Introduction

In daily life, we rely on environmental cues to choose the course of action that would maximize positive outcomes and minimize adverse outcomes. This process requires tracking the discrepancies between received and expected outcomes, known as prediction errors(1–4). Prediction errors control how much a given outcome changes the expectation of subsequent outcomes, such that the larger the prediction error, the greater the change. This process, however, becomes complicated when outcomes follow cues or actions unreliably(5–7). Under such noisy conditions, even correct cues and actions occasionally result in large prediction errors. Such anomalies may indicate the need to accelerate learning to cope with a sudden, drastic environmental change. Alternatively, they may fall within a typically expected range of outcome variability and should be dismissed.

These two conflicting interpretations of surprising outcomes lead to “unexpected” and “expected” uncertainty, respectively(5–7). Humans and rodents can use both types of uncertainty to adjust the learning rate properly, and this essential ability relies on neuromodulatory systems, such as serotonin(8–13), noradrenaline(14–21), dopamine(22–26), and acetylcholine (ACh)(20,27–31). Parallel evidence also suggests that adaptive learning and decision-making under uncertainty rely on the integrity of the anterior cingulate cortex (ACC) in humans and monkeys(7,32–35) and its homolog, the medial prefrontal cortex (mPFC), in rodents(36–39). Neurons in these regions track outcome expectations by drastically changing the activity patterns depending on how reliably outcomes follow cues(40–44) or actions(45–48). They also signal prediction errors by scaling the magnitude of responses to outcomes according to how surprising it was(41,49–54). However, how these two systems interact to enable the proper handling of ambiguous outcomes remains unknown.

Here, we investigated the potential role of ACh in modulating prefrontal neural ensemble dynamics during adaptive learning and decision-making in a stable but probabilistic environment. We focused on ACh because (i) theories implicate ACh in computations of expected uncertainty(5,55), (ii) the mPFC receives extensive axonal projections of cholinergic cells in the basal forebrain (BF)(56–58), and (iii) these cholinergic cells emit phasic responses precisely time-locked to reward and punishment(58–65). Here, combining in vivo one-photon calcium imaging of mPFC cells with optogenetic manipulations of local cholinergic terminals in mice, we have shown that threat-evoked phasic cholinergic signals control the content of prediction error coding in the mPFC and limit the learning from surprising threats that are within an expected range of variability.

Results

Phasic cholinergic signals modulate the development of adaptive threat expectations

To investigate how threat-evoked phasic cholinergic signals modulate adaptive threat learning and its neural correlate in the mPFC, we designed a probabilistic spatial learning task in which mice chose one of two paths associated with different probabilities of threats to obtain rewards. Hungry mice went back and forth between two reward sites at the opposite corners of a square-shaped maze (Fig 1A). They always received sugar pellets at the reward sites [P(Reward) = 100%]; however, they also occasionally received an air puff toward the face/body in the middle of two paths connecting the reward sites (threat sites). On one of the paths (high-threat path, ht-p), an air puff was delivered 75% of the time [P(Threat|ht-p) = 75%], whereas on the other path (low-threat path, lt-p), it was delivered 25% of the time [P(Threat|lt-p) = 25%]. The mice received air puffs at the same probability on a given path regardless of which direction they ran. The timing of reward and threat delivery was controlled based on real-time tracking of mouse’ position.

Fig 1.

During the probabilistic spatial learning task, we conducted cellular-resolution calcium imaging from the prelimbic region (PL) of the mPFC with simultaneous optogenetic manipulations of the BF cholinergic terminals. We stimulated, rather than inhibited, cholinergic terminals for two reasons. First, because the relationships between ACh level and behavior follow an inverted U-shape, in which abnormally high ACh is as detrimental as abnormally low ACh(66,67). Second, we have shown that the stimulation of threat-evoked phasic cholinergic signals in the PL prevents mice from associating the threat with a preceding cue, while their inhibition facilitates association formation(58). Thus, the stimulation, but not inhibition, of threat-evoked cholinergic signals disrupts prefrontal neural processing during threat learning.

To limit the expression of excitatory opsins to cholinergic cells, the viral-mediated gene transfer approach was applied to mice expressing Cre recombinase in cholinergic cells (ChAT_(IRES)-Cre mice; Fig 1B). The mice were injected with conditional adeno-associated viral vectors (AAVs) for expressing a red-shifted excitatory opsin, ChrimsonR, with a florescent reporter (ChrimsonR group, n = 6) or the reporter alone (tdTomato group, n = 4) in the horizontal diagonal band of Broca (HDB) region of the BF (Figs 1B, S1A and S1B; based on our previous anatomical tracing result(58)). All mice were also injected with AAVs carrying a calcium indicator (GCaMP6f) under the control of a CaMKII promotor in the PL. Histological validation confirmed that the expression of ChrimsonR overlapped with cholinergic cells expressing ChAT, and their axons were located near putative excitatory cells expressing GCaMP6f in the PL (Fig 1C).

