A Quantitative Analysis of Context-Dependent Remapping of Medial Frontal Cortex Neurons and Ensembles

Abstract

The frontal cortex has been implicated in a number of cognitive and motivational processes, but understanding how individual neurons contribute to these processes is particularly challenging as they respond to a broad array of events (multiplexing) in a manner that can be dynamically modulated by the task context, i.e., adaptive coding (Duncan, 2001). Fundamental questions remain, such as how the flexibility gained through these mechanisms is balanced by the need for consistency and how the ensembles of neurons are coherently shaped by task demands. In the present study, ensembles of medial frontal cortex neurons were recorded from rats trained to perform three different operant actions either in two different sequences or two different physical environments. Single neurons exhibited diverse mixtures of responsivity to each of the three actions and these mixtures were abruptly altered by context/sequence switches. Remarkably, the overall responsivity of the population remained highly consistent both within and between context/sequences because the gains versus losses were tightly balanced across neurons and across the three actions. These data are consistent with a reallocation mixture model in which individual neurons express unique mixtures of selectivity for different actions that become reallocated as task conditions change. However, because the allocations and reallocations are so well balanced across neurons, the population maintains a low but highly consistent response to all actions. The frontal cortex may therefore balance consistency with flexibility by having ensembles respond in a fixed way to task-relevant actions while abruptly reconfiguring single neurons to encode "actions in context."

Significance statement: Flexible modes of behavior involve performance of similar actions in contextually relevant ways. The present study quantified the changes in how rat medial frontal cortex neurons respond to the same actions when performed in different task contexts (sequences or environments). Most neurons altered the mixture of actions they were responsive to in different contexts or sequences. Nevertheless, the responsivity profile of the ensemble remained fixed as did the ability of the ensemble to differentiate between the three actions. These mechanisms may help to contextualize the manner in which common events are represented across different situations.

Keywords: behavioral flexibility; ensemble analysis; medial prefrontal cortex.

Figure 1.

The effect of context change on action representations by ACC ensembles. A, Left, The apparatus for the maze task was diamond shaped with three unique platforms on three of the tips, each containing a unique manipulandum. The starting or reward platform was the fourth tip. The lights above each manipulandum provided the rat with the cue to select the correct direction to circuit the maze. Upon the completion of each action, the light above the manipulandum on the next platform in the sequence was illuminated, indicating that that manipulandum was now active. Right, Schematic of the operant-box apparatus. Rats responded on the same manipulanda in the same order as on the maze, also following the guidance of cue lights. B, Left, A schematic drawing outlining the context-switch task: the animals performed several trials in the maze (Context 1) before being switched to the box (Context 2) and performed more trials involving the same actions in the same sequential order. Right, Schematic outline for the sequence-switch task: several trials in the WT → LP → NP order (Sequence 1) were completed before the animals had to switch to perform the same actions in the reverse order on the maze: NP → LP → WT (Sequence 2). C, An example MSUA space constructed from the iFRs of 46 ACC neurons recorded during a context-switch session where the rat performed the NP → LP → WT sequence first in the maze (Context 1) and subsequently in the operant box (Context 2). The full space of all 46 dimensions—one for each neuron—was reduced to three dimensions using multidimensional scaling for visualization. Each dot is a population vector representing the activities of the entire ensemble during the 1 s period when the rat produced the correct responses. Dots are colored red if they were time bins associated with NPs, yellow if they were time bins associated with LPs, and blue if they were time bins associated with WTs. The combined activity states associated with all three actions in the first context (black dots, left panel) was distinct from the combined state of the same three actions in the second context (gray dots). D, For all context-switch sessions, the distances (D_Mah in MSUA space) between the activity states associated with the same operant action in Context 1 versus Context 2 were calculated in the full space, and compared with the distances for shuffled blocks (5 time bins/block) within each context. The average separation between different contexts (black bars) was significantly larger than the within-context control distances (white bars). E, The activity states associated with NPs, LPs, and WTs in Context 1 shifted by the same distance (D_Mah in MSUA space) following the switch in context. F, The level of performance remained equivalent whether the animal was in the first or the second context. G, On average the D_Mah between the activity-state clusters associated with the first and the second actions was the same in both contexts. H, Given the difference in the size of the maze and the box, the time between the first and second actions was shorter in the box than in the maze (black bars). This was not reflected as a difference in D_Mah as shown in G. Additionally, the animals tended to spend less time between the second and the third actions than between the first and the second actions, especially in the maze (left bars). I, The D_Mah in the MSUA space between the clusters associated with the first and the second actions was nevertheless the same as that between the second and the third actions on the maze. J, The average level of activity across the ensembles remained unchanged from the first to the second context. Error bars indicate the standard error of the mean. *p < 0.05.

