Dlg paralogs confer differential capacities for simple forms of conditioning and associative learning

The first part of our strategy was to use mice to ask if the duplications and divergence of the 4 Dlg genes had conferred differences in their function. Heterozygous mice of the 4 knockout mouse lines were intercrossed, and consistent with published literature, Dlg1^−/− homozygous mice were embryonic lethal in contrast to homozygous mutations of Dlg2, Dlg3 or Dlg4, which were viable^12-14. Homozygous Dlg^−/− mutations in Drosophila¹⁵ and C. elegans¹⁶ are lethal as are homozygous mutations of its murine ortholog Dlg1^−/−17, suggesting vertebrate Dlg1 retained its ancestral roles whereas the function of Dlg2-4 diversified. At the level of protein sequence, the average similarity of the 4 paralogs is approximately 75% (in either mouse or human, Supplementary Fig. 1a-c). We proceeded to test the homozygous Dlg2, Dlg3 and Dlg4 mutant mice and heterozygous Dlg1^+/− mice (as they were viable) in a battery of touchscreen tasks of increasing cognitive complexity (Fig. 1b). Across all the tasks we found that the presence of a single copy of the Dlg1 gene (Dlg1^+/−) was sufficient for normal behavior (see Supplementary Fig. 2) and hereafter we will focus our data on the differential roles of Dlg2-4.

Two simple forms of associative learning are classical (Pavlovian) and operant (instrumental) conditioning where two or more events become linked or associated, such as two stimuli or a stimulus and a response. The cognitive tasks in the rodent touchscreen battery are built on the simple instrumental conditioned response of nose-poking a stimulus displayed on the touchscreen in order to obtain a reward. The first element of the battery was therefore the acquisition of this simple form of operant conditioning by training animals through several phases in the touchscreen (see methods for details). Dlg2^−/− and Dlg3^−/Y mice displayed normal rates of completing each training phase relative to wild-type (wt) littermate controls, indicating these genes were not essential for operant conditioning (Fig. 2a). In contrast, Dlg4^−/− mice showed a marked deficit in acquisition of operant conditioning. They were able to successfully retrieve and eat reward pellets when delivery of the reward did not rely on a direct response (Phases 1 and 2, see methods), but were unable to complete the required trials when the reward was contingent on an instrumental operant response (i.e., touching the screen to attain a reward (Phase 3, see methods). To further investigate this phenotype in Dlg4^−/− mice, we employed another simple associative learning task, a test of Pavlovian conditioned approach behavior (‘autoshaping’)¹⁸. In this task, a spatially localized conditioned stimulus (CS) reliably signals an appetitive unconditioned stimulus (US), a food reward. Mice were presented with a stimulus (white vertical rectangle) displayed on either the left or the right side of the screen (Fig. 2b) and when the stimulus was displayed on the left side for example, a food reward was delivered (CS+), whereas appearance of the stimulus on the right side was never rewarded (CS−). After repeated stimulus location-reward pairings, mice normally begin to display the Pavlovian conditioned response of approaching the CS+ more often than the CS−, with the number of discriminative approaches to the CS+ and CS− serving as a measure of how well the animals have learned the association between the CS+ and the reward. Rodents show this conditioned response behavior even though there is no contingency that requires the animal to approach the stimulus in order to receive the food reward. Wt mice robustly demonstrate associative learning and develop a strong conditioned response (making greater number of approaches to the CS+ and decreasing the number of approaches to the CS− with increased training sessions, as well as showing improved approach latencies to the CS+ compared to the CS−) (Fig. 2b). In contrast to wt mice, Dlg4^−/− mice failed to demonstrate this discriminative capacity and showed equivalent approaches to the CS+ and CS− and no differences in response latencies to either the CS+ or CS−. These data so far highlight the divergence of Dlg paralogs in their contribution to simple forms of learning and information processing: unlike Dlg2 and Dlg3, Dlg4 is required for simple forms of associative learning. This requirement for Dlg4 was further highlighted in the next phase of testing where we examined all the Dlg mutant mice on a battery of tasks that involved more complex perceptual stimuli and/or required more complex response control. Dlg4^−/− mice were incapable of performing the simple operant response underlying any of the more complex tasks, consistent with the view that simple forms of associative learning are a fundamental basis and prerequisite for at least some, more complex forms of cognition.

