Differential contribution of distinct neuronal populations to danger representations

9 The recognition of specific stimuli and contexts in dangerous situations determines the expression 10 of behaviors needed to appropriately cope with each threatening encounter. Moreover, the 11 detection of common features shared by different dangerous situations allows eliciting general 12 brain states and is necessary for both the expression of preparatory reactions and adaptive 13 behavioral responses in a timely manner. However, it is unknown how general and specific danger 14 representations emerge from the combined activity of different neuronal populations to elicit the 15 expression of adaptive defensive responses. Using a behavioral paradigm that exposes mice to 16 multiple threatening situations and calcium imaging recordings in freely moving mice, we 17 investigated the role of different dmPFC neuronal populations in the generation of general and 18 specific neuronal representations. Our results suggest that the population of somatostatin positive 19 (SST+) interneurons generates specific representations while those arising from parvalbumin 20 positive (PV+) interneurons are mainly unspecific. Together, this data suggests the presence of 21 distinct information in different dmPFC neurons allowing a collective encoding of both general and 22 specific danger representations. 23 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 2

24 The recognition of discrete features of stimuli and contexts informs about the specificities of 25 animal surroundings. The activity of neuronal populations devoted to the formation of specific 26 brain representations is essential to express optimal behavioral responses to properly cope with 27 a particular threatening situation1-4. Moreover, the ability to detect and classify individual features 28 into general categories is computationally advantageous and allows the generation of broader 29 brain representations 5-7 that are necessary for the expression of behaviors shared by distinct 30 situations3,4,8,9. This is especially relevant when animals face threatening situations, as both 31 preparatory reactions to danger and adaptive defensive responses to specific threats are 32 essential to cope with the incoming danger. The survival of an organism hinges significantly on 33 its capacity to recognize potential threats and to select suitable defensive responses to properly 34 cope with danger. The expression of adapted defensive behaviors is key to help individuals avoid 35 predation, escape from threats and protect themselves and their offspring. However, dysfunctions 36 in the response to threats are implicated in various neurological and psychiatric conditions such 37 as anxiety disorders, phobias, or post-traumatic stress disorder10. Thus, understanding the neural 38 mechanisms underlying the expression of defensive behaviors in dangerous situations is 39 essential to unravel the neuronal bases of these severe conditions11-16 40 In order to select the most advantageous defensive response, animals must be able to detect 41 threatening situations and to discriminate the type of threat that is confronted. To achieve this, 42 multiple interconnected brain regions shape survival neuronal circuits17,18 to identify the presence 43 of a dangerous situation and accurately recognize the specificity of the threat encountered. 44 Traditional fear-conditioning and active avoidance paradigms have been widely used to study the 45 brain mechanisms underlying emotional processing of threat-related information 11,17,19-25. 46 However, these paradigms entail a reductionist design in which a single threat is presented and 47 a single defensive response is examined. This has long been a limitation when trying to 48 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 3 disentangle whether neuronal populations recorded in specific brain regions encode information 49 specific to each threatening situation and/or more general information relative to the presence of 50 a dangerous stimulus and the defensive state elicited. Nevertheless, a new generation of 51 behavioral paradigms has been developed in which mice are not exposed to a single threat but 52 also to multiple threats1,26 and/or stimuli of positive valence1,27-33. Similarly, we recently developed 53 a behavioral paradigm, the Multi input-output (MIO) paradigm, in which mice are presented with 54 multiple threats and elicit different defensive behaviors 3. Using this task while performing 55 electrophysiological recordings in the dorsal division of the medial prefrontal cortex (dmPFC), 56 which has largely been evidenced as a key regulator of defensive behavior13,14,23,27,34, we showed 57 that the dmPFC simultaneously encodes a general danger representation but also specific 58 information about the identity of each threatening situation3. 59 Although the PFC contains distinct classes of neurons with different functional properties35, most 60 of the literature on the study of defensive behaviors has either employed mixed populations or 61 focused on the role of the most abundant neurons in the PFC, the excitatory pyramidal 62 neurons3,34,36,37. Nonetheless, the activity of pyramidal neurons (Pyr) is extensively orchestrated 63 by a network of GABAergic interneurons (INs) with diverse morphology, firing properties, 64 postsynaptic targets and protein expression patterns 35,38. Indeed, it is known that dmPFC 65 processing of emotionally relevant information to control the expression of defensive behaviors 66 does not solely depend on Pyr neuronal activity, but on the interaction between Pyr neurons and 67 a heterogeneous set of interneuron populations 23,39-42. Parvalbumin-expressing (PV +) 68 interneurons, the most abundant of these interneuron types, mostly convey feedforward inhibition 69 onto the soma and axon initial segment of Pyr neurons 35,38. Somatostatin-expressing (SST +) 70 interneurons are less abundant and mainly deliver feedback inhibition targeting the apical 71 dendrites of excitatory Pyr neurons35,38. There is a growing body of evidence showing that, within 72 the dmPFC, these two IN populations play a critical role in the expression of defensive 73 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 4 behaviors23,40-42. However, it is still unclear how dmPFC Pyr neurons, SST+ and PV+ interneurons 74 cooperate to generate general and specific representations to control the expression of defensive 75 behaviors. To investigate the specific role of dmPFC neuronal populations in the encoding of 76 general and specific representations of danger, we used calcium imaging to record the activity of 77 dmPFC Pyr neurons, SST+ and PV+ interneurons populations in mice facing different threatening 78 situations in the MIO paradigm. Our results suggest that although all the studied neuronal 79 populations are modulated by the presence of threatening situations , each population of 80 interneurons contributes to general and specific encoding of information differently. The activity 81 of the population of SST+ interneurons allowed us to specifically discriminate between threatening 82 or non-threatening trials as well as between the sensory inputs setting each threatening situation 83 apart, while the population of PV+ interneurons encoded, in a more unspecific manner, the general 84 presence of stimuli independently of their emotional value, the tone or the context. 85 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 5

