Polygenic risk scores for predicting outcomes and treatment response in psychiatry: hope or hype?

Abstract

Over the last years, the decreased costs and enhanced accessibility to large genome-wide association studies datasets have laid the foundations for the development of polygenic risk scores (PRSs). A PRS is calculated on the weighted sum of single nucleotide polymorphisms and measures the individual genetic predisposition to develop a certain phenotype. An increasing number of studies have attempted to utilize the PRSs for risk stratification and prognostic evaluation. The present narrative review aims to discuss the potential clinical utility of PRSs in predicting outcomes and treatment response in psychiatry. After summarizing the evidence on major mental disorders, we have discussed the advantages and limitations of currently available PRSs. Although PRSs represent stable trait features with a normal distribution in the general population and can be relatively easily calculated in terms of time and costs, their real-world applicability is reduced by several limitations, such as low predictive power and lack of population diversity. Even with the rapid expansion of the psychiatric genetic knowledge base, pure genetic prediction in clinical psychiatry appears to be out of reach in the near future. Therefore, combining genomic and exposomic vulnerabilities for mental disorders with a detailed clinical characterization is needed to personalize care.

Keywords:

• Polygenic risk score
• genetics
• schizophrenia
• depression
• environment
• prediction
• personalized medicine

Introduction

Decades of twin, family, and adoption studies have established that mental disorders aggregate in families and are substantially heritable (Smoller et al., 2019). Molecular genetics studies have been previously restricted to a limited set of ‘candidate genes’ that are selected based on biological hypotheses about putative etiopathology, with a vast overestimation of the effect size that such loci were likely to have. However, the genetics of mental disorders appear to be far more complex, as was first postulated more than 50 years ago: the polygenic theory of schizophrenia (Gottesman & Shields, 1967). According to this theory, each individual in the general population has varying degrees of quantifiable genetic and environmental liability for schizophrenia but develops schizophrenia only when the combined liability exceeds the threshold on the continuum. More recently, the common disease-common variant (CDCV) hypothesis argued that numerous common genetic variants of very small individual effects would substantially contribute to genetic susceptibility to common disorders (Reich & Lander, 2001).

In the context of a rapid progression of novel molecular, technological, and statistical methods, there has been a rapid flourishing of the ‘genome-wide association studies (GWAS)’ – aimed at scanning variations across the DNA of a person (the ‘genome’) and identifying genetic variations (i.e. single-nucleotide polymorphisms, SNPs) associated with a quantitative trait or a disease phenotype. Based upon the CDCV paradigm, the GWAS captures common genetic variants only, while it does not include rare genetic variants, such as copy number variations (CNVs), it is now established that they could play a considerable role in the susceptibility to mental disorders (Singh et al., 2022).

The GWAS have produced robust and replicable associations between DNA variations and a range of mental disorders, including more than 200 variants significantly associated with schizophrenia (Trubetskoy et al., 2022) and more than 100 with major depressive disorder (MDD) (Howard et al., 2019). Ultimately, the combination of ‘risk’ associated with individual variants into a single metric score has brought in a polygenic risk score (PRS). PRSs have been tested for a wide range of applications in psychiatry, including risk prediction, diagnosis, and prognosis (Lewis & Vassos, 2020). The present narrative review will specifically focus on the potential utility of PRS in predicting treatment response and outcomes in psychiatry. After providing a brief framework on the PRS construct, we will summarise the main studies conducted so far on three major mental disorders (i.e. psychosis, major depressive disorder [MDD], and bipolar disorder [BD]). Finally, we will discuss the limitations of currently available PRSs and provide hints for future research, such as the integration of genetic and environmental vulnerability for mental disorders with the clinical characterization of patients.

Polygenic risk score

PRSs represent measures of the genetic predisposition of an individual towards the development of a certain trait. PRSs are calculated by summing the weighted allele counts of independent SNPs associated with a trait or disorder in a GWAS. SNPs are selected based on their GWAS association statistical significance level (i.e. p-values).

