-ancestry participants generally showing higher completion rates and a notably larger case-control completion gap (mean Δ = 17.1 percentage points) than was observed in African (Δ = 4.4 pp) or admixed American (Δ = 5.4 pp) ancestry groups (all χ2 p < 0.001). Consequently, subsequent analytic sample sizes ranged from n = 17,098 (ACEs) to n = 64,009 (CRMN), with four of six surveys yielding samples of approximately 49,000. Full sample sizes by survey and ancestry group are reported in Table S2.

Table 1.

Cohort demographic and social risk characteristics

	Overall	Affected individuals	Control subjects	LRT χ2	LRT p value
n	115,275	7,000	108,275	N/A	N/A
Age, mean (SD)	50.5 (13.2)	48.9 (12.3)	50.6 (13.2)	640.2	<0.001
Cis-female, n (%)	76,185 (66.1)	4,409 (63.0)	71,776 (66.3)	171.6a	<0.001
Cis-male, n (%)	38,241 (33.2)	2,433 (34.8)	35,808 (33.1)	N/A	N/A
SGM, n (%)	849 (0.7)	158 (2.3)	691 (0.6)	N/A	N/A
European, n (%)	67,424 (58.5)	4,054 (57.9)	63,370 (58.5)	–b	N/A
African, n (%)	25,967 (22.5)	1,953 (27.9)	24,014 (22.2)	–b	N/A
Latino/admixed American, n (%)	19,217 (16.7)	944 (13.5)	18,273 (16.9)	–b	N/A
East Asian, n (%)	2,667 (2.3)	49 (0.7)	2,618 (2.4)	–b	N/A
EHR length, days, mean (SD)	4,289.0 (2,621.0)	4,722.2 (2,613.7)	4,260.9 (2,619.0)	4.9	0.027
EHR depth, visits, mean (SD)	110.6 (163.7)	234.9 (356.5)	102.5 (138.7)	2,431.6	<0.001
Mean PRS (SD)	0.00 (1.00)	0.18 (1.01)	−0.01 (1.00)	228.4	<0.001
Perceived stress (PSS) (mean, SD) (n = 48,881)	14.3 (7.7)	20.5 (8.1)	14.0 (7.5)	907.5	<0.001
Missing, n	66,394	4,914	61,480	N/A	N/A
Discrimination in medical settings (DMS) (mean, SD) (n = 49,402)	11.1 (4.5)	13.8 (5.7)	11.0 (4.4)	381.3	<0.001
Missing, n	65,873	4,887	60,986	N/A	N/A
Social cohesion (SCNS) (mean, SD) (n = 49,478)	15.2 (3.0)	13.6 (3.5)	15.2 (3.0)	402.6	<0.001
Missing, n	65,797	4,880	60,917	N/A	N/A
Perceived neighborhood disorder (PNDS) (mean, SD) (n = 48,900)	21.4 (6.8)	24.7 (7.9)	21.2 (6.7)	337.0	<0.001
Missing, n	66,375	4,923	61,452	N/A	N/A
Cost-related medication nonadherence (CRMN) (mean, SD) (n = 64,009)	0.6 (1.2)	1.0 (1.5)	0.6 (1.1)	108.5	<0.001
Missing, n	51,266	4,117	47,149	N/A	N/A
Adverse childhood experiences (ACEs) (mean, SD) (n = 17,098)	2.3 (2.1)	3.8 (2.2)	2.2 (2.0)	131.0	<0.001
Missing, n	98,177	6,300	91,877	N/A	N/A

Data are mean (SD) or n (column %). The LRT χ2 for each term was obtained by comparing the full joint logistic model of bipolar disorder case status with the model omitting that term only; p is from the χ2 distribution. χ2, chi-square test; LRT, likelihood-ratio test; EHR, electronic health record; PRS, polygenic risk score; SD, standard deviation; SGM, sexual and gender minority.

^aSex/gender is tested as a single multi-category term (degrees of freedom [df] = 2).

