Abstract
Working memory updating is an important executive process. Here, we study the single-neuron mechanisms involved in updating versus protecting memory from distractors in the macaque prefrontal cortex. We recorded single-neuron activity from the lateral prefrontal cortex (LPFC) and prearcuate cortex (PAC) while male monkeys performed a task that required them to update their memory of target locations while ignoring distractors. Our findings revealed that neurons in the PAC signaled updated memory locations ∼100 ms after stimulus onset, significantly faster than the ∼400 ms observed in the LPFC. Additionally, PAC neurons exhibited longer encoding of distractor information. Population decoding analyses further indicated that distractor information was maintained in orthogonal subspaces from target information in both regions, minimizing interference. These results demonstrate the distinct temporal dynamics in memory updating processes between the PAC and LPFC and highlight the interplay between robust memory maintenance and updating, suggesting that local neural mechanisms may contribute to these processes.
Significance Statement
Working memory is a fundamental cognitive function. It stored information in the short term, and this information can be manipulated to allow intelligent behaviors. The lateral prefrontal cortex is involved in this process, but the mechanisms of working memory manipulation remain unclear. Here, we studied one type of manipulation, working memory updating, which refers to the exchange of one memory for another. We found that two adjacent regions in the lateral prefrontal cortex show different updating times: while a posterior region updates memory content very fast, the more anterior region takes significantly longer. These results show that working memory updating may involve multiple operations, such as updating of memory or attention versus updating of motor plans.
Introduction
Working memory is the ability to maintain and manipulate information derived from past sensory experiences, such as an object's location. Mechanisms that allow the robust maintenance of working memory information are essential since maintaining behaviorally relevant information protected from irrelevant (or distracting) influences is necessary for cognitive operations. However, working memory information may also need to be updated quickly if the task requires it. Updating of working memory has been proposed to be one of three primary central executive processes (Miyake et al., 2000) and a strong predictor of fluid intelligence (Friedman et al., 2006; Chen and Li, 2007).
Functional magnetic resonance imaging and electroencephalography have been used to study the neural mechanisms of working memory updating (D’Ardenne et al., 2012; Nir-Cohen et al., 2020; Yu et al., 2022; Tortajada et al., 2024; Trutti et al., 2024). These studies revealed that the prefrontal cortex, thalamus, and basal ganglia are involved in memory updating (Nir-Cohen et al., 2020). However, thus far, this process has not been studied at the level of single neurons. The lateral prefrontal cortex (area 9/46, LPFC) and the prearcuate cortex (area 8a, PAC) are key regions in the brain network associated with working memory and motor preparation (Fuster and Alexander, 1971; Funahashi et al., 1989; Riley and Constantinidis 2016; Murray et al., 2017; Parthasarathy et al., 2017, 2019; Tang et al., 2020; Panichello and Buschmann, 2021). Therefore, in this study, we measured single-neuron activity in the lateral prefrontal cortex (PAC and LPFC) of macaques trained to perform a task requiring spatial memory updating. We studied the response properties of single neurons and the population of recorded neurons using population decoding and dimensionality reduction.
We found that while single neurons in the PAC quickly signaled the location of the updated memory (∼100 ms after stimulus onset), those in LPFC tended to take longer to update (∼400 ms after stimulus onset). At the same time, distractor information was maintained for longer in the PAC than that in the LPFC. The timing of the delayed updating in LPFC is consistent with the known delays involved in attentional selection of spatial memories, thus suggesting that LPFC activity is related to working memory maintenance and attentional selection, while the faster PAC updating, and the longer distractor encoding, is consistent with a role in the updating of motor preparation.