After the recovery period, all mice underwent fifteen days of training, divided into three stages (Fig S1C), pre-training [T1-5; P(Threat|ht-p) = 0%, P(Threat|lt-p) = 0%], probabilistic [P1-5; P(Threat|ht-p) = 75%, P(Threat|lt-p) = 25%], and random stages [R1-5; P(Threat|ht-p) = 50%, P(Threat|lt-p) = 50%]. During the probabilistic and random stages, we stimulated local cholinergic terminals unilaterally by delivering orange LED stimulations (300 msec, 20Hz, 8 mW/mm²) to the imaged window. The LED stimulation was applied during ∼50% of the air-puff delivery on both paths, allowing for monitoring threat-evoked activity with and without the LED stimulation (Fig S1D). During the pre-training stage (T5), both groups chose two paths with similar likelihood (t-test, t(8) = 1.497, p = 0.173; Fig 1D). When air-puff delivery and LED stimulation commenced (P1-5), tdTomato-expressing mice learned to choose the low-threat path preferentially (two-way mixed ANOVA, Session×Group interaction, F(4,32) = 3.347, p = 0.021; follow-up one-way repeated measure ANOVA, tdTomato, F(4,12) = 12.469, p < 0.001; Fig 1D). Initially, they approached both threat sites at comparable speeds (Fig S2A). After learning, they ran more slowly on the high-threat path than the low-threat path (three-way mixed ANOVA, Group×Path interaction, F(1,8) = 9.225, p = 0.016; follow-up t-test, tdTomato, path type, t(38) = 5.576, p < 0.001; Figs 1E and 1F). In contrast, ChrimsonR-expressing mice continued to frequently choose the high-threat path even after five days of training (ChrimsonR, F(4,12) = 1.837, p = 0.161; Fig 1D) and did not differentiate the running speed between the paths (ChrimsonR, path type, t(58) = 1.495, p = 0.140; Figs 1E, 1F and S2A). In the random stage (R1-5), both groups gradually reduced the probability of choosing the original low-threat path and no longer showed any differences in the path choice (Session×Group interaction, F(4,32) = 0.588, p = 0.674; Session, F(4,32) = 7.747, p < 0.001; Group, F(1,8) = 2.269, p = 0.170.; Fig 1D) or speed (Fig S2A).

Because tdTomato-expressing mice learned to differentiate the speed toward the threat sites depending on threat probabilities (Fig 1E), we used the speed variations across laps to gauge the mice’s expectation of threats: a slower speed indicated that they expected to receive threats, while a faster speed implied that they did not (Fig 1G). The comparison of the running speed between these two lap types revealed that the mice reacted differently to the same outcomes depending on their expectations. Specifically, mice accelerated more robustly after surprising threats than expected threats (Fig 1H, left). They also took longer to accelerate after surprising omissions than expected omissions (Fig 1H, right). Similar patterns were also detected in P1 and R5 (Fig S2B). In all these sessions, cholinergic terminal stimulation did not affect the reactions to expected or surprising outcomes (Threat: three-way mixed ANOVA, Session×Expectation×Group interactions, p = 0.285; Session, F(2,16) = 5.626, p = 0.014; Expectation, F(1,8) = 37.918, p < 0.001; Group, F(1,8) = 0.039, p = 0.849; Omission: Session×Expectation×Group interactions, p = 0.360; Session, F(2,16) = 12.840, p < 0.001; Expectation, F(1,8) = 33.280, p < 0.001; Group, F(1,8) = 0.156, p = 0.703; Fig 1I). These findings suggest that phasic cholinergic signals do not modulate the development of threat expectation but are essential for evaluating different probabilistic outcome contingencies.

Enhanced phasic cholinergic signals augment the threat-evoked activity of PL cells

To probe neural mechanisms behind the observed behavioral effects, we used a head-mounted single-photon epifluorescence miniature microscope to monitor Ca²⁺ activity of PL cells via a gradient-index (GRIN) lens (Fig 2A). The biological and optical crosstalk of optogenetic stimulation on calcium imaging is negligible(68). We first investigated how cholinergic terminal stimulation affected the threat-evoked activity of PL cells by using the data during the first session with probabilistic air-puff delivery with LED stimulation (P1; Fig 2B). ∼15% of PL cells increased Ca²⁺activity in response to the combined delivery of air puff and LED stimulation (Fig 2C; tdTomato, 14.7%; ChrimsonR, 17.9%). In the absence of LED stimulation, the magnitude of their puff-evoked activity was reduced in ChrimsonR-expressing mice but not in tdTomato-expressing mice (Fig 2C). Overall, the LED stimulation increased the puff-evoked activity more robustly in ChrimsonR-expressing than tdTomato-expressing mice (Kolmogorov-Smirnov test, w/ LED, p = 0.016; Fig 2D). This group difference was not detected without the LED stimulation (w/o LED, p = 0.386), confirming that the enhanced puff-evoked activity was due to the stimulation of the BF cholinergic terminals.

Fig 2.

We also examined the effect of cholinergic terminal stimulation on PL cell activity in the absence of air-puffs. In one pre-training session (T3), LED stimulation was applied when mice entered the corner of the maze, a location that would become the threat site in subsequent days (Fig S3A). Some cells changed their activity at the corner, and the LED stimulation disrupted the location-selective activity. However, the degree of the LED-induced change was comparable between the groups (Kolmogorov-Smirnov test, p = 0.237; Fig S3B). Thus, the optogenetic stimulation barely affected PL cell activity unless it coincided with the natural activation of cholinergic cells by threats(58,59).

Enhanced phasic cholinergic signals suppressed learning-dependent changes in outcome-related PL cell activity

