Functional genetic variation in the Rev-Erbα pathway and lithium response in the treatment of bipolar disorder


  • M. J. McCarthy,

    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    2. VA San Diego Healthcare System, San Diego, CA, USA
    Search for more papers by this author
  • C. M. Nievergelt,

    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    Search for more papers by this author
  • T. Shekhtman,

    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    2. VA San Diego Healthcare System, San Diego, CA, USA
    Search for more papers by this author
  • D. F. Kripke,

    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    2. Scripps Clinic Sleep Center, La Jolla, CA, USA
    3. Center for Chronobiology, University of California, San Diego, La Jolla, CA, USA
    Search for more papers by this author
  • D. K. Welsh,

    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    2. VA San Diego Healthcare System, San Diego, CA, USA
    3. Center for Chronobiology, University of California, San Diego, La Jolla, CA, USA
    Search for more papers by this author
  • J. R. Kelsoe

    Corresponding author
    1. Department of Psychiatry, University of California, San Diego, La Jolla, CA, USA
    2. VA San Diego Healthcare System, San Diego, CA, USA
    Search for more papers by this author

J. R. Kelsoe, Department of Psychiatry, University of California, San Diego, 9500 Gilman Drive #0603, La Jolla, CA 92093, USA. E-mail:


Bipolar disorder (BD) is characterized by disruptions in circadian rhythms such as sleep and daily activity that often normalize after lithium treatment in responsive patients. As lithium is known to interact with the circadian clock, we hypothesized that variation in circadian ‘clock genes' would be associated with lithium response in BD. We determined genotype for 16 variants in seven circadian clock genes and conducted a candidate gene association study of these in 282 Caucasian patients with BD who were previously treated with lithium. We found that a variant in the promoter of NR1D1 encoding Rev-Erbα (rs2071427) and a second variant in CRY1 (rs8192440) were nominally associated with good treatment response. Previous studies have shown that lithium regulates Rev-Erbα protein stability by inhibiting glycogen synthase kinase 3β (GSK3β). We found that GSK3β genotype was also suggestive of a lithium response association, but not statistically significant. However, when GSK3β and NR1D1 genotypes were considered together, they predicted lithium response robustly and additively in proportion to the number of response-associated alleles. Using lymphoblastoid cell lines from patients with BD, we found that both the NR1D1 and GSK3β variants are associated with functional differences in gene expression. Our findings support a role for Rev-Erbα in the therapeutic mechanism of lithium and suggest that the interaction between Rev-Erbα and GSK3β may warrant further study.

Bipolar disorder (BD) is a psychiatric illness characterized by severe, episodic mood disturbances that include distinct periods of mania and depression. In both phases of the illness, BD perturbs daily rhythms in sleep, energy, appetite and activity. Furthermore, mood symptoms can vary with time of day and season, and bright light is both an effective antidepressant and a potential trigger for mania. For these reasons, circadian clock abnormalities have been suspected in BD for decades (reviewed by McClung 2007). Previous studies from our laboratory and others have supported this hypothesis, reporting that polymorphisms in a number of genes with circadian rhythm roles are associated with BD including ARNTL, CRY1, CRY2, NR1D1 and RORB among others (Kripke et al. 2009, Mansour et al. 2005, Mansour et al. 2009, McGrath et al. 2009, Nievergelt et al. 2006, Shi et al. 2008, Sjöholm et al. 2010, Soria et al. 2010). Moreover, models for mania have been developed based on experimental manipulations of the circadian clock in animals (Le Niculescu et al. 2008, Mukherjee et al. 2010, Roybal et al. 2007).

The circadian clock is composed of a complex, interacting network of genes (‘clock genes') that collectively maintains rhythmic expression over approximately 24-h cycles (reviewed by Takahashi et al. 2008). The protein products of CLOCK and ARNTL are the primary transcriptional activators, whereas the protein products of PER1/2, CRY1/2 and NR1D1/2 (Rev-Erbα/β) are the major transcriptional inhibitors that complete negative feedback loops within the clock (Kume et al. 1999, Liu et al. 2008). Interestingly, these genes are expressed widely throughout the brain and peripheral tissues, maintaining rhythms throughout the body.

Among the effective treatments for BD, lithium (Li) is commonly used. However, a good treatment response to lithium is not universal, and among BD patients, approximately 50% fail Li by the second year of treatment (Vestergaard & Schou 1988). The clinical profiles of Li responsive (Li-R) and nonresponsive (Li-NR) BD patients have been reported to differ. For example, euphoric mania and a positive family history are considered to be the factors favoring Li response, whereas dysphoric mania, the presence of comorbid psychiatric conditions and a negative family history are considered poor prognostic indicators (Kukopulos et al. 1980, Grof et al. 2003). On the basis of these and other observations, it has been proposed that Li-R and Li-NR BD may be genetically distinct forms of illness (Grof et al. 2003, 2009). However, the biological basis for these differences in Li responsiveness remains unknown.

