Somatostatin-expressing interneurons enable and maintain learning-dependent sequential activation of pyramidal neurons

Summary

The activities of neuronal populations exhibit temporal sequences that are thought to mediate spatial navigation, cognitive processing and motor actions. The mechanisms underlying the generation and maintenance of sequential neuronal activity remain unclear. We found that layer 2/3 pyramidal neurons (PNs) showed sequential activation in the mouse primary motor cortex during motor skill learning. Concomitantly, the activity of somatostatin (SST)-expressing interneurons increased and decreased in a task-specific manner. Activating SST interneurons during motor training, either directly or via inhibiting Vasoactive Intestinal Peptide-expressing interneurons, prevented learning-induced sequential activities of PNs and behavioral improvement. Conversely, inactivating SST interneurons during the learning of a new motor task reversed sequential activities and behavioral improvement that occurred during a previous task. Furthermore, the control of SST interneurons over sequential activation of PNs required CaMKII-dependent synaptic plasticity. These findings indicate that SST interneurons enable and maintain synaptic plasticity-dependent sequential activation of PNs during motor skill learning.

eTOC blurb:

Adler et al. reveal mechanisms underlying learning-dependent sequential activation of pyramidal neurons in the primary motor cortex. SST-expressing interneurons and CaMKII-dependent synaptic plasticity control the establishment of sequential activity during motor training and prevent the interference from new learning.

Introduction

Sequential activation of pyramidal neuron (PN) populations is believed to be important for a wide variety of brain functions such as episodic memory formation, decision making and motor behavior (Wehr and Laurent, 1996, Yu and Margoliash, 1996, Peters et al., 2014, Pfeiffer and Foster, 2013, Pastalkova et al., 2008, Harvey et al., 2012). This sequential neuronal activation is characterized by distinct segregation of neuronal activities such that different neurons are active at different time periods of an animal’s behavior. While the sequential neuronal activity profile is dynamic during the process of learning (Manns et al., 2007, Ziv et al., 2013, Hainmueller and Bartos, 2018), its stability increases over time and is associated with behavioral improvement and performance stereotypy (Peters et al., 2014, Okubo et al., 2015, Hainmueller and Bartos, 2018, Pastalkova et al., 2008). Therefore, the establishment of stable sequential activity pattern is likely critical for information encoding and storage. Nevertheless, the mechanisms that generate and maintain learning-dependent sequential activation of PNs are poorly understood.

Inhibition can control and shape activity profiles of PNs, leading to increased temporal precision and tuning in response to sensory stimuli (Wehr and Zador, 2003, Pouille and Scanziani, 2001, Wilson et al., 2012). Network modeling suggests that the parsing of PNs into sequentially active groups depends on inhibition (Rabinovich et al., 2008, Klausberger and Somogyi, 2008, Gibb et al., 2009). Recently, a circuit motif of dis-inhibition, through activation of Vasoactive Intestinal Peptide (VIP)-expressing and inactivation of somatostatin (SST)-expressing GABAergic interneurons (INs), has been suggested to enable information processing and enhanced excitability of PNs (Pi et al., 2013, Fu et al., 2014, Lee et al., 2013, Urban-Ciecko and Barth, 2016, Pfeffer et al., 2013, Gentet et al., 2012). Whether and how inhibition involving SST INs and VIP INs affects learning-induced sequential activities of PNs remain unknown.

In addition to inhibition, network modeling studies suggest that the establishment of temporal sequence of PNs depends on spike timing-dependent plasticity (STDP) mechanisms (Blum and Abbott, 1996). A key concept of such STDP rules is that synaptic strength would be potentiated or de-potentiated depending on if the presynaptic neurons fire prior or after the postsynaptic neurons (Bi and Poo, 1998, Markram et al., 1997). One prediction from STDP-dependent synaptic strengthening is that the firing of presynaptic neurons would cause postsynaptic neurons to fire earlier in the late phase than the initial phase of learning (Mehta et al., 2000, Blum and Abbott, 1996). In support of this prediction, studies revealed a backward shift in the center of mass of the CA1 place fields (Mehta et al., 2000) and head direction tuning curves of the anterior thalamus (Yu et al., 2006). Together, these studies strongly suggest that activity-dependent synaptic plasticity is involved in establishing sequential activation of PNs. However, direct experimental evidence demonstrating the role of synaptic plasticity in the generation and maintenance of temporal sequences is lacking.

In this work, we investigated the roles of inhibition and synaptic plasticity in motor learning-dependent sequential activation of layer 2/3 (L2/3) PNs in the mouse primary motor cortex (M1). We observed a temporal shift and progressive stabilization in the sequential activation of L2/3 PNs during motor skill learning. Concurrently, SST INs displayed both increases and decreases in activity during motor training. Activating SST INs directly or indirectly by inactivating VIP INs prevented the learning-dependent shift in sequential activation of PNs and the animal’s behavioral improvement. On the other hand, inhibiting SST INs during the learning of a new motor task de-stabilized the sequential activity of PN population and reversed behavioral improvement established during previous learning. Furthermore, CaMKII-dependent synaptic plasticity was critically involved in the modulation of PNs sequential activity by SST INs. These findings demonstrate the important roles of SST INs and synaptic plasticity in enabling and maintaining sequential activation of PNs during motor skill learning.

Results

Motor learning-induced sequential activity of layer 2/3 PNs

To investigate the mechanisms underlying learning-related sequential activity of PNs, we first examined the population activity of L2/3 PNs in the M1 of head-restrained mice while they were trained to run forward on a treadmill at a fixed speed (Yang et al., 2014, Cichon and Gan, 2015) (Fig. 1A). In this training paradigm, mice changed their gait patterns over time such that with training they spent higher fractions of the running epoch in structured running, as opposed to unstructured movement: drag of feet, sweep and wobble (Figs. S1 and Video Sl). Significant performance improvement could be observed within 20 minutes of forward running (FR) training (Fig. 1B-D). To measure the running related activity of L2/3 PNs, we used transgenic mice expressing a genetically encoded Ca²⁺ indicator (GCaMP6s) in these cells and performed two-photon Ca²⁺ imaging before and following 20-minute FR training (Fig. 1A). The vast majority of PNs displayed low levels of activity when the treadmill was turned off and the animals were stationary. When the treadmill was turned on, L2/3 PNs increased their activity significantly (95% of all cells) during FR both before and after training (Fig. S2A-E). The increase of PN activity during running was transient, displaying short response durations relative to the running time (Fig. 1E, F, Fig. S2F; average response duration: 3.54±0.06s and 3.62±0.07s during pre- and post-training respectively in 45s running trials). Notably, the activity of individual PNs showed variable timing relative to the movement onset, such that different cells peaked their activity across running trials at different time points (for example, Fig. 1E, F, Fig. S2G, M). When the cells within an imaging plane were ordered according to the timing of their average peak activity, we observed a temporal sequence of neurons across the entire running epoch (Fig. 1G, and Fig. S2J, K). A similar temporal sequence of neuronal activity was observed when pooling all imaging planes from all animals together (Fig. 1I and Fig. S2L-P).

Figure 1. Motor learning induces temporal reorganization in the population activity of L2/3 PNs.

A. Left: Schematic of Ca²⁺ imaging of L2/3 PNs during treadmill training. Right: image of GCaMP6s-expressing L2/3 PNs. Yellow and cyan: cells depicted in E and F.

B. Percentage of different FR gait patterns pre- and post-training. 24 mice.

C. Percent of structured running increases after training (P< 0.0001; 24 mice).

D. Distance between animal’s front paws (stride width) during structured running decreases after training (P<0.0001; 24 mice).

E-F. Running related activity of two PNs (left and right columns) before (E) and after (F) training. 4 individual trials (color) and average ΔF/F ± s.e.m (black and gray) are shown (Arrow: treadmill on). Bottom row in (F): average activity during pre- (pink) and post-training (blue) zoomed on the peak (Arrows: time differences between peaks). Scale bars: 10s and 100% ΔF/F.

G-J. Maximum-normalized average activity to FR from all neurons (rows) imaged from a single animal pre- (G) and post- (H) training, and from multiple animals (508 cells, 17 animals) pre- (I) and post-training (J). Time zero: treadmill on. Cells are ordered according to the time of their peak responses.

K. Cumulative sum of the time of the peak activity of the neurons in (I) and (J), P<0.0001.

L. Peak activity time of PNs from (I) (x-axis) and (J) (y-axis). Each dot represents one PN. PNs peaked earlier in the running trial following training (P<0.0001).

M. Activity onset time of PNs from (I) (x-axis) and (J) (y-axis). Response onset occurred earlier post-training, P<0.0001.

N. The amplitude (ΔF/F) of the peak activity of PNs was not different before (x-axis) and after training (y-axis), P=0.13.

O. Cumulative sum of the time of the peak activity of PNs before (dashed red/purple) and after 20-minute (dashed blue/light blue) FR in naive mice or in mice after extensive training respectively (508 cells, 17 mice and 111 cells, 5 mice). Pre-training peak activity time is significantly different from all other groups (P<0.0001). After extensive training there is no longer a temporal shift in the population sequential activity.

P. Following extensive training the correlation between the ranks of the neurons pre- and post- FR training (0.33, red line) is significantly higher compared with random shuffling (P<0.001). 111 cells, 5 mice.

Q. The sequential activity profile and structured running are correlated (P<0.001). Each black dot represents the median of the times of peak activity of PNs vs. the percentage of structured running for a single running trial. The median is taken as a measure of the pattern of the sequential activity of PNs, i.e. amount of shift towards movement onset. 38 trials, 4 mice.

R. Similar as in (Q) except for wobble gait pattern. Sequential activity profile and wobble gait pattern are not correlated (P=0.056).

Statistical results are in Table S1. See also Figures S1, S2 and S8

Previous studies have shown that when M1 is selectively silenced, motor performance is largely intact but behavioral improvement after motor training is impaired (Cichon and Gan, 2015, Kawai et al., 2015, Peters et al., 2014). These findings suggest a view that motor cortex modulates the activity and plasticity of subcortical motor circuits such as basal ganglia and spinal cord, which are directly responsible for generating rhythmic locomotor behaviors (Kawai et al., 2015, Peters et al., 2014, Belanger et al., 1996, Shenoy et al., 2013, Lemon, 2008, Grillner, 2006). Consistently, we found that even though mice showed repetitive movement behaviors on the treadmill, L2/3 PNs exhibited sequential rather than repetitive activation during the 45s running period.

