Person-level contributions of bipolar polygenic risk score to the prediction of new-onset bipolar disorder in at-risk offspring

Abstract

Background:

Previous work indicates that polygenic risk scores (PRS) for bipolar disorder (BD) are elevated in adults and youth with BD, but whether BD-PRS can inform person-level diagnostic prediction is unknown. Here, we test whether BD-PRS improves performance of a previously published risk calculator (RC) for BD.

Methods:

156 parents with BD-I/II and their offspring ages 6–18 were recruited and evaluated with standardized diagnostic assessments every two years for >12 years. DNA was extracted from saliva samples, genotyping performed, and BD-PRS calculated based on a 2021 meta-analysis. Using a bootstrapped and cross-validated penalized Cox regression, we assessed whether BD-PRS (alone and interacting with clinical variables) improved RC performance.

Results:

Of 227 offspring, 38 developed BD during follow-up. The penalized regression selected BD-PRS and interactions between BD-PRS and parental age at mood disorder onset (AAO), depression, and anxiety. The resulting RC discriminated offspring who developed BD (vs. those that did not) with good accuracy (AUC = 0.81); removing BD-PRS and its interaction terms was associated with a significant decrement to the AUC (decrement = 0.07, p = 0.039). Further exploration of selected interaction terms indicated that all were significant (p-values<0.02), indicating that BD-PRS has a larger effect on the outcome in offspring with depression and anxiety, whose affected parent had a younger AAO.

Conclusions:

The addition of BD-PRS to clinical/demographic predictors in the RC significantly improved its accuracy. BD-PRS predicted BD on the person-level, particularly in offspring of parents with earlier AAO who already had symptoms of anxiety and depression at intake.

Bipolar Disorder (BD) is a serious mood disorder that is associated with impairment in psychosocial function, substance abuse, and suicide. Early diagnosis and treatment are key, but often treatment delays are on the order of a decade or more (Joslyn et al., 2016; Dagani et al., 2017). Such delays frequently coincide with late adolescence and early adulthood, important developmental periods during which milestones may not be reached if BD is undiagnosed or poorly treated. Furthermore, BD is often preceded by months to years of subthreshold mood symptoms (Van Meter et al., 2016; Hafeman et al., 2016; Duffy et al., 2019), which can be impairing and interfere with developmental milestones. Improved prediction of who will go on to develop BD can optimally inform clinical decision-making and open avenues for early intervention (e.g. (Miklowitz et al., 2020)).

BD runs in families with estimates of 70–90 % heritability (Smoller and Finn, 2003; Gordovez and McMahon, 2020). Work over the past decade indicates that this risk is not inherited through simple Mendelian transmission, but rather via numerous variants with small effect size, as identified via Genome Wide Association Studies (GWAS). Approximately 25 % of heritability has been attributed to measured genetic variance (Gordovez and McMahon, 2020). Using results of a meta-analysis of GWAS studies, polygenic risk scores (PRS) for BD and other disorders and traits have been constructed (Mullins et al., 2021). Previous work, including our own, has found that high PRS for BD (BD-PRS) is associated with risk of BD in adults and youth (Birmaher et al., 2022; Boies et al., 2018), transition from depression to BD in adults (Musliner et al., 2020), and development of BD in offspring of parents with BD (Birmaher et al., 2022; Fullerton et al., 2015).

While BD-PRS is elevated, on average, in individuals with and at risk for BD, it currently has limited utility at the individual level, at least when used alone. Across psychiatric disorders, relevant PRSs only account for 2–11 % of the variance, depending upon disorder (Murray et al., 2021). Using metrics such as positive and negative predictive values (PPV and NPV, respectively), previous studies show that PRSs have limited benefit when applied alone to predict psychiatric or medical illness, particularly in the absence of other risk factors (e.g. (Fullerton and Nurnberger, 2019; Groenendyk et al., 2022)). However, it is possible that when added to other clinical factors, PRS may improve risk prediction. For medical disorders, it has been noted that PRS is not strongly associated with other clinical and demographic predictors (Lambert et al., 2019); thus, even a small effect of PRS can lead to improved risk prediction, above and beyond existing factors. There is also evidence across disorders, including BD, of interaction between PRS and other risk factors (e.g. clinical, environmental) to cause the outcome of interest, such that high PRS may increase the effect of other risk and vice-versa (Stocker et al., 2023; Hindy et al., 2018; Wilcox et al., 2017; Tamman et al., 2024). Recent work indicates that PRS for schizophrenia moderates the effect of genetic rare variants, with clinically meaningful increases in PPVs for 22q11.2 deletion syndrome (Davies et al., 2020). The clinical utility of PRS for a variety of medical illnesses and, more recently, psychosis, is starting to be explored (Torkamani et al., 2018; Perkins et al., 2020). A recent analysis of the PRS for psychosis found that it improved risk prediction in clinical high-risk youth (Perkins et al., 2020).

