Abstract
The motor cortex not only executes but also prepares movement, as motor cortical neurons exhibit preparatory activity that predicts upcoming movements. In movement preparation, animals adopt different strategies in response to uncertainties existing in nature such as the unknown timing of when a predator will attack—an environmental cue informing “go.” However, how motor cortical neurons cope with such uncertainties is less understood. In this study, we aim to investigate whether and how preparatory activity is altered depending on the predictability of “go” timing. We analyze firing activities of the anterior lateral motor cortex in male mice during two auditory delayed-response tasks each with predictable or unpredictable go timing. When go timing is unpredictable, preparatory activities immediately reach and stay in a neural state capable of producing movement anytime to a sudden go cue. When go timing is predictable, preparation activity reaches the movement-producible state more gradually, to secure more accurate decisions. Surprisingly, this preparation process entails a longer reaction time. We find that as preparatory activity increases in accuracy, it takes longer for a neural state to transition from the end of preparation to the start of movement. Our results suggest that the motor cortex fine-tunes preparatory activity for more accurate movement using the predictability of go timing.
Significance Statement
Anticipating when to move is important in movement preparation. However, it is unclear how the motor cortex prepares movement depending on how easy that anticipation is. To answer this, we examine the motor cortical activity of mice during a delayed-response task. While motor cortical activity rapidly reaches a “movement-ready” state with unpredictable timing of a go signal (go timing), it does so more gradually when go timing is predictable. Moreover, when go timing is more predictable, the motor cortex produces more accurate movement with, unexpectedly, a longer response time. This suggests that rodent motor cortical neurons can resource time information.
Introduction
Movement is prepared before its execution and is often initiated by environmental cues instructing “go.” The firing activities of motor cortical neurons during movement preparation are referred to as “preparatory activity” (Vyas et al., 2020). In the dynamical system view, the final state of preparatory activity right before the go cue (i.e., the preparatory end-state) determines the subsequent neural dynamics of movement execution. As such, ensuing behavior [e.g., reaction time (RT) and arm speed] can be predicted from the preparatory end-state (Churchland et al., 2010; Afshar et al., 2011). In this regard, preparatory activity has been understood as a neural process to achieve an appropriate preparatory end-state to produce the desired movement (Churchland et al., 2006, 2010, 2012; Shenoy et al., 2013; Michaels et al., 2016).
In movement preparation, it is important to predict when to move. For instance, responses are faster, and decisions are more accurate when a go cue is given at the expected timing (Jaramillo and Zador, 2011). However, little is known about how the motor cortex leverages this temporal prediction to enhance behavioral performance and whether the availability of temporal prediction influences movement preparation strategies.
Previous studies have shown that when go timing is predictable, the time left until go timing influences how quickly preparatory activity progresses toward the end-state (Kilavik et al., 2014; Murakami et al., 2014). On the other hand, when go timing is unpredictable, preparatory activity quickly reaches the end-state and sustains it until the go cue is given (Inagaki et al., 2019). However, whether and how the predictability of go timing affects the dynamic profile of the preparatory activity is unknown.
Therefore, we aimed to study whether the dynamics of preparatory activity in the motor cortex alters according to the predictability of go timing and, if so, how the altered dynamics cope with both unpredictable and predictable go timing. We analyzed the firing activity of the anterior lateral motor cortex (ALM) while mice performed one of the two different auditory delayed-response tasks: random- or fixed-delay task. In the random-delay task, delay length was randomly given across trials for unpredictable go timing, whereas in the fixed-delay task, prediction of go timing was enabled via fixed delay.
If the preparatory end-state—which reflects the dynamics of preparatory activity—is identical between the two tasks, ensuing behavior should be similar. However, we observed differences in behavioral accuracy and RT between the tasks, as well as in the degree of selectivity at the end-states. These results indicate that the dynamics of preparatory activity vary based on the predictability of go timing. We selected and examined a putative component of ALM population activity assumed to represent movement preparation specifically. In the random-delay task, the preparation-specific component largely featured “readiness to move,” exhibiting negative correlations with RT. In contrast, in the fixed-delay task, preparatory activity reached the end-state at a later stage of the delay. The preparation-specific component was found to have positive correlations with behavioral accuracy, demonstrating motor decision. Intriguingly, the component also showed positive correlations with RT, exhibiting a similar pattern to the positive correlation between RT and accuracy (speed-accuracy trade-off). Our results suggest that the motor cortex can fine-tune its firing dynamics during movement preparation based on given temporal resources.
Materials and Methods
Behavioral tasks
In this study, we analyzed the open dataset of 11 male mice [aged from postnatal day (P) >60]. The datasets were made publicly available by the Svoboda laboratory at FigShare (https://doi.org/10.25378/janelia.7489253). A detailed description of the data collection procedure can be found in Inagaki et al. (2018, 2019). In brief, the mice were trained to learn the auditory delayed-response task (Fig. 1A,B). Five mice were trained to perform the auditory delayed-response task with randomized delay lengths—the random-delay task. At the beginning of the trial, an auditory stimulus was presented at one of two frequencies, 3 or 12 kHz, which informed the water port location to the right or left, respectively. After 1.15 s of presenting the auditory stimulus, a delay of a certain length was given. Then, given a nonselective auditory go cue, the mice had an opportunity to lick right or left based on the given auditory stimulus. Delay lengths were randomly selected from eight values: 0.3, 0.5, 0.7, 0.9, 1.2, 2, 3.2, and 4 (or 5) s. We excluded the trials with delay lengths of 4 or 5 s consistent time conditions across mice (see below for more details). The probability of delay lengths followed the cumulative distribution function of exponential distributions (τ = 1.4 s) with a 0.2 s offset (Inagaki et al., 2019). This keeps the hazard function of the exponential distribution constant, disabling the mice from predicting when the go cue would be delivered (Zariwala et al., 2013). Trials in which the mice did not lick within 1.5 s after the go cue were classified as no-response trials. We excluded these no-response trials from the analysis. On average, each mouse performed 4.60 ± 3.36 sessions over multiple days, with each session consisting of 121.43 ± 7.57 right-lick trials and 123.30 ± 27.84 left-lick trials on average (mean ± SD).