86 To investigate the encoding of general and specific danger information within different dmPFC 87 neuronal populations we used the multi input-output (MIO) paradigm 3, a behavioral task that 88 allowed us to study the neuronal activity of freely moving mice facing different threatening 89 situations within the same behavioral session (Fig. 1). To monitor the activity of distinct neuronal 90 populations within the dmPFC of mice performing the MIO paradigm we expressed the calcium 91 indicator GCaMP6f in either CaMKII+, SST+ or PV+ neurons. We implanted gradient-index (GRIN) 92 lenses in the dmPFC and used miniscopes (Inscopix) to image the calcium levels of the different 93 neuronal types (Fig. 1b). To monitor the calcium levels of CaMKII-expressing pyramidal neurons 94 we employed an AAV5-CaMKII-GCaMP6f viral construct in C57/BL6 mice. To image the calcium 95 levels of SST-expressing and PV-expressing interneurons we used a C re-dependent GCaMP6f 96 viral construct (AAV5-Syn-Flex-GCaMP6f) in mice lines expressing Cre in SST + or PV+ neurons 97 respectively. These mice were then trained in the MIO paradigm. In this paradigm mice are freely 98 exploring a rectangular behavioral box divided in 2 parts, the shelter and the arena ( Fig. 1a). 99 During active trials a conditioned stimulus (a constant tone of 7 kHz for a maximum of 7 seconds, 100 the CS+1) is presented when mice are in the arena and it is associated with a mild foot shock (the 101 unconditioned stimulus; US) that can be avoided by shuttling to the shelter. In contrast, during the 102 passive trials the CS+1 is presented when mice are in the shelter and mice must stay in the shelter 103 for the whole duration of the CS (7 seconds), otherwise they receive a foot shock if shuttling to 104 the arena. During the inverse trials mice have to shuttle to the arena to avoid the US in response 105 to a different tone (a constant tone of 2 kHz for a maximum of 7 seconds, the CS +2). Altogether, 106 mice are presented with 3 different threatening trials requiring the integration of both tone and 107 context to elicit the correct defensive responses in each threatening situation (Fig. 1c). Mice are 108 also presented with control trials (a constant white-noise tone lasting 7 seconds, the CS−) that are 109 never associated with a foot shock independently of mice location or behavior. Overall, mice 110 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 6 learned to perform the MIO paradigm, eliciting the correct defensive responses in each 111 threatening situation ( Fig. 1c). The dmPFC neuronal activity of mice was imaged during the 112 different training sessions and the day of best performance for each mouse was used to evaluate 113 the activity of the different neuronal types during each trial type ( Fig. 1d). As observed by the 114 representative field of views (FOVs) of Fig. 1d, imaging pyramidal neurons gave a higher yield 115 (~140 neurons per FOV) than imaging SST+ and PV+ interneurons (~35 and 25 neurons per FOV 116 respectively). The higher number of pyramidal neurons within the imaged FOVs is consistent with 117 the higher proportions of this neuronal type with in the PFC 35. Using this method we could then 118 image the activity of pyramidal, SST + and PV + neurons ( Fig 1d ). We observed that: (1) All 119 neuronal populations studied contained neurons responsive to CS + and control trials, (2) single 120 neuronal responses presented strong variability and that (3) response profiles varied greatly 121 between neurons. These features are displayed in the 3 representative neurons shown in Fig. 122 1d. Together this data shows that the use of the recently developed MIO paradigm coupled with 123 in vivo calcium imaging allows us to investigate the activity of different dmPFC neuronal types 124 during multiple threatening situations and, therefore, the encoding of information related to either 125 general or specific danger in pyramidal neurons, SST+ interneurons and PV+ interneurons. 126 To investigate responses of the different neuronal types during the different trials we looked at 127 the activity of all the recorded neurons. We imaged 1253 pyramidal neurons from 9 mice, 211 PV+ 128 neurons from 8 mice and 208 SST+ neurons from 6 mice (Fig. 2). We used this data to investigate 129 the modulation of the different neuronal populations during the CS onset and the CS offset of 130 active, passive, inverse and control trials (Fig. 2a). Aligning the neuronal activity to the CS onset 131 allowed us to investigate the engagement of each neuronal populations at the onset of the 132 threatening and control trials, while aligning their activity at the CS offset allowed us to investigate 133 the neuronal modulation related to the expression of the different defensive behaviors ( Fig. 2). 134 Pyramidal neurons showed a clear increase in activated and inhibited neurons at the CS onset 135 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 7 during threatening trials and also during control trials (although with briefer responses in control 136 than in threatening trials; Fig. 2b). We also assessed the modulation of pyramidal neurons before 137 and after the CS offset of the 3 threatening trials (active, passive and inverse) but not at the offset 138 of the control trials (Fig. 2b), consistent with the lack of defensive response elicited during these 139 trials. Then, we investigated the neuronal activity of PV + interneurons during the different trial 140 types. Similarly to pyramidal neurons, PV+ interneurons were activated and inhibited at the onset 141 of the CS during all trial types. However, the responses of PV + neurons seemed more sustained 142 and appeared before and after the CS offset for all trial types (including control trials; Fig. 2c). 143 Finally, SST + interneurons showed response profiles strikingly different to those observed in 144 pyramidal and PV+ interneurons (Fig. 2d). The population of SST+ interneurons contained a higher 145 proportion of activated neurons when compared to the other investigated neuronal types. SST + 146 neurons showed stronger and more sustained modulations both at the CS onset and before and 147 after the CS offset (Fig. 2d and e). Together, this data shows that all studied neuronal types are 148 modulated (both through activation and inhibition) during threatening trials, suggesting that 149 pyramidal neurons, SST + and PV + interneurons participate in the processing of threatening 150 information in the dmPFC. Interestingly, CS+ trials strongly affected the activity of SST + 151 interneurons, suggesting a greater engagement of this neuronal population in the processing of 152 information related to threatening trials than the other investigated neuronal types. Whereas this 153 data shows that all the studied neuronal populations are modulated during threatening trials, it 154 does not address the question of whether their activity is devoted to the encoding of information 155 related to either general or specific danger. 156 In order to address this, we used a population decoding approach based on linear kernel SVM 157 classifiers applied to the neuronal activity of the different populations of neurons (See Methods). 158 This strategy is illustrated in Fig. 3a, in which the calcium levels from the ensemble of neurons of 159 each type of the recorded populations at a given time point t (the so-called population vector) is 160 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 8 compared between conditions. In the scheme, active and control trials are represented; we 161 repeated the same strategy between the following conditions: active, passive, inverse, control 162 and no-CS trials (trials lacking any tone presentation). To obtain results comparable between the 163 different neuronal populations we made classifiers using 100 randomly selected neurons of each 164 type ( Fig. 3a ). Using this approach we were able to significantly decode with high decoding 165 accuracies all threatening trials (active, passive and inverse trials) from the no-CS trials using 166 either the population of pyramidal neurons or the populations of SST + or PV+ interneurons (Fig. 167 3b). Moreover, we addressed whether using this approach we could decode the threatening trials 168 from the control trials. Interestingly, we observed high and significant decoding accuracies at CS 169 onset when the populations of pyramidal neurons and SST + interneurons were challenged to 170 classify threatening from control trials. However, the population of PV + interneurons presented 171 only low decoding accuracies