Edited by: Denise Manahan-Vaughan, Ruhr University Bochum, Germany
Reviewed by: Bruno Poucet, Université de Provence, France; Andrea Burgalossi, Werner Reichardt Centre for Integrative Neuroscience, Germany
*Correspondence: A. David Redish, Department of Neuroscience, University of Minnesota, 6-145 Jackson Hall, 321 Church St. SE, Minneapolis, MN 55455, USA e-mail:
This article was submitted to the journal Frontiers in Behavioral Neuroscience.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
The rodent prelimbic cortex has been shown to play an important role in cognitive processing, and has been implicated in encoding many different parameters relevant to solving decision-making tasks. However, it is not known how the prelimbic cortex represents all these disparate variables, and if they are simultaneously represented when the task requires it. In order to investigate this question, we trained rats to run the Multiple-T Left Right Alternate (MT-LRA) task and recorded multi-unit ensembles from their prelimbic regions. Significant populations of cells in the prelimbic cortex represented the strategy controlling reward receipt on a given lap, whether the animal chose to go right or left on a given lap, and whether the animal made a correct decision or an error on a given lap. These populations overlapped in the cells recorded, with several cells demonstrating differential firing to all three variables. The spatial and strategic firing patterns of individual prelimbic cells were highly conserved across several days of running this task, indicating that each cell encoded the same information across days.
The rodent prelimbic cortex (PL) plays an important role in cognitive processing and the solving of decision-making tasks (Kolb,
How PL satisfies all of these different roles remains unclear. Do they co-occur on a single task? Are the different firing correlates particular to the task, or could several of them be observed on the same task if required? And do these different firing correlates arise from different populations of cells, or do the populations somehow overlap? We tested these questions by recording from the PL of rodents as they attempted to solve a spatial decision-making task.
The Multiple-
We recorded neural ensembles (up to 30 neurons simultaneously) from 3 rats through the 6-day strategy-switch sequence. Using established techniques (Schmitzer-Torbert and Redish,
3 male Fisher Brown Norway (FBNF-1) rats aged 8-12 months at the start of behavior were used in this study. Animals were housed on a 12 h light-dark cycle and all experiments for a given rat were run at the same time each day during the rat's light phase. Prior to task training, animals were handled and trained to eat the food pellets used as reward on the task. During training and recording periods, animals received all of their food during behavior or during handling immediately afterward. Animals had free access to water in their cages throughout the experiment. The rats' weights were monitored daily and maintained at or above their 80% free food weight by hand-feeding after running if necessary. All training procedures were approved by the Institutional Animal Care and Use Committee at the University of Minnesota and in accordance with the National Institutes of Health guidelines.
The Multiple-
A daily session on the task began when the rat was placed at the base of the first
Pellet rewards at the feeder locations were provided if the animal went to the correct side as defined by the current reward
The Multiple-
Initial training of the animals was conducted with one of the sides blocked off at the start of maze and choice point, so the animals were forced to run laps around a single loop to either the left or right. Animals ran 40 min sessions to both left and right on blocked tracks for 1–2 weeks until they had run over 50 laps in a session on both the left and the right sides. For these sessions, the available feeders were always rewarded when the animal completed a lap. Once they had reached this 50 lap criterion, the blocks were removed and the animals continued training on the open track initially with just the L and R contingencies pseudo-randomly chosen for each day. Once the animals were running over 50 laps per day and over 80% correct laps, they were introduced to the A contingency as well. When animals were running all 3 contingencies consistently with more than 80% of laps correct and 50 or more laps per session, they underwent the surgical implantation of a 12-tetrode micro-array drive targeted to the pre-limbic region of cortex (PL). Following surgery, the animals were allowed to recover for 2–4 days before returning to running on the track. When performance had returned to previous levels, the tetrodes had been advanced to the recording locations, and the cellular ensemble had been maximized, we began the switch sequence. This paper contains only data recorded during the 6 day switch sequence for these 3 rats.