Notably, when we compared the temporal activity profile of the same PNs before and following the 20-minute FR training, we found that the temporal sequence of the population activity shifted in time such that most cells reached their peak activity earlier during the running epoch (Fig. 1E-L, and Fig. S2G-K). While the timing of the peak (or onset) activity was significantly shifted (Fig. 1K, L, M), the peak values themselves remained similar following training (Fig. 1N). When the activities of PN population were examined after extensive training over 2 days, we found that the temporal sequence was stabilized with no additional shift in timing (Fig. 1O). Additionally, the rank of the cells in the sequence (i.e. the identity of the individual cells in the ordered sequence) was significantly correlated before and after a 20-minute FR session (Fig. 1P).

The temporal shift in the population sequential activity is thought to reflect synaptic potentiation important for generating new sequences during learning (Blum and Abbott, 1996, Yu et al., 2008, Mehta et al., 2000). Given the importance of early neuronal spikes in information processing (Bolding and Franks, 2018, Gollisch and Meister, 2008, Thorpe et al., 2001), such a temporal shift in sequential activity may enhance information transfer from motor cortex to subcortical regions to modulate behavioral performance. Consistently, we found that the shift in the sequential activity of PNs paralleled behavioral improvements in the FR task (Fig. 1C, D and K). The temporal profile of PN sequential activity and the percent of structured running, but not wobbling, were positively correlated on a trial by trial basis (i.e. the more cells peaked earlier in the running trial, the more structured running performed, Fig. 1Q, R). Taken together, these results suggest that the sequential activity, its leftward temporal shift and stabilization may regulate the activity of subcortical circuits and facilitate structured running.

SST INs increase and decrease activities during motor learning in a task-specific manner

What are the circuit mechanisms underlying motor learning-induced sequential activity of L2/3 PNs? SST INs target the dendrites of PNs and have been suggested to regulate the excitability and information flow of neuronal circuits (Murayama et al., 2009, Chiu et al., 2013), specifically via a circuit motif of dis-inhibition involving VIP IN activation (Lee et al., 2013, Pfeffer et al., 2013, Fu et al., 2014). To test whether SST INs are critical for motor learning-dependent sequential activation of L2/3 PNs, we first examined their responses to motor training (Fig. 2A, B, and Video S2). In contrast to the increased activity of PNs in response to FR, the activity of L2/3 SST INs in M1 was diverse, displaying both decreases below and increases above baseline activity (Fig. 2C-G).

Figure 2. SST INs exhibit diverse and task specific responses to motor learning.

A. Schematic of Ca²⁺ imaging of L2/3 SST INs in M1.

B. Left: image of SST INs expressing GCaMP6m before and after FR. Right: FR induced changes in ΔF/F. Arrow: treadmill on. Top 4 traces: ΔF/F of cells on the left; bottom: average ± s.e.m (gray envelope). Scale bar: 100% ΔF/F and 10s.

C-D. Distribution of the average and peak baseline activity of SST INs (150 cells, 21 mice).

E. Average activity of SST INs (rows) in response to FR (150 cells, 21 mice). Baseline activity was subtracted from the average activity traces. Time zero: treadmill on.

F. Population average response ± s.e.m (gray envelope) of all SST INs with significantly reduced (left) or increased (right) activity to FR.

G. Percentages of SST INs with significantly decreased (blue) or increased (red) activity or no change (gray) during FR.

H-J. Similar to E-G, but during BR. Cells are reordered according to their BR response profiles. 150 cells, 21 mice.

K-M. Examples of SST IN responses to FR and BR averaged over 5 trials ± s.e.m (gray envelope). Responses to FR and BR are significantly different (P=0.049, 0.0056 and 0.006 for K, L and M).

N. Ratio of SST INs (out of all INs with significantly different responses to FR and BR, which constitute 50% of the cells) with opposite or similar response profile directions to FR and BR.

Statistical results are in Table S1. See also Figure S3

To further understand SST INs’ response diversity, we examined whether the different responses were cell autonomous or task dependent by imaging the activity of SST INs in response to both FR and backward running (BR). SST INs’ responses to BR were also diverse (Fig. 2H-J). Notably, when we compared the response profiles of the same SST INs to FR and BR, we found that among the cells with significantly different response profiles (see example cells in Fig. 2K-M), the vast majority (75%) had an opposite polarity (Fig. 2N). Thus, SST INs increasing their activity during FR were likely to reduce their activity during BR (Fig. 2L) and vice versa (Fig. 2M), indicating that the diversity in SST INs’ responses was largely task dependent.

We further imaged SST INs after a 20-minute FR training and found that most cells did not change their response profiles (76.7%, Fig. S3A-H) or their response onset time (average response onset time during pre-training: 3.51±0.65s; and post-training: 5.9±0.94s; Fig. S3I, J). However, the time of their peak responses were delayed (rightward shift) (average peak time during pre-training: 8.59±0.95s; and post-training: 15.42±1.09s; Fig. S3K). Because SST INs are directly connected to PNs (Urban-Ciecko and Barth, 2016, Wang et al., 2004, Kepecs and Fishell, 2014), the rightward shift of peak SST neuronal activity following training suggest a reduction of inhibition onto PNs at the start of running trials, which could contribute to the leftward shift of PN sequential activities and animals’ behavioral improvement.

Decreased activity of SST INs is required for motor learning-induced temporal shift in sequential activation of PNs

To test the potential role of SST INs in regulating motor learning-induced sequential activity of PNs, we imaged the running related activity of PNs while manipulating the activity of SST INs using the Designer Receptor Exclusively Activated by Designer Drug (DREADD) system (Dong et al., 2010). Specifically, we crossed SST Cre mice with transgenic mice expressing GCaMP6s in PNs and virally infected SST INs locally within M1 with either the activating (hM₃D(Gq)) or inhibiting (hM₄D(Gi)) form of the DREADD. While the DREADD/clozapine-N-oxide (CNO) system led to the expected changes in activities of SST INs and PNs (Fig. 3A-I and Fig. S4), modulating SST IN activity by CNO administration alone without 20-minute training did not affect sequential activation of L2/3 PNs (Fig. S4Q, U, and Y) or behavioral performance (Fig. S5).

Figure 3. Activation of SST INs blocks the temporal shift in sequential activation of L2/3 PNs and behavioral improvement.

A. Images of SST INs infected with DREADD Gi in M1. Image on the right: enlarged rectangle area.

B. Number of SST INs infected with DREADD in the densest 1 mm² area. 6 mice.

C. Images of L2/3 SST INs expressing both GCaMP6m (green) and hM3D(Gq)-mCherry (red).

D. Schematic of Ca²⁺ imaging of SST INs infected with DREADD Gq.

E. Two examples of SST IN activity to FR (arrow: treadmill on) before (left) and after (right) CNO administration in mice infected with DREADD Gq. Gray lines: single trials; black line: average.

F. The level of SST IN activity (ΔF/F) during FR is higher following CNO administration in mice infected with DREADD Gq (P<0.0001). 26 cells from 2 mice.

G-I. Similar as in D-F only for animals infected with DREADD Gi. (I): The level of SST IN activity is lower following CNO administration (P<0.01). 13 cells from 3 mice.

J. Schematic of Ca²⁺ imaging of PNs in SST Cre mice infected with DREADD Gq or Gi and injected with saline.

K. Maximum-normalized average activity of PNs pre- (left) and post- (right) training.

L. Cumulative sum of the time of the peak activity of PNs in (K). Post training saturates faster (P<0.0001).

M. Peak activity time of PNs (dots) from (K) pre- (x-axis) and post-training (y-axis). PNs peaked earlier following saline injection and FR training (P<0.0001).

N. Percent structured running is higher post-training (P<0.001; 16 mice).

O-S. Similar to J-N, except that mice were infected with DREADD Gq and received CNO. PR: 174 cells, 6 mice; (S): 8 mice. Peak activity timing did not shift (P=0.18, P=0.85 for Q and R). Mice did not display behavioral improvements (P=0.97 for S).

T-X. Similar to J-N, except that mice were infected with DREADD Gi and received CNO. UW: 140 cells, 4 mice; (X): 10 mice. Peak activity timing shifted toward treadmill on (P<0.0001, P<0.0001 for V and W). Mice displayed behavioral improvements (P=0.015 for X).

Statistical results are in Table S1. See also Figures S4 and S5

Notably, unlike the control condition (saline i.p injection, Fig. 3J-M), activating SST INs (CNO i.p injection, Fig. 3O) during the 20-minute FR training session prevented the temporal shift of the population sequence to earlier time points in the running trial (Fig. 3P-R), and the animals’ behavioral improvement (Fig. 3S, compare with Fig. 3N). In contrast, when the activity of SST INs was inhibited (CNO i.p injection, Fig. 3T) during the 20-minute FR, the temporal shift in sequential activation (Fig. 3U-W) as well as behavioral improvement (Fig. 3X) were not disrupted. Together, these findings suggest that the reduction in SST IN activity during motor training is important for enabling learning-induced sequential activation of PNs and performance improvement.

Increased activity of SST INs is required for maintaining previous motor learning-induced sequential activity of PNs

Since SST IN activity was not uniformly reduced during motor learning and inhibiting SST INs did not prevent the leftward temporal shift in the sequential activation of PNs nor the behavior improvement, the role of SST INs’ active inhibition remained unclear. Recent studies have shown that global deletion of SST INs resulted in the interference of previously-learned skills by new learning (Cichon and Gan, 2015). Because FR and BR training induced largely opposite activity patterns in SST INs (Fig. 2), this raised the possibility that the inhibitory activity of SST INs during one motor learning task may be involved in protecting the sequential activation pattern of PN populations induced by a different motor learning task. To test this possibility, we first imaged the activity of PNs before and following a 20-minute FR training session to establish the initial leftward temporal shift. Subsequently, animals received either saline injection or CNO injection to inhibit the activity of SST INs (DREADD hM₄D(Gi)), followed by an additional 20-minute training session in which mice were trained with the same FR task or a new task, BR. Finally, the activity of the same PNs in response to FR was imaged to determine the stability of the sequential activation pattern of PNs established in response to the initial FR training (Fig. 4A).