The calculation of a PRS requires two key datasets: (1) discovery dataset, that is the GWAS, consisting of summary statistics of SNPs; and (2) target dataset, consisting of genotypes, and usually also phenotypes, in individuals from an independent sample. In the target sample, PRS analyses are performed by computing PRSs in all the target individuals, calculating associations between the PRSs and phenotypes/outcomes of interest, or predicting individuals’ risk of disease in clinical and non-clinical settings (Choi et al., 2020; Wand et al., 2021).

PRSs typically follow a Gaussian distribution at a population level, with the distributions of cases and controls overlapping substantially. Therefore, meaningful risk predictions can only be expected for extreme percentiles of the PRSs distributions. Established PRS software and accessible discovery GWAS summary statistics from collaborative groups such as the Psychiatric Genomics Consortium (PGC) are available. Moreover, genome-wide genotyping and calculation of PRSs are now relatively easy in terms of cost and time (Lewis & Vassos, 2020).

Models testing the association between PRSs and outcomes are typically adjusted for covariates, such as genetic principal components (PCs) to account for the confounding effect of ancestry. When reporting the results of a PRS analysis, rigorous standards should be followed ideally. Authors should define in detail the study type (e.g. development or validation), purpose (e.g. risk prediction vs. prognosis), and predicted outcomes (e.g. psychosis). The discovery of GWAS and the methods for PRS construction and estimation should be provided, along with the assessment of the risk model’s predictive ability, calibration, and discrimination described with effect size, variance explained (R²), reclassification indices, and metrics, such as sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Classification ability can be presented in terms of the area under the receiver operating characteristic (AUROC) or precision-recall curve (AUPRC), or the concordance statistic (C-index). Finally, a comprehensive interpretation of the models is essential along with a discussion of the limitations and the generalisability of the findings (Wand et al., 2021).

Utility of PRS in predicting outcomes and treatment response

A large portion of patients with mental disorders, such as psychosis, MDD, and BD, do not fully respond to treatment, undergo frequent relapses, or suffer from severe adverse events. The mechanisms underlying the inter- and intra-individual responses to treatment are mostly unclear. Genetics may likely play a crucial role in determining outcome trajectories. In this regard, PRSs could ideally help identify subgroups of patients. This would subsequently enable the implementation of stratification and personalized therapies, with improved efficacy and reduced side effects (Murray et al., 2021). In the following paragraphs, we will review the main studies conducted on the utility of PRS in predicting outcomes and treatment responses in major mental disorders. The key findings are summarised in Table 1.

Table 1. Summary of the key findings on the association between PRSs and outcomes or treatment response in major mental disorders.

Psychosis spectrum disorders

In a FEP cohort, PRS-SCZ did not predict symptom change over 20 years but was regarded as a predictor of shift from affective to non-affective psychosis with poor accuracy. PRS-SCZ did not perform better than clinical information in predicting proxies of poor outcome in two multi-ethnic cohorts. In a meta-analysis of FEP cohorts, higher PRS-SCZ significantly predicted worse 12-week symptom scores and patients with low PRS-SCZ were more likely to be treatment responders than those with high PRS-SCZ. However, results were not confirmed in all cohorts when each were analysed individually. Findings on the association between PRS-SCZ and treatment-resistance or clozapine use were contrasting. A PRS model for TRS was developed with detected heritability (1-4%) and correlation with history of clozapine use, although the variance explained was very low.

Major depressive disorder

PRS-SCZ, PRS-MDD, and PRS-BD did not predict symptoms improvement, response to antidepressants or esketamine, remission, or treatment resistance in individual studies or meta-analyses. Higher PRS-ADHD was associated with TRD but not with percentage change in depression score, treatment response, and symptom remission after 4-week treatment with esketamine. PRS for openness was positively associated with non-remission and non-response after antidepressant treatment. Higher PRSs for cardiometabolic disorders were associated with poorer response/remission. No association between PRS-MDD and ECT response were found, but higher PRS-SCZ was associated with symptoms decrease and remission.

