Automated, Quantitative Cognitive/Behavioral Screening of Mice: For Genetics, Pharmacology, Animal Cognition and Undergraduate Instruction

The system can and should be used to run protocols tailored to the aims of the individual investigator or classroom teacher. However, we have developed a suite of 3 protocols that should prove useful in large-scale screening of genetically manipulated mice and large-scale pharmacological testing: the matching protocol, the 2-hopper autoshaping protocol, and the switch protocol. The matching protocol measures the mouse's capacity to estimate incomes (food pellets per unit time) at two different locations, to remember which location yields which income, to match the ratio of its average visit durations to the ratio of the incomes, and the rapidity with which it detects and adjusts to changes in the income ratio. The 2-hopper autoshaping protocol measures rates of instrumental and classical conditioning (associative learning rates). The switch protocol measures interval timing accuracy and precision, the ability to estimate probabilities (relative frequencies) and to adjust to changes in probabilities (risk assessment). It also gives a measure of impulsivity.

Matching. The feeders in each of the two lateral hoppers (the feeding hoppers) are independently armed at an exponentially varying latency following each pellet delivery. A pellet is delivered whenever the feeder is armed and the mouse triggers pellet release by poking into the hopper, thereby interrupting the infrared (IR) beam. The intervals that elapse before the next arming are drawn from exponential distributions, whose expectations are specified in the parameter file (Table 1). In this protocol, mice rapidly begin to cycle between the two feeding hoppers, poking for a while into one, then moving to the other, then back to the first one. The ratio of the average durations of their stays at the two hoppers approximately matches the ratio of the pellets per min that they are obtaining from the two hoppers (N.B. Pellets per min in the environment, that is the "incomes" from hoppers, not pellets per min of poking into the hopper, that is the "returns" from their investments in the two hoppers). Thus, if they get on average 3 times as many pellets from Hopper 1 as from Hopper 2, then the average duration of a visit to Hopper 1 is approximately 3 times as long as the average duration of a visit to Hopper 2. We run this protocol first because of the rapidity with which the matching behavior appears⁸. It is usually measurable within the first 24 hr of testing, except, of course, in the occasional mouse that is too timid to visit the hoppers within the first 24 hr.

Matching behavior implies that the mechanisms for measuring duration and number and computing relative rates are intact, as are the spatial localizing mechanisms that enable a mouse to associate appropriately the two different rates with the two different hopper locations. Thus, it is an excellent quick test of basic cognitive function. The precision with which a mouse matches the ratio of its expected visit durations to the ratio of the experienced incomes is an indication of the precision with which the mouse can represent quantities.

The precision with which a mouse matches is graphically visualized by the extent to which two concurrently plotted cumulative records have the same slope (Figure 6). A cumulative record is the cumulating sum of a sequence of measurements. In this case, one plot (the red ones in Figure 6) is the cumulative record of the pellet-by-pellet income-imbalance measure; the other (the black ones in Figure 6) is the cumulative record of the pellet-by-pellet visit-imbalance measure. For a single inter-feeding interval, the income imbalance is either -1 or +1 depending on whether the pellet obtained at the conclusion of the interval came from Hopper 1 (imbalance = -1) or 2 (imbalance = +1). The slope of the cumulative record of these imbalances is the average income imbalance (the imbalance per feeding). If in 10 intervals, the mouse gets 5 pellets from Hopper 1 and 5 from Hopper 2, then the average income imbalance over that those 10 inter-feed intervals is ; if all 10 pellets came from Hopper 2, then it is ; and if all 10 came from Hopper 1, then it is . Thus, the average income imbalance can range from -1 to 1, and it is 0 when there is on average no imbalance. The visit imbalance during an inter-feeding interval is the time spent at Hopper 2 (T2) minus the time spent at Hopper 1 (T1), divided by the total of the two times. For a single inter-feeding interval, this measure can take on any value between -1 and 1, unlike the income imbalance, which must be either -1 or 1. However, the slope of the cumulative record of this measure reflects its average value, just as is the case for the income measure. When the two slopes of the two cumulative records (the red and the black records in Figure 6) are equal, the average visit imbalance equals the average income imbalance, which means that the mouse is matching the ratio of its average visit durations to the ratio of its incomes from the two hoppers.