Different neural dynamics appeared depending on the predictability of go timing. A, Task schematic for the random-delay task. As the trial starts, a 3 or 12 kHz auditory cue is given for 1.15 s (sample period). After a temporal delay (delay period), mice lick right or left based on the auditory cue given during the sample period (response period). Delay length is randomly selected among eight delay lengths across trials. B, In the fixed-delay task, the task structure is the same as that of the random-delay task, but the delay length is fixed as 2 s across trials. C, Example of neural responses during the random-delay task. Firing activities increased from the sample period and persisted in the ready state in the delay period. The vertical black lines indicate the start of the sample (−1.15 s), delay (0 s), and response periods. The top and middle are spike raster plots, and the bottom shows averaged firing rates across trials when a 2 s delay was given. In spike raster plots, trials are randomly sampled for each delay length. The shading denotes SEM. D, Example of neural responses during the fixed-delay task, similar to C. Unlike persistent dynamics in the fixed-delay task, neurons showed ramping dynamics during the delay period. E, Similarity between firing activities during 2 s (horizon) and 3.2 s (vertical) delay lengths in the random-delay task. The similarity was calculated at every time point and averaged across sessions. The dotted white lines separate each period (S, sample; D, delay; R, response). F, Similarity between firing activities in the fixed-delay task. The trials were randomly divided in half, and the similarity between the average firing rates of each set was calculated at every time point—these were averaged across sessions. The dotted white lines separate each period (S, sample; D, delay; R, response). G, We established the reference similarity as the similarity between firing rates at close time points
Separately, six mice were trained to perform the fixed-delay task, in which the delay length was fixed at 2 s, allowing the mice to predict go timing at the start of each trial. Each mouse performed 3.33 ± 1.32 sessions over multiple days, with each session consisting of 124.95 ± 30.10 right-lick trials and 117.55 ± 27.21 left-lick trials on average.
Extracellular action potential analysis
Action potentials (spikes) were simultaneously recorded in the left ALM using 64-channel Janelia silicon probes. Extracellular traces were recorded from the left ALM and were bandpass filtered (300–6,000 Hz). Spike width was calculated as a trough-to-peak interval in the averaged spike waveform. In the random-delay task, 841 of 1,005 recorded neural units were stably recorded through whole trials. Units with a width narrower than 0.35 ms were classified as putative fast-spiking neurons, whereas units with a width wider than 0.5 ms were defined as putative pyramidal neurons. Among the total of 841 units recorded during the random-delay task, 106 units were putative fast-spiking neurons, and 718 units were putative pyramidal neurons. Other units with intermediate spike widths (0.35–0.50 ms, 17 out of 841) that could not be classified into either of these cell types were excluded from the analyses. As a result, 35.83 ± 9.80 units (mean ± SD) per session were used for analysis. In the fixed-delay task, 594 of 755 recorded neural units were stably recorded through whole trials. A total of 62 units were putative fast-spiking neurons, 521 were putative pyramidal neurons, and 11 were excluded from the fixed-delay task. Thus, 29.15 ± 16.14 units per session were used for analysis. Firing rates were computed with a 100-ms-sized squared bin at every 1 ms.
Similarity between firing rate vectors at different time points
If population activities show persistent dynamics, activity patterns at different time points will be similar. On the other hand, if the dynamics of population activities fluctuate, activity patterns at close time points will be similar but those at distant time points will be different. Based on this idea, we analyzed how the firing rates of a population varied over time by measuring the similarity between the firing rates of ALM populations at different time points.
In the random-delay task, we selected two sets of trials with two different delay lengths (e.g., 2 s and 3.2 s). The delay lengths of 2 and 3.2 s were the longest durations used in the random-delay task. Therefore, it was suitable to examine persistent activity across the delay period. Additionally, selecting different delay lengths allowed us to gather more trials, reducing the potential noise effect when averaging firing rates. For each set, we averaged the firing rate vector for N neurons across trials, at each time point. Each of these averaged firing rate vectors was then normalized to have a unit length. We calculated the dot-product (i.e., similarity value) between the trial average vectors of every possible time point pair, and then averaged those across sessions. In the fixed-delay task, the similarity was calculated as in the random-delay task, except that we randomly divided the trials into two sets per session.
Selectivity
We first normalized the firing rate using the baseline mean and standard deviation as follows:
Alignment index
We quantified how well the firing rates of neural populations were aligned on an arbitrary unit vector
v using the alignment index (AI; Elsayed et al., 2016), which is given as follows:
Putative motor preparation and execution subspaces
Nonhuman primate studies have shown that neural trajectories of motor cortical neurons lie within different subspaces for different movement phases. With previous studies confirming that ALM populations show coordinated activities throughout movement preparation and execution (Elsayed et al., 2016; Lara et al., 2018), subspaces representing those coordinated activities can be different between movement preparation and execution (Economo et al., 2018; Inagaki et al., 2022a). Here, we called a subspace representing ALM population activities during movement preparation as a preparation subspace (P) and that during movement execution as a movement subspace (M). Preparation-specific activity and movement-specific activity are defined as neural activity projected on the preparation and movement subspaces, respectively. Preparation-specific activity is a latent component of the neural population related to movement preparation during the delay period, predicting upcoming movement information without making overt movement. Movement-specific activity is a similar latent component directly linked to movement execution.
