Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Article
  • Published:

Prefrontal circuits encode both general danger and specific threat representations

Abstract

Behavioral adaptation to potential threats requires both a global representation of danger to prepare the organism to react in a timely manner but also the identification of specific threatening situations to select the appropriate behavioral responses. The prefrontal cortex is known to control threat-related behaviors, yet it is unknown whether it encodes global defensive states and/or the identity of specific threatening encounters. Using a new behavioral paradigm that exposes mice to different threatening situations, we show that the dorsomedial prefrontal cortex (dmPFC) encodes a general representation of danger while simultaneously encoding a specific neuronal representation of each threat. Importantly, the global representation of danger persisted in error trials that instead lacked specific threat identity representations. Consistently, optogenetic prefrontal inhibition impaired overall behavioral performance and discrimination of different threatening situations without any bias toward active or passive behaviors. Together, these data indicate that the prefrontal cortex encodes both a global representation of danger and specific representations of threat identity to control the selection of defensive behaviors.

This is a preview of subscription content, access via your institution

Access options

Buy this article

Prices may be subject to local taxes which are calculated during checkout

Fig. 1: New MIO behavioral paradigm.
Fig. 2: Single-neuron activity during different trials.
Fig. 3: Encoding of general danger representation at the dmPFC population level.
Fig. 4: General danger and threat-specific neuronal representations in dmPFC neuronal populations.
Fig. 5: Behavior analyses in the MIO paradigm.

Similar content being viewed by others

Data availability

The data that support the findings of this study are available from the corresponding authors on request.

Code availability

Custom written codes used to analyze data from this study are available from the corresponding authors on request.

References

  1. LeDoux, J. E. Emotion circuits in the brain. Annu. Rev. Neurosci. 23, 155–184 (2000).

    CAS  PubMed  Google Scholar 

  2. Kim, J. J. & Fanselow, M. S. Modality-specific retrograde amnesia of fear. Science 256, 675–677 (1992).

    CAS  PubMed  Google Scholar 

  3. Haubensak, W. et al. Genetic dissection of an amygdala microcircuit that gates conditioned fear. Nature 468, 270–276 (2010).

    CAS  PubMed  PubMed Central  Google Scholar 

  4. Milad, M. R. & Quirk, G. J. Neurons in medial prefrontal cortex signal memory for fear extinction. Nature 420, 70–74 (2002).

    CAS  PubMed  Google Scholar 

  5. Likhtik, E., Popa, D., Apergis-Schoute, J., Fidacaro, G. A. & Pare, D. Amygdala intercalated neurons are required for expression of fear extinction. Nature 454, 642–645 (2008).

    CAS  PubMed  PubMed Central  Google Scholar 

  6. Courtin, J. et al. Prefrontal parvalbumin interneurons shape neuronal activity to drive fear expression. Nature 505, 92–96 (2014).

    PubMed  Google Scholar 

  7. Likhtik, E., Stujenske, J. M., Topiwala, M. A., Harris, A. Z. & Gordon, J. A. Prefrontal entrainment of amygdala activity signals safety in learned fear and innate anxiety. Nat. Neurosci. 17, 106–113 (2014).

    CAS  PubMed  Google Scholar 

  8. Herry, C. et al. Switching on and off fear by distinct neuronal circuits. Nature 454, 600–606 (2008).

    CAS  PubMed  Google Scholar 

  9. Hagihara, K. M. et al. Intercalated amygdala clusters orchestrate a switch in fear state. Nature 594, 403–407 (2021).

    CAS  PubMed  PubMed Central  Google Scholar 

  10. Tovote, P. et al. Midbrain circuits for defensive behaviour. Nature 534, 206–212 (2016).

    CAS  PubMed  Google Scholar 

  11. Jercog, D. et al. Dynamical prefrontal population coding during defensive behaviours. Nature 595, 690–694 (2021).

    CAS  PubMed  Google Scholar 

  12. Moscarello, J. M. & LeDoux, J. E. Active avoidance learning requires prefrontal suppression of amygdala-mediated defensive reactions. J. Neurosci. 33, 3815–3823 (2013).

    CAS  PubMed  PubMed Central  Google Scholar 

  13. Rogan, M. T., Staubli, U. V. & LeDoux, J. E. Fear conditioning induces associative long-term potentiation in the amygdala. Nature 390, 604–607 (1997).

    CAS  PubMed  Google Scholar 

  14. Nader, K., Schafe, G. E. & Le Doux, J. E. Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Nature 406, 722–726 (2000).

    CAS  PubMed  Google Scholar 

  15. Nabavi, S. et al. Engineering a memory with LTD and LTP. Nature 511, 348–352 (2014).

    CAS  PubMed  PubMed Central  Google Scholar 

  16. Chen, B. K. et al. Artificially enhancing and suppressing hippocampus-mediated memories. Curr. Biol. 29, 1885–1894 (2019).

    CAS  PubMed  PubMed Central  Google Scholar 

  17. Sengupta, A. & Holmes, A. A discrete dorsal raphe to basal amygdala 5-HT circuit calibrates aversive memory. Neuron 103, 489–505 (2019).