^bGenetic ancestry is not tested separately, as it is captured by the 16 principal components.

PRS performance

We computed individual PRSs (PRS_BD) in our sample of 7,000 individuals with BD and 108,275 control subjects, observing a positive association between PRS_BD and odds of BD case status across ancestries, though predictive performance varied by population (Table 2; Figure 2). In the all-ancestries analysis, each standard deviation increase in PRS_BD was associated with a 22% increase in odds of BD case status (odds ratio [OR] = 1.22; 95% CI: 1.19–1.25; p = 5.91 × 10⁻⁵²). Liability-scale R² was highest in the European group (1.86%; 95% CI: 1.29–2.47), followed by Latino/admixed American (1.65%; 95% CI: 0.68–3.16), with lower performance in the African-ancestry group (0.60%; 95% CI: 0.22–1.12). These values are substantially attenuated relative to the PGC target cohorts, where O’Connell et al. reported a median R² liability of 9.0% in European-ancestry groups and 2.2% in African-ancestry groups.¹² In absolute terms, observed BD prevalence ranged from 4.4% in the lowest PRS decile to 8.6% in the highest decile in the all-ancestries sample (risk ratio = 2.0), with the strongest separation observed in the European subsample (4.2% vs. 8.8%; risk ratio = 2.1). Theoretical absolute risk curves illustrate the translational population-level risk estimates by ancestry, ranging from 1% to 5% across the PRS distribution (Figure S7).

Table 2.

Polygenic risk score performance across ancestry groups

Ancestry	n (affected individuals)	OR per SD (95% CI)	R2 liability, % (95% CI)	p
All ancestries	115,275 (7,000)	1.22 (1.19–1.25)	1.53 (1.15–2.01)	5.91 × 10−52
European	67,424 (4,054)	1.22 (1.18–1.26)	1.86 (1.29–2.47)	1.00 × 10−36
African	25,967 (1,953)	1.18 (1.11–1.25)	0.60 (0.22–1.12)	1.31 × 10−7
Latino/admixed American	19,217 (944)	1.21 (1.13–1.30)	1.65 (0.68–3.16)	4.56 × 10−8

Models adjusted for age, age², sex/gender, EHR length, EHR depth, and 16 genetic PCs. OR per SD, odds ratio per 1-SD increase in the standardized polygenic risk score; R² liability, variance explained on the liability scale (assuming 2% population prevalence); CI, confidence interval; N, affected individuals; analytic N, number of individuals affected by bipolar disorder.

Figure 2.

Bipolar disorder polygenic risk score risk stratification across ancestry groups

(A) Adjusted odds ratios (ORs) of bipolar case status across PRS deciles. ORs were calculated relative to the middle quintile (deciles 5–6, indicated by the shaded region) using a logistic regression model adjusted for age, age², sex/gender, EHR length, EHR depth, and 16 genetic PCs. Trend lines are split by genetically predicted ancestry group, faceted out to the right.

(B) Error bars represent the 95% confidence intervals (95% CIs) derived from the adjusted logistic regression models. Gray background lines show the trends of other groups for comparison.

In a sensitivity analysis comparing PRS-CSx to PRS-CS-auto, we observed slightly lower liability-scale R² when using PRS-CSx across all ancestry groups (Table S1), supporting our use of PRS-CS-auto as the primary scoring method.

To evaluate model discrimination, we assessed AUC across three covariate sets: a minimally adjusted model matching O’Connell et al. (set 1: sex + 5 PCs), our primary fully adjusted model (set 2: age + age² + sex/gender + EHR length + EHR depth + 16 genetic PCs), and models additionally including each social risk variable (set 3). For each covariate set, we additionally computed change in AUC between the covariate-only model and the same model with PRS added, isolating the individual PRS contribution in each scenario. Across covariate sets, PRS contributed modestly to model discrimination, with change in AUC in set 2 ranging from 0.002 to 0.009 across ancestry groups. Full results are reported in Table S4. Due to the apparent low additional contribution of the PRS to model accuracy compared to that achieved using only demographic covariates, we also tested for relationships between the included covariates and PRS_BD. Among the model covariates, EHR depth contributed most to model discrimination (LRT χ2 = 2,431.6), consistent with the substantially higher healthcare utilization among individuals with BD (mean: 234.9 vs. 102.5 visits for control subjects), followed by age (χ2 = 640.2) and sex/gender (χ2 = 171.6).