Despite its common use for over 50 years, the precise therapeutic mechanism of Li remains unknown, but may involve the circadian clock. In support of this notion, various molecular components of the circadian clock are known to be Li sensitive. Specifically, Li has been reported to lengthen the period of circadian rhythms in several species (Kripke and Wyborney 1980, Kripke et al. 1979, Welsh and Moore-Ede et al. 1990) and to alter circadian rhythms in neuronal firing (Abe et al. 2000). Li also inhibits glycogen synthase kinase 3β (GSK3β, Klein & Melton 1996), which in turn regulates protein turnover and stability in the clock network by phosphorylating target proteins, thereby modulating circadian rhythms (Iitaka et al. 2005, Yin et al. 2006). Of particular interest to the current study, GSK3β inhibition by Li has been shown to facilitate the degradation of Rev-Erbα, an important clock signaling molecule (Yin et al. 2006). This evidence suggests that Li affects clock genes, and that within the clock, the relationship between GSK3β and Rev-Erbα (transcribed from the gene NR1D1) may be of particular interest. Supporting this notion, the variant rs2314339 within NR1D1 has been recently associated with Li response (Campos-de-Sousa et al. 2010). However, in this study the relationship of NR1D1 to GSK3β was not examined.

Presently, we sought to identify functional genetic variants in the circadian clock that may be useful in predicting lithium response in BD. We found clock gene variants that are nominally associated with clinical Li response in patients with BD, including two variants from the well-characterized clock signaling pathway that combine for a more robust association. We then used lymphoblastoid cell lines (LCLs) to model the effects of genetic variation within the clock on Li-induced gene expression in Li-R and Li-NR patients with BD.



Subjects (N = 282) were recruited for genetic studies through a mood disorder clinic at the Veterans Affairs San Diego Healthcare System (VASDHS). Research was approved by the UCSD/VASDHS IRB and all participating subjects provided informed consent. BD can be subdivided into types I and II, based primarily on illness severity and the presence of mania (in BD I) or hypomania (in BD II), whereas BD not otherwise specified (NOS) is used when one or more factors limit the ability to make a specific BD I/II diagnosis (American Psychiatric Association 2000). As Li is an accepted therapy for all forms of BD, subjects with any BD subtype were included. The majority (91%) had BD I, whereas a minority had BD II (7%) or BD NOS (2%). All diagnoses were established using one or more standardized instruments including the Structured Clinical Interview for DSM-III-R or DSM-IVTR (SCID) and the Diagnostic interview for genetic studies (DIGS). All subjects were of self-declared Caucasian ancestry.

Clinical assessment

Li response was determined retrospectively as described previously (Bremer et al. 2007). Briefly, a detailed review of medication trials was conducted using information from the patient, medical records and other informants. A panel of clinicians blind to genotype then rated this information. Episode frequency and severity were compared between periods on/off Li. Only Li trials of 3 months or longer were considered. Response was rated as good if there was a 50% reduction in the frequency and/or severity of symptoms on Li. These subjects were classified as Li responders (Li-R, N = 148). Those with less than 50% symptom reduction were considered Li nonresponders (Li-NR, N = 134). The majority of subjects were also scored on the 10 point Alda scale of Li response (Grof et al. 2002). Average Alda A criterion score (±SD) for Li-R (N = 115) was a 6.6 ± 1.5, corresponding to a response of approximately 81.5%, and for Li-NR (N = 118) was 2.5 ± 1.5, corresponding to a response of approximately 22.5%. In each group, average age (year ± SD) was similar (Li-R: 45 ± 12.9 and Li-NR: 44 ± 12.0). Among the Li-R group, 82/148 (55%) were men, compared with 60/134 (45%) men in the Li-NR group [χ2 = 3.1(1), P = 0.08].


Seven genes were selected for analysis based on their critical roles in the core clock mechanism, both in the positive and negative feedback limbs of the system (Table 1). Sixteen single nucleotide polymorphisms (SNPs) were selected based on predicted function and position within the gene. We focused on regions showing epigenetic histone marks associated with regulatory elements and/or phylogenetic conservation (UCSC Genome Browser:, March 2006 Assembly). We also used evidence from a case–control study performed previously by our group (Kripke et al. 2009) and unpublished association data. This set of markers does not provide comprehensive haplotype coverage but tests multiple candidate gene regions while minimizing multiple comparisons. All SNPs were transmitted independently across genes. Within ARNTL, rs1982350 and rs2279287 were correlated (r2 = 0.36, and all six SNPs in PER3 showed varying degrees of correlation (r2 = 0.21–0.84). Thus, the Bonferroni correction employed for multiple testing was considered conservative (Nyholt 2004).

