Simultaneous behavioral characterizations: Embracing complexity

The advent of transgenic technologies in mice has dramatically impacted the genetic analysis of behavior. The first knockout mice evaluated for behavioral responses were reported in 1992 (1, 2), and since these initial reports a significant number of scientists have performed behavioral studies with mutant mice each year. In fact, a literature search by using simply the words “knockout” and “behavior” and “mice” results in almost 4,000 hits. The powerful molecular technique has not simply increased the amount of behavioral genetic studies performed across the globe, but has fundamentally altered the strategies and methodologies of standard approaches for assessing rodent behavior. Not only have standard experimental designs been altered to successfully work with transgenic mice, but scientists have also developed technologies to simultaneously and automatically quantify fine behavioral patterns in home cage settings, as illustrated by a study in this issue of PNAS (3). This study clearly demonstrates the capacity to capture and characterize complex behavior in mice coupled with sophisticated learning algorithms to provide a type of detailed analysis that has not been previously available.

Automated Behavioral Analysis

Over the past several years a number of related technologies for automatically capturing rodent behavior over long time periods have become available (e.g., refs. 4–8), and recent studies have started to show the utility of such systems with transgenic mice (e.g., ref. 9). The findings by Tecott and colleagues (3) have pushed the envelope even further by pulling together resources from vastly different scientific disciplines such as animal behavior, molecular biology, genetics, engineering, and mathematics to create a unique automated system that captures and quantifies levels of behavioral organization and behavioral structure in mice.

Behavior is the final functional output that reflects the integration and orchestration of internal biological pathways and events, and as such, behavior can often be complicated and very difficult to interpret. Traditionally, studying behavior in rodents has resorted to applying a more reductionistic approach; by using a behavioral setting that simplifies the behavioral response in an attempt to capture a single, or limited domain of CNS function. If researchers needed to evaluate additional multiple domains of CNS function, they would examine the behavior of an animal on another distinct test. Although this strategy is important and has successfully resulted in a number of “test batteries” commonly used for mutant mice (e.g., ref. 10) it is limited by the fact that an isolated test (i) typically only captures behavior for a relatively limited period, (ii) is often evaluated by using a novel setting that will itself influence behavior and make interpretations more challenging, and (iii) does not readily allow a researcher to simultaneously integrate responses (e.g., ref. 11) across behavioral tests. Whereas a reductionistic approach is a critical and necessary scientific methodology, it is important that we continue to strive to develop strategies that capture the complexity of an entire system in a modifiable environment, in this case a freely moving mouse in its home cage, so that we devise ways to systematically understand and integrate the complexity that is the whole organism in a less confined and more natural/normal state.

This study clearly demonstrates the capacity to capture and characterize complex behavior in mice.

The report by Goulding et al. (3) clearly demonstrates that it is possible to simultaneously evaluate complex behavioral structures and elements with powerful temporal and spatial resolution in a fully integrated home cage system. Use of a home cage setting to study behavior has a number of advantages including the reduction in “stress- and anxiety-related” responses often associated with moving a mouse from its home cage to test apparatus that is often unfamiliar. Another major advantage of the home cage environment is the ability to measure various behaviors that are critical for a normal physiological function such as eating and sleeping, and how these integrate with cage exploration. Last, the ability to integrate multiple behavioral responses across daily 24-hr cycles allows for the evaluation of responses in the context of, and influence by, normal circadian biology. It is important to note that each of these independent features (i.e., cage activity, feeding, sleeping, drinking, and circadian activity) can be captured by using more traditional individual assays, but it is virtually impossible to study these behavioral processes together to obtain a unified assessment of the integration of these particular domains of CNS function, something the Tecott system can readily handle. More importantly, the learning algorithms developed by Tecott and his colleagues are capable of handling and analyzing in a coherent manner the massive amount of data acquired during an experiment designed to simultaneously capture and measure these types of behavioral variables with spatial resolution and across time.

To put the advantage of an automated behavioral system with an integrated learning algorithm into perspective—if my laboratory were to try and reproduce the same experiments reported by Goulding et al. (3) we would likely require at least three distinct pieces of major equipment (i.e., a device to measure eating, a device to measure drinking, and a device to measure exploratory activity). In addition, a system for high-resolution video recording would be needed, and personnel to “hand-score” the active and inactive states necessary to obtain some of the resolution provided by the Tecott system. Scoring the videotapes would likely require at least 4 full-time research staff per cage, per day of the experiment. Although we might be able to capture many of the same behaviors, we still would not be able to readily perform the sophisticated analysis on the massive amount of data because we would need to develop the learning algorithm necessary to integrate all of the information into an interpretable and manageable output. Thus, it would be expensive and very labor intensive, and possibly still would not provide the same refined data that the Tecott system provides. Therefore, the availability of this type of automated behavioral recording system to the research community should provide scientists with a new tool to address experimental questions that would have been difficult by using current technologies.

