Asymmetric directed functional connectivity within the frontoparietal motor network during motor imagery and execution

Both imagery and execution of motor controls consist of interactions within a neuronal network, including frontal motor-related regions and posterior parietal regions. To reveal neural representation in the frontoparietal motor network, two approaches have been proposed thus far: one is decoding of actions/modes related to motor control from the spatial pattern of brain activity; another is to estimate directed functional connectivity, which means a directed association between two brain regions within motor regions. However, directed connectivity among multiple regions of the motor network during motor imagery (MI) or motor execution (ME) has not been investigated. Here, we attempted to characterize the directed functional connectivity within the frontoparietal motor-related networks between the MI and ME conditions. We developed a delayed sequential movement and imagery task to evoke brain activity associated with data under ME and MI via functional magnetic resonance imaging scanning. We applied a causal discovery approach, linear non-Gaussian acyclic causal model, to identify directed functional connectivity among the frontoparietal motor-related brain regions for each condition. We demonstrated higher directed functional connectivity from the contralateral dorsal premotor cortex (dPMC) to the primary motor cortex (M1) in ME than in MI. We mainly identified significant direct effects of the dPMC and ventral premotor cortex (vPMC) to the parietal regions. In particular, connectivity from the dPMC to the superior parietal lobule in the same hemisphere showed significant positive effects across all conditions. Contrastingly, interlateral connectivities from the vPMC to the superior parietal lobule showed significantly negative effects across all conditions. Finally, we found positive effects from A1 to M1 in the same hemisphere, such as the audio-motor pathway. These results indicated that the sources of motor command originated in d/vPMC influenced M1 and parietal regions as achieving ME and MI. Additionally, sequential sounds may functionally facilitate temporal motor processes.


Abstract:
1 Both imagery and execution of motor controls consist of interactions within a 2 neuronal network, including frontal motor-related regions and posterior parietal 3 regions. To reveal neural representation in the frontoparietal motor network, two 4 approaches have been proposed thus far: one is decoding of actions/modes 5 related to motor control from the spatial pattern of brain activity; another is to 6 estimate directed functional connectivity, which means a directed association 7 between two brain regions within motor regions. However, directed connectivity 8 among multiple regions of the motor network during motor imagery (MI) or 9 motor execution (ME) has not been investigated. Here, we attempted to 10 characterize the directed functional connectivity within the frontoparietal motor-11 related networks between the MI and ME conditions. We developed a delayed 12 sequential movement and imagery task to evoke brain activity associated with 13 data under ME and MI via functional magnetic resonance imaging scanning. We 14 applied a causal discovery approach, linear non-Gaussian acyclic causal model, 15 to identify directed functional connectivity among the frontoparietal motor-16 related brain regions for each condition. We demonstrated higher directed 17 functional connectivity from the contralateral dorsal premotor cortex (dPMC) to 18 the primary motor cortex (M1) in ME than in MI. We mainly identified significant 19 direct effects of the dPMC and ventral premotor cortex (vPMC) to the parietal 20 regions. In particular, connectivity from the dPMC to the superior parietal lobule 21 in the same hemisphere showed significant positive effects across all conditions. 22 Contrastingly, interlateral connectivities from the vPMC to the superior parietal 23 1. Introduction 1 Achievements of motor imagery (MI) and motor execution (ME) consist of 2 interactions within a neuronal network, including frontal motor regions and 3 posterior parietal regions. According to previous studies, MI is defined as a 4 mental simulation or a mental rehearsal of movements without any overt 5 physical movements (Hanakawa et al., 2008;Lorey et al., 2014;Pilgramm et al., 6 2016). Neural activities during MI can be predominantly modulated by tasks 7 containing visual, auditory, and kinesthetic aspects (Hanakawa, 2016). Neural 8 substrates of the motor system have common and specific representations 9 corresponding to MI or ME. 10 11 Previous studies have reported that information associated with MI can be 12 We followed a preprocessing protocol from a previous study (Ogawa et al., 7 2018). The data were processed using SPM12 (Wellcome Trust Centre for 8 Neuroimaging). The first four volumes were discarded to allow for T1 9 equilibration. The remaining data were corrected for slice timing and realigned 10 to the mean image of that sequence to compensate for head motion. Next, the 11 structural image was co-registered to the mean functional image and 12 segmented into three tissue classes in the MNI space. Using associated 13 parameters, the functional images were normalized and resampled in a 2 × 2 × 14 2 mm grid. Finally, they were spatially smoothed using an isotropic Gaussian 15 kernel of 8 mm full-width at half maximum. 16 17

