Recent advancement of resting-state functional connectivity magnetic resonance imaging (MRI) has provided a method for drawing boundaries of brain areas. However, it remains to be elucidated how the parcellated areas in the association cortex relate to the spatial extent of the brain activation which ought to reflect a functional unit in the neural network supporting that particular function. To address this issue, in the present study, we first mapped boundaries and 2 adjacent activations in the human inferior frontal cortex, and then examined the spatial relationship between the boundaries and the 2 activations. The boundaries mapped with high-resolution functional magnetic resonance imaging revealed a collection of micromodules, the size of which was approximately only 12 mm on average, much smaller than the Brodmann areas. Each of the 2 activations associated with 2 functions, response inhibition and feedback processing, was smaller in size than the micromodules. By comparing the spatial patterns between the boundaries and the 2 activations, it was revealed that the brain activations were less likely to be located on the boundaries. These results suggest the functional relevance of the areas in the association cortex delineated by the boundary mapping method based on resting-state functional connectivity MRI.
Although the neural mechanisms for low-frequency fluctuations in the blood oxygenation level-dependent (BOLD) signal during resting state are not fully understood, the use of functional connectivity based on the low-frequency fluctuations in the BOLD signal has revealed a robust interregional interaction between distant brain regions that constitute large-scale regional networks (Biswal et al. 1995; Greicius et al. 2003, 2004; Leopold et al. 2003; Fox et al. 2005, 2009; Damoiseaux et al. 2006; Fair et al. 2007; Vincent et al. 2007; Buckner et al. 2009; Bullmore and Sporns 2009; Honey et al. 2009; Krienen and Buckner 2009; Smith et al. 2009; Schölvinck et al. 2010; Van Dijk et al. 2010; Zhang and Li 2010; Hutchison et al. 2011, 2012; Power et al. 2011). Prior studies of structural connectivity measured with diffusion tensor imaging have analyzed the voxel-wise relationship in the whole brain, which was used to classify the voxels into distinct regions (Behrens et al. 2003; Johansen-Berg et al. 2004; Anwander et al. 2007; Klein et al. 2007; Mars et al. 2011). Recent advancement of resting-state functional connectivity magnetic resonance imaging (rs-fcMRI) has provided the method of delineating areal boundaries based on the spatial pattern of connectivity between each voxel and the rest of the brain (Margulies et al. 2007; Cohen et al. 2008; Biswal et al. 2010; Nelson et al. 2010; Hirose, Watanabe et al. 2012; Zhang et al. 2012). Successful applications of the boundary mapping method have been reported in several areas such as the cingulate cortex, the supramarginal/angular gyrus, the lateral parietal cortex, the inferior frontal cortex (IFC), the medial superior frontal cortex, and various regions in the whole brain (Margulies et al. 2007; Cohen et al. 2008; Biswal et al. 2010; Nelson et al. 2010; Hirose, Watanabe et al. 2012; Zhang et al. 2012).
Despite the successful delineation of areal boundaries using rs-fcMRI, it is not clear whether the delineated areas in the association cortex reflect functional units in the neural network that implement particular functions. Although early sensory and motor areas are organized along several disciplines such as somatotopy (Felleman and van Essen 1991), the functional organization in the association cortex is much less understood. One feasible landmark for a functional unit in the association cortex would be the brain activation, which ought to reflect a neural substrate for information processing in the neural network. A recent study where the boundary mapping method was applied to the lateral parietal cortex (Nelson et al. 2010) has shown that different areas in the parietal cortex had distinct temporal timecourse profiles of the brain activity. However, the relationship between the delineated areas and the spatial extent of the brain activations remains to be elucidated.
To examine the spatial relationship in more detail, in the present study, we utilized high-resolution functional magnetic resonance imaging (fMRI), typically with the spatial resolution of 2 mm or less (Kim et al. 2000; Zeineh et al. 2001; Moon et al. 2007; Yacoub et al. 2008; Carr et al. 2010). The higher resolution is intended to reveal micromodules, regions delineated based on rs-fcMRI, without signal contamination from the other bank of a sulcus. To compensate for low signal-to-noise ratio caused by the small voxels, a large number of volumes were collected for each individual subject. The boundary mapping method was applied to the posterior part of the inferior frontal cortex (pIFC), which is known to be activated during both response inhibition (RI) and feedback processing (FP), with the 2 activations associated with the 2 functions separated by only approximately 1 cm (Hirose et al. 2009). This situation provides a unique opportunity of examining 2 activations in a small field of view in detail, overlaying the boundary on to the activation maps. By combining the boundary mapping method and high-resolution fMRI, we delineated the micromodules, and compared them with the 2 activations in the right pIFC.