Bipolar disorder

People with higher PRS-MDD were less likely to respond to lithium compared to patients with low polygenic load in the multi-ethnic and European samples but not in an Asian sample. Higher PRS-SCZ and PRS-MDD were associated with poorer lithium response, whereas PRS-BD had no association with treatment outcome. A combination of PRS-SCZ and PRS-MDD could improve response prediction. Genomic stratification slightly improved variance in a clinical model of treatment response.

Abbreviations. ADHD: attention deficit hyperactivity disorder; BD: bipolar disorder; ECT: electroconvulsive therapy; MDD: major depressive disorder; PRS: polygenic risk score; SCZ: schizophrenia; TRD: treatment-resistant depression; TRS: treatment-resistant schizophrenia.

Psychosis spectrum disorder

Psychosis spectrum disorder encompasses a heterogeneous phenotype with large variability in terms of course and outcomes, roughly with 30% of patients experiencing severe, chronic symptoms, 30% experiencing relapses, and 30% recovering completely (Harrow et al., 2005). Exploring the genetic substrate might thus be crucial to understanding predictors of clinical response and outcome (Zhang & Malhotra, 2018).

Longitudinal cohort studies could be particularly useful to understand long-term outcomes. A cohort study assessed 249 first-admission patients (85% of European ancestry) with psychosis over 20 years (Jonas et al., 2019). The authors found that PRS for schizophrenia (PRS-SCZ) did not predict changes in symptom severity over time. When PRS-SCZ scores were split into ‘high’ and ‘low’ using the median score, it was observed that the trajectories were stable across the two groups, although with diverse severity levels. Specifically, PRS-SCZ was positively associated with interindividual differences in avolition (i.e. ‘high’ PRS-SCZ indicated stable severer avolition) and negatively associated with differences in global functioning (i.e. ‘high’ PRS-SCZ indicated stable poorer global functioning) over the 20-year follow-up period. Moreover, PRS-SCZ was negatively associated with the latent cognitive factor at 24 months and 20 years, but not with the change between the two time points. PRS-SCZ was also the strongest predictor of diagnostic shifts from affective to non-affective psychosis over the 20-year follow-up, although with a poor discriminant ability (AUC = 0.62). At the highest decile, PRS-SCZ detected those who transitioned into the non-affective psychosis group with 68% accuracy (sensitivity = 16%, specificity = 92%).

Landi et al. (2021) evaluated the contribution of PRS-SCZ to improving the prediction of poor outcomes in clinical practice. Data from two unique cohorts of adults of African, admixed American, and European ancestries were analyzed. Analyses were first conducted in 762 individuals with psychosis from the BioMe cohort, an ancestrally diverse biobank in a New York City health system (Belbin et al., 2021), and findings were subsequently replicated in 7779 cases from the Genomic Psychiatric Cohort (Pato et al., 2013). Six outcomes were selected as proxies for the poor clinical course: the need for inpatient psychiatric treatment, prescription of two or more unique antipsychotics, prescription of clozapine (irrespective of the total number of unique antipsychotics), aggressive behaviour, self-injurious behaviour, and homelessness. Findings suggested that the PRS-SCZ would not significantly improve the prognostic ability of information collected during a routine psychiatric interview. These observations were consistent across different case ascertainment strategies and ancestral backgrounds.

Zhang et al. (2019) aimed to understand whether PRS-SCZ would predict symptom severity after 12 weeks of treatment in multiple cohorts of first-episode psychosis, two of European ancestry and two including European, African, Asian, and mixed ancestries. A combined random-effect meta-analysis (n = 510 participants) revealed that higher PRS-SCZ (at p < 0.01) significantly predicted worse 12-week symptom scores (3.24% of variance explained). When considering the cohorts individually, findings were confirmed in the discovery cohort and replicated in two out of three replication cohorts. Patients with low PRS-SCZ were more likely to be treatment responders than those with high PRS-SCZ (OR = 1.91 in European ancestry). Given that repeating the analysis using PRSs for non-psychiatric disorders (e.g. diabetes mellitus) did not yield significant results, researchers argued for the likely specificity of the association between PRS-SCZ and treatment outcomes.