Figure 4. Inhibiting SST INs during new learning de-stabilizes previous learning-dependent sequential activity of PNs.

A. Training protocol to test the role of SST IN inhibition in stabilizing sequential activation of PNs.

B. Schematic of Ca²⁺ imaging of PNs in SST Cre mice infected with DREADD Gi and receiving saline prior to BR training.

C-E. Maximum-normalized average activity to FR from all PNs before (C), after 20 minutes FR training (D) and after saline injection/20 minutes BR training (E). 123 cells, 5 mice.

F. Cumulative sum of the time of the peak activity of PNs from (C), (D) and (E). Changes in sequential activity are maintained. P<0.0001 (C and D); P=0.01 (C and E); P=0.13 (D and E).

G. Peak activity time of single PNs (dots) is delayed before (x-axis) compared with after FR training (y-axis) (P<0.001).

H. Peak activity time of single PNs (dots) after FR training (x-axis) is comparable with after BR training following saline administration (y-axis).

I. Behavioral improvement is maintained after BR training. (P=0.002; 8 mice).

J-Q. Similar to B-I, except that mice were infected with DREADD Gi and received CNO before BR training. K-P: 120 cells, 5 mice; (Q): 8 mice. Leftward shift in temporal sequence (P<0.001 (K and L), P<0.01 (K and M), P<0.0001 (L and M)) and behavioral improvement (P<0.001) are reversed.

R-Y. Similar to B-I, except that mice were infected with DREADD Gi and received CNO before the second FR training. S-X: 70 cells, 4 mice; (Y): 8 mice. Leftward shift in temporal sequence (P<0.0001 (S and T), P<0.0001 (S and U), P=0.72 (T and U)) and behavioral improvement (P<0.001) are maintained.

Statistical results are in Table S1. See also Figures S4 and S5

In control mice receiving saline injection, the leftward temporal shift in the population sequential activation after FR training (Fig. 4B-D) was maintained following BR training (Fig. 4E-H). In parallel, the animals increased the percent of time spent in structured running following FR training, and this increase was also maintained after BR training (Fig. 4I). In contrast, in mice injected with CNO to inhibit the activity of SST INs during the introduction of BR training (Fig. 4J), the initial leftward shift in the population temporal sequence as well as behavioral improvement were reversed (Fig. 4K-Q). Notably, in mice injected with CNO to inhibit the activity of SST INs during a second FR training, the initial temporal shift in the population activity and behavioral improvement were maintained (Fig. 4R-Y).

Taken together, the findings above suggest that active inhibition of SST INs during motor skill learning is important for preventing the de-stabilization of previously-learned sequential activation pattern and for maintaining behavioral improvement when a new motor skill task is introduced.

Increased activity of VIP INs is required for enabling, but not for maintaining, motor learning-induced sequential activation of PNs

It has been suggested that during active sensation and complex behaviors, VIP INs are activated to inhibit SST INs, which in turn lead to dis-inhibition of PNs and potentially the induction of plasticity during learning (Pfeffer et al., 2013, Fu et al., 2014). Consistent with this notion, we found that VIP INs were stereotypically and significantly activated in response to FR and BR (Fig. 5A-D). Inhibiting the activity of VIP INs (Fig. 5E-H) during the 20-minute FR training prevented the temporal shift in sequential activation and animals’ behavioral improvement (Fig. 5I-Q). Because VIP INs exert strong inhibition onto SST INs (Pfeffer et al., 2013, Lee et al., 2013), this finding is in line with the impact of increased SST IN activity on preventing learning-induced sequential activation and performance improvement (Fig.3). Furthermore, when we reduced the activity of VIP INs during the introduction of the new BR (Fig. 5R), the changes in sequential activity and behavior improvement that occurred in response to initial FR training were still maintained (Fig. 5S, T). This finding is also in line with the importance of active SST IN inhibition for maintaining previously-learned sequential activation pattern and behavioral improvement when a new motor skill task is introduced (Fig. 4).

Figure 5. Increased activity of VIP INs is required for establishing, but not for maintaining, learning-induced sequential activation of PNs.

A. Schematic of Ca²⁺ imaging of VIP INs in VIP Cre mice infected with PSAM in M1.

B. Left: average activity of VIP INs to FR pooled across 74 cells from 12 mice. Baseline activity was subtracted from average activity traces. Time zero: treadmill on. Right: percentages of VIP INs with significantly decreased (blue) or increased (red) activity or no change (gray) during FR.

C. Similar as in (B) except during BR (74 cells, 12 mice).

D. Ratio of VIP INs (out of all INs with significantly different responses to FR and BR, which constitute 60.3% of the cells) with opposite or similar response profile directions to FR and BR.

E. Image of VIP INs infected with PSAM in M1.

F. Number of VIP INs infected with PSAM in the densest 1 mm² area. 6 mice.

G. Two examples of VIP IN activity to FR (arrow: treadmill on) before (left) and after (right) PSEM³⁰⁸ administration. Gray lines: single trials; black line: average.

H. The level of VIP IN activity (ΔF/F) following PSEM³⁰⁸ administration is lower (P<0.0001; 3 mice, 40 cells).

I. The level of PNs activity (ΔF/F) following PSEM³⁰⁸ administration is lower (P<0.001; 3 mice, 89 cells).

J. Peak activity time of PNs is not significantly different before and following PSEM³⁰⁸ administration (P=0.1; 3 mice, 89 cells).

K. Percent structure running is not significantly different before and following PSEM³⁰⁸ administration (P=0.69; 8 mice).

L. Left: training protocol to test the effect of ACSF administration on learning-induced sequential activity. Right: schematic of Ca²⁺ imaging of PNs in VIP Cre mice infected with PSAM and receiving ACSF locally in M1.

M. Cumulative sum of the time of the peak activity of PNs pre- (red) and post-training (blue). Post-training saturates faster (P<0.0001; 170 cells, 4 mice).

N. Percent structure running is significantly higher post-training (P<0.001; 7 mice).

O-Q Similar as in (L-N) except that mice were administered PSEM³⁰⁸. Peak activity time of PNs (P, 204 cells, 4 mice) and percent structured running (Q, 8 mice) are not significantly different pre- and post-training (P=0.08 and P=0.97)

R. Left: training protocol to test the function of VIP IN inhibition in the stabilization of the sequential activation of PNs. Right: schematic of Ca²⁺ imaging of PNs in VIP Cre mice infected with PSAM and receiving PSEM³⁰⁸.

S. Cumulative sum of the time of the peak activity of PNs pre- and post- FR training and after PSEM³⁰⁸ administration/BR training. Changes in sequential activity are maintained. P<0.0001 (pre- and post training); P=0.13 (post-training: FW and BW PSEM); 131 cells, 4 mice.

T. Percent structured running pre-, post- FR training and after BR training following PSEM³⁰⁸ administration. 8 mice. Behavioral improvement maintained after BR training (P<0.001).

Statistical results are in Table S1.

Taken together, these results suggest that VIP IN activity (likely by inhibiting SST INs) is important for enabling, but not maintaining learning-induced sequential activity of PNs and behavioral improvement.

Increase and decrease of SST IN activity are important for promoting and maintaining synaptic activity of L2/3 PNs

Computational neural network models have suggested a critical role for synaptic plasticity in the emergence of sequential activity patterns of PNs and their temporal shift (Jun and Jin, 2007, Fiete et al., 2010, Rajan et al., 2016, Blum and Abbott, 1996). Additionally, SST INs, which directly target the dendrites of PNs (Urban-Ciecko and Barth, 2016, Wang et al., 2004), have been shown to control Ca²⁺ signals at individual L2/3 pyramidal dendritic spines (Chiu et al., 2013) as well as dendritic Ca²⁺ spikes associated with synaptic plasticity (Cichon and Gan, 2015, Murayama et al., 2009). To investigate whether synaptic plasticity is involved in SST IN-dependent sequential activity of PNs, we first imaged Ca²⁺ activity of dendritic spines on apical dendrites of L2/3 PNs in M1 with or without manipulating the activity of SST INs (Fig. 6A-C, Fig. S6A-F and Video S3). Ca²⁺ activity of spines (but not dendritic shafts) was significantly different following 20-minute FR training (for spines, Fig. 6D. For dendritic shaft, Paired Wilcoxon rank sum, U=0.8 P=0.3; 48 dendrites). When SST INs were activated by CNO/DREADD hM₃D(Gq) during FR training, spine Ca²⁺ activity was comparable before and after training (Fig. 6E, F). Thus, FR training resulted in changes in synaptic activities that were blocked by the elevation of SST IN activity.

Figure 6. Inhibiting SST INs during new motor learning reverses previous learning-induced changes in spine Ca²⁺ activity of PNs.

A. Images of PN dendrite expressing GCaMP6s during FR at two individual time points (t₁ and t₂). Boxed area is enlarged on the left of each image. Left: single spine is active, right: Ca²⁺ activity occurs along dendritic shaft invading all spines. Arrow heads: spines with activity traces in B.

B. Individual traces for spines (color) and shaft (black) showing changes in ΔF/F in response to FR.

C. Spine activity after removing contributions of Ca²⁺ signal due to back-propagating action potentials.

D. Peak Ca²⁺ activity of PN dendritic spines pre- and post- FR training. 155 spines, 19 mice. Spine Ca²⁺ activity following FR training is significantly different compared with before (P<0.05). Inset: population average spine activity (ΔF/F) pre- and post-training (P<0.0001). Population average was taken over spines for which the normalized activity following FR training was equal to or higher than before training.

E. Schematic of Ca²⁺ imaging of dendritic spines in SST Cre mice infected with DREADD Gq and injected with CNO.

F. Similar as in (D), except that mice infected with DREADD Gq received CNO prior to FR training. 39 spines, 6 mice. Spine Ca²⁺ activity is comparable before and after FR training (P=0.5).