Prefrontal cells are known to change responses to outcomes throughout learning, indicating that they encode values of expected outcomes(69–73) and the discrepancies between expected and actual outcomes (i.e., prediction errors)(41,49–54). Therefore, we next asked how PL cells changed outcome-related activity with learning and how this process was affected by cholinergic terminal stimulation. We categorized laps into four types based on which path the mice took and whether they received an air puff. For laps with air-puff delivery, we only used ones without LED stimulation to isolate the learning-dependent changes from the direct effects of LED stimulation (Figs 2C and 2D). In each lap, the cell activity was aligned to the moment when the mouse’s position tracking entered the threat site. In the first (P1) and last (P5) sessions of the probabilistic stage, a sizable proportion of PL cells changed their activity in response to threat delivery or omission (Figs 3Aa and Ba). In both groups, cells responding to outcomes on the high-threat path were distinct from those on the low-threat path (P1, Fig S4A; P5, Fig 3Ab and Bb), suggesting that cholinergic terminal stimulation did not affect the differentiation of outcome representations between two threat sites. In tdTomato-expressing mice, the omission-evoked activity on the high-threat path was shifted toward excitation from P1 to P5, while the threat-evoked activity was unchanged (Kolmogorov-Smirnov test, threat on ht-p, p = 0.485; omission on ht-p, p = 0.005; Fig 3Ac). Threat- or omission-evoked activity on the low-threat path was not changed across sessions (threat on lt-p, p = 0.106; omission on lt-p, p = 0.718). In ChrimsonR-expressing mice, the distribution of neither threat-nor omission-evoked responses was changed from P1 to P5 on either path (threat on ht-p, p = 0.265; omission on ht-p, p = 0.055; threat on lt-p, p = 0.279; omission on lt-p, p = 0.473; Fig 3Bc). The comparisons between the groups revealed that cholinergic terminal stimulation had no effects on threat- or omission-evoked activity in P1 (threat on ht-p, p = 0.731; omission on ht-p, p = 0.682; threat on lt-p, p = 0.404; omission on lt-p, p = 0.142; Fig S4B, left); however, it blocked the learning-dependent increase in the responses to rare threats omissions on the high-threat path (omission on ht-p, p < 0.001; Fig S4B, right). Thus, enhanced phasic cholinergic signals do not change how PL cells respond to subsequent outcomes within minutes; instead, they modulate the gradual restructuring of outcome representations over days.

Fig 3.

Phasic cholinergic signals augmented the selectivity of PL cells for signed prediction errors

The strengthening of cell activity to rare threat omission alluded to the sensitivity of PL cells to the discrepancy between expected and actual outcomes, i.e., prediction errors. To further address this point, we investigated whether PL cells differentiated the outcome-related activity depending on the lap-by-lap fluctuations in threat expectations. Using the speed toward the threat site (Fig 1G), laps with air-puff delivery were split into surprising (faster speed) and expected (slower speed) threat laps. In parallel, laps with air-puff omission were split into surprising (slower speed) and expected (faster speed) omission laps. We then quantified the degree to which individual cells differentiated the outcome-related activity depending on expectations by subtracting the activity for expected outcomes from the activity for surprising outcomes. We observed strong differentiations of threat- and omission-evoked activity in all three sessions (Fig 4A). In both groups, cells showed diverse patterns of differential activity (Fig 4B). Some cells increased the activity in response to surprising threats but decreased the activity in response to surprising threat omission (Fig 4C, Up-Down), while others showed reversed patterns (Down-Up). Thus, these cells changed the response direction based on whether outcomes were better or worse than expected, a signature of signed prediction errors. In parallel, other cells showed the same direction of responses to surprising threats and surprising omission (Up-Up, Down-Down), suggesting that they were sensitive to unsigned prediction errors. In parallel, other cells showed expectation-dependent differentiations only to threats or their omissions (Fig S5). In P1, the proportion of cells encoding signed prediction errors was comparable to that encoding unsigned prediction errors (Fig 4D, left). These cells were much fewer than others that encode errors of one outcome type. All these cell types were detected in both groups in similar proportions. In P5, however, cholinergic terminal stimulation increased the proportion of cells encoding signed prediction errors (binominal test, p < 0.05; Fig 4D, middle). As a result, cells that encoded signed prediction errors became twice as common as those that encoded unsigned prediction errors. The overall proportion of cells encoding prediction errors was slightly increased. Similar patterns were also detected during R5 (binominal test, p < 0.05; Fig 4D, right). These findings suggest that enhanced phasic cholinergic signals augmented the encoding of signed prediction errors in the PL.

Fig 4.

Enhanced phasic cholinergic signal impairs adaptive learning of uncertain, but not certain, threats

The analyses of cell activity revealed that enhanced threat-evoked cholinergic signals modified the encoding of prediction errors in the PL. To further investigate how such changes in PL outcome coding affected behavior, we conducted follow-up behavioral experiments with bilateral excitation of the BF cholinergic terminals in the PL. AAVs were infused into the HDB of ChAT-cre mice to express either channelrhodopsin (ChR2, n = 10), ChrimsonR (n = 8) with control fluorophore, or control fluorophore alone (No-opsin, n = 14) in cholinergic cells (Figs S6A, 5A and 5B). Their terminals in the PL (Fig S6B) were excited during every air-puff delivery in all sessions of the probabilistic paradigm [Fig 5Ca; P(Threat|ht-p) = 75%, P(Threat|lt-p) = 25%]. Consistent with the effect of the unilateral manipulation (Fig 1D), enhanced threat-evoked cholinergic terminal activity impaired the development of preference for the low-threat path (two-way mixed ANOVA, Session×Group interaction, F(8,116) = 2.249, p = 0.029; follow-up one-way repeated measure ANOVA, No opsin, F(4,52) = 11.190, p < 0.001; ChR2, F(4,36) = 0.627, p = 0.646; ChrimsonR, F(4,28) = 0.856, p = 0.502; Figs 5Cb and 5Cc). Because the task performance of the ChR2 and ChrimsonR groups was comparable, these groups were collapsed into a single group (opsin group) in subsequent analyses. Closer examinations of the movement pattern revealed that mice in the no-opsin group took longer to reach the threat site on the high-than the low-threat path (three-way mixed ANOVA, Path×Group×Session interaction, F(4,96) = 0.508, p = 0.730; Path×Group interaction, F(1,24) = 13.54, p = 0.001; follow up t-test, no-opsin, path type, t(108) = 3.979, p < 0.001; Fig S7Aa). In contrast, mice in the opsin group took comparable time on both paths (path type, t(148) = 0.570, p = 0.569; Fig S7Aa). In contrast, the time taken from the threat site to the reward site was similar between the groups, and it became shorter across the sessions (Session, F(4, 96) = 5.091, p < 0.001; Group, F(1, 24) = 3.846, p = 0.062; Path, F(1, 24) = 0.297, p = 0.591; Fig S7Ab). Similarly, both groups initially increased and later decreased the time to initiate a subsequent lap (Session, F(44, 96) = 7.474, p < 0.001; Group, F(1, 24) = 0.529, p = 0.474; Path, F(1, 24) = 0.653, p = 0.427; Fig S7Ac).

