A dendritic disinhibitory circuit mechanism for pathway-specific gating

Abstract

While reading a book in a noisy café, how does your brain 'gate in' visual information while filtering out auditory stimuli? Here we propose a mechanism for such flexible routing of information flow in a complex brain network (pathway-specific gating), tested using a network model of pyramidal neurons and three classes of interneurons with connection probabilities constrained by data. We find that if inputs from different pathways cluster on a pyramidal neuron dendrite, a pathway can be gated-on by a disinhibitory circuit motif. The branch-specific disinhibition can be achieved despite dense interneuronal connectivity, even with random connections. Moreover, clustering of input pathways on dendrites can naturally emerge through synaptic plasticity regulated by dendritic inhibition. This gating mechanism in a neural circuit is further demonstrated by performing a context-dependent decision-making task. The model suggests that cognitive flexibility engages top-down signalling of behavioural rule or context that targets specific classes of inhibitory neurons.

Figure 1. Dendritic disinhibitory circuit as a mechanism for pathway-specific gating.

(a) Subcellular microcircuit motif for gating through dendritic disinhibition. Dendrites of pyramidal neurons are inhibited by SOM interneurons, which are themselves inhibited by VIP interneurons. A control input (representing a context or a task rule) targeting VIP interneurons (and potentially SOM neurons) can thereby disinhibit pyramidal neuron dendrites, opening the gate for excitatory inputs targeting these dendrites. (b) Circuit configuration for pathway-specific gating. Pyramidal neurons receive converging inputs from multiple pathways, for example, visual and auditory. Single neurons in these areas are selective to multiple stimulus features, indicated here by colour and frequency. The processing of each pathway is regulated by the control input. (c) Inputs from different pathways target distinct subsets of dendrites of these pyramidal neurons. A pathway can be gated-on by specifically disinhibiting the dendrites that it targets, corresponding to an alignment between excitation and disinhibition. Disinhibition is represented by dashed lines.

Figure 2. Context-dependent gating of specific pathways.

(a) A reduced compartmental neuron with a somatic compartment connected to multiple, otherwise independent, dendritic compartments (only three shown). (b) Excitatory inputs can generate a local, regenerative NMDA plateau potential in the dendrite. As number of activated synapses increased, there is a sharp nonlinear increase in the evoked dendritic membrane depolarization (V_D). (c) Disinhibition of the targeted branch opens the gate for the excitatory input. (d) A pyramidal neuron receives converging inputs from multiple pathways carrying different stimulus features, giving it selectivity to a preferred stimulus for each feature dimension. Each input pathway targets separate dendrites, which are disinhibited correspondingly in each context by top-down control inputs (not modelled here). (e) Tuning curve for input pathway 1, when only this pathway is activated. The input pathway encodes a stimulus feature, for example, motion direction, with a bell-shaped tuning curve for the input. The preferred feature value corresponds to higher input firing rate. When gate 1 is open by disinhibiting the dendrites targeted by input pathway 1, the neuron exhibits strong tuning (light blue). When gate 2 is instead open, the neuron exhibits weak tuning for the feature (dark blue). The amount of inhibition reduced for a disinhibited dendrite, that is, the disinhibition level, is 30 Hz. (f,g) Two-dimensional tuning curves when both pathways are activated. (f) In the default context, no dendrites are disinhibited and both pathways are gated-off. The neuron exhibits weak responses regardless of the stimulus features. (g) When gate 1 is open by disinhibiting branches targeted by pathway 1, the response of this neuron is dominated by tuning to the pathway 1 stimulus, although pathway 2 has a residual impact.

Figure 3. Characterization of gating selectivity in pyramidal neurons.

(a) Schematic of gating, presenting pathway 1 input when gate 1 is opened (left) or gate 2 is opened (right). There are N_dend available dendrites in total. Each input pathway targets N_disinh dendrites. To gate a pathway on, these exact N_disinh dendrites are disinhibited, creating an aligned pattern of disinhibition. Each pathway selects dendrites randomly and independently from other pathways, which can result in overlap of the excitation–disinhibition patterns across pathways. When N_disinh is large, projections from different pathways are more likely to overlap. The neuron's firing rate is r_on and r_off in response to the preferred stimulus of the gated-on (left) and gated-off (right) pathway, respectively. The gating selectivity is defined as (r_on−r_off)/(r_on+r_off), which is 1 for perfect gating and 0 for no gating. (b) Gating selectivity increases as excitation/disinhibition patterns become sparser, that is, with a smaller proportion of targeted and disinhibited dendrites for a pathway (N_disinh/N_dend). Diamonds mark the case of non-overlapping excitatory projections, corresponding to the limit of maximal sparseness. (c) Gating selectivity is higher with stronger disinhibition, for all sparseness levels.

Figure 4. Gating selectivity as functions of SOM–pyramidal circuit parameters.

