Striatal Neurons Are Recruited Dynamically into Collective Representations of Self-Initiated and Learned Actions in Freely Moving Mice

Self-initiated movement in freely moving mice strongly modulates the activity of HaR striatal neurons

To examine the response reliability of striatal neurons (presumably to cortico- and thalamostriatal inputs) during naturalistic behaviors, we used GCaMP mice that serendipitously expressed the genetically encoded Ca²⁺ indicator (GECI) GCaMP6f in a sparse population of striatal neurons consisting of a majority of ∼60% SPNs and ∼30% PV interneurons (see Materials and Methods, Discussion, Fig. 1-1)—two striatal populations that are HaR (Koós and Tepper, 2002; Hammond, 2015; Tepper and Koós, 2016) and typically driven to spike by few large synaptic inputs (Johansson and Silberberg, 2020). Seven mice were implanted with a GRIN lens and mounted with microendoscopes to simultaneously visualize multiple sparsely distributed neurons in the dorsal striatum of freely moving mice during innate behavior (Fig. 1A–C). The kinematics of the mice were monitored using video cameras, an accelerometer, and tracking software that were synchronized with the Ca²⁺ imaging, enabling us to compare the collective neuronal activity in different behavioral states.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

The collective neuronal activity is strongly modulated by self-initiated movement in freely moving mice. A, A 1-mm-diameter GRIN lens is implanted into the dorsolateral striatum. B, Implanted mouse with a microendoscope and an accelerometer mounted on its head moves freely in a behavior chamber. C, Image via lens in freely moving mouse reveals signals from dozens of striatal neurons (see Fig. 1-1 for distribution of striatal subtypes). D, Ca²⁺ signals of two individual neurons (top, green), the average Ca²⁺ signal across all imaged neurons (56 neurons, middle, green) and the simultaneously recorded total body acceleration of a representative mouse (bottom, black). The two individual neuron exhibited were chosen because they were both shown to respond significantly to changes in the acceleration: with the top one reducing its Ca²⁺ signal and bottom one increasing its Ca²⁺ signal. E, Color-coded matrix showing activity around movement onset (Fig. 1-2), averaged across all movement onset events (top). Average Ca²⁺ activity across the population of imaged neurons and average total body acceleration across movement onset events are represented by green and black traces, respectively (middle). PSTH of Ca²⁺ events centered around the movement onset (bottom). Bin width: 67 ms. See Fig. 1-3 for description of positively and negatively modulated neurons. F, Same as E but around movement offset. G, Rates of detected neurons that did (green) or did not (white) significantly modulate their Ca²⁺ signals following movement onset (Fig. 1-2). Shaded areas represent S.E.M. g in scale bar is the gravitational constant 9.81 m/s².

We began by examining the activity of SPNs and PV interneurons around the onset of self-initiated movement. Six-hundred thirty-seven putative somata were detected in seven mice [two imaging sessions per mouse, 45.5 ± 33.4 (mean ± SD) neurons per session]. Movement onset and offset times were detected by setting a threshold on the total body acceleration of the mouse [see Materials and Methods, Fig. 1-2A,B, 171 ± 91 (mean ± SD) movements per session]. For each neuron, we calculated the event triggered average (ETA) by averaging the Ca²⁺ signal in a 5 s time window around movement onset across all movement events. We found that the average ETA across the population of imaged neurons increased following movement initiation (Fig. 1D–F). The Ca²⁺ signal around the onset of self-initiated movement followed the total body acceleration and peaked approximately 130 ms after the acceleration peak. The PSTH centered around the same events of movement onset exhibited a similar peak, indicating that the population of imaged neurons produced more Ca²⁺ events following movement onset (Fig. 1E). Repeating the analysis around movement offset times generated a mirror image of the movement onset responses (Fig. 1F), suggesting that neurons modulate their activity when movement is initiated and that, at the population level, the modulation is maintained throughout the movement. To determine whether the activity of an individual neuron was significantly modulated around movement onset, we devised a bootstrapping-based method (see Materials and Methods, Fig. 1-2C–E). Twenty seven percent of detected neurons were found to be significantly responsive around the onset of self-initiated movement (Fig. 1G). Thus, while movement responsiveness is evident at the population level, only a subset of the imaged neurons consistently modulate their Ca²⁺ signals around movement onset in a given imaging session.

