We first asked how IL PV neurons respond during active avoidance behavior. To target this population for fiber photometry, we selectively expressed a genetically encoded calcium indicator in IL PV neurons by injecting AAV-CAG-Flex-GCaMP6f into IL in PV-Cre mice (Hippenmeyer et al., 2005; Chen et al., 2013). In control mice, we expressed GFP in IL PV neurons by injecting AAV-CAG-Flex-GFP. We implanted an optical fiber over IL to monitor calcium-dependent fluorescence, and recorded IL PV population activity via fiber photometry (Figure 1A and B; Cui et al., 2013; Gunaydin et al., 2014).

We used a two-way signaled active avoidance paradigm (Figure 1C). When an auditory cue (constant tone at 12 kHz or 8 kHz) was played, mice were required to cross the midline of the behavioral testing chamber within 5 s of tone onset to avoid an impending foot shock. If mice crossed the chamber during this 5 s period, no foot shock was delivered and the trial was scored as a successful avoidance. If not, a 2 s foot shock was delivered. Mice could terminate the foot shock early by crossing the chamber, which was scored as an escape. The tone terminated at either successful avoidance or shock offset. Prior to avoidance training, mice received two tone-shock pairings to learn the association between the auditory cue and the foot shock.

As IL inactivation impairs avoidance (Moscarello and LeDoux, 2013; Halladay and Blair, 2017), we predicted that the activity of inhibitory IL PV neurons would be low during successful avoidance trials. Contrary to expectations, we found that IL PV neural activity rose upon the initiation of avoidance movements and peaked at chamber crossing (Figure 1D–G and G: N=8 mice, p=0.0077, paired t-test; Figure 1—figure supplement 1A–F,B: N=8 mice, p=0.0013, paired t-test, F: N=8 mice, p=0.4682, paired t-test). The trial-by-trial variability in the amplitude of IL PV neural activity was not correlated with variability in the speed or latency of the avoidance movement (Figure 1—figure supplement 1G–L, H: N=8 mice, p=0.3541, one-sample t-test, K: N=8 mice, p=0.4132, one-sample t-test). IL PV neural activity during the avoidance movement was not attenuated by learning or repeated reinforcement (Figure 1—figure supplement 1M and N, N=8 mice, p=0.8886, one-way ANOVA).

During successful avoidance trials, two events happen simultaneously: the mouse crosses the chamber, and a shock-predicting tone is terminated. To determine whether PV neural activity better reflects the avoidance movement or the termination of the predictive auditory sensory cue, we designed a version of the avoidance task with additional trial types to uncouple these events. Regular trials (80%) were interleaved with trials with shortened (10%) or lengthened (10%) tones (Figure 2A). In short-tone trials the tone lasted for only 1.5 s, and in long-tone trials the tone was not terminated until 1.5 s after successful avoidance (Figure 2A). IL PV neural activity at chamber crossing did not differ between regular and long-tone trials (Figure 2B and D, N=7 mice, p=0.9546, paired t-test). Usually, the chamber was not crossed in short-tone trials, so short-tone trials were not included in this analysis. PV neural activity at tone offset was significantly different among short-tone, regular, and long-tone trials (Figure 2C and E, N=7 mice, p=0.005, one-way repeated ANOVA. Multiple comparison, R and S, p=0.0045, R and L, p=0.0398, S and L, p=0.4620.). These results indicate that PV neural activity primarily reflects the avoidance movement and not tone offset.

We extracted characteristic neural responses to each event without interference from other events close in time. We constructed a linear regression model to extract the isolated neural responses to chamber crossing and tone termination (Parker et al., 2016; Musall et al., 2019), and all major events were included in the model. We assumed (1) that neural responses would be similar for the same events and dissimilar for different events, and (2) neural responses to events can be summed up linearly to form the recorded signals. With these assumptions and through linear regression, isolated neural responses to each event were extracted. The model-extracted isolated neural response to tone offset was flat, while the isolated avoidance signal peaked at chamber crossing (Figure 2F and G). To further demonstrate that IL PV activity can be better accounted for by the action to avoid the shock than tone termination, we compared the explanatory power (R 2 , coefficient of determination) of the reduced model with shuffled time points to the full model. Shuffling the avoidance time points significantly reduced the explanatory power of the model (ΔR 2 different from zero, N=7 mice, p=0.0125, one-sample t-test), while shuffling the tone offset time points had no effect (Figure 2H, N=7 mice, p=0.0802, one-sample t-test). Thus, IL PV neural activity at chamber crossing reflects the avoidance behavior and not the cessation of the predictive auditory cue.