Figure 2.

Comparing action representations across two sequences. A, An example of a reduced MSUA space constructed from the iFRs of 58 ACC neurons recorded during the first part of a single sequence-switch session, where the WT → LP → NP sequence (Sequence 1) was rewarded in the first half of the session while the NP → LP → WT sequence (Sequence 2) was rewarded in the second half. The combined activity states associated with all three actions in the first sequence (black dots, left panel) were distinct from the combined state of the same three actions in the second sequence (gray dots). Upper right panel breaks down the cluster into the three distinct action representations in the first sequence while the lower right panel shows the three distinct action representations in the second sequence. B, For all sequence-switch sessions, the D_Mahs in MSUA space between the activity states associated with the same operant action in Sequence 1 versus Sequence 2 were calculated in the full space, and compared with the distances for shuffled blocks (5 time bins/block) within each sequence. The average separation between different sequences (black bars) was significantly larger than the within-sequence control distances (white bars). C, The activity states associated with NPs, LPs, and WTs in Sequence 1 all shifted by the same distance (D_Mah in MSUA space) following the sequence switch. D, The level of performance remained unchanged from the first to the second sequence. E, The animals tended to take less time to traverse between the second and the third actions than between the first and the second actions, especially in the first sequence (left bars). F, The D_Mah in the MSUA space between the clusters associated with the first and the second actions was nevertheless the same as that between the second and the third actions in the first sequence. G, The average level of activity across the ensembles remained unchanged from the first to the second sequence. Error bars indicate the standard error of the mean. *p < 0.05.

Figure 3.

Examples of the various responses of individual neurons to actions performed across contexts on the context-switch sessions. A, A neuron that responded similarly when the rat performed WTs in the two contexts. B, A different neuron that responded to NPs in Context 1 but not 2. C, A third neuron that responded to LPs but only in the second context. In each case, raster plots are shown in the top panels and peristimulus time histograms in the bottom panels. The action occurred at time 0.

Figure 4.

Quantification of the neuronal responses to actions during the context-switch sessions. A, Raw spike counts for all neurons recorded during eight context-switch sessions. Each dash represents the spike count of a neuron in the period surrounding an action. A case is the sum across all 26 trials for one neuron for one of the three actions (444 neurons × 3 actions = 1332 cases). The sorting order was the same in A and E and was based on the difference in RT counts of the neurons across trials for the same action in the two contexts. The gray dotted line separates the 13 trials in Context 1 from the 13 trials in Context 2. B, PCA was performed on the matrix of raw spike counts in A. PC1 exhibited an abrupt transition between the last trial of Context 1 and the first trial of Context 2. PC2–PC6 are shown in the inset. C, Distribution of counts of cases in which raw spike counts changed across trials in Context 1 versus 2. Cases losing spikes are on the left of the distribution (n = 573) while cases gaining spikes (n = 584) are on the right. The gains and losses were not significantly different, indicating that the context changes were highly balanced across the population. D, Pie chart illustrating the relative proportions of active trials (trials with nonzero spike counts) in each context versus the proportion of nonactive trials (trials with zero spike counts). E, Distribution of RTs plotted in an identical manner to A. Each trial was classified as an RT or an NRT based on the difference in spike counts for the action periods relative to the baseline periods (data not shown) for each neuron. F, When PCA was performed on the RT matrix in E, PC1 exhibited an abrupt transition just after the context switch. PC2–PC6 are shown in the inset. G, Distribution of counts of cases in which RT counts changed across trials in Context 1 versus 2. Cases losing RTs are on the left of the distribution (n = 256) while cases gaining RTs (n = 228) are on the right. The gains and losses were not significantly different. H, Pie chart illustrating the relative proportions of RTs in each context versus the proportion of NRTs.