Table 1.  Variants selected from circadian clock genes and Li response association results
GeneSingle nucleotide polymorphismChromosomePositionLocation N Minor alleleMajor alleleMinor allele frequencyWald statisticsOdds ratio additive P additiveOdds ratio dominant P dominantOdds ratio recessive P recessive
  1. Gene and SNP information including the MAF for the rare allele is indicated. Additive, dominant and recessive genetic transmission models were used, the nominal P-values (P) and ORs are shown for each (Add, Dom and Rec). Uncorrected P < 0.05 was considered nominally statistically significant (bold text) with 1 df for each analysis. In some cases, exclusions for quality control slightly reduced total sample size.

  2. NA, not available.

ARNTL rs22792871113298485Enhancer281AG0.270.561.110.571.100.681.300.57
ARNTL rs19823501113350131Enhancer281TC0.35−0.180.970.861.040.890.800.56
ARNTL rs22787491113397878Promoter282AG0.190.721.180.471.110.692.670.24
CLOCK rs1801260456301369Exon281CT0.260.321.070.751.320.260.450.13
CLOCK rs3736544456309992Exon279AG0.36−1.640.750.100.680.120.680.27
CLOCK rs34897046456325365Exon282CG0.050.751.340.451.520.31NANA
CRY1 rs819244012107395106Exon278AG0.41−2.50 0.64 0.01 0.54 0.02 0.590.11
GSK3β rs64385523119631814Enhancer281CT0.421.381. 1.82 0.08
NR1D1 rs20714271738254492Promoter282AG0.281.95 1.45 0.05 1.90 0.01 0.910.83
PER2 rs23046722239186589Exon281GC0.09−1.650.590.100.600.12NANA
PER3 rs22872917845695Intron282AG0.310.691.140.491.250.360.970.95
PER3 rs22864217863293Intron282CT0.430.561.110.581.060.831.270.46
PER3 rs22866617868725Intron282CT0.37−0.510.920.610.950.840.780.46
PER3 rs22869717887579Exon281GC0.110.441.120.661.100.751.580.61
PER3 rs285938817888215Promoter282GA0.360.261.050.790.990.971.260.55
PER3 rs264090917890117Exon277CT0.33−0.720.880.470.830.450.870.73

For genotyping, we used SNPlex with an ABI 3730 48-capillary DNA analyzer following the manufacturer's protocols (Applied Biosystems, Foster City, CA, USA). Three SNPs (rs2304679, rs3736544 and rs8192440) were genotyped by Taqman PCR (Applied Biosystems) using an ABI 7900 thermocycler. Among our cohort of 282 BD subjects, 184 (92 Li-R and 92 Li-NR, 65% of subjects) have been genotyped previously using a distinct and nonoverlapping set of markers (Bremer et al. 2007). A different set of 56 BD subjects (26 Li-R, 30 Li-NR, 20% of subjects) was genotyped previously using some of the same markers as part of a BD case–control association study (Kripke et al. 2009). In the present case–case analysis, no healthy controls were included, and all samples were regenotyped. Using the latter 56 samples as a quality control measure, we found 100% agreement in genotype calls between our previous study and the results presently reported. All of the 16 SNPs used for the Li association experiment also passed the following quality controls: call rate >0.98, minor allele frequency (MAF) >0.05 and genotypes in Hardy–Weinberg equilibrium.

Genetic association analysis

We used a panel of 41 ancestry-informative genetic markers to control for population stratification. Continental ancestry was determined using 944 Human Genome Diversity Panel subjects as a reference. Within subjects of European ancestry, the multidimensional scaling method (Purcell et al. 2007) was used to infer continuous axes of genetic variation to be used as covariates to control for residual stratification.

Genetic association was determined using logistic regression. The Wald test was used to calculate odds ratios (OR) for each variant with Li-R/Li-NR as the dependent variable, allelic frequencies of the SNPs as the predictors and gender with two MDS components as covariates. As we had no a priori hypothesis regarding the specific inheritance patterns that determine lithium response, we explored three common mechanisms, analyzing our genetic association data using additive, recessive and dominant genotypic models for each SNP. A haplotype analyses using a moving window approach was performed but did not result in any significant associations (not shown). All associations were performed in PLINK version 1.0.7 and SPSS version 16.

