Neurons remap to represent memories in the human entorhinal cortex

The entorhinal cortex (EC) is known to play a key role in both memory and spatial navigation. Despite this overlap in spatial and mnemonic circuits, it is unknown how spatially responsive neurons contribute to our ability to represent and distinguish past experiences. Recording from medial temporal lobe (MTL) neurons in subjects performing cued recall of object–location memories in a virtual-reality environment, we identified “trace cells” in the EC that remap their spatial fields to locations subjects were cued to recall on each trial. In addition to shifting its firing field according to the memory cue, this neuronal activity exhibited a firing rate predictive of the cued memory’s content. Critically, this memory-specific neuronal activity re-emerged when subjects were cued for recall without entering the environment, indicating that trace-cell memory representations generalized beyond navigation. These findings suggest a general mechanism for memory retrieval via trace-cell activity and remapping in the EC.


Introduction 27
The ability to organize our past experiences is a defining aspect of memory, and a crucial component of this 28 is distinguishing between overlapping experiences for memory retrieval. For example, imagine that you 29 have been asked to recommend things to do in a city you have visited frequently-the question elicits your  (Squire et al., 1993). Spatially modulated neuronal activity has also provided evidence for a 41 mechanism that might support memory representation and differentiation. Place fields "remap"-changing . This suggests remapping serves as a mechanism linking spatially responsive 50 neurons to memory, such that cells remap in responses to changes in memory states. 51 We hypothesized that recalling different memories would elicit remapping of neuronal activity to the 52 location of the specific memory being recalled. We further hypothesized that the memory-specific neural 53 activity associated with remembered locations would be accessible even when subjects were not moving 54 through the environment. In this way, we theorized that neurons in the MTL integrate the content and context 55 of past experiences to represent and differentiate between memories-neuronal representations that persist 56 beyond that environment for general memory retrieval. 57 To test this hypothesis, we recorded and analyzed the activity of single neurons from the MTL of human 58 epilepsy patients as they performed a cued spatial-memory task in which they recalled the locations of 59 cued objects while moving through a virtual environment. We observed a unique population of cells in 60 the entorhinal cortex and cingulate, which we refer to as trace cells. Specifically, trace cells remap their 61 activity to locations near the cued object-location memory, indicating that their neural activity related to the 62 specific location relevant for the memory cued on each trial. Furthermore, as subjects moved through the 63 cued object's remembered location, the firing rate of entorhinal trace cell could decode the cued object for 64 that trial, and this memory-specific neuronal activity was also present even when subjects were not moving 65 through the environment. Trace cell activity in the entorhinal cortex thus illustrates a potential neural basis 66 for the representation and differentiation of experiences for memory retrieval.

68
We recorded from 295 neurons in the entorhinal cortex, hippocampus, amygdala, and cingulate cortex of 19 69 neurosurgical patients performing an object-location memory task in a virtual, linear track environment (Fig. 70 1A). In this task, subjects were instructed to learn the locations of different objects along the track and then 71 to recall the locations when the objects were removed. The task consisted of separate encoding trials and 72 retrieval trials. Encoding and retrieval trials follow the same general structure and task instruction, except 73 that objects are visible on the track during encoding trials, and are absent during retrieval trials. Each trial 74 begins with an "cue period," in which subjects view text instructions indicating the cued object for that trial. 75 Following this is the "hold period," during which subjects remain stationary at the entrance to the track for 76 4 seconds. Then, the "movement period" begins and the subject is moved automatically down the track. 77 During encoding trials, the object remains visible on the track, allowing the subject to easily press a button as 78 they approach the object's location (Fig. 1B). During retrieval trials, the object is absent and subjects press a 79 button at the location where they remember the cued object being present. Figure 1C shows that subjects 80 performed this task accurately because they pressed the button within 2.8 virtual units of the correct location 81 on average (7% of the track length). 82 We examined the activity of each neuron in the task during retrieval trials by computing its firing rate   : Place cell activity. A) Raster plot and mean firing rate for two representative place cells recorded from the hippocampus. Box and shading indicate location where the cell activity significantly increased. Dotted line represents the significance threshold, assessed with a shuffling procedure (see Methods). B) Distribution of mean firing rates among place fields. C) Distribution of field sizes as a percentage of the track. D) Proportion of place cells recorded in each brain area. A = amygdala, H = hippocampus, EC = entorhinal cortex, C = cingulate. Asterisks indicate location with a significant proportion of place cells (binomial test, p < 10 −4 ). Bars indicate the 95% confidence interval from a binomial test. E)Number of responsive cells with more than one spatial field. Inset shows an example of multi-peak cell recorded from the cingulate. as a function of the subject's virtual location along the track. To assess the modulation of neuronal activity, 84 we used a two-way repeated-measure ANOVA to identify neurons whose activity varied as a function of the 85 subject's location during retrieval trials, the retrieval cue, and their interaction. This analysis revealed two 86 groups of neurons with distinct firing patterns. We found neurons with firing rates that varied as function 87 of subject location alone ( Fig. 2A) Dostrovsky, 1971). We also found a distinct cell type, which we call "trace cells," that exhibited spatial firing 89 fields that remapped to different locations along the track according to the retrieval cue on each trial (Fig. 2B, 90 S2).