Materials and Methods
Subjects and surgical procedures
Two adult male macaques (Macaca fascicularis) were used in this experiment: Monkey A (age 12) and Monkey B (age 12). All animal procedures were approved by, and conducted in compliance with, the standard of the National University of Singapore Institutional Animal Care and Use Committee (NUS IACUC #R18-0295). Procedures also conformed to the recommendations described in Guidelines for the Care and Use of Mammals in Neuroscience and Behavioral Research. Each animal was first implanted with a titanium head-post (Crist Instrument) before arrays of intracortical microelectrodes (Microprobes) were implanted in multiple regions of the left frontal cortex. In Monkey A, two arrays were placed over the dorsal and ventral aspect of the LPFC (Area 9/46), one array was placed over PAC (Area 8A), and one array was placed over the pMA (not included in the analyses here), with 32 electrodes in each. In Monkey B, two arrays of 32 electrodes were placed over the PAC, and two arrays of 32 electrodes each were placed over the LPFC (as shown in Fig. 1b). The arrays consisted of platinum-iridium wires with 200 µm separation, 1–5.5 mm long, and with 0.5 MΩ of impedance, arranged in 8 × 4 grids. For arrays positioned in the prearcuate region (PAC), we did not conduct electrical microstimulation to confirm that saccades could be elicited with low currents; thus, we are unable to confirm that these electrodes are located in the frontal eye field.
Recording techniques
The neural signals in both monkeys were recorded using a 128-channel Grapevine recording system (Ripple Neuro) at 30 kHz sampling rate. The wideband signals were bandpass-filtered between 300 and 3,000 Hz, and spikes were detected on each channel separately using an automated sorting algorithm based on hidden Markov modeling (Herbst et al., 2008). We recorded the eye positions of each subject using the EyeLink 2000 (SR Research) on another standalone computer. We designed and ran the behavioral tasks using the PsychoPy in Python (Peirce et al., 2019) on a third computer connected to the recording computer using parallel ports for event mark synchronization during recording.
Behavioral tasks
Both monkeys performed a two-item delayed saccade task. Each trial started with a mandatory fixation interval (500 ms) during which the animal gazed at a white dot at the center of the screen. A first stimulus (Item 1) was presented at one of four possible locations (Fig. 1A). Each location was located at 10° of visual angle from the fixation. Item 1 was presented for a fixed interval (Monkey A, 400 ms; Monkey B, 300 ms), followed by a delay of 1,000 ms (Delay 1) during which the animal was expected to remain fixated on the screen center. At the end of Delay 1, a second stimulus (Item 2) was presented at one of the three remaining locations that were different from the first target location for the same length as Item 1 (400 or 300 ms). For each trial, Item 1 was always a task-relevant target (red square), while Item 2 could be either a task-irrelevant distractor (T/D trials) or a new target (T/T trials; 50% chance for each). The target and distractor stimuli were identical in shape and size except for their colors (target, red; distractor, green). Item 2 was again followed by a 1,000 ms delay (Delay 2). At the end of Delay 2, the center fixation dot was removed, signaling the animal to give a saccadic response to the location of the most recent target presented. In other words, the animal needed to make an eye movement to the location of Item 1 when Item 2 was a distractor or to the location of Item 2 when it was a new target. A juice reward was only delivered to the animal when a correct response was made within a 500 ms post-go-cue window (i.e., the removal of center fixation) and if the animal fixated on the correct target for longer than 200 ms. Item 1 and Item 2 were never repeated at the same location in a trial. In total, 24 conditions were available from the combinations of Item 1 × Item 2 locations (4 × 3) and by Item 2 types (2).
For each correct trial, the amount of juice delivered varied as a function of the distance between the saccade end point and the center of the target. The target size is 3° of visual angle. If the saccade endpoint was further than 8° from the target center, there was no reward given. At 8° the reward solenoid opened for 250 ms. This length was linearly increased up to 800 ms at the center of the target. On different sessions animals drank between 0.5 and 1.5 lts, depending on the number of trials performed.
