Descending pathways increase sensory neural response heterogeneity to facilitate decoding and behavior

Abstract

The functional role of heterogeneous spiking responses of otherwise similarly tuned neurons to stimulation, which has been observed ubiquitously, remains unclear to date. Here, we demonstrate that such response heterogeneity serves a beneficial function that is used by downstream brain areas to generate behavioral responses that follows the detailed timecourse of the stimulus. Multi-unit recordings from sensory pyramidal cells within the electrosensory system of Apteronotus leptorhynchus were performed and revealed highly heterogeneous responses that were similar for all cell types. By comparing the coding properties of a given neural population before and after inactivation of descending pathways, we found that heterogeneities were beneficial as decoding was then more robust to the addition of noise. Taken together, our results not only reveal that descending pathways actively promote response heterogeneity within a given cell type, but also uncover a beneficial function for such heterogeneity that is used by the brain to generate behavior.

Keywords: Behavioral neuroscience; Neuroscience; Sensory neuroscience.

Graphical abstract

Figure 1

Experimental setup and relevant neural circuitry (A) The animal is placed in an otherwise empty tank and behavioral responses (EOD; bottom left) as well as neural activity (top right) are recorded simultaneously. The stimuli consisted of amplitude modulations (AMs) of the animal’s own EOD (gray, middle left) whose amplitude (i.e., the envelope, green, middle left) was modulated sinusoidally at different frequencies. The animal’s behavior (bottom left) follows the detailed timecourse of the envelope stimulus. (B) Simplified circuit diagram. The envelope stimulus is transduced by electroreceptor afferents that project to pyramidal cells within the electrosensory lateral line lobe (ELL). ELL pyramidal cells are organized in columns consisting of six neurons (one ON- and OFF-type superficial, intermediate, deep per column) that in turn project to the midbrain torus semicircularis (TS) and indirectly to higher brain areas mediating behavioral responses. Although all ELL pyramidal cells project to TS, only deep ON- and OFF-type cells also directly (gray) project to the nucleus praeeminentialis (nP). ELL pyramidal cells also receive large amounts of descending input (i.e., feedback; pink arrow) from nP.

Figure 2

ELL pyramidal cells display highly heterogeneous responses to envelope stimuli that are strongly attenuated by pharmacological inactivation of feedback pathways (A) Simplified circuit diagram showing that pharmacological inactivation of feedback (pink arrow) onto ELL pyramidal cells was achieved by injecting lidocaine bilaterally into nP (red cross). (B) Envelope stimulus waveform (top, green) and raster plots showing the activities of the same ELL pyramidal cell population before (middle, blue) and after (bottom, red) feedback inactivation in response to this stimulus. Overall, it is seen that responses were much more similar to one another after feedback inactivation. (C) Histograms of the pairwise correlation coefficients between neural activities before (blue) and after (red) feedback inactivation. Note the increased probability of obtaining large correlation coefficient values near unity after feedback inactivation. Both distributions were significantly different from one another across our datasets (inset: two-sample Kolmogorov-Smirnov test, p = 1.04∗10⁻⁶, D = 39).

Figure 3

Heterogeneities are functionally beneficial, as decoders optimized to reconstruct the envelope stimulus’ detailed timecourse are more robust to noise addition before feedback inactivation (A) Schematic showing optimal decoding. Neural responses to the envelope are weighted and the weights are chosen such as to minimize the mean-squared error between the weighted sum of neural activities (i.e., the predicted stimulus) and the actual stimulus. (B) Actual (green) and predicted stimulus waveforms when the weights are optimized before (blue, top left) and after (red, top right) feedback inactivation. To test decoding robustness, independent normally distributed random numbers (i.e., noise) was added to each weight and the standard deviation of the distribution was progressively increased (see middle column). Increasing noise intensity increased the error between predicted and actual stimulus waveforms to a lesser extent before (blue, bottom left) than after (red, bottom right) feedback inactivation. (C) Performance as a function of noise intensity before (blue) and after (red) feedback inactivation. It is seen that performance is more greatly attenuated after feedback inactivation. Inset: The rate of increase of performance attenuation was greater after feedback inactivation (Wilcoxon signed rank test; p = 0.047, N = 7).

Figure 4

Downstream decoders take advantage of the beneficial function of heterogeneities to generate behavior (A) Schematic showing decoding. Neural responses to the envelope are weighted and the weights are chosen such as to minimize the mean-squared error between the weighted sum of neural activities (i.e., the predicted behavior) and the actual behavior. (B) Actual (brown) and predicted behavioral responses before (blue, top left) and after (red, top right) feedback inactivation. To test decoding robustness, independent normally distributed random numbers (i.e., noise) was added to each weight and the standard deviation of the distribution was progressively increased (see middle column). Increasing noise intensity increased the error between predicted and actual stimulus waveforms to a lesser extent before (blue, bottom left) than after (red, bottom right) feedback inactivation. (C) Performance as a function of noise intensity before (blue) and after (red) feedback inactivation. It is seen that performance is more greatly attenuated after feedback inactivation. Inset: The rate of increase of performance attenuation was greater after feedback inactivation. (Wilcoxon signed rank test; p = 0.016, N = 7).