Further analyses of the NR1D1 and GSK3β genotypic combinations were made post hoc, examining each of the nine genotype possibilities individually. Calculations for ORs were made for each possible genotype combination and compared with the OR of Li-R/Li-NRs for all of the other genotypes, combined in an iterative manner for each of the nine possibilities. In all analyses, a higher OR corresponds to better chance of treatment response. A χ2 test was used to calculate ORs with significance defined as P < 0.05. Gene–gene interaction power analyses were performed using QUANTO version 1.2.

Lymphoblastoid cell cultures

For the NR1D1 expression studies, LCLs were selected from Li-R (N = 13) and Li-NR (N = 18) subjects with BD I on the basis of NR1D1 genotype. For GSK3β expression studies, additional samples were required to obtain sufficient numbers of each genotype, and not all of these cells had donors' Li response history available. In all cases, LCLs were cultured from frozen vials in standard RPMI growth medium with 10% fetal bovine serum (FBS) for 10–14 days to a density of approximately 1 million cells/ml. Cells were then collected, rinsed in phosphate buffered saline (PBS) and stored at −80°C. For time–course experiments, cells were serum shocked in medium containing 50% FBS for 90 min to synchronize circadian clock gene expression (Liu et al. 2007), then resuspended into standard medium in parallel cultures. After a 20-h equilibration period, flasks were harvested every 3 h over the next 24 h. For the Li studies, additional LCL cultures were grown in fresh medium ± Li chloride (1 mm) for 72 h, a duration that simulates the early phase of a clinically relevant drug exposure, but short enough to eliminate the need for manipulations that could introduce artifacts between cultures. Parallel cultures from the same donor ± Li were compared, allowing each cell line to serve as its own control.

RNA preparation and cDNA synthesis

RNA extraction from LCLs was performed with the RNeasy Mini Kit (Qiagen, Hilden, Germany). RNA (500 ng) was used to synthesize complementary DNA (cDNA) from each sample using a High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems).

Assessment of clock gene expression

Expression of individual genes was measured by real-time polymerase chain reaction (RT-PCR) using commercial primers (Applied Biosystems) for GSK3β (Hs00275656_m1) and GAPDH (Hs99999905_m1). NR1D1 has been previously reported to contain an alternative promoter, generating two functional transcripts of different lengths. For this reason, two distinct PCR primer pairs were used to measure NR1D1 expression. The first (Hs00253876_m1) targeted the junction of the first and second exons, thereby measuring only the full-length transcript. The second (Hs00897531_m1) targeted the junction of exons six and seven, thereby measuring both the full-length and truncated transcripts. RT-PCR reactions were performed in an ABI 7900 thermocycler with all samples run in duplicate. Gene expression was calculated as a ratio of the target gene to GAPDH, and analyzed using the comparative Ct method (Schmittgen & Livak 2008). Where appropriate, data were normalized as a percent of baseline.

Statistical analysis

Time–course data were analyzed using a one-way analysis of variance (anova) with post hoc t-tests. The least squares method was then used to determine the best line or nonlinear function to fit the data. Changes in gene expression following Li were determined by two-way anova with repeated measures using commercial software (Graphpad Prism, version 5, La Jolla, San Diego, CA, USA).


Genetic association of lithium response in patients with BD

Among the 16 clock gene variants, two SNPs (rs8192440 and rs2071427) were nominally associated with Li response. The rs8192440 (CRY1) association was strongest using an additive transmission model with the common G allele favoring good response (Table 1). The rs2071427 (NR1D1) association was strongest using a dominant transmission model with the rare A allele favoring good response (Table 1). Previous reports have suggested that GSK3β polymorphisms are associated with Li response (Adli et al. 2007, Benedetti et al. 2005). In our sample, the GSK3β variant rs6438552 showed only a suggestive trend toward association with Li response using a recessive inheritance model (Table 1), whereby homozygosity at the rare C allele favored Li response. The results obtained using a univariate analysis also showed nominal lithium response associations with rs8192440 from CRY1 and rs2071427 from NR1D1, but no significant association with the GSK3β variant rs6438552 (Table S1, supporting information). In neither analyses did any of the nominal associations remain significant after Bonferroni correction for 16 tests.

Association analysis of combined NR1D1 and GSK3β variants

In model systems, GSK3β has been reported to phosphorylate multiple proteins within the circadian clock including Rev-Erbα (Yin et al. 2006), ARNTL (also called BMAL, Sahar et al. 2010), PER2 (Iitaka et al. 2005) and CRY2 (Harada et al. 2005). Unlike CRY2, evidence suggests that CRY1 is not a substrate for GSK3β phosphorylation (Kurabayashi et al. 2010).

