Genetic Dissection of Behavioral Flexibility: Reversal Learning in Mice

Subjects

A total of 176 adult male mice (2–5 mice per strain), selected from 51 BXD recombinant inbred strains, including both older and newer strains (39), plus the relevant founder strains (C57BL/6 and DBA/2), were obtained from the University of Tennessee Health Science Center (UTHSC). The sample size per strain is justified because the value used in mapping is the mean performance of all mice that have inherited either the C57BL6 or DBA2 alleles (all mice from all strains that inherit the B allele are compared that to all mice from all strains that inherit the D allele). Thus, mapping relies on the mean of these genotype-grouped means, which will be very accurate across 51 strains, being based on about 25 strain means per genotype mean (40).

Subjects were habituated to the UCLA facility for 4–6 wks after importation, group-housed by strain in standard cages on a 14/10 light/dark schedule, initially with food and water ad libitum. Mean age at onset of training was 121+2 days. Three to five days prior to the onset of training, mice received restricted access to food in their home cage, sufficient to achieve a ~10–15% decline of their free-feeding weights at the onset of testing, adjusted up to a 20% decline during testing.

Two BXD strains have accumulated spontaneous mutations that could have influenced performance. BXD24 (JAX stock 005243) has a known mutation in the Cep20 gene that drastically impairs vision, but the BXD24 animals we studied do not carry the mutation and have normal vision. BXD29-Tlr4<lps-2J> (henceforth simply BXD29) has a known mutation in Tlr4 (JAX stock 000029). This strain also has a second mutation associated with a cortical heterotopia in the caudal neocortex (41). Despite this strain’s marked cortical abnormality, learning and reversal scores were near average (Fig 1A), and we therefore opted to include this strain in all statistical analyses and mapping.

Procedures were consistent with the Public Health Service’s Guide for the Care and Use of Laboratory Animals and were approved by the Chancellor’s Animal Research Committee at UCLA.

Training/Testing

Mice were trained and tested daily in individual Med Associates (St Albans VT) operant conditioning chambers fitted with a horizontal array of 5 nose poke apertures on one side of the box and a photocell-equipped food-delivery magazine on the other. The subjects were trained to initiate trials by nose poking into the central aperture, triggering illumination of flanking apertures. Responding in one of the two illuminated apertures was reinforced (1 food pellet; termed a correct response), while responses at the other aperture earned a time-out (all lights extinguished for 5 seconds; termed an incorrect response). The aperture reinforced during initial training was randomized across subjects and strains. All animals were tested in daily sessions that ended after: 1) one hour, 2) 126 trials, or 3) the performance criterion (16 out of 20 correctly-completed trials, computed as a moving window) was reached. Computing the performance criterion as a moving window of 20 trials ensured that mice were consistently sampling from the reinforced hole, ruling out generic hyperactivity or disinhibited responding as a factor contributing to successful stage completion. Once a performance criterion was met within a session, testing on that day was stopped and the reinforcement contingencies were reversed beginning in the following session. All mice completed the initial discrimination phase and a reversal phase.

Genetic Correlations and QTL analyses

The primary outcome measures were the number of trials required to reach performance criterion (TTC) during the initial acquisition and reversal conditions. It is customary in QTL studies, particularly of recombinant inbred lines, to control for the potential influence of potential non-genetic factors on the dependent variable(s) by a regression strategy; typically each factor is entered into a regression model and factors showing a significant relationship are entered into a multiple regression model, with the resulting residual scores submitted to genomic analysis as the phenotype. Individual subject performance data was therefore assessed for variation due to age at onset of testing, free-feeding weight, average percentage of weight decline during testing at each phase, transport batch, and operant chamber assignment. The categorical variables (i.e. chamber assignment and transport batch) were analyzed by “dummy coding” each variable. None of the known non-genetic factors accounted for a significant portion of the variation in acquisition of the initial choice or reversal (see Table S1 in the Supplement); therefore, the unadjusted TTC measures were assumed to represent the acquisition and reversal learning phenotypes. Each phenotype was then submitted to one-way ANOVA by strain to estimate a broad-sense heritability, calculated as the amount of variance explained by strain membership, reported as R² (sum of squared errors between strains divided by the overall sum or squared errors in the dataset) using SPSS v. 17.0 software. Mean TTC values were computed for each strain and used for all subsequent analyses.

The phenotypes associated with this paper are available at GeneNetwork.org (www.genenetwork.org) in the BXD Phenotype database under the record ID numbers 12730, 12731, and 12762. A comprehensive list of BXD genotypes, phenotypes, and BXD gene expression data gathered by other researchers is also available at GeneNetwork.org (select Species = Mouse, Group = BXD). These multiscalar data were used to perform both hypothesis-driven testing and global exploratory analysis. The percentage of variance explained by identified QTL was estimated using R (software version 2.12.0) and the R/qtl package (42); the locus specific heritability is calculated as 1 – 10^−2LOD/N.

