Angelman syndrome (AS) is a neurodevelopmental disorder caused by loss of expression of the maternal copy of the UBE3A gene. Individuals with AS have a multifaceted behavioral phenotype consisting of deficits in motor function, epilepsy, cognitive impairment, sleep abnormalities, as well as other comorbidities. Effectively modeling this behavioral profile and measuring behavioral improvement will be crucial for the success of ongoing and future clinical trials. Foundational studies have defined an array of behavioral phenotypes in the AS mouse model. However, no single behavioral test is able to fully capture the complex nature of AS—in mice, or in children. We performed multidimensional analysis (principal component analysis + k-means clustering) to quantify the performance of AS model mice (n = 148) and wild-type littermates (n = 138) across eight behavioral domains. This approach correctly predicted the genotype of mice based on their behavioral profile with ~95% accuracy, and remained effective with reasonable sample sizes (n = ~12–15). Multidimensional analysis was effective using different combinations of behavioral inputs and was able to detect behavioral improvement as a function of treatment in AS model mice. Overall, multidimensional behavioral analysis provides a tool for evaluating the effectiveness of preclinical treatments for AS. Multidimensional analysis of behavior may also be applied to rodent models of related neurodevelopmental disorders, and may be particularly valuable for disorders where individual behavioral tests are less reliable than in AS.
Rodent models have enabled mechanistic insights into the genetic causes and circuit-level manifestations of single-gene neurodevelopmental disorders (NDDs) [1,2,3,4,5,6,7]. As mechanism-based treatments are developed for NDDs, the effectiveness of such treatments is often first tested preclinically by assessing improvements in mouse behavior. Behavioral phenotypes span multiple domains in individuals with NDDs (e.g., cognitive, motor, seizures, sleep), and a wide range of corresponding behavioral assessments have been developed and deployed in mouse models [8,9,10,11]. Accurately measuring phenotypic severity across multiple behavioral domains is critical to properly assess the effectiveness of treatments in rodent models of NDDs.
We hypothesized that multidimensional analysis of mouse behavior would enable quantification of overall behavioral severity, aggregated across multiple domains. Here we define multidimensional analysis as the multi-step process of: (a) reducing the dimensionality of large behavioral datasets using principal component analysis (PCA) [12, 13], (b) clustering data in principal component space using k-means clustering, and (c) assessing whether behaviorally defined clusters align with animal genotype. Dimensionality reduction and clustering have been validated in various mouse behavioral contexts [14,15,16,17,18,19,20,21,22]. Here, we tested the hypothesis that multidimensional analysis of mouse behavioral data could accurately distinguish the genotype of Ube3a mutants (a model of Angelman syndrome (AS)) from wild-type littermates. AS is an ideal disorder for testing the effectiveness of multidimensional analysis because of recent progress in developing mechanism-based treatments  and because behavioral testing can be performed reliably in Angelman model mice .
AS is a NDD caused by lack of expression of the maternal allele of UBE3A, an E3 ubiquitin ligase located on chromosome 15 [25,26,27]. Individuals with AS have a multifaceted behavioral phenotype that typically includes cognitive impairment, motor impairment, lack of speech, seizures, and disrupted sleep [28,29,30]. While mutations in maternal UBE3A are sufficient to cause AS, the majority of cases (~70%) are caused by deletions of a region of maternal chromosome 15q11–13 spanning UBE3A and neighboring genes . The paternal UBE3A allele is epigenetically silenced in neurons by expression of a UBE3A antisense transcript (UBE3A-ATS) [31, 32], and neuronal paternal Ube3a imprinting is conserved  in the mouse model of AS (Ube3am-/p+) . Multiple approaches have successfully unsilenced the dormant paternal Ube3a allele in mice [35,36,37,38,39,40,41,42], and one such approach (antisense oligonucleotides targeted to UBE3A-ATS) is currently in early-stage clinical trials (NCT04428281, NCT04259281, NCT05127226). Other AS treatments such as gene replacement therapy and targeting downstream processes are also in preclinical development . Developing a pipeline to test new AS treatments in mice will be critical to evaluate their success, regardless of treatment mechanism. Recent work established a “gold standard” behavioral battery consisting of five tests (rotarod, open field, marble burying, nest building, forced swim) that are reliably impaired in Ube3am-/p+ mice  and are sensitive to treatment [40, 44,45,46]. Using this battery, we hypothesized that multidimensional analysis: (a) would enable quantification of behavior across multiple domains as a single “severity score,” (b) that this severity score would be a reliable indicator of Ube3a genotype, and (c) that this severity score is sensitive to treatment.
Methods and materials
We performed multidimensional analysis using three mouse behavioral datasets. All datasets assessed behavior in male and female AS model mice (Ube3am-/p+)  and wild-type littermate controls (WT; Ube3am+/p+) with experimenters blind to genotype. For all datasets, experimental WT and Ube3am-/p+ littermates were generated by crossing female Ube3am+/p- mice and male WT mice. Dataset 1 used experimental mice on an F1 hybrid 129S2-C57BL6/J background, and Datasets 2 and 3 used experimental mice on a congenic C57BL6/J background. All experimental protocols were conducted in accordance with the European Commission Council Directive 2010/63/EU (CCD approval AVD101002016791; Dataset 1), or were approved by the Institutional Animal Care and Use Committee (IACUC) of Children’s National Medical Center (Dataset 2). Dataset 3 contained only previously published data and no additional mouse experiments.
