Different Multidimensional Representations across the Amygdalo-Prefrontal Network during an Approach-Avoidance Task

SUMMARY

The prelimbic (PL) area and basolateral amygdala (lateral [LA] and basolateral [BL] nuclei) have closely related functions and similar extrinsic connectivity. Reasoning that the computational advantage of such redundancy should be reflected in differences in how these structures represent information, we compared the coding properties of PL and amygdala neurons during a task that requires rats to produce different conditioned defensive or appetitive behaviors. Rather than unambiguous regional differences in the identities of the variables encoded, we found gradients in how the same variables are represented. Whereas PL and BL neurons represented many different parameters through minor variations in firing rates, LA cells coded fewer task features with stronger changes in activity. At the population level, whereas valence could be easily distinguished from amygdala activity, PL neurons could distinguish both valence and trial identity as well as or better than amygdala neurons. Thus, PL has greater representational capacity.

In Brief

Both the PL and BLA regulate emotional behaviors and have similar extrinsic connectivity. However, it is unclear whether they encode distinct aspects of emotional events. Kyriazi et al. show that neurons in these structures encode similar emotional variables but at the population level tend to represent different task dimensions.

Graphical Abstract

INTRODUCTION

Of the structures regulating memory and emotions, two nodes stand out for their remarkably similar connectivity: the medial prefrontal cortex (mPFC) and basolateral amygdala (BLA; Ongür and Price, 2000; Pitkanen, 2000). Not only do these structures form dense reciprocal connections with each other, but they both have access to associative sensory information, share close ties with the insula, project to overlapping striatal territories, and target a common array of brainstem nuclei involved in neuromodulation, emotional expression, and autonomic regulation (Allen et al., 1991; Gabbott et al., 2005; Hoover and Vertes, 2007; Krettek and Price, 1977, 1978; McDonald, 1998; Ongür and Price, 2000).

Consistent with their similar connectivity, the functions of the amygdala and mPFC are tightly intertwined. For instance, they are both required for the expression and extinction of conditioned responses (CRs) to cues (conditioned stimuli [CSs]) that predict aversive outcomes (unconditioned stimuli [USs]) such as behavioral freezing and active avoidance (AA) (Bravo-Rivera et al., 2014; Fanselow and Poulos, 2005; LeDoux, 2000; Moscarello and LeDoux, 2013; Sierra-Mercado et al., 2006). Moreover, they both regulate cued reward-seeking CRs (Ambroggi et al., 2008; Burgos-Robles et al., 2013; Ishikawa et al., 2008; Peters et al., 2009). If the roles of the BLA and mPFC overlap so much, what is the computational advantage of having both? One possibility is that mPFC and BLA neurons code information differently, allowing them to implement distinct sets of input-output functions.

Although many emotional conditioning studies have examined the coding properties of BLA and, to a lesser extent, mPFC neurons, most relied on simple tasks that featured one or just a few CSs and CRs, giving the misleading impression that neurons encode single, easily interpretable dimensions (Paré and Quirk, 2017). In contrast, studies relying on more complex or diverse experimental paradigms revealed that BLA (Gründemann et al., 2019; Kyriazi et al., 2018; Saez et al., 2015) and mPFC neurons ( Ma et al., 2014, 2016) encode various conjunctions of task features or behaviors, in effect implementing high-dimensional representations that may support higher cognitive functions.

Unfortunately, other than simple conditioning studies, the coding properties of mPFC and amygdala neurons have not been examined in the same task yet. A few studies compared the coding properties of amygdala and anterior cingulate or orbitofrontal neurons in non-human primates (Klavir et al., 2013; Livneh and Paz, 2012; Livneh et al., 2012; Munuera et al., 2018; Pryluk et al., 2019; Saez et al., 2015), revealing significant differences between them. For instance, whereas in the amygdala, the same neurons encode social hierarchical rank and the reward value of non-social stimuli, in the orbitofrontal and anterior cingulate cortex, they encode only reward value (Munuera et al., 2018).

Thus, the present study was undertaken to compare the coding properties of mPFC and amygdala neurons using a task that requires rats to flexibly select between different defensive or appetitive behavioral outputs on the basis of the location of the CS on each trial.

RESULTS

Overview of the Experiments

Rats were trained on the recently introduced risk-reward interaction (RRI) task (Figure 1A; Kyriazi et al., 2018). In this task, rats learn to obtain rewards and avoid shocks depending on the location of a light CS. Light activation below one of the three floor sectors (CS-S1, CS-S2, and CS-S3) signals an impending foot shock (Figure 1A, left), whereas light activation behind the left or right wall (CS-R1 or CS-R2, respectively) signals the availability of a liquid reward at that site (Figure 1A, right; blue, water port).

Figure 1. Risk Reward Interaction (RRI) Task, Histological Verification of Recording Sites, and Neuronal Classification.

(A) RRI task apparatus: a rectangular arena with high walls, no ceiling, and a floor made of metal bars. Light-emitting diodes (LEDs) at different positions signal an upcoming reward (blue; behind left and right walls, CS-R1 and CS-R2, respectively) or an impending foot shock (red; under each of three different floor sectors, CS-S1, CS-S2, and CS-S3).

(B) Acquisition of conditioned active avoidance (B1) and reward seeking (B2). Gray lines, individual rats (n = 4). Colored lines, average ± SEM.

(C) Histological verification of recording site. Coronal section stained with cresyl violet. A small electrolytic lesion (arrow) marks the last recording site.

(D) Classification of recorded PL cells as presumed projection cells (PNs; blue) or fast-spiking interneurons (ITNs; red) on the basis of their spike duration (y axis, peak-to-trough interval, cutoff of 0.5 ms) and firing rate (x axis, cutoff of 10 Hz). Cells that did not meet both criteria (UNCL; gray) were not considered further. Cg, cingulate cortex; DP, dorsal peduncular cortex; fmi, forceps minor of the corpus callosum; IL, infralimbic cortex; PL, prelimbic cortex. See also Figure S1.

Rats could emit different CRs depending on their locations with respect to the CSs. On aversive trials, if they were located in the shock sector, they had to move to a different part of the arena to avoid it (Figure 1A, left), a behavior termed AA. Otherwise, they had to remain away from the shock sector until the trial ended. Rats also showed various amounts of freezing, especially on trials that called for an AA response. On appetitive trials, two CRs could be distinguished: reward approach (RA) and reward anticipation (RAnt). RA refers to rats’ running to the activated water port from a different part of the arena (Figure 1A, right). RAnt refers to rats’ placing their front paws on the water port in anticipation of reward delivery.

Once rats (n = 12) reached criterion (>85% correct trials on all CSs combined), generally in less than 1 week (Figure 1B), they were implanted with a silicon probe just dorsal to the prelimbic (PL) sector of the mPFC (n = 4) or the lateral nucleus of the amygdala (LA; n = 11). Three of these rats had probes aimed at both structures. Within-session performance was stable in implanted rats (Figure S1). Throughout the recording sessions, we monitored the rats’ position, head direction, and movement velocity using an overhead camera. The probes were lowered after each recording session by at least 140 μm to obtain new cells the following day. At the conclusion of the experiments, the location of the probe was marked with small electrolytic lesions for subsequent histological verification of recording sites (Figure 1C).

CS Responsiveness and Behavioral Correlates

The following includes only neurons histologically determined to have been recorded in PL (n = 526), LA (n = 365), and the basolateral nucleus of the amygdala (BL; n = 307). As detailed in STAR Methods, on the basis of earlier reports (Amir et al., 2015, 2018; Barthó et al., 2004; Kyriazi et al., 2018), we classified PL and BLA cells as putative principal neurons (PNs) or fast-spiking interneurons (ITNs; Figure 1D) on the basis of their firing rates and spike durations. However, some neurons could not be classified, resulting in samples of 464 PL (452 PNs, 12 ITNs), 315 LA (284 PNs, 31 ITNs), and 259 BL (214 PNs, 45 ITNs) neurons. Because the paucity of ITNs recorded in PL precluded their comparison between regions, this report deals exclusively with PNs.