    CAS  PubMed  PubMed Central  Google Scholar 

  18. Grewe, B. F. et al. Neural ensemble dynamics underlying a long-term associative memory. Nature 543, 670–675 (2017).

    CAS  PubMed  PubMed Central  Google Scholar 

  19. Taub, A. H., Perets, R., Kahana, E. & Paz, R. Oscillations synchronize amygdala-to-prefrontal primate circuits during aversive learning. Neuron 97, 291–298 (2018).

    CAS  PubMed  Google Scholar 

  20. Sotres-Bayon, F., Sierra-Mercado, D., Pardilla-Delgado, E. & Quirk, G. J. Gating of fear in prelimbic cortex by hippocampal and amygdala inputs. Neuron 76, 804–812 (2012).

    CAS  PubMed  PubMed Central  Google Scholar 

  21. Bravo-Rivera, C., Roman-Ortiz, C., Brignoni-Perez, E., Sotres-Bayon, F. & Quirk, G. J. Neural structures mediating expression and extinction of platform-mediated avoidance. J. Neurosci. 34, 9736–9742 (2014).

    CAS  PubMed  PubMed Central  Google Scholar 

  22. Penzo, M. A. et al. The paraventricular thalamus controls a central amygdala fear circuit. Nature 519, 455–459 (2015).

    CAS  PubMed  PubMed Central  Google Scholar 

  23. Dejean, C. et al. Prefrontal neuronal assemblies temporally control fear behaviour. Nature 535, 420–424 (2016).

    CAS  PubMed  Google Scholar 

  24. Rozeske, R. R. et al. Prefrontal–periaqueductal gray-projecting neurons mediate context fear discrimination. Neuron 97, 898–910 (2018).

    CAS  PubMed  Google Scholar 

  25. Kyriazi, P., Headley, D. B. & Pare, D. Different multidimensional representations across the amygdalo–prefrontal network during an approach–avoidance task. Neuron 107, 717–730 (2020).

    CAS  PubMed  PubMed Central  Google Scholar 

  26. Padilla-Coreano, N. et al. Hippocampal–prefrontal theta transmission regulates avoidance behavior. Neuron 104, 601–610 (2019).

    CAS  PubMed  PubMed Central  Google Scholar 

  27. Choi, J. S. & Kim, J. J. Amygdala regulates risk of predation in rats foraging in a dynamic fear environment. Proc. Natl Acad. Sci. USA 107, 21773–21777 (2010).

    CAS  PubMed  PubMed Central  Google Scholar 

  28. Namburi, P. et al. A circuit mechanism for differentiating positive and negative associations. Nature 520, 675–678 (2015).

    CAS  PubMed  PubMed Central  Google Scholar 

  29. Kennedy, A. et al. Stimulus-specific hypothalamic encoding of a persistent defensive state. Nature 586, 730–734 (2020).

    CAS  PubMed  PubMed Central  Google Scholar 

  30. Zhang, X. et al. Genetically identified amygdala–striatal circuits for valence-specific behaviors. Nat. Neurosci. 24, 1586–1600 (2021).

    CAS  PubMed  PubMed Central  Google Scholar 

  31. Stephenson-Jones, M. et al. Opposing contributions of GABAergic and glutamatergic ventral pallidal neurons to motivational behaviors. Neuron 105, 921–933 (2020).

    CAS  PubMed  PubMed Central  Google Scholar 

  32. Zhang, X. & Li, B. Population coding of valence in the basolateral amygdala. Nat. Commun. 9, 5195 (2018).

    PubMed  PubMed Central  Google Scholar 

  33. Burgos-Robles, A. et al. Amygdala inputs to prefrontal cortex guide behavior amid conflicting cues of reward and punishment. Nat. Neurosci. 20, 824–835 (2017).

    CAS  PubMed  PubMed Central  Google Scholar 

  34. Pare, D. & Quirk, G. J. When scientific paradigms lead to tunnel vision: lessons from the study of fear. NPJ Sci. Learn. 2, 6 (2017).

    PubMed  PubMed Central  Google Scholar 

  35. Mobbs, D. et al. Viewpoints: approaches to defining and investigating fear. Nat. Neurosci. 22, 1205–1216 (2019).

    CAS  PubMed  PubMed Central  Google Scholar 

  36. Karalis, N. et al. 4-Hz oscillations synchronize prefrontal–amygdala circuits during fear behavior. Nat. Neurosci. 19, 605–612 (2016).

    CAS  PubMed  PubMed Central  Google Scholar 

  37. Bagur, S. et al. Breathing-driven prefrontal oscillations regulate maintenance of conditioned-fear evoked freezing independently of initiation. Nat. Commun. 12, 2605 (2021).

    CAS  PubMed  PubMed Central  Google Scholar 

  38. Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks. Nature 497, 585–590 (2013).

    CAS  PubMed  PubMed Central  Google Scholar 

  39. Cunningham, J. P. & Yu, B. M. Dimensionality reduction for large-scale neural recordings. Nat. Neurosci. 17, 1500–1509 (2014).

    CAS  PubMed  PubMed Central  Google Scholar 

  40. Ceko, M., Kragel, P. A., Woo, C. W., Lopez-Sola, M. & Wager, T. D. Common and stimulus-type-specific brain representations of negative affect. Nat. Neurosci. 25, 760–770 (2022).