Fig 5.

As outlined in the introduction, the control of learning rate is particularly essential when threats occur unreliably. To test whether cholinergic modulation of the PL is vital for probabilistic threat learning or threat learning in general, we conducted the same experiment with the deterministic paradigm [P(Threat|ht-p) = 100%, P(Threat|lt-p) = 0%]. We observed that all three groups preferred the path without threats (no-threat path) over the path with threats (threat path; two-way mixed ANOVA, Session×Group interaction, F(4,24) = 3.195, p = 0.031; follow-up one-way repeated measure ANOVA, no-opsin, F(2,12) = 25.83, p < 0.001; ChR2, F(2,6) = 14.39, p = 0.005; ChrimsonR, F(2,6) = 7.199, p = 0.025; Fig 5D). In addition, both groups took longer to reach the threat site on the threat than no-threat path (three-way mixed ANOVA, Session, F(1, 13) = 0.034, p = 0.857; Group, F(1, 13) = 0.758, p = 0.399; Path, F(1, 13) = 26.38, p < 0.001; Fig S7Ba). They also took longer to reach the reward site on the no-threat than threat path (Session, F(1, 13) = 0.008, p = 0.930; Group, F(1, 13) = 0.025, p = 0.877; Path, F(1, 13) = 73.320, p < 0.001; Fig S7Bb). On the contrary, both groups increased the time to initiate a subsequent lap after choosing the no-threat path (Session, F(1,13) = 5.182, p = 0.040; Group, F(1, 13) = 0.089, p = 0.770; Path, F(1, 13) = 48.28, p < 0.001; Fig S7Bc).

Intact learning in the deterministic paradigm (Fig 5D) suggests that enhanced phasic cholinergic signals did not make the mice insensitive to threats or incapable of learning threat locations. To identify specific deficits underlying this behavioral change, we investigated how the mice adjusted their path choice after experiencing threats in the probabilistic paradigm. After receiving an air puff on the high-threat path, both groups chose the other path in the subsequent lap starting from the first session (two-way mixed ANOVA, Session, F(4,56) = 1.054, p = 0.388; Group, F(1,14) = 0.131, p = 0.723; Fig 6A, left). On the other hand, after receiving an air puff on the low-threat path, both groups initially chose the other path in the subsequent lap. Over five days of training, however, the no-opsin group learned to stay on the low-threat path, whereas the opsin group continued to switch to the high-threat path (two-way mixed ANOVA, Session×Group interaction, F(4,44) = 3.442, p = 0.017; follow-up one-way repeated measures ANOVA, No opsin, F(4,16) = 3.291, p = 0.038; Opsin, F(4,28) = 0.579, p = 0.680; Fig 6A, right).

Fig 6.

The above analysis revealed that the no-opsin group learned to ignore rare threats on the low-threat path, but the opsin group did not. This observation led us to theorize that the no-opsin group estimated threat probability by integrating threat occurrences across multiple laps, and this ability might be impaired in the opsin group. To test this idea, we correlated a sequence of path choices against the threat probability estimated with the threat history over different numbers of laps (Fig S7C). Consistent with our view, the path choice of the no-opsin group showed stronger correlations with the estimated threat probability as the greater number of laps was used to calculate threat probability (Fig 6B, left). In contrast, the path choice of most mice in the opsin group was not correlated with the estimated threat probability regardless of the number of laps used to calculate the probability (right). The proportion of mice with significant correlations was greater in the no-opsin group than in the opsin group in both paths (chi-square test, ht-p, Χ²₁ = 4.073, p = 0.044; lt-p, Χ²₁ = 7.748, p = 0.005; Fig 6C). These observations suggest that abnormally strong cholinergic signals during threats made mice adjust their choices after each threat without tracking its occurrence over time. This would make them highly sensitive to a surprising threat, even when it falls within the expected range of variability.

Discussion

The mPFC has been implicated in the evaluation of probabilistic outcome contingencies (36–39) and contains cells sensitive to outcome expectations(69–73) and prediction errors(41,49–54). Here, we have shown that this function and coding property are modulated by phasic cholinergic signals time-locked to outcomes.

During our probabilistic spatial learning task, PL cells robustly responded to threats and their omission (Fig 3Aa). From the beginning, PL cells differentiated outcome-related activity for the two threat sites (Figs S4A and 3Ab), indicating the conjunctive code of outcome and its location. With subsequent learning, omission-evoked activity was strengthened, while threat-evoked activity was unchanged. Notably, such strengthening occurred only for rare threat omissions on the high-threat path but not for frequent omissions on the low-threat path (Fig 3Ac), paralleling the gradual differentiation of threat expectations for these paths (Figs 1D and 1F). Furthermore, we also observed that outcome-evoked responses were sensitive to mice’s outcome expectations on each lap (Fig 4A). Although patterns of activity differentiation were diverse (Figs 4B, 4C and S5), ∼10% of PL cells responded to better-than-expected and worse-than-expected outcomes. Half of them reversed the response direction (i.e., increased v.s. decreased activity) for positive and negative errors, indicating the selectivity for signed prediction errors. Meanwhile, the remaining showed the same response direction for both error types, indicating the selectivity for unsigned prediction errors. These findings align with the selectivity of prefrontal cells for reward prediction errors(41,49–54) and expand the critical role of mPFC in prediction error coding to aversive outcomes.