The task-related activity of PL neurons was similar to BLA in many respects (Figure 2). Although the magnitude of firing rate changes tended to be lower in PL, CS responses at the three sites were similar, ranging from transient or sustained increases in firing rates to long-lasting inhibitory responses (Figures 2 and S2).

Figure 2. CS-Evoked Activity in LA, BL, and PL Neurons.

(A–C) Heatmaps of activity evoked by CS-Rs (top) or CS-Ss (bottom) in all available principal LA (A) (n = 284), BL (B) (n = 214), and PL (C) (n = 452) neurons. Firing rates were Z-scored on the basis of activity during the pre-CS period. Data are plotted with 200 ms bins. Warmer colors indicate higher firing rates (see color bar). In (A1)–(C1), neurons were rank ordered on the basis of the amplitude of their responses to CS-R1, and the same order was kept for CS-R2, respectively. In (A2)–(C2), cells were rank ordered on the basis of the amplitude of their responses to CS-S1, and the same order was kept for CS-S3.

See also Figures S2 and S3.

To assess the statistical significance of CS responses, we compared firing rates in the first second following CS onset with the 5 s pre-CS baseline (50 ms bins) using rank-sum tests with a significance threshold of p < 0.005. We restricted this comparison to the first second after CS onset because other analyses described below revealed that PL neurons, like BLA cells, also fired in relation to CRs occurring later during the CSs. Restricting to increases in firing, the proportion of cells with significantly increased firing rates during at least one of the CSs did not differ significantly at the three sites for the CS-Rs (LA, 7%; BL, 9%, PL, 7%; χ² = 0.31, p = 0.86) or the CS-Ss (LA, 8%; BL, 13%, PL, 11%; χ² = 3.36, p = 0.19) (Figure S3).

Another similarity between LA, BL, and PL neurons was the strong dependence of their CS-related activity on behavior. To examine this question, we referenced unit activity to the onset of CRs instead of CSs, computed peri-event time histograms (PETHs) of neuronal discharges, and determined whether firing rates increased in relation to one or more CRs relative to firing rates preceding behavior onset (rank-sum tests, p < 0.005). Figure 3A illustrates three different PL neurons in which trials were aligned to the onset of CSs or CRs. As we reported in BLA (Kyriazi et al., 2018), some PL cells with robust CS responses lacked behavior-related activity (Figure 3A). Conversely, some PL cells with no CS responses showed strong CR correlates (Figure 3B). And finally, some cells had both CS- and CR-related increases in firing rates (Figure 3C). A detailed analysis of the relation between CS responsiveness and CR-related activity revealed a remarkably similarprofile atthe three sites (Figure S3). The proportions ofcells with increased firing rates during active CRs did not differ significantly between amygdala and PL neurons for the CS-R (RA: LA, 9.9%; BL, 12.6%; PL, 13.2%; χ² = 1.98, p = 0.37) or CS-S (AA: LA, 16.5%; BL, 17.2%; PL, 15%; χ² = 0.64, p = 0.73).

Figure 3. CS-, CR-Related Activity, and Selectivity of PL Neurons.

(A–C) Three different PL neurons. Top: rasters, where ticks represent individual spike times and each row represents a trial. Bottom: Z-scored peri-event histograms (PEHs) of firing rates. In (A1)–(C1), rasters and PEHs are referenced to CS onset, whereas in (A2)–(C2), they are referenced to CR onset. (A) Cell with a CS response but no CR correlate. (B) Cell with no CS response that fires in relation to the CR. (C) Cell with both, a CS response and a CR correlate.

(D) Comparison between Z-scored averaged firing rate ± SEM of PL cells during AA (black) versus no avoidance (green), referenced to CS onset. Only cells increasing their firing rates during AA are included in this panel.

(E) Comparison between Z-scored average firing rate ± SEM of PL cells during correct (black) and error (red) CS-S trials. The number of cells included in this analysis is smaller than in (D) because error trials did not occur in all recording sessions. This figure is based on ten error trials from seven recording sessions.

(F) Comparison between Z-scored average firing rate ± SEM of PL cells during correct (black) and error (red) CS-R trials (60 error trials from 12 sessions).

(G1) Comparison between the incidence of cells with excitatory (red) or inhibitory (blue) responses to only one CS-R (top) or CS-S (bottom) among LA (left), BL (middle), and PL (right) neurons. Percentages represent selective cells out of total number of responsive cells. The total numbers of CS-R excited cells were 24 in LA, 22 in BL, and 40 in PL. The total numbers of CS-S excited cells were 31 in LA, 26 in BL, and 47 in PL. The total numbers of CS-R inhibited cells were 25 in LA, 52 in BL, and 49 in PL. The total numbers of CS-S inhibited cells were 20 in LA, 46 in BL, and 38 in PL.

(G2) Example of PL cell with inhibitory response to CS-R2 but not to CS-R1. Comparison between the incidence of cells with excitatory (red) or inhibitory (blue) CR-related activity in relation to only one CS-R (top) or one CS-S (bottom). The total numbers of RA-excited cells were 28 in LA, 27 in BL, and 60 in PL. The total numbers of AA-excited cells were 47 in LA, 37 in BL, and 68 in PL. The total numbers of RA-inhibited cells was 35 in LA, 44 in BL, and 148 in PL. The total numbers of AA-inhibited cells were 47 in LA, 37 in BL, and 68 in PL.

(H1) Comparison between the incidence of cells with excitatory (red) or inhibitory (blue) responses to only one RA (top) or AA (bottom) among LA (left), BL (middle), and PL (right) neurons. Percentages represent selective cells out of total number of responsive cells. The total numbers of RA excited cells were 28 in LA, 27 in BL, and 60 in PL. The total numbers of AA excited cells were 47 in LA, 37 in BL, and 68 in PL. The total numbers of RA inhibited cells were 35 in LA, 58 in BL, and 94 in PL. The total numbers of AA inhibited cells were 45 in LA, 44 in BL, and 148 in PL.

(H2) Example of PL cell activated in relation to AA elicited by CS-S1 but not CS-S3.

(G2 and H2) Top: rasters, where ticks represent individual spike times and each row represents a trial. Bottom: Z-scored PEHs of firing rates.

In fact, further analyses revealed that, as we previously observed in BLA (Kyriazi et al., 2018), much of the CS-related increases in PL firing rates were linked to the behavior they elicited. For instance, when we compared the same cells with AA-related activity on CS-S trials that called for AA versus no avoidance responses, we found that the CS-S-related activity was drastically higher on trials that called for avoidance responses (Figure 3D). Similarly, when we compared the activity of AA or RA cells on correct versus error trials, their firing rate during the CS was lower on error than correct trials (Figures 3E and 3F).

Differences between the Selectivity of PL and BLA Neurons

Despite the similarities, there were also clear differences. First, the cells’ responses to different CSs of the same valence were more dissimilar in PL (Figure 2C) than in LA (Figure 2A) or BL neurons (Figure 2B). To quantify this, we first averaged the Z-scored response of each cell during the first second after CS onset and then calculated the correlation between CS-R1 versus CS-R2 and CS-S1 versus CS-S3 for each brain region. The correlation was lower in PL than LA or BL neurons for the CS-Rs (LA, 0.53; BL, 0.69; PL, 0.42) and CS-Ss (LA, 0.5, BL, 0.58; PL, 0.45).