    CAS  PubMed  Google Scholar 

  41. Grundemann, J. et al. Amygdala ensembles encode behavioral states. Science 364, eaav8736 (2019).

  42. Gehrlach, D. A. et al. Aversive state processing in the posterior insular cortex. Nat. Neurosci. 22, 1424–1437 (2019).

    CAS  PubMed  Google Scholar 

  43. Riberto, M., Paz, R., Pobric, G. & Talmi, D. The neural representations of emotional experiences are more similar than those of neutral experiences. J. Neurosci. 42, 2772–2785 (2022).

    CAS  PubMed  PubMed Central  Google Scholar 

  44. Stringer, C. et al. Spontaneous behaviors drive multidimensional, brainwide activity. Science 364, 255 (2019).

    PubMed  PubMed Central  Google Scholar 

  45. Musall, S., Kaufman, M. T., Juavinett, A. L., Gluf, S. & Churchland, A. K. Single-trial neural dynamics are dominated by richly varied movements. Nat. Neurosci. 22, 1677–1686 (2019).

    CAS  PubMed  PubMed Central  Google Scholar 

  46. Courtin, J. et al. A neuronal mechanism for motivational control of behavior. Science 375, eabg7277 (2022).

    CAS  PubMed  Google Scholar 

  47. Resnik, J., Sobel, N. & Paz, R. Auditory aversive learning increases discrimination thresholds. Nat. Neurosci. 14, 791–796 (2011).

    CAS  PubMed  Google Scholar 

  48. Mathis, A. et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nat. Neurosci. 21, 1281–1289 (2018).

    CAS  PubMed  Google Scholar 

Download references

Acknowledgements

We thank T. Bienvenu, P. Meyrand and J. Viellard for discussions and reading the manuscript; K. Deisseroth and E. Boyden for sharing material; and S. Laumond, J. Tessaire and the technical staff of the housing and experimental animal facility of the Neurocentre Magendie. Microscopy was performed in the Bordeaux Imaging Center of the CNRS–INSERM and Bordeaux University, a member of France BioImaging. This work was supported by EMBO (ALTF 200-2018) and HFSP postdoctoral fellowships (LT0000658/2019) to M.M.-F., grants from the French National Research Agency (ANR-10-EQPX-08 OPTOPATH), the Conseil Regional d’Aquitaine and the Fondation pour la Recherche Médicale (FRM-Equipes FRM 2017). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Author information

Authors and Affiliations

Authors

Contributions

M.M.-F. and A.P.M. performed electrophysiological recordings. M.M.-F., A.P.M. and G.L.-F. performed behavior and optogenetic experiments. A.P.M., D.G. and G.L.-F. performed histology. M.M.-F., A.P.M., G.L.-F., H.-R.K. and D.J. performed surgeries. D.J. and N.W. performed electrophysiology and analyses associated with the appetitive–aversive conditioning paradigm. M.M.-F. and C.H. designed experiments. M.M.-F. designed single-unit and population analyses and wrote software codes. C.D. and M.M.-F. designed and wrote code for LFP analysis. M.M.-F. and C.H. analyzed data and wrote the paper. All the authors read and edited the manuscript.

Corresponding authors

Correspondence to Mario Martin-Fernandez or Cyril Herry.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Neuroscience thanks the anonymous reviewers for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Behavior analyses in the MIO paradigm.