We assumed that subspaces P and M should meet three requirements based on the findings from other studies. First, preparation-specific activity at the delay end (the preparatory end-state) should at least partially predict movement-specific activity right before making movement (Fig. 3A), as the preparatory end-state was an initial condition of subsequent movement. Second, P and M should be near-orthogonal to each other so that preparation-specific activity precludes muscle activity (Fig. 3A; Kaufman et al., 2014; Elsayed et al., 2016; Stavisky et al., 2017; Kao and Hennequin, 2018). Note that the second requirement does not violate the first requirement: Trial-by-trial correlations can exist between the preparatory end-state and movement-specific activity right before making movement even when P and M are orthogonal to each other (Fig. 3A). Third, P and M should capture large variance of the population activity during movement preparation and execution, respectively. This is necessary because accurate licking is a goal of the task; thus, neural components for preparing and executing licking should be prominently manifested in the population activity.
To estimate preparation- and movement-specific activity, we inferred the projection vectors
p and
m, which linked high-dimensional population activities to low-dimensional activities on P and M, respectively (Fig. 3B). We designed an objective function for learning
p and
m as follows:
A parameter λ balances between the third term and the first two terms. A larger λ puts more weight on increasing correlation during the optimization process, but this could hinder maximizing the first two terms. We empirically determined a minimum λ such that the resulting correlation after learning in the third term was significant for every session (p < 0.05). In this study, λ was set to 0.76 in the random-delay task and to 0.8 in the fixed-delay task.
To initialize
p in the random-delay task, we conducted a principal component analysis (PCA) on trial-averaged firing activity matrix
After maximizing the objective function, we utilized the learned
p and
m to infer preparation-specific activity,
Random AI
We aimed to assess the extent of neural variance that could be attributed to task-irrelevant subspaces. To achieve this, we computed a measure called chance alignment (AIrand) using randomly generated projection vectors,
Distance between the neural states on the P–M dimension
Assuming that P and M could represent movement-related information, we inspected whether the preparation-specific activity and movement-specific activity could predict RT on a single-trial basis (movement-related information is encoded in the preparation and movement subspaces section in Results). We reasoned that preparation- and movement-specific activities at delay end with the shortest RT across trials would be an optimal neural state to generate movement rapidly. Thus, we assumed that RT would increase as neural activities on P and M deviate more from this optimal neural state. We defined
Selectivity captured by the subspace
To estimate how much selectivity was captured by P and M, we first calculated the firing rate matrix,
Correlation between AI and behavioral outcomes
We calculated the correlation between AI and behavioral outcomes by measuring AI at each time point for each session. The AI vector was sized by the number of sessions (sessions × 1), and there were corresponding behavioral outcomes, such as decision accuracy and RT, also sized by the number of sessions (sessions × 1). We then measured the correlation between AI and behavioral outcomes at every time point. We applied a correction for the false discovery rate (FDR) to the p-values for multiple comparisons (Benjamini and Hochberg, 1995).
Reconstruction of firing rates from preparation- and movement-specific activity
Assuming that the population activity can be decomposed into neural components of movement preparation
Distance between firing rates at different time points
We estimated how much firing activity at the delay end, which generates no movement, would have to change to another one that can generate movement. To this end, we measured the distance between firing rates at 50 ms before the delay end
Results
Behavioral outcomes and neural dynamics depend on the predictability of go timing
Aligning with other studies, in the random-delay task, ALM neurons’ firing activities rapidly increased and then plateaued until the go cue (Fig. 1C; Inagaki et al., 2019). Meanwhile, in the fixed-delay task, ALM neurons demonstrated gradually increasing activity throughout the delay period (Fig. 1D; Inagaki et al., 2019).
To analyze these different activity patterns at the population level, we measured the similarities between firing rate vectors at every time point using dot-product (Fig. 1E,F). We observed that in the random-delay task, similarity remained consistently high throughout the delay period (Fig. 1E). Conversely, in the fixed-delay task, similarity levels between adjacent time points were higher than those between more distant time points in the delay period (Fig. 1F).
To statistically evaluate this difference in neural dynamics, we firstly introduced a variable
As a result, for the fixed-delay task, we found that as
We also observed differences in behavioral outcomes. Rodents demonstrated higher decision accuracy and longer RT in the fixed-delay task than in the random-delay task for every delay length (one-tailed paired t test; p < 0.05; Fig. 2A). To investigate the underlying neural mechanism of this difference, we calculated selectivity, indicating the degree of bias toward a particular licking direction in neuronal activity (right vs left; Li et al., 2015, 2016; Fig. 2B). We expected that the degree of selectivity would be related to decision accuracy and elucidate the accuracy difference. Indeed, the degree of selectivity at the preparatory end-state was greater in the fixed-delay than in the random-delay task (one-tailed t test, p < 0.001; Fig. 2C). Further, the degree of selectivity at the delay end increased as animals licked more accurately, only in the fixed-delay task (Pearson’s correlation coefficient, r = 0.716, p < 0.001; Fig. 2D, right). There was no significant correlation between the degree of selectivity and behavioral accuracy in the random-delay task (p > 0.1; Fig. 2D, left).