(a) A simplified model for a cortical column of SOM and pyramidal neurons. We only modelled the SOM-to-pyramidal connections. The model is subject to experimentally measured constraints of the following parameters: number of SOM neurons (N_SOM), connection probability from SOM to pyramidal neurons (P_SOM→pyr), and the number of dendrites on each pyramidal neuron (N_dend). We consider the ‘worst-case' scenario that the SOM-to-dendrite connections are random. Finally, we assume for now that control input for each pathway suppresses a random subset of SOM neurons. The different contrasts used are for illustration purpose only. (b) A critical parameter for the SOM-to-pyramidal circuit is the number of SOM neurons targeting each dendrite (N_SOM→dend). This parameter can be calculated using other experimentally measured parameters under the assumption of random connectivity, . (c–e) Gating selectivity only weakly depends on N_dend (c), N_SOM (d) and P_SOM→pyr (e) if N_SOM→dend is kept constant by co-varying another parameter. The plotted curve marks the mean and the shaded region marks the bottom 10% to top 10% of the neuronal population. (f) Gating selectivity is high when each dendrite is targeted by a few SOM neurons. Given experimental measurements of P_SOM→pyr≈0.6, N_dend≈30 and N_SOM≈160, we obtained N_SOM→dend≈5, leading to relatively high gating selectivity ∼0.5. Total strength of inhibition onto each pyramidal dendrite is always kept constant when varying parameters.

Figure 5. Two possible schemes of control for the interneuronal circuit.

We built a simplified circuit model containing VIP, SOM and pyramidal neurons. (a–c) Control signals target only VIP neurons. (a) In this scheme, for each pathway, control inputs target a random subset of VIP neurons. And the connection probability from VIP to SOM neurons is P_VIP→SOM. (b,c) Good gating selectivity is only achieved when a small subset of VIP neurons is targeted by control inputs (b), and when the VIP–SOM connections are sparse (c). (d–f) Control signals target both VIP and SOM neurons. (d) In this scheme, we assume that for each pathway control inputs target a random subset of VIP and SOM neurons. (e) Gating selectivity depends on the proportion of SOM (blue) but not VIP (green) neurons targeted by control input. (f) Gating selectivity does not depend on P_VIP→SOM. Curves and shaded regions are as in Fig. 4.

Figure 6. Somatic inhibition improves gating selectivity.

(a) PV neurons project to the somatic areas of pyramidal neurons, and are inhibited by SOM neurons and themselves. Suppression of SOM neurons causes disinhibition of PV neurons, therefore an increase in somatic inhibition onto pyramidal neurons. (b) A moderate increase in somatic inhibition always improves gating selectivity. We included PV neurons and their corresponding connections in the model of Fig. 5d. Gating selectivity increases as a function of the SOM-to-PV connection weights (w_SOM→PV) in a wide range (see Supplementary Note 1 for a proof). However, when gating selectivity is low without PV neurons (light curve), the peak of this increase is lower and the slope is sharper. Gating selectivity starts to decrease when the SOM-to-PV connection, therefore the somatic inhibition, is too strong that the responses of many pyramidal neurons are completely suppressed.

Figure 7. Learning to gate specific pathways.

(a) Model schematic. Pre- and post-synaptic spikes both induce calcium influx. The overall synaptic weight change is determined by the amount of time the calcium level spends above thresholds for depression (θ_d) and potentiation (θ_p). The model is fitted to the experimental data, and is able to quantitatively predict results not used in the fitting. (b) Dendritic inhibition makes potentiation harder to induce. With background-level inhibition (light blue), synaptic weight change shows three regimes as a function of excitatory input rate: no change for low rate, depression for medium rate and potentiation for high rate. With a medium level of inhibition (dark blue), potentiation requires a higher excitatory input rate. With relatively strong inhibition (black), potentiation becomes impossible within a reasonable range of excitatory input rates. The post-synaptic rate is fixed at 10 Hz. (c) Learning paradigm. Left: excitatory synapses from each pathway are initialized uniformly across dendrites. When pathway 1 is activated, specific branches of the neuron are disinhibited (dashed line), that is, gate 1 is open. During learning, only one pathway is activated at a time. Right: after learning, activated excitatory synapses onto the disinhibited branches are strengthened, while activated synapses onto inhibited branches are weakened, resulting in an alignment of excitation and disinhibition patterns. Synaptic weights of non-activated synapses remain unchanged (not shown). (d) Response properties of the neuron before learning. Top: tuning curve of the neuron when only pathway 1 is presented. The neuron shows no preference to the gate opened prior to learning. Bottom: two-dimensional tuning curve of the neuron when both pathways are simultaneously presented and gate 1 is open. See Fig. 2 for the definition of the tuning curves. (e) Response properties of the neuron after learning. Top: the neuron shows strong tuning to pathway 1 input when gate 1 is open. Bottom: when both pathways are presented, the neuron's response is primarily driven by pathway 1 stimulus, although pathway 2 stimulus also affects the neuron's firing.

Figure 8. Pathway-specific gating in an example context-dependent decision-making task.

(a) A flexible decision-making task. Depending on the context, the subject's behavioural response should be based on either the colour or the motion direction of the stimulus. (b) The circuit model scheme. Motion and colour pathways target associative sensory neurons, which are subject to context-dependent disinhibitory control. Neurons preferring colour and motion evidence for the same target project to the corresponding neural pool in the decision-making circuit. (c) Associative sensory neurons receive converging inputs from both motion and colour pathways, and are controlled by the dendrite-targeting interneuronal circuit. (d–f) Fit and prediction of behavioural performance. Behavioural performance in the motion context as a function of motion coherence (d) and colour coherence (e) for a monkey (dots), and the model's fit (line). Experimental data are extracted from ref. . The model can capture the behavioural performance of a monkey. (f) In the model, the impact of the irrelevant pathway (colour) is strongest when the relevant pathway signal is weak (with low motion coherence).