when decoding control from threatening trials ( Fig. 3b ). To 172 investigate the number of neurons needed by the different neuronal populations to decode the 173 threatening trials from either basal conditions (no-CS) or control trials, we constructed classifiers 174 with a variable number of randomly selected neurons of each type (from 1 to 200 neurons of each 175 type) and calculated the mean decoding accuracies obtained at CS onset ( Fig. 3c). Using this 176 technique we observed that SST + neurons provided higher decoding accuracies even when the 177 classifiers used lower number of neurons. Moreover, while the population of pyramidal neurons 178 and SST + interneurons obtained similar decoding accuracies when decoding threatening trials 179 from either control trials or trials with no tone presentation (no-CS trials ; Fig. 3c), the population 180 of PV + interneurons showed a clear decrease in the decoding accuracies when decoding 181 threatening from control trials as compared to threatening vs no-CS trials (Fig 3c). This indicated 182 that, although PV + interneurons could decode the presence of a tone from baseline, their 183 population activity did not allow us to decode whether the tone presented was threatening or not 184 (CS+ or control tone; unlike SST+ and pyramidal neurons. Fig. 3b and c). This suggests that the 185 population of PV + interneurons carried more unspecific information than SST + and pyramidal 186 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 9 neurons. To further confirm this observation, we tested the generalization of the classifiers among 187 threatening and control trials (see Methods). To investigate the threat generalization, we trained 188 a classifier to decode a random threat type (active, passive or inverse) from basal activity (no-CS 189 trials) and challenged that classifier to classify a different threat type (not used to train the 190 classifier) from basal activity. Using this approach, we observed that the classifiers made with all 191 the studied neuronal populations showed high levels of generalization across the threatening trials 192 (Threat-generalization; Fig. 3d). Then we investigated whether these neuronal populations could 193 decode control trials from no-CS trials and observed a transient decoding accuracy (around 500 194 ms) in the pyramidal and PV + neuronal populations and a sustained encoding in the SST + 195 interneuronal population (Fig. 3d). Next, we evaluated the generalization of the patterns of activity 196 presented in the threatening trials to the control trials. For this, we used the same classifier that 197 we trained for the threat generalization to classify control trials from basal activity. We observed 198 that, although SST + interneurons showed the stronge st encoding of control trials, they lacked 199 strong levels of generalization from threatening to control trials ( Fig. 3d). The SST + and the 200 pyramidal neuronal populations only presented a transient control-generalization while PV + 201 neuronal population showed levels of control-generalization similar to those obtained when 202 decoding control trials ( Fig. 3d ). We then measured the accuracy loss as the difference in 203 accuracy obtained in each iteration between the ’threat’ classification accuracy and the ‘threat -204 general’ ( CS+-CS+) and ‘control -general’ ( CS+-control) classification accuracies. Indeed, there 205 was a stronger accuracy loss in control-generalization than in threat-generalization for pyramidal 206 and SST + neuronal populations, while this difference was not present in the PV + neuronal 207 population (Fig. 3e). These results show that the patterns of activity presented in the population 208 of PV + interneurons at the early CS onset (500 ms) are shared among the different trials , 209 independently of whether those are threatening or not. This data, together with the low decoding 210 accuracies obtained between threatening and control trials ( Fig. 3b and c) suggests that the 211 population of PV+ interneurons encodes the presence of trials (or auditory stimuli) in an unspecific 212 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 10 manner. However, the population of SST + interneurons showed stronger decoding accuracies 213 than the other studied neuronal populations when decoding threatening from no-CS trials ( Fig. 214 3b and c), threatening from control trials (Fig. 3 b and c) and control from no-CS trials (Fig. 3d), 215 with low generalization between threatening and control trials (Fig. 3d and e). This points to the 216 encoding of information by SST+ populations being more robust than the encoding present in PV+ 217 and pyramidal neurons in both threatening and non-threatening trials and that this is achieved 218 with different patterns of activity dependent on the emotional value of the stimuli . We then 219 investigated whether the activity of the different neuronal populations encoded information 220 necessary for the correct identification of the threatening trials and the selection of defensive 221 responses. To do so, we took advantage of the design of the behavioral paradigm, in which mice 222 need to discriminate both tone and context to select the correct defensive response in each 223 threatening situation (Fig. 1c). Then we evaluated whether the dmPFC population activity could 224 decode threatening trials that presented different tones (CS +1: active and passive trials; CS +2: 225 inverse trials) and threatening trials in different contexts (Shelter: inverse and passive trials; 226 Arena: active trials) and observed that in both conditions the population activity of SST + 227 interneurons provided stronger decoding accuracies than the populations of pyramidal and PV + 228 neurons (Fig 3f and g). This suggests that the population of SST+ interneurons provides not only 229 a stronger encoding of threatening and non-threatening trials but also more specific information 230 about the different threatening trials presented. Together, this data suggests that, although all the 231 studied neuronal populations are modulated during the threatening trials, their activity encodes 232 different information. While the pyramidal population activity encodes the presence of threatening 233 trials specifically, the population of PV + interneurons encodes the presence of stimuli 234 independently of their emotional value, and the population of SST+ interneurons not only informs 235 about the threatening or non-threatening nature of the trials, but also provides specific information 236 about the different factors differentiating the threatening trials (tone and context). 237 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 11

238 Coupling the recently developed MIO behavioral paradigm with calcium imaging in freely moving 239 mice we were able to investigate the generalization and specificity of neuronal representations in 240 different neuronal populations. Our results suggest that different neuronal types represent 241 different information allowing the dmPFC to encode both general and specific danger. 242 The MIO paradigm was previously developed to investigate the neuronal activity of mice facing 243 different threatening situations during the same behavioral session 3. This task offers an 244 advantage from classical behavioral paradigms, in which mice are presented with a single 245 threatening event, since it permits us to evaluate whether the neuronal responses evoked by a 246 threat are either specific or general for all dangerous situations . We coupled this paradigm with 247 the imaging of calcium activity of different dmPFC neuronal types in freely-moving mice using 248 miniscopes ( Fig. 1 ). We imaged the activity of Pyr neurons, SST + and PV + interneurons, 249 generating the first dataset in which large populations of dmPFC SST+ and PV+ interneurons are 250 imaged in mice performing defensive behaviors. Moreover, coupling this paradigm with calcium 251 imaging, this dataset allowed us to further investigate the specificity/generality of the CS+-evoked 252 neuronal responses in distinct neuronal populations. 