Top-down inputs drive neuronal network rewiring and context-enhanced sensory processing in olfaction

Abstract

Much of the computational power of the mammalian brain arises from its extensive top-down projections. To enable neuron-specific information processing these projections have to be precisely targeted. How such a specific connectivity emerges and what functions it supports is still poorly understood. We addressed these questions in silico in the context of the profound structural plasticity of the olfactory system. At the core of this plasticity are the granule cells of the olfactory bulb, which integrate bottom-up sensory inputs and top-down inputs delivered by vast top-down projections from cortical and other brain areas. We developed a biophysically supported computational model for the rewiring of the top-down projections and the intra-bulbar network via adult neurogenesis. The model captures various previous physiological and behavioral observations and makes specific predictions for the cortico-bulbar network connectivity that is learned by odor exposure and environmental contexts. Specifically, it predicts that-after learning-the granule-cell receptive fields with respect to sensory and with respect to cortical inputs are highly correlated. This enables cortical cells that respond to a learned odor to enact disynaptic inhibitory control specifically of bulbar principal cells that respond to that odor. For this the reciprocal nature of the granule cell synapses with the principal cells is essential. Functionally, the model predicts context-enhanced stimulus discrimination in cluttered environments ('olfactory cocktail parties') and the ability of the system to adapt to its tasks by rapidly switching between different odor-processing modes. These predictions are experimentally testable. At the same time they provide guidance for future experiments aimed at unraveling the cortico-bulbar connectivity.

Fig 1. Computational model.

(A) Network structure emerging after learning 2 training stimuli. The modeled neuronal populations are indicated by circles marked OSNs (sensory neurons), MCs, GCs, and CCs. Excitatory connections in green with arrows, inhibitory ones in red with squares. Connectivity matrices indicate the connectivities between the individual neurons of the populations (black = connection, white = no connection; cf. Fig 2). For the intra-cortical connectivity W^(CC) the colors indicate the synaptic strength of the connections. (B) The GCs mediated disynaptic mutual inhibition of MCs with effective connectivity matrix W^(MM) and disynaptic inhibition of MCs by CCs with effective connectivity matrix W^(MC). The color indicates the number of GCs contributing to the respective connections. (C) Idealized schematic of the network structure emerging after learning two odors: CCs disynaptically inhibit predominantly those MCs that respond to the same odor as the CCs. Line thickness indicates the number of connections.

Fig 2. Model of adult neurogenesis.

In each time step adult neurogenesis added GCs, which made random reciprocal synapses with MCs: the synapses excited the GCs and inhibited the MCs that they connected (left panel). In addition, the GCs received excitatory projections from random CCs. The synapses between the individual MCs and GCs are indicated by black rectangles in the connectivity matrix (middle panel). GCs were removed reflecting a survival probability that depended on GC-activity via a resilience that was summed over the training stimuli (right panel).

Fig 3. Extinction of cortical odor memory eliminates GCs.

(A) Training and probe odors. (B) Extinguishing the cortical memory of odors 1 and 2 was implemented by removing the associative connections between the respective CCs. (C) Extinction induced the apoptosis of many GCs. (D) In each time step we measured the average of the odor response across the GCs that were to be removed in that time step. On average, the GCs that were removed as a result of the memory extinction had strongly responded to odor 1 before the extinction, but not to odor 3. (E) The extinction-induced reduction in inhibition compromised the discrimination between the ‘forgotten’ odors 1 and 2, but not between odors 3 and 4.

Fig 4. Context induces specific GC activity.

(A,B) The 2 training stimuli were associated with contexts 1 and 2, respectively. In (B) CCs with index up to 110 received odor information from MCs. CCs with index above 110 received contextual information. (C) Upper panel: odor-evoked MC- and GC-activities (odor 1: black lines, odor 2: red lines). Lower panel: context-evoked MC- and GC-activities (context 1: black lines, context 2: red lines). (D) Context- and odor-evoked MC activities were anti-correlated, while context- and odor-evoked GC activities were highly correlated (red boxes in (D)).

Fig 5. Context-enhanced odor processing in cluttered environments: Occluding stimulus.