Dataset 1 was used for the majority of analyses (Figs. 1–3, 4a, 5, S1–S8, and S10) to assess the effectiveness of multidimensional analysis at predicting Ube3a genotype based on behavior. Dataset 1 included 286 total mice (WT: n = 148, Ube3am-/p+: n = 138) run across 10 cohorts at Erasmus Medical Center. Behavioral data from 8 of these 10 cohorts (n = 231/286 mice) were previously published  and 2 additional cohorts (n = 55 mice) were also included. A detailed table showing the genotypes, sex, and behavioral tests performed in each of the ten cohorts is shown in Fig. S1. Behavioral testing was performed in P60–P90 mice.
Dataset 2 was used to test the hypothesis that multidimensional analysis of behavior can effectively predict Ube3a genotype using small sample sizes (Fig. 4b and S9). Dataset 2 included 24 mice (WT: n = 12, Ube3am-/p+: n = 12) tested at Children’s National Research Institute. Behavioral testing was performed in P60-P90 mice.
Dataset 3 was previously published (Wolter et al.; Supplementary Fig. 4)  and was used to assess whether multidimensional analysis could detect behavioral improvement in Ube3am-/p+ mice treated with CRISPR-Cas9-based targeted treatment to unsilence the dormant paternal Ube3a allele (Fig. 6 and S11). Mice were treated with either viral expression of a SaCas9 gRNA targeting the Ube3a-ATS locus (Sajw33), or a negative control gRNA that did not unsilence paternal Ube3a. Bilateral i.c.v. AAV delivery of Sajw33 or control occurred at both E15.5 and P1 within the same animals, and behavioral testing began at 4 weeks and continued through 40 weeks . Behavioral data included WT + control (n = 34), Ube3am-/p+ + control (n = 25), and Ube3am-/p+ + treatment (Sajw33; n = 32) groups.
Mice in Datasets 1 and 2 were weighed and then underwent a series of behavioral tests in the same order: rotarod, open field, marble burying, nest building, and forced swim (Fig. 1a). A subset of mice in Dataset 1 lacked data from three of these tests (weight, open field, nest building; Fig. S1a). Methods for Dataset 1 were previously published  and methods for Dataset 2 (below) were based upon Sonzogni et al. , with minor modifications.
Mice were placed on a rotating bar that accelerated from 4 to 40 rpm across 5 min at an acceleration rate of 7.2 rpm2 (Ugo Basile model #47600). Trials were complete once the mouse fell off, if three consecutive wrapping rotations were made, or if 5 min elapsed. Each day, the results of two trials with an inter-trial interval of one hour were averaged together. Experiments were run over the span of five consecutive days, with a two day interval before open field testing.
Open field test
Mice were placed into a 42 cm square open field arena (AccuScan Instruments, Inc., Columbus, OH) and were allowed to freely move for a single 30 min trial. The center square was defined as 21 cm × 21 cm. The data was collected using the open field activity monitoring system (Omnitech Electronics, Inc. SuperFlex Open Field System) which uses photocell emitters and receptors forming an x–y grid of infrared beams. Total distance moved and time spent in the center square were recorded using infrared beam break information.
Mice were placed individually in a 16 × 8 in cage with ~4 cm of bedding (Bed-o’Cobs 1/4” bedding) and 20 black glass marbles arranged in a 5 × 4 array for a single 30 min trial. A marble was considered buried if it was >50% covered with bedding at the end of the trial.
Immediately after marble burying, mice were habituated to single housing as well as new nesting material (Bio-Rad 7.5 × 10 cm extra thick block filter paper; 11 ± 1 g) for 5–7 days prior to testing. During testing, new nesting material was introduced on day 1 and unused material was weighed daily across five days.
Forced swim test
Forced swim testing was performed on the same day immediately following the final day of nest building. Mice were placed in a 9 in cylindrical × 9.25 in tall tank filled with 23 ± 1 °C water to a height of ~60% of the container’s height. Trials lasted 6 min: the first 2 min were an acclimation period and the last 4 min were to record immobility (lack of movement or only necessary movements to keep head above water) as a percentage of total recording time. Immobility was recorded manually using a stopwatch.
Mice in Dataset 3 underwent a different series of behaviors in consistent order: hindlimb clasping, rotarod, open field, marble burying, fear conditioning, rotarod (again) (Fig. S11a). Methods for Dataset 3 were previously published .