Here, we build on a recent analysis of PRS in the Pittsburgh Bipolar Offspring Study (BIOS) that found BD-PRS to be associated with BD in both the parents and offspring (Birmaher et al., 2022). We previously developed a risk calculator (RC) based on clinical factors that predicted, with good accuracy (AUC = 0.76) whether an offspring of a parent with BD would develop new-onset BD (Hafeman et al., 2017). The goal of the current paper is to assess the degree to which BD-PRS, alone and interacting with other clinical measures, improves risk prediction of new-onset BD. These analyses are critical to assessing the degree to which this biological marker may improve person-level prognostication.

1. Methods

1.1. Sample

Methods for the BIOS sample have been described previously (Hafeman et al., 2016; Axelson et al., 2015). Briefly, adults with DSM-IV BD-I/II who had biological offspring ages 6–18 years old were recruited between September 2001 and July 2008, primarily through advertisements. Parents with schizophrenia, autism, intellectual disability (IQ < 70), or mood disorders secondary to other conditions were excluded. All offspring (6–18 years old) of eligible parents were included in the study, except for offspring with intellectual disability (IQ < 70) or other condition(s) that interfered with evaluation. Parents and offspring were followed approximately every two years. Community controls were also included in this study but are not part of the current analysis. Informed consent was obtained from parents and adult offspring; assent was obtained from offspring <18 years old. All study procedures were approved by University of Pittsburgh’s Institutional Review Board.

1.2. Procedures

The Structured Clinical Interview for DSM-IV (SCID) (First and Gibbon, 2004) was utilized to assess lifetime diagnoses in BD parents and a subgroup of biological co-parents (31 %). The Family history Research Diagnostic Criteria (Andreasen et al., 1977) was used to assess family psychiatric history (including co-parents unavailable for interview). Parental age at mood disorder onset (AAO) was assessed via the SCID, and defined as the age of the first major mood episode (major depression or mania) in the proband with BD.

At intake, and approximately every two years across follow-up, offspring disorders (including BD) were assessed using the Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children–Present and Lifetime Version (Kaufman et al., 1997) (KSADS-PL) until age 19, and the SCID thereafter. Specific criteria for BD not-otherwise-specified (BD-NOS) from the Course and Outcome of Bipolar Youth (COBY) study were used (Axelson et al., 2011) (eMethods). All offspring assessments were completed by trained interviewers who were blind to parental diagnosis and then presented to a child psychiatrist or psychologist (also blind to parental diagnosis) for confirmation. The kappa for diagnostic reliability was ≥0.80 across disorders (Hafeman et al., 2016; Axelson et al., 2015; Birmaher et al., 2009).

At intake and follow-up, the KSADS Mania Rating Scale (KMRS) and depression items from the KSADS (KDRS) present version were used to assess mood disorder symptoms during the worst week in the past month. Parents and offspring also completed dimensional assessments of mood lability (Child Affective Lability Scale; CALS) (Gerson et al., 1996) and anxiety (Screen for Child Adolescent Related Disorders; SCARED) (Birmaher et al., 1999). The Child Global Assessment Scale (CGAS) was utilized to rate psychosocial function. In a previous publication (Hafeman et al., 2017), we constructed a RC that used these clinical measures (i.e., KMRS, KSADS, CALS – child report, SCARED – child report, and CGAS), in combination with age and parental AAO, to predict new-onset BD in this sample. Socioeconomic status was determined using the Hollingshead scale (Hollingshead, 1975).

