Fig 1.
Nodes of the graph correspond to modules. Dotted/solid arrows denote the existence of plastic/non-plastic synapses between modules. During Perception, environmental inputs are passed to the ‘engram modules’ HIP, CTX and BAN, activating patterns that may be remembered via Hebbian plasticity, so that the model’s state may converge onto them during later Recall. To allow either fear or safety to be associated with a context, synapses from HIP onto BAN are strengthened when a prediction error occurs. Prediction errors are defined as the difference between the current US input and fear response (CeA output). Errors promote Hebbian plasticity within BAN and on HIP → BAN synapses, allowing the context to be associated with valence. Positive / negative errors further strengthen synapses from BAN onto BAP / BAI, increasing / decreasing the amount of fear associated with the current environment. Nodes labelled with a sigmoid are single units. All other modules are sets of neurons – the BCPNN (Bayesian Confidence Propagation Neural Network), kWTA (k-Winner-Takes-All) and ‘binary’ module types are defined in S1 Appendix. Abbreviations: SC = Sensory Cortex. EC(in / out) = Entorhinal Cortex (Input- / Output layers). HIP = Hippocampus. CTX = Neocortex. BA(N / P / I) = Basal Amygdala (valence-Neutral / fear-Promoting / fear-Inhibiting sets of neurons). CeA = Central Nucleus of Amygdala.
Fig 2.
Dynamics of engram formation and replay in our model.
During Perception (left), groups of engram cells in HIP, CTX and BAN are activated by environmental inputs arriving from SC. Fast Hebbian plasticity (thick, dotted lines) strengthens the HIP and BAN engrams and forms excitatory connections from the HIP onto the CTX and BAN ensembles. Plasticity in (and extending from) CTX is slow (thin, dotted lines), not yet supporting later Recall. During Sleep, HIP independently replays stored engrams, co-activating its CTX and BAN counterparts. This allows the CTX engram to consolidate and to form connections onto the associated BAN engram. In this way, sleep replay drives the long-term retention of fear (and safety) memories in our model.
Fig 3.
Fear acquisition and synaptic homeostasis.
a) Fear acquisition: When a surprising US occurs, active (valence-Neutral) BAN engram cells, encoding the current context, become associated with fear. Left: Over the course of several time steps, synapses from the BAN engram onto fear-Promoting BAP cells are strengthened by various amounts (denoted by the different line weights). Right: The extent to which synapses onto certain P-cells are strengthened is mediated by those cells’ current recruitability (S2 Fig). In this example, the time window of US delivery (time step ) is shaded in red. P-cells 3 and (especially) 4, but not 1, are recruited – i.e., become associated with the current context – as the US signals coincide with their time windows of high recruitability. Synapses onto P-cell 2 are strengthened to some, limited extent as random noise has raised its recruitability for a single time step during conditioning. b) Cubic growth model: The differential equation describing homeostatic changes applied to the strength of our model’s valence-coding synapses during Sleep,
with (exaggerated) learning rate r = 1.4, extinction threshold A = 0.9 and recruitment strength K = 1.7. For
, the strength stably converges to x = 0 or x = K, depending on whether its starting value x(0) lies above or below A. This rule gives rise to the synaptic changes from a) (Left) to c). c) Fear recall after sleep: When the conditioned context is encountered after a Sleep phase, synaptic homeostasis has acted on the synapses strengthened in a). Weak synapses have been pruned, whereas stronger synapses have been normalized towards an appropriate strength. See also Fig 7 and its discussion in later results.
Fig 4.