Strain means for performance during the acquisition and reversal phases were submitted to separate genome-wide linkage analyses using a set of 3,797 markers spanning the genome and the mapping module of the GeneNetwork.org (43). Genome-wide linkage statistics (likelihood ratio scores) were computed using a fast regression-based method (44). Significance thresholds were computed based on 1000 permutations of the trait values (45).

Election of Positional Candidates

We used the informatics resources at GeneNetwork.org to elect positional candidates within the linkage interval (46, 47). The relevant QTL was defined as the 2-LOD support interval centered on the peak LRS (48). Array-based transcriptional phenotypes for multiple brain regions and microarray platforms were available for a subset of the strains measured in the current study. We relied most heavily on expression data for hippocampus (49), neocortex (50), and dorsal striatum (51). We also examined normative data on the mesocorticolibmic system (ventral tegmental area, nucleus accumbers, and prefrontal cortex) generated by Michael Miles and colleagues that are publically available in GeneNetwork.

We computed genetic correlations between the two primary behavioral phenotypes and the levels of mRNA expression, with particular attention to how expression of genes within the chromosome 10 QTL related to reversal learning. We reasoned that some genes and transcripts within this region would have the following attributes: 1) significant sequence differences between the parental strains, and 2) significant correlation between quantitative measures of expression and reversal learning performance, and 3) demonstrate cis-regulation (quantitative expression regulated from the same locus on chromosome 10), determined by submitting quantitative mRNA expression across the panel to QTL (eQTL) analysis. Transcripts that were correlated with reversal learning TTC at an uncorrected p < .05 level in three brain regions were mapped to determine if they were cis-regulated (52) and to assess the degree to which the regulatory region overlapped the reversal learning QTL (50). When a microarray probe targets a transcript with sequence variation between individual animals, anomalous hybridization may occur (one transcript will hybridize more efficiently than the other by virtue of being a better match for the probe, rather than because there are higher levels of it), resulting in spurious eQTLs (53). We therefore filtered probes that overlapped B-vs-D SNPs.

Descriptive Statistics

Figure 1 shows that the initial discrimination acquisition (panel A) and reversal learning (panel B) phenotypes are distributed continuously, though differently, across the BXD strains: an outcome indicative of complexly-determined traits (also see Table S2 in the Supplement). Both phenotypes were moderately heritable (R² = 0.29 for acquisition and R² = 0.31 for reversal), adequate for mapping traits with the current sample sizes (40). A paired-samples two-tailed t-test (TTC_acquisition – TTC_reversal) revealed that reversal (mean = 173.2 trials, SD = 56.4; range 77 to 310.5) was significantly more difficult than acquisition (mean = 106.6 trials, SD = 45.3; range 23.6 to 200, t(174) = − 9.12, p < .001). The strain means were not significantly correlated across experimental phases (N = 51, r = 0.220, p = .12), and several patterns of strain variation between phases are evident. For example, the BXD16 strain performed poorly on both phases, suggesting poor learning or discrimination generally, while BXD2 performed very well on acquisition but very poorly on reversal, suggesting a reversal-specific effect in that strain. More commonly, strain ranks shifted by moderate amounts in either direction across phases (i.e. BXD13 improved by 20 rank positions between phases, while BXD12 dropped back by 22 rank positions, etc.). The diversity of strain variation between experimental phases suggests that the genetic influences on the cognitive and behavioral demands of each phase are at least somewhat independent.

Prior to conducting the genome scan, we considered whether the behavioral measure used in mice (reversal learning) bears mechanistic relationships to related human phenotypes, as hypothesized. If there were common molecular correlates of reversal learning in mice and relevant phenotypes in humans, this would support the idea that the relationship between these measures is more than simply descriptive. In human subjects, trait impulsivity, which is linked to reversal learning and inhibitory control (54), is related to low ventral mesencephalic D2-like receptor availability, measured in vivo (6), and the Genenetwork.org phenotype database contains data on dopamine transporter density, D1 protein density, and D2-like receptor density in prefrontal cortex (PFC), nucleus accumbens, dorsal striatum and ventral mesencephalon in a subset of BXD strains (55). We found that D2-like receptor expression in the ventral mesencephalon and PFC are inversely related to TTC in reversal (poor performance) in our sample (ventral mesencephalon: r = −0.75, p < 0.01, n = 11; prefrontal cortex: r= −0.67, p < 0.01, n=11; dopamine transporter: r = −0.70, p < 0.001; see Figure S1 in the Supplement). Other measures did not significantly correlate. These findings indicate that the current reversal learning phenotype in mice relates to a set of molecular correlates of human traits.