Behavioral outcomes and task-relevant neural characteristics are different between the tasks. A, Behavioral accuracy and RT. Accuracy was higher, and RT was longer during the fixed-delay task than those during the random-delay task (one-tailed t test). The error bar denotes SEM across sessions. B, Selectivity at delay ended. Selectivity is the firing rate difference between licking directions (right, left) normalized by baseline activity. C, Averaged selectivity after taking absolute on selectivity in (B). Selectivity in the fixed-delay task was significantly larger than selectivity in the random-delay task (one-tailed t test, p < 0.001). The error bar denotes SEM. D, Selectivity in the random-delay task did not correlate with behavioral accuracy (p > 0.1) while selectivity was positively correlated to behavioral accuracy in the fixed-delay task (p < 0.001).
The dynamical system, being near-autonomous, assumes that an initial condition largely determines the neural dynamics for the subsequent move (Churchland et al., 2010; Remington et al., 2018; Wang et al., 2018; Kao et al., 2021). Thus, if the preparatory end-state for both tasks is similar, both transitions to the subsequent movement initiation should span a similar time (i.e., similar RT) as well. However, as shown in Figure 2A, RT was significantly longer in the fixed-delay task, indicating that the preparatory end-state is in different forms for the two tasks.
Based on these neural and behavioral results, we hypothesize that different movement preparation strategies would be adopted by motor cortical neurons depending on the predictability of go timing. When go timing is unpredictable, preparatory activity quickly reaches a neural state that can produce a rapid licking movement and persists in that state until the go cue is given. However, it resulted in less accurate behavioral performance (Fig. 2A). On the other hand, when go timing was predictable, neural dynamics were shaped to increase behavioral accuracy (Fig. 2A).
ALM population activity is segregated into preparation and movement subspaces
To test our hypotheses, we extracted preparatory- and movement-specific activities by inferring preparation and movement subspaces P and M (please refer to “Putative motor preparation and execution subspaces” in Materials and Methods).
Population activities on inferred preparation and movement subspace. A, The schematic diagram for correlated neural states of different movement phases. Each dot represents the neural state of a single trial. The neural state in preparation subspace P predicts the neural state in movement subspace M. P and M are orthogonal to each other so that neural activities on P do not show significant activities on M. Note that the neural state on P could be highly correlated to the neural state on M in trial-by-trial basis even P and M are orthogonal. B, An algorithm seeks vectors
p and
m which are projection vectors to maximize the trial-by-trial correlation of the neural states during distinct epochs (N, number of neurons; TR, number of trials). C, AI across delay period in the random-delay task (delay length, 2 s). AI on P increases over the sample period and flattens when the delay starts (left panel). Meanwhile, AI on M remains low during the delay period and rises after the go cue (right panel). The gray shadow represents SEM across sessions. D, AI averaged across sessions in the fixed-delay task. AI on P gradually increases until the go cue (left panel). AI on M remains low during the delay period and rapidly increases after the go cue (right panel). The gray shadow represents SEM across sessions. E, Distance between
Z and
We calculated the AI of every ms (100 ms bin width) to measure how much neural variance the inferred subspaces capture at each time point [please refer to “Alignment index” in Materials and Methods). In the random-delay task, both AI(Delay, P) and AI(Delay, M) remained relatively persistent during the delay period (Fig. 3C). In the fixed-delay task, AI(Delay, P) gradually increased during the delay period (Fig. 3D, left), while AI(Delay, M) was largely absent as in the case of the random-delay condition (random, 0.073 ± 0.015; fixed, 0.066 ± 0.014; mean ± SD across the delay period; Fig. 3C,D).
We found that both AI(Delay, P) and AI(Response, M) were significantly higher than AIrand in every session (one-tailed t test, p < 10−4 for every session), supporting that subspaces P and M captured a significant amount of the variance of population activity during different periods of the task.
Movement-related information is encoded on preparation and movement subspaces
To verify if the inferred subspaces P and M were related to the preparation and execution of directional licking, we tested whether movement-related information such as RT and licking direction were encoded on P and M (please refer to “Distance between the neural states on P–M dimensions” in Materials and Methods). We first examined whether RT could be predicted by preparation- and movement-specific activities. It was identified that both preparation- and movement-specific activities are predictive of RT (Afshar et al., 2011; Meirhaeghe et al., 2023). Firstly, when the preparatory end-state is shaped close to the optimal preparatory end-state that results in the shortest RT, RT becomes shorter. Secondly, any progression of the neural state toward the state that corresponds to the impending movement prior to the end of the delay period results in a shorter time required to initiate the movement. This neural state corresponds to the movement-specific activity. Therefore, if the movement-specific activity is further developed before the delay ends, it leads to a reduction in RT. Therefore, the distance from the neural state that results in the shortest RT will predict variations in RT.
To test this, we defined a 2D vector,
The resulting correlation coefficients were significantly larger than 0 in both tasks (one-tailed t test, p < 0.05), which indicated that the farther
Z deviated from
Next, we quantified how much directional selectivity was captured by P and M (please refer to “Selectivity captured by the subspace” in Materials and Methods). We found that a significant amount of selectivity was captured by both P and M (Fig. 3F). In the random-delay task, 58.90 ± 6.61% of the population selectivity during the delay period was captured by P and M (48.46 ± 6.14% for P, 10.44 ± 1.91% for M) and 78.00 ± 4.28% during response period (14.78 ± 3.08% for P, 63.22 ± 13.18% for M; mean ± SEM across sessions). In the fixed-delay task, 87.54 ± 4.99% of the population selectivity during the delay period was captured by P and M (75.93 ± 5.67% for P, 11.61 ± 2.10% for M) and 87.34 ± 3.47% during the response period (12.08 ± 2.70% for P, 75.26 ± 16.83% for M; mean ± SEM across sessions). The previous study by Li et al. (2016) reported that neural manifolds that were inferred to discriminate licking direction captured 65.6 ± 5.1% of the population selectivity while mice performed a similar perceptual decision task. Thus, the captured selectivity in the present study was on par with that in the previous study. Collectively, we verified that neural activities on P and M could encode movement information such as RT and licking directions.