Despite the modest variance explained, we observed a consistent linear increase in odds of BD case status across PRS deciles (Figure 2). Relative to individuals at median risk (40^th–60^th percentile), individuals in the top decile had elevated odds of BD in the combined analysis (OR = 1.45; 95% CI: 1.33–1.59), with similar elevation seen in European (OR = 1.44; 95% CI: 1.29–1.60), African (OR = 1.30; 95% CI: 1.05–1.60), and Latino/admixed American (OR = 1.45; 95% CI: 1.14–1.83) groups. Full decile-stratified ORs by ancestry are reported in Table S5.

We additionally examined concordance between our EHR-based phenotyping and self-reported personal and family history of BD (see subjects and methods). Across four self-report indicators, the higher PRS was associated with higher probabilities of EHR-defined case status, and EHR-defined individuals with BD self-reported personal or family history of BD at substantially higher rates than control subjects (Figure S8). PRS distributions across EHR case assignment strategies are shown in Figure S9.

Finally, we also examined diagnostic performance metrics for each self-report indicator under each case definition strategy, observing variable concordance with self-reported diagnosis (sensitivity = 28.1%, positive predictive value = 67.7% for the primary EHR case definition; full results are provided in Figure S10 and Table S6).

Association of social risk factors with BD

We next evaluated the relationship between six social risk factors and BD case status. We note that survey completion rates differed across affected individuals and control subjects (Figure S6); thus, the social ORs reported here represent associations specific for this subsampled cohort and cannot be easily generalized to the larger population.

In our sample, however, we observed significant associations between higher social risk and odds of BD case status across all ancestry groups (Figure 3). Among the continuous variables tested, perceived stress (PSS) yielded the largest effect size; in the all-ancestries analysis, every 1-SD increase in PSS was associated with an approximately 2-fold increase in the odds of BD case status (OR = 2.05; 95% CI: 1.96–2.15). Similarly, for every 1-SD increase in social cohesion (SCNS; reverse scored), we observed an approximately 1.5-fold increase in the odds of BD case status (OR = 1.55; 95% CI: 1.49–1.62), with DMS (OR = 1.47) and perceived neighborhood disorder (PNDS) (OR = 1.49) showing comparable effect sizes.

Figure 3.

Association of social risk factors with bipolar disorder across ancestry groups

Forest plot displaying adjusted odds ratios (ORs) and 95% confidence intervals (95% CIs) for the association between social risk measures and bipolar disorder case status. Associations were estimated using logistic regression models adjusted for age, age², sex/gender, EHR length, EHR depth, and 16 genetic PCs. The top image displays results for continuous variables, which were Z score standardized, such that ORs represent the risk associated with a 1-SD increase in the social score. The bottom image displays results for binary indicators.

This pattern was largely mirrored for the examined binary indicators (CRMN and ACEs). Reporting 4 or more ACEs showed the largest effect size of the binary variables tested, particularly in the European group (OR = 2.89; 95% CI: 2.41–3.46; p = 7.54 × 10⁻³¹). Note that binary ORs (ACEs and CRMN) represent the odds associated with exceeding a threshold, whereas continuous ORs (PSS, DMS, SCNS, and PNDS) represent the odds per standard deviation increase; thus, these effect sizes are not directly comparable. While effect sizes were directionally consistent in the Latino/admixed American group, the associations for ACEs (OR = 1.82; 95% CI: 0.96–3.49; p = 0.069) and CRMN (OR = 1.28; 95% CI: 0.98–1.65; p = 0.065) did not reach statistical significance. Notably, effect sizes for social risk variables were consistently smaller in the African-ancestry group compared to the European and admixed American groups. For example, the PSS association in the African-ancestry group (OR = 1.72; 95% CI: 1.52–1.94) was attenuated relative to the European group (OR = 2.14; 95% CI: 2.03–2.26), a pattern observed across all six measures. Detailed association results by ancestry are provided in Table S7.