91
Place cells activate in fixed locations, independent of memory retrieval demands. While subjects 92 moved down the track, place cells activated in fixed locations of the environment ( Fig. 2A, 3A). We defined 93 place cells as those that showed a significant main effect of subject location on firing rate, and had at least one 94 place field. We defined place fields by characterizing contiguous locations in which firing rate significantly 95 exceeded a threshold measured with a permutation procedure (see Methods). A total of 16.9% of cells 96 analyzed (50/295, p < 0.05, binomial test) showed this consistent spatial modulation of firing rate, and we 97 classified them as place cells. A majority of spatial fields were smaller than 10% of the track length and none 98 covered more than 40% of the track (Fig. 3C). We found significant numbers of place cells in the entorhinal 99 cortex, hippocampus, and cingulate ( Fig. 3D; binomial test, p < 0.05).

100
Because 90% of the place cells continued to show this spatial coding even after accounting for potential  Trace cells remap according to cued memory retrieval In addition to place cells, we also observed trace 105 cells whose firing fields remapped depending on the memory retrieval cue for each trial (Figs. 2B, S2). Figure   106 2B depicts two example cells recorded in the entorhinal cortex that showed spatially modulated activity.

107
However, the particular location preference of each cell changed depending on the retrieval cue, or the 108 specific object location that the subjects had been instructed to recall on each trial. Specifically, these cells 109 significantly activated as subjects approached the cued object's location, and then decreased afterwards. 110 We defined trace cells as those that showed a significant interaction effect of the subject's location and the 111 retrieval cue on firing rate, and had at least one trace field. We characterized trace fields as the place field that 112 a trace cell exhibited during the retrieval trials for a particular object location. We found significant numbers 113 of trace cells (43/295; binomial test, p < 10 −11 ), primarily in the entorhinal and cingulate cortices (Fig. 4A). 114 We observed at least one trace cell in 15 of 19 subjects (Supp.  The fact that trace cells remapped in response to changes in the memory retrieval cue seemed to 117 demonstrate a possible mechanism whereby a single cell's activity could maintain distinct representations 118 of different memories. To test whether the particular nature of this remapping related to the specific object 119 location that was recalled, we assessed where trace fields were most prominently located with respect to  Figure 4: Trace-fields remap according to subjects' memory for cued object locations. A) Distribution of trace cells across brain areas. Asterisks indicate significance at p < 10 −5 (binomial test). B) Distribution of trace-field locations relative to object location (indicated by black line). Asterisk indicates the greater prevalance of trace fields immediately before versus after object location (χ 2 (1) = 10.4, p < 10 −3 ). C) Distribution of the counts of unique trace fields exhibited by trace cells. D) Comparison of trace cell's peak firing rate in field (z-scored) between encoding and retrieval trials (t(125) = 15.6, p < 10 −30 ). E) Raster plot of spiking activity and corresponding PSTH for three representative entorhinal cortex trace cells, aligned relative to response location (indicated by blue dotted line). F) Mean firing rate (z-scored) of all trace cells aligned to response location. Shading indicates SEM. Asterisks indicate spatial bins that are significant from baseline (p's< 0.05, one-sample t test, FDR-corrected). G) Pre-response and post-response firing rate (z-scored) compared between encoding and retrieval trials. Asterisks indicate significance from an ANOVA (interaction of pre-vs. post-and encoding vs. retrieval, F(1) = 5.79, p = 0.016).