For three of the mice in Figure 6 (5027, 5015, and 5024), the slope of the cumulative visit imbalance plot (black) closely tracks the slope of the cumulative income imbalance (red), indicating nearly perfect matching of the ratio of the average visit durations to the ratio of the incomes. Note how quickly the ratio of the visit durations adjusts to the change in the ratio of the incomes, at the points of downward inflection. On the other hand, Mouse 5034 (bottom left plot) failed to match during the second feeding session; the slope of the cumulative visit-duration imbalance plot (black) is 0 whereas the slope of the income imbalance (red) is substantially positive. However, during the third feeding phase (after the black vertical), the slopes are parallel, which means that this mouse began abruptly at the beginning of this phase to match its visit ratio to the income ratio. Clearly, therefore, it was not incapable of doing so, but did not do so for some reason during part of the testing. This illustrates the importance of visualizing the entire course of performance rather than relying entirely on a few summary statistics. Mouse 5028 matched precisely during the first income ratio, but when it was reversed, it did not fully reverse. Mouse 5025 "overmatched" during the 3^rd feeding phase (after green vertical), that is, its average visit imbalance was greater than its average income imbalance and it did not fully adjust to the reversal of the income imbalance. Notice, however, that all 6 mice, both the 3 wild types and the 3 heterozygotes matched with precision during the very first 4-hr feeding phase. Notice also that these records cover only 36-48 hr (1.5-2 day/night cycles with 2 feeding phases/cycle), during which one observes not only initial matching but an abrupt response to the reversal of the income ratio. As previously explained, this demonstrates the intact functioning in the heterozygotes of many basic mechanisms of cognition, to wit, duration estimation, number estimation, rate estimation, spatial localization (of income rates to hoppers), and the ability to remember simple abstract quantities. The quantitative results, which are typical, are obtained from automated testing in a live-in environment, with no handling of the mice beyond their initial placement in the environment.

As is evident in Figure 6, mice match the ratio of their visit durations to the ratio of the incomes as soon as they begin to cycle between the hoppers. The rapidity with which the mouse begins to cycle between the hoppers is a measure of what might be called its boldness or tendency to explore. This is visualized by plotting the cumulative record of cycles, that is, the number of cycles completed as a function of session time. Figure 7 shows these plots for the same mice as in Figure 6. Five of the six showed some exploration before the onset of the first feeding phase and began abruptly to cycle between the hoppers as soon as it began. One mouse, however (bottom left), did not show any exploratory behavior at all (not a single cycle) until midway through the 2^nd feeding phase, at which point, it began abruptly to cycle between the hoppers.

Instrumental and Classical Conditioning. Traditional learning theory distinguishes between two kinds of associative learning: classical conditioning (also called Pavlovian conditioning) and instrumental conditioning (also called operant conditioning). The physiologically meaningful parameter in both cases is the learning rate. This is thought to measure the rapidity which synaptic strengths are adjusted in response to the pairing of two sensory events in classical conditioning (e.g. hopper illumination and food delivery) or in response to the pairing of an S-R event and a reinforcing event in instrumental conditioning. The learning rate is measured by the reciprocal of trials-to-acquisition, that is, the number of pairings prior to the appearance of the conditioned behavior. Our second protocol measures these rates. It also teaches the subjects the feeding latencies to be expected at the two pellet-delivering hoppers, which knowledge is tested in the third protocol.