Beyond individual metrics, we also examined the cumulative impact of social risk. BD prevalence remained relatively stable among the lowest-burden groups (∼4%–6% for individuals with 0–2 factors) but exhibited an upward trend with the accumulation of additional risk factors, ultimately reaching ∼22% for individuals reporting all six risk factors (Figure S11).

To ensure our control definition did not artificially inflate these social risk associations, we repeated the primary analyses using a broader control group that excluded only individuals with BD, without additional exclusions based on psychiatric indications. Effect sizes were modestly attenuated across all social risk factors (2.7%–9.3%) but remained significant and directionally consistent (Figure S12; Table S8).

Finally, we evaluated the robustness of the above associations to survey non-response and healthcare utilization biases. First, to address differential survey completion, we conducted a response-weighted sensitivity analysis using IPW. This weighting successfully balanced covariate differences between survey respondents and non-respondents (absolute standardized mean differences < 0.1; Figure S13A), and applying these weights to our primary models did not meaningfully alter the magnitude or direction of the social risk associations in either the all-ancestries or ancestry-stratified analyses (Figure S13B). Second, to ensure our findings were not disproportionately driven by diagnostic opportunity, we stratified the cohort into tertiles of EHR depth. The ORs for all six social risk measures remained stable and directionally consistent across all levels of healthcare utilization (Figure S14).

Combined contributions of polygenic and social risk to BD

To examine the combined contributions of polygenic and social risk, we examined BD prevalence across PRS_BD deciles stratified by social risk group (Figure 4). Prevalence curves for the remaining four social variables are shown in Figure S15. Here, we highlight SCNS and CRMN, one continuously modeled and one binary-modeled variable, as these yielded the strongest interaction estimates and interaction significance values among the six tested variables. Notably, SCNS and CRMN also represent our two largest analytic subsamples among the tested social risk variables (n = 49,478 and n = 64,009, respectively), affording greater statistical power to detect subtle interaction effects. For both neighborhood social cohesion (SCNS; Figure 4A) and CRMN (Figure 4B), individuals with high social risk (red lines) exhibited higher disease prevalence across all genetic risk deciles compared to those with low social risk (blue lines). Notably, individuals in the lowest genetic risk decile with high social burden often had disease prevalence rates comparable to or exceeding those in the highest genetic risk decile who had low social burden, underscoring the magnitude of social risk contributions. This contrast is further illustrated by direct comparison of BD prevalence between extreme combined risk groups (Figure S16). PRS_BD associations stratified by social risk level for all six variables are shown in Figure S17.

Figure 4.

Polygenic risk for bipolar disorder stratified by social risk

Predicted (dashed) and observed (solid) prevalence of bipolar disorder (BD) across PRS deciles, stratified by social risk strata for (A) low neighborhood social cohesion (SCNS; tertiles) and (B) cost-related medication nonadherence (CRMN; any vs. none). Predicted values are derived from logistic regression models adjusted for age, age², sex/gender, EHR length, EHR depth, and 16 genetic PCs. PRS deciles were calculated within each analytic subsample. Error bars represent 95% confidence intervals. Sample sizes for each stratum are shown in the legend.

As has been discussed in the context of polygenic scores, coefficient estimates in logistic models can be biased when relevant predictors are omitted, even in the absence of true interactions.^35,88 We therefore fit models for case status that included social risk, PRS_BD, and their interaction as predictors. Across the six investigated factors, the interaction terms were generally small in magnitude, with 95% CIs often crossing or approaching zero (Table 3).

Table 3.