A closer examination of the ETAs of individual significantly responsive neurons around the onset of self-initiated movement revealed that they exhibit two types of responses (Fig. 1E,F and Fig. 1-3): while a majority of responsive neurons (78%) increase their rate of Ca²⁺ transients as the total acceleration of the mouse increases (positively modulated, top part of matrix in Fig. 1E- and Fig. 1-3A), some significantly responsive neurons (22%) produce less Ca²⁺ transients when the mouse is moving (negatively modulated, bottom part of matrix in Fig. 1E and Fig. 1-3B). Additionally, movement responses of the imaged neurons were lateralized, displaying a strong preference to contralateral turns over ipsilateral ones (Fig. 4A,C,D). These properties are consistent with known activity patterns of SPNs and PV interneurons during spontaneous movement (Albin et al., 1989; Alexander and Crutcher, 1990; DeLong, 1990; Gerfen et al., 1990; Mink, 1996; Hikosaka et al., 2000; Chan et al., 2006; Brown, 2007; Nambu, 2008; Kravitz et al., 2010; Gittis et al., 2011; Cui et al., 2013; Isomura et al., 2013; Jin et al., 2014; Tecuapetla et al., 2014; Perk et al., 2015; Barbera et al., 2016; Klaus et al., 2017; Parker et al., 2018; Gritton et al., 2019).

Portion of HaR striatal neurons responsive to task-related cues increases systematically as training progresses

To evaluate the response reliability of SPNs and PV interneurons during learning, we trained six mice in an operant conditioning paradigm to associate an auditory CS with a reward. Each training session consisted of 42 trials. In each trial, an auditory cue (CS+ or CS−) was presented for 10 s in a pseudo-random manner. To receive a sucrose–water mixture reward, the freely moving mouse was required to enter its head into the reward delivery port while the CS+ was being presented. A head entry during CS− presentation was not rewarded (Fig. 2A).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Responses to task-related cues develop with training. A, Mice were trained in an operant conditioning paradigm to associate an auditory CS with a sucrose–water mixture reward. Each session consisted of 42 trails. In each trial, a cue (CS+ or CS−) was presented for 10 s in a pseudo-random manner. If the mouse entered its head into a designated port during the 10 s time window following CS+ (but not CS−) onset, reward was available for 4 s. B, The learning rate quantified as the portion of cue presentations during which the mouse entered its head into the port (% hits). Blue rectangles mark imaging sessions and the dashed box marks extinction sessions. C, Color-coded matrix showing neuronal activity around CS+ presentation averaged across all CS+ presentation events for naive (left), intermediate level (middle) and expert (right) mice. Average Ca²⁺ activity across the population of imaged neurons and average total body acceleration across CS+ presentation events are represented by pink and black traces, respectively (bottom). See Fig. 2-1 for the development of the total body acceleration profile around CS+ presentation as training progresses. D, Same as C, for CS− presentation. Average Ca²⁺ activity across the population is represented by gray traces. E, Rate of responsive neurons (left) and “hit” rate (right, same data as in panel B) around CS+ (pink) and CS− (gray) presentation on various stages of training. Shaded areas represent the S.E.M. g in scale bar is the gravitational constant 9.81 m/s². See Fig. 2-2 for the relationship between the neural responses around CS+ presentation and total body acceleration or reward delivery.

Mice successfully acquired the association between the CS+ and the sucrose–water mixture reward (Fig. 2B). To quantify learning, we measured for each cue (CS+ and CS−) the portion of cue presentations during which the mouse entered its head into the port attempting to obtain the reward (“hits”). On advanced training days, mice perform head entries upon almost every CS+ presentation, while the rate of CS− “hits” decreases to ∼15%, indicating that the CS+ was associated with the sucrose–water mixture reward while the CS− was not. Once the association was established, the mice underwent three extinction sessions. Those sessions were identical to conditioning sessions, but no reward was delivered. The association between the CS+ and the reward was quickly extinguished, and mice ceased to perform head entries (Fig. 2B).

To characterize the neural activity in various stages of learning, we compared ETAs in a 10 s time window around cue presentation in naive, intermediate level and expert mice (see Materials and Methods). We considered one session per mouse at each learning stage and detected a total of 355, 342 and 373 putative somata on naive, intermediate, and expert training sessions, respectively. In naive mice, the average Ca²⁺ response following CS+ presentation was weak. The population response developed as training progressed and was strong on advanced conditioning sessions when mice were experts at the task (Fig. 2C). In contrast, average CS− responses remained weak throughout training (Fig. 2D). This was evident in the average ETAs representing the population signal, as well as in the ETAs of individual neurons.