Mechanisms of context/fear memory formation.
a) Engram formation: Encoding strength (net synaptic weight between active engram cells, normalized by their total outgoing weights) of a novel context, presented for 50 time steps, in HIP, CTX, and BAN. US inputs occurred on the last 30 time steps. HIP rapidly formed context engrams; BAN did so after the first US; CTX learned slowly. Results were averaged over 10 simulation runs. Shaded regions denote standard deviation (SD) across runs. b) Sleep replay: Distance between original context patterns and neural activity during Sleep in HIP, CTX, and BAN following sequential exposure to 10 novel contexts. In contexts 6 to 10, US signals were delivered. Blue indicates replay of stored engrams. All contexts were replayed in HIP and CTX; BAN replay was limited to contexts paired with aversive stimuli. The distance measure was borrowed from Greve et al. [68]. c) Recall performance: Percentage of contexts successfully recalled by HIP, CTX, and BAN as a function of days since initial encoding. HIP reliably retrieved recently, but not remotely, perceived contexts. CTX was most likely to recall contexts encoded several days ago, reflecting sleep-dependent consolidation. Compared to HIP, CTX featured a recall curve with a lower peak, but flatter drop-off, indicating a more selective, more durable, storage strategy. BAN recalled contexts when supported by either HIP or CTX. Curves were averaged over 100 runs, with 4 contexts assessed per run and day. Shaded regions denote
SD on the percentage of recalled contexts across runs. d) Acquisition and extinction: Fear acquisition and extinction dynamics. In each of 30 simulation runs, a context was presented for 70 steps, with US inputs delivered only during the first 20. The model acquired fear quickly (via BAP activation), which was slowly extinguished following removal of the US through gradual activation of BAI cells. BAN → BAI synapses began growing in strength as soon as the US signals were removed, but fear did not decrease until – after a delay – the first BAI cells became active. Shaded regions denote
SD across runs.
Fig 5.
Results on fear generalisation.
a) Generalisation schema. Fear conditioning in context ‘A’ strengthens synapses between its engram and
cells, activating the latter. In a similar context ‘B’, some of these BAN cells are active, exciting the same
cells. Whether these become active, causing fear to generalize, depends on the margin by which their activation threshold was exceeded in context ‘A’ and on the overlap of the two
engrams. b) Generalisation gradients. After fear acquisition and extinction in context ‘A’, red dots denotes the fraction of active BAP-cells (top), of active BAI-cells (center) or the magnitude of the CeA output (bottom) when the model meets a context whose feature overlap with ‘A’ is shown on the x-axis. Plots were averaged over 10 runs. Fear renewal is likeliest to occur when the model is placed in a context moderately similar to ‘A’. c) Renewal paradigms. From top to bottom, the figure shows the activity of our model’s CeA cell when subjected to the ABA, ABC and AAB fear renewal paradigms, as described in the main text. Plots were averaged over 10 runs. d) Fear generalisation increases over time. After fear conditioning in context ‘A’, fear expressed in a similar, unconditioned context ‘B’ increases as days pass. In a dissimilar context ‘C’, no fear is expressed, showing that the increase in ‘B’ is associative. The dotted ‘Recall score’ (F1 measure) quantifies how accurately the HIP engram retrieved when observing context ‘B’ in Recall encodes its true input features (S1 Fig) – it is inversely proportional to fear expression in ‘B’, indicating the generalisation increase is tied to use of the CTX → BAN pathway for Recall. Curves were averaged over 50 runs. In panels b, c and d, shaded regions denote
SD across runs.
Fig 6.
ABC Renewal with extinction in multiple contexts.
a) Multi-context extinction suppresses renewal of conditioned fear. Each panel shows the average CeA output across acquisition (context A, blue), extinction (B, red), and renewal (C, green) phases. Top: Single-context extinction protocol () with 80% input overlap between A and B, as well as between A and C. Bottom: Multi-context extinction (
), where each Bi, as well as C, shares 80% overlap with A. Shaded regions show ±SD across runs (n = 10). In the multi-context condition, fear expression (CeA activity) at the onset of testing in C is significantly lower than after single-context extinction, suggesting that distributing extinction across several similar but non-identical contexts can reduce fear renewal. b) Fear renewal in C as a function of
similarity. We performed the multi-context renewal paradigm from subplot a), varying the input overlaps
and
. Different lines correspond to different values of
; points represent means across (n = 50) runs; shaded bands denote
. The rightmost points (
) correspond to extinction in A. Generally, distributing extinction across multiple contexts lowers fear renewal in C. For renewal contexts very similar to A, using a set of Bi that share a large overlap with A is most effective. However, as the similarity between A and C decreases, introducing greater differences between A and the Bi becomes beneficial, since additional variation in the Bi allows extinction to generalize more broadly. In the lowest similarity condition (
), renewal is moderate, but unaffected by extinction training, as the Bi are too dissimilar from C for extinction to generalize.