Figure 5.

Cross-context changes in RTs for the three actions. A, Top, Distributions of RTs for each of the three actions across the population. Left, RTs for NPs. Middle, RTs for LPs. Right, RTs for WTs. The sorting order was identical to that used in Figure 4 and was maintained for all three panels. Bottom, RT probability or the average RT count/trial across all neurons. While <10% of neurons exhibited RTs on any given trial, the rate was consistent across trials in both contexts. The vertical gray dotted lines in each panel denote the context switch point, whereas the single horizontal dotted line separates the population of neurons in half. B, Pie charts illustrating the relative proportions of RTs (black) and NRTs (gray) for each half of the population in each context. Note that neurons in the top half exhibited relatively more RTs in Context 2 than 1 whereas those in the bottom exhibited more RTs in Context 2 than 1. C, Pie charts illustrating the distribution of RTs by action type for the top and bottom halves of the population. The color key can be found at the bottom of B. Each half of the population exhibited different proportions of RTs for NP1, LP1, and WT1 from Context 1, and for NP2, LP2 and WT2 from Context 2. D, When the two halves of the population were recombined, the proportions of RTs dedicated to each action in each context were uniform. E, Pie charts showing the distributions of RTs for the three individual neurons whose rasters were shown in Figure 3. The neuron number corresponds to its position in A. While RT allocations were consistent with what was illustrated in the rasters, individual neurons may maintain, lose, or gain responses to actions in different proportions across the two contexts.

Figure 6.

Examples of the various responses of individual neurons to actions performed across sequences on the sequence-switch sessions. A, A neuron that responded similarly when the rat performed WTs in the two sequences. B, A different neuron that responded to NPs in Sequence 1 but not 2. C, A third neuron that responded to LPs but only in the second sequence. In each case, raster plots are shown in the top panels and peristimulus time histograms in the bottom panels. The action occurred at time 0.

Figure 7.

Quantification of the neuronal responses to actions during the sequence-switch sessions. A, Raw spike counts for all neurons recorded during eight sequence-switch sessions. Each dash represents the spike count of a neuron in the period surrounding an action. A case is the sum across all 26 trials for one neuron for one of the three actions (398 neurons × 3 actions = 1194 cases). The sorting order was the same in A and E and was based on the difference in RT counts of the neuron across trials for the same action in the two contexts. The gray dotted line separates the 13 trials in Sequence 1 from the 13 trials in Sequence 2. B, PC1 of the raw spike-count matrix in A was characterized by an abrupt transition between the last trial of Sequence 1 and the first trial of Sequence 2. PC2–PC6 are shown in the inset. C, Distribution of counts of cases in which raw spike counts changed across trials across sequences. D, Pie chart illustrating the relative proportions of active trials versus nonactive trials. E, Distribution of RTs plotted as in A. F, PC1 of the RT matrix exhibited an abrupt transition just after the sequence switch. PC2–PC6 are shown in the inset. G, Distribution of counts of cases in which RT counts changed across sequences. The gains and losses were not significantly different. H, Pie chart illustrating the relative proportions of RTs in each sequence versus the proportion of NRTs.

Figure 8.