IL PV neural activity rises during movement to avoid a predicted future shock (Figure 1E and F), but it is unclear whether this neural activity specifically reflects avoidance, or if IL PV neural activity would be elevated during any movement. If IL PV neural activity reflects locomotor activity, we would expect to see elevated neural activity during movement to obtain rewards. To test this idea, we recorded IL PV neuronal activity during a reward approach task, which mirrored the avoidance design in temporal structure, but shock omission on successful chamber crossing was replaced with reward delivery (Figure 3A). In this task, chamber-crossing movements to obtain a water reward were not associated with elevated IL PV neural activity (Figure 3B-E; E, N=7 mice, p=0.3003, paired t-test; Figure 3—figure supplement 1A-C, C: N=6 mice, p=0.0382, paired t-test; see Figure 1C-G for comparison).

We observed an elevation of IL PV neuronal activity after chamber crossing when the animals started to consume the water reward (Figure 3C, magenta dots, and Figure 3D). To test whether this transient increase in IL PV neuron activity after chamber crossing reflects water consumption or the preceding approach behavior, we introduced randomized water-omission trials in 10% of the trials (Figure 3—figure supplement 1D–E). We found that the PV neuron signal rose both when water was received and when the water reward was omitted, and increases in IL PV activity preceded the first lick to consume water reward delivered after a successful crossing (Figure 3—figure supplement 1E). We speculate that, rather than a reward signal related to water consumption, this signal may reflect the suppression of the habitual chamber-crossing movement in the approach task when the mouse nears the lick port.

To further investigate whether elevated IL PV neuronal activity reflected avoidance or locomotion, we recorded IL PV activity in an open field test (OFT) where animals were allowed to run freely without any defined task structure (Figure 3—figure supplement 2A and B). When we aligned IL PV activity to movement peaks, no elevated activity was observed (Figure 3—figure supplement 2C). We varied the overhead lighting and found that IL PV neurons showed no elevated activity during movement under either a bright ceiling light, which is a more aversive setting to the mice, or a dim ceiling light, which is more comforting to the mice (Figure 3—figure supplement 2D and E, N=7 mice, bright light: p=0.8705, dim light: p=0.1475, paired t-test; Figure 3—figure supplement 2F). It should be noted that the OFT is a relatively neutral context, regardless of lighting conditions, when compared to the appetitive reward approach or aversive active avoidance.

In environments with different emotional valence animals engage in different suites of behaviors. For example, animals in threatening environments spend more time freezing and less time exploring than animals in rewarding environments. We hypothesized that IL neural activity may become positively correlated with movement when animals learn that their environment is threatening, and must suppress behaviors such as freezing in order to engage in behaviors such as exploration.

To test this hypothesis, we analyzed the recordings made on the first day of active avoidance training to determine how movement-related IL PV neural activity evolves as animals learn that the environment is threatening. On the first session, we habituated animals to the chamber for 5 min before the task started. During this habituation period, IL PV activity was not positively correlated with movement (Figure 4A; Figure 4—figure supplement 1A), similar to our observations in appetitive and neutral contexts (Figure 3; Figure 3—figure supplement 1E–I). During avoidance behavior, we found that IL PV activity peaked during movements in the inter-trial interval (Figure 4B; Figure 4—figure supplement 1A). Considering that animals minimize their movements in threatening environments, movement outside of the tone-evoked trial could require suppression of this natural inclination.

The correlation between movement and IL PV neural activity during the inter-trial interval did not emerge until after animals experienced the first foot shock (Figure 4C; Figure 4—figure supplement 1A). The correlation between movement and IL PV neuronal activity emerged immediately after receiving the first shock and did not increase during the learning process, suggesting that avoidance learning is not necessary for the correlation between movement and IL PV neural activity to emerge. The lack of positive correlation between activity and movement in the habituation phase (Figure 4A–C; Figure 4—figure supplement 1A) of the avoidance task is similar to what was observed throughout the approach task (Figure 4D; Figure 4—figure supplement 1B). These data show that the positive correlation between movement and PV neuron activity emerges immediately after animals learn that their current environment is threatening, when movement requires a suppression of freezing.