A pretrial in this protocol begins with the illumination of the middle hopper, the one where pellets are never delivered, because there is no pellet feeder for it (see diagram of Test Box, upper right in Figure 1). When the mouse pokes into the illuminated middle hopper, the light there goes out, and a randomly chosen one of the two flanking hoppers lights up. If it is the left hopper that lights, then a pellet is delivered there 4 sec after illumination, regardless of what the mouse does. If it is the right hopper that lights, then a pellet is delivered there 12 sec after illumination, regardless again of what the mouse does. Thus, the poking into the illuminated middle hopper is instrumental in initiating a trial; when this hopper is illuminated, nothing further happens until the mouse pokes into it; the mouse must poke into this hopper in order to get a pellet. Once the mouse has made this instrumental response, a pellet will be delivered into whichever hopper then lights up, at the delivery-latency peculiar to that hopper, regardless of whether the mouse goes to the hopper or not during the interval that elapses before pellet delivery.

The software automatically makes three plots of performance in this task: 1) the cumulative record of trial initiation speed (Figure 8), the cumulative distribution of trial-initiation latencies (Figure 9) and the cumulative records of the CS-ITI poke-rate differences (Figure 10). And, it derives from these records three sets of summary statistics: 1) trials to acquisition of the instrumental response (instrumental learning rate); 2) the median trial-initiation latency; and 3) trials to acquisition of anticipatory poking into both the short-latency and the long-latency feeding hoppers (autoshaped, that is, classical learning rates).

Timed decisions. The third protocol is the "switch" protocol first used with pigeons by Fetterman and Killeen²⁴ and adapted for the mouse by Balci et al²². As in the previous protocol, trials are initiated when the mouse pokes into the illuminated middle hopper. However, in this protocol, this trial-initiating poke illuminates both flanking hoppers, one of which is silently and randomly chosen by the computer as the "hot" hopper for that trial. The feeding latencies associated with the two hoppers are the same as in the previous protocol (4 sec and 12 sec). However, only the hot hopper delivers, and it delivers only in response to a poke made there at or after "its" feeding latency. If the short-latency hopper is hot, then if the mouse is at that hopper at the end of the short latency, the first poke made there at or after that latency delivers a pellet into that hopper. If the long latency hopper is the hot hopper, then the first poke made there at or after the long latency delivers a pellet into that hopper. Mice rapidly learn to go first to the short-latency hopper on most trials and then, on "long" trials, switch to the long-latency hopper when poking in the short-latency hopper fails to deliver a pellet at the conclusion of the short latency of 4 sec. The mouse obtains its pellets on the great majority of trials. However, if it stays too long at the short-latency hopper on a long trial, with the result that the first poke at or after the long latency is in the wrong hopper, the trial ends without a pellet delivery. Similarly, if it switches prematurely on a short-latency trial, so that the first poke at or after the short latency is in the wrong (long-latency) hopper, the trial ends without pellet delivery. The data of interest are the latencies of the switches on the long trials, defined as the termination-latency of the last poke into the short hopper prior to the first poke into the long hopper.

In this paradigm, we vary the long-latency. We start with a long latency of 12 sec, which is 3 times the 4 sec short latency, because mice learn to switch mush faster when the task is made that easy. We then shorten the long latency to 8 sec or even to 7 sec or 6 sec to test how tightly the mice can time their switching (see Figure 11). The measure of timing precision is the coefficient of variation in the Gaussian component of the distribution of switch latencies. The Gaussian coefficient of variation is the ratio between the standard deviation and the mean (σ/µ in Figure 11).

We also vary the probability of a long trial, thereby varying the relative risk of switching prematurely versus switching too late. When the probability of a long trial is high, the risk incurred by a premature departure from the short-hopper is reduced, because on a long trial, it does not matter how soon the mouse switches. Moreover, the risk incurred by a too-late departure is correspondingly elevated. Conversely, when the probability of a long trial is low, the relative risk of a premature departure versus a too-late departure shifts in the opposite direction. A rational decision maker should shift the distribution of its switches so as to approximately equate the risks, which is what mice in fact do^22,23 (Figure 12). Measuring the shifts induced by changing the probability of a long trial assesses the mouse's ability to estimate probability (of a long versus a short trial), to estimate the variability in the distribution of their switches, and to compute the relative risks implied by these two sources of uncertainty.

