Effects of corticothalamic feedback depend on visual responsiveness and stimulus type

Abstract

In the dorsolateral geniculate nucleus (dLGN) of the thalamus, stimulus-driven signals combine with modulatory inputs such as corticothalamic (CT) feedback and behavioral state, but their impact in shaping dLGN activity is debated. We recorded extracellular responses in the dLGN of mice viewing a movie stimulus, while photosuppressing CT feedback and tracking locomotion and pupil size. Using generalized linear models fit to single neuron responses, we found that including CT feedback and behavioral state improved model predictions, especially for a subpopulation of neurons poorly responsive to the movie. Intriguingly, the impact of CT feedback was stronger without a patterned visual stimulus. Finally, for neurons sensitive to CT feedback, visual stimuli could be more easily discriminated when CT feedback was suppressed. Together, these results show that the effects of modulatory inputs in dLGN depend on visual responsiveness and stimulus type, with CT feedback affecting sensitivity and reliability, potentially to tune the thalamic relay.

Keywords: Neuroscience; Sensory neuroscience.

Graphical abstract

Figure 1

CT feedback and behavior modulate dLGN responses to movies (A) Schematic of the recording setup and photo-suppression of V1 L6 CT pyramidal neurons in Ntsr1-Cre mice with Cre-dependent AAV-stGtACR2-RFP. (B) Histology. Left: Coronal section near the V1 injection site, with stGtACR2-RFP expression (red) in Ntsr1+ somata. Blue: DAPI; scale bar: 1 mm. Inset: Magnification of area marked by dotted rectangle. Scale bar: 50 μm. Right: Coronal section of dLGN recording sites, with electrode tracks for two consecutive recording sessions (arrows 1 and 2) marked with DiI (yellow). Scale bar: 1 mm. Dotted line: dLGN contour. Numbers on top: position relative to Bregma in mm. (C) Snippet of an example dLGN recording. Top to bottom: example frames of the movie stimulus, photo-suppression pulse train (blue), running speed (green), pupil area (yellow), and time varying firing rate (sorted by first principal component) of simultaneously recorded dLGN neurons. (D) Raster plot of responses of an example dLGN neuron, time locked to the onset of CT feedback photo-suppression (blue, OFF-ON transition) and to control periods without photosuppression (gray, OFF-OFF transition). Note that the example neuron illustrates the observed effect of CT feedback photo-suppression, but the size of the effect is not representative of that observed in the population of recorded dLGN neurons. (E) Corresponding PSTHs (solid line: average across trials, shaded area: standard error of the mean). (F) Effects of CT feedback photo-suppression (left), locomotion (middle), and pupil size (right) on dLGN mean firing rates. Example neuron from (D, E) marked with ×. p values denote results of a Wilcoxon signed-rank test, n = 122 neurons. Insets: Histogram of firing rate fold-change relative to control (Δ FR log₂-ratio).

Figure 2

Spline-based GLM captured the RF, the influence of CT feedback, and the impact of behavior on responses of dLGN neurons (A) Schematics of the spline-GLM model architecture. Firing rate of individual dLGN was predicted as a combination of kernel outputs summed at the linear stage and then passed through a softplus nonlinearity. Each modeled neuron had a kernel for the stimulus and kernels for the three modulatory inputs: run speed (green), pupil size (orange), and CT feedback suppression (blue). (B) Schematics of the permutation test to assess the significance of the learned kernels. Model performance was evaluated by comparing the actual correlation (Pearson r) between predicted and observed firing rates to correlations when one of the inputs was taken from an unrelated experimental session (for movie, running, and pupil size) or randomly generated with the same statistics (for CT feedback suppression). Inputs were considered significant if the actual performance differed from the permuted performance with p ≤ 0.05. (C) Three dLGN example neurons, their learned kernels, firing rate predictions, and outcomes of the permutation test (neuron 1 is the same example neuron as in Figures 1D and 1E). (C₁) Spatial and temporal RF components separated by singular value decomposition (SVD, see STAR Methods), along with kernels for the modulatory inputs. (C₂) Observed (gray) versus predicted (black) firing rates during 80 s of movie presentation. (C₃) Actual model performance (Pearson’s r, red dashed line) and performance for permuted stimulus (gray), CT feedback suppression (blue), running (green), and pupil size (orange) inputs. Kernels that contribute significantly to the model’s performance are marked with ←. The box displays the quartiles of the distribution, the whiskers represent its range, excluding outliers. (D) Spatial RFs of example neurons with significant stimulus kernels. Gray: outline of common visual space (azimuth: -35–110 deg; elevation: -35–50 deg); Solid lines: monitor border. (E) Temporal RFs (SVD component of the stimulus kernel) in the recorded dLGN population, sorted by their area under the curve. (F–H) Modulatory kernels in the recorded dLGN population, sorted by their area under the curve, for CT feedback suppression (F), running (G), and pupil size (H). Horizontal bars, side: Neurons with significant kernels based on the permutation test. Panels (E–H) show data from all n = 122 neurons.