To shed light on the origin of this difference, we compared the incidence of cells whose firing rates significantly increased or decreased in response to just one of the CSs of a given type. This method disclosed a clear difference between PL and amygdala neurons. That is, the proportion of cells with inhibitory responses to only one CS of a given type was significantly higher in PL (Figure 3G1). This was true of the CS-Rs (χ² = 18.7, p < 0.0001) and the CS-Ss (χ² = 9.64, p = 0.0081) but did not hold for excitatory responses (Figure 3G1) to CS-Rs (χ² = 1.03, p = 0.59) or CS-Ss (χ² = 0.17, p = 0.92). Figure 3G2 shows an example PL cell with inhibitory response to CS-R2 but not to CS-R1.

PL neurons also showed greater selectivity for CRs. A higher proportion of PL cells were activated in relation to RA evoked by only one of the CS-Rs or AA elicited by only one of the CS-Ss than in LA or BL (RA: χ² = 12.11, p = 0.0024; AA: χ² = 19.99, p < 0.0001; Figure 3H1). The difference in selectivity was less marked (RA) or absent (AA) when considering inhibitory responses (RA: χ² = 20.61, p < 0.0001; AA: χ² = 0.03, p = 0.99; Figure 3H1). Figure 3H2 shows an example PL cell activated in relation to AA elicited by CS-S1 but not CS-S3.

Disentangling the Correlates of Unit Activity with a Generalized Linear Model (GLM)

In the prior section, our analyses were complicated by the presence of multiple sensory and behavioral variables whose relative timing varied between trials. Specifically, a number of factors potentially influenced neuronal activity including CS identity, the rats’ position with respect to the CS, the type and timing of the CRs, the speed at which the rats moved, and potential interactions among these variables. Because PETHs cannot isolate these interacting factors, we turned to a different approach: we inferred the variables PL neurons encode using a GLM.

The GLM relies on variations in the timing, duration, and type of variables occurring on different trials to determine which ones neurons encode. Of note, the group lasso GLM permits dimensionality reduction in correlated data and favors sparsity in the identification of variables related to neuronal activity (Breheny and Huang, 2015; Tibshirani, 1996; Yuan and Lin, 2006). Because of the prevalence of selective responses observed in PL (Figures 3G and 3H), we added interaction terms in the GLM to capture the differential associations between individual CSs and CRs (e.g., CS-R1 with RA).

To ensure that the GLM captured the observed coding parameters, we compared the model’s output with PETHs referenced to different variables. One example cell shown in Figure 4 had a complex profile of responses, including a transient inhibition at the onset of all CSs, a selective increase to RA only on CS-R2 trials, and an inhibition related to AA. As indicated in the first six columns of Figure 4, the model accurately captured all of these responses, including selective responses with interaction terms such as CS-R2 RA. Additionally, it identified that a proportion of the response was related to speed, a parameter that is tracked throughout the sessions. Locomotion effects on firing rates are a factor that traditional PETH analyses cannot account for. Often, apparent behavioral correlates are in fact related to changes in the animal’s speed (Figure S4A). The GLM allows us to factor speed out and attribute firing related to CRs without the confounding influence of movement speed.

Figure 4. Coding of Task Variables by Example PL Cell as Estimated by the GLM.

The first six columns show GLM-estimated spiking (blue lines) for each task variable (gray lines). The last column on the right consists of the estimated spiking (blue lines) superimposed on the observed spiking of the cell (red lines) for each CS. CS-Rs and associated behaviors (blue letters) are shown in the first two rows, and CS-Ss and associated behaviors (red letters) shown in the bottom two rows. AA, active avoidance; Frz, freezing; RA, reward approach; RAnt, reward anticipation.

See also Figure S4.

Another benefit of the lasso GLM is its ability to set parameters to zero when they do not contribute to the firing rate prediction of a given cell. For example, in Figure 4, the freezing behavior variables are all set to zero, including both main and interaction effects, because the firing rate of the cell was unaffected by freezing, even though the animal spent several trials freezing to CS-S1. This possibility prevents overfitting and favors sparsity in identifying variables related to firing rates.

The GLM output was used to characterize coding by LA, BL, and PL neurons. As a control, we performed the same analyses on putative principal striatal cells (n = 237) that were recorded dorsal to the amygdala. The model’s estimation accuracy was comparable among striatal, amygdala, and PL cells, although the fit for BL cells was superior to the other regions (Figure S4B; χ²[3] = 25.81, p < 0.0001). On the basis of the model predictions, we computed the proportion of cells with inhibitory and excitatory modulations (see Figures 5A and S5 for response latencies) as well as the average absolute magnitude of the modulations (Figure 5B). With a few exceptions, the encoding of task variables was similar in the four regions. At the four sites and for each task feature, >35% of cells (Figure 5A) showed excitatory or inhibitory modulations with the exception of freezing, a variable that was especially under-represented among PL neurons.

Figure 5. Comparison between Coding of RRI Task Variables by Striatal, LA, BL, and PL Neurons as Determined Using a Generalized Linear Model (GLM).

(A1–A4) Proportion of presumed PNs (y axis) in the striatum (A1), LA (A2), BL (A3), and PL (A4) that exhibited excitatory (red) or inhibitory (blue) coding of different task variables used in the GLM (x axis).

(B) Absolute average ± SEM modulation of the firing rates of striatal (B1), LA (B2), BL (B3), and PL (B4) PNs in relation to each variable used in the GLM. AA, active avoidance; Frz, freezing; RA, reward approach; RAnt, reward anticipation. See also Figure S5.

A two-way ANOVA with recording site and feature type as factors on the modulation of responses revealed a significant main effect for recording site and feature (F_[3,17] = 96.77 and 41.8, respectively, p < 0.0001; Figure 5B). A distinguishing feature of PL neurons was the generally low magnitude of their modulations. A Tukey-Kramer multiple-comparison test confirmed that the magnitude of the modulations differed among all recording sites at p < 0.0001 with the exception of BL versus PL (p = 0.1545).

To further characterize coding at the four sites, we correlated the cells’ peak modulations by all task variables (Figure 6A). As we reported previously (Kyriazi et al., 2018), LA cells tended to show correlated modulations by task events of the same valence (Figure 6A2) and lower or negative correlations between modulations by task events of different valences (Figure 6A2). In contrast, this pattern was less apparent among BL neurons (Figure 6A3) and even less so in PL cells (Figure 6A4). To quantify this aspect, we computed the d0 score (Keene et al., 2016; McKenzie et al., 2014), which measures the separation between different coding dimensions on the basis of the distributions of the correlation coefficients in the similarity matrices. The d' metric was calculated for each coding dimension by comparing the degree to which within-valence correlations exceeded those between valence. Consistent with the impression gained from visual inspection of the matrices, valence coding (Figure S6A) was significantly lower in PL (0.75) than LA (1.66) and BL (1.43) (bootstrap resampling, p < 0.05) but did not differ from the striatum (0.83). In contrast, the d' metric for active versus passive CRs (Figure S6B) was significantly higher in the striatum (1.26) than in LA (0.50) and BL (0.22) (bootstrap resampling, p < 0.05) but did not differ from PL (0.86) (bootstrap resampling, p > 0.05).

Figure 6. Contrasting Representation of Task Variables in the Striatum, Amygdala, and PL.

Presumed PNs in the striatum (1), LA (2), BL, (3), and PL (4).

(A) Spearman correlation matrices between activity elicited by different task variables. As indicated by the color scale on the right, warmer and cooler colors indicate positive and negative correlations, respectively.

(B) Frequency distributions of Gini index (top) and dimensionality index (bottom).