After training, animals were implanted with 12-tetrode hyperdrives (Kopf). Anesthesia was initiated with sodium pentobarbital (Nembutal, 50 mg/kg, delivered IP) and maintained while on the stereotax via isoflurane mixed at 0.5–2% into medical grade oxygen delivered via a nosecone. Two different implantation techniques were utilized for the hyperdrive itself. The first rat (R193) was implanted with a single bundle drive targeted to the right pre-limbic and infra-limbic (IL) regions of mPFC using a technique based on the one described by Euston and McNaughton (
For the subsequent two rats (R195, R199), we utilized a surgical technique we developed specifically for these experiments. We implanted a dual bundle drive (1 mm spacing between bundles) with 6 tetrodes and 1 reference electrode targeted to each hemisphere of the PFC. The bundle was designed to be implanted on both sides of the saggittal sinus, which was localized by opening a large craniotomy at AP + 3.0 mm from Bregma and extending approximately 1.4 mm laterally on either side of the midline and approximately 1.4 mm A-P. We created this craniotomy by slowly grinding away the skull with a burr in a high speed drill (Foredom, 45,000 RPM) until we could visualize the central sinus, then we were able to carefully remove the final layers of skull and dura matter over the brain on either side and target each of our bundles to one side of cortex. Tetrodes were advanced at least 2 full turns (640 μm) as soon as possible following surgery (~15 min after removal from the stereotax), and were advanced every day subsequently until the switch sequence began. Recording locations were verified histologically to be in the inferior prelimbic cortex, see Figure
In each of these surgeries, a ground screw was secured to the skull along with at least 6 anchor screws, and the drive was held in place by a dental acrylic base secured to the anchor screws. All animals were given a 3-day course of antibiotics (Baytril) and returned to free food following surgery for several days until they were ready to return to running on the track.
Prior to surgical implantation, the animal's position on the maze was tracked via an overhead camera from an LED light on an in-house-designed backpack. After surgery, the rat's position was tracked by LEDs mounted on the headstage plugged into his implant. Local field potentials and unit spiking activity were recorded on a 64-channel analog Cheetah recording system (Neuralynx, Bozeman, MT, USA). Spikes were identified and recorded online using built in filters, then were sorted into individual units offline. Pre-clusters were formed automatically using KlustaKwik (KD Harris), then sorted into individual putative units with the MClust 3.5 software package (AD Redish). Cellular identity and spiking times were registered to the animal's position on the maze and feeder food delivery events that were recorded by the Neuralynx software. A total of 330 cells were identified, predominantly from the inferior PL. Based on assessment of tetrode depths, ~10% of cells may have come from superficial IL, but no clear distinctions were seen. Additionally, based on firing rate measures, few putative interneurons were recorded (less than 10%). Accordingly, all neurons were pooled for all analyses. Subsequent analysis (see below) revealed that many of these cells were instances of the same cell being recorded across multiple days. Waveform matching analyses (see below) left us with 205 putative individual cells.
For analysis purposes, the track on which the animal ran was divided into 6 sections (SoM, NS, CP, Top, Fed, Bot) defined by pixel coordinates relative to experimenter identified locations in the animal's tracking data and the zones used to automatically trigger the feeders in the experiment (see Figure
A lap was defined as a complete cycle from SoM back around to SoM. Animals always started the session at the SoM for their first lap, but the end of the session occurred whenever the 40 min time expired, so animals may not have made it all the way back to SoM on their last putative lap. If the animal made it from SoM to at least the first feeder on their last journey of the session, it was considered to be the last lap. Conversely if they did not make it as far as the first feeder the journey was not counted as a lap. Laps were labeled as
Each cell's firing rate was determined for each maze section on each lap by counting the number of spikes recorded during that section divided by the time elapsed during that section, and the
We measured differential responses to binary task parameters (going left vs. going right, strategy before vs. strategy after the switch, choosing correctly vs. choosing incorrectly) in two ways.
We measured the firing rate difference between the average firing rate for the two components and divided it by the standard deviation of one of the two components. e.g.,
Where
This pseudo-
As an alternate analysis capable of directly comparing the firing of a cell in response to the effect pairs described above, we used a Kolmogorov-Smirnov test to determine whether the distribution of average firing rates on laps identified in one category was different from that of the other. We counted the number of cells with significant (α = 0.05) distributions to the binary pair. We compared this count to the expected count from two controls, an ISI-shuffled control (ISI) and a strategy-shuffled control (ID). To generate the ISI-shuffle, we randomly reordered the inter-spike intervals of each individual cell and recalculated the KS-test significance, counting the number of cells with a significant difference between the binary alternatives using firing rates generated from shuffled ISIs. To generate the strategy-shuffled control, we took the original firing rates, but changed the identification of which lap belonged to which category randomly. Again, we recalculated the cells with a significant difference as measured through a KS-test. In both cases we calculated the mean and standard deviation of the expected number of significant cells from a distribution of 1000 different random shuffles. We used a
Previous research (Euston and McNaughton,
Additionally, we controlled for running speed differences by fitting a robust regression of normalized running speed in each section of the maze to the recorded firing rates in that section of the maze over all laps. Any cells that were found to have a significantly non–zero regression coefficient were removed from the count of cells with significant differential firing as assessed by the KS-test method described above.