The mice changed their behavior as training progressed. Specifically, as the association between the CS+ and the reward strengthened, mice were quicker to move towards the reward port when the CS+ was presented, resulting in more stereotypical acceleration profiles across CS+ presentations (Fig. 2-1), as well as a larger average acceleration change around CS+ onset (Fig. 2C, compare the amplitude of the black trace as sessions progress from naive to expert). This raises the concern that the CS+ responses that we observed are in fact responses to the initiation of the accompanying movement. To address this issue, we divided CS+ presentations on the final conditioning session into two groups according to the size of the associated change in total body acceleration (see Materials and Methods). The average acceleration profiles were very different for the large (4th quartile) and small (1st quartile) acceleration change event groups, and even exhibited opposite change directions (Fig. 2-2A). However, the shape and amplitude of the Ca²⁺ responses were very similar in the two cases. In contrast, when we repeated the analysis for movement initiations on free movement sessions, the average Ca²⁺ trace followed the average acceleration trace, with a substantially smaller average Ca²⁺ amplitude for small acceleration change movements compared to large acceleration change movements (Fig. 2-2B). Thus, while neuronal responses following CS+ presentation may be affected by the mouse’s movement, these results indicate that they are not determined by the total body acceleration. Finally, CS+ responses continue to appear and increase with training on unrewarded trials and are therefore not dependent on reward delivery (Fig. 2-2C).

The event rates of neurons significantly responsive to CS+ or CS− tightly corresponded to task performance (Fig. 2E). In naive mice, very few neurons responded around either cue. In intermediate level mice, towards the middle of training, there was an increase in the portion of significantly responsive neurons for both CS+ and CS− presentations. On advanced conditioning sessions, when mice were experts at the task, a larger portion of detected neurons significantly modulated their activity following CS+ presentation, while very few responded to CS− presentation. This profile echoed the progression of learning as represented by the rates of CS+ and CS− “hits”, suggesting that the activity of these neurons is meaningful in the context of the task. Notably, while both the average population signal and rate of significantly responsive neurons increase systematically as training progresses, even in expert mice, only a subset (∼30%) of detected neurons significantly modulated their activity in response to CS+ presentation.

Similarly to movement-responsive neurons, neurons that significantly modulated their signals around task-related cues could be divided into two functional groups—positively and negatively modulated neurons (Fig. 2C,D and Fig. 1-3F). Here too, the majority of responsive neurons increased the rate of their Ca²⁺ transients following cue presentation, while the rest exhibited less Ca²⁺ events. Thus, subsets of SPNs and PV interneurons modulate their activity around self-initiated movements as well as task-related cues in an operant conditioning paradigm.

Individual HaR striatal neurons are recruited to respond to the same behavioral events in a nonreliable fashion across and within imaging sessions

We next examined the reliability of neuronal responses around various behavioral events. Will a neuron that is significantly responsive around a specific behavior in a given imaging session continue to code for it on subsequent sessions? Our experimental approach allowed us to track individual neurons across multiple imaging sessions (Fig. 3A, see Materials and Methods). To examine the stability of movement responses, we detected neurons that were active on two free movement sessions. Out of 88 such neurons, 22% significantly responded around movement onset on the first session, 27% responded on the second session, and only 9% significantly modulated their signals around self-initiated movement on both free movement sessions (Fig. 3D). While the 9% overlap is greater than the expected rate if neurons are randomly recruited to respond around movement onset (e.g., 6%), the identity of significantly responsive neurons appears to be highly dynamic across imaging sessions.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Responses around movement onset and cue presentation are dynamic across imaging sessions. A, Four example cells that were detected on 3 imaging sessions. The images on the bottom represent the footprints produced by CNMF-E of one example cell across the various sessions. B, Average Ca²⁺ signal of an example neuron on the first (top) and second (bottom) free movement imaging sessions that exhibited significant responses (according to our bootstrapping criterion, see Materials and Methods) around movement onset on both sessions. C, Same as A, around CS+ presentation. D, Percentage of neurons significantly responsive on the first (light green) or second (medium green) free movement imaging sessions, both of them (dark green) or neither (white). See Table 3-2 for the time between the two free movement sessions for each mouse. E, Percentage of neurons from mouse 1 (left) and mouse 2 (right) significantly responsive on the first (light green) or second (medium green) free movement imaging sessions, both of them (dark green) or neither (white), that were at most 3 d apart. F, Percent of neurons significantly responsive on one (light), two (medium) or three (dark) advanced conditioning sessions around CS+ (left, pink) and CS− (right, gray) presentation. See Fig. 3-1 for analysis of the reliability around CS+ presentation within a single session. G, PDF of the percent of movement onset (green) and CS+ presentation (pink) events after which the neuron produced a Ca²⁺ event for positively modulated neurons. H, PDF of the percent of neurons that responded with a significant Ca²⁺ transient following each movement onset (green) of CS+ presentation (pink). I, Examples of neurons exhibiting significant and nonsignificant responses around movement onset and CS+ presentation. Cell 1 responded significantly only to CS+ presentation. Cell 2 responded significantly only to movement onset. Cell 3 responded significantly to both. J, Percentage of neurons significantly responsive around movement onset (green), CS+ presentation (pink), both (dark purple) or neither (white) out of the neurons that were detected both in a free movement session and in an advanced conditioning session.