Other studies have focussed on treatment-resistant schizophrenia (TRS) and response to clozapine, which is one of the most effective antipsychotics and the sole medication approved for TRS (Wagner et al., 2021). Wimberley et al. (2017) found no significant associations between PRS-SCZ and treatment resistance. Conversely, Werner et al. (2020) recently showed in 321 patients with schizophrenia that higher PRS–SCZ was significantly associated with TRS (OR = 1.5). The model yielded a sensitivity of 29.6% and a specificity of 90.6%, explained variance of 14.6%, and remained significant also after excluding clozapine users. Frank and colleagues reported that PRS-SCZ was higher in patients with a history of clozapine use compared with those who had never used clozapine (R² = 0.008 for p < 0.2), but this finding was not confirmed after introducing clinical variables into the model (Frank et al., 2015).

Recently, a PRS model for treatment-resistant schizophrenia (PRS-TRS) was developed based on two training GWAS of schizophrenia, that is, the CLOZUK and PGC datasets. The TRS interaction summary statistics of both GWASs were used to estimate SNPs variation effect size differences between individuals with and without TRS. PRS-TRS was then tested in prevalence (Cardiff Cognition in Schizophrenia) and incidence (Genetics Workstream of the Schizophrenia Treatment Resistance and Therapeutic Advance) samples of 817 and 563 individuals, respectively. PRS-TRS showed detectable heritability (1–4%) and was correlated with several cognitive traits (with genetic correlation in the range of 0.41–0.69). Moreover, analyses showed a positive association between PRS-TRS and a history of taking clozapine both in the prevalence (R² = 2.03%) and incidence (R² = 1.09%) samples. These findings provide support for a polygenic contribution associated with TRS that seems distinct from liability to schizophrenia per se (Pardiñas et al., 2022).

Major depressive disorder

MDD is a common mental disorder with a heterogeneous clinical presentation. Treatment options vary from psychological to pharmacological to somatic therapies, such as electroconvulsive therapy (ECT) or transcranial magnetic stimulation (Maj et al., 2020). At least 30% of people with MDD have treatment-resistant depression (TRD), which leads to poorer clinical trajectories (McIntyre et al., 2014).

Meerman and colleagues recently systematically reviewed the evidence on the association between PRSs and antidepressant response, including 30 different traits examined in 11 papers reporting inconsistent results (Meerman et al., 2022).

García-González et al. (2017) tested whether PRS-MDD and PRS-SCZ predicted response to antidepressants. First, PRS-MDD and PRS-SCZ were constructed from antidepressant trials (i.e. the GENDEP [n = 736] and the STAR*D [n = 1409] studies) to see whether they predicted symptom improvement or remission. No significant results were obtained, also after treatment stratification. However, the authors acknowledged that this analysis might have been underpowered. Second, using the GWASs from the PGC, they tested whether PRS-MDD or PRS-SCZ predicted symptom improvement in GENDEP, STAR*D, and five other studies (n = 3756). This analysis similarly yielded no evidence of associations in the meta-analysis or in each individual study. Stratifying by antidepressant did not alter the results.

A meta-analysis of three cohorts (n = 760) suggested that higher genetic loading for MDD and neuroticism might predict a poorer response to antidepressants after 4 and 8 weeks, although the results were not statistically significant (Ward et al., 2018). More recently, Fanelli et al. (2022) performed a larger meta-analysis of relevant PRSs and antidepressant non-response (n = 3637) and non-remission (n = 3148) in MDD. The strongest association was found for PRS-MDD and non-remission, with more likely non-remission in patients in the highest PRS quintile compared with those in the lowest PRS quintile (OR = 1.5). Associations were also found between PRS-MDD and non-response, as well as between PRS-SCZ and non-remission. However, none of the significant findings survived correction for multiple testing.