Figure 3

Adding model predictors for modulatory inputs improves the performance for a subgroup of poorly visually responsive dLGN neurons (A) Comparison of model performance for models with different sets of inputs indicated in the table on the left, sorted by performance. Bonferroni corrected p-values of paired Wilcoxon signed-rank test: ∗ ∗ ∗ ∗ p ≤ 1.0 × 10⁻⁴; “ns” non-significant. Error bars: 95% confidence intervals. (B) Comparison of model performance (Pearson’s r) for the “Full model” (black, inputs: stimulus, CT feedback suppression, run speed, pupil size) and the “Stimulus only” model (gray). Arrow heads: mean performance (n = 122 neurons). (C) Improvement in model performance (“Full model”−“Stimulus only”) for neurons grouped by their performance of the “Stimulus only” model. Bonferroni corrected p-values of paired Wilcoxon test: ∗p ≤ 0.05; ∗ ∗ ∗ ∗ p ≤ 1.0 × 10⁻⁴. Dotted lines indicate the mean of each group. The box displays the quartiles of the distribution, the whiskers represent its range, excluding outliers. (D) Comparison of model performance for all neurons in the “Stimulus only” model and the “Full model.” Arrows indicate the group of neurons that benefits from adding CT feedback, running, and pupil size (“Modulation-explained”) and the group that does not improve (“Stimulusexplained”). (E) Relationship between the amount of joint modulation by CT feedback, running, and pupil size estimated directly from the data without a model (see STAR Methods), and the improvement in model performance when adding predictors for modulatory inputs (“Full model”−“Stimulus only”). (F) Comparison of model performance for the “stimulus only” model and a model without the stimulus (“modulation only”). Arrows indicate “modulation-explained” and “stimulus-explained” neurons. In panels (D, F), darker colors indicate stronger joint modulation by CT feedback, running, and pupil size (MI_Joint) estimated directly from the data without a model (see STAR Methods). Panels (D–F) show data from all n = 122 neurons.

Figure 4

The effect of CT feedback is dependent on the presence or absence of the visual stimulus (A) Schematics of the hypothesis that the absence of a patterned visual stimulus elicits weaker stimulus-driven input to dLGN, which is in turn accompanied by stronger CT feedback. Note that the schematics should not imply stronger L6 CT pyramidal neuron firing, but that the net effect of CT feedback on dLGN firing is stronger in the absence of a patterned visual stimulus. (B) Schematics of fitting the spline-GLM model separately during movies (top) vs. blank periods (bottom). (C) Effects of CT feedback suppression. (C₁) Mean PSTHs time locked to the onset of photosuppression for one example dLGN neuron (same example neuron as in Figures 1D and 1E) during blank periods (gray screen) flanking the movie presentation. The shaded area represents the SEM. (C₂) Same as (C₁), during movie presentation. Purple, blue: PSTH during CT feedback suppression (OFF-ON transition), light gray, dark gray: PSTH during control condition (OFF-OFF transition). (D) Comparison of MI_{CT FB supp.} during blanks vs. movies for the recorded dLGN population (number of neurons n = 122). Note that the results do not depend on the exact metric used for the comparison of CT feedback suppression effects, and also hold if raw differences or simple ratios are considered instead of the MI_{CT FB supp.} (E) Percentage of recorded dLGN neurons modulated by CT feedback during the two stimulus conditions. Three modulation metrics were separately considered to count the modulated neurons. A neuron was considered modulated (1) based on data: |MI_{CT FB supp.}| from (D) ≥ 0.1, (2) based on model predictions: |MI_{CT FB supp.}| from (H) ≥ 0.1, or (3) based on model performance: permutation test p ≤ 0.05. Notably, all three modulation metrics consistently revealed a higher proportion of neurons displaying CT feedback modulation during the blank condition compared to the movie condition. (F) CT feedback suppression kernel for the example dLGN neuron in (C₁). The model was trained on either the data during movie presentation (dark blue) or blank periods (light blue). (G) Same for all significantly CT feedback modulated neurons (n_movie = 12, n_blank = 21). Solid lines represent the mean of the kernels, and transparent surrounds represent the standard error of the mean (SEM). (H) Comparison of MI_{CT FB supp.} for blanks vs. movies calculated from simulated data using the fitted model for the dLGN population to the two stimuli. (I) Comparison of the Rate of Change (RoC) of model-predicted neurons’ responses to CT feedback suppression during movie presentation vs. blank periods. (J) Same as (I), for the recorded data. Panels (H–J) show data from all n = 122 neurons.

