Male adult C57BL/6 mice (SLAC, as wild-type) were used for the current study (8–40 weeks of age, weighted between 20 g and 30 g). Wild-type mice were provided by the Shanghai Laboratory Animal Center (SLAC), Chinese Academy of Sciences (CAS), Shanghai, China. Mice were group-housed (4–6/cage) under a 12-h light-dark cycle (light on from 5 a.m. to 5 p.m.). Before behavioral training, mice were housed in stable conditions with food and water ad libitum. After the start of behavioral training, the water supply was restricted. Mice could drink water only during and immediately after training. Care was taken to keep mice body weight above 80% of a normal level. The behavioral results reported here were collected from a total of 25 wild-type mice. All animal studies and experimental procedures were approved by the Animal Care and Use Committee of the Institute of Neuroscience, Chinese Academy of Sciences, Shanghai, China.

Mice were anesthetized with analgesics (Sodium pentobarbital, 10 mg/mL, 80 mg/kg body weight) before surgery. All surgery tools, materials and experimenter-coats were sterilized by autoclaving. Surgery area and materials that cannot undergo autoclaving were sterilized by ultraviolet radiation for more than 20 min. Aseptic procedures were applied during surgery. Anesthetized mice were kept on a heat mat to maintain normal body temperature. Scalp, periosteum and other associated soft tissue over skull were removed. Skull was cleaned by filtered artificial cerebrospinal fluid (ACSF) with cotton applicators. After skull was dried out, a layer of tissue adhesive was applied on the surface of the skull. A steel plate was placed on the skull and then fixed by dental Cement.

HATS was composed of a mouse containing, head-fix, odor delivery and reward delivery, Arduino based control and data acquisition units (diagram in Figure 1A, photo in Figure 1B). All valves and motors were controlled by Arduino based processors and customized software. The 3d printing files, a step by step instruction for hardware assembling, the source code for behavior training and the data acquisition source code were publically available^1,2.

Three-dimensional printing technique was used to generate the small components in the system (Figure 1C). The training tube was used to maintain the relative position of mouse body to the water- and odor-delivery ports. The motor slot held a direct-current motor to move the water port forward or backward. The water tube slot held a metal needle with a blunt tip, from which mice obtained water as a reward. The odor-tube slot connected the odor tube from the odor-delivery unit.

A movable water port was connected to a peristaltic pump, which was controlled by an Arduino board. The volume of water reward was controlled by changing the duration of the output signal to peristaltic pump from the Arduino board. Peristaltic pumps of different setups were calibrated for the stable volume of water delivery in each trial (5 ± 0.5 μL).

Water- and odor-delivery units were both controlled by an Arduino board. During behavior training, detailed timing information of events was sent back to the computer via the USB-simulated serial-port interface and stored by a customized Java program. The stored events included an odorant valve on/off, peristaltic pump on/off, and licking start/end. Licking event was detected by a capacity detector. Infrared LED-based licking detectors were used for electrophysiological recording if required. An infrared camera was placed under the water port to monitor behavioral states of mice.

The olfactometer was designed to efficiently and reliably mix and deliver odor. Air source was a pump that provided air flow with the flow rate of ~120 L/min. The filter was applied to eliminate moisture and dust. Eight training setups shared one set of pump and filter. For each setup, pure air with the flow rate of 2 L/min is constantly delivered to mice during the entire process. The air input to each air route could be turned on and off by a manual valve (labeled as “M” in Figures 2A,B). The flow rate was adjusted by a needle valve (labeled as “V” in Figures 2A,B). As shown in the Figure 2B, one type of odorant in liquid state was stored in one airtight bottle. The air-in tube was placed right above the surface of the liquid odorant. Two-way solenoid valves were used to switch the odor to either mouse or flow mater. In the standby state (no odor was delivered, Figure 2A), the valve to odorant bottle (labeled as “O”) was closed, and that to the flow meter (labeled as “F”) was opened. Therefore, no odor will be mixed with pure air and delivered to the mouse. In the working state that odor was delivered (Figure 2B), “O” was open and “F” was closed. Therefore odor was mixed with constant air and delivered to the mouse. Four kinds of odorants were used in the behavior tasks, 1-Butanol, Methyl butyrate, Hexanoic acid and Octane. The relative volume ratios of these odorants in the pure air were 10%, 2.5%, 15% and 5%, respectively. The difference was due to the distinct evaporation pressure of different odorant molecules at room temperature (see Table 1 for detailed rising/decay and residual time of the odorants). The odor tubes after “O” valves and before mixture chamber had an inner diameter of 0.5 mm. The odor tube for constant air before mixture chamber had an inner diameter of 2.5 mm.