Interaction between polygenic risk score and social risk factors

Variable	N	N affected individuals	PRS or (95% CI)	SDOH OR (95% CI)	Interaction β (95% CI)	LRT χ2	LRT df	LRT p value	LRT p (FDR)
Perceived stress (PSS)	48,881	2,086	1.20 (1.15–1.26)	2.04 (1.95–2.14)	−0.043 (−0.087 to 0.001)	3.67	1	0.055	0.096
Discrimination in medical settings (DMS)	49,402	2,113	1.22 (1.17–1.27)	1.46 (1.41–1.52)	0.007 (−0.028 to 0.042)	0.16	1	0.692	0.692
Cost-related medication nonadherence (CRMN)	64,009	2,883	1.23 (1.18–1.27)	1.54 (1.42–1.67)	−0.085 (−0.164 to 0.006)	4.50	1	0.034	0.096
Low social cohesion (SCNS)	49,478	2,120	1.21 (1.16–1.27)	1.55 (1.48–1.62)	−0.050 (−0.091 to 0.010)	5.85	1	0.016	0.093
Perceived neighborhood disorder (PNDS)	48,900	2,077	1.22 (1.17–1.28)	1.49 (1.43–1.55)	−0.031 (−0.070 to 0.008)	2.41	1	0.120	0.144
Adverse childhood experiences (ACEs)	17,098	700	1.23 (1.14–1.33)	2.64 (2.24–3.11)	−0.146 (−0.301 to 0.009)	3.43	1	0.064	0.096

PRS OR and SDOH OR: odds ratios (95% CI) from the main-effects model (PRS + SDOH + covariates, no interaction term). Interaction β: coefficient (95% CI) for the PRS × SDOH product term on the log-odds scale; a negative value indicates sub-multiplicative interaction. LRT χ2 (df), p: likelihood-ratio test comparing the interaction model to the main-effects model. p(FDR): Benjamini-Hochberg corrected across the six interaction tests. Covariates: age, age², sex/gender, 16 genetic ancestry PCs, EHR length, and EHR depth. ACEs and CRMN are modeled as binary; all other variables are modeled as continuous Z scored measures.

On the multiplicative scale, two variables showed nominally significant interactions, including SCNS (n = 49,478; p_int = 0.016). Notably, the association with social risk (OR = 1.55) was larger than that with genetic risk (OR = 1.21), and the negative estimate of the interaction coefficient (−0.05) indicated that a slightly attenuated association with genetic risk was observed in the high-social-risk group. We observed a similar pattern for CRMN (n = 64,009; p = 0.034), where the association with social risk (OR = 1.54) exceeded the association with genetic risk (OR = 1.23), and the negative interaction coefficient (−0.09) again suggested a slightly attenuated association with genetic risk in the high-social-risk group. However, neither of these interaction terms remained significant after correction for multiple testing (Benjamini-Hochberg false discovery rate [FDR] > 0.05). Overall, across all six interaction models, the OR for the social risk term (range: OR = 1.46–2.04 for continuous measures; OR = 1.54–2.64 for binary measures; Table 3) exceeded that of the PRS (range: OR = 1.20–1.23), reinforcing that social risk factors were strongly associated with BD beyond what was captured by polygenic risk alone.

To complement the multiplicative analysis, we also assessed additive interaction using the RERI (Table S9). DMS was the only variable that showed significant additive interaction (RERI = 0.62; 95% CI: 0.24–1.00; synergy index = 1.42; 95% CI: 1.14–1.77). The attributable proportion (AP) of the interaction of PRS and DMS was about one-fifth (AP = 0.20; 95% CI: 0.09–0.31), that is, one-fifth of the combined risk among dually exposed individuals could be attributed to the interaction itself (Figure S18). No other social risk variable showed significant departure from additivity after correction for multiple testing. Interaction estimates were broadly consistent when re-estimated under alternative case definition strategies (Figure S19).

