Do Polygenic Scores Inform Psychiatric Disease Risk After Considering Family History?

Psychiatric disorders are heritable, and thus both family history and an individual’s genetic makeup are related to their risk of illness. It has long been known that family history can help predict psychiatric outcomes (1), yet individual prediction of risk has remained challenging. In some circumstances, family history can predict substantial increase in risk; for example, the risk of schizophrenia is increased about eightfold for those who have a first-degree relative with the illness (2). However, most people do not have a first-degree relative with any given illness, and for them family history is not very informative. Even among those with a family history, most people will remain unaffected.

Recently, leveraging the results of large genome-wide association studies of psychiatric traits has enabled the computation of individual polygenic scores (PGSs) for those traits from a person’s genome (3). PGSs, often referred to as polygenic risk scores when they are used to quantify risk of a particular disorder, are currently not predictive enough to be used in most clinical settings (3), but show consistent associations with the target traits in research studies. Given sufficient knowledge about effects of specific genetic variants on disease risk, a direct measurement of an individual’s genome should be more informative than family history (which ignores which parental alleles were transmitted to the offspring) regarding the genetic contribution to a person’s phenotype. Thus, it is pertinent to ask, Do current PGSs provide information on disease risk beyond what can be inferred from family history?

In this issue, Zwicker et al. (4) report on a study that explored this question in the context of onset of major mood and psychotic disorders (major depressive disorder, bipolar disorder, and schizophrenia). The outcome was measured prospectively in a cohort ranging in age from 6 to 36 years at most recent follow-up. To predict the outcome, the authors considered a binary family history variable representing parental history of major mood and psychotic disorders, as well as PGSs for depression, bipolar disorder, anxiety, attention deficit hyperactivity disorder, schizophrenia, “p factor” (an index of general genetic liability to psychopathology), neuroticism, and subjective well-being. They observed that family history as well as many of the PGSs were associated with onset of major mood and psychotic disorders, but only the PGSs for neuroticism and subjective well-being remained statistically significant predictors of the outcome after accounting for family history.

As the authors remarked, this may be partly due to the transdiagnostic nature of these phenotypes. In fact, of all the considered PGSs, the neuroticism PGS showed the strongest association with the clinical outcome, even before adjustment for family history. It is also worth noting that (unsurprisingly) the family history measure reflected major mood and psychotic disorders, rather than neuroticism or subjective well-being. It may not be feasible, or constructive, to measure family history of more general psychological or personality traits, but subclinical psychopathology and related measures can be directly assessed in patients and can be important predictors of subsequent development of psychiatric disorders (5). Thus, it would be interesting to evaluate whether, for example, a neuroticism PGS would be a significant predictor of psychiatric disorder outcomes after accounting for measured neuroticism in at-risk young people.

Zwicker et al. used a simple binary definition of family history based on the presence of a major mood or psychotic disorder in a biological parent, which does not account for disease status of other relatives. This single family-history variable also made no distinction between different mood or psychotic disorders, meaning that, for example, parental history of major depressive disorder, was treated equally with parental history of schizophrenia. While a more nuanced or even quantitative measure of family history could be an even better predictor of risk, possibly further reducing the relative importance of PGSs, it is difficult to collect more extensive family history accurately. Moreover, reduction of family sizes has made extended family history measures less informative. Furthermore, the direct measurement of family history in the Zwicker et al. study, which was largely through diagnostic interviews of parents, likely provided a much more accurate predictor of risk than could be achieved in clinical practice, where family history collection usually relies on patient self-report. Thus, it is reasonable to expect that in clinical settings, family history would contribute less to prediction of risk, and therefore the relative contribution of PGSs after accounting for family history would be greater.

In the analysis presented by Zwicker et al., aside from the PGS for subjective well-being, all PGS effects were attenuated after accounting for family history, but not fully. For example, the hazard ratio for the schizophrenia PGS effect was 1.10 (95% CI=1.00–1.22) with family history adjustment, as opposed to 1.15 (95% CI=1.04–1.26) without family history adjustment. Given the observed PGS effects and their confidence intervals, it is reasonable to expect that in an analysis of a larger sample, family-history-adjusted PGS effects would remain statistically significant, even for current PGSs with their limited predictive value. In fact, other psychiatric studies have already found that specific-disorder PGSs are significant predictors over and above family history, or within cohorts with family history (6, 7). Outside of psychiatry, a recent large study of 24 common diseases in the FinnGen sample found that family history and PGSs provide complementary information on inherited disease susceptibility, with both making independent contributions to risk prediction (8).

Although some studies have already demonstrated statistical significance of PGSs as predictors of psychiatric outcomes after accounting for family history, statistical significance does not imply clinical significance. In the Zwicker et al. study, neither family history nor the different PGSs contributed greatly to overall risk prediction, with each explaining less than 5% of the variance in time until diagnosis, which again emphasizes the limited clinical utility of current PGSs. More importantly, given the poor performance of current PGSs in populations underrepresented in research, until genetic research becomes more representative of ancestral diversity of patients, application of PGSs in clinical settings would only exacerbate current health disparities (9). Finally, to be useful in clinical settings, PGSs must be converted to absolute or relative risks. Although methods were recently proposed to achieve this (10), they also rely on input parameters that are representative of the population in which the method is applied. Thus, as the research community works toward more predictive PGSs and develops tools for using them in clinical settings, we must ensure that this research is conducted in diverse research cohorts to allow for equitable application of research findings across populations.

Psychiatric disorders are only partially heritable. Therefore, ultimately PGSs—and for some conditions, specific genetic risk factors, such as copy number variants—will need to be combined with other relevant information, including environmental factors and social determinants of health, to enable more precise risk prediction. Because family history reflects not only shared genetics but typically also shared environment, it will remain a relevant risk predictor for many outcomes even as PGSs become more predictive. Some studies have begun exploring the contribution of PGSs to risk prediction based on multiple predictors; for example, in a high-risk population, Perkins et al. (11) found that a schizophrenia PGS showed promise in improving risk prediction of a psychosis risk calculator that included certain environmental and clinical risk factors in addition to family history of psychosis. Studies have also shown that incorporating multiple PGSs simultaneously can improve prediction of psychiatric outcomes (12). More research should be directed toward development of improved predictive models that capture the full range of environmental and genetic risk factors, include multiple PGSs, and consider potential interactions among risk factors.

It is also important to consider prediction of outcomes beyond broad diagnostic categories or their groupings. In the Zwicker et al. study, which aimed to identify stable predictors across diagnostic categories, little was gained from considering PGSs after accounting for family history. However, many studies have demonstrated that beyond diagnostic categories, PGSs are often associated with specific subphenotypes within diagnostic groups (13). Thus, PGSs may ultimately prove helpful for prediction of more nuanced clinical outcomes, including illness trajectories, predominant symptoms, and treatment outcomes, where it becomes increasingly difficult to collect relevant family history.

The relative value of family history data and PGSs as predictors of risk depends on several factors, including the genetic architecture of the outcome being predicted, the ease and feasibility of collecting precise family history data, the availability of genetic data, and the precision of PGSs derived from these data. Critically, PGSs need to be developed using data from more diverse patient populations to ensure that they provide valuable information for all patients. Other risk factors, including environmental and social determinants of health, and in some cases specific genetic variants, will have to be incorporated in the predictive models (14). As genetic data continue to become more readily available and the precision of PGSs further improves, the relative contribution of PGSs will increase, but optimal prediction will always require a holistic approach.