Other studies focussed specifically on TRD. For instance, Wigmore et al. (2020) reported that PRS-MDD and PRS-SCZ at different p-value thresholds were positively associated with antidepressant treatment resistance. Moreover, PRS-MDD was associated with the stages of antidepressant resistance. However, these findings did not remain significant after multiple test corrections. Fabbri et al. (2021) found that PRS for attention-deficit hyperactivity disorder (ADHD) was positively and significantly associated with TRD (defined as the prescription of more than two different antidepressants within 14 weeks) vs non-TRD (OR = 1.09), with consistent results across different p-value thresholds for PRS-ADHD. Other PRSs tested were not associated with TRD (SCZ, MDD, BD) or did not survive Bonferroni correction (neuroticism, subjective wellbeing, intelligence). The authors argued that this finding might indicate that TRD could hide undetected ADHD. However, Li et al. (2020) failed to confirm these findings and found no associations of PRS-ADHD with the percentage change in depression score, treatment response, and symptom remission after 4-week treatment with esketamine. A suggestive negative correlation was instead found between PRS for depressive symptoms and the esketamine treatment response outcome (i.e. percentage change from baseline in the MADRS scores). Suggestive positive associations were reported also between PRS for depressive symptoms and both esketamine responder, as well as remission status. None of these associations were of study-wide statistical significance.

Although undiagnosed BD has been regarded as a potential cause for TRD, no significant associations of PRS-BD with the antidepressant response (Fanelli et al., 2021; Tansey et al., 2014) or antidepressant treatment resistance have been demonstrated so far (Wigmore et al., 2020).

Higher PRS for openness was associated with non-remission and non-response after 4 weeks of antidepressant treatment in MDD (Amare et al., 2018). Interestingly, higher PRS for Coronary Artery Disease and PRS for obesity were also associated with poorer responses (Amare et al., 2019). Moreover, PRS for cardioembolic stroke and large vessel strokes were associated with a reduced probability to achieve symptom remission; however, the performance improvement compared to a clinical model was minimal (Marshe et al., 2021). These preliminary results may indicate that the utilization of non-psychiatric PRSs may shed further light on the association between MDD and cardiovascular disorders.

Electroconvulsive therapy (ECT) is a therapeutic option, particularly for patients with TRD. Meta-analytical findings report ECT response rates of 58% for patients with TRD and 70% for those without TRD (Haq et al., 2015). Two studies investigated the association between PRSs and response to ECT in patients with MDD (Foo et al., 2019; Luykx et al., 2022). Both studies found no significant associations between PRS-MDD and ECT response. However, Luykx et al. (2022) reported that higher PRS-SCZ was positively associated with a larger Hamilton Depression Rating Scale decrease over time (R² = 6.94%) and remission (R² = 4.63%). These findings were in accordance with evidence showing that psychotic features predicted a favourable response to ECT (van Diermen et al., 2018). Interestingly, the direction of the association between PRS-SCZ and response to ECT is opposite to that observed between PRSs and response to pharmacological treatment.

Bipolar disorder

Making accurate predictions on the course and outcomes of BD is often challenging, as BD is characterized by episodic relapses over the lifespan, also making pharmacological management highly complex. Lithium is a cornerstone first-line treatment for BD, both for acute manic episodes, as well as for long-term maintenance and relapse prevention (Yatham et al., 2018). However, only about 30% of people with BD can be considered full responders, and there is evidence that these patients tend to cluster in families (Alda, 2017).

Amare and colleagues investigated the association between PRS-MDD and response to lithium treatment in 2586 bipolar patients in the Consortium on Lithium Genetics Study. People with higher PRS-MDD were less likely to respond to lithium compared to patients with low polygenic load (the lowest vs the highest PRS quartiles) in the multi-ethnic sample (OR = 1.54) and the European sample (OR = 1.75) but not in the Asian sample (Amare et al., 2021).