Multidimensional analysis of mouse behavior
Multidimensional analysis of behavior consisted of a series of steps: data selection, standardization, principal component analysis (PCA), k-means clustering, and validation (Fig. 1i) . Data selection: Eight total measures (weight, rotarod day 1, rotarod day 5, open field distance traveled, open field center time, marbles buried, nest building, forced swim floating time) from a series of six tests (weight, rotarod, open field, marble burying, nest building, forced swim) were included in multidimensional analysis. Tests with multiple measures include rotarod (day 1 and day 5 performance) and open field (total distance and time spent in the center of the arena). Measures considered redundant (e.g., intermediate, non-independent timepoints for rotarod and nest building) were excluded a priori from multidimensional analysis. For the majority of analyses (Figs. 2, 3a–d, 4a, 5, S3–S6, S8c, and S10), we included the subset of animals that underwent all six tests (WT: n = 88, Ube3am-/p+: n = 82). One outlier was excluded from analysis based on its position in 2PC space (PC1: −2.49, PC2: 8.62) using Grubbs’ test for outliers (alpha = 0.05), leaving a total n = 81 Ube3am-/p+ mice. Standardization: All measures were standardized using a z-score (z = (data point − group mean)/standard deviation) to account for different units across tests. Prior to standardization, we tested whether each behavioral measure showed significant sex differences using a two-way ANOVA with sex and genotype as factors (Fig. S2). All behaviors that showed either a significant main effect of sex or sex × genotype interaction were standardized separately in male and female animals (Fig. 1i), except where noted (Fig. S6; a separate analysis to assess the effectiveness of multidimensional analysis if sex is not accounted for). Measures with no sex differences were standardized using the entire group. Principal component analysis: We performed PCA using the pca() function in MATLAB and calculated the amount of variance explained by each PC using a Scree plot (Figs. S3, S9h, and S11b) and the loading distribution of principal components using the coefficient outputs from PCA (Figs. 5a, b and S11c). k-means clustering: k-means clustering was performed in principal component space using the kmeans() function in MATLAB with k = 2 clusters. Except where noted (Figs. 5, 6c, S4, and S11d, g), clustering was performed in 2PC space. In Figs. 5c, 6c, and S11d, g, one PC was used for clustering. Figure 5d, e evaluated using three and four PCs for clustering. In Fig. S4, clustering was performed on raw data (not in PC space). Validation: We compared the actual genotypes of animals to their assigned cluster and calculated the percentage correct (accuracy of clustering, Figs. 2c, 3b, c, f–h, 4b, 5c–e, S5a, S6, S10, and S11d).
In order to assess the generalizability of multidimensional analysis, we varied the input parameters of analysis in three ways (Fig. 3). First, we performed multidimensional analysis on Dataset 1 (n = 169; standardized by sex) in 2 PC space under 11 conditions (Fig. 3a–c). Each condition represented the removal of a single behavioral measure or a complete behavioral test. Next, we performed multidimensional analysis using every possible combination of 3–8 total measures (Fig. 3d). Finally, we performed multidimensional analysis on different combinations of data where more animals could be included with partial behavioral profiles (Fig. 3e–h and S1). This approach resulted in four conditions with the number of behavioral measures ranging from 4 to 8 and the number of animals ranging from 169 to 286 per condition.
To assess the minimal sample size needed for multidimensional analysis to effectively classify Ube3a genotype, we performed a bootstrap analysis of Dataset 1 (n = 169) using 2 PCs for clustering (Fig. 4a). For n = 3 to n = 30 with a step of n = 1, we randomly selected n animals per genotype from the overall dataset (n = 169), with replacement, and performed multidimensional analysis. We repeated 10,000 trials for each n and report the average clustering accuracy.
To determine whether multidimensional analysis results in false positive effects, we performed multidimensional analysis in a homogenous sample (all 36 wild-type females). We randomly assigned mice to two groups (to model two genotypes with no true behavioral difference; Fig. S7).
Behavioral rescue with Ube3a reinstatement by Cas9 gene therapy
Behavioral data was available for a total of n = 91 mice in Dataset 3. However, every behavioral test was not performed in every animal. Thus, we performed multidimensional analysis in two subsets (“conditions”) of the total sample (Fig. S11e). Condition 1 (Fig. 6) included n = 33 total mice where behavioral data was available for each of thirteen total measures. Condition 2 (Fig. S11f, g) included n = 86 total mice where behavioral data was available for six of thirteen total measures. Data were standardized by sex for measures where statistically significant sex differences were observed between WT + control and AS + control groups in the total sample (P90 weight, open field center time).
Statistical analysis was performed using GraphPad Prism 9 and MATLAB R2019a, R2021a, and R2022a (Mathworks). Group comparisons on individual behavioral tests (Figs. 1 and S9) were made using Student’s t-tests (weight, open field distance, open field center time, marble burying, forced swim) and two-way RM ANOVA (rotarod, nest building). Assessment of sex differences (Figs. 2, S5c, and data not shown to accompany Fig. 6/S11) were made using two-way ANOVA. Effect of treatment was assessed using one-way ANOVA and post hoc Tukey’s multiple comparison test in 1 PC space (Figs. 6c and S11g). Figure 3d used one-way ANOVA and post hoc Tukey’s multiple comparison test. Figure S7d used Student’s t test. Fisher’s exact test was used to determine if clustering accuracy was statistically different between test conditions (Figs. 3a–c, S4, and S6). For all figures, *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001.