Discussion

Summary of principal findings

In this study, we evaluated the contributions of PRSs and social risk factors to BD case status in the AoU Research Program. We successfully reproduced the association between PRS and BD case status across all ancestries, though with low and variable predictive performance. We also evaluated the effect of several social risk factors on BD liability: perceived stress (PSS), experiences of DMS, perceived neighborhood disorder (PNDS), neighborhood social cohesion (SCNS), nonadherence to prescribed medication due to cost barriers (CRMN), and reports of ACEs. Although we hypothesized that social risk might make those with the highest polygenic risk even more susceptible to development of BD, our results did not yield statistically significant multiplicative interactions after correction. Instead, we observed primarily additive contributions of genetic and social risk, with a significant super-additive interaction observed between genetic risk and DMS.

PRS performance and the phenotyping gap

A notable finding in our study was the attenuation of PRS predictive power compared to prior literature. While the latest PGC meta-analysis for BD reported R² liability-scale estimates of 9% in clinical cohorts, our analysis yielded estimates of just 1.86% in European ancestry and 0.60% in African ancestry. This discrepancy likely reflects the “phenotyping gap” observed between curated research cohorts and EHR-linked population biobanks. Discovery GWASs typically rely on strict, structured clinical interviews to define affected individuals, whereas our study relied on EHR diagnostic codes and medication records. Psychiatric disorders are prone to high rates of diagnostic flux and misclassification—estimates suggest that up to 40% of patients with BD are initially misdiagnosed, with the rate and nature of misdiagnosis varying by racial/ethnic group.^28,89 The misclassification inherent in EHR-based phenotyping inevitably dilutes the genetic signal. Furthermore, recent simulations demonstrate that diagnostic inaccuracies based on EHR codes and self-report instruments do not simply attenuate signal but also introduce heritable confounders that generate systematic off-target genetic associations, which cannot be overcome by increasing sample size alone.⁵¹ The extent to which such confounders operate differently across ancestry groups—for instance through differential patterns of healthcare utilization, diagnostic practices, or case ascertainment—remains an open question that may further contribute to the ancestry-stratified differences in both PRS performance and social risk effect sizes observed in our study. Consistent with this, the PGC BD GWAS found that SNP heritability differed markedly by ascertainment method (h²SNP_clinical = 0.22; h²SNP_EHR = 0.05; h²SNP_self-report = 0.08) and that genetic correlations between clinical and self-report BD cohorts were well below unity (genetic correlation [rG] = 0.47), indicating partially distinct genetic architectures.^12,51 Finally, PRSs derived from shallow phenotypes have been shown to predict many phenotypes beyond the target disorder, suggesting that higher R² values in curated research cohorts may partly reflect nonspecific genetic signals. Our results may therefore provide a more realistic estimate of how current PRSs perform when applied to the heterogeneous clinical reality of general healthcare systems.

The impact of social risk

Consistent with established epidemiological work, our results underscore the sizable role of social and environmental risk factors in psychiatric outcomes.^90,91 The effect sizes we observed for social variables were substantial; for instance, reporting four or more ACEs was associated with a nearly 3-fold increase in the odds of BD in European ancestry (OR = 2.89; 95% CI: 2.41–3.46). Similarly, perceived stress (PSS) showed a strong linear association across all groups, with each standard deviation increase in PSS corresponding to an approximately 2-fold increase in BD odds (OR = 2.05; 95% CI: 1.96–2.15 in the all-ancestries analysis).

When modeling the combined contributions of PRSs and social risk factors, we observed that social risk consistently contributed substantial additional explanatory power beyond genetic risk alone. In several models, the magnitude of social risk contributions rivaled or exceeded that of genetic risk, with high social burden conferring substantial risk even among individuals in the lowest deciles of polygenic liability. Across the six investigated social risk factors, multiplicative interaction terms between PRS and social risk were generally small in magnitude and did not survive multiple-testing correction (FDR > 0.05). Notably, CIs for PRS associations overlapped across social risk strata for all variables (Figure S17), consistent with the non-significant interaction terms. Interestingly, the nominally significant negative interaction coefficients observed for SCNS and CRMN suggest that the association between genetic risk and BD is slightly attenuated in high-social-risk groups.