AA, active avoidance; Frz, freezing; RA, reward approach; RAnt, reward anticipation.

See also Figure S6.

To compare how coding of the various dimensions is allocated within neurons in the striatum, LA, BL, and PL, we computed the Gini index, a measure of statistical dispersion originally used in economics to measure income equality (Gini, 1921). A high Gini index indicates a less distributed code whereby neurons encode different dimensions unevenly, whereas a low Gini index indicates a more distributed code whereby neurons encode different dimensions more equally. As shown in Figure 6B, striatal and LA neurons tended to have higher Gini indices than BL and PL neurons. Wilcoxon rank-sum tests confirmed that the Gini indices of LA and striatal neurons did not differ from each other (p = 0.091) but were significantly higher than in BL (BL versus LA, p = 0.011; BL versus striatum, p < 0.0001) and PL neurons (PL versus LA, p = 0.026; PL versus striatum, p < 0.0001), with no difference between PL and BL cells (p = 0.44). Hence, BL and PL neurons form a more distributed representation of task features than striatal and LA neurons.

Complementing these results, the dimensionality index, which measures how many features each neuron encodes (see STAR Methods) was lowest among striatal and LA neurons and highest among BL and PL cells (Figure 6B). Wilcoxon rank-sum tests confirmed that the dimensionality indices of striatal neurons were lower than those of LA (p = 0.023), BL (p < 0.0001), and PL (p < 0.0001) neurons. In addition, these tests revealed that the dimensionality indices of LA cells were lower than that of BL cells (p = 0.002) but did not differ from PL cells (p = 0.217), with BL cells having higher dimensionality indices than PL neurons (p < 0.007).

Population Analyses Reveal Multidimensional Representations

Together, the results obtained so far suggest that neurons in the four regions encode information differently. Striatal and LA neurons tend to have strong modulations and encode few task features. In contrast, PL and BL neurons represent multiple features through small fluctuations in their firing rates.

To assess the impact of these differences in the representation of task features at the population level, we performed dimensionality reduction using principal-component analysis (PCA) across the different trial types (Figure 7). Given that the timing of CRs varied among trials and that firing rates fluctuated depending on the CRs emitted, we divided each trial into four equal epochs and time-normalized each epoch by assigning them a fixed number of bins. The four epochs represented the baseline period, the CS-onset period (before any CRs occurred), the CR period (RA or AA), and the rest of the trial until US onset.

Figure 7. Principal Component Trajectories of Trial Types.

(A) Time-normalized trajectories in three-dimensional PCA space starting 5 s before to 10 s after CS-onset for STR (A1), LA (A2), BL (A3), and PL (A4) neurons. Colors go from pale to dark across time. The black circle denotes the onset of the CS. The three-dimensional space has been rotated for each structure to show the best dimension of separation.

(B) Mean Euclidean distances calculated between each trial type for striatal (B1), LA (B2), BL (B3), and PL (B4) neurons.

See also Figure S7.

Figure 7A shows the PCA trajectories for each brain region, starting 5 s before trial onset until 10 s after trial onset (pale to dark colors, respectively). The PCA trajectories revealed striking differences between the four structures. Striatal and LA cells showed trial trajectories separated by valence (Figures 7A1 and 7A2). That is, within CS types (e.g., CS-R1 and CS-R2), trials followed similar paths, whereas between CS types (e.g., CS-R versus CS-S), trial trajectories diverged (Figures 7A1 and 7A2). In PL by contrast, trial paths differed within CS types and by valence (Figure 7A4). Finally, BL neurons showed an intermediate pattern in which trial paths were well separated for only CS-Ss, while showing divergence as a function of valence (Figure 7A3). Notably, the divergence observed in PL trajectories does not appear to be driven by a larger proportion of cells modulated by position information (Figure S7).

To quantify the visual representation of PCA trajectories, we compared the Euclidean distances between trajectories (Figure 7B). Distances were significantly different between structures (χ²[3] = 85.81, p < 0.0001), with PL (Figure 7B4) showing higher distances compared with striatum (Figure 7B1), LA (Figure 7B2), and BL (Figure 7B3; Tukey-Kramer post hoc test, p < 0.01 for all). Additionally, all other distances differed except for LA and BL (p = 0.9669).

As our PCAs indicate that population activity diverges on the basis of trial type in PL, a linear decoder should identify trial type more reliably on the basis of the activity of PL than striatal, BL, or LA neurons. Algorithms such as support vector machines implement decoders that are physiologically plausible (Baker, 2003; Poirazi et al., 2003). Thus, to test this prediction, we trained a support vector machine with 10-fold cross-validation on a resampled population of 150 cells from each brain region and compared decoder performance on individual trial types (Figure 8). Because a binary decoder allows only two-class comparisons, we trained one decoder on identifying CS-R1 versus CS-R2 trials and another decoder on CS-S1 versus CS-S3 trials. We then pooled the performance accuracy of the two decoders to calculate a performance score for trial identity.

Figure 8. Trial Type and Valence Decoders.

(A1) Decoder accuracy of the decoders for trial type (CS-R1 versus CS-R2 and CS-S1 versus CS-S3) during the CS period before behaviors are initiated. Dashed line represents shuffled trial type, indicating chance. Error bars represent the SEM on the basis of 50 decoder repetitions.

(B1) Decoder accuracy of trial type during active CRs (CS-R1 RA versus CS-R2 RA and CS-S1 AA versus CS-S3 AA).

(C1) Decoder accuracy of valence (CS-Rs versus CS-Ss) during the CS period before behaviors are initiated.

(D1) Decoder accuracy of valence active CRs (RA versus AA).

(A2–D2) Decoder accuracy as population size increases. Dark-colored lines represent observed decoder performance. Light-colored lines represent decoder performance with shuffled trial labels.

Related to Figure S8.

(E) Gradients of representation indicated by color scale for valence, active-passive code, distributed code, dimensionality, mixed selectivity coding, and representational capacity in the striatum, LA, BL, and PL. The measures indicated on the right were used to calculate the color gradients for each representation.

The decoders were superior at identifying trial type on the basis of the activity of PL than BLA or striatal neurons, during both the CS (Figure 8A1) and CR (Figure 8B1) epochs. Consistent with the higher response selectivity we observed earlier in PL using peristimulus time histograms (PSTHs) (Figures 2 and 3), CS identity was decoded with higher accuracy by PL neurons (93%) than BL (70%), LA (68%), and striatal (55%) neurons (Figure 8A1). Although decoding accuracy during the behavior epoch was high with all cell types (striatal, 74%; LA, 84%; BL, 92%), PL neurons still outperformed the others with 98% accuracy (Figure 8B1). Yet all decoders performed above chance, as determined by comparing decoding accuracy with actual versus shuffled trial labels.

Next, we tested whether valence could be decoded from the same population activity. Accordingly, we collapsed CS-R1 and CS-R2 trials into a CS-R category as well as CS-S1 and CS-S3 trials into a CS-S category and then trained the decoder to distinguish between CS-R versus CS-S trials. Decoding accuracy for valence was above 90% in all brain regions for both CS (Figure 8C1) and CR (Figure 8D1) periods.