a, To evaluate whether passive responses in the MIO task are instrumental responses we trained animals in a modified version of the MIO task in which the passive trials (grey; d0 vs d1 p = 0.012, d0 vs Learned p = 0.034 and d1 vs learned p < 0.001) were not reinforced (n = 16; that is foot shocks were never delivered when the CS+ 1 was presented in the shelter). Active (red; d0 vs d1 p = 0.011, d0 vs learned p < 0.001 and d1 vs learned p < 0.001) and inverse trials (purple; d0 vs d1 p = 0.42, d0 vs learned p < 0.001 and p < 0.001) were reinforced. Graphs depict mean performance of the different trial types across sessions. Shaded areas and error bas represent S.E.M. b, Left: Performance across learning of passive trials in the passive non-reinforced MIO modification (n = 16) and littermates run in the standard version of the MIO paradigm (n = 16) in which passive responses were reinforced. Right: Mean Performance of passive trials in the passive non-reinforced (P-nr) and in the standard version of the MIO in which passive responses are reinforced (P-r). We show data in the habituation (d0), first day of training (d1) and last day of training (d7) and observed a difference between P-nr and P-r during d7 but not during d0 or d1, suggesting that passive responses are instrumental after training but not early in training. Shaded areas and error bars represent S.E.M. Two-way ANOVA with post hoc Tukey, d0 p = 0.75, d1 p = 0.97 and d7 p < 0.001. c, Distribution of day of best performance day over training sessions in the mice implanted with electrodes for recordings of neuronal activity. d, Performance of mice implanted with recording electrodes (n = 24) in active, passive and inverse Trials, Shuttling in CS trials and Global Performance in the habituation (d0), first day of training (d1) and best performance day (Learned). P values d0 vs d1, d0 vs learned and d1 vs learned are: 0.95, <0.001 and <0.001 for Active trials, <0.001, <0.001 and 0.326 for passive trials, <0.001, 0.72 and <0.001 for inverse trials, 0.015, 0.1 and 0.69 for CS minus shuttling, and 0.028, <0.001 and <0.001 for global performance. e, Behavioral discrimination of tone and context in the habituation (d0), first day of training (d1) and best performance day (Learned; n = 24. Tone discrimination: d0 vs d1 p = 0.51, d0 vs learned p = 0.022 and d1 vs learned p < 0.001. Context discrimination: d0 vs d1 p < 0.001, d0 vs learned p < 0.001 and d1 vs learned p < 0.001). For d and e bar plots represent mean values across mice and grey dots individual value of each mice. Error bars represent S.E.M. f, Latency to shuttling in active correct, passive error and inverse correct trials (n = 24) measured as the time difference between the onset and the offset of the CS. Central dot represents mean and error bar the standard deviation. Violin plots represent relative distribution of single trials latencies. g, Preference of mice for the shelter over the arena measured as the Shelter ratio across training sessions (number trials in which mice were in the shelter at the CS onset minus the number of trials in which mice were in the arena at the CS onset, over the total number of trials. Since trials are presented randomly during the session, the proportion of trials in the shelter serves as a measure of the proportion of time spent in the shelter). Shaded areas represent S.E.M. h, Top: Number of active, passive and inverse trials in the day of best performance for implanted mice (n = 24). Bottom: Number of correct trials for active, passive and inverse trials in the day of best performance for implanted mice. Error bars represent S.E.M. i, Top: Median speed (shaded areas are 95% confidence intervals) across trials around the CS offset in active, passive, inverse and CS trials. Bottom: Single trial speed for active, passive inverse and CS trials. j, Mice median speed during the CS onset for error active, passive and inverse trials. Shaded areas represent 95% confidence intervals. k, Change of median speed over baseline for the correct and error trials of active (red; correct trials: 304 trials from 24 mice; error trials: 89 trials frim 24 mice), passive (blue; correct trials: 421 trials from 24 mice; error trials: 84 trials from 24 mice), inverse (magenta; correct trials: 231 trials from 24 mice; error trials: 311 trials from 24 mice) and CS (gray; 720 trials from 24 mice) trials in basal conditions, at the onset of the CS (from 0 to 2 seconds) and at the CS offset (from −2 to 0 seconds). Shaded areas correspond to S.E.M. l, Schematic representation showing the position of the markers tracked with DLC. m, Median speed of the different markers tracked in all trial types aligned to the CS onset. Shaded areas represent 95% confidence intervals. *P < 0.05; One-way and Two-way ANOVA with post hoc Tukey.

Extended Data Fig. 2 dmPFC recordings in the MIO paradigm.

a, Left: Representative coronal slice (of n = 24 mice) showing the location of the recording electrode (red) in the dmPFC region (orange). Right: Schemes indicating the location of the recording electrodes in the dmPFC. Scale bar 1 mm. b, Representative mean waveforms of two sorted units (green and pink) and the unsorted waveforms (grey) and the projection of these waveforms in PC1-3 space. Shaded areas represent standard deviation c, Autocorrelograms of the two sorted units shown in b (time bin 1 ms). d, Z-scored dmPFC LFP power aligned to the CS onset of n = 19 mice during the different trial types in the habituation (Day 0), first day of training (Day 1) and day of best performance (Learned). For Day 1 and Learned day, only correct trials are used. e, Z-scored dmPFC LFP power aligned to the CS onset during error trials of the different trial types. f, Average z-scored 4-Hz power at the onset of the CS (n = 19; 4–7 Hz from 0 to 2 seconds after the CS onset) for the different trial types in the habituation (D0), first day of training (D1) and day of best performance (Learned; L). The p values for the d0 vs d1, d0 vs Learned and d1 vs Learned comparisons are 0.41, 0.003 and 0.1 for active trials, 0.99, 0.043 and 0.05 for passive trials, 0.035, <0.001 and <0.001 for inverse trials and 0.08, 0.012 and 0.69 for control trials. g, Average z-scored 4-Hz power at the onset of the CS for n = 14 mice during active, passive and inverse error trials and for the control trials. P values: Control vs Active=0.004, control vs passive=0.003, control vs inverse=0.004. All other comparisons p > 0.05. h, Heat maps of z-scored activity aligned to the CS offset in different trial types for the n = 546 dmPFC recorded neurons from 24 recorded mice. i, Percentage of activated (red) and inhibited (blue) neurons around CS offset (250 ms time bins) for the different trial types. Green bar marks the offset of the CS, j, Z-scored dmPFC LFP power relative to the CS offset of the different trial types for n = 19 mice. Box plot center line represents the median; box limits are 25th and 75th percentiles; whiskers are more extreme non-outliers data points; ‘+’ markers represent outliers. *p < 0.05; One-way ANOVA with Tukey post hoc.