A meta-analytic PRS (MET-PRS) was developed from combinations of highly correlated psychiatric traits, namely schizophrenia, MDD, and BD (Schubert et al., 2021). Higher PRS-SCZ and PRS-MDD were associated with poorer lithium response, whereas PRS-BD had no association with treatment outcome. The combined PRS (MET2-PRS) comprising of schizophrenia and MDD variants and a model using PRS-SCZ and PRS-MDD sequentially improved response prediction, compared to single-disorder PRSs and a combined score using all three traits (MET3-PRS). Patients in the highest decile for MET2-PRS loading had 2.5 times higher odds of being classified as poor responders than patients with the lowest decile of MET2-PRS (Schubert et al., 2021). Recently, Cearns et al. (2022) developed a non-linear model with only clinical variables explaining 8.1% of the variance in lithium response. A priori genomic stratification improved the model performance to 13.7% and improved the binary classification of lithium response.

Discussion

The findings presented in our review are substantially inconclusive. They highlight that, at present, the standalone applicability of PRS in clinical practice is very limited. Nevertheless, we already possess other (non-genetic) instruments that can help foresee – at least in part – outcomes and treatment responses for personalized psychiatry; the role of genetic vulnerability cannot be completely understood if detached from the rest of the picture. To enhance the validity of PRS and work towards its incorporation in the real world, we propose to (1) overcome the limitations of currently available genetic measures; (2) integrate genetic risk with environmental and clinical features for personalized treatment; (3) improve the methodology of the studies; (4) measure meaningful outcomes for patients and clinicians.

Limitations of PRS

PRSs have certainly a series of advantages. First, they represent stable trait features. Second, they have a normal distribution in the general population. Third, they can be relatively easily calculated in terms of time and costs.

However, they still present several limitations that reduce their real-world applicability. First, they explain only a small part of the heritability for major mental disorders. For instance, PRS-SCZ – which is considered the best performing PRS in psychiatry – only captures a median of 7.3% of the variance in liability attributable to schizophrenia, with the SNP-based heritability being around 24% (Trubetskoy et al., 2022). These estimates are considerably below the 60–80% heritability previously demonstrated in family and twin studies of schizophrenia (Hilker et al., 2018).

One of the reasons is that the standard PRS approach is based on common genetic variants only. However, rare variants and rare structural changes (e.g. CNVs, deletions, insertions) may also confer risk for mental disorders (Singh et al., 2022) and contribute to the variability in treatment response (Ruderfer et al., 2016). Rare variants are more difficult to genotype and analyze, as genome sequencing rather than genome-wide genotyping is needed. To overcome this limit, next-generation approaches like whole-genome sequencing (WGS) and whole-exome sequencing (WES) have been proposed. WGS consists of the sequencing of the entire genome to identify both polymorphism (i.e. responsible for interindividual phenotypic variability) and pathogenic variants. The identification of these variants in affected individuals can be crucial, although their interpretation is not easy (Dewey et al., 2014). WES is the sequencing of the protein-coding regions of genes and targets approximately 3% of the whole genome. However, it allows identifying variations in the protein-coding region of any gene, rather than in only a select few genes, contrary to PRSs. Because most known mutations that cause disease occur in exons, WES may represent an efficient method to identify possible disease-causing mutations (Suwinski et al., 2019). However, the implementation of these alternative approaches requires larger datasets than currently available SNPs-based GWASs.

Another important limitation of PRSs is the lack of population diversity in the current studies. It has long been known that population stratification is a major confounder in genetic research. European ancestry consistently represents the largest population: 96% of participants in GWAS by 2009 were of European descent (Popejoy & Fullerton, 2016). Current PRSs perform poorly in ancestries different from those of the GWAS training dataset. The active non-inclusion of a vast proportion of the population worldwide inevitably limits the validity and thus the predictive performance of PRSs in non-European samples. To improve the generalisability and applicability of the genomic-based prediction models, there is urgent need to achieve ‘racial’ and ethnic diversity in psychiatric genetics (Burkhard et al., 2021).

