Evaluation of ambiguous associations in the amygdala by learning the structure of the environment

Abstract

Recognizing predictive relationships is critical for survival, but an understanding of the underlying neural mechanisms remains elusive. In particular, it is unclear how the brain distinguishes predictive relationships from spurious ones when evidence about a relationship is ambiguous, or how it computes predictions given such uncertainty. To better understand this process, we introduced ambiguity into an associative learning task by presenting aversive outcomes both in the presence and in the absence of a predictive cue. Electrophysiological and optogenetic approaches revealed that amygdala neurons directly regulated and tracked the effects of ambiguity on learning. Contrary to established accounts of associative learning, however, interference from competing associations was not required to assess an ambiguous cue-outcome contingency. Instead, animals' behavior was explained by a normative account that evaluates different models of the environment's statistical structure. These findings suggest an alternative view of amygdala circuits in resolving ambiguity during aversive learning.

Figure 1

Reduced CS-US contingency results in reduced CS memory, irrespective of trial order and with or without changes in context memory. (a) Experimental design. Animals underwent threat conditioning on day 1 and were tested for contextual and tone (or tone and contextual) memories 24 hours later. (b) Conditioning protocols. Rats received sequences of tone-shock pairings (P) and unsignaled shocks (U), or tone-shock pairings only. The boxes indicate time spent in the conditioning chamber, ~7 min for CTL II group and ~31 min for all other groups. (c) A 20% CS-US contingency during training leads to significantly lower CS-induced freezing compared to 100% contingency, whether unpaired shocks are given intermixed with, or after with tone-shock pairings (n=22, 22, 17, 18, 2-way ANOVA, no significant interaction F_1,75 = 1.63, p=0.21, main effect for contingency F_1,75 = 18.02, * p=0.00006, simple effects for contingency F_1,75 = 15.0, p=0.0002, F_1,75 = 4.48, p=0.038 for spaced and massed condition respectively)‥ (d) Context memory strengths for the same animals. Reduction of CS memory with degraded CS-US contingency is not explained by changes in context memory strengths, as there was no difference between context memories of Control II and Pairings First groups (2-way ANOVA, significant interaction F_1,75 = 6.44, p=0.013, simple effect for contingency F_1,75 = 0.00008, p=0.81, not significant). Error bars indicate s.e.m.

Figure 2

Conditioning to the Context is not required for contingency degradation. (a) Experimental design depicting pharmacological inactivation of NMDA receptors in DH before conditioning. (b) Hippocampal APV injections have no effect on learning the reduced auditory CS-US contingency (n=8, 11, 10, 7, 2-way ANOVA, no significant interaction F_1,32 = 0.07, p=0.79, main effect for contingency F_1,32 = 11.98, p=0.0015, simple effect for contingency F_1,32 = 8.60, * p= 0.011, F_1,32 = 5.45, ** p= 0.026 for Vehicle and APV groups respectively). (c) NMDA receptor blockade impairs the acquisition of contextual aversive memories (2-way ANOVA, no significant interaction, F_1,32 = 0.38, p= 0.54, main effect for drug treatment, F_1,32 = 9.47, * p=0.0043). d) Similarly to the Pairings First case, contingency degradation to the auditory stimulus is unaffected in the Intermixed condition by APV infusion in DH (n=9, 9, unpaired sample t-test, t₁₆ = 2.14, * p=0.048). (e) Impaired contextual aversive memory formation after NMDA receptor blockade in the Intermixed condition (n=7, 9, unpaired sample t-test, t₁₄ = 2.31, * p=0.037). Error bars indicate s.e.m.

Figure 3

Activation of LA pyramidal cells during unsignaled USs is required to learn the degraded CS-US contingency. (a) Optogenetic inhibition of shock evoked firing rate (Hz) responses in single LA neurons. Perievent time rasters (top) and histograms (bottom) show shock evoked (red bars) responses in 2 example neurons without (left) and with optical inhibition (right, laser on denoted by green bar). (b) Design of optogenetic behavioral experiments. (top) Graphical depiction of lentivirus injection and example of Arch-T expression in LA pyramidal neurons (scale bar = 160 um). Virus expression within LA was verified for all experimental animals included in study. Experimental protocols with laser illumination either coinciding, or offset from UUSs (bottom). 3 tone-shock US pairings (P) were presented either intermixed with, or prior to 12 unsignaled USs (U). (c) Inactivation of LA pyramidal neurons during, but not offset from UUSs prevents the learning of the degraded auditory CS-US contingency (n=7, 8, 8, 10, 2-way ANOVA, no significant interaction, F_1,29 = 0.28, p=0.60, main effect for inactivation, F_1,29 = 11.41, p=0.0021, simple effects for inactivation F_1,29 = 7.02, * p=0.013,, F_1,29 = 4.45, ** p=0.044 for Pairings First and Intermixed groups respectively) (d) Context memory strength was unaffected by optogenetic manipulation (2-way ANOVA, significant interaction, F_1,29 = 4.39, p=0.045, main effect for inactivation F_1,29 = 0.0297, p=0.86, not significant, simple effects for inactivation F_1,29 = 2.37, p=0.13, F_1,29 = 2.03, p=0.17, not significant). Error bars indicate s.e.m.

Figure 4

Degraded CS-US contingency leads to reduced enhancement of CS processing in the LA. (a) Experimental design for the in-vivo physiology experiment. (b) Population-averaged A-LFP traces before (HAB), and after (LTM) conditioning in the Control II group (left) and Pairings First group (right). Vertical line at t=0 indicates CS onset, red arrow indicates peak depolarization. (c) Population averaged post training A-LFP % pre-training baseline. Reduced CS-US contingency leads to reduced potentiation of auditory CS processing (n=10, 8, unpaired t-test, t_11,1 = 2.54, * p=0.028). (d) Animals in the in vivo physiology experiment also showed reduced CS memory in the reduced contingency (Pairings First) condition (unpaired t-test, t_11,0 = 3.07, * p=0.011). Error bars indicate s.e.m.

Figure 5

Comparison of the fit of the structure learning model (SLM) and behavioral data. (a) The six different graphical representations of statistical associations in the environment that are compared in SLM. (b) Context and tone CS memory strengths (Behavioral data, top and SLM, bottom). Data points are % freezing at long term memory, after different conditioning protocols. Context memory: (I), 2, 3, 6, 10 or 15 unsignaled USs (UUSs). (II), 2 or 3 UUS followed by no or 1 unreinforced ITI in the conditioning chamber (III), 2, 3, 6, 10 or 15 CS-US pairings. (IV), 3 CS-US pairings followed by 0, 3, 6, 9, or 12 UUSs. Intermixed condition, and Pairings Last condition (3 CS-US pairings + 12 UUS). Tone memory: (V), Same groups as in III. (VI), Same groups as in IV. (SLM, bottom) Memory strength as predicted by SLM. Values are averages of 20 runs using the best-fit parameters. Model was fitted to data in I-VI. Best-fit parameters were then used to predict the effects of neural interventions (see Online Methods and Supplementary Modeling) and freezing scores for CTL I and Cover stimulus groups.