As Rev-Erbα is a substrate for GSK3β phosphorylation, and Li has been shown to inhibit this interaction (Yin et al. 2006), we examined whether multiple genetic variants within this known signaling pathway could interact to better predict Li response outcomes. Considering the NR1D1 and GSK3β variants together, the total number of Li response-associated alleles (range 0–4) was highly correlated with clinical outcome (Fig. 1a; r2 = 0.97, F1,3 = 98, P < 0.0025). Subjects who were doubly homozygous for the favorable A allele at rs2071427 and C allele at rs6438552 had a 75% chance of being in the Li-R group. In comparison, those who were doubly homozygous for the unfavorable G allele at rs2071427 and T allele at rs6438552 had only a 44% response rate. Those with three, two or one favorable alleles were intermediate with progressively lower chances of favorable clinical outcomes. Two additional analyses were conducted to evaluate the Li response rates of specific genotype combinations. The first, a 2 × 2 analysis using only the best fitting transmission models (dominant for NR1D1 and recessive for GSK3β) maximized statistical power. The second, a 3 × 3 analysis using each of the nine possible genotypic combinations allowed for higher resolution comparisons of the each genotypic combination (Fig. 1b,c). In the 2 × 2 analysis, those with NR1D1:GSK3β genotypes of AA-AG:CC responded significantly more often than did the comparison group [χ2(1) = 9.36, OR = 7.5, 95% CI 1.68–33.46, P < 0.0025]. Conversely, those with GG:TT-TC genotypes responded significantly less often [χ2(1) = 7.40, OR = 0.52, 95% CI 0.32–0.83, P < 0.0075]. The more detailed 3 × 3 analysis suggested that the rs6438552 CC genotype at GSK3β nominally favored Li response, regardless of NR1D1 genotype. The statistical analysis showed that subjects with NR1D1:GSK3β genotype of AG:CC fared best [χ2(1) = 8.75, OR 11.87, 95% CI 1.52–92.25, P < 0.003], whereas those with GG:TC genotype fared worst [χ2(1) = 3.85, OR 0.58, 95% CI 0.33–1.00, P < 0.05]. Only four subjects had the doubly rare AA:CC genotype. While this combination showed a trend toward favorable Li response its OR was not significantly different from unity [χ2(1) = 0.84, OR 2.77, 95% CI 0.28–26.98, P < 0.36]. Similarly, while the Li response rate of the AG:CC genotype was nominally higher than the AA:CC genotype, these groups had overlapping confidence intervals and could not be statistically discriminated. None of the other genotype combinations differed significantly from unity in the post hoc OR analyses (range 0.66–2.77).

Figure 1.

Additive effects of GSK3β and NR1D1 alleles on lithium response. (a) All possible genotype combinations at rs2071427 and rs6438552 were determined for all Li-R and Li-NR subjects and scored (0–4) for the number of alleles associated with favorable lithium outcomes. Dashed lines indicate 95% confidence intervals for the regression line. Rates of response for each allelic combination are the following: 4 alleles, 75% (N = 4); 3 alleles, 69% (N = 26); 2 alleles, 57% (N = 95); 1 allele, 44% (N = 108) and no allele, 47% (N = 48). (b) OR for lithium response calculated for each of 2 × 2 GSK3β and NR1D1 genotype combinations using dominant and recessive models. Sample sizes for each subgroup are as follows: AA-AG:CC, N = 17; AA-AG:TT-TC, N = 116; GG:TT-TC, N = 119 and GG:CC, N = 29. (c) OR for lithium response calculated for each of 3 × 3 GSK3β and NR1D1 genotype combinations. Sample sizes for each subgroup are as follows: AA:CC, N = 4; AA:TC, N = 13; AA:TT, N = 7; AG:CC, N = 13; AG:TC, N = 59; AG:TT, N = 37; GG:CC, N = 29; GG:TC, N = 71 and GG:TT, N = 48.

Examination for epistatic interaction

To more precisely define the combinatorial nature of the relationship between NR1D1 and GSK3β, a logistic regression analysis was performed to predict Li response including both rs2071427 and rs6438552. For each gene, this analysis used the most significant genetic model based on the results from the single SNP association analyses (dominant model for rs2071427, recessive model for rs6438552, see Table 1) and included covariates for gender and population stratification. This analysis showed a significant additive effect [χ2(5) = 15.92, P < 0.007] but fell short of providing evidence for a synergistic gene–gene interaction [χ2(1) = 2.228, P = 0.136]. We conducted a power analysis to assess our ability to detect epistasis using the model parameters that corresponded to the strongest single SNP associations (rs2071427: dominant model, MAF 0.28, OR 1.90 and rs6438552: recessive model, MAF 0.42, OR 1.82) for 148 Li-R and 134 Li-NR at a baseline Li response rate of 0.47. Under these assumptions, we had 80% power to detect only very large interaction effects of OR > 25. To detect interaction effects in the range 7–12, as can be seen in Fig. 1b,c, we would need at least 246 Li-R and 224 Li-NR, or 187 Li-R and 171 Li-NR, respectively.