Integrating genetic, environmental and clinical features for personalized medicine

The ‘heritability gap’ between proxy and molecular genetic studies strongly indicates that the pathoetiology of mental disorders relies not only on genetics but also on environmental risk factors (van Os et al., 2010). To investigate the entirety of environmental exposures (i.e. the totality of the nongenetic component) from conception onward – the effects of which accumulate over time – it has been proposed to apply the ‘exposome’ paradigm to mental health (Guloksuz, Rutten, et al., 2018; Guloksuz, van Os, & Rutten, 2018). The exposome paradigm aims to complement the genome by considering the complex network of exposures (Wild, 2005; Wild, 2012). Guided by this paradigm, a specific aggregate environmental vulnerability score for schizophrenia, the exposome score for schizophrenia (ES-SCZ), has been developed and tested in different samples (Pries et al., 2019). The ES-SCZ was associated with the psychosis risk level (Guloksuz et al., 2020) and showed a good performance for identifying schizophrenia in the general population (Pries et al., 2021). Moreover, the ES-SCZ longitudinally predicted poorer broad mental and physical well-being in a population cohort study (Pries, van Os, et al., 2020) and was associated with overall levels of functioning in individuals with schizophrenia (Erzin, Pries, van Os, et al., 2021), and overall levels of functioning and symptomatic improvement in first-episode psychosis (Erzin, Pries, Dimitrakopoulos, et al., 2021). The example of ES-SCZ underlines the potential predictive capacity of environmental risk factors for determining not only the onset but also clinical trajectories of major mental disorders. These findings indicate that integrating PRS-based prediction with exposome data, such as ES-SCZ, will improve predictive power (Pries, Dal Ferro, et al., 2020).

Evidence has increasingly shown that mental disorders, such as schizophrenia, lack clear boundaries, with large phenotypic and pathoetiological overlap (Guloksuz & van Os, 2018). It is thus crucial to accomplish a complete clinical characterization of patients that goes beyond diagnostic silos and mere symptom checklists of operationalized diagnostic classification systems. It should be noted that clinical characterization mainly relies on psychiatric interview skills and clinical reasoning (Guloksuz & van Os, 2020). A detailed description of clinical and behavioural variables, such as symptom severity, cognition, and comorbidities, as well as a clinical contextualization of PRSs could improve prediction in clinical practice. This could lead to the harmonization of the populations targeted for molecular genetic research studies, in which diagnostic characteristics are sometimes too broad and overlapping.

Socio-demographic factors, such as sex and age, should be also considered in PRS-based prediction. For instance, it has been shown that several clinical variables differ between men and women with psychosis (Ferrara & Srihari, 2021; Pence et al., 2022) and MDD (Girgus & Yang, 2015; Parker & Brotchie, 2010), suggesting that sex might play a crucial role not only in the onset but also in outcome and treatment response.

Methodology

Studies with longitudinal design are crucial to understanding the utility of PRS in predicting outcomes. Equally important is the collection of genetic samples from large cohorts of medication-naïve individuals to understand the actual efficacy of treatment. Nonetheless, the cross-sectional recruitment of participants with mental disorders – classified as responders and non-responders – and healthy controls, may eventually lead to a rapid accumulation of data for the development of specific PRSs for treatment response or resistance (see for instance: Amare et al., 2021; Pardiñas et al., 2022).

The validity of PRS has been questioned when it is used in combination with clinical factors that are not completely independent of the outcome (e.g. social impairments to predict psychosis). In this case, the PRS may capture both the risk for the outcome (e.g. psychosis) and the clinical factor (intermediate phenotype) inserted in the prediction model (e.g. social impairment). Thus, it has been argued that the utilization of a ‘residual’ PRS – that is a PRS capturing the genetic liability to the outcome minus the liability for the clinical factor – might improve the validity (Janssens, 2019).

Another limitation to the PRS validity is the assumption that SNPs would have the same impact on disease risk in all populations of the same ethnicity, although the GWAS estimates are pooled across multiple studies that differ in study design and study population and that may even differ in the diagnostic criteria and assessment of the disease of interest (Janssens, 2019). For this reason, international genetic consortia should further work towards the harmonization of samples.