Materials and Methods
Informed consent was obtained from 6 healthy right-handed subjects (3 males; 3 females, age: 20–25 years). They were scanned by fMRI using experimental procedures approved by the institutional review board of the University of Tokyo School of Medicine.
Imaging data were collected at a 3-T MRI system (Philips Achieva X 3T Rel. 2.6, Best, The Netherland). T1-weighted images were obtained for anatomical reference [repetition time (TR) = 6.2s, echo time (TE) = 3.0msec, 120 slices, slice thickness = 1.2 mm, in-plane resolution = 1.2 × 1.2 mm]. For functional imaging of the resting state, gradient echo-planar sequences were used (TR = 6.0sec, TE = 35msec, flip angle = 90°, cubic voxel of 2 mm, 34 slices without a slice gap that covered the pIFC). The data were sampled using the cubic voxels of 2 mm to minimize signal contamination from the other bank of the sulcus (Kim et al. 2000; Zeineh et al. 2001; Moon et al. 2007; Yacoub et al. 2008; Carr et al. 2010). Sixty-four runs were collected for each of the subjects, and each run contained 54 volumes (later excluding the first 4 volumes), to compensate for lower signal-to-noise ratio caused by the smaller voxels. In the resting-state experiment, subjects were instructed to fixate on a cross presented at the center of the screen. Functional imaging was also conducted for the investigation of the task-related brain activation with the same spatial resolution (TR = 3.0s, TE = 35msec, flip angle = 90°, 34 slices without the slice gap that covered the pIFC). Sixteen to 52 runs were collected, and each run contained 64 volume images, later excluding the first 4 images. The total imaging time for the resting-state images and task-related images in each subject amounted to 397–512min (interrun intervals not included).
Preprocessing of Resting-State Data
Functional images were realigned and were slice-timing corrected using SPM2. Neither spatial normalization nor spatial smoothing was conducted to keep the higher spatial resolution. The following preprocessing is essentially the same as that used in previous literatures on rs-fcMRI (Fox et al. 2005; Fair et al. 2007). Temporal band-pass filtering (0.009 Hz < f < 0.08 Hz) was applied using FSL (Smith et al. 2004), and a general linear model (GLM; Worsley and Friston 1995; Miezin et al. 2000) was used to regress out nuisance signals that correlated with head motion, whole-brain signal, averaged ventricular signal, and averaged white matter signal.
Generation of Probabilistic Boundary/Center Maps
The probabilistic boundary maps were generated based on the boundary mapping method (Cohen et al. 2008; Nelson et al. 2010; Hirose, Watanabe et al. 2012). The right pIFC was flattened into the 2D space using Caret (http://brainmap.wustl.edu/caret; Van Essen et al. 2001). Each pixel in the 2D space was used as the seed to calculate a correlation with the target voxels in the 3D space (Fig. 1A). Rather than calculating correlations with all the voxels in the whole brain as is done in the original method by Cohen et al. (2008), the target voxels were restricted, in the present study, to only those in the contralateral region (Hirose, Watanabe et al. 2012) (Fig. 1B). This modification of the original method is based on the observation that a region has the strongest functional connectivity with the corresponding contralateral region (Stark et al. 2008) and is shown to be efficient in a signal to noise ratio, with a modest calculation time (Hirose, Watanabe et al. 2012). For each seed voxel (Fig. 1B, left), a sphere (radius: 2 mm) was generated in the contralateral region, and the spheres were combined to form a collection of target voxels (Fig. 1B, right). A correlation map was generated by applying the Fisher's z transformation to the correlation coefficient (Fox et al. 2005; Fair et al. 2007).
The procedures after generation of correlation maps are basically the same as those described in the previous studies (Cohen et al. 2008; Nelson et al. 2010). Briefly, the pixels where the spatial pattern of the correlation maps changes abruptly are thought to represent the boundaries between functional areas. To quantify the change of the spatial pattern of the correlation maps, the similarity of the correlation maps was estimated between the seeds by calculating η2. The Canny edge detection algorithm (Canny 1986) was applied to η2 maps to generate a gradient map and then to detect edges (Fig. 1A). Averaging across the entire set of binary edge maps generates a probabilistic boundary map in which intensity represents how likely a location is to be an edge. In the present study, the probabilistic map was also generated for the centers, besides the boundaries. The local minima in the gradient map were detected, and the resultant binary local minimum maps were averaged to generate a probabilistic center map.