Persistent activity during the random-delay task rapidly responds to the go cue under temporal uncertainty
We hypothesized that, in the random-delay task, neural dynamics would quickly reach a neural state ready to initiate licking movement anytime and persist until the go cue (persistent dynamics, Fig. 1C,E). To test our hypothesis, we estimated when neural activity became ready to make a licking movement during the delay period. Especially, we examined when movement-related information such as RT could be predicted from preparation-specific activity.
Lara and colleagues reported that preparation-specific activity increased and peaked several milliseconds before movement initiation. This incremental occupancy of preparation-specific activity among total neural activity indicated that preparation-specific activity became more prevalent as the population activity approached the preparatory end-state (Lara et al., 2018). Thus, we used this occupancy to estimate the extent to which the population activity was ready to initiate a movement. We assumed that high occupancy of preparation-specific activity predicts shorter RT. We used AI(Delay, P) to measure occupancy. Furthermore, to investigate whether movement-specific activity during the delay period reaches the desired state faster, leading to quicker movements after the go cue, we tested if AI(Delay, M) was negatively correlated with RT. Hence, if neural activity at a given time point during the delay period is ready to move, its AI on P and M would show negative correlations with RT.
We calculated correlations between AI and behavioral outcomes (RT, accuracy) across sessions. Specifically, we measured AI as in Figure 3C (please refer “Correlation between AI and behavioral outcomes” in Materials and Methods). We then examined the correlation between AIs and RT (or accuracy) at each time point across all sessions. Here, we conducted a correlation analysis across sessions because decision accuracy and AI were calculated once in each session. In the random-delay task, we found that decision accuracy was not correlated with AI(Delay, P) nor AI(Delay, M) (Fig. 4A). However, negative correlations between AI(Delay, P) and RT were observed before delay started, until the end of delay (FDR corrected; Fig. 4B, top). This early appearance of a negative correlation also existed between AI(Delay, M) and RT (Fig. 4B, bottom). We measured a time lag from trial onset to the first emergence of a significant negative correlation. The first significant correlation appeared even during the sample period (Fig. 4C). Although the correlation coefficient was slightly jittered during the sample period, it became stable when delay started (Fig. 4B). Also, we observed that the average correlation between the first appearance of negative correlation to go cue was negative regardless of delay lengths (Fig. 4D).
Preparation- and movement-specific activity appears right after the delay starts. A, Example correlation between AI and decision accuracy (delay length: 2 s). S, sample; D, delay. B, Example correlation between AI and RT (left panel: 2 s delay, right panel: 3.2 s delay). The magenta-colored dot represents the time point at which the correlation is significant (p < 0.05, FDR corrected). A significant and negative correlation between RT and AI appears before the delay period. C, The first significant and negative correlation between AI and RT appears before the delay begins regardless of delay length. The black dotted line indicates the time point when the delay begins (0 s). The magenta-colored dotted line indicates the shortest delay length given in the random-delay task (0.3 s). D, Average correlation coefficient between AI and RT from the time when the first significant and negative correlation appears to the moment the go cue is presented (mean ± SEM). The dotted line indicates a zero correlation coefficient. E, Counts of sessions where MAE showed no statistical difference with MAE at delay end (two-tailed t test; p = 0.05). The total number of sessions is 23 for the random-delay task. S, sample; D, delay.
However, this correlation between AI and RT may only reveal a general increasing or decreasing trend between them. Thus, we further analyzed when RT can be predicted to examine when RT can be predicted through trial-by-trial analysis. To this end, we trained a linear regressor to predict the RT in each trial using the preparation-specific activity at the last 50 ms of the delay period. We obtained two distinct regressors, one for predicting RT for right-lick trials and another for those in left-lick trials. These regressors were then applied to the entire duration of the task, and the mean absolute error (MAE) was calculated by comparing the predicted RTs to the true RTs. We conducted a statistical test to test if the MAE values at each time point are significantly different from those just before the go cue (two-tailed paired t test, p = 0.05). Our hypothesis was that MAE at RT can be predicted from the motor cortical dynamics and would not exhibit statistical difference with the MAE just before the go cue. We found that a statistically significant difference with MAEs at delay started to vanish from the sample period in several sessions (Fig. 4E). More notably, this significant difference disappeared before the shortest delay length (0.3 s) for all delay lengths, except for 3 sessions among total 23 sessions at most (Fig. 4E). This result aligns with the findings in Figure 4B, indicating that the neural state rapidly reaches and consistently maintains the state predictive of upcoming movement.