1.3. Genotyping and polygenic risk score

Methods for DNA extraction and genotyping have been previously described (Birmaher et al., 2022). Quality control procedures were followed, as described in (Birmaher et al., 2022); participants with low-quality data were excluded. The BD-PRS was constructed using PRSice-2 (Choi and O’Reilly, 2019) and the results of meta-analyses of GWAS of BD, using the p-value threshold that maximally captured variance in the discovery GWAS sample (BD = 0.05) (Mullins et al., 2021). To construct the PRS, the contribution of each allele was weighted by the effect size of its association with each phenotype in the reference sample GWAS. Since the existing discovery GWAS samples included only individuals of European ancestry (Mullins et al., 2021), we only included offspring of this ancestry for this analysis. While there have been recent efforts to develop a trans-ancestry BD-PRS, these efforts are still in early stages and explain <2 % of the variance (Bigdeli et al., 2020); thus, for the purposes of person-level prediction, such an approach would unfortunately be premature.

1.4. Statistical methods

We first tested whether BD-PRS, both alone and interacting with clinical and demographic characteristics, significantly improved our previously published RC’s ability to predict new-onset BD. Given that PRS does not fluctuate over time, we did not use baseline-resetting Cox regression as we did previously (Hafeman et al., 2017), but instead focused on baseline predictors of new-onset BD evaluating risk predictions at the median age of follow-up before BD diagnosis or right-censoring (21 years old) by estimating time-dependent area under the receiver operating characteristic curve (AUC). Given the large number of predictors in a model including BD-PRS and its possible interactions, we used a penalized (LASSO) Cox proportional hazards regression to select variables that substantially contributed to the model. We used three-fold cross-validation to train and test the RC such that within each iteration, all model-tuning/predictor selection was conducted in the training set, and AUC was computed using risk estimates in the testing set. We then bootstrapped this process with 1000 resamples to estimate the stability of LASSO predictor selections (i.e., percentage of bootstrap iterations in which each predictor was selected), estimate confidence intervals for AUCs, and test the statistical significance of AUC decrements after sequentially removing predictors to evaluate predictor importance. Given that our sample included siblings, and BD-PRS is highly correlated across siblings, all siblings were confined to the same folds during cross-validation. Further, to ensure balance in fold size and BD rates, we designed an algorithm that first generated 10,000-fold randomizations, observed only randomizations in which fold sizes were within 1 participant of one another (75–76 participants per fold), and selected the randomization with optimal balance in BD rates (16–17 % within-fold). We fit a model with baseline clinical characteristics and BD-PRS (plus interactions), and next a model without BD-PRS, assessing the decrement in AUC. We also assessed the percent of variance in the dichotomous BD outcome explained by each model using Nagelkerke’s pseudo-R². Finally, we tested model calibration via Hosmer-Lemeshow test and by plotting and comparing observed vs. predicted BD risk.

To assess BD-PRS as a predictive tool, we next tested three thresholds for a “positive” test: top 25 %, top 33.3 %, and top 50 %. While previous PRS studies have used a wider range of thresholds (e.g., top 5 % or 10 %), given our sample size, such imbalanced thresholding led to insufficient numbers in each group. We assessed sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for BD-PRS and the demographic/clinical variables that it interacted with, both alone and in combination. We also calculated the “false omission rate”, which is simply 1 minus the NPV and indicates the probability of BD onset in an individual with a negative test (i.e., BD-PRS below threshold value).

We finally tested interactions between BD-PRS and LASSO-selected clinical/demographic variables to predict new-onset BD in the offspring. While primary analyses utilized continuous variables, variables were dichotomized at the median for purposes of Kaplan-Meier survival curve visualization. Supplemental analyses were conducted to further probe positive findings (eMethods).

2. Results

2.1. Sample

The final sample consisted of 227 offspring of 146 parents with BD. Of these offspring, 38 developed BD over follow-up (from 33 different families). Table 1 shows the demographic and clinical characteristics of this sample according to incidence of BD. Participants were a median age of 10.5 years old at intake and were followed an average of 7.0 times over a median of 14.2 years.

Table 1.

Sample characteristics.

Demographics at Intake	Full Sample (n = 227)	Developed BD (n = 38)	Right-Censored (n = 189)	Test Stat	p-value
Age	10.5 (4.7)	9.8 (4.5)	10.7 (4.8)	t = 1.06	0.2913
Male, % (n)	48.5 (110)	47.4 (18)	48.7 (92)	χ2 = 0.02	0.8829
SES	37.3 (14.2)	37.5 (16.0)	37.3 (13.9)	t = 0.09	0.9294
Maternal Age at Offspring’s Birth	28.5 (5.7)	28.1 (5.6)	28.6 (5.7)	t = 0.43	0.6694
Paternal Age of Offspring’s Birth	30.9 (6.8)	31.5 (6.0)	30.8 (6.9)	t = 0.55	0.5802
Follow-up Statistics
Number of Assessments, median	7.0	7.0	7.0	t = 0.31	0.7586
Duration of Follow-up, median	14.2	13.6	14.2	t = 0.25	0.8012
Duration between Assessments, median	2.2	2.2	2.2	t = 0.27	0.7906