Ube3a m-/p+ mice have robust behavioral impairments
Ten independent cohorts totaling 286 mice (Fig. S1; 26–30 mice per cohort) performed a series of behavioral tests in order: weight, rotarod, open field, marble burying, nest building, and forced swim (Fig. 1a). Behavioral data from eight of these cohorts (n = 231) were previously reported by Sonzogni and colleagues . In the complete dataset, Ube3am-/p+ mice showed increased weight (Fig. 1b; t(226) = 4.428, p < 0.0001), impaired rotarod performance (Fig. 1c; main effect of genotype: F(1,284) = 103.8, p < 0.0001), and impaired rotarod learning (Fig. 1c; genotype × time interaction: F(4,1136) = 6.792, p < 0.0001) relative to WT controls. Ube3am-/p+ mice were hypoactive in an open field (Fig. 1d; t(226) = 8.874, p < 0.0001) despite normal time spent in the center of the arena (Fig. 1e; t(226) = 1.348, p = 0.1789). Ube3am-/p+ mice also showed impairments on marble burying (Fig. 1f; t(284) = 16.79, p < 0.0001), nest building (Fig. 1g; main effect of genotype: F(1,207) = 176.1, p < 0.0001), and forced swim tests (Fig. 1h; t(284) = 15.73, p < 0.0001) relative to WT controls. Prior work found significant sex differences in a subset of these behaviors , and we confirmed with the complete dataset that six of eight behavioral measures showed statistically meaningful sex differences (Fig. S2).
Multidimensional analysis of behavior predicts Ube3a genotype with high accuracy
From the group of six behavioral tests, we selected eight non-redundant measurements to include in multidimensional analysis (Fig. 1i). We performed multidimensional analysis using data from mice where all behavioral tests were performed within individual animals (n = 169/286; 6/10 cohorts; Fig. S1). Multidimensional analysis consisted of four steps: standardization, principal component analysis (PCA), k-means clustering, and validation (Fig. 1i). First, all measures were standardized using a z-score and measures with sex differences were standardized separately by sex. PCA revealed that two principal components (PCs) are likely sufficient to capture a majority of variance in the dataset (Fig. S3); thus, we performed k-means clustering in 2PC space (Fig. 2a). Validation comparing clusters (Fig. 2a) to the actual genotype of animals (Fig. 2b) demonstrated that multidimensional analysis was 94.7% accurate in predicting Ube3a genotype from behavior (Fig. 2c). Multidimensional analysis performed better than k-means clustering of performance on individual assessments, which ranged from ~55–85% accuracy (Fig. S4). Male and female mice were equally distributed in PC space (Fig. S5), confirming that sex differences in behavior were accounted for by standardization, and likely do not represent an additional source of meaningful variance. Not accounting for the sex dependence of behaviors slightly reduced the accuracy of multidimensional analysis (from 94.7% to 91.1%), though this difference was not statistically meaningful (Fig. S6). To confirm the validity of multidimensional analysis, we demonstrated that this approach does not detect “false positive” differences in behavior between two randomly assigned groups within a homogenous group of animals (all wild-type females) (Fig. S7).
Multidimensional analysis remains effective with different combinations of behavioral input
For multidimensional analysis to be a valuable tool to quantify behavior in mouse models of AS and related disorders, it should generalize across multiple combinations of behavioral input. To address this question, we first performed multidimensional analysis on the same dataset (n = 169) under conditions where each individual measurement (e.g., rotarod day 5) or overall test (e.g., rotarod) were excluded from analysis. Removal of individual measures or tests resulted in a range of 92.3–96.4% accuracy, which was not statistically different from the baseline of 94.7% (Fisher’s exact tests; lowest p = 0.5092; Figs. 3a–c and S8a). Next, we assessed the clustering accuracy in each of 219 possible combinations of including between three and eight behavioral measures in analysis. Clustering accuracy decreased on average as the number of measures decreased (Fig. 3d; one-way ANOVA, F(4,261) = 21.31, p < 0.0001). We also expanded the dataset to include all animals (n = 286), and defined four conditions where different combinations of behavior were available for different subsets of animals (Fig. S1). As the number of measures decreased, the accuracy of multidimensional analysis decreased despite the sample size increasing (Figs. 3e–h and S8b). Together, these data suggest that multidimensional analysis generalizes well across different combinations of behavioral data in the Ube3am-/p+ mouse model, and gains effectiveness as more behavioral tests are included in analysis.
Multidimensional analysis accurately predicts Ube3a m-/p+ genotype with reasonable sample sizes for mouse behavior
In practice, a behavioral study requiring >80 mice per genotype to detect group differences would likely be time and cost prohibitive. Here, we asked whether multidimensional analysis would retain high accuracy in groups with smaller sample sizes. To address this question, we performed a bootstrap analysis to predict the clustering accuracy that could be achieved across sample sizes ranging from n = 3–30 per group (Figs. 4a and S8c). We then performed multidimensional analysis on a behavioral cohort tested prospectively in a different laboratory with 12 animals per group (Dataset 2, Fig. S9). The bootstrap analysis predicted an accuracy of 90% using n = 12, and multidimensional analysis achieved 100% accuracy in this new cohort (Figs. 4b and S9). We hypothesized that accuracy was higher in a new cohort because the bootstrap analysis pulls animals randomly across multiple cohorts, introducing inter-cohort variability. We tested this hypothesis by performing multidimensional analysis separately in each of six individual cohorts from Dataset 1 (Fig. S9). Accuracy in each of these cohorts also outperformed the expectations of the bootstrap analysis, ranging from 92–100% with sample sizes of ~13–15 animals per genotype (Fig. 4a; open circles). The cohort sizes tested here are comparable to the sample sizes determined by Sonzogni et al. to be required to detect group level differences on individual behavioral tests (n = 7–21, depending on test) . Overall, this work confirms that multidimensional analysis of behavior is highly accurate in Ube3a mice using typical sample sizes for mouse behavior.