The two free movement sessions in each mouse were between 1 and 29 days apart (see Table 3-2). To test if a large interim could account for the high variability in the populations of responsive neurons, we focused on two mice whose free movement sessions were 1 and 3 days apart (mice 5 and 6 in Table 3-2, Fig. 3E). This analysis revealed very similar percent overlap between the populations of neurons significantly responsive on the two sessions, suggesting that the low reliability is not due to changes occurring slowly over time.

To assess the reliability of the responses around task-related cues, we examined the signals of neurons that were detected on three advanced training sessions (2–4 d apart) around the presentation of the CS+ and CS– (Fig. 3F). Of the 233 detected neurons, 32% significantly responded to CS+ presentation on a single session, 9% were responsive on two sessions, and only one neuron significantly modulated its signal following CS+ presentation on all three sessions. Similarly, 17% of the neurons were significantly responsive around CS− presentation on a single session, 3% were responsive on two sessions, and one neuron was responsive on all three sessions. These percentages of overlap are consistent with neurons being randomly and independently recruited to respond around cue presentation on different imaging sessions.

An important caveat is that the movement events that we considered, detected by setting a threshold on the total body acceleration, can include a wide range of movements. Thus, the changes in the identity of significantly responsive neurons between the first and second free movement sessions could be due to differences in the behavior of the mice (Liberti et al., 2022). To address this concern, we analyzed video recordings of mouse behavior to detect grooming initiations and turning events in the ipsi- or contralateral direction (see Materials and Methods, Fig. 4). Repeating the analysis for these more consistent, strictly defined movements, we found again that the rate of neurons significantly responsive around a given movement type on both free movement sessions was most consistent with random recruitment (Fig. 4I).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Responses around ipsi- and contralateral turns and around grooming are dynamic across imaging sessions. A, Color-coded matrix showing neuronal activity around the onset of contra- (left) and ipsilateral (right) turns averaged across all turning events for all mice. Average Ca²⁺ activity across the population of imaged neurons is represented by green and purple traces for contra- and ipsilateral turns, respectively. The average total body acceleration across the relevant turning events are represented by black traces (bottom). B, Same as A, around grooming initiation. C, Percentage of positively (light) and negatively (dark) modulated neurons out of the neurons significantly responsive around contra- (green) and ipsilateral (purple) turns on the two free movement sessions. D, Percentage of neurons significantly responsive around ipsilateral turns only (purple), contralateral turns only (green), both turning direction (gray) or neither (white) on the two free movement sessions. E, PDF of the percent of contra- (green) and ipsilateral (purple) turns and grooming initiation (orange) events after which the neuron produced a Ca²⁺ event for positively modulated neurons. F, Average Ca²⁺ signal of an example neuron on the first (top) and second (bottom) free movement imaging sessions that exhibited a significant response around contralateral turns on the second session but not on the first. G, Same as F, for ipsilateral turns. H, Same as F, for grooming initiation. I, Percentage of neurons significantly responsive around contralateral turns (green), ipsilateral turns (purple) and grooming initiations (orange) on the first (light) or second (medium) free movement session only, both (dark) or neither (white) out of the neurons that were detected on both sessions. g in scale bar is the gravitational constant 9.81 m/s².

This result is further supported by the dynamic identity of CS+ responsive neurons over advanced training days. As mice become experts in the task, the acceleration profile of their behavior around task-related cues becomes much more consistent (Fig. 2-1). Our analysis of CS+ and CS− responses across advanced imaging sessions indicates that while the percentage of significantly responsive neurons increases with training, their identity is dynamic and shifts from session to session (Fig. 3F). Thus, while behavioral variability may still contribute to the shifting identities of significantly responsive cells, our results suggest that it is not the sole cause.

To examine the reliability of the responses within a given session, we considered significantly responsive neurons whose signals were positively modulated around movement onset. Positively modulated neurons are the majority of significantly responsive neurons and tend to generate more Ca²⁺ events following movement onset compared to prior to it (Fig. 1-3). For every such neuron, we calculated the portion of movement initiations after which it produced a Ca²⁺ event. Interestingly, even neurons that were categorized as positively modulated for a given imaging session only produced Ca²⁺ events in an average of 7.2% of movements (Fig. 3G). Repeating the analysis for neurons that were positively modulated around CS+ presentation (Fig. 2D and Fig. 1-3F) produced a higher percent overlap—neurons responded with a Ca²⁺ event for an average of 46.4% of CS+ presentations, with some cells responsive following 100% of cue presentations. The higher reliability of the responses following CS+ presentation compared to movement onset may reflect a difference in the organization of striatal neurons around spontaneous movement and learned behaviors. Alternatively, the disparity may be due to the lower number of cue presentation events (23 cue presentations per session compared to an average of 171 movement initiation events), as well as the more homogeneous behavior associated with these events (Fig. 2-1). Repeating this analysis for grooming, ipsi- and contralateral turns produced very similar results (Fig. 4E), indicating that reliability within a training session is highly dynamic even for more stereotyped behaviors.

Next, for each spontaneous movement initiation and CS+ presentation, we quantified the percentage of HaR neurons that produced a Ca²⁺ transient in a 1.5 s time window following the onset of the behavioral event (Fig. 3H). We found that an average of 2.8 and 16.2% of the imaged neurons were active following movement onset and CS+ presentation, respectively, suggesting that a small portion of the HaR population is sufficient to encode each repetition of the behavior. Furthermore, our data indicate that the specific neurons responsive to each repetition of a behavioral event can vary (Fig. 3D–G), but the percentage of responsive neurons remains relatively consistent (0–22.7% for movement onset and 0–39.3% for CS+ presentation).

Finally, to further test the hypothesis that HaR neurons are dynamically recruited even during the same behavior, we focused on the expert session, where the acceleration profiles of CS+ triggered movements are more consistent (Fig. 2-1). We split the session into two, considering odd and even trials separately, and tested whether the same neurons respond around cue presentation. Sixty-five and 80 neurons were categorized as significantly responsive based on odd and even trials, respectively. Of these, 44 cells were identified as significantly responsive for both trial subsets (Fig. 3-1D). Results were similar when the first or last halves of the trials were analyzed separately, with 92 and 56 neurons significantly responsive based on the first and last half of trials, respectively, and 37 responding significantly in both cases (Fig. 3-1E). Thus, while the identity of neurons recruited to respond around CS+ presentation within a session is more consistent than recruitment across sessions (Fig. 3), even within a single session and with relatively consistent behavior (Fig. 2-1), the identity of responsive striatal HaR cells is not constant.

Our results suggest overall low reliability in the responses of SPNs and PV interneurons, in line with our main hypothesis. We observe this inconsistency in the identity of responsive neurons around both the initiation of spontaneous movement and CS+ presentation, suggesting that neural responses around innate and learned behaviors are equally unreliable, contrary to our predictions.

Individual HaR neurons are recruited to respond to diverse actions in an independent fashion

Imaged neurons significantly modulate their activity following the onset of self-initiated movement and task-related cues (Figs. 1, 2). Do the same neurons respond to the different behavioral events, or are there distinct subpopulations designated to code for each behavior? To address this question, we tracked individual neurons across multiple imaging sessions and examined their responsiveness around movement initiation and CS+ presentation.