Conversely, our assessment of interactions on the additive scale revealed a super-additive effect for DMS (Table S9). Taken together, these findings suggest that while most social factors examined here increase BD risk through primarily additive effects independent of genetic liability, specific stressors may have the potential to compound genetic vulnerability. However, given the attenuated PRS performance in our cohort and the relatively modest sample sizes available for interaction testing, larger cohorts are needed to robustly evaluate these complex gene-environment dynamics.

Ancestry differences, portability, and equity

Our study further highlights the persistent “portability gap” seen in complex trait genetics. Consistent with the well-documented Eurocentric bias in GWAS discovery, PRS discriminative ability was highest in European and admixed American groups and markedly lower in the African-ancestry group. This lower PRS performance likely reflects a combination of the portability gap and ancestry-specific phenotyping and ascertainment biases. For instance, O’Connell et al. found that excluding self-reported affected individuals improved PRS performance in European and East Asian target cohorts, but in an African-ancestry target cohort, inclusion of self-reported affected individuals substantially increased variance explained.¹²

We also explored ancestry-matched LD reference panels via PRS-CSx to address the European-ancestry bias of the discovery GWAS, but this approach did not improve variance explained in any ancestry group (Table S1). Consistent with prior work, this suggests that applying ancestry-specific LD panels during scoring may not overcome portability issues from a fundamental lack of genetic diversity present in the original discovery sample.^19,49

Several additional factors may contribute to our observation that effect sizes for social risk variables were consistently smaller in the African-ancestry group compared to the European and admixed American groups. First, differential rates of misdiagnosis in Black Americans with BD (who are more likely to be misdiagnosed with schizophrenia)^27,28 may attenuate differences between affected individuals and control subjects in this group by including true individuals with BD among control subjects. Second, differential survey response patterns and the healthy volunteer bias inherent in AoU may operate differently across ancestry groups, as indicated by our comparison of survey completion rates across ancestries and case/control groups (Figure S6). Disentangling these contributions remains challenging, underscoring the need to investigate how structural factors influence diagnostic accuracy, genetic signals, and the measurement of social risk in diverse cohorts.

This disparity in predictive performance across ancestry groups has direct implications for health equity: if studies increasingly prioritize the recruitment of under-sampled populations, the clinical implementation of PRS stands to become more equitable, though disparities will remain until this progress is realized at scale. Many groups are already working to expand genetic diversity and equity to non-European and otherwise under-sampled populations, including the Latin American Genomics Consortium, the Ancestral Populations Network, the Nigerian 100K Genome Project, and the Sangre por Salud Biobank, among others.^{92,93,94,95,96,97} Crucially, these efforts must be conducted in partnership with represented communities and with attention to historical harms and the ethical obligations of genomic research in underrepresented populations, prioritizing community trust, partnership, and safety.^{98,99,100,101}

Study limitations

We acknowledge several limitations of the present study. First, because the AoU cohort is not a random population sample and contains unmodeled patterns of missingness in EHR and survey data, our observed effect sizes should be interpreted as associations within this specific analytic sample rather than as direct estimates of population-level risk.¹⁰² The cohort is subject to “healthy volunteer” bias, potentially narrowing the range of social adversity observed in our cohort. This bias is likely compounded in our analyses, which further restrict to the subset of participants who completed optional SDOH surveys.^102,103 However, our sensitivity analyses examining potential selection bias due to differential survey completion (Figure S13) and potential confounding or effect modification by level of EHR utilization (Figure S14) suggested that the observed social risk associations were robust to the patterns of missingness and diagnostic opportunity present within this cohort.