To determine how the size of the population influences decoding accuracy, we increased the number of cells. This approach revealed significant differences between structures for trial identity during the CS (Figure 8A2; χ²[3] = 54.46, p < 0.0001) and CR (Figure 8B2; χ²[3] = 34.29, p < 0.0001) periods as well as for valence during behavior (Figure 8D2; χ²[3] = 18.11, p < 0.001). Decoding trial identity during the CS period resulted in improved performance with growing population size only in PL. In BLA, the number of cells did not contribute to the decoder’s performance, indicating that few cells hold a weak representation of trial identity information at the CS level. Puzzlingly, trial identity performance decreased as the number of striatal cells was increased but remained above chance. A Tukey-Kramer post hoc test showed that all structures differed in decoding performance for trial identity during the CS period (p < 0.05), except LA and BL (p = 0.5345). During the behavior period, decoder accuracy for trial identity increased in all brain structures with increasing population size, but only striatum versus BL (p = 0.001) and PL (p < 0.0001) as well as LA versus PL (p < 0.0001) were significantly different from each other. Finally, decoding of valence during the CS period did not differ between structures (Figure 8C2; χ²[3] = 4.05, p = 0.2557), while decoding of valence during the behavior period was lower for the striatum compared with LA (p = 0.0048) and BL (p < 0.0001).

The prevalence of the valence code across regions was surprising and raises the question of whether the valence code in LA and BL is similar to the one in PL and striatum. One way to address this is to determine whether the coding of valence during the CS period generalizes to the CR period. Cross-temporal generalization analyses capture this by training a decoder during one trial period and testing during another period. Valence codes in LA and BL generalized better between CS and CR periods than those in PL and striatum (Figure S8A). These analyses also showed that the PL code for trial identity generalized between trial periods, suggesting that its ensemble activity codes for trial identity in a manner that is consistent across the entire trial (Figure S8B).

Together the decoder results indicate that PL neurons are better at holding information about multiple representations simultaneously including trial identity and valence, while BLA units are more specialized for valence coding. Additionally, increasing population size influences decoding accuracy differently depending on the brain structure and the variable being decoded.

DISCUSSION

In keeping with their similar connectivity and responsiveness to task, BLA and mPFC neurons are involved in closely related functions, raising the question of what is the computational advantage of having both regions. Here, we tested the hypothesis that mPFC and BLA neurons code information differently, potentially allowing them to represent distinct task dimensions (e.g., valence). Accordingly, we compared the responses of PL, LA, and BL neurons during a task in which rats generated different defensive or appetitive CRs depending on CS location. We found that although specific task events were represented by individual neurons in PL, LA, and BL, cells differed in the combinations of events they responded to. At the population level, these differences scaled up to variations in their proclivity to encode certain task dimensions. Hence, the amygdalo-prefrontal network is not characterized by clear-cut regional differences in the events encoded but by gradations in the representation of task dimensions.

Gradations in Task Representations

On the surface, the task-related activity of PL neurons appeared similar to LA and BL. For instance, a nearly identical proportion of neurons at the three sites significantly increased their firing rate in response to at least one CS. Also, cells with CR-related activity seemed as prevalent in PL, LA, and BL. Moreover, at the three sites, CS-evoked activity showed a strong dependence on behavior, with the same CS eliciting larger responses on trials in which rats emitted the appropriate CR than on error trials. Beneath these superficial similarities, however, there were conspicuous differences. First, the cells’ responses to different CSs of the same valence were more dissimilar in PL than in LA or BL. That is, more PL than LA or BL cells had inhibitory responses to only one CS of a given valence. PL neurons also showed greater selectivity for CRs evoked by different CSs: whereas in LA and BL, most CR-activated cells were recruited similarly irrespective of the CS being presented, the CR-related activity of most PL neurons varied depending on the CS that evoked it. This encoding of specific conjunctions of task features is known as mixed selectivity, which has been described previously in other prefrontal areas (Lindsay et al., 2017; Rigotti et al., 2013, Saez et al., 2015). Furthermore, mixed selectivity is thought to provide a computational advantage for multidimensional representations (Fusi et al., 2016).

That LA, BL, and mPFC neurons encode overlapping task features while showing differences in how they are represented at the population level suggests that these regions serve partially distinct functions. Indeed, directly comparing these population representations among regions does not reveal discrete differences in what task dimensions are encoded but instead differing propensities for the representation of some dimensions over others, hereafter referred to as representation gradients. We observed several instances of this phenomenon (Figure 8E). First, we observed a gradient of valence coding across the amygdalo-striatal-prefrontal network. Although it is well known that LA and BL cells are modulated by valence (Belova et al., 2008; Kyriazi et al., 2018; Namburi et al., 2015; Paton et al., 2006; Sangha et al., 2013), we observed that striatal and PL neurons represent valence to a lesser degree, as evidenced by their low d' index. Nevertheless, valence could be easily decoded from the population activity of all four regions, as indicated by our linear decoder results (Figures 8A–8D). However, the ability for a decoder to derive valence from striatum and PL does not necessarily mean that those areas code for valence in the same way as LA and BL. In LA and BL, neurons tended to respond similarly to all CSs and CRs with the same valence (Figures 6A and S6A), although this was not the case for either striatum or PL. An additional confirmation of this came from cross-temporal generalization analyses of our decoders (Figure S8), which found that striatal and PL decoders trained on the CS period did not perform well during the behavior period. Thus, although the valence code in LA and BL strongly generalizes across conditions, in striatum and PL the generalization is weaker.

Another gradient was found for active versus passive CRs. In this case, striatal cells had the strongest representation, followed by PL and finally amygdala cells (Figure 8E). That striatal cells can easily distinguish different behavioral strategies is consistent with the involvement of the striatum in motor functions (Wall et al., 2013). Similarly, mPFC neurons respond differently to movement excitation and inhibition during fear conditioning (Halladay and Blair, 2015), and an inhibitory signal in PL was found to mediate AA (Diehl et al., 2018). Finally, given that amygdala neurons weakly distinguish between active and passive strategies, it is unlikely that they orchestrate their execution, but rather integrate inputs from mPFC and provide a bias for approaching or avoiding stimuli on the basis of valence (Lázaro-Muñoz et al., 2010; Sangha et al., 2020).

Gradations in the Manner of Population Representations

Besides what is encoded by BLA and PL neurons, there is also the question of how the codes are represented by these cell populations. Here too, we found a continuum in the form of representation across these regions (Figure 8E), as quantified with the Gini and dimensionality indices and event-evoked firing rates. Although individual PL and BL neurons tended to weakly respond to many different task events, LA and striatal cells coded fewer task features with stronger changes in their firing rates. Although it may seem odd that PL was more similar to BL than the neighboring LA nucleus in this respect, this finding is consistent with the connections between these three structures. That is, PL forms stronger reciprocal connections with BL than LA (Krettek and Price, 1977; McDonald, 1991; Mcdonald et al., 1996). Moreover, BL projections to LA are sparse (Krettek and Price, 1978; Paré et al., 1995; Pitkänen et al., 1997).

Another instance of graded regional differences in coding was found in their representational capacity (Figure 8E). Using PCAs, we observed that whereas striatal and LA populations showed trial trajectories only separated by valence (or its associated CSs and CRs), PL exhibited a unique trajectory for each trial type. Between these extremes, BL neurons distinguished valence, but they also differentiated the two types of CS-S trials. The ability of PL neurons to maintain independent representations of each trial type indicates that the population activity of PL neurons holds more information than that of BLA or striatal neurons. The more distinct representations found in PL may prevent interference when learning new associations (Bartolo et al., 2019).

If these representations are physiologically meaningful, then they must be decodable by downstream neurons. Indeed, decoders were superior at identifying trial identity on the basis of the activity of PL than BLA or striatal neurons, during either the CS or CR periods (Figures 8A–8D). Moreover, gradually increasing the size of the cell populations used in these analyses for the CS period improved decoder performance only with PL neurons, suggesting that although BLA and striatal cells hold some information about trial identity, redundancy between cells limits population level decoding of trial types. This may arise because BLA neurons respond similarly to CSs and CRs of the same valence (Figure 6A), in effect causing them to act as "valence detectors." Indeed, smaller sized BLA ensembles outperformed those from PL and striatum for encoding valence (Figure 8D). But as population size increased, PL caught up, achieving comparable performance. These results indicate that even though PL neurons show small task-related fluctuations in firing rates, their population activity has greater representational capacity, as it can encode both trial identity and valence information simultaneously. This is in line with recent observations that the activity of prefrontal neurons forms a geometric pattern in multidimensional space that makes it possible to decode abstract representations such as context and history of reward value (Bernardi et al., 2018).