To investigate the causal role of IL PV neural activity in active avoidance, we bilaterally inhibited IL PV neurons by expressing Cre-dependent halorhodopsin (AAV5-EF1α-DIO-eNpHR; control animals: AAV5-EF1α-DIO-eYFP) in PV-Cre mice (Figure 5A and B). We used the same avoidance and approach tasks described above (Figure 1A; Figure 3A), and introduced interleaved simulation blocks where IL PV neurons were optogenetically inhibited during the interval from 0.5 to 2.5 s after tone onset (Figure 5C). Inhibiting IL PV neuronal activity delayed the avoidance movement (Figure 5D–G; Figure 5—figure supplement 1A and B; Video 1) but did not delay the movement for water reward in the approach task (Figure 5H–K; Figure 5—figure supplement 1C and D) or the speed of voluntary locomotion in the OFT (Figure 5L–O, dim OFT: interaction between opsins and stimulation: N=8 NpHR mice, N=5 eYFP mice, p=0.7636, two-way ANOVA; bright OFT: interaction between opsins and stimulation: N=8 NpHR mice, N=5 eYFP mice, p=0.7469, two-way ANOVA).

Both the speed of the avoidance movement (Figure 5D and E, significant interaction between opsin and stimulation, N=8 NpHR mice, N=5 eYFP mice, p=0.0281, two-way ANOVA) and the probability of successful avoidance (Figure 5F and G, N=8 NpHR mice, N=5 eYFP mice, p=0.0058, unpaired t-test) were reduced by suppression of IL PV neurons. The speed of movement and the probability of successful reward delivery were not affected in the reward approach task (speed: Figure 5H and I, interaction between opsin and stimulation, N=8 NpHR mice, N=5 eYFP mice, p=0.479, two-way ANOVA; probability of approach: Figure 5J and K, N=8 NpHR mice, N=5 eYFP mice, p=0.1204, unpaired t-test). This suggests that IL PV neuronal activity is not just correlated with avoidance behavior but plays a causal role. By further analyzing the video data, we found that freezing during tone presentation increased (Figure 5—figure supplement 1E–G, G: N=8 NpHR mice, N=5 eYFP mice, p=0.0175, unpaired t-test; Video 1). In addition, the ratio of freezing duration to avoidance latency increased during IL PV inhibition (Figure 5—figure supplement 1H, N=8 NpHR mice, N=5 eYFP mice, p=0.0302, unpaired t-test; Video 1), suggesting that increased avoidance latency was primarily due to an increase in freezing. This finding supports the idea that IL PV neural activity promotes avoidance by suppressing freezing (Murphy et al., 2005; Hardung et al., 2017).

We then asked whether inhibiting IL PV neuronal activity had an immediate effect on avoidance during the current trial or if instead inhibition influenced learning. To address this question, we investigated whether IL PV inhibition affected avoidance in trials following inhibition, using the task described in Figure 5C. Contrary to the learning hypothesis, we found that suppressing IL PV neurons delayed avoidance even on the first trial of a stimulation block. The latency of the first trial of inhibition blocks was significantly longer than the latency of pre-stimulation trials (Figure 5—figure supplement 1I and J, N=8 NpHR mice, N=5 eYFP mice, p=0.0022, paired t test), but was not significantly different from the latency of subsequent inhibition trials (Figure 5—figure supplement 1I and J, N=8 NpHR mice, N=5 eYFP mice, p=0.0943, paired t test). The latency of the first non-inhibited trial following an inhibition block is significantly different from the latency of the preceding inhibition trials (Figure 5—figure supplement 1K anf L, N=8 NpHR mice, N=5 eYFP mice, p=0.0343, paired t test) but is not significantly different from the latency of subsequent non-inhibited trials (Figure 5—figure supplement 1K and L, N=8 NpHR mice, N=5 eYFP mice, p=0.8147, paired t-test). These results suggest that IL PV neuronal activity plays a causal role in modulating ongoing behavior.