Variant rs2071427 does not affect rhythmic NR1D1 expression

We next examined whether rs2071427 was associated with functional effects on NR1D1 expression. As a clock gene, expression of NR1D1 (encoding Rev-Erbα) was expected to oscillate rhythmically (Liu et al. 2008). Indeed, we were able to show that NR1D1 expression did vary significantly over time (F7,88 = 3.55, P < 0.005) after synchronization of the cells, and that a nonlinear curve fitted the data better than the best fit line, suggesting an approximate 24-h periodic expression rhythm (Fig. 2a). When NR1D1 expression was analyzed by rs2071427 genotype and by Li treatment outcome (Fig. 2b,c), the expression profiles of NR1D1 did not differ significantly between groups. Both NR1D1 PCR primer sets (directed toward exon junctions 1/2 or 6/7) showed similar results, and the ratios of NR1D1 alternate transcripts did not differ by time (not shown).

Figure 2.

Time course of NR1D1 expression. (a) Expression in LCLs (N = 12) over 24 h after serum shock varies in a nonlinear manner suggestive of an approximate 24-h periodic rhythm. Expression is similar when analyzed by (b) Genotype (N = 6/group) or (c) Li response history (N = 5 Li-R and 7 Li-NR). In all cases data are presented as percentage baseline expression ± standard error of the mean.

Variant rs2071427 and NR1D1 expression after lithium

Genetic variation in NR1D1 might also affect expression in response to Li. To assess this possibility, we measured NR1D1 transcripts in LCLs after 72 h of Li exposure. Under these circumstances, rs2071427 genotype was associated with several differences in NR1D1 expression. First, in samples homozygous for the A allele (AA), expression of the full-length NR1D1 transcript was decreased by approximately 20% following Li, whereas expression in samples homozygous for the G (GG) allele did not change. Statistical analysis showed a significant interaction of genotype × Li treatment, favoring decreased expression selectively in the AA group (F1,18 = 5.4, P < 0.05, Fig. 3a). When expression of both the full and truncated NR1D1 transcripts was measured, the AA samples had a nonsignificant approximately 15% trend toward increased NR1D1 expression after Li treatment, whereas the GG samples again did not show any change in expression (Fig. 3b).

Figure 3.

NR1D1 expression differs by genotype after Li exposure. (a) Full-length transcript expression decreases after Li preferentially in AA homozygotes at the rs2071427 locus. (b) Expression of full-length + truncated NR1D1 transcript trends toward an increase after Li preferentially in AA homozygotes. (c) Ratio of full-length/(full + truncated) NR1D1 transcript is significantly and specifically decreased among AA homozygotes at rs2071427. NR1D1 expression among GG homozygotes does not change significantly after Li. N = 10 LCLs/group. Data reflect normalized averages ± standard error of the mean. *Indicates significance of P < 0.05.

The ratio of truncated to full-length Rev-Erbα protein has been shown previously to affect the stability of the protein (Rambaud et al. 2009). When NR1D1 expression after Li was analyzed as the ratio of full-length transcript to full + truncated transcript, the effects of drug treatment on NR1D1 transcript populations were more pronounced (Fig. 3c), with a significant effect of Li (F1,18 = 5.15, P < 0.05), and a highly significant interaction between Li and genotype (F1,18 = 16.4, P < 0.001). A post hoc t-test showed that Li-treated samples with the AA genotype expressed a significantly lower ratio of full-length/truncated NR1D1 transcript compared with untreated samples (t = 3.58, df = 9, P < 0.01), whereas samples with GG genotype showed no difference after Li treatment.

We examined NR1D1 expression in a smaller number of AG heterozygotes (N = 5) with TT genotype at GSK3β rs6438552 and found expression levels that were on average intermediate between the homozygous groups in all conditions but did not statistically differ from either the AA or GG groups (data not shown).

GSK3β expression: effect of genotype, lithium and clinical history

The SNP rs334558 within GSK3β commonly called −50 T/C (Russ et al. 2001) has been previously shown to affect gene expression and transcription factor binding (Kwok et al. 2005, Lau et al. 1999) but has not been formally studied in samples from patients with BD. The rs6438552 variant that we studied has been linked to rs334558, and we confirmed this linkage in our sample, finding nearly perfect linkage disequilibrium (r2 = 0.95) between genotypes at rs6438552 and rs334558. We examined the association of rs6438552 with gene expression in a sample of 29 LCLs from patients with BD. GSK3β expression differed significantly by rs6438552 genotype, with TT samples having the greatest expression, CC having the least, and TC intermediate (F2,26 = 5.25, P < 0.05; Fig. 4). In 11 LCL samples (all TT genotype) GSK3β expression after Li treatment was analyzed by clinical response history. Exposure of LCLs to Li did not have any significant effect on GSK3β expression in either Li-R (13.59 vs. 15.10) or Li-NR (13.85 vs. 15.59, data reflect normalized GSK3β expression: baseline vs. Li treated, all P-values > 0.05).

