Top-down beta oscillations modulate gamma band
feedforward stimulus drive between V4 and V1
William Hedley Thompson
Research Project 2, University of Amsterdam. Supervisor: Craig G. Richter.
Co-Assessor: Conrado Bosman.
Lab: Fries Lab, Ernst Strungmann Institute for Neuroscience in Cooperation with the Max Planck Society.
Top-down beta oscillations modulate gamma band
feedforward stimulus drive between V4 and V1
The brain attempts to anticipate upcoming percepts. Upon the onset of a difting-grating stimulus, bottom-up Granger Causal (GC) interaction in the gamma band is present from primary visual cortex (V1) to extrastriate visual cortex (V4). This activity has been implicated in feedforward stimulus drive.1,2 Perceptual expectation affects baseline activity.3 Top-down activity is known to affect perceptual states.4 Top-down baseline activity at beta (13-30hz), prior to the onset of a stimulus, from V4 to V1, has been found to modulate the amplitude of the evoked response in V1.5 Beta has also been implicated in anticipatory processes in frontal areas,6 has predicitve abilities in decision making in motor sensitive areas,7 as well as frontal and parietal areas.8 An overall mechanism where beta is singaling expected predictions has been proposed.9 It is however unclear whether this pre-stimulus beta anticipation is targeting specific regions or the entirety of V1. Additionally, the relationship between the top-down pre-stimulus oscillatory activity and feedforward stimulus induced oscillatory activity has not previously been investigated. We found that top-down beta frequency influences prior to stimulus onset are active in the same spatial regions of V1 as those that show the greatest gamma activity after stimulus onset. We found evidence that beta band top-down GC plays a modulatory role of bottom-up gamma frequency GC. Furthermore, we found that the magnitude of the beta influence does not appear to be affected by the deployment of attention. Our results suggest that top-down GC beta activity prior to stimulus onset may recruit perceptual resources in a spatially specific manner. This modulation is not modified by the presence of spatial attention, which may indicate that the process is mechanistically different.
A visual attention task was performed by two monkeys where two drifting-grating stimuli appeared (one blue, one yellow), followed by a cue instructing attention towards one stimulus requiring a reaction to a change in the cued stimulus while ignoring changes in the distractor stimulus (see methods and supplementary figure 1 for task descritpion). Both monkeys were implanted with a 252-channel electrocorticography (ECoG) grid10 covering a large area of cortex (Figure 1A; see Methods). To remove the recording reference and suppress noise from the recording electronics, bipolar derivations were constructed by differentiating activity between neighboring electrodes (see methods).
For this analysis, we focused on the “pre-stimulus” fixation period prior to the stimulus appearing on the screen (600ms prior to stimulus onset) and up until the point the cue could first appear (800ms post-stimulus onset). Our analysis focused on the sites covering extrastriate cortex (V4) (number of sites: Monkey K: 15 & Monkey P: 16) and primary visual cortex (V1) (number of sites: Monkey K: 40 & Monkey P: 54). Figure 1B shows these areas highlighted for
Monkey K. We refer to the average Granger Causal interaction between a single site and all sites in the opposite area as GC strength.
During the pre-stimulus phase, we found that there was a peak in top-down (from V4 to V1) and bottom-up (from V1 to V4) GC strength in the beta band (13-24hz) (Figure 1C). During the post-stimulus period (200-800ms post-post-stimulus), there was a large bottom-up GC gamma peak (55-85hz) (Figure 1D). In some trials only one stimulus appeared and, when this absence occurred in the recorded hemisphere, gamma was absent (see methods, Supplementary Figure 2). The subsequent trials after single-stimulus trials in the opposite hemisphere showed an identical pre-stimulus top-down GC strength spectrum during the pre-pre-stimulus period. This confirms that the pre-stimulus top-down GC beta was not caused by the stimulus from the previous trial. We therefore associate the top-down GC beta with anticipatory control.
During the pre-stimulus period, there were beta oscillations present in both top-down and bottom-up directions. The average GC strength time-frequency spectra during the pre-stimulus period for both top-down beta and bottom-up GC strength show beta frequency activity. (Figure 1E & F). This activity increases dramatically for both during the pre- stimulus period. Prior to this large increase however, there is an increase in the top-down GC strength at beta centered at 340ms. This suggests that the top-down beta influence arises prior to the bottom-up GC beta frequency strength. For this reason, we focused on correlating the top-down beta with the effect of the stimulus. Future analysis should probe the relationship between this top-down and bottom-up pre-stimulus activation.