Neuron firing rates and preprocessing
The single-neuron firing rate was calculated using a sliding 50 ms window with 10 ms overlap. We converted the firing rate to a z-score by normalizing to the mean and standard deviation of the instantaneous firing rates from a 200 ms baseline interval prior to Target 1 onset across all trials. The normalization method was adapted from Parthasarathy et al. (2019). Note that the duration of Item 1 and Item 2 differed between Monkeys A (400 ms) and B (300 ms). For displays of data pooled across both monkeys (Fig. 4, mean tuning; Figs. 5, 6, LDA decoding), we removed 100 ms of each delay period from Monkey B to match the length of time series and align the onsets of both Item 1 and Item 2 (the removed period was in the middle of the delay: 450–550 ms of each delay period).
Neuron selectivity
For each neuron, we calculated the mean firing rates separately within epochs of Item 2 (Monkey A, 1,400–1,800 ms; Monkey B, 1,300–1,600 ms), and the second half of Delay 2 (Monkey A, 2,300–2,800 ms; Monkey B, 2,100–2,600 ms), and applied a Nlocations (4) × Ntypes (2) factorial ANOVA. Neurons with significant main effect(s) or interaction were categorized as selective to the corresponding stimulus feature(s) or combinations.
Then, we defined the preferred spatial location as the location of Item 1 corresponding to the highest mean firing rates during the Target 1 and Delay 1 (0–1,300 ms). We then grouped trials based on whether the stimuli were presented in the preferred location of the neuron (Item 1 IN, Item 2 IN, or None IN) as well as by Item 2 type (Target 2 or Distractor), giving us six tuning conditions (Fig. 4A,B). Within each tuning condition, we compared the firing rates at each time point to the baseline (the average activity over the 200 ms period prior to the onset of Target 1 across all trials) using a t test. To determine the response latency of single cells (Fig. 4C) we identified the earliest time point with significant (t test; p < 0.001) above-baseline activity for at least 100 ms continuously after the onset of targets. To determine the latency of the return to baseline (Fig. 4D), we calculated the time taken for activity associated with Target 1 to return to baseline after Target 2 was presented by identifying the earliest time point after the onset of Target 2 with nonsignificant (p > 0.05) difference from baseline for at least 100 ms. This was done on neurons that showed significant (p < 0.001) above-baseline activity for at least 100 ms during the 200 ms preceding Target 2 onset. Significance of timing differences between LPFC and PAC were examined by t test.
Decoding analysis
To measure the decodable stimulus location information across the population of recorded neurons, we used linear discriminant analysis (LDA) based on the algorithm from scikit-learn (Pedregosa et al., 2011) to predict the locations of Items 1 and 2 (Figs. 5, 6). We created pseudo populations by sampling equal numbers of trials per condition from each session of both monkeys and pooling these trials together, as if these neurons responded together in each pseudo trial. We created 100 pseudo populations separately for the LPFC (89 neurons) and the PAC (129 neurons). Only correct trials were included in the pseudo population. Decoding analyses were applied to the interval from −200 ms before Target 1 to the delivery of go cue. For smoothing, we averaged the time series with a 50 ms nonoverlapping time window before training and testing the decoders. We also applied principal component analysis (PCA) to each pseudo population to reduce the dimensionality across neurons before decoding (see below). Decoding analyses were conducted based on the activities projected onto the PC axes.
Within each pseudo session, we randomly sampled half of the trials as the training set and the other half as the test set. We trained a decoder with the training set PC activities at each time point to predict the corresponding locations of items and then tested the decoder performance with the test set activities at the same time point. Each decoder was bootstrapped 100 times. We iterated this process until all time bins had been trained and tested. Within each pseudo population, we examined the significance of decodability at each time bin using one-sample t test compared with chance-level accuracy. Note that, since Item 1 and Item 2 could never repeat on the same location in each trial, we used a chance-level performance to 33% for the significance tests.
In addition to the decoder performance, we calculated the latency of above-chance decodability (Fig. 5C) in LPFC and PAC. We identified the earliest time bin after the target onset with significant (p < 0.001) above-chance decodability for at least 100 ms continuously. In addition, we calculated the time of return to chance decodability (Fig. 5D) by identifying the earliest time bin after the Target 2 onset with nonsignificant (p > 0.05) decodability of Target 1 location for at least 100 ms continuously. We only included pseudo sessions with significant (p < 0.001) decodability for at least 100 ms during the 200 ms preceding Target 2 onset. Significance of timing differences between LPFC and PAC was examined by t test.