253 To investigate this, we first evaluated the responsiveness of each neuronal type during threat 254 presentation and observed that all of them were modulated by the presence of threatening trials 255 (Fig. 2). This suggests that all the studied neuronal populations are involved in the processing of 256 danger-related information in the dmPFC. However, different neuronal populations showed 257 different profiles of responsiveness. The most striking difference attains the SST+ interneurons: a 258 higher percentage of this neuronal population was modulated from the onset to the offset of both 259 threatening and control trials ( Fig. 2). This suggests that SST + interneurons display a higher 260 degree of engagement in the processing of danger than either PV + interneurons or pyramidal 261 neurons. This could be due to their role in the dmPFC micro-circuitry 35 or to their higher sparsity 262 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 12 in the dmPFC. To further clarify the nature of this higher recruitment of SST+ interneurons further 263 studies investigating the functional interactions of the different neuronal types in the dmPFC 264 micro-circuitry during threat presentations are needed. However, despite the higher modulation 265 of SST+ interneurons during threatening trials, all neuronal types showed evoked responses to 266 threat during all trial types. 267 To investigate whether the activation of the different neuronal types was either specific for each 268 threatening situation or more general, we looked at the population patterns of activity elicited by 269 the different trials in each neuronal population. Although the presence of threatening trials could 270 be decoded by the population activity of all neuronal types, the activity of PV + interneurons 271 provided unspecific information that did not allow us to differentiate between threatening and non-272 threatening (control) trials ( Fig. 3). In strike contrast, the population of SST + interneurons 273 displayed patterns of activity that, not only allowed us to discriminate between threatening and 274 non-threatening trials, but also between threatening trials presenting different tones (CS +1 or 275 CS+2) or contexts (shelter or arena; Fig. 3). Pyr neurons displayed patterns of activity that allowed 276 us to discriminate between threatening and non-threatening situations but that did not allow us to 277 classify threatening trials presenting different features. Although this result is consistent with the 278 primacy of general representations of danger observed in the mixed population of dmPFC3 , it is 279 important to note that the presented results are obtained sampling an equal number of neurons 280 for each population (100 neurons). An alternative approach would be to evaluate the classification 281 accuracies adjusting the numbers of neurons used for each population in a manner that 282 resembled the proportions observed in the dmPFC (approximately 80% Pyr neurons, and 20% 283 interneurons). Therefore, the classification accuracies obtained with a larger population of Pyr 284 neurons may serve as a better comparison for those obtained with 100 SST + and PV + 285 interneurons. Nonetheless, our results show a differential encoding of general and specific 286 representations of danger in the dmPFC interneuron populations. SST+ interneurons show more 287 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 13 specific representations and their population activity allowed us to classify not only the threatening 288 nature of the stimuli, but also the sensory features distinguishing the threatening trials (tone and 289 context). These results suggest that these neurons may provide specific inhibition necessary for 290 the appearance of specific threat-representations in the dmPFC. Moreover, this data further 291 suggests that SST+ activity is necessary for the discrimination of threatening tones and contexts, 292 however, further studies demonstrating the causality of the involvement of SST + interneurons in 293 the selection of defensive responses are needed. The PV+ population activity was largely 294 unspecific and did not allow us to discriminate between control and threatening trials. Further 295 studies are needed to disambiguate whether their activity in threatening situations is passively 296 informing about the presence of an auditory stimuli or whether it promotes the engagement of 297 mice in the behavioral task by favoring processes such as arousal or attention. 298 Overall, our study shows a distinct role of different interneuron populations in the processing of 299 threat-related information within the prefrontal cortex. Maladaptive fear processing may occur 300 both through the generalization of danger-representations to non-threatening situations or 301 through the impairment of the threat-specific representation, which could affect the ability to 302 adaptively cope with threats. A better understanding of the processes that allow the dmPFC to 303 generate both general-danger and specific-threat representations might shed light in the 304 mechanisms responsible of the development of those pathological conditions. 305 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 14 Author contribution 306 A.P.M., G.L.-F and M.M-F. performed calcium imaging recordings. A.P.M., and G.L.-F. performed 307 histology. A.P.M., G.L.-F and M.M-F performed surgeries. M.M.-F. and C.H. designed 308 experiments. M.M.-F. designed population analyses and wrote software codes. A.P.M., G.L.-F, 309 C.H., and M.M-F. analyzed data and wrote the paper. All the authors read and edited the 310 manuscript. 311 Acknowledgments 312 We thank S. Laumond, J. Tessaire and the technical staff of the housing and experimental animal 313 facility of the Neurocentre Magendie. This work was supported by EMBO (ALTF 200-2018) and 314 HFSP postdoctoral fellowships (LT0000658/2019) to M.M.-F., ECO-Contrat doctoral de la 315 Fondation pour la Recherche Médicale (ECO202206015556) to G.L.-F., grants from the French 316 National Research Agency (ANR- 10-EQPX-08 OPTOPATH), the Conseil Regional d’Aquitaine 317 and the Fondation pour la Recherche Médicale (FRM-Equipes FRM 2017). The funders had no 318 role in study design, data collection and analysis, decision to publish or preparation of the 319 manuscript. 320 Competing interests 321 The authors declare no competing interests 322 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 15

323 1 Kyriazi, P., Headley, D. B. & Pare, D. Different Multidimensional Representations across the 324 Amygdalo-Prefrontal Network during an Approach-Avoidance Task. Neuron 107, 717-730 e715 325 (2020). https://doi.org:10.1016/j.neuron.2020.05.039 326 2 Kyriazi, P., Headley, D. B. & Pare, D. Multi-dimensional Coding by Basolateral Amygdala Neurons. 327 Neuron 99, 1315-1328 e1315 (2018). https://doi.org:10.1016/j.neuron.2018.07.036 328 3 Mart in-Fernandez, M. et al. Prefrontal circuits encode both general danger and specific threat 329 representations. Nat Neurosci 26, 2147-2157 (2023). https://doi.org:10.1038/s41593-023-01472-330 8 331 4 Ceko, M., Kragel, P. A., Woo, C. W., Lopez-Sola, M. & Wager, T. D. Common and stimulus-type-332 specific brain representations of negative affect. Nat Neurosci 25, 760-770 (2022). 333 https://doi.org:10.1038/s41593-022-01082-w 334 5 Seger, C. A. & Miller, E. K. Category learning in the brain. Annu Rev Neurosci 33, 203-219 (2010). 335 https://doi.org:10.1146/annurev.neuro.051508.135546 336 6 Roy, J. E., Riesenhuber, M., Poggio, T. & Miller, E. K. Prefrontal cortex activity during flexible 337 categorization. J Neurosci 30, 8519-8528 (2010). https://doi.org:10.1523/JNEUROSCI.4837-338 09.2010 339 7 Cromer, J. A., Roy, J. E. & Miller, E. K. Representation of multiple, independent categories in the 340 primate prefrontal cortex. Neuron 66, 796-807 (2010). 