2.2. Risk calculator

The penalized Lasso regression selected the following predictors of new-onset BD: age, parental AAO, global function (CGAS), mood lability (CALS), BD-PRS and interactions between BD-PRS and parental AAO, depression (KDRS), and anxiety (SCARED) (Table 2). The resulting risk calculator (RC) predicted new-onset BD with good discrimination (AUC = 0.81; 95 % CI = 0.79–0.92). Removing BD-PRS and its interactions from the model resulted in an AUC decrement of 0.07 (p = 0.039), indicating that this variable contributed significantly to the prediction of new-onset BD above and beyond clinical predictors. Further, Nagelkerke’s pseudo-R² estimates indicated that the full model explained 19 % of the variance, whereas the reduced model explained only 16 % of the variance. Predicted and observed BD risk were consistent through the range of risk scores and did not significantly differ (Hosmer-Lemeshow χ² = 12.87, df = 8, p = 0.12), indicating no evidence of miscalibration (eFig. 2). However, inspection of the calibration curve’s right tail does indicate that the model slightly underpredicted risk among cases at the highest risk of developing BD.

Table 2.

Risk calculator to predict new-onset bipolar disorder.

Baseline Predictors	LASSO Hazard Ratio	% of Bootstrap Iterations Retained	AUC Decrement if Removed	Bootstrap p-value
Age	0.71	99 %	0.09	0.004
Parent Mood Onset Age	0.73	99 %	0.07	0.037
CALS (child-report)	1.43	98 %	0.05	0.133
CGAS	0.81	97 %	0.02	0.222
BD-PRS*	1.04	93 %
BD-PRS × Parent Mood Onset Age	0.89	95 %
BD-PRS × DRS	1.30	95 %	0.07	0.039
BD-PRS × SCARED (child-report)	1.13	95 %

Significant (p<.05) values are shown in bold.

^*Any instance of PRS (linear or multiplicative effects) was selected by 99.9 % of LASSO bootstrap iterations.

To assess the clinical significance of these variables, we computed the predicted risk score for new-onset BD according to BD-PRS and the clinical variables with which it interacts (i.e., parental AAO, depressive symptoms, anxiety). Offspring with young parental AAO (25th percentile), anxiety and depressive symptoms (75th percentile), and high BD-PRS (75th percentile) have an estimated 29 % chance of developing new-onset BD (Fig. 1).

Fig. 1.

Predicted risk for bipolar disorder onset by the median age at BD diagnosis/right-censoring (21 years old), stratified according to Lasso-selected variables*.

*For illustrative purposes, estimated risk is calculated for low (25th percentile) versus high (75th percentile) values of clinical symptoms and BD-PRS; and late (75th percentile) versus early (25th percentile) parental age at mood disorder onset.

2.3. Diagnostic assessment

Across the tested thresholds, the optimal threshold was 50 %; this was associated with a PPV of 23 % and NPV of 89 % (Table 3). This was similar to the PPV/NPV with parental AAO alone. Combining these two factors, the PPV increased to 34 %, while maintaining a NPV of 89 % (equivalent to a false omission rate of 11 %). Further adding anxiety and/or depressive symptoms, the PPV increased to 39 %, while the NPV was 89 % (false omission rate of 11 %). This means that 39 % of those who meet this cut-off will develop BD, while only 11 % of those who do not meet the cut-off will develop BD.

Table 3.

Performance metrics for diagnostic assessment.

Decision Rule	Sensitivity	Specificity	Positive Predictive Value	Negative Predictive Value	False Omission Rate
Upper 25 %	0.32	0.76	0.21	0.85	0.15
Upper 33.3 %	0.39	0.68	0.20	0.85	0.15
Upper 50 %	0.68	0.54	0.23	0.89	0.11
Early parental onset (< 18)	0.66	0.52	0.22	0.88	0.12
Elevated (≥ median) anxiety or depression	0.87	0.31	0.20	0.92	0.08
Upper 50 % BD-PRS + Early parental onset	0.50	0.80	0.34	0.89	0.11
Upper 50 % BD-PRS + anxiety or depression	0.61	0.70	0.29	0.90	0.10
Upper 50 % BD-PRS + Early parental onset + anxiety or depression	0.45	0.86	0.39	0.89	0.11