Regions of interest 18
We followed a previous study (Zabicki et al., 2017) which used sixteen brain 19 regions and added bilateral dorsolateral prefrontal cortex (DLPFC) as control 20 regions. Therefore, we defined eighteen ROIs (Table 1)  residual or external input has non-Gaussian distributions, and 3) the 23 corresponding directed dependency graph is acyclic, i.e., the nodes have a 1 natural ordering. Letting xi denote the i-th observation variable (i = 1, …, n), its 2 linear association to the preceding/parent variable can be represented as 3 where bij is the connection weight, quantifying the direct causal effect, and k (i)

Statistics of network strength 21
To evaluate the stability of directed functional connectivity across the subjects, 22 we applied the bootstrapping method to estimate the lower and upper boundary 23 of the strength of the connections with 100 repetitions following a method 1 described by Xu et al. (2014). We assumed that the fMRI data, x, can be 2 linearly represented by Eq. (2). To check the validity of the DirectLiNGAM 3 approach, we confirmed non-Gaussianity of innovations, e = (I−B) x, of the 4 estimated model by the Kolmogorov-Smirnov test (p < 0.05, "kstest.m", in 5 Statistics toolbox, MATLAB). We then estimated 100 models for each condition 6 (R-ME, R-MI, L-ME, and L-MI) and applied the bootstrapped t-test for each 7 condition (t-test, p < 0.001 corrected by the Bonferroni method). Additionally, to 8 evaluate statistical difference of the direct effects between the two conditions, 9 we applied the two-sample t-test (p < 0.001 corrected by Bonferroni method). images were high-pass filtered with a 128 s cutoff period to remove the effect of 17 low-signal drift. Additionally, the six head movement parameters derived from 18 the realignment procedure were used as regressors of no interest. fMRI data of 19 each session were modeled with two regressors of interest (ME and MI) 20 corresponding to the task blocks included in the session. All brain activity during 21 the 7 s of the IP was included in the analysis. To identify the activated brain 22 areas during the MI and ME conditions, contrast images were calculated 23 including (i) R-ME > L-ME; (ii) L-ME > R-ME; (iii) R-MI > L-MI; (iv) L-MI > R-MI; 1 (v) R-ME > R-MI; (vi) R-MI > R-ME; (vii) L-ME > L-MI; (viii) L-MI > L-ME. The 2 contrast images were used as the input data in group-level random-effect 3 analyses using one-sample t-tests. The threshold for statistical significance was 4 set at an uncorrected p-value < 0.001 with a cluster-based family-wise-error 5 correction of p-value < 0.05. 6 7

Directed functional connectivity on fMRI data 9
Applying the DirectLiNGAM algorithm, we estimated the weight coefficient 10 matrices, B, from data x for each condition (R-ME, R-MI, L-ME, L-MI). By 11 shuffling the subjects' order with 100 repetitions for each condition, we 12 illustrated colormaps of the mean weights over repetitions, such that only 13 significant ones were shown (Fig.3A). According to the colormaps, we drew 14 DAGs to visualize directed effects between two regions (Fig.3B). To indicate 15 graphs of the network, we counted number of edges a node has to or from other 16 nodes as a measure of node importance in the network, (i.e., out-degree refers 17 to the number of edges from the node to other nodes, and in-degree refers to 18 the number of the edges from other nodes to the node) in Fig.3CD. 19 20 First, brain regions in the bilateral d/vPMC showed high out-degrees and low in-21 degrees. This indicates that the bilateral d/vPMC strongly influences other 22 regions, but direct effects from other regions are small (out-degree > in-degree, 23  Table  3 1). We identified positive direct effects from the dPMC to M1 ipsilateral to the 4 hand (R-ME and R-MI: R-dPMC  R-M1; L-ME and L-MI: L-dPMC  L-M1). As 5 a bilateral interaction, we found positive direct effects (all conditions: L-vPMC  6 R-vPMC; L-ME, R-ME, L-MI: L-dPMC  R-dPMC). Additionally, we found 7 strong negative direct effects from the L-dPMC to R-M1 in R-ME, but also weak 8 negative effects in the other three conditions (all conditions: L-dPMC  R-M1). 9 Most of positive direct effects in the same hemisphere showed symmetric 10 structure; however, the interhemispheric directed functional connectivity with 11 negative effects showed inhibitions in the target region and an asymmetric 12 structure. 13 14 Second, we also found that brain regions in the bilateral A1 regions showed 15 high out-degrees and low in-degrees ( Fig. 3C and 3D). In the colormap of the B 16 matrix and DAG, we found a significant positive effect, e.g., in the audio-motor 17 network (L-ME and L-MI: R-A1  R-M1; R-ME and R-MI: L-A1  L-M1), and a 18 weak negative connection to the parietal regions. We added A1 into our 19 analysis as control regions based on a previous study (Zabicki et al., 2017), 20 which was not expected motor processes associated with MI and ME. However, 21 our results suggest that A1 may influence not only M1, but also parietal regions 22 similar to d/vPMC during the dSMI task, because we used a sequential pitch to 23 lead finger tapping smoothly for the participants. 1 2 Third, we found that in-and out-degree in the parietal regions were not 3 obviously different compared to that in the frontal, auditory, and control regions. 4 Connections in the frontoparietal network, in particular, from the d/vPMC to 5 SPL/IPS/IPL were identified. We found direct effects from the frontal regions 6 (d/vPMC) to parietal regions (SPL, IPL, IPL), for instance, L-ME (L-dPMC  L-7 Meanwhile, connectivity from the parietal regions to frontal regions was not 11 selected by DirectLiNGAM. The direct effects from the frontal to parietal regions 12 showed an asymmetric structure. In the parietal regions, the bilateral IPS seed 13 regions seemed to influence neighboring regions such as the IPL and SPL 14 across all conditions. Thus, to achieve MI and ME, direct effects may show 15 interactions within the frontoparietal network and intraparietal network. 16 17 Regarding the usage of bilateral FMG and DLPFC as control regions, profiles of 18 their in-/out-degrees were opposite to those of the d/vPMC and A1, such that 19 the in-degrees of these regions were higher than the out-degrees. Additionally, 20 directed connections from the FMG and DLPFC to frontal motor regions, the 21 parietal regions, and A1 were not selected by DirectLiNGAM (Fig.3A) across all 22 conditions. This result indicates that brain activity in the FMG and DLPFC were 23 not caused by the achievements of ME and MI. 1 2