341 https://doi.org:10.1016/j.neuron.2010.05.005 342 8 Reinert, S., Hubener, M., Bonhoeffer, T. & Goltstein, P. M. Mouse prefrontal cortex represents 343 learned rules for categorization. Nature 593, 411-417 (2021). https://doi.org:10.1038/s41586-344 021-03452-z 345 9 Grundemann, J. et al. Amygdala ensembles encode behavioral states. Science 364 (2019). 346 https://doi.org:10.1126/science.aav8736 347 10 Shin, L. M. & Liberzon, I. The neurocircuitry of fear, stress, and anxiety disorders. 348 Neuropsychopharmacology 35, 169-191 (2010). https://doi.org:10.1038/npp.2009.83 349 11 Herry, C. et al. Switching on and off fear by distinct neuronal circuits. Nature 454, 600-606 (2008). 350 https://doi.org:10.1038/nature07166 351 12 Sierra-Mercado, D., Padilla-Coreano, N. & Quirk, G. J. Dissociable roles of prelimbic and infralimbic 352 cortices, ventral hippocampus, and basolateral amygdala in the expression and extinction of 353 conditioned fear. Neuropsychopharmacology 36, 529-538 (2011). 354 https://doi.org:10.1038/npp.2010.184 355 13 Diehl, M. M., Bravo-Rivera, C. & Quirk, G. J. The study of active avoidance: A platform for 356 discussion. Neurosci Biobehav Rev 107, 229-237 (2019). 357 https://doi.org:10.1016/j.neubiorev.2019.09.010 358 14 Karalis, N. et al. 4-Hz oscillations synchronize prefrontal-amygdala circuits during fear behavior. 359 Nat Neurosci 19, 605-612 (2016). https://doi.org:10.1038/nn.4251 360 15 Wolff, S. B. et al. Amygdala interneuron subtypes control fear learning through disinhibition. 361 Nature 509, 453-458 (2014). https://doi.org:10.1038/nature13258 362 16 Moscarello, J. M. & LeDoux, J. E. Active avoidance learning requires prefrontal suppression of 363 amygdala-mediated defensive reactions. J Neurosci 33, 3815-3823 (2013). 364 https://doi.org:10.1523/JNEUROSCI.2596-12.2013 365 17 LeDoux, J. E. Emotion circuits in the brain. Annu Rev Neurosci 23, 155-184 (2000). 366 https://doi.org:10.1146/annurev.neuro.23.1.155 367 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 16 18 Malezieux, M., Klein, A. S. & Gogolla, N. Neural Circuits for Emotion. Annu Rev Neurosci 46, 211-368 231 (2023). https://doi.org:10.1146/annurev-neuro-111020-103314 369 19 Kim, J. J. & Fanselow, M. S. Modality-specific retrograde amnesia of fear. Science 256, 675-677 370 (1992). https://doi.org:10.1126/science.1585183 371 20 Haubensak, W. et al. Genetic dissection of an amygdala microcircuit that gates conditioned fear. 372 Nature 468, 270-276 (2010). https://doi.org:10.1038/nature09553 373 21 Milad, M. R. et al. Neurobiological basis of failure to recall extinction memory in posttraumatic 374 stress disorder. Biol Psychiatry 66, 1075-1082 (2009). 375 https://doi.org:10.1016/j.biopsych.2009.06.026 376 22 Likhtik, E., Popa, D., Apergis-Schoute, J., Fidacaro, G. A. & Pare, D. Amygdala intercalated neurons 377 are required for expression of fear extinction. Nature 454, 642-645 (2008). 378 https://doi.org:10.1038/nature07167 379 23 Courtin, J. et al. Prefrontal parvalbumin interneurons shape neuronal activity to drive fear 380 expression. Nature 505, 92-96 (2014). https://doi.org:10.1038/nature12755 381 24 Likhtik, E., Stujenske, J. M., Topiwala, M. A., Harris, A. Z. & Gordon, J. A. Prefrontal entrainment of 382 amygdala activity signals safety in learned fear and innate anxiety. Nat Neurosci 17, 106-113 383 (2014). https://doi.org:10.1038/nn.3582 384 25 Hagihara, K. M. et al. Intercalated amygdala clusters orchestrate a switch in fear state. Nature 385 594, 403-407 (2021). https://doi.org:10.1038/s41586-021-03593-1 386 26 Kennedy, A. et al. Stimulus-specific hypothalamic encoding of a persistent defensive state. Nature 387 586, 730-734 (2020). https://doi.org:10.1038/s41586-020-2728-4 388 27 Bravo-Rivera, C., Roman-Ortiz, C., Brignoni-Perez, E., Sotres-Bayon, F. & Quirk, G. J. Neural 389 structures mediating expression and extinction of platform-mediated avoidance. J Neurosci 34, 390 9736-9742 (2014). https://doi.org:10.1523/JNEUROSCI.0191-14.2014 391 28 Choi, J. S. & Kim, J. J. Amygdala regulates risk of predation in rats foraging in a dynamic fear 392 environment. Proc Natl Acad Sci U S A 107, 21773-21777 (2010). 393 https://doi.org:10.1073/pnas.1010079108 394 29 Namburi, P. et al. A circuit mechanism for differentiating positive and negative associations. 395 Nature 520, 675-678 (2015). https://doi.org:10.1038/nature14366 396 30 Zhang, X. et al. Genetically identified amygdala-striatal circuits for valence-specific behaviors. Nat 397 Neurosci 24, 1586-1600 (2021). https://doi.org:10.1038/s41593-021-00927-0 398 31 Stephenson-Jones, M. et al. Opposing Contributions of GABAergic and Glutamatergic Ventral 399 Pallidal Neurons to Motivational Behaviors. Neuron 105, 921-933 e925 (2020). 400 https://doi.org:10.1016/j.neuron.2019.12.006 401 32 Zhang, X. & Li, B. Population coding of valence in the basolateral amygdala. Nat Commun 9, 5195 402 (2018). https://doi.org:10.1038/s41467-018-07679-9 403 33 Burgos-Robles, A. et al. Amygdala inputs to prefrontal cortex guide behavior amid conflicting cues 404 of reward and punishment. Nat Neurosci 20, 824-835 (2017). https://doi.org:10.1038/nn.4553 405 34 Jercog, D. et al. Dynamical prefrontal population coding during defensive behaviours. Nature 595, 406 690-694 (2021). https://doi.org:10.1038/s41586-021-03726-6 407 35 Anastasiades, P. G. & Carter, A. G. Circuit organization of the rodent medial prefrontal cortex. 408 Trends Neurosci 44, 550-563 (2021). https://doi.org:10.1016/j.tins.2021.03.006 409 36 Bravo-Rivera, C., Roman-Ortiz, C., Montesinos-Cartagena, M. & Quirk, G. J. Persistent active 410 avoidance correlates with activity in prelimbic cortex and ventral striatum. Front Behav Neurosci 411 9, 184 (2015). https://doi.org:10.3389/fnbeh.2015.00184 412 37 Diehl, M. M. et al. Active avoidance requires inhibitory signaling in the rodent prelimbic prefrontal 413 cortex. Elife 7 (2018). https://doi.org:10.7554/eLife.34657 414 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 17 38 Kupferschmidt, D. A. et al. Prefrontal Interneurons: Populations, Pathways, and Plasticity 415 Supporting Typical and Disordered Cognition in Rodent Models. J Neurosci 42, 8468-8476 (2022). 416 https://doi.org:10.1523/JNEUROSCI.1136-22.2022 417 39 Lucas, E. K. & Clem, R. L. GABAergic interneurons: The orchestra or the conductor in fear learning 418 and memory? Brain Res Bull 141, 13-19 (2018). 419 https://doi.org:10.1016/j.brainresbull.2017.11.016 420 40 Cummings, K. A., Bayshtok, S., Dong, T. N., Kenny, P. J. & Clem, R. L. Control of fear by discrete 421 prefrontal GABAergic populations encoding valence-specific information. Neuron 110, 3036-3052 422 e3035 (2022). https://doi.org:10.1016/j.neuron.2022.07.004 423 41 Cummings, K. A. & Clem, R. L. Prefrontal somatostatin interneurons encode fear memory. Nat 424 Neurosci 23, 61-74 (2020). https://doi.org:10.1038/s41593-019-0552-7 425 42 Xu, H. et al. A Disinhibitory Microcircuit Mediates Conditioned Social Fear in the Prefrontal Cortex. 426 Neuron 102, 668-682 e665 (2019). https://doi.org:10.1016/j.neuron.2019.02.026 427 428 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 18