Among these coding properties of PL cells, cholinergic terminal stimulation 1) abolished the learning-dependent increase in the responses to rare threat omission (Figs 3Bc and S4B) and 2) increased the proportion of cells encoding signed prediction errors (Fig 4D). Notably, these effects emerged over days but not across minutes within the first day, suggesting that enhanced cholinergic signals did not immediately change how PL cells responded to subsequent threats. Instead, it modulated the gradual restructuring of outcome representations with learning. An open question is how phasic cholinergic signals achieve such modulation. Cholinergic modulation of the mPFC is highly complex because both nicotinic and muscarinic ACh receptors are expressed in pyramidal cells and interneurons in a layer-dependent manner(74,75). Indeed, in vitro slice preparations, the application of ACh inhibits pyramidal cells in superficial layers but excites pyramidal cells in deep layers(76–78). These mixed effects at least partially explain why our optogenetic stimulation did not result in a drastic increase in PL cells responding to threats (Fig 2) or the corner of the maze (Fig S3).

In parallel to the modulation of cell excitability, ACh also controls the plasticity of glutamatergic synapses(79,80) by modulating glutamate release onto pyramidal cells(81,82) or inhibitory synaptic transmission(83). One potential function of such modulation is to adjust the relative impacts of various long-range inputs on PL cells. For example, some BLA cells encode unsigned prediction errors at the time of outcomes(84,85) and outcome-predictive cues(86). Notably, activating the M1 subtype of muscarinic ACh receptor induces long-term depression of field excitatory post-synaptic potentials in BLA-PL pathways(87). Therefore, enhanced cholinergic signals may reduce the influence of BLA inputs carrying unsigned prediction errors via the activation of M1 receptors, thereby increasing the impact of others carrying signed prediction errors, such as dopamine inputs from the ventral tegmental area (VTA). Although some controversy exists on how dopamine neurons respond to aversiveness(88,89), several studies have reported that dopamine neurons are inhibited by unexpected threats(90,91) and excited by unexpected threat omission(92). In addition, dopaminergic signals at the time of outcome modulate aversive learning(93,94). Investigating such hetero-synaptic plasticity would be a promising future direction to deepen our understanding of intricate cholinergic modulation of prefrontal neural computations.

Alternatively, phasic cholinergic signals may convey signed prediction errors by themselves and directly improve the selectivity of PL cells. However, this is unlikely for several reasons. First, BF cholinergic cells do not differentiate their responses between expected and surprising threats(59,65). Second, cholinergic terminals in the PL were activated by threats but not by surprising threat omission(58) (Fig S8). These observations suggest that phasic cholinergic signals covey the presence of aversive stimuli to PL cells but not prediction errors.

To decipher specific deficits that emerged from the modified prediction error coding in the PL, we conducted additional behavioral experiments (Fig 5) and analyzed mice’s choices on a lap-by-lap basis (Fig 6). We observed that cholinergic terminal stimulation did not affect the likelihood of correcting the choice after threats on the high-threat path (Fig 6A), suggesting that enhanced phasic cholinergic signals do not affect the learning from aversive outcomes. This observation explains why the manipulation did not affect learning in the deterministic paradigm, in which mice needed to correct the choice after every threat (Fig 5D). The intact learning also eliminates the possibility that cholinergic terminal stimulation affected the sensitivity to threats, the motivation to obtain the reward, the discrimination between the two paths, and the learning of threat locations. In contrast, the enhanced cholinergic signals reduced the likelihood of staying on the low-threat path after a surprising threat (Fig 6A). Also, it decoupled the choice from the estimated threat probability based on threat occurrence across multiple laps (Figs 6B and 6C). Theories link the handling of surprising outcomes to the processing of prediction error signals(7,35,95). Specifically, when threats occur reliably, the estimate of threat probability should be updated with each prediction error signal to reduce the magnitude of error signals (Fig 7). In contrast, when threats occur probabilistically, the magnitude of prediction error signals massively fluctuates depending on whether a threat occurred or not on each occasion. With such large variability, the estimate of threat probability should not be updated by error signals with unusually large magnitudes. Also, to assess whether an error signal is an anomaly, one must track the absolute magnitude of error signals over time(7,35,95). The present behavioral results are consistent with these theories and identify that threat-evoked phasic cholinergic signals serve as an essential input enabling PL cells to properly handle prediction errors for adaptive threat learning.

Fig 7

A question remains open as to whether this function is achieved solely by phasic cholinergic signals within the PL. Although we stimulated cholinergic terminals in the PL, the stimulation could retrogradely activate their cell bodies, leading to additional release of ACh at their terminals in other regions. Indeed, ∼75% of PL-projecting cholinergic cells also send axons to the anterior cingulate and orbitofrontal cortex(96). We, therefore, cannot eliminate the possibility that the coordinated release of ACh in multiple prefrontal regions might also contribute to the detected manipulation effects on PL cell activity and adaptive threat learning.

In conclusion, the evidence from cell activity and behavior suggests that the threat-evoked phasic cholinergic signal modulates the gain of learning from surprising threats by controlling the content of outcome coding in the PL. Our findings extend the cholinergic modulation of cognition to an intricate evaluation process of surprising outcomes and its control over adaptive learning and decision-making.

Materials and methods

Ethics statement

All methodological procedures were conducted under the regulation of the University of Toronto Animal Care Committee and the Canadian Council on Animal Care (AUP20012042).