Extended Data Fig. 3 dmPFC population analyses.

a, Heatmaps of neuronal activity aligned to the CS onset of the different trial types for the 546 dmPFC recorded neurons from 24 mice. Neurons are sorted based on their activity at the onset of the active (top left), passive (top right), inverse (bottom left) or CS minus (bottom right) trials and the order is kept for all the display of neuronal activity during the rest of trial types. b, Sub-categorization of mixed responsive neurons obtained in Fig. 2d. Mixed responsive neurons are subcategorized in 4 categories: Auditory-threat neurons are those neurons showing modulation at the onset of the CS+1 independently of the mice location (that is during passive and active trials) but not by CS+2. Spatially-selective neurons are those modulated at the onset of threats within the shelter but not outside of the shelter (that is only during passive and inverse trials). Motor-planning neurons are those activated at the onset of only the threats eliciting shuttling responses (that is only during active and inverse trials). Other mixed responsive neurons are the neurons showing mixed responses not belonging to any of the previous 3 categories (these neurons are modulated by the CS minus and additionally either one or two CS+ trials). c, Left: Mean single-neuron decoding accuracies obtained across the recorded dmPFC neuronal population aligned to CS onset. Note that the mean increase at CS onset reaches ~55% decoding accuracy, showing that globally dmPFC neurons have limited decoding power. Shaded areas represent 95% confidence intervals. Right: Percentage of neurons exhibiting significant decoding accuracy as during active, passive, inverse or Control (CS) trials. d, Schematic depicting the approach used to investigate differences of population vector activities at CS onset. e, Activity of neuronal population vectors at CS onset for the different trials sorted according to CS onset neuronal activity of the different trial types (n = 546 neurons from 24 mice). f, Top Left: Mean euclidean distance at CS onset between population vectors for different trials (active, passive, inverse and CS minus) and the Pre-CS basal activity. Distances among threats are shown inside of the green square. Top Right: Mean Euclidean distance across time relative to CS onset between population vectors for different trials (active, passive, inverse and CS minus) and the Pre-CS basal activity. Bottom Left: Mean Euclidean distance across time relative to CS onset among population vectors for the different threatening trials (active, passive, and inverse). Bottom Right: Mean Euclidean distance across time relative to CS onset between population vectors for different threatening trials (active, passive, and inverse) and CS trials. Shaded areas represent standard deviation. g, Left: Mean decoding accuracy matrix for comparisons between active, passive, inverse and CS trials at CS onset (from 0 to 250 ms after the onset of the CS). Differences among threats are shown inside of the green square. Right: Mean decoding accuracies across time relative to CS onset between population vectors for different threatening trials (active, passive, and inverse) and CS trials. h, Effect of changing the number of trials used for decoders in the obtained decoder accuracy at the CS onset of active, passive and inverse trials. Increasing the minimum number of trials forced us to exclude mice and neurons from the analysis. The decoders made with 2, 3 and 4 trials used 546 neurons from 24 mice. The decoder made with 5 trials used 440 neurons from 20 mice. The decoder made with 6 trials used 362 neurons from 17 mice. The decoder made with 10 trials used 104 neurons from 6 mice. Note that with 4 trials decoders reach ~ 0.9 accuracy and no much improvement is achieved by increasing the numbers of trials at the cost of losing neurons. i, Top: Average decoding accuracies of CS+ versus Pre-CS basal activity (top, red) and average CS+ vs CS+ decoding accuracies (bottom, green) around CS onset. Shuffle accuracies in grey. Bottom: Decoding accuracies of CS+ vs Pre-CS (red) and among threats (green) at CS onset (from 0 to 250 ms after CS onset, n = 100 repetitions). Box plot center line represents the median; box limits are 25th and 75th percentiles; whiskers are more extreme non-outliers data points; ‘+’ markers represent outliers. j, Top: Schematic depicting the strategy used for Threat-identity decoding accuracies (AC: active correct trials; PC: passive correct trials; IC: inverse correct trials). Bottom: Threat-identity decoding accuracies relative to CS onset (green) and shuffle decoding accuracies (grey).k, Top: Schematic depicting the strategy used for Global-danger decoding accuracies. Bottom: Global-danger decoding accuracies relative to CS onset (orange) and shuffle decoding accuracies (grey). l, Top: Median speed of the immobility events detected during the inter-trial intervals (ITIs). Note a sharp decrease in mice speed at the onset of the events. Bottom: Heatmap of the individual immobility events detected during the inter ITIs. m, Top: Schematic depicting the strategy used to decode the immobility events from the pre-immobility periods. Bottom: Immobility decoding accuracies relative to the immobility onset (Blue) and shuffle decoding accuracies (grey).n, Top Left: Schematic depicting the strategy used to assess the generalization of the global-danger decoders to the immobility periods. Bottom Left: Danger to immobility mean generalization accuracies relative to the immobility and CS onsets (yellow) and shuffle decoding accuracies (grey). Note that the global-danger decoders do not generalize to immobility periods, suggesting that the shared patterns of activity across CS+ are not explained by the decrease in speed observed at CS onset. Bottom Right: Effect of shifting the time of immobility onset on the global-danger decoders obtained at different time points from the CS onset. Observe that mean decoding accuracies around chance levels ( 0.5) are obtained regardless of the time of the immobility event employed, showing that global-danger decoders do not generalize to any time point of the immobility periods. Shaded areas represent 95% confidence intervals for g, h, i, j, k, l and m. Horizontal lines indicate significant decoding accuracies (P < 0.05, two-sided permutation test. See Methods). *P < 0.05; Kruskal-Wallis one-way ANOVA on ranks.