Multidimensional analysis remains effective using a single PC for clustering
A potentially valuable use of multidimensional behavioral analysis is the possibility of summarizing an animal’s overall phenotypic severity as a single number (PC1). Such an approach will only be valuable if (a) PC1 represents a substantial amount of the total variability in the dataset and (b) PC1 alone is sufficient to accurately predict Ube3a genotype. PC1 accounts for 37.7% of variance in the dataset (Figs. S3 and 5a) and correlates strongly with each behavioral measure tested (Fig. 5a, b). Clustering accuracy remained high (94.7%) using a single principal component for k-means clustering (Fig. 5c). Increasing the number of PCs to 3–4 (accounting for up to 75% of total variation) provided little additional benefit in predicting Ube3a genotype (Fig. 5d, e). Overall, these results suggest that PC1 alone is sufficient to accurately predict Ube3a genotype.
Multidimensional analysis can detect behavioral improvement following paternal Ube3a unsilencing
Behavioral testing in mouse models of neurodevelopmental disorders is often used to test the preclinical effectiveness of treatments . We performed multidimensional analysis of behavior in a group of mice where CRISPR/Cas9-based targeting of Ube3a-ATS enabled unsilencing of the paternal Ube3a allele and re-expression of UBE3A protein . In this cohort (Dataset 3), the series of behavioral tests performed was similar but not identical to the series of tests performed by Sonzogni and colleagues  (Figs. 6a and S11a). In this dataset, PC1 represented 31.0% of total variance and clustering accuracy was 100% between WT and Ube3am-/p+ non-drug control groups when using 1 PC for clustering (Fig. S11b–d). CRISPR/Cas9 treatment resulted in an amelioration, but not full correction, of overall behavioral severity in Ube3am-/p+ mice as measured by PC1 (Fig. 6b, c; main effect of group: F(2,30) = 93.66, p < 0.0001; post hoc WT/control vs. AS/control: p < 0.0001; post hoc AS/control vs. AS/treatment: p < 0.0001). We observed a significant effect of treatment using PC1 under multiple conditions where different combinations of behavioral measures were included in the analysis (Fig. S11e–g).
PUMBAA: A graphical user interface for multidimensional analysis of behavior
Multidimensional analysis may be valuable for other behavioral datasets in Ube3am-/p+ mice and mouse models of related disorders. To enable widespread use of multidimensional behavioral analysis, we developed a graphical user interface for phenotyping using a multidimensional behavioral analysis algorithm (PUMBAA; Fig. S12; https://github.com/sidorovlab/PUMBAA). PUMBAA runs in a MATLAB environment but does not require users to have prior MATLAB coding knowledge. PUMBAA enables user control of analysis parameters for all steps including data selection, data standardization, principal component analysis, clustering, and validation.
Multidimensional analysis of behavior (Fig. 1) correctly predicted Ube3a genotype in a mouse model of AS with high accuracy (Fig. 2). This approach retained high accuracy with multiple combinations of behavioral data (Fig. 3) and in behavioral cohorts of a manageable size (n = ~12–15; Fig. 4). Principal component analysis enabled the simplification of an animal’s overall behavioral profile to a single severity score (PC1, Fig. 5) that demonstrated improvement following treatment (Fig. 6). We propose that multidimensional behavioral analysis provides a generalizable approach to quantify behavioral impairment and screen treatments preclinically in rodent models of neurodevelopmental disorders.
In the context of AS, the primary value of multidimensional behavioral analysis is to assess the efficacy of treatments in rodent models. Multiple promising treatments for AS are currently under development at various stages of clinical and preclinical testing [23, 43]. Current and future treatments will span multiple mechanisms of action, including directly targeting Ube3a expression, targeting downstream Ube3a protein targets, and more generalized symptom-based approaches. Simplifying mouse behavior to an overall severity score will be valuable for measuring overall improvement following treatment, and for assessing the effect of treatment across development. Our analysis found that each behavioral measure in the well-established Sonzogni battery  contributes roughly equally to PC1 (loadings range from 0.22–0.42; removal of single measures resulted in 92.3–96.4% accuracy), and that different combinations of behavioral inputs can be used to achieve high accuracy (Figs. 3 and 5a, b). These results suggest that multidimensional behavioral analysis can generalize well across different behavioral batteries and across different laboratories. In addition, multidimensional analysis is a tool that can generalize across mouse strains, mouse lines, and across species to be applied to the new Ube3am-/p+ rat model . Our study assessed behavior using the Ube3am-/p+ mouse developed by Jiang and colleagues , which has been the most commonly used preclinical model for AS research. However, a limit to this mouse line is that it mimics the loss of UBE3A but not other nearby genes that are also deleted in the majority of individuals with AS. In humans, AS clinical severity is typically greater in individuals with a deletion genotype . Multidimensional analysis can be used in the future to test whether AS mouse models with larger deletions  have a more severe behavioral phenotype than the Ube3am-/p+ model.