Importantly, heterogeneous methodology and scattered outcome evaluation (e.g. treatment resistance/non-resistance, response/non-response, remission/non-remission, number of medications prescribed) might lead to inconsistency and difficulty in interpretation of results. Standardized outcome tools or measures should therefore be adopted by researchers and international consortia. Indeed, methodological flexibility – with the creation of ad-hoc outcomes – and selective reporting may eventually transform what would be ‘negative’ results into ‘positive’ results (Ioannidis, 2005). Additionally, most studies included in our review tended to emphasize nominally significant yet small effect sizes that were less likely to be clinically relevant (Ioannidis, 2008). A critical interpretation of the published literature would convey a more balanced view of the potential clinical utility of PRSs in predicting outcome and treatment response and guide the field in designing future clinically oriented studies.

Finally, the implementation of PRSs in clinical settings is hampered by the lack of reporting standards, which limits the interpretation and the comparison of findings. There is an urgent need for data transparency and availability, as well as the deposit and the share of PRS data to facilitate reproducibility and comparative analysis (Wand et al., 2021).

Meaningful outcomes for clinical practice

Future research should ideally focus on outcomes that can be meaningful for patients and helpful for clinicians in the choice of the best treatment. Thus far, the majority of studies have focussed on the efficacy of treatments, while evidence on the prediction of side effects is still scarce. However, psychotropic medications are frequently burdened by side effects, which may eventually lead to treatment discontinuation. The wide inter-individual variability in side effects suggests an underlying genetic predisposition; for instance, polymorphisms of drug-metabolizing enzymes may lead to changes in the function of metabolizing enzymes and contribute to genetic liability to develop certain side effects. Therefore, the development of PRS models aimed at predicting the genetic liability to adverse events is highly needed. However, this requires deeply phenotyped, large clinical cohorts (Gibson, 2019).

Another innovative clinical application could be related to the prediction of trajectories in help-seeking youths. Indeed, some studies have reported significant associations between PRSs for mental disorders and psychiatric, cognitive, and behavioural phenotypes in children and adolescents (Jones et al., 2016; Mistry et al., 2019; Pain et al., 2018). More recently, Murray et al. (2021) generated PRSs for three mental disorders and four non-psychiatric traits in a cohort of 158 adolescents and young adults presenting to youth mental health services and compared them to a control sample. Large mean differences between the two groups were found for PRS-SCZ, PRS-MDD, and PRS-BD, although the PRS distributions were substantially overlapping. Nevertheless, it was observed that around 24% of the clinical sample’s PRS-SCZ were in the top 10% of the PRS-SCZ distribution of the population sample. This analysis supports the notion that integrating PRSs with specific clinical profiles could provide valuable hints for decision-making. Moreover, PRSs may have some potential in predicting mental health trajectories in youths with early-stage pluripotential syndromes (Shah et al., 2022). However, replication in larger samples is needed.

Conclusions

In the present review, we summarised the studies that investigated the association of PRSs with treatment response and outcomes of major mental disorders. Although the use of genetic risk scores for prediction might offer some potential in the future, the small phenotypical variance explained by PRSs, the heterogeneity of the methodology including samples and outcome definitions, and the inconsistency of findings so far make the clinical applicability of PRSs very premature as of now. Clearly, research examining the association between PRSs and psychopathology needs to consider the ethnic, geographic, and clinical diversity of people with mental disorders and harmonize clinical samples to maximize the specificity and accuracy of genetics-informed prediction. Studies with large sample sizes, accurately captured outcome measures, and meta-analytic estimates are needed to estimate specific treatment response PRSs. Even with the rapid expansion of the psychiatric genetic knowledge base in the last decade, it is clear that pure genetic prediction in clinical psychiatry appears to be out of reach in the near future – if at all. Therefore, combining genomic and exposomic vulnerability for mental disorders with a detailed clinical characterization is needed to personalize care.

Disclosure statement

The authors report no conflicts of interest.

Additional information

Funding

Dr. Fusar-Poli was co-funded by the European Union FSE-REACT-EU, PON Research and Innovation 2014-2020 DM 1062/2021 [08-I-17631-1].