Article Figures & Data
Figures
- Figure 1.
Circuit model. The olfactory pathway comprises three feedforward layers. Olfactory input activates an odor-specific combination of PNs. The connectivity matrix from PNs to the large population of KCs is divergent-convergent and sparse. The inhibitory feedback mechanism ensures population sparseness of 5% active cells in the KC layer. KCs are fully connected with four MBONs in the output layer. A positive or negative reinforcement stimulus directly excites the dopaminergic PAM or PPL1 neuron, respectively. Lateral inhibition between the MBONs and excitatory feedback from MBONs to DANs is crucial for the reward prediction and extinction mechanisms. Activation of the MVP2 output neuron mediates approach behavior, activation of MV2 mediates avoidance behavior. Behavioral preference toward an odor is calculated as the imbalance between the activations of MVP2 and MV2.
- Figure 2.
Differential conditioning leads to CS+-specific odor preference. A, Experimental protocols for simulation of associative conditioning experiments. The model was trained in a classical conditioning paradigm, in which a conditioned odor stimulus (CS+) was paired with a reward (green) or punishment (red). Subsequently, a second unconditioned odor stimulus (CS–) was presented without reinforcement. The standard protocol comprises 12 training trials. During the retention test, both odor stimuli were presented alone without reinforcement. Preference index and performance index in the model were computed after each trial and during the retention test. B, Dynamics of the odor preference index across 24 appetitive (top) and aversive (bottom) learning trials, respectively averaged across 10 networks. C, left, Combinatorial response patterns in the PN population. Each odor stimulus activates 50% of all PNs. Similarities between the CS+ reference odor and five novel odors are defined by their overlap in the PN activation pattern (x-axis). Right, Activation patterns across the subpopulation of 100 KCs (5%) activated by the CS+ odor. Pattern overlap in the KC population reduces rapidly with decreasing odor similarity as expressed in the percentage of PN pattern overlap (x-axis). D, Generalization to different odors after associative conditioning to the CS+ odor. The model shows a significant CS+ approach (green) or avoidance (red) after 12 training trials as expressed in the preference index (n = 15), which diminishes rapidly with decreasing odor similarity. Boxplots show the median and the lower and upper quartiles, whiskers indicate 1.5 times interquartile range, outliers are marked with + symbol.
- Figure 3.
CS+ odor reactivation leads to memory extinction. A, Training and test protocol as in Figure 2A. In the extinction protocol, the training phase is followed by an odor reactivation phase, in which the CS+ was presented alone for 12 consecutive trials before performing the retention test. B, The appetitive conditioning protocol (light gray) induced a strong and significant preference for the CS+ odor and a weak and significant preference for the CS– odor. Extinction learning partly abolished the CS+ preference during the retention test (dark gray). The preference index for the CS+ odor was significantly lower than after initial appetitive conditioning, for the CS– odor no more significant preference was observed after extinction learning. C, Symmetric results were observed for aversive conditioning. Again, extinction learning significantly reduced the learned preference for the CS+ odor. Boxplots show the median and the lower and upper quartiles, whiskers indicate 1.5 times interquartile range, outliers are marked with + symbol. D, E, The performance indices were significantly reduced after extinction compared with the initial appetitive and aversive memories. Data are presented mean ± SD; ***p < 0.001; n = 15 independent models. For preference and performance indices of the model tuned toward asymmetric appetitive and aversive learning pathways, see Extended Data Figure 3-2.
- Figure 4.