Finally, to measure the decodable distractor location information across the population of recorded neurons, we trained decoders to predict distractor location during distractor presentation and Delay 2. We estimated the duration of distractor presence by counting the number of time bins with significant above-chance decodability (p < 0.001). Differences between LPFC and PAC were examined by t test.
Principal component analysis
We applied principal component analysis (PCA) to each pseudo population to reduce the dimensionality before decoding, using the PCA algorithm from the scikit-learn library in Python (Pedregosa et al., 2011). To preserve the variations between conditions, we first calculated mean firing rates in each condition and averaged across the full Delay 1 and Delay 2 intervals. This returned an activity matrix X of size Nconditions (24) × Nneurons. We fitted a PCA model to this matrix and then applied the transformation to the full time series in each pseudo trial so that the original pseudo population states were projected into the PC space. For each pseudo population, we defined the full space of the transformed data as the projections on the first 15 PCs, which counted for at least 90% of the total explained variance (EVR) on average in both regions (mean EVR: LPFC = 0.911, PAC = 0.916). We further divided these PCs into two sets for subspaces analysis: a two-dimensional subspace consisted of PCs 1 and 2 (mean EVR: LPFC = 0.450, PAC = 0.540) and an orthogonal subspace consisted of the rest of PCs (mean EVR: LPFC = 0.461, PAC = 0.370).
Results
We trained two monkeys to perform a delayed saccade memory updating task (Fig. 1A). In this task, monkeys had to report the location of the last target seen (red square) with a saccadic eye movement. Two types of trials were randomly intermixed: Target 1/Distractor (T/D) trials and Target 1/Target 2 (T/T) trials (Fig. 1A). After an initial fixation period (500 ms), they were shown a visual stimulus (Item 1) for 400 ms (Monkey A) or 300 ms (Monkey B). Item 1 was a target (Target 1) in both T/D and T/T trial types. After a 1,000 ms delay period, a second visual stimulus (Item 2) was shown for 400 ms (Monkey A) or 300 ms (Monkey B). In T/D trials, Item 2 was a distractor (green square), while in T/T trials, Item 2 was a new target (Target 2). After a second 1,000 ms delay period, the fixation dot disappeared. This served as a go-cue to initiate the response. For T/D trials, monkeys had to report the location of Target 1, while for T/T trials, monkeys had to report the location of Target 2. The overall performance of the monkeys was higher than 50% (Monkey A: overall M = 53.271%, STDV = 1.148%, T/D M = 50.619%, STDV = 1.807%, T/T M = 55.881%, STDV = 0.834% and Monkey B: overall M = 56.303%, STDV = 3.637%, T/D M = 56.272%, STDV = 4.061%, T/T M = 56.947%, STDV = 6.509%). This is a conservative estimate of performance, since we are excluding trails in which the animals made a saccade to the correct target location but were either too slow (>500 ms) or failed to maintain fixation on the correct target location for longer than 200 ms (combined: T/D M = 20.3%, STDV = 10.9%, T/T M = 15.8%, STDV = 12%). These types of errors, however, do not reflect a lack of understanding of the task, but rather a failure of attention (for responses slower than 500 ms) or a failure of self-control (for when the animals did not maintain fixation on the target for >200 ms). If we recalculate the performance of the animals recategorizing these trials as successful, the overall performance would be as follows: Monkey A: T/D M = 68.8%, STDV = 2.2%, T/T M = 71.1%, STDV = 3.2%; Monkey B: T/D M = 62.6%, STDV = 3.4, T/T M = 61.1%, STDV = 6.3%. For subsequent neural data analysis, we only included correct trials that were rewarded (i.e., when the animals responded faster than 500 ms and fixated on the correct target for longer than 200 ms).