Multidimensional analysis revealed that overall behavioral severity was improved but not fully corrected by paternal Ube3a unsilencing at E15.5 + P1. Incomplete behavioral improvement is consistent with the results of Wolter et al. on certain individual tests (e.g., rotarod, brain weight), though they did report full correction of impairments in hindlimb clasping . We hypothesize that overall behavioral rescue was incomplete due to the amount of UBE3A reinstatement achieved using CRISPR/Cas9-based unsilencing: Wolter et al. achieved reinstatement of UBE3A protein to ~40% of WT levels in a subset of animals where Western blotting was performed . An advantage of multidimensional analysis is that for future treatment studies, PC1 (as a readout of overall behavioral severity) can be correlated with the degree of UBE3A reinstatement achieved within individual animals.
Beyond AS, this study provides proof of concept that multidimensional behavioral analysis can be applied to rodent models of related disorders. Prior studies have generally applied principal component analysis to rodent behavioral data in two contexts: (a) to assess which subset of behavioral measures are most relevant or valuable [14, 17, 18], and (b) to attempt to categorize two or more groups based on their behavior [15, 16, 22]. Here, we applied both approaches to quantify behavior in Ube3a mutants. Multidimensional analysis was especially effective in Ube3am-/p+ mice because behavioral impairments on individual tests are so reliable and widespread in this line . However, behavioral phenotypes in other lines are often less robust. We hypothesize that multidimensional analysis will be particularly valuable for detecting subtle behavioral differences in established mouse models and for screening behavior in new mouse models of rare disorders. Multidimensional analysis used in this context is unlikely to result in false positive effects of genotype (Fig. S7). To enable widespread use of multidimensional behavioral analysis, we developed a graphical user interface (PUMBAA; Fig. S12) to simplify and generalize analysis methods. PUMBAA enables users to perform customized multidimensional analysis in a MATLAB environment without any prior coding knowledge.
Our work identified a number of practical considerations for future studies using multidimensional analysis to quantify behavior in rodent models of NDDs. First, multidimensional analysis can only be performed in datasets where longitudinal behavioral testing is performed within animals. This requirement places certain limits on experimental design, such as the inclusion of tests that may be terminal (e.g., audiogenic seizures). Longitudinal testing also presents challenges when assessing the efficacy of treatment. For example, a four-week longitudinal behavioral battery would not be appropriate for a treatment expected to last two weeks. In addition, accounting for sex differences is an important consideration for multidimensional analysis, as sex differences in behavior have been reported in Ube3am-/p+ mice and in rodent models of related disorders (Fig. S2) [24, 45, 50,51,52,53]. Accounting for sex differences in behavior resulted in a slight but not statistically meaningful improvement in the accuracy of multidimensional analysis from ~91% to ~95% (Figs. 2 and S6). Finally, our results suggest that cross-cohort behavioral variability decreases the accuracy of multidimensional analysis performed across multiple behavioral cohorts (Fig. 4). The inclusion of behavioral assessments with no group differences would not “dilute” the effectiveness of PCA; thus, multidimensional analysis is well-suited for analysis of broad behavioral phenotyping regimens [12, 13]. Future studies using multidimensional analysis of behavior may also consider additional methods (beyond percent accuracy) to quantify the degree to which two genotypes can be distinguished in PC space. Other potentially valuable parameters include the distance between cluster centroids and the spread of data within individual clusters.
Overall, multidimensional behavioral analysis enables representation of behavior in Ube3am-/p+ mice as a single severity score that is reliably different from wild-type controls and is sensitive to treatment. Multidimensional behavioral analysis represents a tool that may be used to evaluate the effectiveness of preclinical treatments for AS and related neurodevelopmental disorders.
Lombardi LM, Baker SA, Zoghbi HY. MECP2 disorders: from the clinic to mice and back. J Clin Investig. 2015;125:2914–23.
Braat S, Kooy RF. The GABAA receptor as a therapeutic target for neurodevelopmental disorders. Neuron. 2015;86:1119–30.
Krueger DD, Bear MF. Toward fulfilling the promise of molecular medicine in fragile X syndrome. Annu Rev Med. 2011;62:411–29.
Delling JP, Boeckers TM. Comparison of SHANK3 deficiency in animal models: phenotypes, treatment strategies, and translational implications. J Neurodev Disord. 2021;13:55.
Rotaru DC, Mientjes EJ, Elgersma Y. Angelman syndrome: from mouse models to therapy. Neuroscience. 2020;445:172–89.
Winden KD, Ebrahimi-Fakhari D, Sahin M. Abnormal mTOR activation in autism. Annu Rev Neurosci. 2018;41:1–23.
Golden CE, Buxbaum JD, De Rubeis S. Disrupted circuits in mouse models of autism spectrum disorder and intellectual disability. Curr Opin Neurobiol. 2018;48:106–12.
Silverman JL, Ellegood J. Behavioral and neuroanatomical approaches in models of neurodevelopmental disorders: opportunities for translation. Curr Opin Neurol. 2018;31:126–33.
Sukoff Rizzo SJ, Crawley JN. Behavioral phenotyping assays for genetic mouse models of neurodevelopmental, neurodegenerative, and psychiatric disorders. Annu Rev Anim Biosci. 2017;5:371–89.
Silverman JL, Yang M, Lord C, Crawley JN. Behavioural phenotyping assays for mouse models of autism. Nat Rev Neurosci. 2010;11:490–502.
