Noise Gated by Dendrosomatic Interactions Increases Information Transmission

We study how noise in active dendrites affects information transmission. A mismatch of both noise and refractoriness between a dendritic compartment and a somatic compartment is shown to lead to an input-dependent exchange of leadership, where the dendrite entrains the soma for weak stimuli and the soma entrains the dendrite for strong stimuli. Using this simple mechanism, the noise in the dendritic compartment can boost weak signals without affecting the output of the neuron for strong stimuli. We show that these mechanisms give rise to a noise-induced increase of information transmission by neural populations.

Popular Summary

The brain must reliably process a lot of information, despite working in a noisy environment. Variability is an intrinsic feature of the brain’s underlying molecular machinery. But when it comes to preserving information, noise can play two roles in physical and biological systems. Noise can corrupt strong signals, but it can also help the detection of signals that would otherwise go unnoticed. The nervous system might, therefore, have found a way to turn an opponent into an ally. We studied how information transmission is affected by intrinsic variability. By developing computer simulations of a population of nerve cells, we showed that variability can be controlled by interactions within each nerve cell in a manner that enhances information transmission.

We argue, on computational grounds, that neurons enhance transmission by actively toggling between low- and high-noise regimes. The proposed mechanism hinges on the interaction between two neuronal compartments that possess different levels of intrinsic noise: a small and noisy dendrite (the thin branches extending from a neuron), and a large and deterministic soma (the bulbous main cell body). The dendrite sustains noise to detect low-intensity signals, whereas the somatic compartment reliably processes high-intensity signals. When assembled into a population, our model neurons showed an increased capacity to process incoming signals spread over a broad range of intensities.

These results indicate a role for dendrites as noise-assisted encoders of low-intensity signals. More generally, our study suggests that well-tailored interactions between a system’s subparts could control noise in order to enhance information transmission.

Figure 1

Exchange of leadership in a simplified biophysical model of the dendrite-soma system. (a) The dendrite-soma system (see the Appendix) receiving a weak input in both the dendrite and the soma. The membrane potential of the soma (blue trace) and the dendrite (red trace) shows action potentials initiated first in the dendritic compartment. (b) For a strong input, the somatic compartment fires regularly and generally leads the dendritic action potentials. (c) The fraction of somatic spikes that are preceded by a dendritic spike within 8 ms (blue full line) and the fraction of dendritic spikes that are preceded by a somatic spike by at least 8 ms (orange dashed line) are shown. Spike timing is taken to be the time of crossing $-30\mathrm{mV}$ from below. (d) The firing rate of the coupled system is shown as a function of the input strength (black full line). Isolating the compartments by fixing the coupling conductance to zero shows that the dendrite-soma system interpolates between the isolated dendritic compartment (red dashed line) and the isolated somatic compartment (blue dash-dotted line).

Figure 2

Noise gating in coupled integrate-and-fire units. (a) Membrane potential response of $Y$ (somatic, blue dashed line) and $X$ (dendritic, red full line) units for $s=0.95$ and $D_X=3D_Y$. Inset: Expanded view near threshold, where the $X$ unit crosses first (arrow). (b) Same as (a) but for stronger input $s=1.15$. (c) Entrainment probability for a range of input strengths, as described in the caption of Fig. 1 The time window used to establish leadership in the integrate-and-fire model is of the same order as the integration time step. The input intensity where entrainment probabilities cross is the switching point. (d) Dependence of the switching point on $D_X/D_Y$ for different values of $D_Y$. Parameters are described in [36].

Figure 3

The firing statistics of isolated units constitute asymptotic curves for the dendrite-soma system. (a) Firing rate of coupled system (black line), $X$ unit alone (i.e., dendrite alone, red line), and $Y$ unit alone (i.e., soma alone, blue line) for different input strengths. (b) The interspike interval variance of the coupled system follows the minimum between isolated $X$ and $Y$ units. (c) Fisher information about the input strength. Parameters are described in [36].

Figure 4

Stochastic resonance for bimodal input distributions. (a) Schema illustrating a neural ensemble made of dendrite-soma systems (black), somatic units alone (blue), or dendritic unit alone (red). Input distribution for (b) the JDP and (c) a GP with mean matched to that of the JDP. (d) For the JDP, $\mathcal{M}$ calculated from the summed activity of 8000 coupled units (black line) surpasses that of isolated $Y$-unit compartments (blue dot-dashed line) and of isolated dendritic compartments (red dashed line) for a broad range of dendritic noise. (e) The resonance for a GP is reduced with respect to the resonance of obtained using the JDP (d). We use $\tau =10\mathrm{ms}$ to represent $\mathcal{M}$ in bits per second; see [36] for all other parameters.

Figure 5

Stochastic resonance for wide input distributions. (a) Input distribution of the GP chosen to cover a large range around a high mean $\overline{s}$. Using the same color code as in Fig. 3, 3 shows $\mathcal{M}$ and (c) shows $\mathcal{E}$ (in bits per spikes) for different $D_X$ calculated with input distribution in (a). We restrict the range for intrinsic noise such that the standard deviation of $u_X$ remains smaller than 0.5, which would correspond to substantial membrane potential fluctuations in physical units. (d) $\mathcal{M}$ and (e) $\mathcal{E}$ for different values of the mean input and for a fixed level of intrinsic noise $D_X=15D_Y$. Parameters are described in [36].