2.4. Univariate effect of BD-PRS and LASSO-selected interaction effects

Both BD-PRS and parental AAO independently predicted BD in the offspring (parental AAO χ² = 4.83, p = 0.02; BD-PRS: χ² = 6.01, p = 0.01). LASSO-selected interactions between BD-PRS and parental AAO, depression, and anxiety were also significant (Fig. 2a–c; BD-PRS × parental AAO Log-Rank χ² = 18.46, p = 0.0004; BD-PRS × depressive symptoms Log-Rank χ² = 10.23, p = 0.02; BD-PRS × anxiety symptoms Log-Rank χ² = 12.24, p = 0.007).

Fig. 2.

Cumulative Incidence of BD stratified according to BD-PRS (median split) interactions selected by LASSO. a. Depression Score (median split). b. Anxiety Score (median split). c. Parental AAO (<18 vs. ≥18).

2.5. Supplemental analysis: interaction between BD-PRS and parental AAO

Further analyses were conducted to probe the significant interaction between BD-PRS and parental AAO. First, AAO and BD-PRS were not associated in parents or, separately, in the offspring (p-values>0.3). Second, this association was not explained by tested confounds (see eMethods). While earlier parental AAO was associated with more parental anxiety and living with only one parent, the interaction between parental AAO and BD-PRS remained significant even after adjustment for these variables (eTable 1). Second, we found a modest correlation between parental and offspring AAO (r = 0.34, p = 0.04), indicating that offspring of parents with older AAO may not have aged into their highest risk period. To probe this possibility, we tested whether a similar interaction (BD-PRS x parental AAO) would predict mood disorder onset in offspring, since this outcome often precedes BD. We found that this interaction did not significantly predict offspring mood disorders (eFig. 1).

3. Discussion

In this analysis, we extended previous work showing elevated BD-PRS in offspring of parents with BD and found that BD-PRS, interacting with early parental AAO (i.e., <18 years old), anxiety, and depression, adds to a previously published RC to predict new-onset BD. Specifically, including BD and its interactions improved the RC AUC from 0.74 to 0.81, a statistically and clinically relevant difference. Furthermore, the combination of elevated BD-PRS (above the median), early parental AAO, and anxiety or depression led to clinically relevant prediction: 39 % of offspring with these risk factors will go onto develop BD (PPV), compared to only 11 % of offspring without these factors (false omission rate). Thus, despite that BD-PRS explains <10 % of the variance in our sample (and in other studies), these findings indicate that, combined with other important predictors, BD-PRS may enhance person-level prediction.

The current analyses focus primarily on the impact of BD-PRS on person-level metrics, as opposed to standard models of statistical significance. Ultimately, these person-level metrics are critical to the evaluation of clinical utility. An AUC of 0.80 is considered “good”, and standard RCs utilized in other areas of medicine have AUCs in this range, e.g., AUC = 0.75–0.82 for cardiovascular disease risk (Badawy et al., 2022) and AUC = 0.80 for a widely used RC for neonatal sepsis (Puopolo et al., 2011; Kuzniewicz et al., 2016). This finding, if replicated, indicates that BD-PRS may contribute to person-level prediction of new-onset BD. While BD-PRS does not effectively distinguish on its own who will develop BD, in combination with other factors, it contributes clinically and statistically to improved prediction. This is similar to previous reports in cardiovascular disease (O’Sullivan et al., 2022) and psychosis (Perkins et al., 2020), which found that adding a relevant PRS to clinical variables improved risk prediction, despite limited variance explained alone. Of note, this RC only included demographic and clinical predictors from our previous RC (Hafeman et al., 2017) to test specifically whether BD-PRS improved risk prediction; while other predictors (e.g. socioeconomic status) may also improve the AUC, such exploration is not within the aims of this paper and could lead to model overfitting.

Interestingly, BD-PRS interacted significantly with anxiety, depression, and parental AAO of mood disorders to predict new-onset BD. The interaction between BD-PRS and anxiety/depression indicates that these non-specific clinical symptoms are more likely to predict BD onset in the context of elevated polygenic risk, and vice versa. Similarly, we found that BD-PRS had a greater effect in offspring whose affected parent developed a mood disorder at a younger age. We and others have found that parental AAO of mood disorder is an important predictor of BD onset in the at-risk offspring (Hafeman et al., 2016; Preisig et al., 2016), and parental AAO was the most influential variable in our previously developed RC (with a 0.05 decrement in AUC if removed) (Hafeman et al., 2017).