Extended Data Fig. 4 dmPFC neuronal population analyses at CS offset and CS+ to CS minus generalization at CS onset.

a, Activity of neuronal population vectors (n = 546 neurons from 24 mice) at CS offset (from 250 to 0 ms before CS offset) sorted by trial type. b, Mean decoding accuracy matrix for comparisons between active, passive, inverse and CS trials at CS offset (from 250 to 0 ms before CS offset). Differences among threats are shown inside of the green square. c, Mean decoding accuracies around CS offset between CSs and Pre-CS basal activity (left), among CS+s (middle) and between CS+s and CS (right). Shuffle accuracies in grey. d, Median speed of recorded mice aligned to the CS offset. Note different locomotion states before the CS offset during active (red), passive (blue) and inverse (purple). e, Threat decoding (T; brown), Threat-general decoding (T-G; orange) and Threat-specific decoding (T-S; blue) average accuracies obtained around CSs offset. f, Mean generalization of the neuronal activity patterns elicited during the shock (US) to the onset of the CS+. A classifier is trained to decode dmPFC population activity around the offset of error trials from the Pre-CS basal activity. Then, the decoder is challenged to classify either the error trials offset (Shock decoding, yellow) or CS onset (brown). Note no generalization from dmPFC neuronal population activity patterns observed during US to CS+. Shuffle accuracies in grey. Shaded areas represent 95% confidence intervals for c, d and f. Horizontal lines indicate significant decoding accuracies (P < 0.05, two-sided permutation test. See Methods).

Extended Data Fig. 5 Encoding of threat-related information across training.

a, Heatmaps of z-scored neuronal activity evoked by different trials aligned to the CSs onset during habituation (Day 0; n = 374 neurons from 18 mice). b, Decoding accuracy matrix showing mean decoding accuracies for comparisons between active, passive, inverse and CS trials at CS onset (from 0 to 250 ms after the onset of the CS) during habituation. Differences among threats are shown inside of the green square (top). Mean decoding accuracies around CS offset between the different CSs and Pre-CS basal activity (middle) and among CS+s (bottom) during habituation. Shuffle accuracies in grey. c, Heatmaps of z-scored neuronal activity evoked by different trial types aligned to the CSs onset during the first training session (Day 1; n = 353 from 21 mice). d, Decoding accuracy matrix showing mean decoding accuracies for comparisons between active, passive, inverse and CS trials at CS onset (from 0 to 250 ms after the onset of the CS) during the first training day. Differences among threats are shown inside of the green square (top). Mean decoding accuracies around CS offset between different CSs and Pre-CS basal activity (middle) and among CS+s (bottom) during the first training day. Shuffle accuracies in grey. e, Global-danger mean decoding accuracies (orange) during habituation, first training day (day 1) and during the best performance day (Learned). Shuffle accuracies in grey. f, Threat-identity mean decoding accuracies (green) during habituation, first training day (day 1) and during the best performance day (Learned). Shuffle accuracies in grey. g, Threat (brown), threat-general (yellow) and threat-specific (blue) and CS generalization (dark grey) mean decoding accuracies during habituation, first training day (day 1) and during the best performance day (Learned). Shuffle accuracies in light grey. Shaded areas represent 95% confidence intervals. Horizontal lines indicate significant decoding accuracies (P < 0.05, two-sided permutation test. See Methods).

Extended Data Fig. 6 Decoding accuracies for activated and inhibited neuronal subpopulations.

a, Schematic of the strategy used to investigate decoding accuracies for activated and inhibited neuronal populations. b, CS onset decoding matrices obtained for neurons activated at the onset of active trials (left), passive trials (middle) and inverse trials (right). c, CS onset decoding matrices obtained for neurons inhibited at the onset of active trials (left), passive trials (middle) and inverse trials (right). Differences among threats are shown inside of the green square. d, Schematic of the strategy used to investigate the generalization of the decoding accuracies for activated and inhibited neuronal populations. e, Generalization of CS onset decoding accuracies of the decoders trained with the neurons activated at the onset of active trials (left), passive trials (middle) and inverse trials (right). The black square shows the result of the test performed with the classes used to train the classifier (threat decoding accuracy). f, Generalization of CS onset decoding accuracies of the decoders trained with the neurons inhibited at the onset of active trials (left), passive trials (middle) and inverse trials (right). The black square shows the result of the test performed with the classes used to train the classifier. g, Threat decoding accuracies obtained with activated, inhibited and non-modulated neurons at the onset of each threat type (n = 1000 decoding accuracies obtained with randomly selected trials and 10 randomly selected neurons). P values: excited vs non-modulated <0.001, excited vs inhibited <0.001 and inhibited vs non-modulated <00.1 h, Threat decoding Generalization index obtained with neuronal populations of activated, inhibited and non-modulated neurons at CS onset for each threat type (n = 1000 decoding accuracies obtained with randomly selected trials and 10 randomly selected neurons). The generalization is measured as the unspecific threat accuracies (obtained when decoding the threats not used to train the classifier from the Pre-CS basal activity) minus the Threat to Threat accuracies (obtained when decoding the threats not used to train the classifier from the threat used to train the classifier). P values: excited vs non-modulated <0.001, excited vs inhibited <0.001 and inhibited vs non-modulated=0.1 i, Schematic representation of the methodology employed to study threat decoding in subpopulations of activated and inhibited neurons. The same procedure described in Fig. 3d is followed but using a classifier trained with Activated, Inhibited or Non-modulated neurons and trials from the randomly selected threat type. j, Threat (brown), Threat-general (yellow) and Threat-specific (blue) mean decoding accuracies obtained using Activated, Inhibited and Non-modulated neurons. Shuffle accuracies in grey. k, Mean Threat-general and Threat-specific decoding accuracies of Activated, Inhibited and Non-modulated neurons (red, blue and grey, respectively). Small dots represent individual decoding accuracies obtained with different iterations. Shaded areas represent 95% confidence intervals. Box plot center line represents the median; box limits are 25th and 75th percentiles; whiskers are more extreme non-outliers data points; ‘+’ markers represent outliers. *p < 0.05; Two-way ANOVA with post hoc Tukey.