cued object locations during retrieval trials (when the object is no longer on the track). We found that 121 trace fields were predominantly located preceding the cued object's location (χ 2 (1) = 10.4, p < 10 −3 ; Fig.   122 4B), which indicated to us that the activity of these cells could be driven by the memory for the object's 123 location. Critically, trace cells did not represent multiple remembered object locations simultaneously, instead 124 switching between trace fields depending on the specific cued object (see Fig. 2B, S2). Trace cells did not 125 always remap to the location of every cued object, with trace cells exhibiting anywhere from 1-4 trace fields 126 throughout the session (Fig. 4C). These observations suggest that human trace cells remapped according 127 to the retrieval cue-evidence that top-down memory retrieval demands influence remapping of trace-cell 128 activity. object on the track. We thus compared neural responses between retrieval and encoding trials because it 136 allowed us to control for effects unrelated to memory retrieval. We examined trace cell firing rates as subjects 137 passed through the center of each trace field during encoding versus retrieval trials and found that trace-cell 138 firing activity was significantly greater during retrieval than encoding (t(125) = 15.5, p < 10 −30 ; Fig. 4D).

139
This significant increase in activity during retrieval suggests that trace cell activity reflected memory for 140 object locations rather than visual responses to the object or it's location.

141
These observations suggested that trace cells remap to cued object locations during memory retrieval, but 142 did not directly link trace-cell activity to subjects' memories for object locations. In order to assess whether 143 trace-cell activity supports memory retrieval directly, we next assessed trace-cell activity relative to subjects' 144 response locations. Aligning trace-cell activity to subjects's responses on retrieval trials, we found that trace An alternate explanation for these findings is that trace cells were activating in anticipation of subjects' 152 motor response (i.e., the button press). As before, we tested this possiblity by examining encoding trials when 153 the motor demands identical to retrieval. During encoding trials we found significantly smaller changes in 154 firing rates around the response location (ANOVA F(1) = 5.79, p = 0.016; Fig. 4G). This diminished effect 155 in encoding trials indicates that trace-cell activity does not reflect anticipatory motor responses (see Supp.

156
Analyses for additional controls).

157
The firing rates of entorhinal trace cells distinguish between separate memories. In everyday life we t-tests p < 0.05), indicating that trace cells were possibly engaged by memory retrieval or maintenance 168 related to the cued object during this period, even though subjects were not moving in the environment.

169
If trace cells activate after cue presentation during the hold period, we hypothesized that this activity was 170 related to the neural patterns associated with retrieval of cued object locations? If so, this would support the 171 idea that trace-cell activity organizes memories in space but also generalizes beyond navigation to distinguish 172 memories for retrieval. We therefore assessed if the trace-cell activity during the hold period correlated with 173 the activity during the "response period" on the same trial, which is the period during movement when subjects 174 responded to indicate the remembered object location. If trace cells were exhibiting a memory-specific rate 175 code in response to the different retrieval cues, we reasoned that this level of neuronal activity should remain 176 intact over both these periods. Consistent with our predictions, we found that trace-cell activity was positively  Figure 5: Trace-cell activity is correlated between the hold period and response period. A) Mean firing rate (z-scored) across all trace cells by task period. Asterisks indicate p < 0.05 (FDR corrected t tests); † indicates p < 0.1. B) Relation between firing rates between hold-and response periods for six representative trace cells. Black line denotes the robust linear regression fit. C) Distribution of Pearson correlation coefficients for trace-cell firing rates between hold and response periods (mean = 0.31). Dotted line denotes control distribution (see Methods and Supp. Analyses). Asterisk indicates significant difference (t(42) = 6.5, p < 10 −9 ). D) Mean normalized firing rate during hold and response periods for each object cue, for a representative entorhinal cortex trace cell.
when subjects are held stationary at the entrance to the environment (e.g., Fig. 5D & S4).