429 Subject detail 430 Male C57BL6/J (8-24 weeks old, Janvier), SST-IRES-Cre (8-24 weeks old, Jackson Laboratory, 431 Ssttm2.1(cre)Zjh/J) and PV-IRES-Cre (8-24 weeks old, Jackson Laboratory, B6;129P2-432 Pvalbtm1(cre)Arbr/J) mice were single-housed for at least three weeks before experimentation, 433 under a 12-h light-dark cycle, and provided with food and water ad libitum . The housing 434 temperature was 22 ± 1ºC, and housing humidity was 60% ± 5%. All procedures were performed 435 per standard ethical guidelines (European Communities Directive 86/60-EEC) and were approved 436 by the Animal Health and Care committee of Institut National de la Santé et de la Recherche 437 Médicale and French Ministry of Agriculture and Forestry (authorization #A3312001). Mice were 438 handled and habituated to be connected for three days before the experiment started. All data 439 correspond to implanted and connected animals. 440 Behavior 441 Experiments were run in a behavioral box (40 x 10 x 30 cm) composed of Plexiglas walls and a 442 grid floor used to deliver footshocks (0.6 mA) located in a sound-attenuating chamber. This box 443 was divided in the shelter zone (5 x 10 cm) and the arena zone (35 x 10 cm) by a small plastic 444 hurdle (0.5 cm high). The shelter zone was marked with visual cues to facilitate its recognition. 445 The box was equipped with infrared beams to detect mice shuttling between the shelter and arena 446 zones. The behavioral box was enclosed in an acoustic isolated box containing speakers, a light 447 source and a video camera (30 frames per second; Cineplex, Plexon) above the behavioral box. 448 Three different auditory cues were delivered at a volume of 75 dB: CS +1 (7 kHz), CS +2 (2 kHz) 449 and CS− (white noise). The auditory cues were presented for 7 s or until mice shuttled from one 450 zone to the other (shelter and arena) in the case of the CS +. The CS − was presented for 7 s 451 independently of mouse behavior. 452 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 19 The presence of the auditory cue and the location of the mice in either the shelter or the arena 453 established the different trial types. In active trials, a CS+1 was presented when mice were in the 454 arena. Mice had to shuttle to the shelter before the end of the CS (7 s) to avoid a foot shock (1 455 s). In passive trials, the CS+1 was presented when mice were in the shelter. Mice had to remain 456 in the shelter for the duration of the CS (7 s) to avoid a foot shock (1 s). In inverse trials, a CS +2 457 tone was presented when mice were in the shelter. Mice had to shuttle to the arena before the 458 end of the CS (7 s) to avoid a foot shock (1 s). In control trials, the CS − was presented either in 459 the shelter or in the arena without any reinforcement and independently of mouse behavior. The 460 foot shocks lasted for a maximum of 1 s (shuttling to the correct area of the box would end the 461 foot shock) and co-terminated with the CS when mice did not perform the correct defensive 462 response for each CS + trial type (error trials). Therefore, at the end of error trials, there was an 463 extra second of CS+–US co-presented. The different trial types were intermingled and dependent 464 on mouse location and tone presentations. In a behavioral training session, mice were exposed 465 to 64 intermingled and pseudorandomized ‘tone-presentation events’ of three different types: 24 466 type A (CS+1 presentations: active or passive trials according to mouse location), 24 type B (CS+1 467 presentation if mice were in the arena and CS+2 if mice were in the shelter: active or inverse trials, 468 respectively) and 16 control (CS− presentation independent of mouse location). Additionally, mice 469 were exposed to 8 intermingled control trials with a duration of 7 seconds in which no tone was 470 presented (no-CS trials). The ITI was pseudorandomized, ranging from 32 to 75 s. A behavioral 471 session therefore lasted approximately 75 min on average. The training sessions consisted of a 472 habituation session (identical to a normal training session but lacking shock presentations) and a 473 maximum of nine training sessions until mice reached high behavioral performance levels for the 474 three threatening trial types (active, passive and inverse trials). We calculated the performance 475 of active, passive and inverse trials as the number of correct trials over the total number of trials 476 for each trial type. The experimental box was cleaned with 70% ethanol before and after each 477 session. 478 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 20 Surgical procedure 479 Mice (8-10 weeks old) were anesthetized with isoflurane (induction 3%, maintenance 1.75%) in 480 O2. Body temperature was maintained at 37 ºC with a Temperature Controller System (FHC), and 481 eyes were hydrated with Ocry-gel (TVM). For analgesia, a subcutaneous injection of 0.05 mL of 482 Metacam (5 mg/kg body weight) was administered 30 min before anesthesia. Additionally, 0.1 mL 483 of local lidocaine anesthesia (Lidor, 20 mg/mL diluted with sodium chloride at 0.5%) was applied 484 under the scalp before incision. Mice were placed in a stereotaxic frame (Kopf instruments), and 485 three stainless steel screws were attached to the skull before craniotomy to secure all implants. 486 Virus injection and GRIN lens implantation 487 For GCaMP6f expression we used 280 nL of the following GCamP6f viral constructs: For CamKII+ 488 neurons we used the AAV5.CamKII.GCaMP6f.WPRE.SV40 virus, titer 2.3x1013, Adgene 100834, 489 injected in C57BL6/J mice; For SST + or PV + neurons we used the 490 AAV5.Syn.Flex.GCaMP6f.WPRE.SV40 virus, Adgene 100833, titer 3.81x1013, in SST-IRES-Cre 491 mice and PV-IRES-Cre mice, respectively). Following craniotomy, the corresponding viral solution 492 was unilaterally injected into the dmPFC (+2.15 mm AP, ±0.35 mm ML relative to the bregma, 493 and −1.4 mm DV from the dura) using a micromanipulator (Scientifica) and pulled glass pipettes 494 (tip diameter ~25 mm). After injection, a track above the imaging site was opened with a sterile 495 needle (26G, 0.45 mm outer diameter, Terumo) to assist in inserting the GRIN lens (Gradient-496 index optics lens, 4 mm long x 0.5 mm wide). The needle was inserted at the exact coordinates 497 of the injection, except for the DV, at 1.0 mm. The GRIN lens was then implanted at the following 498 coordinates: +2.15 mm AP, ±0.35 mm ML relative to the bregma and −1.25 mm DV from the dura. 499 Super-Bond cement was used to secure the GRIN lens. The upper surface of the skull was 500 protected with Kwik-Sil silicone adhesive (World Precision Instruments). 501 Calcium imaging recordings 502 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 21 Two weeks after surgery, to allow for optimal GCaMP6f expression, a miniaturized fluorescence 503 endoscope (miniscope) baseplate was attached to the implant in anesthetized mice. T he 504 miniscope was used for GCaMP6f fluorescence detection so that the baseplate was fixed to the 505 implant based on an optimal field of view within the dmPFC. The miniscope was detached, and 506 the baseplate was covered with a cover. Experiments did not start earlier than 2 weeks after 507 baseplating to allow further viral expression and recovery. One week before starting the 508 behavioral training, mice were habituated to the mounting procedure and the weight of the 509 miniscope for four consecutive days. For the first three days, animals were restrained, the 510 baseplate cover removed, a miniscope dummy was attached to the baseplate, and mice were left 511 with it for at least 30 minutes in their home cage. On the fourth day, mice were mounted with the 512 miniaturized endoscope to set the miniature endoscope light-emitting diode (LED) power and 513 electronic focus. Imaging data were acquired using the nVoke software at a frame rate of 20 Hz 514 (exposure time, 50 ms) with a LED power from 1 to 1.5 mW/mm2, a gain from 1 to 3 and a spatial 515 downsampling by a factor of 2. Recordings of calcium imaging activity during behavioral sessions 516 were recorded in segments around the trial presentations. Imaging started 7 seconds prior to the 517 onset of every trial (CS +, CS− and no-CS trials), continued for the whole duration of the trial and 518 ended 3 seconds after the end of the trial (7 seconds pre-trial + 1-8 seconds CS/US or no-CS + 519 3 seconds post-trial). Therefore, each recording segment lasts from 10 to 18 sec. 520 Processing of calcium imaging data 521 The recordings were processed using the Inscopix Data Processing Software. The time series 522 were joined, pre-processed, spatially filtered and motion corrected before identifying the cells in 523 the field of view. Cell identification was performed through a CNMFe ( Constrained Nonnegative 524 Matrix Factorization for microEndoscopic data; Zhou et al., 2018 ) algorithm and each trace was 525 validated or rejected by the experimenter based on the identified morphology, activity trace and 526 calcium transients. We obtained the single-cell denoised calcium dF/F traces from each accepted 527 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 22 neuron. Neural activity of all recorded mice was aligned to trial onset and the activity of each 528 individual neuron was z-scored using a reference period of -2 to 0 seconds before CS onset. To 529 quantify the percentage of modulated cells at CS onset and offset, we calculated the mean z-530 score activity aligned at CS onset (0 to 1 s from CS onset) or CS offset (0 to 1 s from CS offset) 531 and considered individual neurons to be active or inhibited if their mean activity was above 1.96 532 or below -1.96, respectively. We then calculated the percentage of modulated cells as the total 533 number of active and inhibited cells over the total number of recorded cells for each recorded 534 mouse. 