Estimation of the Distance Between the Centers of the Adjacent Micromodules
The 2-dimensional autocorrelation function, as implemented in the Signal Processing Toolbox of the MATLAB software suite, was applied to the probabilistic center map after spatial smoothing ([full width at half maximum (FWHM) = 4 mm]. To account for possible variation of the autocorrelation results across the central positions of a 50 × 50 mm sliding window, the probabilistic center map was first generated in 80 × 80 mm space, which extended maximally within the brain regions surrounding the right pIFC. The window of 50 × 50 mm slid within the space, and the autocorrelation was computed repeatedly for each window position and was averaged across the positions. The autocorrelation value was further clumped along the same radial coordinates to neglect the directions.
The go/no-go task (Chikazoe et al. 2009) and the modified version of the Wisconsin Card Sorting Task (WCST; Konishi et al. 2002) were also used for mapping brain activations associated with RI (Garavan et al. 1999; Konishi et al. 1999; Bunge et al. 2002; Aron and Poldrack 2006; Li et al. 2006; Brass and Haggard 2007; Zheng et al. 2008) and FP (Monchi et al. 2001; Jimura et al. 2004), respectively (Hirose et al. 2009). Briefly, in the go/no-go task, 3 types of trials were presented: The frequent-go, the infrequent-go, and the no-go trials. In the frequent-go trial (75.0%), a gray circle was centrally presented, and the subjects had to press a button using a right thumb. In the infrequent-go trial (12.5%), a blue circle was presented, and the subjects had to press a button using a right thumb, whereas in the no-go trial (12.5%), a green circle was presented, and they had to withhold the response. The colors of the 2 types of trials (blue and green) were counterbalanced across subjects.
In the modified WCST, a 5-card array was displayed until the subjects selected 1 of the 4 reference card stimuli placed at the corner of the screen. They pressed 1 of the 4 buttons using the right thumb by matching an attribute of a central card to 1 of the 4 references on the basis of the dimension of color, form, or number. A feedback stimulus (correct: O, incorrect: X) was then presented. The modified WCST included shift events with and without negative FP. The shifts were made from one dimension (color, form, or number) to another after 5–8 successive correct trials in the former dimension block. At the shift events with negative FP, the negative feedback stimulus (X) was presented, and then a subsequent dimension was instructed by the visual presentation of the word “color”, “form,” or “number.”
Analysis of Brain Activation Data
Similarly to the resting-state data, functional images were realigned and were slice-timing corrected, but neither spatial normalization nor spatial smoothing was applied. Event timing was then coded into a GLM. To extract the brain activation associated with RI, the no-go minus infrequent-go trials was calculated (Chikazoe et al. 2009). For the FP, the shift events with negative FP were compared with the shift events without negative FP (Konishi et al. 2002; Jimura et al. 2004).
To evaluate the significance level of the activation of individual subjects, a small volume correction for spherical ROIs (radius: 1 cm) was applied, using the coordinates determined in our previous studies of RI (52, 12, 20) and FP (50, 18, 26; Hirose et al. 2009). To provide the coordinates in the Talairach space (Talairach and Tournoux 1988), the spatial normalization using the standard template image was applied in a separate analysis. The activations were significant at P < 0.05, for each of the 6 subjects and each of the 2 contrasts.
Boundaries and Brain Activations
To make a quantitative comparison of probability with which brain activations exist inside/outside the boundaries, we defined the Boundary pixels and Background pixels. The Boundary pixels were on the ridge of the probabilistic boundaries, whereas Background pixels were off that ridge. To determine the Boundary pixels, the direction of any one pixel was first determined out of 4 directions by summing up the pixel values of 3 linear pixels (the pixel itself and 2 adjacent pixels) along each of the 4 directions and selecting a direction with the greatest value. The pixel was regarded as a Boundary pixel: 1) when the pixel value was the greatest of the 3 linear pixels (the pixel itself and the 2 adjacent pixels) along the line orthogonal to the selected direction, and 2) when such pixels were contiguous by 3 or more. All the pixels other than the Boundary pixels were regarded as Background pixels. On the other hand, a pixel was regarded as a Center pixel when the pixel was within a circle (diameter = 3 mm) surrounding local maxima of the probabilistic center map. All the pixels other than the Center pixels were regarded as Surround pixels. As for the brain activation, a pixel was defined as activated when a pixel exceeded t = 2.0 with 3 or more contiguous pixels.