A similar control was conducted to remove the effect of cells with a continuously variable firing rate over time from the before vs. after switch selective cells. A robust regression of firing rate vs. lap number was fit to all cells, and cells with a significantly non–zero regression coefficient were removed from the count of cells with significant differential firing as assessed by the KS-test method described above.
In order to analyze spatial information through decoding, positions were first linearized separately for the left and right loops from start of maze back to start of maze. Spatial extent was normalized between key landmarks (the start of maze, the choice point, feeder1, feeder 2, and the start of maze again at the end of the lap). Paths were projected onto this linear coordinate system as per previous studies (Schmitzer-Torbert and Redish,
Bayesian decoding was conducted according to standard methods (Zhang et al.,
The same analyses were conducted on example hippocampal cells recorded from a different group of rats run on the same task. The data collection, processing, and behavior of these rats were equivalent to the rats presented here, but for more detailed methods see van der Meer and Redish (
In order to identify when the same cell was recorded across several days, we combined techniques previously described by Schmitzer-Torbert and Redish (
To construct an actual decision-threshold, we calculated the
In order to check for topographical organization of cells by firing correlates in rodent PFC, we used an analysis described by Redish et al. (
In order to verify that the rodents did indeed learn the M
Prelimbic cortex has been implicated in encoding representations of abstract rules or strategies involved in solving a task (Rich and Shapiro,
The populations of cells with firing rate differences reflecting strategic differences were significantly larger than those expected by chance across all six maze locations: [SoM : (ACTUAL = 75, ISI-control = 10.4 ± 3.3
In order to ensure that subtle path differences on the navigation sequence were not playing a role in the differential firing reported here, we checked for significant differences in the path the animals took on laps before the switch vs. laps after the switch (see Methods). We found 6 sessions in which there were significant path differences, but excluding these sessions from the analysis made no qualitative difference in the percentage of cells that fired significantly differently to laps before vs. after the switch on either the navigation sequence or any other maze section. Additionally, we checked to see if cells could be responding to differences in running speed in the navigation sequence by regressing the firing rate of all cells in the navigation sequence against the animal's speed through the navigation sequence and removing any cells which had a significant effect. This process left us with a significant population of strategy sensitive neurons: ACTUAL = 55,
There is an additional concern that some of the cells we have detected as firing at a different rate before vs. after the switch based on this measure may in fact simply have a continuously variable firing rate that changes over the course of the session. In order to estimate the size of this potential effect, we ran a regression over all cells against lap number and removed any cells which had a significant effect from our population of significantly differentiating cells recorded above. This still left us with a significant population of before vs. after differentiating cells at all maze locations: [Overall: ACTUAL = 109
Examples of cells changing their firing patterns in response to strategic considerations before vs. after the switch have been provided in Figure
On the M
As with the strategy differences, these navigational differences were significant across all six maze locations. The effect was strongest at the feeders (ACTUAL = 100, ISI-control = 17.0 ± 4.5
As described above, we checked for sessions in which there was a significant difference in the path animals took on laps to the left side of the track vs. laps to the right side. We found 5 such sessions, but excluding these sessions from the analysis did not qualitatively change the percentage of cells that fired significantly differently in response to left-going vs right-going laps on any section of the maze, including the navigation sequence. However, removing speed sensitive cells from the Left/Right sensitive population (as above) dropped that population below significance levels for the Navigation Sequence: ACTUAL = 17,
Work from Laubach and colleagues (Narayanan and Laubach,
Unlike the previous two differences examined (before/after switch, left/right), the correct and error laps were only significant for the current lap at the feeder-reward sites (ACTUAL = 131, ISI-control = 24.3 ± 5.5
We found the largest population of cells with differential firing for correct vs. error laps at the feeder regions, which we expected because of the overt cue of reward receipt (or non–receipt on an error lap) at this region. However, the feeder regions are prone to behavioral confounds because behavior might be quite different at these locations when the animal stops to consume food on correct laps instead of running through with no consumptive behavior on error laps. Due to the setup of our experiment it is difficult to control for all of these factors, but we did check to see if cells had a correlation to running speed by regressing firing rate against running speed for all regions for all cells, and subtracting any cells with significant firing rate modulation to running speed from the counts of cells with significant firing rate modulations for correct vs. error laps. After this subtraction we still had significantly above chance populations on the entire track : ACTUAL = 99