The mixed-selectivity code in PL provides a distinct population code for each situation, able to distinguish between circumstances that downstream areas may encode similarly. Indeed, it has been shown that mixed-selectivity coding leads to higher dimensional representations (Fusi et al., 2016), allowing a brain region to encode and communicate a multitude of task demands. Downstream ensembles can benefit from this unique code because it activates distinct populations of synapses for events those ensembles normally cannot distinguish. As a result, these afferent synapses robustly activate the downstream population and expand the range of circumstances the region normally responds to. And once that association is formed, situations that normally would not drive the downstream area gain the ability to activate the appropriate group of cells. Indeed, PL projections to thalamus (Do-Monte et al., 2015) and periaqueductal gray (Rozeske et al., 2018) mediate learned, but not innate, fear expression. Moreover, PL activation contributes to fear expression only after learning (Corcoran and Quirk, 2007). Thus, downstream areas maintain the same codes for valence or CRs, that they expressed before, with the PL afferents expanding their space.

Codes and Computations

Different situations often call for distinct behavioral strategies and correspondingly different computations (Headley et al., 2019). The activity patterns that accompany these processes will reflect these differences and constrain the kinds of information that downstream areas can extract. Our results indicate that PL encodes a wide variety of contingencies, that LA preferentially represents valence, and BL lies between PL and LA. Prefrontal areas such as PL are important for planning and executing context-dependent CRs (Hyman et al., 2012; Ma et al., 2016; Moorman and Aston-Jones, 2015), while LA links stimuli with their value (Repa et al., 2001; Tye et al., 2008). This gradient, from the generic (PL) to the specific (LA), may reflect the distinct cognitive demands placed on these regions (Mobbs et al., 2020). And yet these areas are coactive during even the simplest CRs (e.g., Vetere et al., 2017). Thus, the gradients of representations we observed across the amygdalo-prefrontal network may enable the sharing of information and provide parallel portraits of the task, allowing flexibility and redundancy in the production of emotional behaviors.

STAR★METHODS

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Denis Pare (pare@andromeda.rutgers.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

The full dataset and custom MATLAB code have not been deposited in a public repository because of their size, but are available from the corresponding author on request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

The Institutional Animal Care and Use Committee of Rutgers University approved the procedures described below. Experiments were conducted in compliance with the Guide for the Care and Use of Laboratory Animals. Twelve adult Long Evans rats (RRID: RGD_2308852) of either sex (280 – 350 g) (Charles River Laboratories) were housed individually with ad libitum access to food and water and maintained on a 12 h light/dark cycle. Rats were first habituated to the animal facility for one week and then to handling (15 minutes daily for 3–5 days). Prior to training in the RRI task, rats were placed on water restriction while ensuring that their body weight remained at ≥ 85% of initial values. Water restriction consisted of six consecutive days of restriction followed by one day of continuous access to water. Experiments were conducted during the light cycle.

METHOD DETAILS

Risk Reward Interaction Task

The behavioral apparatus (Figure 1) was a dimly rectangular arena (90 cm in length × 30 cm in width; 10 Lux) with walls 60 cm in height, a floor made of metal bars spaced 8 mm (5 mm outer diameter), and no ceiling. At both ends of the arena was a water port (Figure 1A, blue) with light-emitting diodes (LED) behind it (Figure 1A right; appetitive trial). The floor was divided into three sectors, each 30 cm × 30 cm), and with LEDs underneath (Figure 1A, left; yellow rectangle, aversive trial). A programmable microcontroller (Arduino, Spark-Fun, Niwot, CO) controlled LED activation, water delivery through the water ports (60 μl), and footshock administration in each floor sector (0.4 mA, 10 s).

LED activation (CS-R; 20 s) behind the active water port signaled availability of a liquid reward, 10 s after CS-R onset. Water was delivered through a dipper that retracted at the end of the CS-R. LED activation below one of the floor sectors (CS-S; 20 s) indicated an impending footshock (0.4 mA), starting 10 s after CS-S onset. The shock and CS-S co-terminated. When rats were in the shock sector at the onset of the CS-S, they could avoid the shock by moving to a different sector, a behavior termed active avoidance. If rats were in a different sector at the onset of the CS-S, they could passively avoid the shock. A video camera (40 frames/s) was positioned above the apparatus and continuously recorded the rats’ behavior. White noise masked ambient sounds.

One day prior to the start of training, rats were habituated to the arena for one hour with the lights off and only white noise. Over a period of one week, rats progressively learned to avoid shocks from each of the three floor sectors and retrieve water rewards at each of the two ports, as signaled by the position of the activated LED on each trial. On Day 1 and 2, rats were trained for two 15-min sessions (9 shock trials, 6 reward trials); the first in the morning and the second in the afternoon. Once rats learned to avoid the shocks, sessions were extended to 60 min (48 shock trials – 16 in each sector, 32 reward trials – 16 of each water port). Training continued until day 7 with 1 one-hour session per day until animals reached at least 80% performance, on all trial types combined (Figure 1B). All trials were reinforced with USs (rewards or shocks) during the training and recording sessions. No CSs or USs were delivered during the inter-trial intervals.

Surgical procedures

Once rats reached criterion, they were implanted with silicon probes aimed at the mPFC, BLA, or both. Rats were anesthetized with isoflurane mixed with O₂ and they received atropine sulfate (0.05mg/kg, i.m.) to facilitate breathing. Using nonpuncture ear bars, rats were placed in a stereotaxic frame. The regions of the scalp to be incised were injected with bupivacaine (S.C.), a local anesthetic. Fifteen minutes later, an incision was made to expose the skull and a craniotomy performed over the mPFC or BLA. Silicon probes (Neuronexus) were attached to a 3D-printed microdrive, and aimed at the top of the mPFC or BLA (coordinates in mm relative to bregma: mPFC, AP +2.52 to +3.92, ML 1.0, DV 2.0, 10° angle; BLA, AP −2.2 to −3.6, ML 5.2, DV 6.0). In the new rats used in this study, we used custom-designed probes in the BLA that had 4 shanks with 16 channels on each shank arranged in tetrode formations. Each tetrode was 333 μm from the neighboring tetrode and the inter-shank distance was 250 mm. For the mPFC, we used Buzsaki64L probes, which consisted of eight shanks spaced 200 μm apart. See Kyriazi et al. (2018) for the probes used in the previously recorded cohort of rats in the BLA. Rats were allowed to recover from the surgery for one to two weeks.

Unit recordings and clustering

Unit recordings began after recovery from the surgery. The silicon probe was lowered ≥ 140 mm after every recording session to avoid recording the same cells across days. Electrophysiological data was recorded using Intan Technologies equipment (RHD2000-Series amplifier evaluation system; intantech.com). Data was sampled at 30 kHz and stored on a hard drive. A high pass filter (300 Hz) followed by a median filter (median amplitude subtracted from all channels at each sample point) were applied to the data. Next, Kilosort (https://github.com/cortex-lab/KiloSort) was used to identify spike waveforms by matching them to template waveforms as described in Pachitariu et al. (2016). Finally, spike clusters were refined manually using Klusters (Hazan et al., 2006). Custom code was written to convert Kilosort output files to Klusters-compatible files to allow for manual refinement of clusters.