Previous studies reported that RT becomes shorter when more time is given to prepare movement with predictable go timing, which is referred to as “RT savings” (Fig. 5A; Riehle et al., 1994; Crammond and Kalaska, 2000; Churchland et al., 2006; Churchland and Shenoy, 2007). Accordingly, we tested if RT decreased as delay length increased in the random-delay task. However, statistical analyses revealed that such RT savings were not present in the random-delay task (ANOVA, p > 0.1 for testing the effect of delay length on RT; linear regression p > 0.1; Fig. 5B). Moreover, there was no significant difference in AI(Delay end, P) among different delay lengths (ANOVA, p > 0.1 for testing the effect of delay length on AI(Delay end, P); linear regression p > 0.1; Fig. 5B). This indicated that the absence of an inverse relationship between RT and delay lengths was related to the lack of change in AI(Delay, P) at delay end. To explore why AI(Delay end, P) did not increase with delay length, we examined how AI(Delay, P) evolved from trial onset to delay end. We observed that AI(Delay, P) did not reach the maximum value at the end of delay when the delay was short (e.g., 0.3 s) and decreased after reaching the maximum when the delay was long (3.2 s; Fig. 5C). Yet, we noticed RT variation over delay lengths in a smaller scale (the shortest RT was 0.215 s at a 1.2 s delay, and the longest RT was 0.226 at 3.2 s delay on average, Fig. 5B). We then found an negative correlation between AI(Delay end, P) and RT (Fig. 5D; Pearson’s correlation coefficient, r = −0.799, p = 0.031). Thus, AI(Delay end, P)—which represents how close preparatory activity is to the preparatory end-state—still elucidated variation of RT even in the absence of RT savings.
RT saving earned by preparation time is removed in the random-delay task. A, An illustration of the RT savings where RT is decreased with the duration of movement preparation time. B, RT across delay lengths. There was no significant difference in RTs between delay lengths (ANOVA, p > 0.1; linear regression, p > 0.1 for slope). The error bar indicates SEM across sessions. Also, there was no significant difference in AI between delay lengths (ANOVA, p > 0.1; linear regression, p > 0.1 for slope). C, AI on P in the random-delay task. ALM activities were mostly persistent, but AI on P slightly increased during the early phase of delay. If the delay length becomes longer, AI on P decreased slightly, but this change was not statistically significant (ANOVA, p > 0.1). The shading represents SEM across sessions. D, At the end of delay, AI on P and RT are inversely related (Pearson’s correlation coefficients, r = −0.799, p = 0.031).
The preparatory end-state maximizes accuracy but accompanies long RT in the fixed-delay task
Behavioral accuracy was higher in the fixed-delay task than in the random-delay task (Fig. 2A). Based on this result, we hypothesized that neural dynamics would shape the preparatory end-state in a way that makes the following movement more accurate when the go cue is predictable. To test this, we measured correlations between AI and accuracy in the fixed-delay task: unlike in the random-delay task, AI(Delay, P) gradually increased during the delay period (Fig. 3D, left) and was positively correlated with accuracy (FDR corrected; Fig. 6A, top). This showed that when preparation-specific activity occupied more of the population activity, decision accuracy was higher. It also indicated that the preparatory end-state, which was used to construct P, represents neural information related to motor decision on licking direction. AI(Delay, M) showed no correlation with accuracy (Fig. 6A, bottom).
Preparation- and movement-specific activity encodes behavioral outcomes in the fixed-delay task. A, Correlation between AI and accuracy. The magenta-colored dot indicates the time point when the correlation is significant (p < 0.001, FDR corrected). A significant positive correlation between accuracy and AI(Delay, P) appears right after delay starts (top). Movement-specific activity does not correlate with accuracy (bottom). B, Correlation between AI and RT. RT shows a significant positive correlation with AI(Delay, P) at late delay, which implies that dominant preparation-specific activity results in longer RT (top; p < 0.05, FDR corrected). Similar to the random-delay task, high AI(Delay, M) results in shorter RT but a significant and negative correlation starts to emerge from −1.22 s in the fixed-delay task, which is delayed than the random-delay task (bottom). C, Counts of sessions where MAE showed no statistical difference with MAE at delay end (two-tailed t test; p = 0.05). The total number of sessions is 20 for the fixed-delay task. S, sample; D, delay. D, Behavioral accuracy is positively correlated with the RT in the fixed-delay task (r = 0.69, p < 0.001). E, In the random-delay task, there was no significant correlation between behavioral accuracy and RT for every delay length (ps > 0.05).
We also observed that RT was slower in the fixed-delay task than in the random-delay task. Thereby, we measured correlations between AI and RT in the fixed-delay task and found a negative correlation across sessions (Fig. 6B, bottom). This was similar to the random-delay task as expected, since AI(Delay, M) reflected how close the neural activities during the delay and at movement initiation were. We observed that the time lag from delay onset to the first emergence of a significant negative correlation was longer in the fixed-delay (<0.8 s after delay starts) than in the random-delay task (sample period). This indicates that during the delay period, neural dynamics gets ready to make movements more slowly in the fixed-delay task than in random-delay. AI(Delay, P), on the other hand, was positively correlated with RT near the end of delay (Fig. 6B, top). The positive correlations of AI(Delay, P) with both RT and accuracy indicated that the preparation subspace P would be altered in the fixed-delay task rather than in the random-delay task, even though P was learned in an unsupervised manner without using the licking behavior information.
We further built the regressors to predict RT using a preparation-specific activity to confirm it RT can be predicted in the late period of delay. We again built two regressors for the right- and left-lick trials. Subsequently, we applied these regressors to the entire duration of the task and calculated MAEs by comparing the predicted RTs to the true RTs. We performed a statistical analysis to examine whether the MAE values at each time point were significantly different from MAE just before the go cue (two-tailed paired t test, p = 0.05). In 3 sessions among the total 20 sessions, we observed that this significant difference disappeared shortly after the delay began (Fig. 6C). However, in the majority of sessions, this significant difference gradually vanished during the late period of the delay, approximately 0–1 s before the go cue. These findings, together with the results in Figure 6A,B, suggest that movement preparation in the fixed-delay task tends to occur relatively late in the delay period.