Fig 7.
Histograms of the summed strength of afferent synapses from active BAN units over all P-cells.
The red lines symbolize the least amount of excitation a P-cell must receive to become active. Histograms were computed: a) at the end of fear acquisition in context ‘A’. BAP cells falling on the right of the red line are active, sparking fear. b) upon being immediately placed in an unconditioned, vaguely similar context ‘B’. The BAP cells that were most strongly innervated in ‘A’ are active – expressing some generalized fear. c) upon revisiting context ‘A’ after a Sleep phase. The amount of fear expressed has not changed relative to a) but – thanks to homeostatic synaptic adjustments during Sleep – no BAP cell receives much more excitation than needed to be activated. d) upon revisiting context ‘B’ after a Sleep phase. Fear is no longer expressed in this unconditioned context. A selective normalization / partial reversal of recently strengthened fear-coding synapses in amygdala circuits during Sleep is thus posited to underlie decreases in the generalisation of freshly learned fear associations [83].
Fig 8.
Effects of chronic sleep deprivation Instances of the model were subjected to 7 days filled with exposure to various contexts, joined by US signals of random, generally low, strength.
At the end of the simulation, the model’s fear response was assessed, after the brief exposure (3 time steps) to a moderate US signal (strength 0.6) in a novel environment. a) Instances of the model whose daily Sleep duration was systematically decreased (by a fraction of or 100%) acquire a greater fear response in the novel environment, on average. b) Sleep deprivation is further accompanied by net increases in the strength of synapses from context- onto fear-coding BA cells, measured at the end of the simulation protocol. While increases in fear acquisition only occurr after relatively drastic decreases in Sleep length, the average
synaptic density increases more linearly. In both plots, dots denote the mean, shaded regions one standard deviation across 30 runs of the simulation.
Fig 9.
a) Stress-Enhanced Fear Learning (SEFL). Violin plots describe the magnitude of the model’s fear response to context ‘B’, following the SEFL protocol as described in the main text, across 30 runs – with exposure to ‘traumatic stress’ on the one hand and without (‘control’) on the other. The trauma group acquires stronger fear responses to context ‘B’ (3 time steps, US strength 0.6), showing fear sensitization. The plotted density estimates were cut at the empirical maximum/minimum across the 30 runs. b) SEFL does not occur if fear learning precedes stress. Same as a), but with the order of ‘traumatic stressor’ and ‘moderate US’ exposure reversed. Trauma exposure does not retrospectively enhance fear of context ‘B’. c) Stress enhances fear learning weeks later. Same as a), but with ‘moderate fear acquisition’ in context B carried out 15 Perception-Sleep cycles after traumatic stress. During the delay, the model meets various contexts (each paired with a US signal of generally low strength ). Both groups show increased fear acquisition after the delay, due to fear generalisation; the effect is stronger the trauma group. d) Stress raises the net strength of
synapses. Blue and orange lines show the total change in the summed strength of
synapses of the ‘trauma’ and ‘control’ model instances, since the time of trauma exposure. Red vertical lines denote, from left to right, delivery of the traumatic stressor, the approximate time by which the ‘trauma’ model’s parameters have recovered from stress, and delivery of the ‘moderate US’. Values were recorded at the end of each Perception and Sleep phase; both curves oscillate since Sleep promotes a synaptic weakening [103]. On average, the
synaptic density of the trauma model shows a drastic increase on the day of the trauma, continues increasing over the following Perception-Sleep cycles, and remains far above that of the control model for the remainder of the simulation. Shaded regions denote
SD across runs; variability results, e.g., from the random sampling of US signals over the course of the simulation.