The first of these more complex tasks was a form of learning and memory that requires a choice based on perceptual visual discrimination. Mice were presented with two stimuli simultaneously on the screen and required to learn which one was correct (i.e., rewarded, the S+) and which was incorrect (i.e., unrewarded, the S−, Fig. 2c)¹⁹. In this task the learning rate of Dlg3^−/Y mice was significantly faster than controls, requiring fewer trials and making fewer errors to learn the task (Fig. 2c). In contrast, the performance of Dlg2^−/− mice was indistinguishable from that of wt mice. This result is interesting because it indicates not only are there differential functions of Dlg2 and Dlg3 in visual discrimination learning and that neither mutation impairs basic perceptual processing abilities, but it also indicates that the Dlg3 paralog restrains or attenuates a specific aspect of the cognitive repertoire.

We next increased the associative complexity of the task by incorporating spatial information in an object-location paired associates learning task. In this task the mice were required to learn and remember which of three objects (flower, plane, spider) was associated with one of three locations on the touchscreen (left, centre, right respectively) (Fig. 2d)^{20, 21}. This task therefore requires animals to not only discriminate the objects, but also to learn the paired-association between the shape and the object’s location. On a given trial, only two different objects were presented; one displayed in its correct location (S+), the other in its incorrect location (S−), thereby allowing 6 possible trial types. Unlike the less complex visual discrimination task where the Dlg3^−/Y mutants were faster, in this task they showed normal object-location associative learning and memory. In contrast, Dlg2^−/− mice were significantly impaired and continued to perform at around 50% (chance level) (Fig. 2d). This double genetic dissociation indicates these two different forms of complex learning (visual and visuo-spatial learning) have distinct genetic regulation, each dependent on a different Dlg paralog.

Dlg2 and Dlg3 play opposing roles in cognitive flexibility, inhibitory response control and attentional processing

Complex environments confront animals not only with stable associative relationships between stimuli, responses, and outcomes, but with situations in which these relationships can change. To succeed in such environments, animals require greater response control to be able to adapt to such changes. The genes underlying such flexible behavior are unknown. Thus, having established that Dlg2^−/− and Dlg3^−/Y learn the visual discrimination task, we reversed the reward contingences so that the previously correct option was now incorrect and vice versa (S+ and S− were switched) (Fig. 3a) and thereby probed their ability to inhibit the established dominant or prepotent response and acquire the new association²². Dlg3^−/Y mutants performed normally whereas Dlg2^−/− mice showed significant impairments. When tested with simple visual stimuli, Dlg2^−/− mice showed an impairment in the early trials (the ‘perseverative’ phase of reversal learning when correct responses are low due to continued responses at the previously rewarded stimulus^{22, 23}. However, when challenged with more complex visual stimuli with greater perceptual demands, this impairment in reversal learning was more severe and found across all trials (Fig. 3b) while no differences were again observed in the initial discrimination learning (trials to criterion: Wt 502.36±58.69, Dlg2^−/− 560.43±77.01). These results show a striking dichotomy of function of Dlg2 and Dlg3 in the acquisition and reversal learning of visual discrimination: Dlg3 regulates the acquisition of the discrimination (and the mutation amplifies the rate of learning) while Dlg2 regulates the reversal or flexibility of the learned information (and the mutation attenuates the rate of reversal learning).

To examine if another form of behavioral flexibility has the same genetic requirements as reversal learning, we assessed another task for inhibitory response control using a test for extinction learning, which measure the ability to reduce responses when that response no longer results in a favourable outcome. In the touchscreen extinction task, mice are first trained to make a response to a simple visual stimulus (white square) and obtain a food reward, after which, extinction is tested by no longer rewarding the stimulus²⁴. In the absence of reinforcement, normal mice rapidly decrease their responding (Fig. 3c). Both Dlg2^−/− and Dlg3^−/Y mice displayed normal rates of learning during the acquisition phase of the task (consistent with our earlier findings that these genes are not essential for simple operant learning (Supplementary Fig. 3). However, during the extinction phase, not only were there clear phenotypes for both Dlg2^−/− and Dlg3^−/Y mice but we found evidence of their opposing function: Dlg3^−/Y mice showed faster extinction, whereas Dlg2^−/− mutants showed slower extinction (sessions 4-6) (Fig. 3c).