Animals

Adult ChAT_(IRES)-Cre mice (RRID: IMSR_JAX:028861) were bred as homozygotes. Cre recombinase is expressed under the ChAT promotor in these mice without disrupting endogenous ChAT expression. In total, sixty-eight 8-20 weeks old male ChAT_(IRES)-Cre mice were used in optogenetics and calcium imaging experiments. Mice were housed under a 12-hour light/dark cycle with ad libitum access to food and water. Seven days before the behavioral testing, food was restricted to maintain 85-90% of free-feeding body weight for the experiments.

Viral vectors

AAV9-EF1a-DIO-hChR2(H134R)-EYFP-WPRE-HGHpA (RRID: Addgene_20298), AAV9-EF1a-DIO-EYFP (RRID: Addgene_27056), AAV5-Syn-FLEX-rc[ChrimsonR-tdTomato] (RRID: Addgene_62723) and AAV5-FLEX-tdTomato (RRID: Addgene_28306) were purchased from the Addgene. AAV1-CamKIIa-GCaM6f-WPRE-bGHpA ready-to-image virus (Catalog ID: 1000-002613) was purchased from the Inscopix (Bruker).

Surgery

Viral infusion

Eight to 12 weeks old ChAT_(IRES)-Cre mice were anesthetized (1-2% isoflurane by volume in oxygen at a flow rate of 0.8 L/min; Fresenius Kabi) and administered with meloxicam (20 mg/kg) subcutaneously. Mice were placed in a stereotaxic frame, and holes were made in the skull bilaterally above the basal forebrain (BF; +0.2 mm AP, ±1.3 mm ML)(99). A 33 G microinfusion cannula was connected to a 10-µl Hamilton syringe with Tygon tubing. One of the viral vectors (AAV9-EF1a-DIO-hChR2(H134R)-EYFP-WPRE-HGHpA, AAV9-EF1a-DIO-EYFP, AAV5-Syn-FLEX-rc[ChrimsonR-tdTomato] and AAV5-FLEX-tdTomato; 0.8-1.0 µl/side) was infused to the basal forebrain (BF; -5.4 mm DV) at a rate of 0.1 µl/min for optogenetic experiments. For imaging experiments, AAV5-Syn-FLEX-rc[ChrimsonR-tdTomato] or AAV5-FLEX-tdTomato (0.8-1 µl/side) was infused to the BF (-5.4 mm DV) at a rate of 0.1 µl/min. Another hole was drilled above the right medial prefrontal cortex (mPFC; +1.9 mm AP, + 0.35 mm ML). A 200 nl AAV1-CamKIIa-GCaM6f-WPRE-bGHpA was infused to the prelimbic region (PL; -2.2 mm DV) at a rate of 50 nl/min. After infusion, the cannula was left in the brain for 15 minutes, allowing the viral vectors to diffuse.

Optic fiber implantation

Three to 6 weeks after viral infusion, mice were placed in a stereotaxic frame as described above. Optic fibers (200 µm core diameter, 0.39 NA; Thorlabs) were implanted bilaterally at 30 degrees lateral to the medial angle above the PL (+1.9 mm AP, ±1.4 mm ML, -1.8 mm DV). The optic fibers were secured to the skull with resin cement (3M dental, Relyx unicem). The mice were given 14 days to recover from the surgery before behavioral testing.

Gradient index (GRIN) lens implantation

Three to 6 weeks after viral infusion, mice were placed in a stereotaxic frame as described above. A 1.2 mm diameter craniotomy was made above the mPFC (centered at +1.9 mm AP, + 0.6 mm ML). A 25 G blunt needle was slowly lowered above the PL (+1.9 mm AP, + 0.6 mm ML, -1.5 mm DV) at a rate of 10 µm/s to make a track for GRIN lens implant. Five minutes later, the 25 G blunt needle was slowly removed at a rate of 10 µm/s. A ProView™ integrated lens (1 mm diameter, 4 mm length; Inscopix) with baseplate attached was implanted 100-200 nm above the PL (+1.9 mm AP, + 0.55 mm ML, -1.7-1.9 mm DV; lowered at a rate of 2 µm/s). To prevent dust, the lens and basal plate were secured to the skull with resin cement (3M dental, Relyx unicem) and covered by a magnetic cover (Inscopix). The mice were given at least 21 days to recover from the surgery before behavioral testing.

Behavioral experiments

Probabilistic spatial learning

The apparatus was a 75 cm × 75 cm square-shaped maze with an 8 cm wide track and 15 cm walls. The inner wall was inclined 30 degrees outward to prevent cables from tangling at the corners. A camera was positioned centrally 1 m above the maze to monitor the mice’s movement. At the two opposite corners of the maze, a circle feeding trough (2 cm in diameter) was placed at the rear wall (reward site). A sugar pellet (20 mg, Bio-Serve) was manually delivered to the trough using a Tygon tube accessible through the opening (1 cm diameter) at the rear wall right above the trough. Additionally, two pieces of Tygon tubing (6 mm internal diameter and total length of 20 cm) were attached outside the two sides walls of the other two corners (each side 10 cm long, threat site). The tubes and side walls were perforated with 20 holes × 5 mm diameter. Using plastic Y connectors, these tubes were connected to solenoids and a compressed air cylinder (Whisper-tech). An air puff (300 ms duration, 35 psi) was delivered via this Tygon tubing. During the task, the mouse’s head, body, and tail locations were extracted from a video recorded at 30 frames per second by an overhead webcam (Ethovision software, Noldus Information Technology). The entry to the reward or threat sites were defined as the time when the mouse body center entered the pre-defined region of interest (reward site, 20 cm × 20 cm; threat sites, 8 cm × 20 cm), and the corresponding timestamps were stored. In a subset of laps, an air puff was delivered 100 ms after the entry to a threat site, at which Ethovision software emitted a TTL pulse to activate a solenoid valve connected to the tubes installed at that threat site.