Task and recording areas. A, Task performed by monkeys. After a 500 ms fixation period, an Item 1 period always contained a target (Target 1, red square), which was presented for either 400 ms (Monkey A) or 300 ms (Monkey B). After Item 1 disappeared, there was a 1,000 ms Delay 1 period followed by an Item 2 period. For T/D trial types, Item 2 was a distractor (green square), while for T/T trial types, Item 2 was a new target (Target 2, red square). Item 2 was presented for either 400 ms (Monkey A) or 300 ms (Monkey B). After Item 2 disappeared, there was a 1,000 ms Delay 2 period followed by the disappearance of the fixation spot, which served as a go cue. To get a reward (drop of juice), monkeys needed to saccade to the location of the last target shown (Target 1 in T/D trials and Target 2 in T/T trials). B, Location of implanted electrodes in the lateral prefrontal cortex. We chronically implanted 64 electrodes in the LPFC of each monkey and either 32 (Monkey A) or 64 (Monkey B) electrodes in the PAC.
We recorded a total of 89 neurons in LPFC (Monkey A, 17; Monkey B, 72) and 129 neurons in PAC (Monkey A, 3; Monkey B, 126) across 11 sessions (Monkey A, 3; Monkey B, 8). To characterize the selectivity of single neurons, we performed a two-way ANOVA on the neuronal firing rates between task type (T/D and T/T) × target location (four locations). During the Item 2 period, ∼30% of neurons showed selectivity in LPFC and ∼70% in PAC (Fig. 2A). While the majority of selective neurons in LPFC (64% of selective cells) showed location selectivity independent of the trial type, which means that these neurons responded similarly to Item 2 when it was target or distractor (Fig. 2B), the majority of selective neurons in PAC (53% of selective cells) showed an interaction between location and type, which means that these neurons responded differently to Item 2 when it was target vs a distractor (Fig. 2C,D). For example, some PAC neurons responded selectively to Item 2 only when it was a target but not if it was a distractor (Fig. 2C), while others responded differently to targets and distractors (Fig. 2D). This observation is consistent with the “filtering” of distractor information in these PAC neurons (Cosman et al., 2018). However, PAC also had a large proportion (33% of selective neurons) with location preference independent of trial type (Fig. 2A). These neurons may receive prefiltered inputs from visual areas, suggesting that the filtering occurs locally and quickly in the PAC, for example, by combining the activity of neurons with location and type selective neurons.
Item 2 period single cell selectivity. A, Proportion of neurons in each region with different selectivities during the Item 2 period (Target 2 in T/T trials and Distractor in T/D trials). Type refers to trial type (T/T or T/D), Loc. refers to the location of Item 2, Loc. + Type refers to neurons selective to both but without interaction, Loc. * Type refers to neurons that had a significant interaction. B–D, Example activity of single neurons (z-scored firing rate) aligned to the onset of Item 2 (mean ± standard deviation). Significance bars on top of the plots show periods during which Target 2 differs from Distractor responses (paired t test, p < 0.01), and significance bars on the bottom of the plots show periods where Target 2 (dark lines) or Distractor (light lines) differ from baseline activity (−200 to 0 ms before Item 2 onset; paired t test, p < 0.01). The selectivity of the cell is described in the bottom-right corner.
During the Delay 2 period (the last 500 ms of Delay 2), most neurons across both regions showed a significant interaction between location and type (Fig. 3A). Some of these cells maintained a stable delay period activity starting from Item 2 presentation (Fig. 3B), while others developed the selectivity over time (Fig. 3C). Compared with the Item 2 period selectivity, during Delay 2, there were more neurons with location-by-type interactions and fewer with location-only selectivity, implying that the distractor filtering process develops over time.