Impulsivity. The distribution of switch latencies is a mixture distribution. In the vast majority of mice under most conditions, the distribution is dominated by switch latencies from a Gaussian distribution, as in Figure 11. However, there is often a small component of "impulsive switches," which occur much too soon, appear not to be carefully timed, and are clearly not optimal (unlike the Gaussian component, which is approximately optimally positioned between the temporal goal posts, as may be seen in Figure 11). The frequency of these impulsive switches varies strikingly between mice and with conditions. Making the long trial highly probable increases their frequency in most mice, as does making the task more difficult by moving the long latency closer to the short latency (see Figure 11). The distribution of switch latencies can almost always be fit with impressive accuracy by a "weibgauss" function, which is the mixture of a Weibull distribution and a Gaussian distribution. The measure of impulsivity is the fraction of the switches imputed to the Weibull distribution.

Circadian Memory. Animals record in memory the time of day (phase of their circadian clock) at which motivationally important events happen¹⁷. When the same event routinely recurs at approximately the same circadian phase, their behavior begins to anticipate its recurrence. A commonly studied indication of this ability to remember and thereby anticipate a daily recurrence is feeding-anticipatory behavior²⁵. We measure this while running the above described (or almost any other) protocols, simply by restricting the intervals during which the mice can obtain pellets to two 4 hr intervals surrounding dusk (house light out) and dawn (house light on). During other intervals, the mice are still in the test environment, but no protocol is operative. Thus, they obtain pellets only during the scheduled feeding intervals. In somewhere between 6-15 days after the beginning of testing, anticipatory poking activity appears in the hour or so before the dusk feeding interval (Figure 13). The circa-dusk feeding interval begins without any signal one hour prior to extinguishing of the house light (dusk).

On those days when 3/4 of the behavioral events occur during the final 1/3 of the 9-hr phase of no-food availability preceding the onset of the dusk feeding phase, we count that as an instance of food-anticipatory activity. Anticipatory activity thus scored (blue asterisks in Figure 13) does not occur during the first several days of testing. Our measure of "trials" (days)-to-acquisition of circadian food-anticipatory activity is the day after which the frequency of anticipatory activity shows the first statistically significant increase.

Figure 1. Schematic of screening system. Each cabinet contains 8 test environments, each consisting of a polypropylene mouse tub connected by an acrylic tube to a Mouse Test box with 3 hoppers (H) monitored by IR beams, two of which hoppers have attached pellet feeders (P). A single PC (In-lab computers) running a protocol-control program that controls the environments in two such cabinets (16 environments) and logs time-stamped records of stimulus and response events. A single protocol-control program implements all the different protocols. Which protocol is operative and with which parameters is determined by the contents of a text file (Param File), which is read regularly by the protocol-control program. Changing the contents of this text file changes the operative protocol. The protocol-control program writes the event record every 10 min to another text file. The TSsystem data-analysis software, running on a remote server, harvests the data every 15 min. At regular intervals (typically every 8 or 12 hr), it performs an extensive statistical and graphical analysis of the data obtained so far and puts the results into files on a cloud-synch storage site, where they may be accessed by anyone in the lab at anytime from anywhere. The analysis software automatically moves individual mice on to the next protocol in a specified series of protocols by writing new parameters to the parameter file (Param File) read by the protocol-control program. It does so when the data from that mouse meet user-specified decision criteria. LAN = local area network. Click here to view larger image.

Figure 2A. Screen shot of the TSbroswer GUI. The browser enables browsing through the complex hierarchical data file that contains the raw data and all of the results derived from it. The functions in the TSsystem operate on data in fields of this structure and store the results in other fields within the same structure. Subordinate to the fields at all levels but the lowest are terminal data fields containing summary statistics, notes and numerically indexed lower levels: The principal numerically indexed lower levels are: the Subject Level, the Session Level, and the Trial Level (see subsequent panels in this figure). Click here to view larger image.