To separate clusters reliably, we examined autocorrelograms and cross-correlograms of clusters. Autocorrelograms had to show a refractory period of ≥ 2 ms. Cross-correlograms could not have a refractory period as this betrayed sharing of the same units between clusters. Units with unstable shapes were excluded.

For each unit, spike duration was calculated by selecting the channel with the largest peak-to-trough amplitude. The interval between spike trough and peak was defined as the spike’s duration (Barthó et al., 2004). Presumed fast-spiking interneurons and projection cells recorded in mPFC were classified based on their baseline firing rates (10 Hz cutoff) and trough-to-peak duration (0.5 ms cutoff). BLA units were classified as presumed fast-spiking interneurons or projection cells based on their baseline discharge rates (cutoff of 6 Hz) and trough-to-peak duration (cutoff of 0.55 ms). Striatal units were divided into presumed medium spiny neurons or fast-spiking interneurons based on interspike interval (ISI) and trough-to-peak durations (cutoff of 0.55 ms). After Berke (2008), highly active cells with less than 2% of their ISIs longer than 1 s were classified as interneurons. Units that did not meet both criteria were considered unclassified and excluded from analyses. BLA and striatal units reported in this study include cells from Kyriazi et al. (2018), and any additional units obtained from the 3 rats in the current study with a silicon probe aimed at the amygdala. See Kyriazi et al. (2018; Methods) for an extensive validation of the criteria used in the present study to classify neurons. Only principal cells were analyzed due to the low number of fast-spiking interneurons detected.

Histology

At the completion of the experiments, rats were anesthetized to make electrolytic marking lesions. Lesions were made on the most ventral electrodes of each shank (10 μA for 16 s). Rats were then perfused-fixed through the heart, their brains extracted, and cut on a freezing microtome. Brain sections (80 μm) were counterstained with a 1% thionine solution. Only neurons histologically determined to have been recorded in BLA, striatum, or the infragranular (2/3) layers of mPFC were analyzed.

QUANTIFICATION AND STATISTICAL ANALYSIS

Behavioral Analyses

Prior to each recording session, we attached red and green LEDs to the rat’s headcap. These LEDs allowed us to track the rats’ position and head-direction during the task. The green and red LEDs were always positioned at the caudal and rostral ends of the headcap, respectively. The midpoint between the red and green LEDs was defined as the head position whereas the angle between the midpoint and red LEDs with respect to the apparatus was taken as head direction. Behaviors were analyzed using a custom graphical user interface (GUI) in MATLAB (The MathWorks, Inc., Natick, Massachusetts, U.S.A). We considered four behaviors: reward approach, reward anticipation, active avoidance, and freezing. The start and end times of these behaviors were determined in the GUI by marking the first and last video frame when rats initiated and ended a behavior. For instance, the start of active avoidance was defined as the first frame when the rat began to move off the lit sector. Its end was denoted as the first video frame when the rat ended the continuous avoidance behavior.

Heatmap Stimulus Responses

Firing rates were z-scored to the 5 s baseline period just before CS-onset for each trial type, then ordered by the mean z-scored firing rate during the first second after CS-onset in descending order. The order of the cells was determined based on the magnitude of their response to CS-R1 and the same order was kept for CS-R2. The same applied to the CS-Ss: the cells were ordered based on their response to CS-S1 and the same order was kept for CS-S3.

Selective Cells

Firing selectivity was assessed with respect to CS-evoked activity (first second after CS onset) and behavior-related changes in firing rates. In the case of CS-evoked activity, a cell was considered selective if it increased or decreased its firing rate in response to only one of the two CS-Rs (i.e., excited by CS-R1, unresponsive to CS-R2, or inhibited during CS-R1 and unresponsive to CS-R2, and vice versa). Separately, the same determinations were performed for CS-Ss. With respect to behavioral selectivity, a cell was considered selective if it increased or decreased its firing rate in relation to RA or AA behaviors triggered by only one of the corresponding CSs (i.e., excited by RA CS-R1, unresponsive to RA CS-R2, or inhibited during RA CS-R1, unresponsive to RA CS-R2, and vice versa; excited by AA CS-S1, unresponsive to AA CS-S3, or inhibited during AA CS-S1, unresponsive to AA CS-S3, and vice versa).

Generalized Linear Model (GLM)

To fit the spiking of individual cells, we used a regularized regression, group Lasso, with Poisson distribution (Breheny and Huang, 2015; grpreg R package) and ten-fold cross validation, as described previously (Kyriaziet al., 2018). Briefly, action potentials were binned (50 ms) across the full recording session, indicating when stimuli and behaviors occurred with ones. Behavior and stimulus and events were convolved with basis functions, which were defined by log-time scaled raised cosine bumps separated by π/2 radians (50 ms). A linear combination of basis functions represented each event kernel. The Lasso penalty parameter to the Euclidean (L₁) norm was selected on the basis of the lowest cross-validation error and it was applied to each group, creating sparsity and variable selection at the group level (Breheny and Huang, 2015; Tibshirani, 1996; Yuan and Lin, 2006). For cross-validation, we divided the recording session into ten equal segments.

There were a total of 22 grouped parameters in the GLM: speed, position, CS-R1, CS-R2, CS-S1, CS-S2, CS-S3, US-R1, US-R2, US-S1, US-S2, US-S3, AA, Freezing, RA, RAnt, CS-R2 RA, CS-R2 RAnt, CS-S2 AA, CS-S2 Freezing, CS-S3 AA, CS-S3 Freezing. In order to avoid perfect collinearity with interaction terms, we did not include CS-R1 and CS-S1 variables interaction terms. Interaction terms were computed by multiplying the stimulus variable with the corresponding behavior variable (i.e., CS-S1 * AA = CS-S1 AA)

GLM Basis Functions and Kernels

The GLM kernels for each stimulus and behavior that best fit neuronal activity were constructed by combining a group of pre-determined basis functions, as described previously in Kyriazi et al. (2018). Briefly, two distinct sets of basis functions were used to represent CSs and CRs. Stimulus basis functions covered the initial 10 s, had a sharp onset and narrow width. Across time, they became smoother and wider, reflecting stimulus responses that tend to have sharp onsets and decay gradually. For behaviors, basis functions extended before and during the behavior onset to capture spiking activity related to the planning and execution of the behavior. However, they were bounded by the start of a trial. All interaction terms were treated as behavior basis functions. Then, functions were scaled by the model’s beta values and added within time bins to create a single kernel for each stimulus and behavior. If a cell did not encode a parameter, the group Lasso GLM gave beta values of 0 for the corresponding basis functions. Thus, the model only selects parameters that best fit observed spiking, allowing dimensionality reduction.

Similarity Matrices

Normalized peak firing rate modulations

The absolute peak of each CS and CR kernel was normalized by the baseline firing rate as follows: (Peak-Baseline)/Baseline. The sign of the modulation was preserved in the normalized peak in order to identify cells excited or inhibited in relation to each task feature. Normalized peak modulations ≤ 0.001 were considered non-significant and were set to zero. The average modulations to the task features were computed by taking the mean of the absolute value of all significant peak modulations for each parameter.

Similarity Matrices

We computed similarity matrices using a Spearman correlation of the peak firing rate modulations by all stimuli and behaviors across cells.