We wished to establish a connection between the pre-stimulus activity and how, upon stimulus onset, the activation of the stimulus was processed in V1. First we established that there was a positive correlation between every V1-V4 site pair’s overall pre-stimulus top-down GC at the beta frequency and the post-stimulus bottom-up gamma frequency GC. (Figure 2A; Monkey K: ρ=0.4774; p<<0.05; Monkey P: ρ=0.5887 p<<0.05). This implies that the site-pairs with high top-down beta GC during the pre-stimulus period are followed by high bottom-up gamma GC during the post-stimulus. This suggests co-activation between site-pairs such that those with high pre-stimulus beta GC will also have high post-stimulus gamma band GC.
While co-activation between V1 and V4 site-pairs implies that site-pairs which generally have high beta also have high gamma, but co-activation does not necessarily entail a fluctuation between beta and gamma GC on a single trial basis. It is possible that for a single site a slight increase in beta GC entails an increase or a decrease in the subsequent gamma GC. To test the relation between pre-stimulus top-down beta GC and post-stimulus bottom-up gamma GC over the trial dimension, we developed a single trial jack-knife correlation method, where equally biased single-trial estimates for GC could be found when used in correlations. Here we only used GC activity that showed significant activity at that frequency in both pre-stimulus and post-stimulus periods (see Methods). We correlated the fluctuations over trials between the spectral pre-stimulus top-down GC activity for each frequency bin (between 1-150hz) for each V4-V1
site pair with the post-stimulus GC bottom-up average gamma band in the reverse direction. Figure 2C shows an example for one site pair at one frequency (pre-stimulus top down GC at 20hz in Monkey K and meaned single trial bottom up GC across gamma band (71-91hz); Spearman rank coefficient: 0.16; p<0.05). To quantify the correlation across all site-pairs at each frequency, we summed the positive Spearman rank coefficients (ρ) values that passed a p<0.05 threshold and the negative ρ values that passed the same threshold. Both monkeys showed a large significant beta peaks ρ tally showing that, on a single trial basis, the pre-stimulus beta co-fluctuates positively with the post-stimulus bottom-up GC gamma. Furthermore, it shows that, from all the pre-stimulus top-down GC activity, beta correlates the most with the post-stimulus bottom-up GC gamma. Taken together, this is strong evidence for modulation of the feedforward drive of the post-stimulus bottom-up gamma GC by the anticipatory pre-stimulus top-down beta GC.
One possible concern is that the sites that show co-fluctuation are not necessarily the site pairs that have high co-activation. To control for this we took the site-pairs from pre-stimulus beta and gamma single trial co-fluctuation results and sorted by the post-stimulus bottom-up gamma GC magnitude across all trials. We took the top 10% and bottom 10% and compared the co-fluctuation results between these two groups. The thresholded ρ values from co-fluctuation were significantly higher for the top 10% of post-stimulus bottom-up GC gamma sites than the bottom 10% in both monkeys (Figure 2D).
To check whether this co-activation formed a retinotopic pattern, the mean of the pre-stimulus top-down beta GC across all V4 to each V1 site was projected spatially onto the V1 sites. We took the top one-third of the V1 sites that received the greatest GC influence from V4, which created a binary activation map. The largest cluster of neighbors through triangulation was found. This was contrasted with a binary activation map of the top one-third of V1 sites with the mean GC of the post-stimulus bottom-up gamma mapped onto the same image (Supplementary figure 3). The overlap between the two clusters was calculated (Figure 3). To test for significance we randomly generated two binary activation maps over V1, each with one third of the map activated, and calculated the overlap of the largest cluster for both. This was done 1000 times and compared this with the empirical result (Monkey K overlap of clusters 5 sites; significance overlap threshold: 4 sites; Monkey P overlap of clusters: 8 sites; significance overlap threshold: 5 sites where the significance levels is p<0.05, see Methods). This result implies retinotopic co-activation between the pre-stimulus top-down beta and post-stimulus bottom-up gamma.
With the pre-stimulus top-down beta GC modulation of the post-stimulus bottom-up gamma immediately after stimulus onset established, we wished to test the top-down beta GC role in the gamma activity after the attentional cue to determine if this anticipatatory mechanism varies with attention. When the cue comes on, the monkeys have to deploy attention to one of the two stimuli. We replicated the co-fluctuation analysis between top-down beta GC600ms prior to the cue appearing and the bottom-up gamma GC 200ms-800ms after the cue appearing (see
Methods). In both monkeys we found thresholded ρ values comparable to pre-stimulus beta GC and post-stimulus gamma GC results. There was however no significant difference between the pre-cue and post-cue correlations across attention conditions. The anticipation correlation between pre-cue top-down GC beta and post-cue bottom-up at gamma is the same regardless of whether attention has been deployed in the post-cue period (Figure 4). This suggests that this anticipatory mechanism is separate from attention.