Figure 2B. Screen shot of the browser after selection of a subject. In the lower left window are the data fields directly subordinate to that subject. They contain summary statistics for that subject. In the Sessions window are the numerically indexed sessions for this subject and other fields that, like the Session field, have subordinate fields of their own. The MacroInfo field has subfields giving information about the protocol-control macro, including the path to and file name of the code file read by the process-control software at the start of a session. The Protocols field has subfields giving the parameters of the different protocols run in the selected session, the decision fields containing the data on which the decision to terminate a protocol is based, the code that makes that decision, and the decision criteria. Selecting a data field in the lower left window displays the data in that field in the lower right window. Click here to view larger image.

Figure 2C. Screen shot after selection of a session. The fields immediately subordinate to the selected session are displayed in the lower left window. Among those fields are those containing the raw data and subdivisions thereof and those containing summary statistics extracted from those raw data. When a data field in this window is selected, its contents are displayed in the window at lower right. In this screen shot, the field containing the raw data from the matching protocol has been selected. In the Trial Type window to the right of the session window are displayed the trial types run during the selected session. A "trial" is any user-specified sequence of not-necessarily contiguous events that define multiple stretches of data that are to be analyzed for their content. The user may define arbitrarily many different types of trials. Click here to view larger image.

Figure 2D. Screen shot of the browser after selection of a trial type. The numerically indexed trials of that type are displayed in the Trials window. As always, the data fields immediately subordinate to the selected field are displayed in the lower left window, and, if one of those fields is selected, its contents are displayed in the lower right window. Click here to view larger image.

Figure 2E. Screen shot of the browser after selection of an individual trial. The data fields subordinate to it are displayed in the lower left window and, if one of those fields is selected, its contents are displayed in the lower right window. Among the fields automatically included in every trial are fields specifying the session time at which the trial began, the session time at which it ended, its duration, and the row numbers in the raw data file that compass the trial. The other fields (here ‘Pokes,' ‘NumPks', and ‘AutoLngPkRate' are user defined. The contents of all the fields are computed from the raw data stored in a field at the session level. Click here to view larger image.

Figure 3. Screen shot of the initial prompt from TSbegin. TSbegin uses prompts to lead the user through the process of creating the hierarchical data structure into which the raw data and all the summary statistics and other results derived from those data will go. Click here to view larger image.

Figure 4. Screen shot of help for TSaddprotocol. This interactive GUI leads the user through the process of creating a new version of one experimental protocol. To do this, the user must specify new values for the parameters of that protocol (for example, new values for the concurrent variable interval schedules in the matching protocol). The user must also specify (or choose previously specified) code for making the decision to terminate the protocol and go on to the next protocol. The user must also specify (or choose previously specified) decision fields and decision criteria. Click here to view larger image.

Figure 5. Initial prompt from TSstartsession. TSstartsession is an interactive GUI that leads the user through the process of starting an experimental session. One session typically runs from a day or two to several weeks, with several different protocols (tasks) performed in the course of that one session. Click here to view larger image.

Figure 6. Cumulative imbalance plots for income (pellets obtained, in red) and visit duration (in black) for 3 wild type mice and 3 L1 heterozygotes. The relative richness of the two concurrent VI schedules was reversed at the point where the slopes turn downward (marked by vertical dashed red line). Mouse ID numbers are on each plot. The green verticals mark the onset of the dusk feeding phase; the black verticals mark the onset of the dawn feeding phase. The Er (for error) values are the differences in slope between the income plot and the visit-duration plot. These imbalance errors are twice the error in the Herrnstein fraction. The Herrnstein income fraction is the fraction of total income obtained from a hopper; the visit-duration fraction is the fraction of total time spent at a hopper (relative to combined time spent at both hoppers). A positive sign indicates "overmatching", that is, the time fraction is more extreme than the income fraction; a negative sign indicates "undermatching", that is, the time fraction is less extreme than the income fraction. For details on these experiments see Gallistel et al²⁸. Click here to view larger image.

Figure 7. Cumulative cycles (Hopper 1 visit, followed by Hopper 2 visit, followed by a return to Hopper 1) versus testing time for the same 6 mice as in Figure 6. Dashed green verticals mark the onsets of feeding phases; solid red verticals, the offsets. Click here to view larger image.