180
To more directly demonstrate that the same patterns of neuronal activity in both the hold and response 181 period consistently carried information about the memories being retrieved on each trial, we next used a cross-182 validated decoding framework to test if trace-cell activity was predictive of the content of object-location 183 memories. This decoding analysis not only tested if activity in a single task period was able to reliably 184 decode the cued object-location memory, but also whether a single shared neuronal representation of the 185 current memory persisted across the hold and response periods, indicated by whether decoders trained in 186 different settings reliably generalized to the response period neural activity. We trained decoders to use the 187 normalized (z-scored) trace-cell firing rate from each task period (see Fig. 1A) to predict the identity of the Kubie, 1987). In this way, place cells could potentially index different spatial maps for experiences in a 219 particular spatial context. Given that different maps could conceivably re-activate during retrieval of those 220 experiences, remapping was theorized to be a candidate mechanism for the indexing of different memories as had previously been encountered, suggesting that the activity of these cells represent a non-specific, putative 244 "memory trace" of the objects that the rodent had encountered in the environment, indicating "some object 245 was here once." In contrast, we show that trace-cell remapping in humans is driven by memory demands, 246 leading to memory traces specific to the cued object location for memory retrieval. press a button on the game controller when they reach the location of a specified object ("instruction period").

281
Immediately after receiving this cue, subjects press a button on a game controller to move to the "hold period," 282 in which they are held stationary at the entrance to the track for 4 seconds. Next, the "movement period" 283 begins automatically, in which subjects are moved forward along the track. Subjects are moved passively for 284 56 of 64 trials and on other randomly selected trials control movements with a handheld controller (Supp. Fig.   285 1A )-we did not analyze the manual movement trials here. Individual trials consisted of either encoding or 286 retrieval trials (see 1A). The first two times that subjects encounter a particular object are encoding trials, in 287 which the object is visible during movement so the subjects can learn its location. On the subsequent retrieval 288 trials, the object is invisible during movement and subjects are instructed to recall its location by pressing 289 the controller button when they believe they are at the correct location. Subjects encode and retrieve a total 290 of 4 unique object-location associations (16 trials of each) over the course of a session, with each object 291 located at a different randomly selected location ( Figure 1B). In addition to pressing a button to indicate their 292 memory for the object location, subjects are told to press a button as they enter the "stopping zone" at the end 293 of the track, which is visually delineated by a new floor coloring at the end of the track. Pressing the button 294 in this region ends the movement period, and subjects are then shown a fixation cross for 5 seconds ("fixation 295 period"). Finally, during the "feedback" period at the end of each trial, subjects receive points corresponding 296 to how close they pressed the button to the correct location during movement. Only one object was ever 297 present on the track at any given time. The task was split such that the retrieval cue for the first half of each 298 session could correspond to objects 1 or 2, while retrieval cue for the second half could correspond to objects 299 3 or 4.

300
A distinctive feature of our task is that during movement periods subjects are moved subjects passively 301 while their speed is automatically changed in a seemingly random fashion. These uncontrolled speed changes 302 encourage subjects to attend continuously to their current location because they cannot accurately predict  Figure 1B. When speed changes occur, the speed 306 varies gradually over the course of one second to avoid a jarring transition.