Figure 4.

GSK3β expression varies by genotype in LCLs from patients with BD. Expression was measured in LCLs who were identified as CC (N = 2), TC (N = 9) or TT (N = 18) genotypes at rs6438552. Data were normalized to TT genotype. Post hoc tests showed that the differences between CC vs. TT and TT vs. TC were significant (P < 0.05), but that the differences between CC and TC were not significant.

No interaction between GSK3β and NR1D1 genotypes on gene expression

Finally, we examined whether GSK3β genotype altered Li's effects on NR1D1 expression. In the samples described above, NR1D1 expression for four different GSK3β:NR1D1 genotype combinations (TT:GG, TT:AA, TC:GG, TC:AA, N = 5/group) was reexamined at baseline and after Li using both NR1D1 probes. We were unable to examine the rs6438552 CC genotype. No significant differences were observed (data not shown).



The circadian clock has been implicated in BD and Li response by behavioral, genetic and biochemical studies. However, few data directly address the molecular mechanisms underlying clock dysfunction in human patients with BD. We performed a candidate gene association study of Li responsiveness in BD and made detailed measurements of clock gene function at the molecular level. In doing so, we have established a preliminary link between molecular clock function and clinical Li response in BD.

Study limitations

Our work has two major limitations. First, our retrospective genetic association study was underpowered, and the nominally significant findings did not survive corrections for multiple comparisons. Although replication will be required, the shortcomings of our work are mitigated by several important factors. Namely, the individual variants are associated with functional differences in gene expression, have high ORs, are linked in a known biochemical pathway and are robust in their additive effects, together offering sufficient evidence to warrant further investigation.

The second limitation is the use of LCLs to model BD. As a model of psychiatric illness, LCLs are constrained by the absence of specialized brain circuits, clear differences in gene expression between brain and lymphocytes, and potential artifacts caused by viral transformation of the cells. Nevertheless, we believe LCLs to be a reasonable model for the following reasons. First, the circadian clock is cell-autonomous, and does not depend on brain circuits for its function (Welsh et al. 2010). Second, despite differences, there is substantial overlap among the gene expression profiles of LCLs, whole blood and brain (Rollins et al. 2010), and although the master clock in mammals is located in the suprachiasmatic nucleus, peripheral cells including blood monocytes also express functional clock genes (Archer et al. 2008, Brown et al. 2005, Liu et al. 2007, 2008, Teboul et al. 2005, Yagita et al. 2001). Moreover, fibroblasts and blood cells have been used to show the differences in circadian gene expression between healthy humans and patients with sleep disorders and BD, suggesting that genetic factors influencing clock function are faithfully reflected in peripheral tissues (Archer et al. 2008, Brown et al. 2005, Yang et al. 2009). We found that LCLs express >40 known clock genes (Table S2), and our data suggest that NR1D1 expression is rhythmic. Collectively, this evidence indicates that LCLs are an appropriate model system for the study of clock genes in BD.

Genetic association of clock genes with lithium response

Using the LCL model, we found that two genetic variants are associated with gene expression differences. These differences plausibly link genetic variation in the circadian clock to lithium response. In addition, we also found two nominal associations of clock gene variants with Li response in patients with BD: the NR1D1 variant rs2071427 and the CRY1 variant rs8192440. We found a trend for association of the GSK3β variant rs6438552 with Li response, partially supporting the findings of Benedetti et al. (2005) who previously implicated a variant within the same haplotype. While none of these findings remained significant after correcting for multiple comparisons, there was an additive effect and a trend toward epistasis between variants of NR1D1 and GSK3β, both of which were associated with relevant functional differences in gene expression. Therefore, our analysis suggests that these effects should be reexamined in larger studies of uniformly treated BD patients that have improved power to detect gene–gene interactions. Indeed, larger replication studies are planned in the near future.

Previous pharmacogenetic studies: NR1D1 and lithium

A recent candidate gene study of Li response also reported positive findings for NR1D1, associating rs2314339 with Li-NR (OR = 0.28), but reported no association with rs2071427 (Campos-de-Sousa et al. 2010). This study population (N = 170) had only seven subjects with the minor rs2071427 allele and lacked statistical power to detect a contribution from the variant. Previously implicated in BD, rs2314339 is in partial linkage with rs2071427 (Kripke et al. 2009), and thus it remains possible that the functional variant in NR1D1 lies not at rs2071427 but elsewhere in a risk-associated haplotype.