Gini Index

To understand how the code was distributed across the population we used the Gini index. This index quantifies the coding distribution across the neuronal population and ranges from 0 to 1, with 0 representing a perfectly equally distributed code among neurons, and 1 representing a perfect inequality of code distribution. The Gini index is calculated using the following function:

$$Giniindex=(n+1-2*\frac{{\mathrm{\Sigma }}_{i=1}^n(n+1-i)\mathrm{*}{\mathit{Y}}_i}{{\mathrm{\Sigma }}_{i=1}^n{Y}_i})\mathrm{*}\frac{\mathrm{1}}{n-1}$$

where Y_i represents the absolute peak firing rate modulation for each task feature i, to the total number of task features n. The Gini index was computed for each neuron individually based on the GLM firing rate modulation to each task variable. The distribution of the Gini indices across all cells in each structure is plotted in Figure 6B (top). A rank-sum test was used to test whether the Gini index distributions differed across brain regions (p < 0.001).

Dimensionality Index

To identify how many features cells encoded, we used the significant normalized peak firing rate modulations of each cell for all task features: CS-R1, CS-R2, CS-R1 RA, CS-R1 RAnt, CS-R2 RA, CS-R2 RAnt, CS-S1, CS-S3, CS-S1 Freezing, CS-S1 AA, CS-S3 Freezing, CS-S3 AA. The number of non-zero modulations were counted for each cell and that value represented the number of features encoded. The dimensionality index was then plotted in Figure 6B (bottom) as the distribution of the number of features encoded across all cells.

D-prime Analyses

We used the d-prime (d’) metric to assess the separation between different coding dimensions using the distributions of correlation coefficients in the similarity matrices (Keene et al., 2016; McKenzie et al., 2014). The coding dimensions tested were valence (CS-R1, CS-R2, CS-R1 RA, CS-R1 RAnt, CS-R2 RA, CS-R2 RAnt versus CS-S1, CS-S3, CS-S1 Freezing, CS-S1 AA, CS-S3 Freezing, CS-S3 AA), and active versus passive behaviors (CS-R1 RA, CS-R2 RA, CS-S1 AA, CS-S3 AA versus CS-R1 RAnt, CS-R2 RAnt, CS-S1 Freezing, CS-S3 Freezing). The d’ metric was calculated for each coding dimension using the following formula:

$$\mathit{d}\prime =\frac{{\mathit{\mu }}_{within}-{\mathit{\mu }}_{\mathit{b}etween}}{\sqrt{\frac{1}{2}\mathrm{(}{\mathit{\sigma }}_{\mathit{w}ithin}^{\mathrm{2}}+{\mathit{\sigma }}_{\mathit{b}etween}^{\mathrm{2}}\mathrm{)}}}$$

Where μ_within and μ_between are the means of the correlation coefficients for the within and between coding dimension features, respectively, and their corresponding variances, ${\mathit{\sigma }}_{within}^{\mathrm{2}}$ ands ${\mathit{\sigma }}_{between}^{\mathrm{2}}\mathrm{.}$. To compare if two d’ values between two brain regions were significantly different from each other, we computed a shuffled distribution of d’-difference values (10,000 permutations). If the observed d’-differences were > 95% of the shuffled distribution then the d’-differences were significant.

Principal Component Analysis (PCA)

To visualize the coding of task dimensions at the population level for each region, we mapped the firing rate into a low dimensional space (3-dimensions) created using principal component analysis (PCA) across the trial duration (Rozeske et al., 2018; Zhang and Li, 2018). Spiking was extracted from 5 s before to 10 s after CS-onset. The firing rates for baseline, CS-onset, and behaviors (AA and RA) were normalized by binning the firing rate into 20 bins for each time epoch. This time normalization was conducted to ensure that spiking to behaviors was always aligned despite trial-to-trial variations in the timing of behaviors. The binned spikes were then z-scored using the 5 s baseline period and averaged across trials. A boxcar function convolution was applied to the z-scored spiking before conducting the PCA analysis. The first 3 dimensions of the PCA scores were smoothed with a Gaussian window of 10 bins for visualization. All further calculations using the PCA scores were applied to the raw, unsmoothed scores.

Euclidean Distance Based on PCA Scores

The 3-dimensional Euclidean distance was calculated at every time bin between each trial type (CS-R1, CS-R2, CS-S1, and CS-S3) using the raw, unsmoothed PCA scores. The distances across time were then averaged to calculate a mean distance value for each trial type comparison, which is plotted in Figure 7B.

Support Vector Machine Decoder

Support vector machine analyses were conducted in a time-normalized fashion similar to PCA, but instead of creating 20 bins for each time epoch, 1 bin was used, resulting in 3 bins total (baseline period, CS-onset period, active behavior period (AA or RA)). The SVM decoder was conducted similarly to Meyers et al. (2008). Briefly, a pseudo-population of 150 cells with replacement was selected from each brain region. For each cell, 5 trials of each type were selected (5 CS-R1, 5 CS-R2, 5 CS-S1, and 5 CS-S3 trials). For the trial type comparison, the two appetitive trial types were compared to each other, and the two aversive trial types were compared to each other. This process was repeated 50 times. The appetitive and aversive decoder performance repetitions were combined to create the trial identity category. The mean decoder accuracy and SEM were computed using the 100 total repetitions. For the valence category, the CS-R responses were compared to the CS-S responses using 10 trials of each type.

To test for significantly different decoding accuracies between brain regions, we bootstrap resampled the decoder accuracy repetitions of two brain regions 10,000 times, and each time calculated a difference score. If the 95^th percentile of this null distribution crossed zero then the difference was not considered significant, if it did not cross zero then it was significant.

For the shuffling procedure, the steps were the same as above with the exception of the trial labels being shuffled randomly among the trials before being fed into the decoder. This destroyed any association between observed spiking and trial label, resulting in decoding accuracy at chance levels.

Population size decoder

The size of the pseudo-population was varied from 1, 5, and 10 to 150 cells in increments of 10 cells. Decoder performance was calculated with each pseudo-population size and the mean and SEM were calculated from the 50 repetitions. The shuffling procedure was the same as described above.

EXPERIMENTAL DESIGN

No explicit replication strategy (beyond running multiple animals on multiple training sessions) was performed for this study. No treatments were administered to the animals that required randomization, nor did we stratify subjects for any of the analyses. The experimenters were not blind to the experimental conditions. Sample sizes were chosen based on previous studies examining similar effects (Kyriazi et al., 2018). Animals that failed to reach the pretraining criterion were excluded from this study.

Statistical Analyses

Group data are reported as mean ± SEM. Details of our statistical tests, such as test statistic, degrees of freedom, and p value can be found in the Results section. Unless otherwise specified, statistical tests were performed across cells, with degrees of freedom reflecting either the number of cells or the number of levels for a factor that was being analyzed. All neurons classified as presumed principal cells were included. Firing rates reported logarithmically are natural logarithms. All statistical tests were two sided. Different procedures were used to assess statistical significance depending on the type of data, as specified above. Wherever possible to evaluate significance we used non-parametric statistical tests such as chi square, bootstrap, or permutation. For the group-lasso regression model fitting we used an exponential link function and poisson distribution because the distribution of spiking data is traditionally treated as poisson.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Experimental Models: Organisms/Strains
Long Evans rat	Charles River Laboratories	RGD_2308852
Software and Algorithms
MATLAB	MathWorks	MATLAB R2018a,b
RStudio	https://rstudio.com	Version 1.2.1335
Kilosort	https://github.com/cortex-lab/KiloSort	N/A
KlustaKwik	http://klustakwik.sourceforge.net/	N/A
Klusters	Hazan et al., 2006	N/A
Other
Silicon probe	NeuroNexus	Buzsaki64L Buzsaki32L a4x4tet_10.6mm333_250_121
Intan amplifier	Intan technologies	N/A

Highlights.

• Neurons in LA, BL, and PL encode the same set of emotional variables

• BL and PL neurons tend to encode more events with minor changes in firing rate

• LA neurons tend to encode fewer events with larger changes in firing rate

• LA and BL populations encode valence, while PL encodes trial identity