To further tease apart the role of IL PV neural activity in learning, we inhibited IL PV neurons immediately after successful chamber crossing in the avoidance task, rather than during the tone (Figure 5—figure supplement 1M). With this experimental design, speed was not affected by IL PV inhibition (Figure 5—figure supplement 1N–P, P: N=8 NpHR mice, N=5 eYFP mice, p=0.8846, two-way ANOVA interaction term). The latency of the first trials in inhibition blocks showed no significant difference from the latency of the second trials in inhibition blocks (Figure 5—figure supplement 1Q, N=8 NpHR mice, N=5 eYFP mice, p=0.6944, paired t-test), and the latency of the first trials after the end of the inhibition block also showed no difference to the latency of the second trials after the inhibition block (Figure 5—figure supplement 1Q, N=8 NpHR mice, N=5 eYFP mice, p=0.6745, paired t-test). Thus, post-avoidance inhibition of IL PV neural activity did not affect avoidance latency in subsequent trials.

All procedures conformed to guidelines established by the National Institutes of Health and have been approved by the Cornell University Institutional Animal Care and Use Committee. PV-Cre mouse line (B6.Cg-Pvalb^{tm1.1(cre)Aibs}/J, RRID:IMSR_JAX:012358) acquired from The Jackson Laboratory (Bar Harbor, ME) was backcrossed to C57BL/6 J mice (RRID:IMSR_JAX:000664). Postnatal six weeks to ten months old PV-Cre male mice were used for all photometry and optogenetics experiments. All mice were housed in a group of two to five under 12 hr reverse light-dark cycle (dark cycle: 9 a.m.-9 p.m.).

In the photometry experiment, we used AAV1-CAG-Flex-GCaMP6f (Titer: 1.33×10¹³, Penn Vector Core, 100835-AAV1, Philadelphia, PA) for experimental animals and AAV9-CAG.Flex.GFP (Titer: 3.7×10¹², UNC Vector Core, Chapel Hill, NC) for control. All the viral vectors used in photometry experiment were diluted in eightfold PBS before injection. In the optogenetics experiment, we used AAV5-EF1α-DIO-eNPHR3.0 (Titer: 4×10¹², UNC Vector Core, Chapel Hill, NC) for experimental animals and AAV5-EF1α-DIO-EYFP (Titer: 6.5×10¹², UNC Vector Core, Chapel Hill, NC) for control. All the viral vectors used in optogenetics experiment were used without dilution.

Mice were put under deep anesthesia with isoflurane (5%). Fur above the skull was trimmed, and the mice were placed in a stereotaxic frame (Kopf Instrument, Tujunga, CA) with a heating pad to prevent hypothermia. Isoflurane level was kept between 0.8 to 2% throughout the surgery. Ophthalmic ointment was applied to protect the eyes. 100 μL 1 mg/ml Baytril (enrofloxacin) was given subcutaneously, and 100 μL 2.5 mg/ml bupivacaine was injected subdermally at the incision site. The scalp was disinfected with betadine and alcohol. The skull was exposed with a midline incision. A craniotomy was made above the medial prefrontal (mPFC) cortex.

For fiber photometry animals, virus (AAV1-CAG-Flex-GCAMP6f) was injected into mPFC unilaterally, with half of the animals into the right hemisphere and half of the animals into the left hemisphere (infralimbic cortex (IL) coordinates: –1.55 AP,±0.3 ML, –2.8 to –3.2 DV). A total of 800 nl diluted vector (1:8 dilution) was injected in each mouse. Virus injection was done with a 10 μL Hamilton syringe (nanofil, WPI, Sarasota, FL) and a 33-gauge beveled needle, and a micro-syringe pump controller (Micro 4; WPI, Sarasota, FL) using slow injection speed (100 nl/min). The needle was slowly withdrawn 15 min after injection. After injection, a 4 mm or 6 mm-long optic fiber (diameter: 400 μm, 0.48NA, Doric Lenses, Quebec, Canada) was implanted 0.5–1 mm above the injection site.

For optogenetics animals, virus (AAV5-EF1a-DIO-eNpHR3.0, Lot #: 4806 G, titer: 4.00x10¹²; control: AAV5-EF1a-DIO-eYFP, Lot #: 4310 J, titer: 6.5x10¹², UNC vector core, NC, USA) was injected into mPFC bilaterally (IL coordinates: –1.55 AP,±0.3 ML, –2.8 to –3.2 DV). 500 nl vector was injected into each site on each mouse. After injection, a 4 mm to 6 mm-long optic fiber (diameter: 200 μm, 0.22NA, Thorlabs, NJ, USA) was implanted 0.5–1 mm above the injection site in a 15 degree angle toward midline (AP = 1.55, ML = ±1.05 or 1.15, DV = −2.8 or –2.99, 15 degree angle).