Differences of direct effects due to modes and laterality 3
To identify the direct effects between the two conditions, we evaluated 4 differences between two samples (A: L-ME vs. L-MI, B: R-ME vs. R-MI, C: L-ME 5 vs. R-ME, and D: L-ME vs. L-MI in Fig.4) with a two-sample t-test (p < 0.001 6 corrected by Bonferroni method). Each element showed a difference between 7 two conditions of a direct effect which was displayed only if the difference was 8 statistically significant. First, we found higher effects from the dPMC to M1 9 contralateral to the hand (green dotted boxes in Fig. 4C and 4D). Second, 10 frontoparietal connectivity found L-dPMC  R-IPS and R-vPMC  R-IPS in L-11 ME, and R-dPMC  R-IPS and L-vPMC  R-IPS in R-ME (orange dotted 12 boxes). 13 14 Obvious differences in directed functional connectivity are represented in the 15 edges dPMC  M1 and A1  M1, contralateral to the hand ( Fig. 4A and 4B). 16 The direct effects, dPMC  M1, are more strongly affected in ME than in MI. 17 This edge is essential to achieve motor control and related imagery. Second, 18 the strength of the edges, A1  M1, is also higher in ME and MI. It may 19 represent a clear neural representation in both regions in the case of ME. 20 21 Comparing the laterality of ME (Fig. 4C) and MI (Fig. 4D), the edges dPMC  22 M1 contralateral to the hand clearly show differences in laterality. In comparison 23 to ME, we identified edges L-M1  R-IPL, R-dPMC  R-IPS larger in R-ME, 1 and edges R-IPS  R-IPL in L-ME (Fig.4C). In the case of MI, we identified 2 edges L-M1  L-SPL, R-M1  L-FMG, R-IPL  R-DLPFC larger in R-MI, and 3 edges, L-IPL  L-DLPFC, R-A1  R-M1 larger in L-MI (Fig. 4D). To confirm brain activations between left and right ME/MI, we performed GLM 7 analysis for comparisons of contrasts. First, we compared contrasts between R-8 ME and L-ME and summarized statistical results in Table 2. The brain regions 9 that emerged from a contrast of R-ME > L-ME included the postcentral gyrus These results were consistent with previous findings that brain activity in the 17 contralateral hemisphere is predominantly higher than that in the ipsilateral 18 hemisphere. Therefore, the clusters were symmetrically colocalized in 19 comparisons between lateralities (Hanakawa et al., 2008). 20 21 Next, we compared contrasts between ME and MI such as modes for each 22 hand ( Table 3). The brain regions that emerged from a contrast of R-ME > R-MI 23 included the left postcentral gyrus (L-M1), which means that L-M1 is more 1 activated at R-ME than R-MI. Similarly, we identified clusters with a contrast of 2 L-ME > L-MI including those in the right precentral (R-M1) and right thalamus. corresponding to ME in M1 was higher than that of MI, but brain activity 8 corresponding MI was lateralized in the left hemisphere. 9 10