The capacity to select optimally when confronted with several alternative choices can put a high premium on divided attentional processing. Attention is not a unitary process but subsumes several processes including constructs such as selective and sustained attention and speed of processing. Attentional capacities can be critical for how well an animal is able to adapt and learn information about the environment. The 5-choice serial reaction time task (5-CSRTT) measures sustained, divided attention since animals need to rapidly respond to a brief visual stimulus presented randomly in one of five spatial locations to obtain a reward (Fig. 4a, see methods for detailed description). Accurate responding requires attention in both the temporal and spatial domains and moreover, the 5-CSRTT measures abnormal responses such as premature or perseverative responses, which are thought to model impulsivity and compulsivity respectively²⁵. We used a recently developed touchscreen version of this method²⁶ where mice are first trained to respond to a stimulus displayed for a duration of 2s and after they acquire a stable performance level, the duration of the stimulus is decreased requiring greater attentional capacity to accurately detect it. In this task, we again observed opposing functions for Dlg2 and Dlg3. Dlg3^−/Y mice acquired the stable level of performance at the same rate as controls (Supplementary Table 1), and as the stimulus duration decreased, they showed enhanced attentional selection (increased accuracy Fig. 4a), decreased premature responding (Fig. 4B) and a trend for decreased perseverative responses (Fig. 4c). In contrast, Dlg2^−/− mutants took significantly longer to reach stable performance at a stimulus duration of 2s, as well as the less stringent condition of 4s stimulus duration (Supplementary Table 1a). With shorter stimulus durations, Dlg2^−/− mice showed a significant impairment in accuracy, which was most pronounced at the shortest stimulus duration (0.2s, with the highest attentional load)(Fig. 4a) and made significantly more premature responses (Fig. 4b); however, perseverative responding was unaffected (Fig. 4c). These data show a remarkable divergence of function, with each of the two closely related Dlg2 and Dlg3 genes playing opposing roles on several measures of target detection and responding.

Genetic dissection of cognition meta-analysis

The systematic quantitative comparison of Dlg paralogs provides a basis for asking general questions about the organisation of the behavioral measures with respect to their underlying genetic mechanisms. We can ask three questions: i) are specific genes required for specific components of the cognitive repertoire; ii) are there differences between simple and complex cognitive behaviors, and iii) do any cognitive measures share the same genetic regulation? Figure 5a compares the results of all the touchscreen testing in the 4 lines of mice with the tasks ordered from simple to complex using the organisational scheme in Figure 1b. Enhanced phenotypes are shown in green and attenuated phenotypes in red. This analysis shows that the Dlg family is involved in the majority (8/12) of the measures of simple and complex forms of cognition. The distinct pattern of gene-phenotype relationships shows that diversity in the regulation of cognitive responses in the mouse is conferred by the set of Dlg paralogs. A complementary way to view these data is shown in Figure 5b where the gene-phenotype relationships are clustered showing 4 groups of cognitive functions (referred to as Cognitive Clusters 1-4) where each cluster consists of the behavioral measures with the same gene dependencies. In Cluster 1, simple operant conditioning is characterised by a requirement for Dlg4. Cluster 2 (object-location paired-associates learning, reversal learning, acquisition of 5-CSRTT) only requires Dlg2. In comparison, three behaviors in Cluster 3 (extinction learning, accuracy and premature responding on the 5-CSRTT) are dependent on both Dlg2 and Dlg3, with each of these genes having opposing regulatory functions. Cluster 4 (visual discrimination) requires Dlg3. Thus different Dlg genes either alone or in combination regulate specific sets of cognitive functions.

Conserved cognitive functions of DLG2 in humans

Since mice and humans diverged from a common ancestor 100 Mya, there has been strong conservation in the protein coding of Dlg orthologs (>95% similarity, Supplementary Fig. 1a-c) and other postsynaptic proteins²⁷. To ask if there has also been conservation in gene expression in brain regions of mice and humans, we correlated mRNA levels of each of the four vertebrate Dlg paralogs in 17 brain regions in both species. This analysis showed Dlg2, Dlg3 and Dlg4 were significantly correlated (Pearson’s R=0.71, 0.68 and 0.53 respectively, all p<0.05, Supplementary Fig. 1d). Recent studies of the coexpression patterns of Dlg family proteins and mRNAs indicate their importance in human brain function^28-30. While these data show conservation in protein sequence and gene regulation, it is unknown if the cognitive functions of Dlg genes are also conserved. Indeed, this has been a general problem in behavioural genetics. Although forms of learning and memory appear to be similar between humans and mice, the conservation of their genetic regulatory mechanisms has been difficult to assess in part because assessment of cognition in mice has mostly been restricted to spatial learning and memory, and there has been a limitation in the comparable nature of the cognitive tests available for rodents. Taking advantage of the fact that multiple aspects of cognition can be tested in humans (and other primates) and mice (and other rodents) in the touchscreen system using analogous tasks designed to probe the same cognitive processes, we can ask if the gene-phenotype relationships of the Dlg genes are conserved between species.