Extended Data Fig. 7 Projection of population vector in PC space.

a, Percentage of the variance explained by different principal components. b, Left: Euclidean distance at CS onset between population vectors projected into PC1-2 space for different trials (active, passive, inverse and CS) and the Pre-CS basal activity. Differences among threats are shown inside of the green square. Right: Average distances of the projection of population vectors (of n = 100 repetitions) into PC1-2 space between threatening (active, passive, inverse) and non-threatening conditions (Pre-CS basal and CS) in orange (Threat-Safe or TS) and among threatening conditions in green (Threat-Threat or TT). c, Left: Euclidean distance at CS onset between population vectors projected into PC3-4 space for different trials (active, passive, inverse and CS) and the Pre-CS basal activity. Differences among threats are shown inside of the green square. Right: Average distances of the projection of population vectors (of n = 100 repetitions) into PC3-4 space between threatening conditions (active, passive, inverse) and non-threatening conditions (Pre-CS basal and CS) in orange (TS) and among threatening conditions in green (TT). d, Threat General vs Specific Euclidean distance (difference of distance between threat and non-threatening conditions minus the distance among threats) in PC1-2 (light grey; n = 100 repetitions) and PC3-4 dimensions (dark grey; n = 100 repetitions). e, Top: Euclidean distance across time at the CS onset between population vectors projected into the PC dimensions for different trials (active, passive, inverse and CS) and the Pre-CS basal activity. Bottom: Euclidean distance across time at the CS onset among population vectors projected into the PC dimensions for different threatening trials (active, passive, and inverse). Shaded areas represent standard deviation. f, Dimensionality reduction and projection into the PC space of dmPFC population vectors lacking the unspecific responsive neurons (n = 36), the threat-general responsive neurons (n = 12), threat-specific neurons (n = 104) or threat-mixed responsive neurons (n = 131) at CS onset (from 0 to 250 ms from CS onset) selected as in Fig. 2. g, Projection of the non-normalized single-trials population vectors into the PC space. h, Comparison of the distribution of the PC1-2 and PC 3-4 combined weights across the dmPFC neuronal population (n = 546 neurons). i, Comparison of PC 1-2 and PC 3-4 combined weights among the non-modulated neurons (those that do not show an absolute z-score higher than 2 at the CS onset of any of the trial types; n = 137 neurons. Two-tailed paired t-test. P < 0.001) and the modulated neurons (showing an absolute z-score activity of 2 or higher at the CS onset of at least one trial type; n = 409 neurons. Wilcoxon signed rank test. P = 0.078). j, Combined weights of PC 1-2 and PC 3-4 in the different groups of responsive neurons (n = 35 unspecific neurons, 12 CS+-trial general neurons, 93 CS+-trial specific neurons and 107 mixed responsive neurons and 300 other neurons) obtained as in Fig. 2e. Box plot center line represents the median; box limits are 25th and 75th percentiles; whiskers are more extreme non-outliers data points; ‘+’ markers represent outliers. *P < 0.05; ***P < 0.001; two-tailed t-test in b, c, d, h, and i. Two-way ANOVA with post hoc Tukey in j.