Figure 8. Cumulative records of trial-initiation speeds during the instrumental/classical conditioning protocol, same 6 mice as in Figures 6 and 7. The short heavy vertical dashed lines mark the acquisition trial in each record, defined as the first trial at which there was a significant increment in average speed to a value greater than the average speed in the first segment of the record. Click here to view larger image.

Figure 9. Cumulative distributions of trial initiation latencies, same mice as in previous figures. The distributions have been truncated at 0.7, because there are extremely long initiation latencies from trials when the mouse was not present in the test box when the trial-initiation hopper lit up, or the mouse was satiated. The point above the x-axis at which the horizontal line at 0.5 intersects the cumulative distribution is the median latency to initiate a trial. Click here to view larger image.

Figure 10. Cumulative records of the differences between the poking rate during the illumination of a feeding hopper and the poking rate during the preceding intertrial interval. Same mice as in previous figures. Initially, the slope of this record is 0 or negative, because the mouse pokes less during hopper illumination than during the intertrial interval. The slope turns positive when the mouse begins to poke into the illuminated hopper in anticipation of the impending delivery of a pellet. This anticipatory poking is classically conditioned because the pellet is delivered at the end of the hopper-specific delivery latency whether the mouse pokes into the hopper or not. The solid dots mark the point at which the slope becomes positive, which is the trial at which the conditioned response first reliably appears (trial-to-acquisition). This is an extremely noisy (variable) measure, as is evident in this sample. Click here to view larger image.

Figure 11. The slightly ragged heavy blue lines are empirical cumulative distribution functions, showing the cumulative distribution of switch latencies for Mouse #5024 at three different settings of the feed latency at the long hopper (heavy black verticals at 12, 8, and 7 sec from top plot to bottom plot, respectively). The short feed latency was always 4 sec (heavy dashed vertical). The thin red curves that superimpose almost exactly on the empirical functions are the 6-parameter Weibull-Gaussian functions fit to these data. The heavy red distributions are the corresponding probability density functions for the Weibull-Gaussian mixture distributions. (They are the derivatives of the cumulative distribution functions.) The parameters of the Gaussian component of these mixture distributions are shown on the plots. Shortening the long feed latency caused the mouse to shorten the mean of its switch latencies (µ) and increase their precision (σ/µ). It also caused the appearance of some impulsive switches (the bump in the left tail of the probability density function in the bottom plot). The measure of this impulsivity is the fraction of the mixture attributed to the Weibull component by the best-fitting version of the Weibull-Gaussian mixture distribution. Click here to view larger image.

Figure 12. Switch latencies (small open circles) shift abruptly soon after changes in the relative frequency of short and long trials. The short trials are marked with tiny red pluses at the bottom of the plot, while the long trials are marked with tiny red pluses at the top of the plot. The change in the density of the resulting red streaks at bottom and top indicate the change in the relative frequency (probability) of the short and long trials. Note that when the red at bottom becomes more dense (when the short trials become more probable), the small circles (switch latencies) shift upward and when the red at the top becomes denser (an increase in long-trial probability), the circles shift downward. The blue lines are the medians of the distributions. Excerpted from Figure 2 in Kheifets and Gallistel²³. Click here to view larger image.

Figure 13. 24 hr raster plot of behavioral and environmental events over a 90 day test period. Black dots record pokes; red and green dots record pellet deliveries at the two feeding hoppers. Black right-pointing triangles record house-light offs; black left pointing triangles record house-light ons; cyan left-pointing triangles record the (unsignaled) onsets of the dusk feeding interval, which occur 1 hr prior to dusk itself; cyan right-pointing triangles record the offsets of the dusk feeding interval (3 hr after house-light off); left-pointing magenta triangles record the (unsignaled) onsets of the dawn feeding intervals (2 hr before house-light on); right-pointing magenta triangles record the offsets of the dawn feedings. The 24 hr time axis is conventional midnight to our midnight, not the reverse-cycle time of the test environment in which the mouse lived. The onsets of feeding-anticipatory activity prior to the dusk feeding are marked with blue asterisks. Click here to view larger image.