Humans carrying mutations disrupting the coding region of DLG2 have been reported^31-34 and we assessed 4 of these individuals (see methods) on a set of cognitive tasks using a touchscreen test battery, the Cambridge Neuropsychological Test Automated Battery (CANTAB). We specifically analysed 3 tasks comparable to those within the rodent touchscreen battery: 1) Intra-extradimensional set-shifting (IED) to assess visual discrimination acquisition and cognitive flexibility (tested using the visual discrimination and reversal learning task in mice) (green bars, Fig. 6), 2) Paired Associates Learning (PAL) to examine visuo-spatial learning and memory (tested using the object-location paired-associates learning task in mice) (orange bars, Fig. 6) and 3) Rapid Visual Information Processing (RVP) to assess sustained attention (tested with the 5-CSRTT in mice) (red bars, Fig. 6). Consistent with results from Dlg2^−/− mice, we found that humans with mutations in DLG2 made significantly more errors compared to controls in tests of visual discrimination acquisition/cognitive flexibility (total errors in IED: controls 27.13±1.52, DLG2 subjects 94.25±51.86, p<0.005) and visuo-spatial learning and memory (total errors in PAL: controls 16.68±0.68, DLG2 subjects 38.25±14.57, p<0.005). In addition, humans with mutations in DLG2 also showed decreased accuracy in responding compared to controls in a test for sustained attention (accuracy of target detection in RVP: controls 0.91±0.005, DLG2 subjects 0.8125±0.02, p<0.005), similar to the impaired response accuracy seen in Dlg2^−/− mice. Using the highly comparable performance measurements derived from both the mouse and human touchscreen tests, we analysed the same performance parameter (for example, total errors made) from each test to calculate a standardised performance score (z-score) compared to controls, where a negative score indicates poorer than average performance. As shown in Figure 6, comparison of the profile of cognitive phenotypes observed in human DLG2 mutations shows a striking degree of similarity to that pattern of cognitive phenotypes seen in mice with Dlg2 mutations. This similarity in the human-mouse Dlg2 cognitive profile and its uniqueness compared to the three other Dlg genes is further reinforced by published and unpublished data from another 13 different genetically modified lines of mice tested in some of the same touchscreen tasks, which do not show the selective Dlg2 phenotype profile (data not shown).

Paralog diversification, cognitive complexity and disease

Our genetic dissection in mice suggests how different components of the cognitive repertoire are related at the genetic level, and how genome evolution produced the range of vertebrate behavioral responses. Our test battery comprised 7 tests (with 13 primary measures) and each of these required the function of at least one Dlg paralog, revealing a central role of this gene family across all aspects of cognition tested. Strikingly, each vertebrate paralog had a different phenotypic profile indicating each gene has evolved a unique contribution to the cognitive repertoire. A dramatic example of this divergence was the opposing direction of the phenotypes of Dlg2 and Dlg3 in complex cognitive behaviors. Moreover, while these two genes had no role in simple conditioning, Dlg4 in contrast was uniquely essential for simple forms of learning. A parsimonious model is that Dlg4 retained an ancestral (invertebrate) role in simple forms of learning, whereas the diversification of Dlg2 and Dlg3 provided novel regulation of complex cognitive processes arising in vertebrates. The grouping of different behaviours (Fig. 5b) according to their distinct genetic underpinnings shows it is possible to identify relationships between cognitive functions based on common and distinct genetic mechanisms, which is an approach that can extend previous studies based on neuroanatomy and pharmacology^{35, 36}.

The reciprocal effects of Dlg2 and Dlg3 on complex behaviors reported here suggest these two genes play essential roles in balancing or tuning the synaptic signalling machinery. This is supported by electrophysiological studies of synaptic long-term potentiation (LTP) in CA1 synapses of the hippocampus, where Dlg2^−/− mutants have reduced LTP³⁷ and Dlg3^−/Y mutants show enhanced LTP¹³. Dlg4^−/− mutants show more severe LTP phenotypes^{14, 37} than Dlg2^−/−37 or Dlg3^−/Y13 mutants, which suggest a more severe disruption to activity-dependent synaptic strengthening is reflected in impairments in simple forms of learning. These differential roles likely reflect the distinct intracellular signaling functions mediated by Dlg proteins with their interacting proteins in the MAGUK Associated Signalling Complexes. In the accompanying manuscript³⁸ differential association between Dlg paralogs and NMDA receptors (GluN2 subunits) is reported. Our data showing the conserved role of Dlg2 in cognition in mice and humans, together with the conservation in expression between brain regions and protein sequence, indicates that it is the conservation at the genomic level that maintains these functions between the two species.