307
To measure task performance, we compute subject's distance error (DE) on each trial, which is defined as 308 the distance between the subject's response location and the actual location of the object. We used a median  For each cell, we counted the spikes in each spatial bin and divided this quantity by the time spent in that bin 334 to yield a firing rate estimate. We smoothed this firing rate estimate on the single-trial level using a Gaussian 335 kernel with a width of 8 VR-bins. We excluded the bins in which subjects spent less than 100 ms over the 336 course of the entire task. This excluded several of the bins in the stopping zone, because the movement period 337 ended as soon as subjects pressed the button in the stopping zone. We normalized firing rate for all analyses 338 comparing spiking across different task periods or trial types, such that a z-score of 0 represented a cell's 339 mean firing rate across all periods of the task. 340 We used a 2-way repeated-measure ANOVA to examine the effects of subject location (1-40 VR-bins), 341 object cue (1, 2, 3, 4), and their interaction, on the binned firing rate of each cell. We defined place cells 342 as those that showed a consistent and significant main effect of location on firing rate via the ANOVA, and 343 that also had a place field greater than 5% the size of the track. Additionally, we performed an ANCOVA estimates 500 times and re-analyzing the data. Six cells showed a main effect of object cue on firing rate.

352
These cells were excluded from analyses. 353 We defined trace cells as those cells whose firing rate showed an interaction between subject location and 354 object cue in the ANOVA. Trace fields in trace cells were determined via the same method as for place cells, 355 using a post-hoc test to identify firing fields that were specific to individual object-location associations. A 356 trace field for a particular object cue was considered unique if the peak location did not overlap with that of 357 any other trace field for that cell (Fig. 1D).

358
Decoding analysis. We used a multivariate decoding framework to test whether trace-cell activity reflected 359 information about the content of each object-location memory across different retrieval contexts. This  The purpose of this decoding analysis was to ascertain whether a group of neurons provided a representa-366 tion of the contents of memory that was similar in form across across separate contexts. For this decoding, 367 we used a k-nearest neighbors (kNN) algorithm using a one-vs.-all paradigm for multi-class decoding of the 368 identity of the remembered item from the recorded neuronal activity. Firing rates were binned by task period 369 and normalized. On each trial we computed the "response period" firing rate by normalizing the activity 370 in the 10 VR-bins preceding the response by the 10 VR-bins following the response (Supp Fig. 5). This 371 normalization procedure captured both the pre-response increase and post-response decrease in firing rate 372 described in the results. We used a similar method to compute a matched "control period" utilized in Figure   373 6C, using the 20 VR-bins immediately following the end of the response period. This ensured that the control 374 period was of equal length to the response period, and that the neural activity during this control period did 375 not overlap with the neural activity during the response period. 376 We trained all the different task period decoders on the firing rate during a particular period of the task 377 and tested on the response period neural activity. Additionally, we trained and tested one decoder with 378 the response period firing rate -this decoder was trained using leave-one-out cross validation to assess 379 performance (Supp Fig. 5). We assessed significant decoding accuracy using a binomial test. Chance-level 380 decoding accuracy was at 25%, given the equal presentation of the 4 different objects.   Neurons remap to represent memories in the human entorhinal cortex Qasim et al.