G. Training protocol to examine stabilization of spine Ca²⁺ activity changes with or without reducing SST IN activity.

H. Schematic of Ca²⁺ imaging of dendritic spines in SST Cre mice infected with DREADD Gi and injected with saline.

I. Peak Ca²⁺ activity of dendritic spines to FR: before training, following 20 minutes FR training and following saline injection and 20 minutes BR training. 40 spines, 8 mice. Changes in synaptic activity following FR training are maintained following BR training (P<0.001).

J. Schematic of Ca²⁺ imaging of dendritic spines in SST Cre mice infected with DREADD Gi and injected with CNO.

K. Similar as in (I), except that animals received CNO before BR training. 37 spines, 9 mice. Changes in spine Ca²⁺ activity following FR training are reversed following BR training (P<0.01).

L. Similar as in (I), except that animals received CNO before additional FR training. 45 spines, 5 mice. Changes in spine Ca²⁺ activity following FR training are maintained following second FR training. (P<0.01).

Statistical results are in Table S1. See also Figure S6

We further examined if changes in spine Ca²⁺ activity occurring in response to FR training would be sustained following either FR or BR training while SST INs were inhibited by CNO/DREADD hM₄D(Gi) (Fig. 6G). Under the control condition, changes in spine activity following FR training were maintained following the introduction of new BR task (Fig. 6H, I). However, when SST INs were inhibited during BR training, changes in spine activity that occurred initially following FR training were reversed following BR training (Fig. 6J, K). On the other hand, when SST INs were inhibited during a second FR training, the initial changes in spine activity were maintained (Fig. 6J, L). Taken together, these changes in dendritic spine Ca²⁺ activity suggest that synaptic plasticity likely occurs in learning-induced sequential activation of PNs.

In addition to spine changes described above, we also found that activating or inhibiting SST INs by CNO/DREADD alone, without 20-minute training, significantly reduced or increased respectively the frequency of FR related dendritic shafts Ca²⁺ activity (Fig. S6G, H). Additionally, when SST INs were inhibited, dendrites active to FR were more likely active also to BR (Fig. S6I-K). Because SST INs directly target pyramidal dendrites and spines (Wang et al., 2004, Chiu et al., 2013, Murayama et al., 2009, Urban-Ciecko and Barth, 2016), these results suggest that increased and decreased SST IN inhibition may modulate dendritic spine activity either directly, or indirectly by affecting the frequency and location of dendritic Ca²⁺ signaling important for synaptic plasticity.

CaMKII-dependent synaptic plasticity is important for SST IN-dependent sequential activation and stabilization of PNs

To directly test the involvement of synaptic plasticity in SST IN-dependent sequential activation and stabilization of L2/3 PNs, we manipulated the activity of Calcium/calmodulin-dependent protein kinase II (CaMKII), a key signaling molecule in regulating synaptic plasticity (Lisman et al., 2002, Lledo et al., 1995). In this experiment, we locally infused CaMKII inhibitors (KN-62, KN-93) (Pellicena and Schulman, 2014), or a negative control (KN-92), into M1 prior to 20-minute FR training (Fig. 7A, B, and Fig. S7). The CaMKII inhibitors, but not the negative control, abolished the temporal shift in sequential activity (Fig. 7C, E), and behavioral improvement following training (Fig. 7D, F). We also tested whether synaptic plasticity is involved in SST IN-dependent de-stabilization of sequential activation. In this experiment, we inhibited SST INs (CNO/DREADD hM₄D(Gi)) and infused a CaMKII inhibitor (KN-62) when introducing BR (Fig. 7G, H). The leftward shift in the population activity following the initial 20-minute FR training was now maintained following the introduction of BR (Fig. 7I, J, and compare with Fig. 4K-N). Together, these findings suggest that the establishment and maintenance of learning-dependent sequential activities of PNs involve CaMKII-dependent synaptic plasticity.

Figure 7. Pharmacological blocking of CaMKII activity prevents the establishment and destabilization of SST IN-dependent sequential activity of PNs.

A. Training protocol to test the effect of blocking CaMKII on learning-induced sequential activity.

B. Schematic of Ca²⁺ imaging of PNs after local application of CaMKII blockers.

C. Peak activity timing shifted toward treadmill on with KN-92 infusion (P<0.0001; 114 cells, 4 mice).

D. Percent structured running is significantly different pre- and post-training (P<0.0001; 7 mice).

E. Similar as in (C) for KN-62 (left, 82 cells, 4 mice) or KN-93 (right, 87 cells, 4 mice). Pre- and post-training are not significantly different (P=0.42 and P=0.46 for KN-62 and KN-93).

F. Similar as in (D) for KN-62 (left, 9 mice) or KN-93 (right, 7 mice). Pre- and post-training are not significantly different (P=0.74 and P=0.87 for KN-62 and KN-93).

G. Training protocol to test the role of CaMKII in SST IN-dependent destabilization of sequential activity using KN-62.

H. Schematic of Ca²⁺ imaging of PNs in SST Cre mice infected with DREADD Gi, injected with CNO and infused with KN-62.

I. Maximum-normalized average activity to FR. Left: pre-training. Middle: post 20-minute FR training. Right: post CNO injection, KN-62 infusion and 20-minute BR training. 99 cells, 4 mice.

J. Cumulative sum of peak activity time of PNs. Leftward shift is maintained. P<0.0001 (pre- and post training); P=0.13 (post-training: FW and BW/CNO/KN62).

Statistical results are in Table S1. See also Figure S7

In the experiments above, the pharmacological inhibitors (KNs) were applied locally in M1 and blocked CaMKII activities in all cell types. We therefore used a genetically encoded light-inducible inhibitor of CaMKII (paAIP2) to block CaMKII activity specifically in PNs, and only during treadmill training (Murakoshi et al., 2017). In vitro photoactivation of this inhibitor has been shown to acutely block the induction of long-term synaptic plasticity of dendritic spines (Murakoshi et al., 2017). The CaMKII inhibitor, but not the non-functional mutant, blocked the temporal shift in the population sequential activity and the animals’ behavioral improvement following training (Fig. 8A-I). Additionally, the CaMKII inhibitor, but not the non-functional mutant, blocked SST IN-dependent de-stabilization of sequential activation (Fig. 8J-Q).

Figure 8. CaMKII-dependent synaptic plasticity is involved in SST IN-dependent temporal reorganization of PN activity.

A. Images of L2/3 PNs expressing a light inducible CaMKII inhibitor (left) and red genetically encoded Ca²⁺ indicator RGECO (middle).

B. Changes in RGECO fluorescence (ΔF/F) in response to FR from single cells (color). Arrow: treadmill on. Scale bars: 10s and 50% ΔF/F (cell 3 (green) scale bar: 100% ΔF/F).

C. Training protocol to test the effect of blocking CaMKII with the light inducible CaMKII inhibitor.

D. Schematic of Ca²⁺ imaging of PNs infected with the non-functional (control) light inducible CaMKII inhibitor.

E. Peak activity timing shifted toward treadmill on following FR training (P<0.0001; 64 cells, 4 mice).

F. Percent structured running is significantly higher after training (P<0.0001; 8 mice).

G. Schematic of Ca²⁺ imaging of PNs infected with the light inducible CaMKII inhibitor.

H. Similar as in (E) for active inhibitor. 118 cells, 6 mice. Pre-training is not significantly different from post-training (P=0.44).

I. Similar as in (F) for active inhibitor. 8 mice. Pre-training is not significantly different from post-training (P=0.62).

J. Training protocol to test the role of CaMKII in SST IN-dependent destabilization of sequential activity with the non-functional light inducible CaMKII inhibitor.

K. Schematic of Ca²⁺ imaging of PNs in SST Cre mice infected with DREADD Gi in SST INs and control inhibitor in PNs. Mice were injected with CNO and received BR training under blue light.

L. Maximum-normalized average activity to FR. 91 cells, 4 mice. Left; pre-training. Middle; post 20-minute FR training. Right; post CNO injection and 20-minute BR training with blue light illumination.

M. Cumulative sum of the time of the peak activity of PNs. Changes in sequential activity are not maintained following BR with blue light illumination and CNO injection in the control mice. P<0.0001 (pre- and post-training/FW); P=0.15 (pre- and post-training/BW light/CNO); P<0.0001 (post training: FW and BW/light/CNO).

N-Q. Similar as in J-M only for mice infected with active inhibitor. 97 cells, 5 mice. Changes in sequential activity are maintained following BR with blue light illumination and CNO injection. P<0.0001 (pre- and post-training/FW); P<0.0001 (pre- and post-training/BW light/CNO); P=0.42 (post training: FW and BW/light/CNO).

Statistical results are in Table S1.

Together, these findings with pharmacological and optogenetic approaches suggest that SST IN-dependent temporal shift in the sequential activity of PNs requires CaMKII-dependent synaptic plasticity. They also suggest that active inhibition by SST INs is important for reducing synaptic plasticity and stabilizing learning-dependent changes of sequential activation during subsequent new motor skill learning.

Discussion

Learning-dependent sequential activation of neuronal population is associated with behavioral improvement but the mechanisms underlying its generation and stabilization remain unclear. By imaging the activities of PNs and SST INs in the mouse M1, we show rapid establishment of the temporal sequence of L2/3 PNs while SST INs exhibit increases and decreases in activities during motor skill learning. Furthermore, using chemogenetic, optogenetic and pharmacological approaches to manipulate SST IN, VIP IN and CaMKII activities, we reveal that the emergence and maintenance of learning-dependent sequential activation of PNs are controlled by SST INs and CaMKII-dependent synaptic plasticity.