535 Anatomical and histological analyses 536 Mice were euthanized with a solution containing pentobarbital (Exagon, 0.27 mg/g) and lidocaine 537 (Lidor, 0.03 mg/g) and perfused via the left ventricle with 4% w/v paraformaldehyde (PFA) in 0.1 538 M PBS. Following dissection, brains were postfixed in PFA for 24h at 4 °C. Brain sections of 60-539 mm thickness were serially cut using a vibratome, mounted on gelatin-coated microscope slides 540 and protected with vectashield antifade mounting medium (Vector Laboratories) and a coverslip. 541 Slices were imaged using an epifluorescence system (Leica DM 5000) fitted with a 10x dry 542 objective. The location of GRIN lens implantation and extent of the viral injections were visually 543 controlled. Only infections accurately targeting the dmPFC and GRIN lenses terminating within 544 this region were considered for the analyses. 545 Population analyses and decoding 546 To investigate the activity of the population of recorded neurons, the calcium activity from the 547 ensemble of recorded cells at a given time point t was pooled in a vector containing as many 548 dimensions as neurons recorded. This instantaneous population vector contained the calcium 549 activity of all the recorded neurons and mice at each time point. For population activity we used 550 the neuronal activity analyzed in 50-ms time bins and relative to the 2 seconds prior to CS 551 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 23 presentation by subtracting the mean dF/F in that time period from the trial. To assess the 552 decoding accuracies of the dmPFC population, we used linear kernel SVM classifiers in single-553 trial instantaneous pseudo-population vectors (including units and trials from different mice) 554 composed of the calcium activity of each neuron (50-ms time bins) after the onset of the five 555 different trial types (active, passive, inverse, CS − and no-CS trials). For the decoding of threat 556 versus threat, threat versus control, threat versus no-CS or control versus no-CS, the decoding 557 was performed among population vectors at time t in relation to the onset of two different trial 558 types. To obtain comparable accuracies across the classifiers used in the different conditions, we 559 used for all classifiers the same number of trials to train and test decoders. We used a leave-one-560 out cross-validation (LOO-CV) method from a total number of 5 trials for each class. To account 561 for the different number of neurons in each dataset we subsampled 100 random neurons for each 562 iteration of the classifier. Using this method, we first selected 100 random neurons and trained a 563 classifier cross-validated by a LOO-CV method. To obtain the average decoding accuracies, this 564 technique was repeated 200 times, randomly selecting in each iteration 100 different neurons and 565 different trials to construct the single-trial pseudo-population vectors. We used 5 trials of each 566 class for all classifiers to obtain comparable decoding accuracies among the different classifiers. 567 To evaluate the difference between the obtained decoding accuracies with the accuracies that 568 could be obtained by chance, we compared the decoding accuracies of our classifiers with that 569 obtained using a ‘shuffled’ condition. To do this, we employed the same method described above 570 but in which the class identity of the test dataset was randomly assigned. Next, we used a two-571 sided permutation test by computing P values as the proportion of the shuffle repetitions that 572 exceeded the decoding accuracies obtained with the method above using the real identity of 573 classes. Significant decoding accuracies were defined as P < 0.05. To investigate the 574 generalization of the population activity patterns in the different neuronal types we used the 575 previous method to train a classifier to decode a random threat type (active, passive or inverse) 576 from basal activity (no-CS trials) and tested it with the same classes to obtain the ‘threat’ decoding 577 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 24 accuracy. We then challenged that classifier to classify a different threat type (not used to train 578 the classifier) from a no-CS trial, obtaining the ‘threat-general’ decoding accuracy. The same 579 classifier was used to classify the control trials from the no- CS to obtain the ‘control-general’ 580 classification accuracies. We measured the CS +-CS+ and CS +-control accuracy loss as the 581 difference in accuracy obtained in each iteration between the ‘threat’ classification accuracy and 582 the ‘threat-general’ and ‘control-general’ classification accuracies. For tone and context 583 discrimination decoding, the same decoding strategy and statistical methods described above 584 were used. For Tone Discrimination decoding, 5 randomly selected trials among active and 585 passive trials were used as a single class decoded from 5 randomly selected inverse trial. For 586 Context discrimination 5 randomly selected trials among inverse and passive trials were used as 587 a single class decoded from 5 randomly selected active trial. 588 Statistics and reproducibility 589 No statistical methods were used to predetermine sample size. Experiments were not 590 randomized, and investigators were not blinded to allocation during experiments and outcome 591 assessment. We conducted all analyses using either custom routines written in MATLAB 592 (Mathworks) or SigmaPlot (Systat Software). Figures were assembled using Adobe Illustrator. 593 For all datasets, normality was tested using Kolmogorov-Smirnov test (α < 0.05) and homogeneity 594 of variance with Levene’s test ( α < 0.05) to determine whether parametric or non-parametric 595 analyses were required. We used parametric analyses including one-way ANOVA followed by 596 post hoc Tukey’s test if a significant main effect or interaction was observed. For all tests, * P < 597 0.05, **P < 0.01, ***P 0.5; significance was calculated after correction 598 for multiple comparisons. If either homogeneity of variance or normality assumptions were not 599 met, nonparametric analyses were used. When required, we used nonparametric Kruskal-Wallis 600 one-way ANOVA on ranks and Dunn’s multiple -comparison post hoc test to protect from false 601 positive errors. All behavioral and imaging data were collected in a computer-controlled, 602 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 25 automated and unbiased manner. No statistical methods were used to predetermine sample 603 sizes, but sample sizes were similar to those reported in previous publications7,33 604 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 26 Figure legends: 605 Figure 1. Calcium imaging recordings of Pyr, SST+ and PV+ neurons during the multi input-606 output paradigm. a, Representation of the behavioral box employed in the multi input-output 607 (MIO) paradigm. The behavioral box is divided in shelter (green) and arena (grey). b, Schematic 608 drawing showing the miniaturized endoscopes used to record the calcium levels of the different 609 neuronal types in freely moving mice. c, Top: Schematic representations of the three threatening 610 situations presented in the MIO paradigm (active, passive and inverse trials). Bottom: Average 611 performance across mice in the MIO paradigm (n= 23 mice) in active (red), passive (blue) and 612 inverse (magenta) trials. Shaded areas represent S.E.M. d, Left: Representative fields of view of 613 three mice showing the regions of interest detected with CNMFe containing pyramidal (black), 614 SST+ (turquoise) and PV + (orange) neurons. Fields of view obtained with lenses of 500 µm 615 diameter. Right: Individual (grey) and mean (black, turquoise or orange) calcium traces of a 616 pyramidal, SST+ and PV+ example neurons during active, passive, inverse and control trials. Note 617 the variability and heterogeneity of neuronal responses. Scale bar 100% dF/F for single trial traces 618 and 10% dF/F for mean trials. 