Supplementary Analyses
Control: Binning firing rate by space. In order to assess the spatial binning on our results, we calculated the results of our main analyses using 30 and 50 equal sized spatial bins, rather than 40 bins as in our main analyses. The number of place cells and trace cells identified by the ANOVA did not vary significantly as a function of the number of bins (results remained within 95 % confidence interval of binomial test determining significant proportion of place and trace cells). This indicates that our primary results are not determined by the spatial scales of the bins used for data analysis.
Control: Electrodes in epileptic regions. The subject cohort examined in this study has drug-resistent epilepsy. Prior research has supported past work in epileptic cohorts through use of scalp EEG or fMRI (Lachaux et al., 2003). Still, it is important to consider whether electrophysiology research in the epileptic brain is reflective of healthy brain. Approximately 31% of the single-units we analyzed were recorded on microwires localized to clinically determined ictal onset zones. To more rigorously control for any confounding effect of epileptic activity, we re-ran all analyses excluding all neurons recorded from these clinically defined ictal onset zones. Our main findings remained unchanged with respect to the proportion of place cells and trace cells, and their properties. Further, this data exclusion did not change any results with respect to trace-cell activity or decoding.
Control: Independence of multiple sessions by a single patient. Several patients contributed multiple sessions of the task, with each session analyzed independently. However, in order to ensure that patients contributing multiple sessions to this study were not confounding our results (Supp . Table 1), we ran control analyses utilizing only the first session recorded from each patient. This controlled for any confounding effect of multiple sessions. Our main findings remained unchanged with respect to the proportion of place cells and trace cells, and their properties. The results presented here thus utilize all the data.
Trace cell activity follows subjective memory judgment. We sought to understand whether trace-cell activity followed the participants' subjective memory of the object-location regardless of whether that memory was correct or incorrect. We tested this by splitting the retrieval task data into "good" and "bad" memory trials utilizing a median split within each subject. Both good and bad retrieval trials showed the same pattern of trace-cell activity with respect to response locations (pre, post paired t-test t(978) = −0.43, p = 0.66 t(978) = 1.12, p = 0.26; Supplemental Fig. 3C) -firing elevated before subjects' response and decreased after, regardless of whether the trial was from the best half or worst half of the subjects' performance. Given that we determined trace-cell activity was not an effect of the button press action itself, this suggested that the trace cells track a person's subjective memory of the object-location, whereas if these cells were involved in context reinstatement we likely would have observed less activity during bad memory trials.
Control Analysis: Trace cells do not encode time to button press. One alternative explanation of tracecell activity is that it reflects a fixed anticipatory signal for the subjects' motor action, the button press. Given that every trial featured random speed changes, our task controlled for consistent effects of time. This inherently meant that trace cells activating at consistent locations relative to subjects' response were not responding at consistent times relative to that response, as time and location were dissociated across trials. To illustrate this, we assessed trace cell firing as a function of time relative to subjects' response. We analyzed the activity of the trace cells time-locked to button press, rather than aligning trace cell activity by spatial bin/distance to button press as in Figure 2F. Anticipatory motor responses are thought to occur within 1-second preceding the relevant event (Mauritz and Wise, 1986), so we analyzed trace-cell firing in a 3-second window surrounding the response. Supplementary Figures 3D,E shows that trace cells did not show any consistent effect of time, as opposed to Figure 2D, in which trace cells exhibit clear preference for particular spatial positions that preceded retrieval. These results provided further evidence that the activity of trace cells reflected spatial activations at or near the remembered positions of cued objects, rather than simply firing at a fixed time preceding button press.
Control: Trace cell hold-response period correlation does not result from temporal auto-correlation.
Given that the response period activity was calculated by normalizing the pre-response firing rates by the post-response firing rates, the results in Fig. 4B,C already control for the effects of temporal autocorrelation (i.e., the hold period firing rate predicts the firing rate for the rest of the trial). To further ensure that the correlation we observed between the hold period firing rate and the response period firing (see Fig. 4B,C) was not the result of such a confound, we computed the correlation between the hold period firing rate and a "control period". Control period activity was computed using the length of the track following the response period, thus ensuring it used the neural activity in the regions of the track that did not overlap with the response period. This control period firing rate was computed identically to the response period-the mean firing rate of the first 10 VR-bins of the control period were normalized by the mean firing rate of the last 10 VR-bins. We then computed the correlation between the hold and control period firing. The null distribution of correlation coefficients assessed in this way is depicted by the dotted line in Fig. 5C. Mean Error (VR-bins)