In fact, behavioral data revealed that accuracy and RT were positively correlated with each other in the fixed-delay task (Pearson’s correlation coefficients, r = 0.69, p < 0.001; Fig. 6D), but not in the random-delay task (ps > 0.05 for every delay length; Fig. 6E). Taking this with the behavioral observations, we postulated that the preparatory end-state was shaped toward greater behavioral accuracy but with longer RT, in the fixed-delay task.
We further examined why increasing behavioral accuracy accompanied longer RT. We defined
Accuracy increases as distance
If
We further investigated the correlation between RT and decoding accuracy at the single-trial level using preparation- and movement-specific activity. Briefly, we employed a logistic regression classifier trained on a random selection of 60% of the total trials to predict the licking direction (right-lick vs left-lick). Subsequently, we applied this classifier to predict the licking direction for the remaining 40% of the trials. We defined the probability that the classifier predicted the true lick direction as the decoding accuracy of each trial.
Additionally, we calculated the distance L in each trial between firing rates 50 ms before the end of delay and those at 50 ms before the first lick, as in Equation 13. Then, we calculated a correlation coefficient between decoding accuracy and RT across trials and statistically tested whether this correlation coefficient between decoding accuracy and RT significantly decreased when controlling for distance L. For the random-delay task, among a total of 23 sessions, we found that the trial-by-trial correlation between RT and decoding accuracy was significantly lower than 0 in 17 sessions (one-tailed t test, p < 0.05).
For the fixed-delay task, we noted that the trial-by-trial correlation between decoding accuracy and RT was significantly greater than 0 in 13 out of the total 20 sessions (one-tailed t test, p < 0.01). Furthermore, among these 13 sessions, we found that in 10 sessions, the trial-by-trial correlation between RT and decoding accuracy significantly decreased when we controlled for the influence of distance L (one-tailed t test, p < 0.01). This suggests that the positive correlations between RT and decoding accuracy at the single-trial level within a session can, to some extent, be attributed to the distance between firing rates at the end of the delay and at movement initiation. This differential effect of L on the correlations between decoding accuracy and RT among the tasks aligns with our conclusion that preparatory activity is influenced by the predictability of go timing.
Discussion
We demonstrated that preparatory activity is systemically altered depending on the predictability of go timing. We inferred preparation and movement subspaces from ALM population activity to decompose preparation-, and movement-specific activity (Figs. 3–6). When go timing is unpredictable, neural dynamics quickly reach the preparatory end-state and maintain it until movement generation (Figs. 1, 4). On the other hand, neural dynamics reached the preparatory end-state gradually when there was no need to cope with temporal uncertainty (e.g., fixed-delay length) and saliently distinguish between alternative licks (Fig. 2C). Behaviorally, however, predictable go timing resulted in higher accuracy and longer RT (Fig. 6D). This is because, as accuracy increased, the preparatory end-state moved further away from the neural state for the onset of movement, resulting in longer RT (Fig. 7B). Thus, motor cortical dynamics may evolve to generate an end-state for increased movement accuracy, but with no free lunch—the end-state also becomes more distant from movement initiation states, resulting in longer RT in the delayed-response task with fixed-delay length. We observed a similar tendency at a single-trial level from an additional analysis. When preparatory-specific activity exhibited pronounced licking direction information, it resulted in longer RT on a trial-by-trial basis in the fixed-delay task. Our findings suggest a speed-accuracy trade-off in which the motor cortex can flexibly generate preparatory activity using temporal resources to enhance behavioral accuracy at the cost of shifts in the preparatory state.
We find two main differences in the movement preparation process depending on the predictability of go timing: the speed at which the preparatory end-state is reached and the behavioral information that the preparatory end-state encodes. When go timing is unpredictable, neural activity rapidly reaches the end-state and persists there until cued to initiate movement (Fig. 1C). A relatively short time to collect sensory evidence may explain the low decision accuracy in the random-delay task. Meanwhile, if the preparation duration is fixed, preparatory activity ramps up throughout the whole delay period. When the go cue is unexpectedly not delivered at the predicted go timing in the fixed-delay task, preparatory activity persists in its current state, like in the random-delay task (Tanaka, 2007; Inagaki et al., 2022b).
In both tasks, RT became faster as the neural activity progressed into movement subspace M during the delay period (Figs. 4B, bottom, 6B, bottom). It supports the idea that the neural subspace representing movement execution (i.e., movement subspace) would remain consistent between the tasks. However, what and how activities on the preparation subspace encode for upcoming behavior were different between the tasks. In the random-delay task, greater occupancy of preparation-specific activity among population activity resulted in a shorter RT, while it did not predict decision accuracy. In the fixed-delay task, however, the greater occupancy resulted in higher accuracy but longer RT. A recent study confirmed the observation that the same preparation subspace activity appears when pursuing the same movement under different movement initiation conditions (e.g., reaching after a fixed-delay or self-cued reaching; Lara et al., 2018). Therefore, whether the movement is cue- or self-initiated, the same preparation subspace is represented. However, unlike our study, the movement preparation process in Lara et al. (2018)'s study did not involve decision-making, because the target was fully specified. Thus, we surmise that systematic changes in a neural network level would emerge through learning different goal-directed movement tasks, resulting in differential formations of preparatory dynamics.