Figure 5

2AFC decoder applied to modeled and recorded dLGN responses reveals that CT feedback suppression can improve stimulus discrimination (A) Schematics of the decoder: the trained model was used to simulate responses with CT feedback being either on (blue) or off (black; 70 s example trace). The movie stimulus, running, and pupil inputs remained the same between CT feedback on and off, using recorded data. Random 50 ms stimulus fragments (Green: A and yellow: B) were chosen, each with the corresponding simulated responses. The decoder’s task was to determine which stimulus was more likely based on the observed responses. The decision of the decoder was directly derived from the model likelihood for the correct and incorrect pairing (see STAR Methods). (B) Decision matrix for one example neuron with 20 random stimulus fragments and their simulated responses. Green: correct pairing of stimulus and response; yellow: incorrect pairing. Left: control condition; right: CT feedback suppressed condition. (C) Percentage of correct decisions in the control and the CT feedback suppressed condition, same example modeled neuron as in (B). Error bars indicate 95% confidence intervals. (D) Change in decoder performance during CT feedback suppression vs. control conditions, split according to whether the modeled neurons were significantly modulated by CT feedback suppression (dark blue) or not (light gray). The box displays the quartiles of the distribution, the whiskers represent its range, excluding outliers. (E) Relationship between the fold change in trial-by-trial reliability and the fold change in decoding accuracy with CT feedback suppression across the population of dLGN neurons, obtained from the simulated responses in (A–D). (F) Spike rasters for one dLGN example neuron in response to a 5 s natural movie clip during the control condition (left) and CT feedback suppression (middle), obtained from the published dataset by Spacek et al. (2022).Green/orange shading: two random 50 ms, non-overlapping movie fragments used to illustrate the analysis. Right: illustration of the support vector machine (SVM) trained to perform 2AFC discrimination of the movie fragments based on single-trial responses. (G) Principal component analysis of the single trial responses to two random 500 ms movie fragments. (H) Percentage of correct decisions of the SVM trained on the example neuron’s responses to 200 random movie fragment-pairs for held-out trials not used for training, separately for the control condition and during CT feedback suppression. Error bars indicate 95% confidence intervals. (I) Change in SVM discrimination accuracy during CT feedback suppression vs. control conditions, split according to whether the modeled neurons were significantly modulated by CT feedback suppression (dark blue) or not (light gray). Modulated neurons were defined as the neurons whose firing rates were significantly negatively modulated by CT feedback suppression (p ≤ 0.01,n = 200 trials, Mann-Whitney-U test). The box displays the quartiles of the distribution, the whiskers represent its range, excluding outliers. (J) Relationship between the fold change in trial-by-trial reliability and the fold change in decoding accuracy across the population of dLGN neurons, obtained from the published dataset by Spacek et al. (2022) in (F–I). Error bars indicate 95% confidence intervals. Significant results are indicated with ∗ ∗ p ≤ 0.01, ∗ ∗ ∗p ≤ 0.001 and ∗ ∗ ∗ ∗ p ≤ 0.0001.

Figure 6

Synaptic noise framework of CT feedback Schematic illustration of the framework by Wolfart et al. (2005) and Behuret et al. (2015), proposing that CT feedback induced synaptic noise could tune the transfer function of individual dLGN neurons, making them more sensitive to weaker inputs. The noise-induced increased sensitivity, however, comes at the expense of making action potential firing more probabilistic, thus decreasing single-neuron reliability. Such decreased single-neuron reliability could be offset through the strong convergence at the geniculo-cortical synapse,^,, where the pooled afferent signal would not only have increased sensitivity to weak inputs, but also an increased dynamic range and hence better resolution. Adapted from Behuret et al. (2015), originally published under a Creative Commons Attribution License (CC BY).