Trials
Supplementary Figure 1: Trial structure and average response error: A) Schematic depicting one example of the trial structure during the task. The first two trials for each object cue were encoding trials (black), after which subjects had retrieval trials (passive movement, blue, manual movement, red). The cued object switched between objects 1 and 2 during the first half of the task, or objects 3 and 4 for the second half. Across sessions, the trials for each cue was random. B) Response error averaged across the 12 retrieval trials (blue dots seen in panel A) comprising each object-cue block. Shading indicates SEM. Note that subjects learned the object-location association quickly (first two trials), only to decrease in performance upon the introduction of another object location to hold in memory.  Fig. 2F, firing rate here is assessed as a function of time surrounding button press, rather than spatial bin. Notably, these cells do not show significant response period activity at consistent times preceding button press, implying that trace cells were not simply activating at fixed times preceding the button press (ttest by spatial bin, p > 0.05, FDR-corrected C) Pre-and post-response trace-cell firing rate, binned by time as in Panel B. D) Mean firing rate (z-scored) of all trace cells aligned to response time. Shading indicates SEM. No bins show significant difference from baseline. E) Pre-response and post-response firing rate (z-scored) relative to response time. "Ns" indicate non-significance from a t-test (p > 0.05). Arrows denote the start and end of the movement period. Grey shading indicates 5 s around button press. Note that this cell shows an increase in firing rate during the hold period. Increases in activity are also visible during the response period preceding response, with the cell largely ceasing to fire following the response. These encoding and retrieval trials have different durations, which is a result from the differing movement speeds on the two trials. B) Scatter plots illustrating the relations between normalized firing rates in the hold and response periods for six representative trace cells. Black line denotes the robust linear regression fit. C) Hold period firing rate and response period firing rate, for a representative EC trace cell, averaged across all trials for each cue object.  Figure 5: Illustration of the multivariate decoding procedure. Left, top: Training: Pseudopopulations of neural activity were constructed by extracting the activity of cells during the different period of interest. Decoders were trained on this data using a k-nearest-neighbor (KNN) framework to predict the object cue for each trial. Left, bottom: Testing: Response period activity for each trial was computed by normalizing the pre-response firing rate by the post-response firing rate. This measure was extracted for each trial and used as the test set for the decoders trained on each task period. Right: LOOCV decoder using response period. We also trained and tested a decoder using just the response period activity. In order to ensure we had separate train-test data, we used leave-one-out cross-validation (LOOCV). Feature extraction and decoding framework were consistent with those used for the task periods(Left). 1 0 of (6) 0 of (6) 0 of (0) 0 of (0) 0 R1027J 1 1 of (15) 0 of (1) 0 of (0) 0 of (0) 1 R1030J 4 1 of (21) 0 of (1) 0 of (0) 0 of (0) 1 R1092J 3 0 of (0) 0 of (17) 0 of (0) 0 of (0) 0 R1139C 2 0 of (0) 0 of (0) 0 of (0) 2 of (10) 2 R1152C 1 0 of (0) 0 of (2) 0 of (0) 5 of (14) 5 R1182C 1 0 of (0) 0 of (0) 1 of (4) 1 of (3) 2 R1219C 1 0 of (0) 0 of (6) 3 of (14) 0 of (0) 3 R1278E 3 0 of (0) 0 of (0) 20 of (65) 0 of (0) 20 UT048 1 0 of (10) 1 of (2) 0 of (0) 0 of (0) 1 R1268T 1 0 of (0) 0 of (2) 0 of (0) 0 of (0) 0 R1241J 1 1 of (10) 0 of (0) 0 of (0) 0 of (0) 1 R1297T 1 0 of (0) 0 of (3) 0 of (0) 0 of (0) 0 R1299T 1 0 of (0) 1 of (12) 0 of (0) 0 of (0) 1 R1315T 2 0 of (0) 0 of (12) 0 of (0) 0 of (0) 0 EU001 1 0 of (0) 1 of (19) 0 of (0) 0 of (0) 1 R1354E 2 0 of (0) 2 of (14) 0 of (0) 0 of (0) 2 R1362E 3 0 of (0) 0 of (0) 4 of (26) 0 of (0) 4 R1414E 1 0 of (0) 0 of (0) 1 of (4) 0 of (0) 1 Supplementary Table 1: Contribution of subjects and sessions to total cell counts: Table indicates