One hundred and three neurons were detected on both a free movement session and an advanced conditioning session. Of these neurons, 21% exhibited significant responses around movement onset only, 21% significantly modulated their activity following CS+ presentation only, and 3% were responsive around both movement initiation and cue presentation (Fig. 3I). The percent of neurons significantly modulating their signals around both behavioral events is consistent with neurons being randomly and independently recruited to respond to movement initiation and CS+ presentation.

Taken together, our results indicate that while HaR SPNs and PV interneurons consistently modulate their activity around self-initiated movement and CS+ presentation on the population level, the identity of responsive neurons is not maintained. Rather, neurons are dynamically recruited to code for these different behavioral events on separate occurrences.

Signal correlations between pairs of HaR neurons are independent of distance

Are responsive neurons selected at random with each repetition of the behavior? Or is there a more complex internal organization that our analysis is unable to capture, perhaps due to the vague definition of the behavior we are examining (e.g., movement initiations)? Recent studies have shown that SPNs form spatially compact neural clusters in which nearby neurons are co-active and code for the same type of movement (Barbera et al., 2016; Parker et al., 2018; Shin et al., 2020). Notably, this influential conclusion was drawn from the observation that the correlations between pairs of SPNs decay with distance. Thus, to test whether the unreliability of individual responses in striatal HaR neurons is compensated for by co-activation, we calculated pairwise signal correlations among our imaged neurons and the dependence of signal correlations on the distance between the neurons’ centers.

To test whether nearby neurons tend to exhibit similar activity patterns, we calculated signal correlations between the Ca²⁺ traces of pairs of neurons during rest and movement. When all simultaneously imaged neuronal pairs were included, the average correlation (averaged over 10 µm distance bins and then across bins, see Materials and Methods) is very low during both rest (0.004, Fig. 5A) and movement (0.005, Fig. 5C). When positively and negatively modulated neurons are considered separately, pairwise correlations during both movement states are larger by an order of magnitude (but still weak, 0.03–0.05) and slightly (but not significantly) larger during movement compared to during rest (Fig. 5B,D). Notably, all of these average values are considerably lower than what was previously reported (Barbera et al., 2016; Shin et al., 2020) (see Discussion). Strikingly, for all functional groups and both movement states, pairwise correlations did not depend on the distance between the neurons’ centers (Fig. 5A–D; all cells during rest: mean = 0.004, m = 2.32^.10⁻⁵, R²= 0.05, p = 0.03^a; all cells during movement: mean = 0.005, m = 1.77^.10⁻⁵, R²= 0.05, p = 0.03^b; positively modulated during rest: mean = 0.032, m = 7.6^.10⁻⁶, R² = −0.01, p = 0.62^c; negatively modulated during rest: mean = 0.012, m = 6.27^.10⁻⁵, R²= 0, p = 0.33^d; positively modulated during movement: mean = 0.053, m = 1.88^.10⁻⁵, R²= 0.02, p = 0.15^e; negatively modulated during movement: mean = 0.03, m = 2.89^.10⁻⁵, R² = −0.02, p = 0.5^f; n = 80 distance bins, linear regression).

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

Pairwise correlations are not distance dependent. A, Pairwise correlations as a function of the distance between the neurons’ centers for all pairs of co-imaged neurons during rest. Each point is the average correlation across all neuronal pairs belonging to the relevant distance bin. Dashed line is the linear fit to the data points. B, Same as A, divided into positively (left) and negatively (right) modulated pairs. C, Same as (A), around movement onset. D, Same as (B), during movement. E, Distribution of neuronal pairs for the various distances for all pairs (left), positively (middle) and negatively modulated pairs (right). Bin width is 10 µm. F, Same as (B), around CS+ presentation on intermediate training session. G, Same as (F), divided into positively (left) and negatively (right) modulated pairs. H, Same as (F), on expert training session. I, Same as (G), on expert training session. See Fig. 5-1 for analysis of the correlations in the microendoscopy background neuropil signal and their dependence on distance.