Effect of Vascular Anatomy
To examine the effect of vascular anatomy on the periodic patterns of the probabilistic boundaries, 2 analyses were conducted. First, the susceptibility-weighted imaging (PRESTO, TR = 18msec, TE = 29msec, 160 slices, slice thickness = 0.5 mm, in-plane resolution = 0.5 × 0.5 mm), that was invented to measure small veins (Reichenbach et al. 1998), was conducted in 6 subjects (3 males; 3 females, age: 21–27 years). Out of the 6 subjects who participated in the resting-state scans, only 1 subject was available and participated in vascular imaging. The vascular images were subject to the procedures applied to the resting-state images, including flattening into the 2D space using Caret, spatial smoothing in a 2D space (FWHM = 4 mm), and the autocorrelation analysis. Secondly, the resting-state images were analyzed in a voxel-by-voxel basis to calculate the standard deviation (SD) in 1 run, and the SD images were averaged across the 64 runs for each subject. The SD images were similarly subject to the procedures applied to the resting-state images.
Size of Micromodules
We applied the boundary mapping method to the right pIFC (Fig. 1C) and generated the probabilistic center maps (Fig. 2A). Figure 2B shows a complementary relationship between the probabilistic center map and the probabilistic boundary map. The size of the micromodules was estimated by applying autocorrelation to the probabilistic center map in 6 subjects. The probabilistic center map, not boundary map, was used for the size estimation, because the probabilistic boundary map was generated using the Canny edge detection algorithm where the continuity of the boundaries was emphasized and was less suitable for the distance estimation. The autocorrelogram, as shown in Figure 2C, was further processed by averaging across the same radius to neglect the direction of the field of view. The result from the average of all 6 subjects is presented in Figure 2D. The peak of the curve was observed around 12 mm in radius. These results suggest that the overall pattern of probabilistic centers was separate between adjacent peaks by approximately 12 mm in the pIFC.
Effect of Vascular Anatomy
To test whether the periodic pattern of the probabilistic boundary maps was derived from the vascular anatomy, first, we collected susceptibility weighted images and conducted the autocorrelation analysis. As presented in Figure 2E,F, no peaks were detected. Secondly, we calculated the SD of the resting-state images and conducted the autocorrelation analysis. As presented in Figure 2G,H, no peaks were detected either. These results suggest that the periodic pattern of the probabilistic boundary maps cannot be explained by the vascular anatomy.
To confirm that the susceptibility-weighted images and the SD images commonly reflected a vascular component, the 2 images were compared on the one subject who participated in both the resting-state imaging and the susceptibility-weighted imaging. The vein regions were defined as clusters of 5 or more contiguous pixels above top 5% in the susceptibility-weighted images, and the resultant 21 clusters were placed on the SD images to test whether the cluster regions contained higher SD than the others. As indicated by the green dots in Figure 2G, the vein regions appeared to be located in higher SD pixels. A t-test of the SD using across-cluster variance revealed a significantly greater SD in the cluster regions (t(20) = 2.9, P = 0.008), confirming the reliability of the 2 imaging analyses.
Two Activations Associated with RI and FP
Functional data were also collected while subjects performed 2 different tasks: the go/no-go task (Chikazoe et al. 2009) and the modified WCST (Konishi et al. 2002) that are used to extract the brain activity associated with RI and FP, respectively (Fig. 3A). Table 1 presents a summary of the activation results. Significant activation was found in the right pIFC during RI and FP, and the results were consistent with our previous study (Hirose et al. 2009).