Cross-sequence changes in RTs for the three actions. A, Top, Distributions of RTs for each of the three actions across the population. Left, RTs for NPs. Middle, RTs for LPs. Right, RTs for WTs. The sorting order was identical to that used in Figure 7 and was maintained for all three panels. Bottom, RT probability or the average RT count/trial across all neurons. The vertical gray dotted lines in each panel denote the context switch point, whereas the single vertical dotted line separates the population of neurons in half. B, Pie charts illustrating the relative proportions of RTs (black) and NRTs (gray) for each half of the population in each sequence. Note that neurons in the top half exhibited relatively more RTs in Sequence 2 than 1, whereas those in the bottom exhibited the opposite pattern. C, Pie charts illustrating the distribution of RTs by action type for the top and bottom halves of the population. The color key can be found at the bottom of B. Each half of the population exhibited different proportions of RTs for WT1, LP1, and NP1 from Sequence 1, and for NP2, LP2, and WT2 from Sequence 2, executed in the order listed respectively. D, When the two halves of the population were recombined, the proportions of RTs dedicated to each action were similar in each sequence. E, Pie charts showing the distributions of RTs for the three individual neurons whose rasters were shown in Figure 6. The neuron number corresponds to its position in A.

Figure 9.

Trial-by-trial decoding of switching dynamics. An SVM approach was used to classify each action on each trial as being more similar to the neuron's typical response to that action versus its typical baseline response. For each action, a trial was selected and, using the remaining trials, an SVM was trained based on the three bins surrounding the action versus an equal number of randomly sampled consecutive three-bin periods from the baseline (nonaction and nonreward) periods. The selected trial was then classified as being associated with the action or the baseline period. The confidence of this classification was given by the SVM posterior and gives what can be considered as a confidence score about whether the neuron was firing more like its typical action response or more like its baseline response. The SVM was trained only on Sequence 1 trials but was tested on both Sequence 1 and 2 trials. A, A neuron responsive to NPs in Sequence 1 (Neuron 3 from the distribution shown in Fig. 7A) was chosen to illustrate SVM decoding. Raster (top) and peristimulus time histograms (bottom) for all NPs performed in Sequence 1 (left) and Sequence 2 (right). The NPs occurred at time 0. The 500 ms bin centered on the NPs and the two 500 ms flanking bins were used for SVM classification. B, Left, Trial 6 (red X) was classified as being associated with the NP rather than the baseline period, when the classifier was trained using all the other trials in Sequence 1 (red dots) versus randomly selected nonaction, nonreward baseline periods (black dots). Successful classification was indicated by the test trial (red X) being on the same side of the hyperplane (blue) as the red dots. The confidence of this classification was given by the SVM posterior scores. While only a single set of random baseline period bootstraps was illustrated here, the selection of baseline periods was typically repeated 100× and the results averaged. Right, By contrast, Trial 25 from Sequence 2 was classified on the “baseline” side of the hyperplane (black X), indicating that this neuron did not respond to NP in Sequence 2 in the same fashion as it did in Sequence 1. C, For this neuron, the mean SVM posterior scores were high for most NPs in Sequence 1 (red line), as would be expected, but dropped and became inconsistent for NPs in Sequence 2. The thin black line denotes decoding during arbitrary surrogate baseline periods (i.e., the SVM posterior scores calculated by creating a group of randomly sampled baseline “trials” and performing decoding on these trials in the identical manner as done for the actual action periods). The thick gray horizontal line denotes the threshold for SVM posterior scores used for all neurons that classify trials as action RTs (SVM RTs). D, The light gray line is the average RT probability for all neurons in the sequence-switch sessions derived from the data shown in Figure 7A. RTs were evaluated for each trial independently and were constant across sequences. The black line is the average SVM RT count across the same set of neurons. SVM RTs were also evaluated on a trial-by-trial basis for each neuron, but since the classifier was trained only on Sequence 1 trials, the drop illustrates that the classifier, which was effective separating action responses in Sequence 1 from baseline firing, was no longer effective at doing so in Sequence 2.