We have demonstrated a modulation across time, frequency and direction. We have further shown that top-down beta GC could be implicated as a neural correlate of anticipation and is not modulated by attention. We have shown that this component of anticipation retinotopically modulates the evoked response and the feed-forward activity that has been linked to perceptual processing.
Large networks of oscillatory activity are considered to be present in the brain.11 To fully understand such networks, it is imperative to see the interactions and influences across frequencies band. Here we show how oscillatory activity implicated in anticipatory control in the beta-band interacts and modulates stimulus-driven feed-forward gamma band activity. The retinotopic specificity of this anticipatory modulation suggests that top-down beta frequency oscillations are dynamically modulating V1 in spatially specific manner, such that anticipated stimulus are processed more efficiently.
Methods: Experiment:
The project was a data-analysis project. For experimental and recording details see ref 1 and 2. Data-analysis
General
All analyses were performed using Matlab (MathWorks, MA; http://mathworks.com) and the FieldTrip toolbox12 (http://fieldtrip.fcdonders.nl).
Pre-stimulus data was -600ms-0ms to stimulus onset. Post-stimulus GC data was calculated on 200ms-800ms post-stimulus onset. This 200ms delay is required so that the Granger causality is not calculated over the event related potential. All Granger causality except the time-frequency analysis was calculated using the non-parametric spectral factorization (NPSF) method.13 Coherence and NPSF Granger used 7hz smoothing applied to attain the Fourier coefficients14 and the data was demeaned.
For the comparison between anticipation and attention, data was taken -600ms-0ms prior to the cue onset and 200ms-800ms after the cue onset. No change in any of the stimuli had occurred up until 800ms after cue onset. The same spectral analysis parameters as the pre-stimulus data were applied.
Bipolar derivation created artifical channel sites between actual chanels by subtracting the difference from one site from the other.
Time-frequency analysis
For the time-frequency auto-regressive (AR) models were calculated with a model order 18. The data was down-sampled to 250 samples per second The data had an anti-aliasing low-pass filter of 125hz applied to it. The model order was obtained using the Akaike Information Criterion (AIC) and comparing the power spectrum obtained with the AR model to the power spectrum obtained using the Fourier method. The AR Granger model was calculated on 200ms of data and slid 10ms along the pre-stimulus data from beginning at -600ms prior to stimulus onset and ending at -200ms.
Co-activation
Co-activation was calculated by calculating the granger causality per site pair for each V4-V1 site pair (Monkey K: 600 site pairs, Monkey P: 864) over all trials using the non-parametric spectral factorization method for both the pre-stimulus period and post-stimulus period. The average GC value at beta during the pre-stimulus (Monkey K: 16-24hz; Monkey P: 14-22hz) and
gamma (Monkey K: 71-81hz, Monkey P: 54-74hz) during the post-stimulus. These were then plotted on a log-log axis for visibility purposes.
Co-fluctuation
Only site-pairs that showed significant GC activity in either the pre-stimulus and post-stimulus period were used. This threshold of significance was calculated by creating surrogate data for each site-pair by randomizing the trial order of the V1 site in the V4-V1 site-pair 1000 times. For each permutation, the highest value in the Granger spectrum of the V4-V1 was taken to correct for multiple comparisons.15 The empirical data had to exceed the 95% value from the surrogate data to be counted. This was done for each frequency bin for the pre-stimulus. For the post-stimulus, if any frequency bin in the gamma range was significant, then that site pair passed the threshold.
For the single trial “co-fluctuation” analysis, for each site-pair, the Granger causality was calculated in the pre-stimulus and post-stimulus data every trial. This was done by a jack-knife approach for both pre-stimulus and post-stimulus which were then correlated. For the trial x, the trial number, x, was left out the calculation for both variables. This effectively inverts both distributions. When these distributions are then correlated, the inversion is corrected. For an in-depth analysis of the jack-knife correlation method for single trial estimates see ref. 16.