Gulinello M, Mitchell HA, Chang Q, Timothy O’Brien W, Zhou Z, Abel T, et al. Rigor and reproducibility in rodent behavioral research. Neurobiol Learn Mem. 2019;165:106780.
Lever J, Krzywinski M, Altman N. Principal component analysis. Nat Methods. 2017;14:641–2.
Jolliffe IT, Cadima J. Principal component analysis: a review and recent developments. Philos Trans A Math Phys Eng Sci. 2016;374:20150202.
Zimprich A, Ostereicher MA, Becker L, Dirscherl P, Ernst L, Fuchs H, et al. Analysis of locomotor behavior in the German Mouse Clinic. J Neurosci Methods. 2018;300:77–91.
Vannoni E, Voikar V, Colacicco G, Sanchez MA, Lipp HP, Wolfer DP. Spontaneous behavior in the social homecage discriminates strains, lesions and mutations in mice. J Neurosci Methods. 2014;234:26–37.
Colomb J, Winter Y. Creating detailed metadata for an R Shiny analysis of rodent behavior sequence data detected along one light-dark cycle. Front Neurosci. 2021;15:742652.
Sierra C, De Toma I, Cascio LL, Vegas E, Dierssen M. Social factors influence behavior in the novel object recognition task in a mouse model of Down syndrome. Front Behav Neurosci. 2021;15:772734.
Heinz DE, Schottle VA, Nemcova P, Binder FP, Ebert T, Domschke K, et al. Exploratory drive, fear, and anxiety are dissociable and independent components in foraging mice. Transl Psychiatry. 2021;11:318.
Muehlmann AM, Bliznyuk N, Duerr I, Lewis MH. Repetitive motor behavior: further characterization of development and temporal dynamics. Dev Psychobiol. 2015;57:201–11.
Karamihalev S, Brivio E, Flachskamm C, Stoffel R, Schmidt MV, Chen A. Social dominance mediates behavioral adaptation to chronic stress in a sex-specific manner. Elife. 2020;9:e58723.
Inayat S, Singh S, Ghasroddashti A, Qandeel, Egodage P, Whishaw IQ, et al. A Matlab-based toolbox for characterizing behavior of rodents engaged in string-pulling. Elife. 2020;9:e54540.
Popovitz J, Mysore SP, Adwanikar H. Neural markers of vulnerability to anxiety outcomes after traumatic brain injury. J Neurotrauma. 2021;38:1006–22.
Elgersma Y, Sonzogni M. UBE3A reinstatement as a disease-modifying therapy for Angelman syndrome. Dev Med Child Neurol. 2021;63:802–7.
Sonzogni M, Wallaard I, Santos SS, Kingma J, du Mee D, van Woerden GM, et al. A behavioral test battery for mouse models of Angelman syndrome: a powerful tool for testing drugs and novel Ube3a mutants. Mol Autism. 2018;9:47.
Kishino T, Lalande M, Wagstaff J. UBE3A/E6-AP mutations cause Angelman syndrome. Nat Genet. 1997;15:70–3.
Matsuura T, Sutcliffe JS, Fang P, Galjaard RJ, Jiang YH, Benton CS, et al. De novo truncating mutations in E6-AP ubiquitin-protein ligase gene (UBE3A) in Angelman syndrome. Nat Genet. 1997;15:74–7.
Buiting K, Williams C, Horsthemke B. Angelman syndrome - insights into a rare neurogenetic disorder. Nat Rev Neurol. 2016;12:584–93.
Thibert RL, Larson AM, Hsieh DT, Raby AR, Thiele EA. Neurologic manifestations of Angelman syndrome. Pediatr Neurol. 2013;48:271–9.
Bird LM. Angelman syndrome: review of clinical and molecular aspects. Appl Clin Genet. 2014;7:93–104.
Willgoss T, Cassater D, Connor S, Krishnan ML, Miller MT, Dias-Barbosa C, et al. Measuring what matters to individuals with Angelman syndrome and their families: development of a patient-centered disease concept model. Child Psychiatry Hum Dev. 2021;52:654–68.
Rougeulle C, Cardoso C, Fontes M, Colleaux L, Lalande M. An imprinted antisense RNA overlaps UBE3A and a second maternally expressed transcript. Nat Genet. 1998;19:15–6.
Yamasaki K, Joh K, Ohta T, Masuzaki H, Ishimaru T, Mukai T, et al. Neurons but not glial cells show reciprocal imprinting of sense and antisense transcripts of Ube3a. Hum Mol Genet. 2003;12:837–47.
Judson MC, Sosa-Pagan JO, Del Cid WA, Han JE, Philpot BD. Allelic specificity of Ube3a expression in the mouse brain during postnatal development. J Comp Neurol. 2014;522:1874–96.
Jiang YH, Armstrong D, Albrecht U, Atkins CM, Noebels JL, Eichele G, et al. Mutation of the Angelman ubiquitin ligase in mice causes increased cytoplasmic p53 and deficits of contextual learning and long-term potentiation. Neuron. 1998;21:799–811.
Huang HS, Allen JA, Mabb AM, King IF, Miriyala J, Taylor-Blake B, et al. Topoisomerase inhibitors unsilence the dormant allele of Ube3a in neurons. Nature. 2011;481:185–9.