Neuron activation rates and behavioral performance across trials. A, Reward value (US, top) and DAN activation rates (bottom) across 12 initial conditioning trials (left) and 12 extinction trials (right). Trial zero indicates neuron activation rates in the naive model before learning. Reward presentation resulted in a prominent peak activation of the PAM during the very first training trial, i.e., when the actual reward strongly deviates from the predicted reward. Omission of the reward during the first extinction trial (right) resulted in a step-like reduction of PAM activation the simultaneous increase in PPL1 activation, which is crucial for extinction learning. B, Summed synaptic KC input to MBONs (top), MBON activation rates (middle), and model preference index (bottom) for each simulated test trial that followed the respective training trial (see Materials and Methods). The strong PAM activation during the first conditioning trial resulted in a strong change of the synaptic weights between KCs and MV2 and M6, and consequently in a step-like reduction of the synaptic input and the output activation rates. Consequently, the model preference index represents the imbalance between M6 and MV2 activation rates showed a step-like increase predicting a switch-like expression of a CR behavior after only a single conditioning trial. Additional training led to a further gradual reduction of gross synaptic input to and activation of MV2 and M6, paralleled by the gradual increase of the preference index. Extinction learning led to a gradual reduction of synaptic weights in the KC::MVP2 and KC::V2 pathway. This reduces the difference between M6 and MV2 activation and leads to a gradual extinction of the preference index. C, D, Same as A, B for aversive conditioning (left) and subsequent extinction (right). Simulation results are shown for a single network. This was initiated identically before appetitive (A, B) and aversive (C, D) conditioning to enforce identical initial conditions stressing the symmetric mechanism of reward prediction.
- Figure 5.
Two separate memories underlie initial and extinction learning. CS+ odor-induced activation rates of MBONs were measured before and after extinction training to locate associative and extinction memories. A, B, Appetitive learning led to a relative depression of CS+ odor-induced activation rates of MV2 and M6, but not MVP2 and V2 MBONs. The memory trace in MV2/M6 stayed intact even after appetitive memory was extinguished. Extinction decreased the CS+ response of MVP2/V2, establishing a parallel memory trace. C, D, Aversive conditioning led to a reduced CS+ input into MVP2/V2, but not into MV2/M6. After memory extinction, the memory trace in MVP2/V2 remained, and we observed an additional decrease in CS+ response in MV2/M6. Boxplots show the median and the lower and upper quartiles, whiskers indicate 1.5 times interquartile range, outliers are marked with + symbol; n.s. = not significant, **p < 0.01. Results across simulation of n = 10 networks in all panels.
- Figure 6.
Blocking specific neuron groups can reverse memory extinction. The computational model was used to reproduce the neurogenetic manipulations with temperature-sensitive shibirets1 mutant flies. The activation of individual neurons was suppressed during extinction training only. This allows to identify neurons that are crucial to the process of establishing the extinction memory in our model. A, Knocking down PPL1 or V2 during extinction training led to a completely abolished the appetitive extinction memory. Blocking PAM, M6, MV2, or MVP2 neurons during odor re-exposure had no effect on the appetitive extinction memory. Blockage of all KCs (100%) prevented the model from extinction learning. However, when only 50% of the KC population was blocked during odor reactivation, extinction learning remained unaffected. Further effects of partial KC blocking on extinction learning are presented in Extended Data Figure 6-1. B, Suppressing the activity of PAM DANs, M6 neurons, or KCs (100%) during CS+ odor re-exposure after aversive conditioning prevents the model from forming the aversive extinction memory. However, blocking PPL1 DANs, MV2, V2, MVP2, or KCs (50%) does not diminish the memory extinction. Data are presented as mean ± SD; n.s. = not significant, **p < 0.01, ***p < 0.001; n = 15.
Extended Data
Extended Data Figure 3-1
Comparison between experimental and model results. The experimental data on appetitive and aversive conditioning is extracted from Felsenberg et al. (2017) and Felsenberg et al. (2018), respectively. For statistical analysis of the simulated data see Methods. Experimental values are shown as mean ± SEM, model values are shown as mean ±STD. *The behavioral data include two control groups, the first carried the GAL4 gene only, the second the UAS gene only. The test group carrying both, the GAL4 and the UAS gene, showed cell-specific expression of the temperature-sensitive UAS-Shibirets1 transgene Download Figure 3-1, EPS file.
Extended Data Figure 3-2
Independent tuning of the model parameters for the appetitive and aversive pathways allows to match the experimental results more accurately. In order to fit the model to the asymmetry in the experimental results, the PPL1 output was tuned towards a lower activation threshold with the sigmoid transfer function
Extended Data Figure 6-1
The percentage of blocked KCs during CS+ odor reactivation determines the effect on extinction memory formation. When all KCs are blocked during appetitive or aversive extinction training, no extinction memory is formed. However, reducing the percentage of blocked KCs gradually increases the ability to establish extinction memory. When 50% or less KCs are blocked, memory extinction formation is not impaired. Data is presented as the mean ± SD; n = 15. Download Figure 6-1, EPS file.
In this issue