After the fiber implant, a layer of metabond (Parkell, Inc, Edgewood, NY) and dental acrylic (Lang Dental Manufacturing, Wheeling, IL) was applied to hold the implant in place. Buprenorphine (0.05 mg/kg), carprofen (5 mg/kg), and lactated ringers (500 μL) were administered subcutaneously after surgery. Photometry recording was done no earlier than three weeks later to allow for virus expression.

Fiber photometry was implemented with a fiber photometry console (Doric Lenses, Quebec, Canada). An FC-FC optic fiber patch cord (400 μm diameter, 0.48NA, Doric Lenses, Quebec, Canada) was connected to implanted fiber with a zirconia sleeve. 405 nm and 475 nm were measured for calcium-independent and calcium-dependent GCaMP signals and were measured with digital lock-in frequency to the input at 208 Hz and 530 Hz, respectively. Photometry signals were collected at 12 kHz and filtered with a 12 Hz low pass filter.

For photometry animals, four GCamp6f animals were tested only on active avoidance. Seven GCamp6f animals and two GFP control animals were tested in the following order: open field test, reward approach test, reward approach with omission (except two GFP and two GCamp6f animals), active avoidance, active avoidance with shortening and extended tone, active avoidance extinction.

In the behavioral experiment, two external FC-FC optic fiber patch cords (200 μm diameter, 0.22 NA, Doric Lenses, Quebec, Canada) were connected to two implanted fibers, respectively, each with a zirconia sleeve. These patch cords were then connected to a 1x2 fiber-optic rotary joint (FRJ_1x2i_FC-2FC_0.22, Doric Lenses, Quebec, Canada) for unrestricted rotation and to prevent tangling. Another FC-FC optic fiber patch cord was used to connect the rotary joint to a 100 mW 594 nm diode-pumped solid-state laser (Cobolt Mambo 100 594 nm, HÜBNER Photonics, Sweden) for optogenetic stimulation. The power of the laser was programmed by the software and fine-tuned by a continuous filter (NDC-50C-2M, Thorlabs, NJ, USA) to 10 mW at the end of the patch cord (~71.59 mW/mm² at the end of the implanted fiber). The stimulation timing was controlled by a shutter (SRS470, Stanford Research System, Sunnyvale, CA) and a Master-8 stimulus generator (A.M.P.I., Jerusalem, Israel). In avoidance and approach, a total of 72 trials were divided into 12 alternating blocks (OFF-ON-OFF…), and 2 s continuous stimulation was delivered 0.5 s after the onset of a tone during stimulation blocks. In the open field test, a 32 min test was divided into eight alternating stimulation blocks (OFF-ON-OFF…), and a 2 s continuous stimulation was delivered every 40 s during the stimulation blocks.

For optogenetics animals, eight experimental and five controls were tested in the following order: open field test under dim light, reward approach, active avoidance, open field test under bright light.

After experiments, animals were deeply anesthetized with pentobarbital at a dose of 90 mg/kg and perfused with 20 ml PBS, followed by 20 ml 4% paraformaldehyde solution. Brains were soaked in 4 °C 4% paraformaldehyde for 20 hr and then switched to 30% sucrose solution for 20–40 hr until the brains sank. Brains were sectioned coronally (40–50 μm) with a freezing microtome and then washed with PBS and mounted with PVA-DABCO. Images were acquired using a Zeiss confocal with 5 x air, 20 x water, and 40 x water objectives.

All data analysis and statistical testing were performed using custom-written scripts in MATLAB 2019 (MathWorks, Natick, MA). For all behaviors, location and movement were tracked using Ethovision XT10 (Noldus Leesburg, VA).

Error bars and shaded areas in figures report standard error of the mean (s.e.m.). All statistical tests were two-tailed. Within-subject analyses were performed using paired t-test, and between-subject analyses were performed using an unpaired t-test.