We further assessed the possible reasons for the observed interaction between BD-PRS and parental AAO. First, it is possible that early onset of BD is associated with higher genetic risk. However, we did not find an association between AAO and BD-PRS within the parents or, separately, within the offspring. This is in line with a large previous study in adults with BD that found higher BD-PRS to be unrelated to earlier AAO (Kalman et al., 2019). Second, it is possible that parental AAO is associated with other demographic or clinical factors that interact with BD-PRS, which in turn explains the observed finding. While we found demographic and clinical correlates of early parental AAO (i.e., living with one parent, co-morbid anxiety), these factors did not explain the observed association. Thus, the most likely possibility is that the offspring of parents with later AAO in our study have not passed through the full risk period. Indeed, we find that parental and offspring AAO are correlated; to our knowledge, this relationship has not been previously assessed in the literature. Furthermore, a similar interaction is not present with mood disorders, which often precede a diagnosis of BD. Interestingly, while not significant, the risk in high BD-PRS/late parental AAO does appear to increase at the older ages of follow-up (eFig. 1).

Given the possibility of clinical utility, this leads to several important questions about clinical application, many of which require further investigation. First, prior to wide adaptation of BD-PRS utilization (and usage of the RC, in particular), it is important to weigh the risks and benefits for patients (and families) to know this information. This includes whether patients benefit from having more prognostic information (which is, by definition, imperfect); and, relatedly, whether there are actionable steps to take based on the outcome (Latham et al., 2021). Second, it is important to acknowledge that current analyses were in offspring of European ancestry, given that this is the population within which the BD-PRS was developed, and population stratification limits its applicability to individuals not of European ancestry (Lewis and Green, 2021). Thus, there is the very real possibility that clinical tools described here, if useful, could exacerbate existing racial inequities. Ongoing efforts to scale up GWAS in non-European populations (O’Connell, 2021) and/or methodological advances (Ruan et al., 2022) are critical to the development of improved BD-PRS in diverse samples. Third, such genetic information will need to be protected, given the possibility that such prognostic information may be used in ways that are not helpful to the individual (e.g., by insurance companies to exclude coverage). While these considerations are outside the scope of the current paper, it is important to contextualize these findings, which may have clinical application, in this broader framework and, with that, suggest caution.

In addition to these considerations, there are several limitations that should be considered when interpreting these results. First, we have not externally validated the current RC (including BD-PRS) in an independent sample. However, we used cross-validation to ensure that all model-tuning/predictor selection was conducted in the training set, and all predictions were evaluated in the testing set, thus protecting against overfitting. External validation in an independent sample will be an important future direction, particularly to assess whether observed interactions remain predictive. Second, while BIOS is the largest and longest running study of its kind, we still had a limited number of individuals with BD in this sample. Nevertheless, the effective sample size for this analysis is 126 (given 227 at-risk offspring, 38 of whom converted); this exceeds the minimum recommended sample size of 100 for PRS studies (Choi et al., 2020). Third, this sample was recruited based on a family history of BD. While this design maximizes the frequency of new-onset BD, results cannot be applied to youth without a clear family history of BD. This is a critical consideration, since prevalence in the population affects statistics such as PPV and false omission rate, as well as calibration of the RC (Abu-Akel et al., 2018). Furthermore, youth that might benefit most from the RC may not have a well-documented family history, e.g. because of lack of treatment-seeking or disconnection from family of origin. Thus, a critical future direction will be to assess the degree to which BD-PRS improves risk prediction in a clinically at-risk sample (e.g (Van Meter et al., 2021),.).

4. Conclusion

Despite these limitations and important considerations for future clinical applications, these findings indicate that BD-PRS may improve person-level risk prediction. Combined with parental AAO, anxiety, and depression, BD-PRS leads to a clinically and statistically significant improvement in RC performance and diagnostic prediction. Importantly, the larger effect of BD-PRS in parental AAO may be related to incomplete ascertainment through the BD risk period in offspring of parents with older AAO. Prior to clinical integration of these findings, they should be externally validated in an independent sample, and the clinical and ethical implications of such usage should be fully appraised.

Data availability

Data are available through the National Data Archive and by request from the authors.