|Case||Response inhibition||Feedback processing||Distance (mm)|
|Case||Response inhibition||Feedback processing||Distance (mm)|
We then compared the spatial relationship between the activations associated with the 2 contrasts, RI and FP. Figure 3B shows the activation peaks in 6 subjects. The distance between the 2 peaks in the 6 subjects is listed in Table 1, and the distance ranged from 6.8 to 25.6 mm. Although the boundaries of the micromodules cannot necessarily be defined rigidly, the 2 activations appeared to belong to 2 adjacent micromodules, or appeared to be intervened by 1 or 2 micromodules. Although the across-subject variability was large, the average distance was 13.9 mm and was close to the estimated distance between the adjacent centers of the micromodules (∼12 mm), which allowed us to compare the size of the activation with that of the micromodule. To inspect the signals in-between the activation peaks, signal profiles were calculated, after normalizing the distances across the 6 subjects (Fig. 3C). The signals in-between the peaks were well below the t = 2 level (the dashed white line that indicates P = 0.05) and were close to the 0% signal baseline for the most part of the profile. These results suggest that the average size of the individual activations associated with RI and FP was smaller than that of the micromodule, approximately 5 mm.
Boundaries and Brain Activations
Figure 4A shows the comparison between the activations and the probabilistic boundaries in one subject. The activation shown in the overlaid map appeared to avoid the probabilistic boundaries. When the activation did overlap with the boundaries, the boundaries appeared to be low probabilistic ones. To quantify these features, pixels were classified into 2 categories (Boundary pixels and Background pixels) according to whether or not one pixel is on the ridge of the probabilistic boundaries. Note that the Boundary pixel did not depend on the absolute height of the boundary probability, but on the relative height among neighboring pixels (see Materials and Methods), so the Boundary pixels included both high- and low-probabilistic boundaries. We calculated the probability, with which activation existed in each of the Boundary and Background pixels. Figure 4B presents the probability of activation existence after averaging across the 2 tasks. There was a significant tendency for the activation to exist less likely in the Boundary pixels, as assessed by the 2-way repeated-measures analysis of variance (ANOVA) using tasks and Boundary/Background pixels as main effects. The main effect of pixels was significant (F1,5 = 11.5, P = 0.02), and neither the main effect of tasks nor the pixel-by-task interaction was significant.
We next tested whether the activation avoided the higher probabilistic boundaries to the greater degree than the lower probabilistic boundaries. The probability of activation existence was calculated further within the Boundary pixels, with the pixels of high- to low-probabilistic boundaries divided into 10 bins. Figure 4C shows the results of correlation between the probability of boundaries and the probability of activation existence in one subject. One-way repeated-measures ANOVA of the z-values of the correlation coefficients (after Fisher's transformation) using tasks as a main effect did not reveal a significant effect. It was further assessed whether the correlation is significantly smaller than zero, and the group analysis using a 1-sample t-test after averaging across tasks revealed significant anticorrelation (t(5) = 2.7, P = 0.04). These results indicate that the more robust the probabilistic boundaries, the more likely the activation tended to avoid that probabilistic boundaries and suggest that activation tends to belong to one micromodule.
We also investigated the probabilistic centers. Pixels were classified into 2 categories (Center pixels and Surround pixels) according to whether or not one pixel is near the local maxima of the probabilistic centers (see Materials and Methods). The difference between Center and Surround pixels was not significant but approached significance (F1,5 = 5.3, P = 0.07), likely reflecting the reverse side of the results from the Boundary/Background pixels. The probability of activation existence was calculated further in the Center pixels, with the pixels of high- to low-probabilistic boundaries divided into 10 bins. However, the correlation in the Center pixels between the probability value of the centers and the probability of activation existence was not significant (t(5) = 0.6, P = 0.6), suggesting that the activation does not necessarily stick to the high-probabilistic centers in the micromodule.
Behavioral Data and Spatial Pattern of Boundary/Activation
We finally examined the correlations between the behavioral data and the spatial pattern of the boundary/activation. In the go/no-go task, the efficiency index, which was invented to assess the ability to inhibit a response in the go/no-go task (Hirose, Chikazoe et al. 2012), was calculated as a behavioral measure based on the data set in our previous study (Hirose, Chikazoe et al. 2012). In the WCST, the difference between the reaction time at the dimension changes with negative feedback and the reaction time at the dimension changes without negative feedback, which reflects negative FP, was calculated as a measure. For the spatial pattern of the boundary/activation maps, 2 aspects of the maps were highlighted. First, the signal to noise ratio of probabilistic boundaries (Hirose, Watanabe et al. 2012) was calculated, and it was examined whether good performance correlated with the clear contrast of the probabilistic boundaries. Secondly, the difference in the probability of brain activation existence between the Boundary pixels and Background pixels was calculated, and it was examined whether good performance correlated with the clear distribution of activated pixels. The 2 × 2 combinations of correlation analyses revealed no significant correlations (P > 0.05, corrected for multiple comparisons).