Human mutations in DLG2 and DLG3 have been reported in psychiatric disorders^{31-34, 39} and mouse models of psychiatric diseases rely on conservation of mechanisms with humans. Rare human DLG2 mutations have been associated with schizophrenia in three independent studies of copy number variants^31-34 and three of the four subjects in our study have this disease (the fourth subject is the youngest and at increased risk of developing the illness). The cognitive impairments observed in Dlg2^−/− mice parallel those observed in schizophrenia patients, such as deficits in reversal learning^{40, 41}, object-location paired-associates learning⁴², extinction⁴³ and attentional function⁴⁴. Cognitive impairments are also observed in humans with Dlg3 mutations³⁹ and we found that Dlg3^−/Y mutant mice displayed enhanced visual discrimination ability and augmented attentional function and response control. In humans, enhanced or superior performance in some cognitive domains, particularly those associated with perceptual processing is reported in autistic individuals⁴⁵. It is noteworthy that Dlg proteins interact with Neuroligin, Shank, DLGAP2 and GluN2 proteins, which are mutated in autism⁴⁶. Mutations in the Dlg family and their interacting proteins cause other diseases with a spectrum of cognitive and motor phenotypes^{27, 47}.

Our data support the model that genome duplication and diversification at the base of chordates around ~550 Mya was a driver of the expansion in complexity of the cognitive repertoire of vertebrates. This genomic mechanism, known to be important in generating complexity in other vertebrate biological systems^48,49, expanded the complexity of vertebrate synaptic signalling processes⁵⁰ before the anatomical diversification in many brain regions and encephalisation that characterises the tetrapod brain. Evidence that expansion of vertebrate postsynaptic signalling proteins is a general mechanism driving vertebrate behavioural complexity is supported by a study of GluN2 paralogs³⁸. Importantly, conservation of Dlg2’s role in human and mouse cognition over the ~90 My since these two mammals shared a common ancestor suggests that genomic mechanisms underpin these (disease relevant) behaviours despite 1000-fold differences in brain size. While on one hand these results show that genome duplication in Dlg and other postsynaptic gene families endowed vertebrates with an expanded and flexible set of cognitive functions, on the other hand it indicates that any benefits to the behavioral repertoire came at the price of susceptibility to mental illness because disease-causing mutations occur in these novel genes. Our comparative touchscreen approach also demonstrate the feasibility of co-clinical trials, using humans and mice carrying the same mutations, aimed at identifying treatments for these illnesses. Together with human genome sequencing, the quantitative testing of human cognitive functions using computerized touchscreen test batteries should aid in understanding the genetic basis of cognition and its diseases.

Animals

Dlg1 heterozygous mice (+/−) and wt littermates were generated from Dlg1 heterozygous intercrosses and maintained on 129S5/SvEvBrd background. Homozygous knockout mice (denoted by ^−/− with the exception of male Dlg3 null mutants which are denoted by ^−/Y as Dlg3 is located on the X chromosome) and wt littermates were generated from heterozygous intercrosses of Dlg2¹², Dlg3¹³ and Dlg4¹⁴ mice and maintained on a C57BL/6J background. Male and female knockout mice from all lines developed normally to adulthood, exhibited normal body size and no gross abnormalities. Mice were housed in mixed groups of wt and knockouts on a 12h light/dark cycle and all behavioural testing conducted during the light phase of the cycle. Two separate cohorts of male mice (n=10-15 for each cohort) from each knockout line were used for cognitive testing on the touchscreen tasks. Mice were maintained on a restricted diet at or above 85% of their free-feeding body weight during behavioural testing. Water was available ad libitum throughout the experiment. All experimentation was conducted in accordance with the United Kingdom Animals (Scientific Procedures) Act (1986).

Cognitive testing in the touchscreen operant system

Human DLG2 analysis - Subjects and Experimental Procedure

Four individuals with copy number variations (CNVs) within DLG2 participated in the current study (see Supplementary Fig. 4). Initial discovery of 4 unrelated cases of DLG2 CNV carriers was made in the ISC GWAS ³³ from 1115 Scottish schizophrenia cases (0.36%). From 978 Scottish control individuals screened, none were found to have this CNV. Expanding the pedigree of one of the individuals discovered in GWAS led to finding 2 more individuals within the same family with DLG2 CNVs (2 daughters; 1 diagnosed with Schizophrenia and 1 unaffected). To further explore the GWAS results, two different Multiplex Amplicon Quantification (MAQ) assays⁵² were used: the first assay included a number of chromosomal regions previously shown to contain CNVs associated with schizophrenia^{31, 32} and a second assay focused specifically on DLG2. Twelve target amplicons comprising exons of DLG2 and 11 reference amplicons were used (see Supplementary Table 2). The study was approved by the Multi-Centre Research Ethics Committee for Scotland and patients gave written informed consent for the collection of DNA samples for use in genetic studies.