Sequential activation patterns of neurons have been observed in various brain regions and behavioral states (Peters et al., 2014, Wehr and Laurent, 1996, Yu and Margoliash, 1996, Pastalkova et al., 2008, Harvey et al., 2012). Indeed, the sequential activity profile can be generated in response to sensory stimuli, during motor actions and sleep replay (Hoffman and McNaughton, 2002, Pastalkova et al., 2008), and is dynamic over minutes to weeks (Manns et al., 2007, Ziv et al., 2013, Hainmueller and Bartos, 2018, Driscoll et al., 2017). The ability of the brain to generate sequential activities switching from one co-activated group of neurons to the next is likely critical for information transfer, memory consolidation, and animals’ behaviors (Kleinfeld and Sompolinsky, 1988, Amari, 1972). While inhibitory circuits have powerful control over a variety of brain functions (Wehr and Zador, 2003, Haider et al., 2013, Pouille and Scanziani, 2001, Royer et al., 2012), the involvement of inhibition in the generation and maintenance of sequential neuronal activities is not clear. Previous studies have established a role for VIP INs and SST INs in information flow in the pyramidal neuronal circuit during sensory processing and learning. Specifically, a canonical circuit in cortical and sub-cortical regions was described, whereby activated VIP INs lead to the dis-inhibition of PNs via inactivation of SST INs during active sensation and reward guided tasks (Pi et al., 2013, Fu et al., 2014, Lee et al., 2013, Pfeffer et al., 2013). Our results support a role for this circuit motif in learning-dependent sequential activity of PNs. In line with the suggested microcircuit, we found that increasing, but not decreasing, SST IN activity either directly with the DREADD approach or indirectly via inhibition of VIP INs, prevents the temporal shift of PNs’ sequential activity and behavioral improvement. Without training, increasing or reducing SST IN activity alone did not induce the temporal shift or the behavioral improvement. Together, these results suggest that VIP INs inhibit SST INs, which in turn dis-inhibit PNs to enable motor learning-induced sequential activity of PNs and behavioral improvement.

In addition to the roles of SST INs and VIP INs in enabling learning-induced sequential activation, we also found that previously-established temporal sequence would be de-stabilized when mice were trained with a new task and SST IN activity was inhibited with the DREADD system. This effect of reducing SST IN activity on temporal sequence de-stabilization only occurred when the new BR task, but not the previous FR task, was introduced. This finding indicates that without the inhibition from SST INs new learning would alter existing sequential activity patterns of L2/3 PNs. Consistently, when we inhibited VIP IN activity while introducing new BR motor training, the initial FR-related changes in sequential activity and behavior improvement were not interfered. Together, these results suggest an important role of increased SST IN inhibition in protecting and stabilizing the previously formed sequential activity profile and behavioral improvement.

Unlike a rather uniform reduction in SST IN activity in sensory cortices (Lee et al., 2013, Gentet et al., 2012, Fu et al., 2014), our results show that SST INs in L2/3 of M1 display both reduced and elevated activities in response to FR and BR. Because nearly all VIP INs exhibit increased activity to both FR and BR, VIP INs are likely responsible for the decrease, but not increase, in SST IN activity. We found that individual SST INs exhibit different and opposite activity patterns in response to different motor learning tasks, suggesting that motor task-specific inputs are important for the increased SST IN activity. Taken together, these findings suggest that VIP INs provide global inhibition to SST INs and reduce their activity for learning-dependent sequence emergence, while excitatory inputs may increase the activity of SST INs in a task-related manner for protecting the recently formed sequential activity pattern of PNs. Our results thus point to the dual functions of SST INs in the motor cortex: on the one hand, dis-inhibition of SST INs facilitates the information flow and learning-dependent sequence emergence. On the other hand, increased SST INs inhibition protects the previously formed sequential activity profile.

Our study indicates that SST INs and synaptic plasticity are two important components for enabling and maintaining learning-dependent sequential activity patterns. SST INs send extensive axonal projections to L1 where they contact apical dendrites of PNs (Urban-Ciecko and Barth, 2016). They directly block the initiation of dendritic Ca²⁺ spikes occurring in response to sensory stimuli (Murayama et al., 2009) and disrupt branch specific Ca²⁺ spikes during motor skill training in L5 PNs (Cichon and Gan, 2015). Additionally, SST INs exert a high level of spatial precision controlling Ca²⁺ signals at the level of individual L2/3 pyramidal dendritic spines (Chiu et al., 2013). Consistent with these studies, we found that SST INs control the frequency and motor task specificity of L2/3 pyramidal dendritic Ca²⁺ activity, suggesting that reduced SST IN activity may increase dendritic Ca²⁺ activity to promote synaptic plasticity during new learning, while increased SST IN activity may reduce dendritic Ca²⁺ activity to protect synaptic plasticity established earlier. It is thus possible that dis-inhibition and inhibition from SST INs promote and prevent synaptic changes of L2/3 PNs via Ca²⁺ spike-timing dependent mechanisms. Future studies are needed to better understand the link between SST IN activity and dendritic Ca²⁺ signaling in regulating motor learning-induced plasticity and maintenance of sequential activity. Furthermore, in our study, the sequential activity of PNs corresponds to repetitive running action spanning tens of seconds, whereas in previous work sequential activity was observed in M1 during a single, typically 1 second long, motor action (Peters et al., 2014). Future studies are also needed to investigate whether similar mechanisms are responsible for generating and maintaining sequential activities of PNs under different motor learning paradigms.

Computational modeling studies have demonstrated that a randomly connected network can produce sequential activity patterns based on the plasticity of synaptic weights (Jun and Jin, 2007, Fiete et al., 2010, Rajan et al., 2016). They also suggest a critical role of the asymmetric STDP rules in the temporal sequence generation and predicted a learning-dependent shift of the temporal sequence (Blum and Abbott, 1996, Fiete et al., 2010). Rapid establishment and temporal shift of sequential activities have been observed in CA1 place fields and head-directional cell tuning curves in the thalamus (Mehta et al., 2000, Yu et al., 2006), as well as in our studies of L2/3 PNs in M1, suggesting that common STDP mechanisms may underlie the generation of learned sequences. Activated CaMKII has been shown to be both sufficient and necessary for long-term potentiation induction and for certain forms of learning and memory (Lisman et al., 2002, Lledo et al., 1995). Our results of blocking CaMKII activities provide strong experimental evidence for a critical role of synaptic plasticity in the generation and destabilization of learning-dependent sequential activation of PNs. In addition to pharmacological reagents, we blocked CaMKII activity during treadmill training with a genetically encoded light-inducible inhibitor of CaMKII, which enabled us to inhibit CaMKII activity acutely and specifically in PNs and to potently target only CaMKII (Murakoshi et al., 2017). It will be of interest to use this novel approach to investigate whether learning-dependent sequential activation and stabilization in other brain regions also involve CaMKII-dependent synaptic plasticity.

Lastly, while our studies reveal the roles of inhibition and synaptic plasticity in generating and maintaining learning-dependent sequential activity of L2/3 PNs in the M1, the precise functions of this sequential activity and its temporal shift in the context of motor behaviors remain to be further investigated. Recent studies have shown that the motor cortex is not required for locomotion or execution of the learned skills but is essential for generating plasticity in downstream motor circuits during learning (Cichon and Gan, 2015, Peters et al., 2014, Kawai et al., 2015). Based on previous work, our results, and the fact that L2/3 PNs are several synapses away from the muscle motor output, it is likely that the sequential activity of L2/3 PNs only modulates, rather than determines, the activity and plasticity in the basal ganglia and spinal cord for motor performance. Furthermore, given the importance of early neuronal spikes in information processing (Bolding and Franks, 2018, Gollisch and Meister, 2008, Thorpe et al., 2001), it is possible that the leftward shift in sequential activity enhances information transfer from M1 to subcortical regions, perhaps for earlier anticipatory adjustments of posture and motor preparation. Consistent with this view, we observed that the temporal profile of the sequence and the percent structured running performed by the animals were correlated on a trial by trial basis. However, it should be noted that when we examined each session independently, we found that the temporal pattern of the sequence and structured running were significantly correlated only after, but not before, the 20-minute FR training (Fig. S8). The population activity of L2/3 PNs showed reduced variability with training, specifically by stabilization of the temporal pattern and in the order of the sequence post 2-day training. Based on our findings and in line with previous studies (Peters et al., 2014, Rokni et al., 2007), we speculate that during the initial training phase, multiple activity patterns of PNs in the M1 can lead to a similar behavioral motor output. The temporal shift and reduced variability in sequential activity occur with training, leading to a more consistent behavioral output. Future studies are needed to manipulate sequential activity of L2/3 PNs directly to better understand its role in reducing the variability of behavioral output and in behavior improvement during motor learning.

STAR Methods

Contact for Reagent and Resource Sharing

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Dr. Wen-Biao Gan (Wenbiao.Gan@med.nyu.edu).

Experimental Model and Subject Details

Experimental animals.

Transgenic mice expressing GCaMP6s in PNs (Thy1.2-GCaMP6s Line 3) and Cre recombinase in SST INs (SST-IRES-Cre) or in VIP INs (VIP-IRES-Cre) were group-housed (5 per cage) in the Parasitology Central Animal Facility or in the Smilow Animal Facility at New York University Medical Center. SST-IRES-Cre (Sst^{tm2.1(cre)Zjh}) and VIP-IRES-Cre (Vip^tm1(cre)Zjh) mice were purchased from the Jackson Laboratories. Mice were maintained at 22 ± 2°C with a 12-hour light: dark cycle. All experiments were conducted during the light cycle between 08:00-18:00. Food and water were available at libitum. 1-1.5 month-old animals of both sexes were used for all the experiments in accordance with ethical regulations and the New York University Medical Center (NYUMC) guidelines, and approved by the Institutional Animal Care and Use Committee (IACUC) at NYUMC. No association was found between animals’ sex and experimental results.

Methods Details

Treadmill training.