Each site-pair correlated the single trial estimates for each pre-stimulus frequency bin with the mean of the post-stimulus bottom-up GC gamma-band activity. The Spearman rank for all the site pairs, at each frequency, were collected and divided between positive and negative. The sum of both the positive groups and negative groups was calculated by only taking the site pairs whose correlation had a p<0.05. To establish the noise floor of this correlation, the post-stimulus GC gamma trial orders was permuted in a random order 1000 times. As with the empirical data, the positive and negative ρ values that had a p < 0.05 were summed. To correct for multiple comparisons the maximum summed ρ value of the positive correlations and the minimum summed ρ value of the negative correlations was taken at each permutation. A distribution was then created with the noise floor calculated at the 950th value.
The same method was done with the cue analysis for both the attention and non-attention condition. Only the mean ρ values through the beta were used to calculate the summed thresholded ρ at beta. To test for significance between the two conditions a permutation test where the ρ values in both conditions were randomly assigned and summed 1000 times. The 950th value was considered significant.
Retintopic co-activation
To find retinotopic overlap between the co-activation sites, pre-stimulus GC at beta and Post-stimulus GC at gamma were calculated as outlined for co-activation above. The mean activity to each V1 sites during the pre-stimulus and the mean activity from V1 during the post-stimulus
was projected onto a topographic map of V1. Through triangulation, the neighbours of each V1 site were calculated. For both time epochs, a topographic binary distribution was created whereby the top-third of V1 sites was given a binary value 1 and the bottom two thirds a value of 0. The largest cluster of 1s where the immediate neighbor was also a 1 was calculated. These clusters represent two retinotopic V1 maps of activated areas, one for the pre-stimulus activation to V1 at beta and one for the post-stimulus activation to V4 at gamma. The overlap between these two clusters was calculated by counting how many sites, if any, shared membership in both clusters (Monkey K: 5 sites; Monkey P: 8 sites). To show that this overlap was significant, we randomized 1000 times the two locations of the two binary distributions, the largest cluster was calculated and then the overlap between these was then found. If the empirical overlap was equal or greater to the 950th highest value (Monkey K: 4 sites; Monkey P: 5 sites), it was considered significant.
Figure Legends
Figure 1 | Top-down GC beta is present in the pre-stimulus and bottom-up GC gamma is present in the post-stimulus. A. 3D image of Monkey K’s brain with the implanted ECoG grid marked on top. B. Topographic image of Monkey K’s ECoG grid and site channels. V4 channels marked in green. V1 channels marked in red. C. Pre-stimulus GC strength spectra for Monkey K and Monkey P. Red shows top-down from V4 to V1. Blue shows bottom up from V1 to V4. D. Post-stimulus GC strength spectra for Monkey K and Monkey P. Colours show same directions as C. E. Time-frequency graph of pre-stimulus top-down GC strength from V4 to V1 for
Monkey K. F. Time-frequency graph of pre-stimulus bottom-up GC strength from V1 to V4 for Monkey K.
Figure 2 | Pre-stimulus top-down GC beta and post-stimulus bottom-up GC gamma positively correlate across channels and trials. A. Each point is a site-pair’s pre-stimulus top-down GC beta plotted against its post-stimulus bottom-up gamma for Monkey K and Monkey P. Each point represents one site-pair. B. Example correlation between one V4-V1 site’s pre-stimulus top-down GC at 20hz and the post-pre-stimulus bottom-up GC gamma. Each point represents one trial (ρ =0.16; p<0.05). C. Summed ρ values for all site-pairs pre-stimulus top-down correlation at each frequency bin with post-stimulus bottom-up gamma where p<0.05. Dashed line marks noise floor (p<0.05, corrected). D. Sorted by overall gamma, top 10% of summed thresholded ρ correlations from single trial correlation of pre-stimulus beta and bottom 10% of summed thresholded ρ (p<0.05).
Figure 3 | Co-activation of pre-stimulus top-down GC beta and post-stimulus bottom-up GC gamma retinotopically overlaps. Red: largest cluster of top 30% V1 site’s receiving pre-stimulus top down GC beta. Blue: largest cluster of top 30% V1 site’s sending post-pre-stimulus bottom-up GC gamma. Black: marks sites where there was overlap between blue and red. Figure 4 | The top-down GC beta correlates with post-stimulus bottom-up GC gamma regardless if attention is present. Single trial correlation of pre-cue top-down GC beta with post-cue bottom-up GC gamma across all pairs for attention deployed after the cue and no attention deployed after cue. Plot shows the sum thresholded ρ values (p<0.05) for all site pairs between V4 and V1.
Supplementary Figure Legends
Figure S1 | Task details and focus of the analysis. Diagram showing this analysis focus in the experiment, the time, regions of interest and directionality of the pre-stimulus and post-stimulus periods.