Another report failed to associate NR1D1 with Li response at two other variants previously associated with BD: rs12941497 and rs939347 (Manchia et al. 2009). These SNPs are not strongly linked with rs2071427 (Campos-de-Sousa et al. 2010, Kishi et al. 2008). The only genome-wide association study of Li response to date failed to identify any Li-associated genetic variants in NR1D1 (Perlis et al. 2009).

Structure and function of NR1D1

Rev-Erbα, transcribed from NR1D1, is of particular interest to Li's pharmacological mechanism because of its role as a substrate for phosphorylation by GSK3β. Within NR1D1, rs2071427 is located in the first intron, a conserved region enriched in epigenetically modified chromatin (Ripperger et al. 2006). This region has been shown to serve as an alternative promoter for NR1D1, generating a truncated, functional form of Rev-Erbα protein (Rambaud et al. 2009, Triqueneaux et al. 2004). Interestingly, truncated Rev-Erbα lacks the amino acids targeted for phosphorylation by GSK3β, and additional sequences that facilitate protein turnover (Rambaud et al. 2009). As a result, truncated Rev-Erbα is a poor substrate for GSK3β but has enhanced protein stability when compared with the full-length protein. Li enhances proteasomal degradation of Rev-Erbα by inhibiting its phosphorylation by GSK3β (Yin et al. 2006). For this reason, Rev-Erbα truncation on the one hand opposes the action of Li by increasing protein stability, but on the other reinforces Li's effects by preventing Rev-Erbα phosphorylation, which could affect clock function by mechanisms other than regulating protein stability. We did not assess Rev-Erbα protein content in our study, but this may be an interesting topic for future research.

In light of these past findings, we propose that differential sensitivity of NR1D1 gene expression to Li and altered ratios of NR1D1 transcripts may improve Li response among BD patients carrying the rs2071427 A allele. The direction of the NR1D1 expression changes we report in response to Li suggest that in sensitive subjects, Li increases the ratio of truncated NR1D1 transcript relative to the full-length transcript, which could alter Rev-Erbα stability, or perhaps other processes like subcellular localization (which depends on phosphorylation in other clock proteins, e.g. PER2, Vanselow et al. 2006).

Interactive genetic effects of GSK3β and NR1D1

Our genetic association results showed that GSK3β and NR1D1 are associated with Li response additively. Our study lacked sufficient power to show epistatic gene–gene interactions, but we did find a trend in this direction. GSK3β and NR1D1 also interact biochemically in a well-characterized pathway within the circadian clock, in which Li blocks phosphorylation of Rev-Erbα by inhibiting GSK3β (Fig. 5). Therefore, any process that reduces the phosphorylation of Rev-Erbα could in principle amplify the effects of Li. As discussed previously, truncated Rev-Erbα cannot be phosphorylated by GSK3β. The mood stabilizing properties of Li have been attributed to GSK3β inhibition, and lower GSK3β expression (as seen with the CC variant of rs6438552) may facilitate this inhibition by reducing the number of active GSK3β molecules available. As both mechanisms ultimately converge on a reduction in phosphorylated Rev-Erbα, they could be combinatorial in their effects. Our genetic association results, while preliminary, suggest that this may be the case.

Figure 5.

Proposed model for the putative additive NR1D1-GSK3β genetic contribution to Li response. (a) Variant rs2071427 is associated with altered NR1D1 transcript ratios in the presence of Li. This may lead to an increase in the truncated Rev-Erbα protein isoform which is more stable, but unable to serve as a substrate for GSK3β phosphorylation. (b) Variant rs6438552 is associated with lower GSK3β gene expression, which may subsequently reduce GSK3β protein levels. (C) These mechanistically independent processes could additively oppose Rev-Erbα phosphorylation, altering protein stability and/or limiting Rev-Erbα's role in processes requiring this modification.


Our work provides an example of how using targeted functional studies of known biological interactions can enrich the study of genetic associations. As many complex processes are polygenic, composed of multiple genetic effects of small size, this convergent approach may continue to be useful in the future. Because the clock system is well characterized and accessible for direct study in samples from clinical populations, further work to clarify its role in BD may be fruitful.


This research was funded by grants from NIH (MH92758 and MH078151) and the VA (Merit Award CX-09-008) to J.R.K.; from NARSAD (Young Investigators Award) to D.K.W.; and from NIH (MH092758, AG030474), VA (HSR&D) and DOD (BUMED, CDMRP) to C.N. M.J.M. was supported by NIH T32 funding (MH018399). None of the authors have conflicting interests to report.