Animals were trained to perform a treadmill motor task in which head-fixed mice were forced to run (forward or backward) at fix speed (8 cm/s or 4cm/s respectively) on a free-floating treadmill (Cichon and Gan, 2015, Yang et al., 2014). Animals were positioned on a custom-made head holder/body support device that allowed the micro-metal bars (attached to the animal's skull) to be mounted so that the mouse forelimbs were positioned on top of the treadmill belt and its head was below the microscope objective (Yang et al., 2013). Mice were allowed to only move their forelimbs for locomotion to minimize motion artifacts in the images during treadmill running. Therefore, only the forelimb region of M1 was imaged, and the movement of forelimbs was monitored. The free-floating treadmill (96 cm × 56 cm × 61 cm dimensions) was custom built such that it could fit into the imaging apparatus (Yang et al., 2014, Cichon and Gan, 2015). The treadmill belt was cut out of black rubber sheets (0.031”) and was connected to a motor (Dayton, Model 2L010, Grainger, USA) driven by a DC power supply (Extech; Calgary, Canada). The treadmill belt was mounted on top of a metal frame. The treadmill itself and all its supporting units were not in contact with either the microscope stage or the supporting air table. During motor training and at the onset of a running trial, the motor was turned on and the belt speed gradually increased from 0 cm/s to 8 cm/s within ~3s. The speed of 8 cm/s was kept for the rest of the trial as the animals were forced to run. Each running trial lasted 45s (or 30s for imaging of SST and VIP INs) followed by an inter-trial interval (ITI, treadmill turned off) of 15s. During the ITI the treadmill was turned off and the animals were stationary, and during the trial the treadmill was turned on and the animals were forced to run throughout the entire period (45s). Treadmill training was performed in 20 minutes session blocks and testing was performed based on a minimum of 4 trials per session.

To assay gait patterns, mouse forelimbs were first coated into ink (Speedball, Stateville, NC, USA) before running on white construction paper. Animals running on the treadmill produce four distinct characteristic gait patterns which are identifiable from forelimb ink prints on white construction paper (Fig. S1). These gait patterns include: drag, wobble, sweep, and steady run. During drag, mice fail to execute steps resulting in their forelimbs being dragged on the treadmill belt thus producing a smearing of ink on the paper in the running direction (in the direction of the moving treadmill belt). During wobble, mice lose their balance and make many disorganized chaotic steps thus producing a highly variable ink pattern both in the running direction and lateral to running direction. During sweep, mice make steps lateral to the running direction thus producing ink marks that are perpendicular to the running direction. Finally, during steady run, mice make distinct steps in the running direction thus producing relatively evenly spaced distinct ink marks centered in the direction of running. Consistent with previous studies (Yang et al., 2014, Cichon and Gan, 2015), the distribution of gait patterns changed with training, such that during pre-training condition in early FR trials mice spent a higher fraction of the running trials performing drag (29.68±3.65%), wobble (22.94±1.95%) and sweep (13.13±1.31%) gait patterns (total of 65.77±3.78%) compared with steady run (34.22±3.78%). However, with training mice improve and refine their gait patterns: switching from drag (6.58±1.32%), wobble (5.31±0.86%) and sweep (5.14±0.82%) gait patterns (total of 17.04±1.83) into steady run (82.95±1.83%) as the prevalent gait pattern (Fig. 1B, and Fig. S1).

Since the steady structured running is a strong indication of behavioral improvement, we have used the percent of time animals spent in structured running and the stride width to describe animals’ behavior. When mice performed structured running, their step sizes (within limb stride distance) ranged from 1.5-2 cm (25%-75% of stride distance distribution across animals) at the speed of ~4-5.5 running steps per second. However, this periodicity was disrupted by other unstructured gait patterns, especially during pre-training.

Previous studies have suggested that this treadmill motor task involves motor skill learning as other motor learning tasks. First, similar to rotarod training and single-pellet reaching tasks, mice subjected to the treadmill running demonstrated rapid (over hours) and slow (over days) phases of behavioral improvement as measured by the percent of time spent in structured running and by the structured running stride width before and after treadmill training (Cichon and Gan, 2015). Second, the behavioral improvement is abolished if treadmill training is executed in the presence of muscimol in the primary motor cortex (M1) (Cichon and Gan, 2015). Third, treadmill training induces rapid new spine formation similar to rotarod training and single-pellet reaching tasks (Yang et al., 2014).

Surgical preparation for imaging awake, head-restrained mice.

We performed Ca²⁺ imaging in awake, head restrained mice through a cranial window. 24 hours prior to imaging, mice underwent surgery to attach a head holder and to create either an open-skull or a thinned-skull cranial window as previously described (Yang et al., 2013). Specifically, mice were deeply anesthetized with an intraperitoneal injection of ketamine (100 μg/g) and xylazine (10 μg/g). The mouse head was shaved and the skull surface was exposed with a midline scalp incision. The periosteum tissue over the skull surface was removed without damaging the temporal and occipital muscles. Two parallel micro-metal bars were attached to the animal’s skull to serve as the head holder to help restrain the animal’s head and reduce motion-induced artifact during imaging. A skull region of ~0.2 mm in diameter, was located over the primary motor cortex based on stereotactic coordinates (at 0.2 mm anterior to bregma and 1.2 mm lateral to midline) and marked with a pen (Tennant et al., 2011). A thin layer of cyanoacrylate-based glue was first applied to the top of the entire skull surface. Next the head holder was mounted with dental acrylic cement (Lang Dental Manufacturing Co., IL, USA) such that the marked skull region was exposed between the two bars. The marked region for imaging was kept exposed, uncovered with dental acrylic cement.

Next, we created the cranial window over the marked region. After the dental cement was completely dry, the head holder was screwed to two metal cubes that were attached to a solid metal base. A high-speed drill was used to carefully reduce the skull thickness under a dissecting microscope (skull thickness ~20μm). The skull was immersed in artificial cerebrospinal fluid (ACSF) during drilling. For open skull preparation a small circular craniotomy (1.0-1.5 mm diameter) was made and covered with a round glass coverslip (World Precision Instruments, cover slips, 5 mm diameter) custom to the size of the bone removed. The coverslip was glued to the skull to reduce motion of the exposed brain. Before imaging, mice were given one day to recover from the surgery related anesthesia, and habituated for a few times (10 minutes each) in the imaging apparatus to minimize potential stress effects due to head restraining and awake imaging.

Calcium imaging in L2/3 neurons expressing GCaMP.

Genetically-encoded Ca²⁺ indicator, GCaMP6 was used for Ca²⁺ imaging of somas, dendrites and spines of L2/3 PNs and of somas of L2/3 SST-expressing and VIP-expressing INs in the primary motor cortex. Transgenic mice expressing GCaMP6s under the control of Thy1 promoter were used for imaging somas of L2/3 PNs. In all other experiments GCaMP6 was expressed with recombinant adeno-associated viruses (AAVs). In experiments in which we imaged dendrites/spines of L2/3 PNs, GCaMP6s was expressed with recombinant AAV under the human sysnapsin-1 promoter (AAV-syn-GCaMP6s; serotype 2/1; >2×10¹³ (GC/ml) titer; produced by the University of Pennsylvania Gene Therapy Program Vector Core). 0.2 μl of AAV virus was diluted 5X in ACSF (or in conjuncture with other viruses) and injected (Picospritzer III; 20 p.s.i., 20 ms, 0.2 Hz) into L2/3 (depth of 200-400 μm) of the primary motor cortex using a glass microelectrode around the coordinate of 0.3 cm anterior and 1.5 cm lateral to bregma (Tennant et al., 2011). The glass microelectrode was advanced into the brain using a micromanipulator (M3301 World Precision Instruments), at a 60 degree angle relative to the skull surface. The virus dilution allowed better spread through L2/3 and sparse neuronal labeling. We examined the distribution of viral infection across layers using histology in some of the animals. GCaMP infected cells were distributed across cortical layers with 4.7±1.7% of cells located within 0-100μm, 30.9±6.8% of cells located within 100-200μm, 30.3±4.8% of cells located within 200-300μm, 17.9±4.8% of cells located within 300-400μm, 8.4±5.2% located within 400-500μm, 4.6±3.1% located within 500-600μm and 2.8±2.7% located >600μm (numbers are averages ± s.e.m for 7 animals). Thus, the vast majority of infected cells were in layers 2/3. Proper targeting of L2/3 neurons by AAV infection was ascertained via Zstack imaging across L1-5 at the end of the experiments. Dendrites of PNs in L1 were ascertained via spine morphology.

In experiments in which we imaged somas of L2/3 SST-expressing or VIP-expressing INs, a Cre-dependent AAV-GCaMP6m (AAV-CAG-Flex-GCaMP6m or AAV-syn-Flex-GCaMP6m; serotype 2/1; >2×10¹³ (GC/ml) titer) was injected into SST-IRES-Cre or VIP-IRES-Cre mice which expressed Cre recombinase exclusively in SST-expressing or VIP-expressing INs, respectively. Mice were injected at post-natal day 21 and imaged 14 days later.

In vivo two-photon Ca²⁺ imaging of motor cortex was done at the depth of 200-350 μm below the pial surface for detecting L2/3 somas of PNs and SST-expressing or VIP-expressing INs and at the depth of 20-70 μm below the pial surface for detecting apical dendrites and spines of L2/3 PNs. In each animal, time-lapse imaging was performed at one focal plane for PNs and at one to three focal planes for SST/VIP INs. Imaging was performed with an Olympus Fluoview 1000 two-photon system (tuned to 920 nm) equipped with a Ti:Sapphire laser (MaiTai DeepSee, Spectra Physics). The average laser power on the tissue sample was ~5–15 and 20-30 mW for imaging in L1 and L2/3 of the cortex respectively. All experiments were performed using a 25X objective (N.A. 1.1) immersed in ACSF solution and with a 2X (soma) and 5X (dendrites) digital zoom. All images were acquired at frame rates of 2 Hz (2-μs pixel dwell time). The duration of imaging and number of repetitions were kept to a minimum that enabled us to perform the experiments from one hand, but did not cause photo damage on the other (Hopt and Neher, 2001). Typical imaging window for Ca²⁺ imaging of neuronal somas was ~322 μm by 161 μm and of tuft dendrites was ~129 μm by 65 μm. Image acquisition was performed using FV10-ASW v.2.0 software and analyzed post hoc using NIH ImageJ and Matlab (Mathworks) software.

Manipulating SST neuronal activity.