ΔF/F was calculated with equation (Equation 1). The signal measured from the 405 nm reference channel was linearly fitted to the 475 nm signal and was subtracted from the 475 nm signal. The difference was then divided by an exponential function (a∙e−b+c) fitted to 475 nm signal.(1)ΔFF=475nmsignal−fitted405nmsignalexponentialdecayfittedto475nmsignal×100

Mouse speed was first grouped into three clusters, nonmovement, low movement, and high movement, using k-means. The movement threshold was thus defined by the lowest speed of the low movement cluster. Then the peak speed of the movement was detected by finding local maxima with absolute peak value larger than a threshold and at least 5 cm/s larger than the baseline. (This was performed using findpeak function in MATLAB and with ‘MinPeakProminence’ set to 5 and with ‘MinPeakHeight’ set to movement threshold.) Movement initiation was defined by when the speed first went above 10% of the peak speed within 2 s before the epoch. Only epochs with at least 1 s of less than threshold speed before movement initiation were included. Group photometry analyses compared mean ΔF/F 4–2 s before and 2 s around the peak speed of the movement. Optogenetic analyses compared speed differences of NpHR- and eYFP-expressing mice between mean speed during the 2-s stimulation period and mean speed at the 2-s period right before stimulation.

Freezing is detected from video based on changes in pixels by a MATLAB function written by David A. Bulkin and Ryan J. Post. To detect the freezing of animals, we used the code first to convert the video into grayscale, crop the window to include only the bottom of the chamber, and then set the pixels belonging to a mouse to 1 s by thresholding pixel values below 27 out of 255. Once the pixels belonging to a mouse were assigned, the code compared the number of pixels changed between frames to get raw movement data. Raw movement data was then filtered with a bandpass filter between 0.01 and 0.9 to ensure the frequency of switching between freezing and non-freezing states matches with the behavior observed. Freezing was then marked when less than 190 mouse pixels were changed in filtered movement data. For Figure 5—figure supplement 1H, duration of freezing was calculated by the total number of frames when the animals were detected freezing during avoidance latency.

Photometry signals were first filtered by a lowpass IIR filter with half-power frequency at 7 Hz and then interpolated to match video recording timeframes (15 Hz). Each segment included photometry and speed data from 5 s after tone offset to 5 s before the next tone onset. Cross-covariance was calculated with an offset of ±2 s for each segment by the xcov function in MATLAB. The cross-covariance value with the largest absolute value within the offset range was selected to be the cross-covariance value of the segment (Seo et al., 2019).

The calcium signal was modeled as a linear combination of characteristic neural responses to each sensory or action event following the scheme of Musall et al., 2019 and Parker et al., 2016 (Musall et al., 2019; Parker et al., 2016). These events included tone onset, tone offset, avoidance, escape, shock onset, and chamber crossing in the intertrial intervals. The model can be written as the following equation, where ΔF/F (t) is the recorded calcium dynamics at time t, k_i and j_i are the kernel coefficients of event type I, τ is the relative time points in a kernel, n1 is the total number of action events, and n2 is the total number of sensory events. Each event kernel k_a or j_b was placed at the time point where the event a or b occurred. For action events, including avoidance, escape, and chamber crossing in intertrial intervals, we used kernels (k_i) ranging from 1 s before to 2 s after the event to cover the action from initiation to termination. For sensory events, including tone onsets, tone offsets, and shock onsets, kernels (j_i) started right at the moment where the event happened and lasted for 2 s afterwards to model the sensory responses. (2)ΔF/F(t)=∑i=1n1(∫τ=−1s2ski(τ)∗(t−τ)dτ)+∑i=1n2(∫τ=0s2sji(τ)∗(t−τ)dτ)+error

ΔF/F (t) was modeled as the linear combination of all event kernels. Given the event occurrence time points of all event types, we can use linear regression to decompose characteristic kernels for each event type. Kernel coefficients of the model were solved by minimizing the mean square errors between the model and the actual recorded signals. To prove that kernel k_i is an essential component for the raw calcium dynamics, we compared the explanation power of the full model to the reduced model where the time points of the occurrence of event k_i were randomly assigned. Thus, the kernel coefficients should not reflect the response to the event in the reduced model. The coefficients of determination (R²) were compared between the reduced model and the full model to estimate the unique contribution of certain events to the explanation power of the model. More details of the methods can be viewed in Musall et al., 2019; Parker et al., 2016.