The performance of the correct rate (referred to as “performance” in labels of figures) of each bin was defined by:

Performance correct rate = (num. hit trials + num. correct rejection trials)/total number of trials

Hit, False choice and CR rates were defined as follows:

Hit rate = num. hit trials/(num. hit trials + num. miss trials)

False choice rate = num. false choice trials/(num. false choice trials + num. correct rejection trials)

Correct rejection rate = num. correct rejection trials/(num. false choice trials + num. correct rejection trials)

Mean correct rate (CR rate/FA rate) was calculated as an averaged correct rate (CR rate/FA rate) between different mice.

Error bars from the mean value of the correct rate (CR rate/FA rate) was calculated by the standard error of the mean. N represents the number of mice.

The licking rate was calculated as lick numbers within each time bin (bin size:100 ms). The curve was smoothed by smooth function from Matlab with a span size of five bin.

Discriminability (d′) was defined by:

d′ = norminv (Hit rate) − norminv (False choice rate). The norminv function was the inverse of the cumulative normal function. Conversion of Hit or False choice rate was applied to avoid plus or minus infinity (Macmillan and Creelman, 2005). In conversion, if Hit or False choice rate was equal to 100%, it was set to [1 − 1/(2n)]. Here, n equals to a number of all possible Hit or False choice trials. If Hit or False choice rate was zero, it was set to 1/(2n).

Licking efficiency = rewarded licking number/(rewarded licking number + unrewarded licking number).

A number of trials to criterion was calculated as the trial numbers before reaching 80% correct rate for 24 consecutive trials. “NRC” in Figures 6–8 represented Not Reaching Criterion, which indicated that mice did not reach the above criterion for that day.

In our previous study (Liu et al., 2014), mice were manually taught to lick for water and shaped for a DNMS task. The goal of the current study was to allow fully automatic training. The only things human operators need to perform were to fixate mice onto head-fix bars, close doors of training boxes, and run computer software controlling training protocols. The current study fulfilled the goal by designing HATS for olfactory and odor-based cognitive behavior in head-fixed mice. HATS was composed of a mouse containing, head-fix, odor- and water- delivery, Arduino based control, and data acquisition units (diagram in Figure 1A, photo in Figure 1B, 3D-printed parts in Figure 1C). Optogenetic, chemogenetic, recording, and imaging methods can be easily integrated into HATS. All valves and motors were controlled by Arduino based processors and customized software. The daily routine was composed of system adjustment, head-fixation of mice, choosing a protocol, and training mice a given behavior (Figure 1D).

In studying olfactory behaviors, it is critical to have fast rise and decay for odor delivery.

Our olfactometer (Figures 2A–D) exhibited fast response and stable performance. The reaction time constant for the onset of these odors was between 11 ms and 71 ms (Table 1), measured with a photoionization detector (PID). Another key parameter was the time constant for decay after the offset of the odor-delivery unit, which was especially important in working memory-related tasks. The current odor-delivery unit exhibited fast decay (time constant: 20–41 ms, Figure 2E, Table 1). Moreover, odor concentration remained stable following more than 200 trials of odor delivery (Figure 2F), which was important for behavioral and recording experiments.

To achieve fully automatic training, we developed a step-by-step training protocol. The protocol was separated into two preparatory steps (water deprivation and habituation) and three training phases (lick-teaching, shaping and learning, Figure 3).