Genome Scan

No significant or suggestive QTL were observed for performance during the acquisition of the discrimination (data not shown), but a genome-wide significant QTL was mapped to a locus on chromosome 10 (77–82.4Mb) for the reversal learning phenotype (see figure 2A). Peak LRS-values were noted around marker CEL-10_86482491 (86.197002Mb; maximum LRS = 19.29, genome-wide p< 0.05). The locus-specific heritability was estimated at 31.5% based on a single QTL model. While variability within strain was generally low throughout the sample, a few strains in the reversal condition had notably high standard errors. If the two strains with the highest errors BXD2 and BXD29 (exceeding 60% of the strain mean) are removed, the QTL scan generates statistically identical results (LRS = 18.567 around marker CEL-10_86482491, genome-wide p < 0.05), demonstrating the robustness of the effect.

Despite the lack of correlation between the acquisition and reversal phenotypes, it remained possible that some small amount of the variance driving the reversal QTL is due to general learning abilities rather than the specific ability to learn a reversal. To rule out this possibility, TTC for the reversal phase was regressed on TTC for the acquisition phase at the individual subject level, and the mean residual scores per strain (the portion of variance in the reversal data that could not be accounted for by variance in the acquisition data) were submitted to a second genome scan, yielding statistically identical results on chromosome 10, around marker CEL-10_86482491 (LRS = 17.487, p < 0.05; see figure 2B). It is therefore reasonable to assume that the chromosome 10 QTL is specific to those cognitive functions required for reversal, but not acquisition, namely inhibitory control over a pre-potent response.

The Region of Interest and Specification of Positional Candidates

The confidence interval around the peak LRS defining the genomic ROI for reversal learning ranged from 77Mb to 92.4Mb on mouse chromosome 10, which is syntenic with portions of human chromosomes 12, 19 and 22. The ROI contains 288 previously-named transcripts according to the GeneNetwork.org interval analysis tool (http://genenetwork.org/webqtl/main.py?FormID=intervalAnalyst), including both known and predicted genes. Of these 288 transcripts, 141 contained sequence variation in the BXD panel (see Table S3 in the Supplement); array-based expression for each of these transcripts was subjected to correlation analyses with reversal learning performance (see Tables S4-6 in the Supplement for correlated probes by brain region). Expression phenotypes for 7 unique transcripts (29 transcript/probe combinations) correlated with the reversal learning phenotype consistently in 3 different brain regions and across two microarray platforms (uncorrected p < 0.05), indicating that the expression data are reliable and the relationships with the reversal learning phenotype are robust (see table 1). Finally, three of these transcripts resulted in significant eQTL in (Hcfc2, Syn3 and Nt5dc3; uncorrected p < 0.05) that substantially overlapped the reversal learning ROI, also in all three brain regions (see figure 3). According to the Allen Mouse Brain Atlas (http://mousebrain-map.org/welcome.do), all three candidate genes are expressed in brain tissue.

Positional Candidates

Variation in the expression of genes within a QTL may be useful for delineating positional candidates for further analysis (56). This argument stems from the concept that continuous variation in a trait is likely linked to genetically-determined continuous variation in expression of genes within the QTL (47, 57, 58). Transcriptome data were available for three brain regions in a large sample of BXD mice, and we were able to use these data to identify seven genes whose expression consistently correlated with the reversal learning phenotype in datasets generated from three separate regions, despite the fact that the genomic data was obtained by different investigators, sampling from different mouse cohorts and using different array platforms. Of these robustly correlated transcripts, three were advanced as high priority candidates because their quantitative expression is controlled from the same locus that influences reversal learning performance.

Of the positional candidates elected, Syn3 has been previously studied for its regulation of synaptic function. The gene codes for synapsin III, a member of a family of neuron-specific phosphoproteins known to affiliate with the cytoplasmic surface of synaptic vesicles (59), regulate vesicular reserve pools, and influence calcium-dependent synaptic dynamics (60–62). While all of the synapsin proteins influence vesicular trafficking, synapsin III demonstrates unique functions under a variety of circumstances. For example, evoked inhibitory post-synaptic currents are reduced in amplitude in hippocampal neurons of Syn3 knockout mice, while evoked excitatory post-synaptic responses are unaffected (60). Synapsin III knockout mice also demonstrate larger and longer striatal dopamine transients, measured by sub-second voltammetry, in response to electrical stimulation (63), while serotonin function is not affected; this change is consistent with reduced dopamine uptake, and it is noteworthy that reversal learning is correlated with dopamine transporter density in some regions of forebrain. The current findings demonstrate that increased Syn3 mRNA levels relate to better reversal learning performance, which together with Kile et al. invites a DA release dysregulation hypothesis of impulse control, mediated by Syn3 function.