The combination of the boundary mapping method and high-resolution fMRI revealed the micromodules in the right pIFC. The overall size of the micromodules was approximately 12 mm, as estimated by autocorrelation analyses. The size of the activations associated with RI and FP was, on average, smaller than that of the micromodules. The activation significantly tended to avoid the boundaries, particularly the higher probabilistic boundaries, suggesting that one activation tends to belong to one micromodule. These results suggest that the micromodules in the human association cortex reflect a local substrate that forms the neural network supporting a particular function.
Both the 2 activations associated with RI (54, 15, 25) and FP (47, 21, 27) appear to be located in the anterior part of Brodmann area 44 (Brodmann 1909), based on the Y coordinates of the cytoarchitectonic probabilistic map (Amunts et al. 1999, 2004; Petrides and Pandya 2002), suggesting that multiple functional areas exist within one Brodmann area. Indeed, the autocorrelation analyses in the present study revealed that the overall distance of adjacent centers of the micromodules was approximately 12 mm. The average size of Brodmann areas 44 and 45 (approximate pIFC) in one hemisphere is estimated to be approximately 22 cm2 (Maldjian et al. 2003; van Essen et al. 2006). Therefore, Brodmann areas 44 and 45 in one hemisphere may contain approximately, on average, 15–20 such micromodules.
The comparison of the micromodules and activations revealed that the activation tended to avoid high-probabilistic boundaries. Conversely, the activation also tended to exist in the centers, irrespective of its probability. One micromodule is not always surrounded by complete boundaries, and the centers are obviously more stably detected than the boundaries. More critically, the use of centers helps to exclude the boundaries that are less likely to be associated with one particular function. So it follows that the use of the centers in regions of interest analyses is more practical (e.g. Nelson et al. 2010), particularly when a part of the boundaries surrounding a micromodule is not very robust. It is unlikely, however, that delineated boundaries are simple artifacts that resulted from sulcus structure (Hirose, Watanabe et al. 2012) or vascular territories (Fig. 2). It seems also unlikely that the across-subject variance of the boundaries correlates with the individual traits, because previous studies have reported negative results on the correlation between the right inferior frontal activation and the impulsivity scores (Asahi et al. 2004; Hendrick et al. 2011).
Despite the robust reproducibility of the probabilistic boundaries when the data set is split into 2 halves (Hirose, Watanabe et al. 2012), the boundary patterns had large variance across subjects. Most of the previous studies of boundary mapping (Biswal et al. 2010; Nelson et al. 2010; Zhang et al. 2012) applied a group analysis and revealed regions parcellated in a relatively larger scale (∼2 cm or larger), when compared with those of the present study using high-resolution fMRI applied to individual subjects. It is possible that the previous studies that required assumptions about the size and number of parcellated regions were intended to highlight a common aspect of areal parcellation that exhibited across-subject reproducibility, missing smaller parcellated areas that were not reproducible across subjects but were reproducible across data sets in individual subjects (Hirose, Watanabe et al. 2012). This view appears to be supported by the relatively small activation in size measured in the present study (Fig. 3C) that was significant in individual subjects, but similarly had large across-subject variance (Fig. 3B).
The results were not as straightforward as a simple expectation that the activation completely coincides with the micromodules. Rather, the activation was, under the condition of the present study, smaller than the micromodule. One physiological explanation for possible multiple subregions (activations) within one micromodule would be that the subregions are activated by a certain common function under different parameters, such as RI using distinct effector modalities of hand, eye movement, and utterance. Although more investigation is required for full understanding of the physiological basis of the micromodules, the present study emphasizes the potential of functional connectivity to uncover macroscopic information about neural circuitry underlying cognition.
This work was supported by a Grant-in-Aid for Specially Promoted Research (19002010) to Y.M., a Grant-in-Aid for Scientific Research B (22300134) to S.K., Global COE Program “Integrative Life Science Based on the study of Biosignaling Mechanisms” (12601-A03) to Y.M. from the Ministry of Education, Culture, Sports, Science, and Technology, Japan, and a grant from Takeda Science Foundation. S.H. is supported by JSPS Fellowship (20-6285) for Young Scientists.
We thank Ms. Yuriko Suzuki for technical assistance. Conflict of Interest: None declared.