619 Figure 2. Activity of P yr, SST+ and PV + neurons during CS onset and offset. a, Schematic 620 representations of mice behavior during CS onset and CS offset of active, passive, inverse and 621 control trials. b, Z-scored mean activity of pyramidal neurons (bla ck, n=1253 neurons, from n=9 622 mice) during CS onset (left) and CS offset (right) of active, passive, inverse and control trials. c, 623 As b, but for PV+ neurons (orange, n=211 neurons, from n=8 mice). d, As b, but for SST+ neurons 624 (turquoise, n=208 neurons, from n= 6 mice). e, Mean percentage of modulated (activated + 625 inhibited) neurons at CS onset (from 0 to 1 seconds after CS onset) and CS offset (from 0 to 1 626 seconds from CS offset) across mice in active, passive, inverse and control trials (n=9 , 8 and 6 627 mice for Pyr , PV + and SST + neurons respectively). Small points represent percentages of 628 modulated neurons in individual mice. Error bars represent S.E.M. 629 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint 27 Figure 3. Encoding of threat-related information in different dmPFC neuronal populations. 630 a, Schematic representation of the strategy employed to construct the pseudo-population vectors 631 and the support vector machine (SVM) classifiers used to decode the patterns of population 632 activity presented in the different trials. b, Top: Classification accuracies obtained around CS 633 onset for classifiers challenged to decode between CS+ trials and no-CS conditions during active 634 (red), passive (blue) and inverse (magenta) trials. Shuffle decoding accuracies in grey. Significant 635 classification accuracies represented as horizontal bars in red (active trials), blue (passive trials) 636 and magenta (inverse trials). Bottom: As top but showing classification accuracies obtained for 637 classifiers challenged to decode between CS + and control ( CS−) trials. c, Mean classification 638 accuracy between CS+ and no-CS trials (thick lines) and CS+ and control trials (thin lines) obtained 639 with classifiers that used different numbers of neurons. Classifiers used active (top), passive 640 (middle) and inverse (bottom) trials and Pyr (black), SST+ (turquoise) and PV+ (orange) neurons. 641 d, Average threat (brown), threat-general (orange), control-general (pink) and control 642 classification accuracies for classifiers built with P yr, SST + and PV + neurons. Shuffle decoding 643 accuracies in grey. Significant classification accuracies represented as horizontal bars in brown 644 (threat-decoding), orange (threat-general), pink (control-general) and black (control). e, CS+-CS+ 645 and CS+-control accuracy loss (see Methods) in Pyr, SST+ and PV+ neurons. Box plot center line 646 represents median; box limits are 25th and 75th percentiles; whiskers are more extreme non-647 outliers data points. f, Left: Schematic representation of the strategy employed to decode among 648 threats with different tones. Right: Tone-decoding accuracies relative to the CS + onset in Pyr , 649 SST+ and PV+ neurons. g, As f, but for the decoding of threats presented in different contexts. 650 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint Shelter Arena Multi input-output paradigm Figure 1. Calcium imaging recordings of Pyr, SST+ and PV+ neurons during the multi input-output paradigm. a c db CS2: 2 kHz Active trial Inverse trial Passive trial CS1: 7 kHz CS1: 7 kHz Performance (%) Day 0 25 50 75 100 Hab1 2 3 4 5 6 7 Day Hab1 2 3 4 5 6 7 Day Hab1 2 3 4 5 6 7 0 25 50 75 100 0 25 50 75 100 Active Passive Inverse Control -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3-2 -1 0 1 2 3 Trial #Trial mean -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3-2 -1 0 1 2 3 Trial #Trial mean -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3-2 -1 0 1 2 3 Trial #Trial mean Time from CS onset (s) Time from CS onset (s) Time from CS onset (s) Time from CS onset (s) Pyr neurons PV+ neurons SST+ neurons Example Pyr neuron Example PV+ neuron Example SST+ neuron (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint -2 -1 0 1 2 3 1 300 600 900 1200 -2 -1 0 1 2 3 1 300 600 900 1200 -2 -1 0 1 2 3 1 300 600 900 1200 -2 -1 0 1 2 3 1 300 600 900 1200 0 -5 5 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 -2 -1 0 1 2 3 1 50 100 150 200 Active trials CS onset Passive trials CS onset Inverse trials CS onset Control trials % Modulated neurons -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 -2 -1 0 1 2 3 Pyr PV SST Pyr PV SST Pyr PV SST Pyr PV SST Pyr PV SST Pyr PV SST Pyr PV SST Pyr PV SST 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 Time from CS onset (s) Time from CS onset (s) Time from CS onset (s) Time from CS onset (s) CS offset CS onset CS offset CS offset CS offset Active trials CS onset Passive trials CS onset Inverse trials CS onset Control trials CS offset CS onset CS offset CS offset CS offset Time from CS offset (s) Time from CS offset (s) Time from CS offset (s) *** *** * * * ****** Time from CS offset (s) Pyr neuronsPV+ neurons SST+ neurons Z-score 0 -5 5 Z-score 0 -5 5 Z-score a e b c d ***** Figure 2. Activity of Pyr, SST+ and PV+ neurons during CS onset and offset (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint CS+ vs control Pyramidal neurons Pyramidal neurons SST+ neurons SST+ neurons PV+ neurons PV+ neurons CS+ vs no-CS 100 randomly selected neurons x200 LOO-CV Train decoder: Neuron sampling: Test decoder: Active vs control SVM classification accuracy Active trials Control trials Time from CS onset (s) Accuracy Accuracy Other neurons 0 1 0.2 0.4 0.6 0.8 0 0 1 2-1 0 1 2-1 0 1 2-1 1 0.2 0.4 0.6 0.8 Used neuron a d e b c CS Unit1 Time Active trial (t2) Control trial (t1) Control trial Active trial Unit2 Unit3 Unit3 Unitn t1 t2 ... Unit1 Unit2 Unit3 Unitn Time from CS onset (s) Time from CS onset (s) Time from CS onset (s) 0 1 2-1 0 1 2-1 Accuracy loss 0 1 0.2 0.4 0.6 0.8 0 1 2-1 0 1 2-1 0 1 0.2 0.4 0.6 0.8 0 1 2-1 0 1 2-1 0 1 0.2 0.4 0.6 0.8 Threat-decoding Control-decoding Threat-decoding Control-decoding Threat-decoding Control-decoding P trials I trials Pyr PV Number of neurons used A trialsActive Passive Inverse Shuffle SST Pyr PVSST 1 100 200 1 100 200 1 100 200 Accuracy 0 1 0.2 0.4 0.6 0.8 0 1 0.2 0.4 0.6 0.8 0 1 0.2 0.4 0.6 0.8AccuracyAccuracy No-CS Control No-CS Control No-CS Control CS + -CS + CS + -controlCS + -CS + CS + -control CS + -CS + CS + -control 0 0.2 0.1 -0.1 0.3 0.4 0 0.2 0.1 -0.1 0.3 0.4 0 0.2 0.1 -0.1 0.3 0.4 f CS CS+2 trialCS+1 trial Active and Passive trial (t1) Inverse trial (t2) Unit1 Unit2 Unit3 Unitn 0 1 2-1 0 1 2-1 0 1 2-1 Pyramidal neurons Dscrimination of threats with different tones SST+ neurons PV+ neurons Accuracy 0 1 0.2 0.4 0.6 0.8 g Passive and Inverse trials (t1) Active trial (t2) CS Arena trialShelter trial Unit1 Unit2 Unit3 Unitn Discrimination of threats in different context Time from CS onset (s)Time from CS onset (s) 0 1 2-1 0 1 2-1 0 1 2-1 Pyramidal neurons SST+ neurons PV+ neurons Accuracy 0 1 0.2 0.4 0.6 0.8 Threat-generalizationThreat Control Control-generalization Shuffle Figure 3. Encoding of threat-related information in different dmPFC neuronal populations. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 24, 2024. ; https://doi.org/10.1101/2024.01.24.577067doi: bioRxiv preprint