Next, we calculated the signal correlations between the Ca²⁺ traces of neuronal pairs around CS+ presentation. Similarly to movement responses, when all pairs were included, pairwise correlations around cue presentation were very low on both intermediate and expert training sessions (0.001, Fig. 5F,H). When positively and negatively modulated neurons were considered separately, pairwise correlations were again increased by an order of magnitude, but still weak. Importantly, for CS+ responses, at both stages of training and for all functional groups, pairwise correlations did not significantly depend on the distance between the neurons, either. To conclude, we observe weak signal correlations between pairs of our imaged neurons that are difficult to reconcile with the idea of co-activated clusters of nearby neurons. These correlations are also distance independent for both self-initiated movement and CS+ presentation (Fig. 5F–I; all cells intermediate: mean = 0.001, m = 9^.10⁻⁷, R² = −0.01, p = 0.9^g; all cells expert: mean = 0.001, m = −1^.10⁻⁷, R² = −0.01, p = 0.98^h; positively modulated intermediate: mean = 0.065, m = −4.3^.10⁻⁵, R² = −0.01, p = 0.44ⁱ; negatively modulated intermediate: mean = 0.033, m = 2.4^.10⁻⁴, R² = −0.01, p = 0.37^j; positively modulated expert: mean = 0.08, m = 3.3^.10⁻⁶, R² = −0.02, p = 0.95^k; negatively modulated expert: mean = 0.05, m = 5.67^.10⁻⁵, R² = −0.03, p = 0.63^l; n = 80 distance bins, linear regression), in contrast to previous findings of spatially compact SPN clusters (Barbera et al., 2016; Shin et al., 2020). Interestingly, correlations in the background neuropil activity, represented by the Ca²⁺ signals in the annuli surrounding the various somata (see Materials and Methods), did exhibit a clear distance dependence with pairwise correlations decreasing sharply between 0 and 200 µm (Fig. 5-1B, see Discussion).

Trial-by-trial covariation among HaR neurons around movement onset are attributable to variability in behavior

Signal correlations between neuronal pairs reflect the degree to which their average responses are modulated together by the sensory or behavioral event. However, signal correlations do not capture the role of the ongoing moment-by-moment synergy between individual neurons. To investigate the propensity of the imaged neurons to be co-active beyond common increases or decreases in the average rate of Ca²⁺ transients, we calculated joint JPSTHs for the various neuronal pairs (Gerstein and Perkel, 1969; Aertsen et al., 1989; Joshua et al., 2009). The JPSTH is a measure of noise correlations or the degree to which trial-by-trial fluctuations in the response are shared by a pair of neurons. It is derived by subtracting the shift predictor matrix from the product of the PSTHs of the two neurons (Aertsen et al., 1989) (see Materials and Methods), resulting in an estimate of unpredicted correlations.

To assess noise correlations in the responses to self-initiated movement, we calculated the JPSTHs for pairs of neurons around movement onset. The JPSTHs of all neuronal pairs were then normalized and averaged to derive the population JPSTH (Aertsen et al., 1989; Joshua et al., 2009) (see Materials and Methods). The diagonal of the resulting matrix quantifies the average time-varying modulation of zero-lag noise correlation across the population (Fig. 6A). Noise correlation values were similar along the entire diagonal, except for the 1 s time window immediately before movement initiation showing a correlation drop (Fig. 6A). This reduction is likely due to the decreased acceleration and neuronal activity that mark the period leading up to movement onset (Fig. 1E).

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

Behavioral variability more strongly affects trial-by-trial correlations around spontaneous movement than during learned behaviors. A, Population joint JPSTH, averaged across all neuronal pairs, centered around movement onset (top). Bar plot shows values of the JPSTH diagonal (bottom). Solid red line represents the mean value of the JPSTH diagonal in a 10 s time window far removed from movement onset events. Dashed red lines represent the mean ± the SD. B, Same as (A), for pairs of positively modulated neurons. C, Same as (B), for contralateral turns. D, Same as (B), for ipsilateral turns. E, Same as (B), for movement onset events with a small acceleration change. Thick lines mark the time windows for calculation of baseline (black) and post-event (red) maximal values. F, Same as (B), for movement onset events with a small acceleration change. (G) Boxplot of baseline (black) and post-event (red) maximal values of population JPSTH diagonals around movement onset for small acceleration change movements. Each circle represents one mouse (N = 6 mice). H, Same as (G), for large acceleration change movements. I Boxplot of the difference between the maximal value of the population JPSTH diagonal in the post-event and baseline time windows for movement onset with small (left) and large (right) acceleration changes. The bold line is the median and the whiskers are the 25th and 75th percentiles. Red crosses represent outliers. J, Same as (B), for CS+ presentations on expert training session (top). K, Same as (J), for first 10 trials of the first extinction session. L, Same as (E), for CS+ presentations. M, Same as (F), for CS+ presentations. N, Same as (G), for CS+ presentations (N = 5 mice). O, Same as (H), for CS+ presentations. P, Same as (I), for CS+ presentations. The bold line is the median and the whiskers are the 25th and 75th percentiles. Red crosses represent outliers.

