Response inhibition in premotor cortex corresponds to a complex reshuffle of the mesoscopic information network

Abstract

Recent studies have explored functional and effective neural networks in animal models; however, the dynamics of information propagation among functional modules under cognitive control remain largely unknown. Here, we addressed the issue using transfer entropy and graph theory methods on mesoscopic neural activities recorded in the dorsal premotor cortex of rhesus monkeys. We focused our study on the decision time of a Stop-signal task, looking for patterns in the network configuration that could influence motor plan maturation when the Stop signal is provided. When comparing trials with successful inhibition to those with generated movement, the nodes of the network resulted organized into four clusters, hierarchically arranged, and distinctly involved in information transfer. Interestingly, the hierarchies and the strength of information transmission between clusters varied throughout the task, distinguishing between generated movements and canceled ones and corresponding to measurable levels of network complexity. Our results suggest a putative mechanism for motor inhibition in premotor cortex: a topological reshuffle of the information exchanged among ensembles of neurons.

Keywords: Behavior; Brain networks; Complexity; Information theory; Inhibition; Motor control.

Figure 1.

Sequence of Behavioral events characterizing the task. Go and Stop trials were randomly intermixed during each session. The epoch (T) of analysis is shown as a gray bar for both trials. White circles are feedbacks for the touch provided to the animals. RT, reaction time; SSD, Stop signal delay; SSRT, Stop signal reaction time; Go: time of Go signal appearance; Stop: Stop Signal appearance; M_on: movement onset.

Figure 2.

Neuronal activity modulation for different trial types. Neuronal MUAs (mean ± SEM) in the epoch T for the three behavioral conditions for all recording sites (modules) of S6 of Monkey P. Dark green traces show the average activity during Go trials aligned to the movement onset (gray vertical line). Dark blue traces show the average activity during wrong Stop trials aligned to the movement onset (gray vertical line). Red traces show the average activity during correct Stop trials aligned to the Stop signal presentation (light red vertical line) and plotted until the end of the SSRT. For each plot the y-axes are units of MUA while the x-axes mark the time during the epoch T; ticks, every 100 ms, are indicated for panels in the top row only. Numbers above each panel indicate the recording site. The actual spatial mapping of the recording sites is not preserved and are here shown in cardinal order from 1 to 96.

Figure 3.

Average TE matrices of Go, wrong and correct Stop trials for data in Figure 2. Each entry of the matrix is color coded according to the corresponding TE value averaged over trials. Statistically significant driver → target relationships correspond to TE values > 0 for the pair of modules considered. Here, for graphical purposes only, TE is normalized to the maximum across behavioral conditions so to obtain TE ∈ [0, 1] (color bar). Axes labels indicate recording sites.

Figure 4.

Information network of Go, wrong, and correct Stop trials for the recording session of Figure 2 and Figure 3. Each link is a significant TE value (i.e., a nonzero entry in the TE matrix). The number of outgoing connections of each module reflects its VD_out: the number of modules on which it acts as a driver. Each link is color coded according to the cluster to which its source node belongs. Information-spreading hubs are observable: C_1 during Go and wrong Stop trials and C_2 during correct Stop trials only. Here, for illustrative purposes and to better highlight the differences between behavioral conditions, the links of the C_3 nodes are removed and for each network only the 20% of the strongest links are shown (for the complete version including the C_3 see Supporting Information Figure S4). Recording sites 1 and 96 are marked in the Go trials graph. The coordinates of each node on the circle are preserved across behavioral conditions.

Figure 5.

Clusters robustness analysis. Robustness is inferred by monitoring the evolution of VD_out (y-axis) as a function of the information processed (TE, x-axis). Panels show the robustness curves averaged (mean ± SEM) over recording sessions compared between clusters and across behavioral conditions. Curves are color coded accordingly to the corresponding cluster as in Figure 4.

Figure 6.

Percolation analysis. (A) percolation curves over recording sessions compared across behavioral conditions and versus the null model for both animals. Each curve is a recording session. As sketched, the level of complexity of the underlying networks increases traveling the curves from left to right, whereas the hierarchy can be explored climbing down the steps of the curves (the fraction of GSCC, z-axis), measured in units of information (TE, x-axis). (B) comparison between the descent point between behavioral conditions (mean ± SEM) and against the null model distribution. (C) comparison between the measured slopes between behavioral conditions (mean ± SEM) and against the null model distribution. Stop trials resulted more complex and less hierarchical than Go trials while having comparable values between themselves (Kruskal-Wallis, p < 0.01 Bonferroni corrected for multiple comparisons). In all panels, for simplicity and homogeneity, TE is reported normalized to its maximum so to obtain TE ∈ [0, 1]. The null distribution is obtained from 500 runs of the null model (see Materials and Methods). Since the null distributions for the different behavioral conditions were indistinguishable, we only reported one for brevity. (GSCC: giant strongly connected component; NM: null model).

Figure 7.

Information spreading among clusters before movement generation and cancellation. Arrows between groups of modules are scaled according to the corresponding average value of I (among sessions and animals). The level of hierarchy in which each cluster is ranked is graphically represented as the vertical positioning of the nodes; such level depends on the set of measures explored and could be efficiently summarized by the index I. Since the C_3 is the dominant cluster in the network, its interactions are always shown. For the other clusters, only the interactions at least one order of magnitude greater than the others and statistically significant across behavioral conditions are shown. Wrong Stop trials are not shown; see Supporting Information Tables S5 and S6 for further details.