During the first five days (pre-training stage), all mice ran on the maze to collect sugar pellets at the reward sites without any air-puff delivery. Subsequently, each mouse underwent one of three paradigms. In the probabilistic paradigm, air puffs were delivered 75% of the time at one threat site and 25% of the time at the other threat site. In the deterministic paradigm, air puffs were delivered 100% of the time at one threat site and 0% of the time at the other threat site. In the random paradigm, air puffs were delivered 50% of the time at both threat sites. Sugar pellets were always delivered in all stages regardless of path choice and air puff deliveries.

1-photon single-cell calcium imaging with optogenetic manipulations

Calcium imaging videos were recorded using a head-mounted miniature microscope (miniscope, Inscopix nVoke acquisition system v2.0) and accompanying software (IDAS, v1.7.1). The fluorescence power during imaging sessions ranged from 0.5-0.8 mW/mm² at a rate of 20 Hz, with an excitation wavelength of 455 ± 8 nm. Ethovision software was used to trigger and control the Inscopix Data acquisition (DAQ) box in synchrony with behavioral video tracking (30 fps) using a webcam installed near the ceiling.

Before imaging experiments, mice were handled by the experimenter for seven days. The mice were further habituated to the miniscope mounting procedure by attaching a dummy scope for 2-3 sessions (∼10 minutes long). During these sessions, focus depth was adjusted for each animal and was maintained consistently throughout the subsequent imaging sessions. During the same period, these mice were subjected to food restriction to maintain 85-90% of their free-feeding body weight.

The mice underwent the pre-training (5 days), probabilistic (5 days), and random (5 days) stages. To mitigate photobleaching, we only recorded images on a subset of sessions (Fig S1D). A dust cover was removed, and the implanted lens was cleaned with 75% ethanol and double-distilled water before attaching the miniscope to the mouse in each session. The mouse was placed in a cylinder (20 cm diameter, 25 cm height) in the maze’s center for 5 minutes to acclimate to the environment. Subsequently, the cylinder was removed, and the mouse was transferred to one of the reward sites, initiating the session automatically. Each imaging session lasted 40 minutes and comprised two LED-ON and two LED-OFF epochs. The order of epochs was counterbalanced across mice and sessions (ON-OFF-ON-OFF or OFF-ON-OFF-ON). During the LED-ON epoch, Arduino sent TTL signals to turn on solenoids to deliver air puffs when the mouse entered a threat site. Concurrently, another TTL signal activated an orange LED integrated into the miniscope for 300 ms (20Hz, 8 mW/mm², 590-650 nm wavelength). In the LED-OFF epoch, the orange LED was not activated concurrently with air puffs. In a subset of sessions (P2, P4, R2, and R4), the LED stimulation was applied during every air-puff delivery without capturing calcium imaging videos.

Optogenetic experiments

A separate cohort of mice first underwent the pre-training stage and moved on to the probabilistic or deterministic stage, during which they received bilateral optogenetic manipulations of cholinergic terminals in the PL. Daily sessions lasted 15-20 minutes or paused when mice ran over 60 laps. Each lap started when a mouse exited one reward site and ended when it entered the other reward site. A blue laser (300 ms, 20 Hz, 3-5 mW, 473 nm wavelength; Laserglow Technologies) was turned on during every air-puff delivery for mice expressing ChR2. An orange LED (300 ms, 20 Hz, 5 mW, 625 nm wavelength; M625F2, Thorlabs) was turned on during every air-puff delivery for mice expressing ChrimsonR. These light stimulation parameters were chosen based on previous works(68,94). Although we did not validate the release of ACh in the PL following the optogenetic stimulation, previous works have reported that ACh is released following optogenetic excitation of cholinergic terminals with light stimulation parameters comparable to ours (100,101).

Tissue collection, immunostaining, and image acquisition

Three to 5 days after the last session, mice were subcutaneously administered with avertin (20 mg/kg) and transcardially perfused with 0.9% saline followed by chilled 4% paraformaldehyde (PFA). The brains were immersed in 4% PFA for 48-72 hours, transferred to 30% sucrose/PBS, and stored at 4°C. A series of 50 µm coronal sections were collected from the area containing the mPFC (AP: +1.3 to +2.4 mm) and BF (AP: +0.6 to -0.6 mm) using a cryostat (CM3050S, Leica Biosystems). The tissues were washed in PBS three times and incubated with 10% donkey (# D9663, Sigma-Aldrich, RRID: AB_2810235) or goat serum (# ab156046, Abcam) for 2 hours at room temperature (RT) followed by a three-time wash in PBS.

To check the transgene expression, the BF sections were incubated with primary antibodies against choline acetyltransferase (ChAT; Goat, # AB144P, 1:200, Millipore, RRID: AB_90661) for 24-36 hours at 4 °C. Following three times wash in PBS, the brain sections were then incubated with a secondary antibody (Rhodamine-conjugated donkey-anti-goat, # 705-025-147, 1:200, Jackson ImmunoResearch, RRID: AB_2340389) for 2 hours at RT. Finally, the BF sections were mounted using an antifade mountant with DAPI (# P36935, ThermoFisher Scientific). Images for viral spread and optic fiber track identification were collected using a confocal microscope (Zeiss) under a 10× objective.

Behavioral analysis

X- and y-coordinates of a mouse’s position were extracted from the video tracking data. The timestamps were collected when a mouse entered/exited the threat and reward sites. A lap was counted as a “full” lap when a mouse exited a reward site, passed a threat site, and entered the other reward site. A lap was counted as a “half” lap when a mouse exited a reward site, entered a threat site, and returned to the same reward site. If a mouse exited a reward site and returned to the same reward site without entering a threat site, it was not counted as a half lap. Adaptive path choice was defined as the percentage of the laps in which a mouse chose the path associated with a lower threat probability among all full and half laps.