Meng L, Person RE, Huang W, Zhu PJ, Costa-Mattioli M, Beaudet AL. Truncation of Ube3a-ATS unsilences paternal Ube3a and ameliorates behavioral defects in the Angelman syndrome mouse model. PLoS Genet. 2013;9:e1004039.
Meng L, Ward AJ, Chun S, Bennett CF, Beaudet AL, Rigo F. Towards a therapy for Angelman syndrome by targeting a long non-coding RNA. Nature. 2015;518:409–12.
Bailus BJ, Pyles B, McAlister MM, O’Geen H, Lockwood SH, Adams AN, et al. Protein delivery of an artificial transcription factor restores widespread Ube3a expression in an Angelman syndrome mouse brain. Mol Ther. 2016;24:548–55.
Lee HM, Clark EP, Kuijer MB, Cushman M, Pommier Y, Philpot BD. Characterization and structure-activity relationships of indenoisoquinoline-derived topoisomerase I inhibitors in unsilencing the dormant Ube3a gene associated with Angelman syndrome. Mol Autism. 2018;9:45.
Milazzo C, Mientjes EJ, Wallaard I, Rasmussen SV, Erichsen KD, Kakunuri T, et al. Antisense oligonucleotide treatment rescues UBE3A expression and multiple phenotypes of an Angelman syndrome mouse model. JCI Insight. 2021;6:e145991.
Wolter JM, Mao H, Fragola G, Simon JM, Krantz JL, Bazick HO, et al. Cas9 gene therapy for Angelman syndrome traps Ube3a-ATS long non-coding RNA. Nature. 2020;587:281–4.
Schmid RS, Deng X, Panikker P, Msackyi M, Breton C, Wilson JM. CRISPR/Cas9 directed to the Ube3a antisense transcript improves Angelman syndrome phenotype in mice. J Clin Investig. 2021;131:e142574.
Copping NA, McTighe SM, Fink KD, Silverman JL. Emerging gene and small molecule therapies for the neurodevelopmental disorder Angelman syndrome. Neurotherapeutics. 2021;18:1535–47.
Silva-Santos S, van Woerden GM, Bruinsma CF, Mientjes E, Jolfaei MA, Distel B, et al. Ube3a reinstatement identifies distinct developmental windows in a murine Angelman syndrome model. J Clin Investig. 2015;125:2069–76.
Cosgrove JA, Kelly LK, Kiffmeyer EA, Kloth AD. Sex-dependent influence of postweaning environmental enrichment in Angelman syndrome model mice. Brain Behav. 2022;12:e2468.
Judson MC, Shyng C, Simon JM, Davis CR, Punt AM, Salmon MT, et al. Dual-isoform hUBE3A gene transfer improves behavioral and seizure outcomes in Angelman syndrome model mice. JCI Insight. 2021;6:e144712.
Dodge A, Peters MM, Greene HE, Dietrick C, Botelho R, Chung D, et al. Generation of a novel rat model of Angelman syndrome with a complete Ube3a gene deletion. Autism Res. 2020;13:397–409.
Keute M, Miller MT, Krishnan ML, Sadhwani A, Chamberlain S, Thibert RL, et al. Angelman syndrome genotypes manifest varying degrees of clinical severity and developmental impairment. Mol Psychiatry. 2021;26:3625–33.
Jiang YH, Pan Y, Zhu L, Landa L, Yoo J, Spencer C, et al. Altered ultrasonic vocalization and impaired learning and memory in Angelman syndrome mouse model with a large maternal deletion from Ube3a to Gabrb3. PLoS ONE. 2010;5:e12278.
Koyavski L, Panov J, Simchi L, Rayi PR, Sharvit L, Feuermann Y, et al. Sex-dependent sensory phenotypes and related transcriptomic expression profiles are differentially affected by Angelman syndrome. Mol Neurobiol. 2019;56:5998–6016.
Huang HS, Burns AJ, Nonneman RJ, Baker LK, Riddick NV, Nikolova VD, et al. Behavioral deficits in an Angelman syndrome model: effects of genetic background and age. Behav Brain Res. 2013;243:79–90.
van Woerden GM, Harris KD, Hojjati MR, Gustin RM, Qiu S, de Avila Freire R, et al. Rescue of neurological deficits in a mouse model for Angelman syndrome by reduction of alphaCaMKII inhibitory phosphorylation. Nat Neurosci. 2007;10:280–2.
Mossa A, Manzini MC. Molecular causes of sex-specific deficits in rodent models of neurodevelopmental disorders. J Neurosci Res. 2021;99:37–56.
We thank Justin Wolter (UNC) and Juliana Popovitz (Johns Hopkins University) for valuable discussion and guidance in data preparation and analysis.
This work was supported by the Angelman Syndrome Foundation (ASF) and by the District of Columbia Intellectual and Developmental Disabilities Research Center (DC-IDDRC) Award P50HD105328 by NICHD (PI: V. Gallo). Behavioral experiments performed in the Elgersma lab were funded by the ASF, Associazione Angelman, ZonMW (40-46800-98-009), and SFARI (275234).
The authors declare no competing interests.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Tanas, J.K., Kerr, D.D., Wang, L. et al. Multidimensional analysis of behavior predicts genotype with high accuracy in a mouse model of Angelman syndrome. Transl Psychiatry 12, 426 (2022). https://doi.org/10.1038/s41398-022-02206-3