Two weeks prior to Ca²⁺ imaging, Cre-dependent DREADD-hM₄D(Gi) or DREADD-hM₃D(Gq) virus (AAV-hsyn-DIO-hM₄D(Gi)mCherry or AAV-hsyn-DIO-hM₃D(Gq)mCherry; serotype 2; >10¹² (GC/ml) titer; produced by University of North Carolina Vector Core) was injected into SST-IRES-Cre mice or SST-IRES-Cre crossed with Thy1-GCaMP6s mice which express Cre exclusively in SST INs. 0.2 μl of AAV virus was diluted 3X in ACSF (or in conjuncture with other viruses) and injected into layer 2/3 (depth of 200-400 μm) of the primary motor cortex as described above. Activation of the DREADD system was done by an intraperitoneal (i.p) injection of Clozapine N-oxide (CNO, C0832, Sigma Aldrich). CNO was dissolved in saline to a concentration of 0.5 mg/ml. Imaging and behavioral experiments were conducted prior to and 20 minutes following CNO administration (0.3 ml/30 g body weight) to allow activation of the DREADD system. DREADD infection levels in SST INs were determined by counting the number of SST-mCherry positive somas (Fig. 3A, B).

Since the mCherry tag of the virus produced weak fluorescence expression, we performed immunostaining of the brain slices for detecting mCherry fluorescence. SST-IRES-Cre animals injected with Cre-dependent DREADD-hM₄D(Gi) were anesthetized and perfused with 20 Ml phosphate buffered saline (PBS) two weeks post injection. Brain tissue was removed and fixed for 1 hour in 4% paraformaldehyde at 4°C. Tissue was rinsed three times with PBS, embedded in 2% agarose, and sectioned at 200 μm with a vibratome. Sections were permeabilized in 1% Triton X-100 in PBS for 3 hours and blocked with 5% normal goat serum for 1 hour. Sections were incubated overnight with primary antibody against RFP (Rockland Immunochemicals, 1:750 #600401379 (Gross et al., 2000)). Sections were then washed three times with PBS/0.05% Tween-20, and then incubated with Alexa Fluor-conjugated goat antirabbit IgG secondary antibody (Life Technologies, 1:500) in PBS for 2 hours. Sections were washed as before and mounted for imaging. Confocal images were obtained on a Zeiss LSM 700 confocal microscope (10X air objective; numerical aperture, 0.8). Somas were confined to M1 (based on the mouse brain atlas) and were counted manually post hoc using ImageJ software.

DREADD manipulation efficacy in SST INs was determined by Ca²⁺ imaging of SST somas (Fig. 3C-I and Fig. S4A-H). In this experiment, two viruses (AAV2/1-syn-Flex-GCaMP6m and AAV2-hsyn-DIO-hM₄D(Gi)mcherry or AAV2-hsyn-DIO-hM₃D(Gq)mcherry) were mixed at equal volumes and injected into the primary motor cortex of SST-IRES-Cre mice. We imaged the same cells before CNO administration, 20-minute and 40-minute following CNO administration. To characterize the activity of SST INs, we examined the level of Ca²⁺ activity during quiet awake (“resting”) and running periods. We also examined the effect of the DREADD manipulation in SST INs on the activity of PNs (Fig. S4L-U). In this experiment we injected AAV2-hsyn-DIO-hM₄D(Gi)mcherry or AAV2-hsyn-DIO-hM₃D(Gq)mcherry in SST-IRES-Cre mice crossed with Thy1-GCaMP6s mice. We imaged the activity of the same PNs before CNO administration, 20-minute and 40-minute following CNO administration. In addition to examining the level of Ca²⁺ activity as described above for SST INs, we also examined the time of the peak activities.

To control for the acute effect of CNO administration, we examined the activities of SST INs and PNs as described above in SST-IRES-Cre mice not injected with DREADD before and 20-minute after CNO administration (Fig. S4I-K for SST INs and Fig. S4V-Y for PNs). We also compared the animals’ behavior before and after CNO administration in SST-IRES-Cre naïve mice or SST-IRES-Cre mice injected with AAV2-hsyn-DIO-hM₄D(Gi)mcherry or AAV2-hsyn-DIO-hM₃D(Gq)mcherry (Fig. S5).

Manipulating VIP neuronal activity.

Two weeks prior to Ca²⁺ imaging and behavioral testing, Cre-dependent PSAM (Pharmacogenetically Selective Actuator Module) (Magnus et al., 2011) virus (AAV2-syn-CAG-Flex-PSAM-2A-EGFP, Produced by University of Pennsylvania Vector Core) was injected into VIP-IRES-Cre mice which express Cre exclusively in VIP-expressing cells. 0.2 μl of AAV virus was injected into layer 2/3 of M1 as described above. Activation of the PSAM was done by local application of PSEM³⁰⁸ (Pharmacogenetically Selective Effector Module) (150 μM dissolved in ACSF) to the superficial layer of M1 through the cranial window. PSEM³⁰⁸ delivery was made by removing the glass cover slip covering the cranial window. PSEM³⁰⁸ was allowed to diffuse for 10 minutes, and then washed with ACSF. The glass cover slip was glued back to the skull followed by imaging or behavioral testing. Imaging and behavioral testing were conducted prior to and 10 minutes following PSEM³⁰⁸ application to allow activation of the PSAM system. PSAM infection levels in VIP neurons were determined by counting the number of VIP-GFP positive somas (Fig. 5E, F). For better detection of fluorescence expression of EGFP, we performed immunostaining of the brain slices with the primary antibody against GFP (Abcam 1:500) and goat anti-rabbit IgG secondary antibody (Invitrogen 1:500).

PSAM manipulation efficacy in VIP neurons was determined by Ca²⁺ imaging of VIP somas (Fig. 5G, H). In this experiment, two viruses (AAV1-syn-flex-NES-JRGECO1a and AAV2-syn-CAG-Flex-PSAM-2A-EGFP) were mixed at equal volumes and injected into the M1 of VIP-IRES-Cre mice. We imaged the same cells before PSEM³⁰⁸ administration and 10-minute following PSEM³⁰⁸ application. Imaging was performed with an Olympus Fluoview 1000 two-photon system tuned to 1020 nm, equipped with a Ti:Sapphire laser (MaiTai DeepSee, Spectra Physics). To characterize the activity of VIP INs, we examined the level of Ca² activity during running. We also examined the effect of the PSAM manipulation in VIP INs on the activity of PNs (Fig. 5I, J). In this experiment we injected AAV2-syn-CAG-Flex-PSAM-2A-EGFP in VIP-IRES-Cre mice crossed with Thy1-GCaMP6s mice. We imaged the activity of the same PNs before PSEM³⁰⁸ administration and 10-minute following PSEM³⁰⁸ application. To characterize the cell activities, we examined the level of Ca² activity during running and the time of the Ca²⁺ peak activities. Finally, to determine the acute effect of PSEM³⁰⁸ administration we also compared the animals’ behavior before and after PSEM³⁰⁸ administration in VIP-IRES-Cre mice injected with AAV2-syn-CAG-Flex-PSAM-2A-EGFP (Fig. 5K).

Optogenetic CaMKII inhibitor.

3-4 weeks prior to Ca²⁺ imaging, a light-inducible CaMKII inhibitor (AAV9-CaMKIIp-mEGFP-P2A-paAIP2, produced by Max Planck Florida Institute for Neuroscience) or non-functional mutant control (AAV9-CaMKIIp-mEGFP-P2A-paAIP2(R5A/R6A)) virus (Murakoshi et al., 2017) was injected into SST-IRES-Cre mice. The virus was injected into layer 2/3 of M1 as described above in conjecture with red genetically encoded Ca²⁺ indicator RGECO (AAV1-syn-NES-JRGECO1a; produced by the University of Pennsylvania Gene Therapy Program Vector Core). For the experiments described in figure 8(J-Q) the animals were also injected with Cre-dependent DREADD-hM₄D(Gi) (AAV-hsyn-DIO-hM₄D(Gi)mCherry) with a separate second injection into the same location. Imaging was performed with an Olympus Fluoview 1000 two-photon system tuned to 1020 nm, equipped with a Ti:Sapphire laser (MaiTai DeepSee, Spectra Physics). PaAIP2 inhibits CaMKII activity in a blue-light-dependent manner. Blue light illumination was done via the arc-lamp or LED laser (Thor Labs, LED 1B, M470F3, 470nm, 20mW) directly on top of the open skull imaging region only during treadmill training and not during neuronal imaging or behavioral testing. During forward/BR training session, blue light was illuminated for 10s every 1 minute over the course of 20 minutes. Neuronal activity and behavioral performance were tested before and after the 20-minute training session (without blue light illumination).

Drug application.

CaMKII inhibitors KN-62 (I2142, Sigma Aldrich), and KN-93 (K1385, Sigma Aldrich), and control KN-92 (sc-311369, Santa Cruz Biotehcnology) were applied to the cortex at the concentration of 10 μM in ACSF. They were first dissolved in DMSO and then diluted in ACSF with a final DMSO concentration of <0.1%. CaMKII inhibitors and control were applied to the superficial layer of the M1 through the cranial window. Drug delivery was made as described above for PSEM³⁰⁸. Drugs were allowed to diffuse for 10 minutes, and then washed with ACSF. We estimated the extent and time course of CaMKII inhibitors and control diffusion by applying Rhodamine B isothiocyanate-Dextran (Sigma, Lot #SLBS1075V, 1mg in 300 Ml ACSF), and imaging the spread across cortical layers over 20 minutes after drug application. Rhodamine B isothiocyanate-Dextran spread was estimated according to the fluorescence spread normalized to its peak level (brain surface). The fluorescence dropped to 50% at 235±16 μm and remained unchanged following 20 minutes (Wilcoxon rank sum, U=−1.68 P=0.09, 3 animals). We examined the acute effects of the CaMKII inhibitors and control on PNs’ activity and animals’ behavior by comparing the cells’ activity and animals’ behavior before and following drug application (Fig. S7), as described above for CNO.

Quantification and Statistical Analysis

Image analysis.