Next, the analysis was repeated, considering only pairs of positively modulated neurons. Focusing on this population resulted in a larger peak in noise correlation immediately following movement initiation (Fig. 6B). The diagonals of the population JPSTHs for pairs of negatively modulated neurons or pairs with opposite modulation directions did not exhibit significant peaks (data not shown).

Movement onset times were detected by setting a threshold on the total body acceleration of the mouse and therefore include a wide range of movements. To test whether joint trial-by-trial variations in neuronal activity arise from variability in the mouse’s behavior, we repeated the analysis for subsets of more stereotyped movement events. Limiting the analysis to contra- or ipsilateral turns eliminated the peak in the diagonal of the population JPSTH (Fig. 6C,D). Finally, the same calculations were conducted focusing separately on movements with small (1st quartile) and large (4th quartile) acceleration changes (Fig. 6E,F). While the JPSTH diagonal for movements accompanied by a small acceleration change exhibits a small peak, which does not exceed random fluctuations in the values of the diagonal, movements with large acceleration changes induce a stronger peak in noise correlation.

To examine the significance of this effect, we evaluated the peak values of the diagonal of the population JPSTH for each mouse. Specifically, we compared the peak values within a 0.5 s time window preceding movement onset (baseline) to those within a 1.5 s window immediately following movement onset (post). Our analysis indicated that while there was a significant difference in peak values between the baseline and post-event periods for movements with large acceleration changes (Fig. 6H; p = 0.03125^m, SRT, n = 6 mice), this difference was not significant for movements with small acceleration changes (Fig. 6G; p = 0.0625ⁿ, SRT, n = 6 mice). When we subtracted baseline peaks from post-event peaks and compared these values for small and large acceleration change movement onsets, we found that the difference was significant (Fig. 6I; p = 0.03125^o, SRT, n = 6 mice). Taken together, these results suggest that trial-by-trial correlations around the onset of self-initiated movements are mostly due to variability in the movement of the mouse driving a temporal covariation among neurons across repeated events.

Trial-by-trial correlations following the presentation of task-related cues

To investigate trial-by-trial correlations during learning, we calculated the JPSTHs of neuronal pairs around CS+ presentation focusing on the expert training session and the first extinction session. When all neuronal pairs were included, the diagonals of the population JPSTHs exhibited no peaks for any stage of training (data not shown). Similarly to the movement onset JPSTH, focusing on pairs of positively modulated neurons resulted in substantial noise correlation peaks following CS+ presentation in expert mice (Fig. 6J). Importantly, mouse behavior following CS+ presentation, particularly at advanced stages of training, is relatively uniform. First, mice successfully obtain the sugar-water reward in the vast majority of trials during the expert training session (128 out of 137 trials in 6 trained mice). Additionally, mice acceleration profiles following CS+ presentation become less variable as training progresses (Fig. 2-1). Next, the analysis was repeated focusing on the first 10 extinction trials. These trials were unrewarded, like the rest of the extinction session, but they occurred early enough in the session so that the association between the cue and the reward had not yet been lost. In this case as well, noise correlations rise shortly after the presentation of the CS+ (Fig. 6K).

To further reduce the variability in the behavior, we divided expert session trials into those accompanied by small and large acceleration changes (Fig. 6L,M). In both cases, the diagonals of the population JPSTHs maintain their peaks following CS+ presentation, with no significant difference in the change in JPSTH diagonal peaks around CS+ presentations with small and large acceleration changes (Fig. 6P; p = 0.125^p, SRT, n = 6 mice).

In summary, it appears that trial-by-trial correlations around the onset of self-initiated movements arise mostly from variability in the behavior of the mouse. In contrast, noise correlations surrounding the presentation of task-related cues persisted when we examined subsets of behavioral events with more similar characteristics. In particular, our analysis revealed a significant difference in peak JPSTH values for self-initiated movements with small and large acceleration changes (Fig. 6I), but the same analysis did not produce a significant difference when applied to CS+ presentations (Fig. 6P). Thus, while evidence for dynamic formation of correlated assemblies on individual actions is weak, if it occurs it seems to primarily occur for learned behaviors (of acquired value) but not for innate self-initiated behaviors.