The relationship between Syn3 and reversal learning may, therefore, involve intermediate adaptations in dopamine system dynamics. Separately, we showed that D2 receptor expression in the ventral midbrain (autoreceptor expression) negatively correlates with reversal learning; furthermore, Syn3 expression tends to positively relate to ventral midbrain D2 levels (n = 17, r = 0.451, p = 0.069), providing some evidence of a genetic relationship between Syn3 expression and D2 function. Trait impulsivity is related to diminished D2 autoreceptor activity (6), and other animal studies show that reducing D2 activity genetically or pharmacologically interferes with reversal learning (64–66), which is interesting in light of these findings, as together they suggest a common pathway to poor impulse control – namely, Syn3 modulation of dopamine dynamics, leading to behavioral inflexibility and impulsivity.

Human Syn3 resides on chromosome 22 (22q12), near a genomic locus that has been repeatedly implicated in psychosis (67); and deficits in reversal learning are often observed among schizophrenia patients (68, 69). Indeed, human studies of polymorphisms in Syn3 and post-mortem Syn3 expression have sometimes implied a relationship to schizophrenia, but these results are neither consistent nor conclusive (70–75). However, the Syn3 locus spans 545,839 bases and at least 3 haplotype blocks. We estimate that more than 125 SNP markers would be required to fully impute variation in the transcript. The majority of Syn3-schizophrenia studies, interrogating but a few SNPs at most, are therefore inconclusive. While questions about Syn3 and schizophrenia remain, the current finding substantially advances understanding of the genomic locus and its relation to impulse control, which is where the future direction of Syn3 research is best aimed. If there is a relationship between Syn3 and schizophrenia, it is most likely a relationship with a symptom-type that plays a small role in the overall syndrome, but is more characteristic of other neuropsychiatric disorders, like ADHD and behavior addictions.

Nt5dc3 (5′-nucleotidase domain containing 3), a member of a 50-nucleotidase domain containing family, codes for a largely uncharacterized transmembrane protein. Polymorphisms in Nt5dc3 and surrounding loci on human chromosome 12 have been associated with ADHD in two genome-wide mapping studies. Fisher et al. (2002) reported weak evidence for linkage with ADHD diagnosis at markers near the 12q23 locus (D12S78–D12S79, p = 0.0039), and in a more recent study, Lesch et al. (2008) identified a SNP within Nt5dc3 (rs4964805) as one of the top 30 out of ~500,000 SNPs (p = 4.74X10-6). A marker at a locus (rs1862032) near human Nt5dc3 (12q23.3) was also suggestively associated with a performance on a stop signal reaction time task (LOD = 2.627, uncorrected p = 0.0003), another measure of inhibitory control function(76). While none of these findings were significant at genome-wide levels on their own, their convergence was worthy of note even before the current study. The current finding alone, however, does provide convincing evidence for genome-wide significant linkage of a clinically relevant translational phenotype with the homologous locus on the mouse genome. Further investigation of Nt5dc3 as a risk gene for disorders of impulse control, including ADHD, is well warranted, as is basic neuroscience research to specify the functions of the encoded protein.

Hcfc2 (host cell factor C2) is also near the human 12q23 locus and is known to interact with a specific herpes simplex protein, VP16, during the infection process. (77). To date, no regulatory functions or interactions with endogenous proteins have been characterized for this transcript. Therefore, Hcfc2 represents the potential for identifying a completely novel genetic mechanism with implications for cognitive control.

Conclusions

We have implemented, in a murine genetic reference population, a task that measures a clinically relevant cognitive-behavioral phenotype, and have shown that performance on the task shares mechanistic features of the same cognitive-behavioral function in humans. Using this approach, we identified important new genomic mechanisms that explain individual variation in the inhibitory control phenotype and that likely have substantial implications for homologous human phenotypes and related neuropsychiatric disorders. While the methods applied to the election of positional candidates for the current QTL cannot convincingly eliminate other transcripts from consideration as potential contributors to the observed phenotypes, they do provide strong evidence for inclusion of the candidates that were advanced, and vigorous follow up investigation of these candidates and their functions is warranted. In addition to adding converging evidence supporting human findings implicating Syn3 and Nt5dc and their surrounding genomic loci to neuropsychiatric disorders, the current study, definitively and for the first time, demonstrates the power of murine genetic reference populations for uncovering the genomic bases for phenotypes of interest to neuropsychiatry. This demonstration of the translational value of the reversal learning phenotype and the elucidation of its genetic underpinnings strongly suggest that the mouse is a powerful system for advancing the understanding of the etiology of impulse control disorders, and their potential treatments.