During running trials, the lateral movement of the images was typically less than 1 μm. Vertical movements were infrequent and minimized due to flexible belt design, two micro-metal bars attached to the animal’s skull (described above) by dental acrylic, and a custom-built body support to minimize spinal cord movements generated by the hind limbs. All time-lapse images from each individual field of view were motion corrected and referenced to a single template frame using cross-correlation image alignment (TurboReg (Thevenaz et al., 1998) plugin for ImageJ). Regions-of-interests (ROIs) corresponding to visually identifiable somas, spines or apical tuft dendrites were selected manually. Somas, dendrites and spines that could be identified in all imaging sessions were included in the data set. Our data set therefore excludes neurons that did not show any fluorescence transients during any of the imaging sessions. All the pixels inside the ROI were averaged to obtain a time-series fluorescence trace for each ROI. This process was done automatically for somas using custom code in Matlab and manually for spines and dendrites. The manual extraction of fluorescence signal from spines and dendrites was done to overcome contamination of the signal in these small structures by nearby structures due to remaining small lateral movements of the images. Data extraction was done blind to experimental group. Background fluorescence was calculated as the average over the 5^th percentile pixel value per frame and subtracted from the time-series fluorescence traces. The baseline (F₀) of the fluorescence trace was estimated by detecting inactive portions of the trace using an iterative procedure (Peters et al., 2014). Briefly, we smoothed (loess, 1 minute moving average) the raw fluorescence trace and subtracted the smooth trace from the raw trace, denoted as preliminary F₀. Two times the standard deviation of the preliminary F₀ trace was set as a threshold for detecting inactive portions in the raw fluorescence trace. The inactive portions were concatenated and the procedure was repeated once again. The resulting inactive portions were then placed according to their original time points and values were linearly interpolated across the gaps of the active portions yielding the F₀. The ΔF/F₀ was calculated as ΔF/F₀ = (F–F₀)/F₀ × 100%.

Data analysis.

To determine the running related responses of PNs, we assessed the neurons’ activities simultaneously with animal’s running gait patterns within single trials (Fig. S8A-I). Specifically, we examined whether and how the neuronal responses (quantified by average and maximum levels of Ca²⁺ activity) corresponded to specific running gait patterns. For each trial, we measured the average and peak activity of each neuron separately for each of four different running gait patterns before and after 20-minute FR training session. Animals’ behavior during imaging sessions was assessed using video footage. Animals’ forelimbs were videoed with a YI home camera (1080p full HD wireless infrared camera, 8Hz, model: YHS.2116.INT) and aligned with Ca²⁺ imaging frames. Offline analysis of animal’s gait patterns from the videos was done manually frame by frame. Additionally, we determined the running related responses of the neurons by comparing their activities before and after turning the treadmill (pre-movement, stationary vs. movement conditions). Ca²⁺ response profiles of individual neuronal somas to running were characterized by averaging the fluorescence traces (ΔF/F₀) across at least four running trials per condition (pre-training and post-training) aligned to turning the treadmill on (start of running). The baseline activity of each soma was calculated by averaging the activity during the ITI (15 seconds times at least 4 trials, denoted baseline activity), and was subtracted from the average response profile of each neuron. We determined if the average response profiles of individual neurons were significantly modulated to running by identifying time segments in the average response profiles following treadmill on time point that deviated from the baseline activity (above or below) by at least three times the standard deviation (SD) of the baseline activity. A neuron was considered to have a significant response to running if the duration of the deviant segment was at least 3 frames (6 seconds).

Average neuronal response profiles aligned to treadmill on (start of running) were maximum-normalized by subtracting the maximum value of the average response profile and dividing by the difference between the maximum and minimum values. Neurons were then ordered according to the time of their maximum response to create the population sequence. In order to determine that the population temporal sequence was not simply due to ordering the normalized response profiles, we compared the population activity data to a shuffled version of the data (Fig. S2L-P) similar to previous studies (Harvey et al., 2012). The shuffled population activity data was created by rotating the response profiles (ΔF/F₀ traces before averaging) of individual trials for each imaged soma by a random amount relative to the time when the treadmill was turned on. We then averaged the shuffled response profiles of individual cells and maximum-normalized the shuffled data. We compared the temporal profile of the original and shuffled sequences (the cumulative sum of the time of the peak activity of the neurons) and the distributions of ridge to background ratio of the original and shuffled data (Fig. S2O, P). Ridge to background was calculated as the ratio between the average ΔF/F₀ response profile in the peak value ± 3 surrounding time bins to the average ΔF/F₀ of all other time points. We shuffled and compared the data 500 times. Calculating the ridge to background ratios using ± 2 and ± 10 time bins yielded similar results (i.e. shuffled data was significantly different from original data, bootstrap, P<0.0001).

We examined whether the neurons reached their peak activity relative to running onset at the same order during different learning phases, i.e. the stability of the ranking of the neurons in the population sequence (Fig. 1P). To that end we calculated the correlation coefficient between the rank of the neurons according to the maximum normalized activity in one running session and the rank of the same neurons in another session (following 20 minutes running training period). To determine if the correlation was significant, indicating the stability in the order in which cells reached their maximum activity, we compared the correlation coefficient calculated on the original data to that calculated on the shuffled data. In the shuffled data set, we kept the original rank of the first session but randomly ranked the neurons in the second session. Significance was determined if the correlation calculated on the original unshuffled data was higher than >95% of the correlation calculated on the shuffled data.

The selectivity of SST INs’ response profiles to the running tasks was determined by the percentage of SST INs that had significantly different response profiles to forward and BR (Fig. 2). For each SST IN we calculated the absolute average difference between its average response profiles to forward and BR, named “difference measure”. We then pooled the response profiles of all individual trials (of both forward and BR) and randomly assigned the response profiles as originating from either a forward or a BR trial. We then averaged over the randomly assigned trials to create the new shuffled average response profiles and calculated the shuffled difference measure. Assuming no difference between the average response profiles to forward and BR (null hypothesis), randomly assigning the individual response profiles to forward or BR should yield similar average response profiles and difference measure as the actual data. Finally, we repeated the process 500 times for each SST IN and determined significance if the difference measure for the unshuffled data was higher than > 95% of the shuffled difference measure. The polarity of SST INs’ response profiles to the running tasks was determined by the percentage of SST INs with significantly different responses to forward and BR that had an opposite direction of response to the two running tasks (determined according to the above description).

For the measurement of spine Ca²⁺ transients in Fig. 6, we detected dendritic spine Ca²⁺ transients in response to running and removed contributions to the signal due to back-propagating action potentials based on the dendritic shaft Ca²⁺ activity (Fig. 6A-C) (Chen et al., 2013). Specifically, we extracted the fluorescence for each spine of its parent dendritic shaft and calculated the ΔF/F₀ denoted ΔF/F_{0_dendrite} We then estimated the coefficients of the robust linear regression of ΔF/F₀ calculated for the spine vs. ΔF/F_{0_dendrite}. We multiplied ΔF/F_{0_dendrite} by the slope of the fitted regression line and subtracted this scaled version of ΔF/F_{0_dendrite} from ΔF/F₀ calculated for the spine to obtain a back-propagation independent spine Ca²⁺ signal. We found that before the subtraction of back-propagation signals, 48% of spines displayed Ca²⁺ activity that was correlated with dendritic Ca²⁺ activity, whereas after back-propagation subtraction only 10% of the spines showed significant correlation. In the analysis presented in figure 6, we excluded spines whose Ca²⁺ activity were still significantly correlated with dendritic Ca²⁺ activity after back-propagation subtraction. We detected spine transients by identifying time segments following treadmill on time point that deviated from the baseline activity by at least three times the standard deviation (SD) of the baseline activity (as described above for soma activity). To compare spine activity between imaging sessions (e.g. before and after 20-minute FR training), we averaged the peak value of all transients in each spine per training session.

The selectivity of pyramidal dendritic Ca²⁺ transients to the running tasks (Fig. S6K) was determined by subtracting the number of Ca²⁺ transients per minute on individual dendrite in response to BR from that to FR and divided by their sum.

Data is presented as average ± s.e.m unless otherwise noted. Sample sizes were chosen to ensure adequate power with the statistical tests while minimizing the number of animals used in compliance with ethical guidelines. We tested the data for normality using the Shapiro-Wilk test and performed a-parametric or parametric statistical tests if normality was or was not rejected respectively. We used Wilcoxon rank sum test (or t-test) to compare two groups and Kruskal-Wallis (or one-way ANOVA) to compare more than two groups. Kruskal-Wallis (or One-way ANOVA) tests were followed by Tukey-Kramer test for multiple comparisons. We used Kolmogorov-Smirnov test to compare the cumulative sum of the peak activity times and corrected with Bonferroni correction when more than one comparison was made. All tests were conducted as two-sided tests. Significance level was determined at 5%. All statistical details of the experiments can be found in the figure legends and in Table S1.

Data and Software Availability

Code availability.

All of the analysis described above was performed using toolboxes and custom code in Matlab. All of the custom codes used for analysis are available from the authors upon request.

Data availability.

All of the data described in this manuscript is available from the authors upon request.

Supplementary Material

1

Video S1: Example of animal’s running behavior on the treadmill, Related to Figure 1

A video image of a single animal running on the treadmill. The video displays 4 running trials (2 pre-training and 2 post-training). Each trial starts with the treadmill turned off and the animal is stationary, followed by turning the treadmill on and the animal is forced to run forward.

2

Video S2: Example of SST IN imaging, Related to Figure 2

Ca²⁺ imaging of SST INs taken from 2 animals. Imaging was taken during pre-training running. Activity is shown during baseline (treadmill is turned off and the animal is stationary) and during forced running (treadmill is turned on). Scale bar depicts 20μm.

4

Video S3: Example of spine imaging, Related to Figure 6

Ca²⁺ imaging of 2 apical tuft dendrites of L2/3 pyramidal neurons. Imaging was taken during pre-training running. Video displays the original images taken without any post-hoc processing (e.g. turbureg). For each dendrite activity is shown during baseline (treadmill is turned off and the animal is stationary) and during forced running (treadmill is turned on). For each dendrite the activities of the shaft and a few of the spines are depicted at the end of the video. Scale bar depicts 10μm.

Highlights:

• Sequential activity of pyramidal neurons emerges and stabilizes with motor training

• SST interneurons regulate the establishment and stabilization of sequential activity

• VIP interneurons regulate the establishment of sequential activity

• The regulation of sequential activity involves CaMKII-dependent synaptic plasticity