Delay 2 period single cell selectivity. A, Proportion of neurons in each region with different selectivities during the Delay 2 period (last 500 ms of Delay 2; Target 2 in T/T trials and Distractor in T/D trials). Type refers to trial type (T/T or T/D), Loc. refers to the location of Item 2, Loc. + Type refers to neurons selective to both but without interaction, Loc. * Type refers to neurons that had a significant interaction. B–C, Example activity of single neurons (z-scored firing rate) aligned to the onset of Item 2 (mean + standard deviation). Significance bars on top of the plots show periods during which Target 2 differs from Distractor responses (paired t test, p < 0.01), and significance bars on the bottom of the plots show periods where Target 2 (dark lines) or Distractor (light lines) differ from baseline activity (−200 to 0 ms before Item 2 onset; paired t test, p < 0.01). The selectivity of the cell is described in the bottom-right corner.
The averaged mean activity of all selective neurons in LPFC and PAC reveals differences in the temporal dynamics of activity between the regions (Fig. 4). While the activity associated with Target 2 increases quickly in PAC, the activity in LPFC takes longer to increase (Fig. 4A–C; Target 2 response latency; LPFC (9): M = 498 ms, SEM = 128 ms; PAC (37): M = 179 ms, SEM = 21 ms; LPFC-PAC: M = 320 ms, p < 0.001). At the same time, the elevation of activity associated with Target 1 in T/T drops abruptly in PAC after Target 2 is presented, but it takes longer to fall to baseline in LPFC (Fig. 4A,B,D; return-to-baseline latency; LPFC (8): M = 503 ms, SEM = 114 ms; PAC (21): M = 160 ms, SEM = 31 ms; LPFC - PAC: M = 343 ms, p < 0.001). It is worth noting that the average LPFC plot (Fig. 4A) shows a brief (<50 ms) increase in activity shortly after Item 2 onset, which was not captured by our latency criteria of at least 100 ms increase (Fig. 4C). This brief increase is also reflected in a brief period of Item 2 decodability observed in upcoming decoding analyses (Figs. 5A, 6A). We are unsure whether this brief period of activity has functional relevance, so henceforth we focus our interpretations on the later, but more sustained, period of elevated activity. With this caveat in mind, these observations suggest that information updating has different time courses in both regions.
Average activity. Mean z-scored activity across all LPFC (A) and PAC (B) neurons. Red traces represent trials in which Item 1 (Target 1) was located inside the neuron’s preferred location, defined as the location with the highest mean firing rate during Item 1 and Delay 1 period (0–1,300 ms after Item 1 onset). Green traces represent trials in which Item 2 (Target 2 or Distractor) was located inside the neuron’s preferred location. Blue traces represent trials in which neither Item 1 nor Item 2 were located inside the neuron’s preferred location. Continuous lines represent T/T trials and dotted lines T/D trials. Significance bars on the bottom of the plots show periods during which activity is significantly above baseline (−200 to 0 ms before Item 2 onset; paired t test, p < 0.001). Thick lines for T/T and thin lines for T/D trials. C, Response latency when Item 1 (dark) or Item 2 (light) was located inside each neuron’s preferred location. Response latency was calculated as the earliest time point with significant (p < 0.001) above-baseline (200–0 ms before Item 1 onset) activity for >100 ms. D, Return-to-baseline latency after Target 2 in T/T trials. Return-to-baseline latency was calculated as the earliest time point where the activity was not significantly above baseline (p > 0.05 compared with baseline for >100 ms). For this analysis we only included neurons with significant above-baseline activity for >200 ms before Target 2 onset. For C and D **p < 0.01, ***p < 0.001, dashed white vertical lines show the mean, triangle the median, and circles the outliers. Single monkey results can be found in Extended Data Figure 4-1.
Figure 4-1
Analyses on single monkeys. Conventions are the same as those in Figure 3. Download Figure 4-1, TIF file.