One potential source that may contribute to that differentiation is ramping input, which delivers task-relevant timing information (Kunimatsu et al., 2018). Previous studies showed that preparatory activity approaches one of the discrete attractors that each match with a different movement option (Shenoy et al., 2013; Inagaki et al., 2022a). When go timing is predictable, timing information is reflected in preparatory activity via a ramping input from other brain regions (Inagaki et al., 2018, 2019, 2022a; Finkelstein et al., 2021). The input drives the dynamics of preparatory activity to move the attractors apart, predisposing motor cortical neurons to be able to distinguish between movement options (Fig. 2C; Finkelstein et al., 2021). However, the ramping input is absent when go timing is unpredictable (Inagaki et al., 2019). In practice, ALM neurons encoding ramping signals significantly overlapped with those encoding the choice of movement option (Yang et al., 2022). Collectively, the presence or absence of ramping input into ALM circuits would result in modifications of preparatory activity.
It has been assumed that one of the roles of movement preparation is to make rapid responses. For example, longer preparation time yields shorter RT (Rosenbaum, 1980; Riehle and Requin, 1989; Duan et al., 2021). Moreover, when the preparatory end-state is disrupted by electrical microstimulation, RT savings are lost (Churchland and Shenoy, 2007). However, when go timing is unpredictable, we found long RT and persistent decision accuracy even if a longer preparation time is given (Fig. 5B). The level of movement readiness did not increase as well (Fig. 5B). These results suggest that, when delay length is unpredictable, neural dynamics do not proceed further toward the neural state of movement (i.e., neural response just before the first lick), thus yielding constant RT across different delay lengths. Other studies have demonstrated that RT savings appeared when delay length was randomly sampled from a uniform distribution (Duan et al., 2021). Yet, it is known that a hazard rate—the probability that the go cue will be given in the remaining delay period—keeps increasing until the end of delay, in a uniform delay length distribution [see Zariwala et al. (2013) for more information]. Since this allows prediction of go timing, the preparatory activity would proceed toward the neural activity of movement as time passes, and RT would be shortened accordingly. However, in our data, delay length was randomly sampled from an exponential distribution, which is known to make a hazard rate constant over the whole delay period (Inagaki et al., 2019) to rule out the possibility of predicting go timing. Therefore, RT savings may occur only when the probability of presenting the go cue increases as the waiting time increases (Janssen and Shadlen, 2005).
Speed-accuracy trade-off has been explained using a serial accumulate-to-threshold framework (Palmer et al., 2005; Heitz, 2014; Servant et al., 2019). In this framework, inaccurate but fast decisions are made by accumulating only a few pieces of sensory evidence. In contrast, decision-making is accurate but slower when subjects accumulate more sensory evidence. In the fixed-delay task, however, the time given to collect sensory evidence was fixed. Also, temporal delay is interleaved between sample and response periods; thus, what determined a balance between speed and accuracy would not be an accumulation of sensory evidence. We found that the transition from the preparatory end-state to the neural state of movement onset became distant as decision accuracy was high. This could be because modifying neural activities to improve the decision-making process would affect movement preparation. For example, selectivity in ALM becomes salient and correlations between selective neurons are reestablished through task learning (Komiyama et al., 2010). It is unclear why this modification should accompany an unoptimal response RT, but an intrinsic neural structure could affect in shaping the neural dynamics (Sadtler et al., 2014).
Distinct neural components may emerge during the delay period, which is unrelated to preparatory activity. One such component could be neural activity associated with uninstructed movement, which is not essential for performing the task but facilitates subsequent target movement. For example, when mice prepare for licking, they proactively position their jaws in the direction of the upcoming lick direction to reduce RT (Mangin et al., 2023). Previous studies have suggested a potential link between neural responses to uninstructed movements and decision variables (Musall et al., 2019; Stringer et al., 2019). Therefore, it is essential to distinguish between neural components associated with instructed and uninstructed movements using behavioral tasks and computational methods (Zagha et al., 2022). In our study, we aimed to segregate preparatory activity for instructed movement by constraining the preparation subspace to capture the upcoming neural state of the licking. Future research should investigate how neural components of uninstructed movements could be effectively segregated by the devised method in this study.
ALM is more likely regarded to play a role in the premotor cortex in movement preparation. ALM neurons encode abstract aspects of licking such as licking initiation. Licking kinematic variables such as tongue length and licking rates are encoded in other motor cortical regions than ALM such as the tongue and jaw regions of the motor cortex (Xu et al., 2022). Recent studies have found that neural activity in the premotor cortex modulates the rodent’s primary motor cortex during learning new motor skills, emphasizing hierarchical structures in motor areas (Veuthey et al., 2020). Our results suggest that the preparation-specific activity in ALM populations could be modified depending on the predictability of go timing while pursuing the same action in mice, emphasizing the crucial role of ALM as a premotor cortex in adjusting motor task strategy. However, our study does not explain the optimality of the adopted strategy in a given task structure. Future studies could reveal how cognition shapes neural dynamics and how, in turn, intrinsic constraints affect cognition.
Footnotes
We thank the Svoboda laboratory for the contribution of data publicly at the FigShare, a data-sharing website. This study was supported by the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. RS-2023-00302489) and the Alchemist Project Brain to X (B2X) Project funded by the Ministry of Trade, Industry and Energy (No. 20012355; NTIS No. 1415181023).
The authors declare no competing interests.
- Correspondence should be addressed to Sung-Phil Kim at spkim{at}unist.ac.kr or Jeong-woo Sohn at jsohn{at}cku.ac.kr.