So far we have found proportions of prelimbic cells that respond to shifts in the reward contingency of laps, errors made in running the task, and to decisions on which direction to go on any given lap. In order to determine whether these were different populations, we examined the overlap of the populations of cells responding significantly to each of these strategies, i.e., the number of cells that responded significantly differently on both correct vs. error laps, as well as laps before vs. after the switch, etc. These results are shown in a Venn diagram in Figure
Early studies of the firing properties of PL neurons in the rat on spatially-based tasks noted the spatial firing properties of these cells (Poucet,
Whether the spatial firing patterns in rodent mPFC actually encode spatial information remains unclear. Several studies in open field environments have failed to find reliable spatial correlates to firing (Poucet,
There was sufficient spatial information in the prelimbic ensembles to decode position reliably. As shown in Figure
Interestingly, the confusion matrix for prefrontal spatial decoding showed blocks consistent with chunking of the maze, similar to that seen in hippocampal decoding of similar tasks (Gupta et al.,
The implication of these blocks of consistent firing in PL is that various locations on the task may be represented by different population states in PL. Indeed, the confusion matrix is suggestive of a transition between states, which would imply that different parts of the track have different population representation states in PFC. These states could represent different sub-tasks that are solved at different spatial locations of the track, supporting the view that spatial firing patterns in rodent PL reflect spatially consistent cognitive tasks rather than information about space itself.
While individual cells displayed marked variation in firing over the spatial extent of the track, we did not find evidence that any specific region of the track had an increased firing rate at a population level. To measure this we averaged the firing rate recorded at each maze location (and the firing rate over the entire track) for every cell on every lap we recorded over all rats. These firing rates were markedly similar: (Entire track: 2.7 ± 4.9(
Given that there were many sub-populations of cells in PFC that responded in different ways, and that these populations appeared to be randomly distributed, an important question is whether a cell's responding patterns remained consistent across sessions on the same task. Building on the methods used previously (Schmitzer-Torbert and Redish,
Figure
In order to examine the spatial and strategic firing of the cell across an entire session concisely, we created a 7 × 7 grid for each cell for each day with each grid row representing the overall firing rate, firing rate on correct laps and error laps, laps to the left and laps to the right, and laps before and laps after the switch, and each column representing one of the 7 maze locations depicted in Figure
In order to determine whether this consistency persisted across all the cells we recorded, we correlated these 7 × 7 grids across days between all pairs of cells. We then separated these correlation coefficients into matched and unmatched cells. As can be seen in Figure
From this evidence we conclude that while the cells in PL demonstrated a wide array of firing patterns to many behavioral parameters, the responses of each neuron were highly consistent from day to day.
In humans, research has indicated the presence of topographical patterns of responding in the mPFC from anterior to posterior regions (Koechlin et al.,
Prelimbic cells have been reported to have a wide variety of responses to multiple strategic, spatial, and behavioral signals on a variety of tasks. Using a decision-making task that allowed us to study these previously described responses on a single task, we found cells with all of these previously described correlates: Cells in PL had distinct changes in their firing patterns in response to changes in the rule required for solving our task, the future direction the animal was planning to go, and whether the animal had recently made an error or a correct decision. The cells that responded to these different strategic parameters came from different but overlapping populations of cells.
Our results are consistent with most of the previous work on the role of the PL, replicating response-patterns previously described, including strategy-differences (Peyrache et al.,
By matching action potential waveforms across days, we were able to match cells likely recorded across multiple days, and found that both the strategic and spatial firing patterns of the cells were consistent across multiple days on the same task, demonstrating that although there are many potential ways in which cells in PL can respond, the cells do seem to maintain a consistent role on a given task.
Recent work (Rigotti et al.,
PL, along with the infralimbic cortex (IL) and anterior cingulate cortex (ACC), comprise the medial prefrontal cortex of the rat (Van Eden and Uylings,
The prefrontal cortex is critical to the ability to integrate multiple dimensions of task-related information in humans (Rushworth et al.,
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank Chris Boldt and Kelsey Seeland for technical support. Work Supported by: NIH R01-MH080318 (A. David Redish), NIH 5T32 HD 007151 (Nathaniel J. Powell), 3M Sciences and Technology Fellowship (Nathaniel J. Powell). We thank Matthijs van der Meer for helping collect the hippocampal data used for comparison in Figure