Decoding full space. Performance of a decoder (LDA) trained to predict the location of Item 1 (Target 1, blue traces) or Item 2 (Target 2 or Distractor, purple traces) in T/T trials (left) and T/D trials (right) using the full population of LPFC neurons (A) or PAC neurons (B). Significance bars on the top of the plots show periods during which decoding performance is significantly above chance (where chance is 33% since Target 2 was never in the same location as Target 1; paired t test, p < 0.001). C, Above-chance decodability latency for Target 1 (dark) or Target 2 (light) location. Above-chance decodability latency was calculated as the earliest time point with significant (p < 0.001) above-chance decodability for >100 ms. D, Return-to-chance decodability latency after Target 2 in T/T trials. Return-to-chance decodability latency was calculated as the earliest time point where Target 1 information was not significantly decodable after Target 2 onset (p > 0.05 for >100 ms). For this analysis, we only included iterations with significant above-chance decoding for >200 ms before Target 2 onset. For C and D, ***p < 0.001, dashed white vertical lines show the mean, triangle the median, and circles the outliers.
Decoding PCs. Performance of a decoder (LDA) trained to predict the location of Item 1 (Target 1, blue traces) or Item 2 (Target 2 or Distractor, purple traces) in T/T trials (left) and T/D trials (right) using either PCs 1 and 2 (top plots) or PCs 3–15 (bottom plots) of LPFC neurons (A) or PAC neurons (B). Significance bars on the top of the plots show periods during which decoding performance is significantly above chance (chance = 33%; paired t test, p < 0.001). Single PC decoding can be found in Extended Data Figure 6-1.
Figure 6-1
Decoding in LPFC and PAC using single principal components. Same conventions as in Figure 5A. Download Figure 6-1, TIF file.
To explicitly quantify the time courses of information updating, we decoded single-trial information of Target 1 or Target 2 location (Fig. 5). Consistent with the averaged activity timings shown in Figure 4, we found that while information associated with Target 2 increases quickly in PAC, the information in LPFC takes longer to increase (Fig. 5A–C, Target 2 above-chance decodability timing; LPFC: M(100) = 962 ms, SEM = 25 ms; PAC(100): M = 100 ms, SEM = 1 ms; LPFC-PAC: M = 862 ms, t(99) = 34.8, p < 0.001). Furthermore, the elevation of information associated with Target 1 in T/T drops abruptly in PAC after Target 2 is presented, but it takes longer to fall in LPFC (Fig. 5A,B,D; return-to-chance decodability timing; LPFC(99): M = 385 ms, SEM = 22 ms; PAC M = 116 ms, SEM = 4 ms; LPFC - PAC: M = 269 ms, t(99) = 12.2, p < 0.001).
The decoding analysis revealed that distractor information emerged briefly after the distractor presentation in PAC (Fig. 5B, right). If distractor information is encoded in the same activity subspace as the targets, it would interfere with target information. On the other hand, if target and distractor information are encoded in orthogonal subspaces, information about both can be maintained without interference. To determine whether target and distractor information are maintained in the same or in orthogonal activity subspaces, we ran a PCA analysis and analyzed target and distractor information contained in different PCA components (Extended Data Fig. 6-1). In both the PAC and LPFC, the first 2 PCs contained target information but no distractor information, while the rest of the PCs (combined) contained significant distractor information (Fig. 6). The only difference was that PAC distractor information lasted longer (Fig. 6; LPFC: M = 90 ms, SEM = 4.5 ms; PAC: M = 387 ms, SEM = 8.9 ms; LPFC-PAC: M = −298, t(99) = −29.7, p < 0.001). This analysis confirms that information about the target and distractor are maintained in orthogonal subspaces, thus minimizing the interference between target and distractor information.