The speed approaching the threat site was calculated as the average speed in a 500-msec window immediately before the threat site entry during each lap. In each session, the median of the approaching speed was used to categorized laps into fast and slow laps (Fig 1G). The reaction to outcomes was quantified as the average speed in a 500-msec window immediately after the threat site entry.

Latency to threats was defined as the time difference between the reward site exit and the threat site entry. Latency to reward was defined as the time difference between the threat site exit and the reward site entry. The latency to initiate a subsequent lap was defined as the time in a reward site before initiating a new lap.

To examine if threat presence in a previous lap affected path choice in the subsequent lap, we used the data in the probabilistic stage, in which air puffs were delivered 75% of the time on one path (high-threat path) and 25% of the time on the other path (low-threat path). We first selected laps with air-puff delivery on each path in each session. If the number of these laps was less than two in a session, that session was discarded. We then selected mice with usable data in all five sessions. For each path, we calculated the proportion of laps in which a mouse chose the low-threat path in the subsequent lap.

To investigate how many laps mice used to estimate threat probabilities associated with the paths, we examined the relationship between a lap-by-lap path choice and the threat probability estimated based on the history of threats over different numbers of laps. In each mouse, laps were concatenated across all five sessions of the probabilistic stage. We then coded path choice in each lap as 1 for the high-threat path and 0 for the low-threat path (Fig S7C). We also calculated threat probability based on threat occurrence in the past 1-50 laps. The Pearson correlation coefficient was calculated between the path choice and the estimated threat probability. We then counted the mice that showed a significant correlation with threat probability calculated with any lap number.

PL imaging data analysis

Cell extraction

All imaging data were imported into the Inscopix Data Processing Software (IDPS) to spatially down-sample by a factor of 2 and temporally down-sample by a factor of 2. Pre-processed movies underwent spatial bandpass filer (low cut-off = 0.005 pixel-1, high cut-off = 0.500 pixel-1) and motion correction (mean frame) using IDPS. We used an established cell-detection algorithm, constrained nonnegative matrix factorization (CNMFe) in motion-corrected movies for cell detection. The applied CNMFe setting was minCorr = 0.8, minPeak2Noise = 10, Merging Threshold = 0.7. Each extracted region of interest (ROI) was manually checked to select the potential neuronal traces from the noise.

Outcome-evoked responses

In each cell, Ca2+ activity in 3-second windows starting from 1 second before the threat site entry was collected and binned into 100-millisecond bins. The binned activity was averaged across laps and converted into z-scores with the mean and standard deviation (SD) of activity during a 1-second window before the threat site entry (first ten bins). This normalization procedure was applied to each lap type separately. Specifically, to test the effect of LED stimulation on threat-evoked activity (Fig 2), laps were categorized into three types: air-puff omission, air-puff delivery, and air-puff delivery with LED stimulation. To examine how PL cells changed threat- and omission-evoked responses with learning (Fig 3), laps were categorized into four types: by which path a mouse took and whether it received air puff or not. To isolate learning-dependent changes in puff-evoked responses from the direct effect of LED stimulation, laps with air-puff delivery and LED stimulation were not included in this analysis. Lastly, to investigate how PL cells differentiated outcome-evoked responses depending on outcome expectation (Fig 4), laps were categorized into four types based on how fast a mouse approached the threat site and whether it received air puff or not.

The response magnitude was calculated by averaging z-normalized activity over a 500-msec window starting from the threat site entry. Cells were deemed to respond to threats or their omission (Figs 3Ab and Bb) if the response magnitude was greater than 2. Cells were deemed to differentiate outcome-evoked responses depending on outcome expectation (Fig 4C) if the difference in the response magnitude between surprising and expected outcomes was greater than 2 or smaller than -2.

Statistical analysis

The sample size was chosen based on previous studies using similar experimental approaches(102–104) . The data were presented as the group mean ± standard error of the mean unless otherwise specified. Statistical analyses were performed with GraphPad Prism, JASP (https://jasp-stats.org/), or MATLAB. To determine the statistical significance, we used one-way (repeated) analysis of variance (ANOVA), two-way mixed/repeated-measures ANOVA, unpaired t-test, chi-square test, Kolmogorov-Smirnov test, and binominal test. Degrees of freedom were corrected with Greenhouse-Geisser correction when sphericity was violated. For post hoc analyses, Tukey’s multiple comparison test was used following one-way repeated measures ANOVA. Significance was defined as *p < 0.05, **p < 0.01, ***p < 0.001.

Acknowledgements

This work was supported by NSERC Discovery Grant, CFI Leaders Opportunity Fund (Kaori Takehara-Nishiuchi), and NSERC graduate fellowship (Gaqi Tu).

Author Contributions

Conceptualization: Kaori Takehara-Nishiuchi, Gaqi Tu.

Data curation: Gaqi Tu, Peiying Wen, Adel Halawa.

Formal analysis: Kaori Takehara-Nishiuchi, Gaqi Tu.

Funding acquisition: Kaori Takehara-Nishiuchi, Gaqi Tu.

Investigation: Gaqi Tu, Kaori Takehara-Nishiuchi.

Methodology: Gaqi Tu, Peiying Wen, Adel Halawa.

Project administration: Gaqi Tu.

Supervision: Kaori Takehara-Nishiuchi.

Visualization: Gaqi Tu, Kaori Takehara-Nishiuchi.

Writing – original draft: Kaori Takehara-Nishiuchi.

Writing – review & editing: Gaqi Tu, Kaori Takehara-Nishiuchi.