(A,D) Training stimuli 1,2 and 3,4 were associated with contexts 1 and 2, respectively. (B,C) Learned disynaptic inhibitory connections among MCs (W^(MM)) and from CCs to MCs (W^(MC)). The color indicates the number of GC mediating the respective connections. (E) Probe stimuli consisted of a weak target odor (red line) with or without stimulus 1 as a strong occluding odor (black line). (F,G) MC activities resulting from the probe stimuli with and without context 1. Red: target with occluder, black: occluder alone. Context 1 reduced the response to the occluder, enhancing the detectability of the target. (H) Context 1 increased the Fisher discriminant ${\mathcal{F}}_{opt}$ when the occluding odor was present, but reduced ${\mathcal{F}}_{opt}$ when it was absent. With top-down input blocked (w_GC = 0) the occluder is suppressed even less than without any context, deteriorating the performance further. The error bars indicate the standard deviation in ${\mathcal{F}}_{opt}$ resulting from fluctuations in the network connectivity.

Fig 6. Context-enhanced odor processing in cluttered environments: Distracting stimulus.

(A,D) Training stimuli 1,2 and 3,4 were associated with contexts 1 and 2, respectively. (B,C) learned disynaptic inhibitory connections among MCs (W^(MM)) and from CCs to MCs (W^(MC)). Colors indicate the number of GCs mediating that inhibition. (E) The probe stimuli (right panel) consisted of a pair of similar, weaker target odors (open black and solid red symbols, left panel) on top of training stimulus 3 (black solid line without symbols, middle panel) as a strong distractor. Sketch of connections indicate an optimal read-out that focuses on the target odors and a non-optimal, random read-out. (F,G) MC activities in response to the probe stimuli with and without context. (H) The correct context suppressed the distractor and enhanced the Fisher discriminant ${\mathcal{F}}_{nonopt}$ of the random read-out. In contrast, the incorrect context reduced ${\mathcal{F}}_{nonopt}$. With top-down input blocked (w_GC = 0) ${\mathcal{F}}_{nonopt}$ is slightly reduced compared to the context-less case. The error bars quantify the fluctuations in the network connectivities.

Fig 7. Cortical task switching enhances discrimination.

(A) The training stimuli consisted of 2 pairs of dissimilar stimuli. (B,C) Resulting bulbar and cortico-bulbar connectivity. (D) The probe stimuli consisted of two mixtures of the training stimuli. (E) Cortical connectivities after cortical training on the individual components (left) and on the mixtures (right). (F) MC activities with cortex trained on the components (left) and on the mixtures (right), respectively. After training on the mixtures the inhibition by the top-down input reduces the MC amplitudes without reducing the difference between the two mixtures, enhancing their discriminability. (G) Optimal Fisher discriminant ${\mathcal{F}}_{opt}$ with cortex trained on the components (blue) and on the mixture (green). The error bars reflect the ongoing network fluctuations.

Fig 8. Role of top-down input in the discrimination and detection of stimuli with distractor.

The system was trained with the stimuli of Fig 6A. (A) Intra-bulbar disynaptic inhibition among MCs (W^(MM)) became less selective with increasing top-down weight w_GC. (B) The disynaptic inhibition of MCs by CCs (W^(MC)) was selective for an intermediate range of w_GC, but not for small or large values of w_GC. (C) Fisher discriminant ${\mathcal{F}}_{opt}$ for the non-optimal discrimination of two similar stimuli in the presence of a distractor (cf. Fig 6 and upper panel of (D)) as a function of w_GC. The correct context enhanced the discrimination only when the connectivities W^(MM) and W^(MC) were selective. Even with the top-down input blocked by setting w_GC = 0 during the probing (but not during the network development) the discriminability of the odors was reduced when w_GC was increased during the learning (grey line), reflecting the decrease in selectivity of W^(MM). (D) Stimuli used for probing in (C,E). (E) Fisher discriminant ${\mathcal{F}}_{opt}$ for the non-optimal detection of a weak stimulus in the presence of a distractor. Grey line: w_GC = 0 during probing but not during training of the network.