Second, reliably disentangling temporal relationships between social risk measures and BD onset remains a critical challenge for this field. Several of the social risk measures examined here, including perceived stress and CRMN, may both influence and be influenced by BD, complicating causal interpretations. ACE is unique in that it captures retrospective reports of childhood experiences that typically precede adult BD onset. This aspect of temporality supports a somewhat more directional interpretation than is possible for the other measures, but it is important to note that a causal interpretation cannot be assumed. Still, the nearly 3-fold OR observed for ACE in the European-ancestry group is consistent with a growing body of literature linking childhood maltreatment to earlier BD onset and greater illness severity.^21,26 Studies that utilize longitudinal cohort designs with repeated social risk measurements will be essential for establishing temporal relationships among these exposures.

Finally, with respect to phenotyping, BD is characterized by substantial symptomatic heterogeneity, encompassing manic, depressive, and mixed episodes with variable severity and course. Our EHR-based strategy does not distinguish between these clinical presentations, and it is plausible that genetic and social risk factors differentially influence specific symptom dimensions. Stricter case definitions could improve specificity but would further reduce the already limited sample sizes, particularly for non-European-ancestry groups. Future studies incorporating symptom-level phenotyping or clinical subtyping may reveal more nuanced patterns of gene-environment interactions, though larger, more ancestrally diverse, and more deeply phenotyped cohorts are needed to achieve this.

Conclusions

Our results highlight the importance of holistic models of psychiatric epidemiology that integrate genetic risk with the social and environmental context that may shape disease manifestation. The consistent and large effect sizes observed for social risk factors, which often exceeded those for genetic risk, suggest that precision psychiatry frameworks that fail to include social or environmental risk information provide an incomplete picture of individual-level risk. Importantly, many of the social risk factors examined here are modifiable or addressable through targeted or community-level interventions, unlike genetic liability. Joint genetic and social risk models may contribute to a more comprehensive understanding of individual-level risk, which could inform targeted prevention efforts and facilitate earlier intervention. Future work should expand these analyses to larger and more diverse cohorts, incorporate additional modifiable risk factors, and evaluate whether combined genetic-social risk models improve clinical outcomes. Together, these findings support continued development of integrated frameworks that combine genetic liability with social context to more accurately and equitably characterize psychiatric risk.

Data and code availability

This study used data from the AoU Research Program’s Controlled Tier Dataset v.8, available to authorized users on the Researcher Workbench. The BD GWAS summary statistics used to construct the polygenic scores are publicly available from the PGC via Figshare: https://doi.org/10.6084/m9.figshare.27216117 (O’Connell et al.¹²; v.2, deposited January 31, 2025) and are subject to the PGC data-use terms. All original analysis code is publicly available at https://github.com/thelovelab/bipolarPRSAOU and archived on Zenodo at https://doi.org/10.5281/zenodo.21775289.

Acknowledgments

We gratefully acknowledge AoU participants for their contributions, without whom this research would not have been possible. We also thank the National Institutes of Health’s AoU Research Program for making available the participant data examined in this study. We thank the Bipolar Disorder Working Group of the PGC and its contributing cohorts and participants for generating and making available the bipolar disorder GWAS summary statistics used to construct the polygenic scores in this study. This work was supported by the National Science Foundation Graduate Research Fellowship Program under grant no. DGE-2439854 awarded to R.R.S. and M.R.H. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the National Institutes of Health.

Author contributions

R.R.S. conceived the study, curated and analyzed the data, secured funding, created all figures, and wrote the manuscript. M.R.H., R.G.M., L.G., and L.M.R. provided relevant expertise on the study design and reviewed the manuscript. M.I.L. co-conceived the study, supervised the work, and reviewed the manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used Claude to organize ideas and format the manuscript. The authors reviewed and edited all AI-generated content and take full responsibility for the content of the publication.

Web resources

• All of Us Researcher Workbench, https://www.researchallofus.org/data-tools/workbench/

• GitHub, https://github.com/thelovelab/bipolarPRSAOU

• OMIM, https://www.omim.org/

• Psychiatric Genomics Consortium, https://pgc.unc.edu/for-researchers/download-results/