Discussion
Here, we studied the activity associated with memory updating at the single-neuron level in the primate prefrontal cortex. Our main finding is that memory updating occurs faster in the PAC than that in the LPFC. Single-neuron analyses and population decoding give slightly different estimates of these timings. If we take population decoding as a more reliable estimate, we found that information updating in the PAC occurred earlier than 150 ms after Target 2 onset (100 ms for new information to emerge and 116 ms for old information to disappear), while in LPFC updating took longer (962 ms for new information to emerge and 385 ms for old information to disappear). The PAC plays a crucial role in motor preparation and execution (Bruce and Goldberg, 1985), and a recent study showed that PAC is involved in motor planning but not in working memory maintenance (Jonikaitis et al., 2023). On the other hand, the LPFC plays a known role in working memory maintenance (Funahashi et al., 1993; Panichello and Buschman, 2021). Our finding that updating occurs hundreds of milliseconds earlier in PAC than that in LPFC suggests that while the animals may be able to report the updated location very quickly if necessary, updating the working memory could take longer. Our results help explain behavioral observations in humans. For example, human subjects take ∼500 ms to use a retro-cue to select stored working memories and remove unnecessary memories (Tortajada et al., 2024), consistent with the updating delays in the LPFC. However, in tasks that require updating a single item, updating is very fast and automatic (Kessler et al., 2023), which is consistent with the updating delays in the PAC.
The tension between reliably maintaining a memory (and protecting it from distractors) and the ability to update it has been hypothesized to involve a gating mechanism that regulates the flow of information into working memory (McNab and Klingberg, 2008; Chatham and Badre, 2015). However, the specific role of subcortical gating mechanisms is debated (Trutti et al., 2024). Our observations that different neurons in prefrontal regions show unfiltered and filtered responses to distractors, as well as a transition over time to have more filtered responses (more Target 2 and less Distractor information), suggest that in addition to subcortical gating mechanisms, local mechanisms may also contribute to the filtering and updating of information.
A conceivable alternative to distractor filtering is that PAC shows a more robust representation of target information, which would in turn prevent distractor information from emerging (e.g., if the attractor basin encoding the target was deeper). We believe that alternative explanation of this is unlikely, for two reasons. First, if the robust representation in PAC leads to a smaller distractor response in T/D trials, then it should also lead to a smaller or more delayed response to Target 2 in T/T trials, which is not what we observe (Figs. 4 and 5). Second, even though PAC does show larger changes in activity associated with the encoded targets, it also shows more information associated with the distractor (Fig. 5), which would not be expected from this alternative explanation. As such, our results are not consistent with this alternative explanation, but rather suggest that these regions are engaging in distractor filtering.
Distractor suppression has been observed previously in PAC neurons (Lennert and Martinez-Trujillo, 2011; Suzuki and Gottlieb, 2013; Jacob and Nieder, 2014; Qi et al., 2015; Cosman et al., 2018). Here, we extend the results presented in these studies in three ways. First, we show that while the stimulus (target or distractor) is presented to the animals, both LPFC and PAC contain neurons that respond similarly to targets and distractors, cells that respond differently to targets and distractors, and among them, cells that only respond to targets, and not to distractors (Fig. 2). During the delay period, this activity primarily transitions to encoding target information (Fig. 3). Second, we show that both regions encode distractor location in a subspace orthogonal to the one used to encode target information, thus protecting target information from the influence of distractors (Fig. 6). Third, we show that the PAC encodes distractor information for a longer period compared with the LPFC, suggesting that distractor influences on motor preparation may be more pronounced than in working memory maintenance (if we assume that PAC is primarily involved in motor preparation and LPFC in working memory maintenance). When human subjects perform a task that requires either updating a memory or ignoring a distractor, they respond faster to the updated item than when they report the original memory after a distractor (Kessler et al., 2023). They interpreted this difference as a slowdown of responses due to the need to remove unnecessary memories. Our results are consistent with this interpretation, since PAC shows fast updating, but distractor information persists in PAC beyond this updating, possibly slowing down the response.
Footnotes
We thank Shih-Cheng Yen for useful discussions and Clement Lim for his help with the experiments.
Ministry of Education - Singapore (MOE-T2EP30121-0010, MOE2017-T3-1-002).
The authors declare no competing financial interests
- Correspondence should be addressed to Camilo Libedinsky at camilo{at}nus.edu.sg.