The foundation and critical aspect for the success by Goulding et al. (3) is the ability to discriminate between active and inactive states while simultaneously examining structures of within-state bouts of behavior. The ability to dissociate active and inactive states provided a means to ultimately develop supervised learning algorithms to detect, measure, and quantify across rapid timescales behaviors such as eating and drinking, and how these responses “cluster” into isolated bouts. The authors show that studying the integrated structure of behavioral responses in different states and across multiple bouts provides a sensitive means to identify previously unidentified phenotypes in 2 genetic mouse models.

Critical Validation

To validate the utility of the system Tecott and colleagues (3) evaluated 2 distinct single-gene mutant mouse models with different alterations in energy balance regulation. The ob/ob (OB) mice have a number of phenotypes including significant obesity and reduced activity. In contrast, the htr2c (5-HT2C) null mutant mice have increased food intake, obesity, and hyperactivity. Not only did the algorithms developed by Tecott and colleagues identify those phenotypes that were already known to be present in these 2 mouse models, they also revealed abnormalities that had not been previously reported. For example, the time budget analysis of the OB mice, which would not have been easily obtained without an automated system like that developed by Tecott and colleagues, found that this mutation leads to increased inactive states with reduced time spent in locomotion, but the a significant portion of the active state was dedicated to feeding and drinking. In contrast, the 5-HT2C mutant mice had decreased inactive states and increased food intake. More importantly, the researchers were able to dissociate the nature of intake between OB and 5-HT2C mice by showing that OB mice had similar patterns of food intake compared with wild-type mice (i.e., peaks of intake at dusk and dawn), but that OB mice increased the amount of time spent consuming during each bout of intake. In contrast the 5-HT2C mice displayed a disproportional increase in their intake specifically in the light cycle; but relative to controls the amount of their intake was not increased per bout.

Importantly, to verify the analyses of their automated system, the researchers scored a number of responses by hand and performed correlation analysis to demonstrate that the automated system consistently agreed with the observations of an experimenter. This is a critical and key feature of the report to ensure that the automated system of capturing and scoring behavior is actually recording what behaviors mice are displaying—the trained “experimental eye” to detect and interpret behavior can never be fully replaced by automation.

The current work by Tecott and colleagues clearly demonstrates that they have developed a new automatic system that provides an opportunity to study a number of behavioral characteristics and how they integrate with the demands of whole-animal physiology. The fine-grain analysis using supervised learning algorithms should provide researchers with a detailed analysis that had not been readily available and an additional insight into the global organization of behavior in a home cage setting. In fact, the real power of this important new technological development is likely to be revealed when it is used to study how additional experimental manipulations on the animal impacts home cage behavior. For example, it will be interesting to determine how stress, acute and/or chronic, will alter the time allocation mice spend in various behavioral states and how this interacts with particular mutations. In addition, it will be interesting to determine how mice respond when they are on food restriction; do mice budget their time differently when food is available for limited, or intermittent periods of time? With the development of automated systems such as that reported in this issue of PNAS by Goulding et al. (3), scientists will likely be able to address questions in ways previously not fathomed because of the availability of learning algorithms that can adapt to changing experimental demands and manage large amounts of data that would have been daunting by using more traditional approaches.

It is important to point out that, although this type of analysis is unique and powerful, it will not replace more traditional behavioral test settings. The current application is for home cage activity that includes feeding, drinking, “sleeping,” and exploration with a single-housed mouse. There are a wide range of domains of CNS function that are currently not readily evaluated with this sort of automated setting, such as general motor coordination and skill learning, sensorimotor capacities, higher-order cognitive abilities, social interactions, and “psychiatric-related” responses such as anxiety-based behaviors. Future developments with automated home cage systems such as that reported by Goulding et al. (3) may ultimately provide technologies that could integrate with the more traditional “single-test setup” to capture additional features of mouse behavior in a dynamic manner that will lead to increased understanding of the nature of behavioral differences in mice.