Figure S2 | Pre-stimulus top-down beta is not caused by the previous trial. Single stimulus trials where one stimulus was presented outside of the recording site did not elicit any induced response in the recording sites. In the pre-stimulus period on subsequent trials following these trials, top-down beta is still present in the recording sites. Spectra showing Granger Strength from V4 to V1 in Monkey K and Monkey P respectively. Red shows top-down activity for all two-stimuli trials. Blue shows top-down activity following single stimulus trials outside of the recording site.
Figure S3: Method for deriving retinotopic clusters. The average top-down GC activity to each site was taken for the pre-stimulus (top left). The sites with the highest one-third activation were placed in binary (marked in red). The same process was done for the post-stimulus activity where the average bottom-up GC activity from each V1 site was first mapped, one third hisgest activation was made into an activation matrix and how well this clustered (right, activity matrices marked in blue.)
References
1. Bosman, C,A., …. Neuron. (In print)
2. Bosman, C.A., BastosA.M., Schoffelen, J.M., Brunet, N., Oostenveld R., Womelsdorf, T., Rubehn, B., Stieglitz T., Weerd, P. & Fries, P. Gamma and beta coherent networks reveal dynamic bottom-up and top-down Processing in the Neocortex. (manuscript) 3. Lagner, R., Kellermann, T., Boers, F., Sturm, W., Willmes, K. & Eickhoff, S.B.
Modality-specific perceptual expectations selectively modulate baseline activity in auditory, somatosensory, and visual cortices. Cerebral Cortex, 21. (2011)
4. Gilbert, C.D. & Sigman, N. Brain states: top-down influences on sensory processing.
Neuron 54. (2007)
5. Richter, C.G., Coppola, R. & Bressler, S.L. Top-down influences predict early visual evoked response magnitude in V1. (manuscript)
6. Liang, H., Bressler, S.L., Ding, M., Truccolo, W.A. & Nakamura, R. Synchronized activity in prefrontal cortex during anticpation of visuomotor processing. Neuroreport 13. (2002)
7. Donner, T.H., Siegel, M. & Fries, P. Buildup of choice-predictive activity in human motor cortex during perceptual decision making. Curr. Biology 19. (2009)
8. Donner, T.H., Siegel, M. Oostenveld, R., Fries, P., Bauer, M. & Engel, A.K. Population Activity in the Human Dorsal Pathway Predicts the Accuracy of Visual Motion
Detection. J. Neurophysiol. 98. (2007)
9. Engel, A.K. & Fries, P. Beta-band oscillations: signaling the status quo? Curr. Opin
Neurobiol. 20. (2010)
10. Rubehn, B., Bosman, C., Oostenveld, R., Fries, P. & Stieglitz, T. A MEMS-based flexible multisite ECoG-electrode array. Journal of Neural Engineering 6. (2009) 11. Siegel, M., Donner, T. H., & Engel, A. K. Spectral fingerprints of large-scale neuronal
interactions. Nat. Rev. Neurosci. 13. (2012)
12. Oostenveld, R., Fries, P., Maris, E. & Schoffelen, J.M. FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data.
Computational Intelligence and Neuroscience, (2011).
13. Dhamala, M., Rangarajan, G. & Ding, M. Estimating Granger Causality from Fourier and Wavelet Transforms of Time Series Data. Phys. Rev. Lett. 100 (2008)
14. Mitra, P.P. & Pearson, B. Analysis of dynamic brain imaging data. Biophysical Journal 76. (1999)
15. Maris, E. & Oostenveld, R. Nonparametric statistical testing of EEG and MEG-data. J.
Neuro. Meth. 164. (2007)
16. Richter, C.G., Thompson, W.H. & Fries, P. Jack-knife correlation tool for single trial correlations of poor single-trial metrics. (in progress)
Fig 1
hz hz
hz hz
Fig 2 Pre-stimulus Top-Down GC Beta
Po st -s tim ul us b otto m -up GC g amma Po st -s tim ul us b otto m -up GC g amma
Pre-stimulus Top-Down GC Beta
hz hz Sum m ed t hr es ho lde d ρ 1S um m ed t hr es ho lde d ρ Top 10% GC. Bot 10% GC.
Fig 4 Sum m ed t hr es ho lde d ρ
Fig S1 Pre-stimulus (-600-0ms) Top-Down Granger Causality Post-stimulus (200-800ms) Bottom-Up Granger Causality