11
In this study, we demonstrated the asymmetric representation of the direct 12 functional connectivity during ME and MI. By applying DirectLiNGAM and GLM 13 to the fMRI data, we characterized functional asymmetry in the d/vPMC and A1, 14 parietal regions, and prefrontal control regions. We found higher direct 15 functional connectivity from the contralateral dPMC to M1 compared to that from 16 the ipsilateral dPMC to M1, as well as activations in M1. Regarding the following 17 of finger tapping with beep sounds, A1 showed high out-degrees that affected 18 M1 predominantly, as well as other regions. Regarding the parietal regions, 19 strong connections were observed mainly from the d/vPMC, as well as intra-20 connections. The frontal control regions were affected by other regions, but not 21 vice versa. However, higher activations during MI were distributed in the left 22 hemisphere than ME in both cases of the right and left hand, such that MI 23 showed functional asymmetry. Here, we would like to discuss the network 1 structure with directed functional connectivity and its brain activations across 2 hands and modes. 3 4

Functional roles of d/vPMC evidenced by activation patterns and 5 causal discovery during ME and MI 6
Our results of the GLM in Tables 2 and 3 showed that the brain activity is 7 generally contralateral to the movement/imagery side, and that brain activity in 8 M1 during ME was higher than that during MI. They were consistent with 9 findings from previous neuroimaging studies ( Kawashima  demonstrated that the out-degrees of the dPMC contralateral to the performing 18 hand were higher than the in-degrees regardless of ME/MI tasks with right/left 19 hand. Our results also showed that the out-degrees of dPMC were higher than 20 the in-degrees. This indicates that dPMC is one of the generators of neural 21 representation associated with ME/MI, not just in M1 but also in the parietal 22 regions. 23 1 We also found positive effects, dPMC  M1, ipsilateral to hand (Supp. Table 1).  2 Our task requires the participants to tap or imagine tapping with one hand; 3 therefore, we did not expect to identify the ipsilateral connections between the 4 dPMC and M1 in this analysis. Comparisons the right and left hemispheres, the 5 contralateral connections were higher than the ipsilateral ones in both ME and 6 MI.

Audio-motor network during ME and MI 17
Regarding the audio-motor network, we found i) high out-degree and low in-18 degree, ii) positive effects, A1  M1 bilaterally, and iii) negative effect, A1  19 parietal regions, in particular, connections within the same hemisphere. We did 20 not expect to find connections related to A1 when we designed this study 21 because we assumed that A1 was a control region. The decoding accuracies of 22 modality (ME vs. MI) in A1 obtained by multi-voxel pattern analysis were 23 significant in a study by Zabicki et al. (2017). It suggests that the voxel pattern 1 in A1 contains information that can distinguish between ME and MI. Lima et al. Additionally, regarding the comparisons of brain activity, the left parietal regions 21 had significantly greater activation during MI compared to ME (Table 2 and 3).

Limitation of DirectLiNGAM 1
The DirectLiNGAM algorithm is a suitable approach to explore the causal 2 relationships between multiple nodes (Shimizu et al., 2011). The conventional 3 methods, such as DCM or GCM, requires large computational costs in treating 4 four or more time series. One main limitation is that if brain regions A and B are 5 both driven by region C but with a different lag, an effective connectivity will be 6 shown an edge between A and B in the DCM. To address this problem, the 7 GCM method has been proposed and applied to field potential data from 8 macaque monkeys when performing a go/no-go visual pattern discrimination

Conclusions
6 By applying the causal discovery approach to the fMRI data, we demonstrated 7 that DAG represented the frontoparietal motor network underlying ME and MI. 8 To support the estimation of the directed functional connectivity, we also 9 confirmed spatial brain activity with GLM during ME and MI. We developed a 10 dSMI task based on that described by Hanakawa et al. (2008) to acquire 11 enough fMRI data samples under the ME and MI conditions. We found 12 predominant directed functional connectivity, d/vPMC  M1, which is a core of 13 the motor control. Additionally, we identified audio-motor connections from A1. 14 The parietal regions and DLPFC showed functional asymmetry during ME and 15 MI. The parietal regions may not only receive movement related information 16 from d/vPMC but may also communicate with neighboring regions. MI can 17 cause higher cognitive load than ME. Our approach has the limitations of

Results for the contrasts between ME and MI [R-ME > R-MI], [R-MI > R-ME], 2
[L-ME > L-MI], and [L-MI > L-ME]. 3 MNI coordinate is corresponding to the peak voxels within each cluster. 4 Clusters were set at a threshold of p < 0.001 and cluster level family-wise-error 5 at p < 0.05. Note: Brain regions were labelled by the AAL toolbox in the SPM 6