The first step of training was to automatically teach licking freely from water tube (Figures 4A,B). Moveable water port (Figures 4C–E) was located 5 mm away from the mouth of a mouse. The flow chart of the lick-teaching protocol was plotted in Figure 4A. At the start of a teaching bout, water port would deliver 10 μL water and then moved forward until contacting the mouth, thus encouraging the licking. If mouse licked, 4 μL water would be rewarded for every three licks. After no licking was detected for consecutive 2 s or water of 200 μL was delivered, one bout of teaching was completed, and the water port was moved back to the initial position. The teaching bout was repeated for several times until water of 400 μL was rewarded in total. The volume of water rewarded in each day was plotted in Figure 4B.

The second step of training was shaping for a specific task. This phase was designed to allow mice to be familiar with the temporal structure of the tasks and the involved sensory stimuli, without experiencing the full task. In shaping, only the trials with water reward were applied. Specifically, for DNMS task, only non-matched odor pairs were applied to mice (Figures 5A,B). For DPA task, only paired trials were applied. For GNG task, only Go cue was applied. Two types of trials were designed, self-learning and teaching trials. In self-learning trials, water delivery was triggered by licking in the response window (Figure 5C, left box). In teaching trials, water port moved forward and delivered water automatically during response window (Figure 5C, right box). These two types of trials were designed to switch automatically. The condition for switching from learning to teaching trial was that mice missed five trials in 35 trials. The condition for switching from teaching to learning trial was that mice licked within the response window in the last teaching trial. Daily shaping phase ended when mice performed 100 hit trials in total. This phase lasted for 3 days.

We trained eight head-fixed mice to perform an olfactory DNMS task (Liu et al., 2014; Figure 6A). In this design mouse needed to temporally maintain information during the delay period before behavioral choices and motor planning. After the shaping protocol, we added the non-rewarded matched trials, which induced false choice and reduced performance to chance level (Figure 6B). Gradually the performance, CR and discriminability (d′) progressively increased, whereas the hit rate remained at a ceiling level (Figures 6B–E). After the training of 5 days (600 trials), the performance showed significant increase (ANOVA, p < 0.0001, F = 775.89). Mice experienced a certain level of relearning each day, with a decreased number to criteria (defined as a correct rate above 80% in 24 consecutive trials) each day through learning (Figure 6F). Most of the licking responses were associated with non-match odor and expectation of water reward (Figure 6G). There were licks associated with the first odor delivery in the early phase of learning (Figure 6G, black curve), which were declined through learning (Figure 6G, blue curve). Also, the licking efficiency (defined as the ratio of successful licks resulting water reward) was increased progressively through learning (Figure 6H).

The second set of head-fixed mice was trained to perform an olfactory DPA task (Figure 7A). As in the DNMS task, the performance, CR and discriminability (d′) progressively increased, whereas the hit rate remained at ceiling level (Figures 7B–E). After the training of 5 days (600 trials), the performance showed significant increase (ANOVA, p < 0.0001, F = 1139.03). Mice also experienced a certain level of relearning each day (Figure 7F). Most of the licking responses were associated with paired odor and expectation of water reward (Figure 7G). There were licks associated with the first odor delivery in the early phase of learning (Figure 7G, black curve), which were declined through learning (Figure 7G, blue curve). Such early licks associated with the sample odor were lower than that in the DNMS task. The licking efficiency also increased progressively through learning (Figure 7H).

The third set of head-fixed mice was initially trained to perform an olfactory GNG task (Figure 8A, above), then subsequently sensory-cue reversal task (Figure 8A, below). The performance, CR and discriminability (d′) progressively increased, whereas the hit rate remained at ceiling level (Figures 8B–E). After the training of 2 days (200 trials), the performance showed significant increase (ANOVA, p < 0.0001, F = 3455.17). Mice also experienced a certain level of relearning each day (Figure 8F). Most of the licking responses were associated with paired odor-pair and expectation of water reward (Figure 8G). The licking efficiency also increased progressively through learning (Figure 8H). After 2-days of GNG training, the odor-reward relationship was reversed (Figure 8A, below). The performance, CR, discriminability (d′) and licking efficiency were decreased initially, and then progressively increased (Figures 8B–E). The hit rate remained at ceiling level (Figure 8D) and relearning was evident from the number of trials to criteria (Figure 8F).