Extended Data Fig. 8 Decoding accuracies obtained in the PC space and alternative decoding approaches.

a, Left: Threat (brown), Threat-General (yellow) and Threat-specific (blue) mean decoding accuracies obtained by projecting the dmPFC population vector in the PC 1 and PC 2 space. Right: as in Left but for the PC 3-4 space. b, Average decoding accuracies among CS+ and CS minus trials. The decoders were trained with correct trials and then tested with correct (green) or error (yellow; brown) trials. The decoders were performed in PC1-PC2 space (Left) and in the PC3-PC4 space (Right). c, Left: Average decoding accuracies among inverse correct and inverse error trials aligned to CS onset in the population vector projected into the PC 3-4 space. Right: Threat-identity generalization of inverse correct trials to inverse error trials in the PC 3-4 space. Mean decoding accuracies aligned to CS onset. c, Average decoding accuracy matrix for comparisons between Active, Passive, Inverse and CS trials at CS onset (from 0 to 250 ms after the onset of the CS) in PC 1-2 space. Differences among threats are shown inside of the green square. d, Average decoding accuracies among CS+ and CSaligned to CS onset. e, Global-danger and Threat-identity (as described in Extended Data Fig. 3) mean decoding accuracies in PC 1-2 space. f, Average Pre-CS basal activity vs. CS+ decoding accuracies (red) and average decoding accuracies among CS+s (green). g, As c but for PC3-4 space. h, As d but for PC3-4 space. i, as e but for PC3-4 space. j, As f but for PC3-4 space. k, As c but for PC 1-4 space. l, Top: CS vs. Pre-CS basal activity mean decoding accuracies aligned to CS onset for population vectors in PC1-4 space. Bottom: Average decoding accuracies among CS+s aligned to CS onset for population vectors in the PC1-4 space. m as in j but for the PC1-4. For f, j and m n = 100 decoder iterations and box plots center lines represents the median; box limits are 25th and 75th percentiles; whiskers are more extreme non-outliers data points; ‘+’ markers represent outliers. Shaded areas represent 95% confidence intervals. *P < 0.05; ***P < 0.001; one-way ANOVA with post hoc Tukey. Horizontal lines indicate significant decoding accuracies (P < 0.05, two-sided permutation test, see Methods).

Extended Data Fig. 9 Optogenetic manipulation of dmPFC neuronal activity.

a, Schemes showing the location of the fiber placements for the ArchT (top) and GFP control groups (bottom). b, Two representative neurons of the ArchT group inhibited during laser activation across trials (top) and for individual trials (bottom). c, Z-scored neuronal activity heatmap for a representative ArchT mouse upon laser activation. d, Two representative neurons of the GFP-control group non affected by laser activation across trials (top) and for individual trials (bottom). e, Z-scored neuronal activity heatmap of a representative GFP-control mouse upon laser activation. f, Median speed aligned to CS onset in mice expressing ArchT during Light ON trials (green) and Light OFF trials (grey) across all CS+ trials. g, Median speed relative to CS onset in mice expressing ArchT during light ON trials (green) and light OFF trials (grey) across all CS trials. h, Median speed aligned to CS onset in mice expressing ArchT during Light ON trials (green) and Light OFF trial (grey) in active, passive and inverse correct trials. i, Effect of optogenetic inhibition of the dmPFC cortex in the preference for shuttling behavior when mice are in Shelter but facing different tones (In shelter) and when facing the CS+ 1 but in different contexts (To CS+ 1) for n = 11 mice. j, Shuttling in response to the CS for mice expressing ArchT during Light ON and Light OFF trials when mice are in the arena or in the shelter for n = 11 mice. k, Effect of light stimulation in n = 8 mice expressing GFP in the MIO paradigm (Global performance: orange) and in active (red) passive (blue) and inverse (magenta) trials. Circles represent average mice performance and lines individual mice behavioral performance during Light OFF (white) and Light ON (yellow) conditions. l, Effect of light stimulation in mice (n = 8) expressing GFP on the selection of the correct defensive response when discriminating among tones (Tone, green) or contexts (Context, blue). Circles represent average across mice and lines individual mice in Light OFF (white) and Light ON (yellow) conditions. Shaded areas correspond to 95% confidence intervals. *P < 0.05 **P < 0.01, ***P < 0.001; Two-way ANOVA with Tukey post hoc test or two-tailed paired t-test.

Extended Data Fig. 10 Appetitive-aversive behavioral paradigm.

a, Performance levels of mice (n = 4) trained to perform avoidance and approach responses to a CS+ (7 kHz, red) and to CSr (13 kHz, blue) respectively. A CS(white noise, grey) was used as a control. Avoidance p = 0.013 and approach p = 0.036. b, Mean decoding accuracies obtained between the aversive CS+ and preCS activity. c, Mean decoding accuracies obtained between the reward-guided cue CSr and preCS activity. d, Mean decoding accuracies obtained between the aversive CS+ and the reward-guided CSr. e, We performed the same dimensionality reduction method as in Fig. 4 and observed a clear separation between aversive and reward-guided behaviors in the first principal components. Large circles represent population vectors of dmPFC mean activity evoked by the different trials types. Small circles represent resampled population vectors using 25% of the trials. Shaded areas represent 95% confidence intervals. Horizontal lines indicate significant decoding accuracies (P < 0.05, two-sided permutation test. See Methods). *P < 0.05; Two-tailed paired t-test or two-tailed paired t-test.

Supplementary information

Reporting Summary

Supplementary Table 1

Table with description of statistics.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Martin-Fernandez, M., Menegolla, A.P., Lopez-Fernandez, G. et al. Prefrontal circuits encode both general danger and specific threat representations. Nat Neurosci 26, 2147–2157 (2023). https://doi.org/10.1038/s41593-023-01472-8

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s41593-023-01472-8

Search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing