Using Medical Records to Investigate the Genetics of Treatment-Resistant Depression Across Health Care Systems

Depression is a psychiatric disorder affecting a large proportion of individuals, with major social and economic consequences for families and society. In the past 10 years, large-scale studies have revolutionized our understanding of depression predisposition. In particular, genome-wide association studies (GWASs) of millions of participants have uncovered hundreds of loci contributing to the polygenic architecture of depressive disorders and symptoms (1, 2). To achieve the statistical power needed to detect the small effects of individual variants, investigators have combined cohorts with diverse characteristics and depression phenotypes derived from different data sources (3). Effects detected in these large-scale GWASs are informative of molecular mechanisms shared across the depression spectrum. However, they may be less informative of specific types of depression (4). This means that more efforts are needed to understand the biology of certain depression outcomes. Among them, treatment-resistant depression (TRD) surely represents a priority, because of its major clinical implications. Indeed, a large proportion of depression-affected patients do not respond to multiple sequential antidepressant therapies (12%–55%, depending on how TRD is defined) (5). Dissecting TRD pathogenesis could have important translational implications leading to the identification of biological processes informative of response to antidepressant treatments.

In this issue of the Journal, Kang and colleagues (6) report on a TRD GWAS of 154,433 participants across four health care systems. This represents a major advance in TRD genetics. So far, a key limitation for TRD studies has been defining the outcome of interest in large cohorts. To overcome this issue, the authors leveraged electronic health records to derive TRD probabilistic phenotypes, considering electroconvulsive therapy (ECT) as the gold-standard target. Among individuals affected by depression, ECT has been demonstrated to be an effective treatment for TRD, showing rapid improvement in depressive symptoms and high remission rates (7). Despite this, ECT utilization remains low because of stigma among patients and clinicians and high costs related to infrastructure and administration requirements (8, 9). The rate of ECT use is only 1.5% in inpatient psychiatric units, and insurance records show that <1% of depression-affected individuals receive ECT (10). These numbers are in line with findings from the present study, with the prevalence of ECT use among individuals with a diagnostic code of depressive disorder or major depressive disorder (MDD) found to be 0.31% in the Mass General Brigham (MGB) database and 0.21% in the Vanderbilt University Medical Center (VUMC) database.

While the number of patients with MDD who received ECT is too low to allow estimation of genetic associations, the data can be used to identify the phenotypic characteristics of TRD. Indeed, electronic health records and phenome-wide analyses can be used to derive prediction models (11) estimating posterior probabilities of an outcome of interest (i.e., the probability that a patient is a case or control for the phenotype of interest, taking into consideration their phenotype risk score). However, this type of approach can be affected by the characteristics of the samples used to train the models (12). To address this key issue, the authors used cohorts from four health care systems. MGB and VUMC cohorts were used for discovery and internal validation, and external validation was performed in cohorts from the Geisinger Health System and the Million Veteran Program. This is a major strength of the study, ensuring the reliability and generalizability of the ECT prediction. Because ECT was used as a proxy to predict TRD more broadly, the investigators showed that the models predict (within and between cohorts) TRD defined using VUMC clinical notes, and also effectively predicted TRD when ECT-treated patients with MDD were excluded.

While cross-cohort prediction confirmed the convergence between the VUMC and MGB models, there was only a partial overlap between them. This is due to the different characteristics of the ECT-MDD cases. In the VUMC data, participants identified as having TRD had a lower mean body mass index (BMI) than control participants, whereas no case-control difference in mean BMI was present in the MGB sample of patients identified as having TRD. Additionally, the patients with MDD who received ECT included a higher proportion of individuals who had received treatment with at least two different antidepressants at VUMC than at MGB (92% vs. 72%). These differences likely contributed to the partial correlation between the phenome-wide association results (r=0.7), which led to the different model performance. The investigators correctly decided to investigate the genetic architecture of quantitative TRD phenotypes derived from both the VUMC and MGB models (TRD_VUMC and TRD_MGB), assessing differences and similarities. Considering single-nucleotide polymorphism–based heritability (SNP-h²) estimates and their standard errors, TRD_MGB was a more genetically informative phenotype than TRD_VUMC (SNP-h² z values of 10.5 and 5.56, respectively). Their genetic correlation (r_g=0.66) was in line with the correlation of the phenome-wide association results (r=0.7). Likely due to the higher statistical power, two genome-wide significant associations were identified for TRD_MGB. One was related to the FTO locus, showing inverse effects between TRD and BMI. The other was an expression quantitative locus (a genetic variant associated with gene expression changes) for the MCHR1 gene, which was in high linkage disequilibrium with a locus that has been linked to bipolar disorder.

While these are important advancements, both TRD_MGB and TRD_VUMC did not show any genetic overlap (r_g values, <0.15; p values, >0.22) with previous GWASs of pharmacologically defined TRD (13) and ECT (14). Conversely, there was a significant genetic correlation between the latter two data sets (r_g=0.75). The lack of TRD_MGB and TRD_VUMC genetic overlap with previous phenotype definitions could be due to multiple dynamics. In their article, Kang et al. highlight the definitions used in these previous studies. The pharmacologically defined TRD included MDD cases having at least two switches between different antidepressant drugs, and also considered criteria related to prescription times (13). The previous ECT GWAS compared individuals who received ECT for a major depressive episode with individuals who had been enrolled as control subjects in a genetic study of anorexia nervosa (who had no history of an eating disorder, MDD, bipolar disorder, schizophrenia, or schizoaffective disorder; 98% females) (14). While the phenotype definitions could partially explain the lack of overlap, other aspects surely contribute. The patterns of genetic correlations with other phenotypes could support the role of factors related to ECT utilization. Both TRD_MGB and TRD_VUMC showed positive genetic correlations with educational attainment and intelligence and negative genetic correlations with attention deficit hyperactivity disorder, alcohol dependence, and smoking traits. The direction of these genetic correlations is the opposite of what was reported in the previous GWASs of pharmacologically defined TRD (13) and ECT (14) and also in previous MDD GWASs (2, 15). In particular, the unexpected positive genetic correlation of TRD_MGB and TRD_VUMC with educational attainment and intelligence suggests that genetic effects may have been partially influenced by dynamics not directly related to TRD but that are more related to which patients decide to undergo ECT. As mentioned above, ECT utilization is low because of stigma and high costs (8, 9). Highly educated individuals may be less affected by these factors. The study conducted by Kang et al. (U.S. cohorts) and the previous ECT GWAS (a Swedish cohort) may also have been influenced by the fact that socioeconomic factors may have differential effects in private versus public health care systems on patients’ decisions to undergo ECT. Previous studies have shown how participation bias can affect genetic associations in large-scale analyses (16, 17). So, it is likely that cohort-driven dynamics related to both ECT and pharmacologically defined TRD contributed to the null genetic overlap observed, also limiting the ability to dissect the underlying biology of TRD. Our field is starting to understand how to identify the genetic signatures of these unwanted effects, and analytic methods and recruitment and assessment strategies have been proposed to correct and avoid them (17, 18).

In their study, Kang et al. did impressive work to ensure the reliability and generalizability of their results, proving the validity of using ECT as a proxy for TRD and validating their models across multiple cohorts. These are important findings on which they and other investigators can build. However, the goal cannot be only to increase the sample size to identify more genetic associations. Because it seems that participation and assessment biases could affect TRD studies, these dynamics should be investigated further to ensure that genomic findings provide a reliable basis for translational research aiming to develop biomarkers and drugs for patients affected by TRD.