Functional MRI Reveals Locomotion-Control Neural Circuits in Human Brainstem

The cuneiform nucleus (CN) and the pedunculopontine nucleus (PPN) in the midbrain control coordinated locomotion in vertebrates, but whether similar mechanisms exist in humans remain to be elucidated. Using functional magnetic resonance imaging, we found that simulated gait evoked activations in the CN, PPN, and other brainstem regions in humans. Brain networks were constructed for each condition using functional connectivity. Bilateral CN–PPN and the four pons–medulla regions constituted two separate modules under all motor conditions, presenting two brainstem functional units for locomotion control. Outside- and inside-brainstem nodes were connected more densely although the links between the two groups were sparse. Functional connectivity and network analysis revealed the role of brainstem circuits in dual-task walking and walking automaticity. Together, our findings indicate that the CN, PPN, and other brainstem regions participate in locomotion control in humans.


Fig. S1. Masks for coregistration and activation analysis.
The coregistration mask was modified from the mask in [1] and incorporated the brainstem and nearby CSF (from z = -3 to z = -62). This mask weighed the cost function in the affine coregistration to improve brainstem coregistration and normalization. With reference of the coregistration mask, the activation mask was created by cutting off outside-brainstem areas and voxels in the border of the brainstem (from z = -10 to z = -59). L R L R  The numbers indicate the MNI coordinates in the x-axis for the sagittal plane or in the z-axis for the axial plane. (A) The red nucleus is shown in violet and projected onto the mean T1 anatomical image. The scope of the red nucleus was manually defined on the basis of the mean functional images of the rest condition, with examples shown in (E). The iron-rich red nucleus has higher magnetic susceptibility than many gray matter nuclei, such as the dentate nucleus and putamen [2] , and present a darker area than surrounding tissues in T2* images as shown in the leftmost slice in (E). A globular shape and a clear boundary of the red nucleus therefore lead to little ambiguity in determining its proper scope [3] . (B) Spatial relation between the red nucleus (in violet) and the cerebello-rubral tract (blue to black) of the superior cerebellar peduncle that runs from a cerebellar hemisphere to the opposite red nucleus, as a validation of the defined scope of the red nucleus. The cerebello-rubral tract is presented using the human probabilistic atlas based on the connectome imaging data [4] .
(C) The scope of the CN (in red) was defined on the basis of the anatomical information [5] , projected onto the mean T1 anatomical image of all subjects. (D) The scope of the PPN (in blue) was also defined on the basis of the anatomical information [5] , projected onto the mean T1 anatomical image of all subjects. (E) The scopes of the CN (red) and PP N (blue) were presented onto the mean functional image of the rest condition from all subjects. All slices containing either of the two ROIs are presented. Note that the shape of each ROI changes due to a lower spatial resolution of the mean functional image in comparison with the high-resolution T1 image. (F) Midline activations in the pons (violet) and medulla oblongata (green) mainly correspond to the raphe nuclei on the basis of [5,6] . The width was arbitrarily set to 3 mm when defined on the T1 images (i.e., the midline plus both of its sides) and automatically transformed to 2 mm when projected on the functional images due to a lower spatial resolution of these images. Lateral activated regions (white) in the ponsmedulla junction zone mainly cover subgroups of the nucleus reticularis on the basis of [5,6] .
(G) The raphe nuclei in the pons and medulla are shown in the sagittal plane in violet and green, respectively. (H) The flowchart defining the scope of the CN and PPN is shown. The histochemically defined atlas consisted of serial axial sections in an axial plane perpendicular to the long axis of the brainstem. The mean T1 image (in the MNI space) was first rotated to the same spatial orientation as the atlas. Afterward, each section of the atlas containing the PPN and/or CN was matched to a corresponding slice of the T1 images on the basis of anatomical landmarks (I). An affine transform was then applied to the section to match the corresponding slice. The scope of the CN/PPN on the MRI slice was defined on the basis of the transformed section. The scopes from all slices constituted an ROI of CN/PPN, which was rotated to the MNI space. (I) Example landmarks are presented on the axial plane (after rotation; that is, each axial slice is perpendicular to the long axis of the brainstem). The landmarks were used to determine which section of the histochemically defined atlas matched to a slice of the T1 image. Red nucleus activations were found in all rhythmic bipedal movement conditions (0.5 Hz, 1 Hz, 2 Hz, and changing speed, all of which use audio cues to pace movement) but not in the free-speed condition (no rhythmic audio cues to pace movement). Functional connectivities between the red nucleus and the cerebellum were detected in 1 Hz, 2 Hz, and changing-speed tasks but not in the 0.5 Hz and free-speed conditions (Fig. 3). The 0.5 Hz (very slow speed) and free-speed (no rhythmic audio cues) tasks should not be fine-tuned movements requiring strong cerebellum control. We thus propose that such functional connectivities indicate a role of the red nucleus in coordinating ankle movements together with the cerebellum for faster (1 and 2 Hz vs. 0.5 Hz) or more complex (changing-speed vs. free-speed) bipedal movements via the rubrocerebellar pathways (e.g. the dentatorubrothalamic tract) [7] .  Fig. S2. (B) Brainstem activations with different Gaussian kernels are shown. Data of the changing-speed condition were smoothed with 3, 4, and 5 mm kernels. The activation maps are thresholded under the p < 0.05 FWE-corrected cluster extent provided by the SPM toolbox, with a primary threshold of p < 0.005. The 5 mm kernel leads to larger activation clusters and therefore appears to be more sensitive. The sensitivity of combining the same normalization method [1] and a similar smooth kernel (4.5 mm) was also proven by [8] .    The nodes in the same module are shown in the same color. The nodes/links are sized according to their strength/weight values. Layouts are generated using the Kamada-Kawai force-spring layout algorithm (http://pajek.imfm.si). The visualization method defines the geometric distance between nodes in a network as the summation of edge weights. Every pair of nodes is connected by a "spring" with a desirable length to reach a state, in which the total spring energy of the network is minimal. The final balanced condition is formulated as the square summation of the differences between desirable distances and real ones for all pairs of vertices.
In the resting state, the topmost nodes constitute a module, and more caudally located nodes, including brainstem regions, constitute a module. During task conditions, the former extends to lower brain regions, such as at least one of two cerebellar nodes, whereas the brainstem nodes constitute two separate modules, that is, a CN-PPN (midbrain) group and a pons-medulla group.

Fig. S7. Mismatch between gait speed and edge numbers.
At constant audio-paced speeds (0.5, 1, and 2 Hz, highlighted with frames), the edge number of the entire network, the inside-brainstem subnetwork, and the outside-brainstem subnetwork increased gradually. However, fewer edges were present between the inside-and outside-brainstem subnetworks in the 2 Hz condition than those in the 1 Hz condition, presenting a mismatch between the behavior (faster speed) and the network feature (less interactions between brainstem and cortical structures). Such mismatch is also seen when comparing the 1 Hz and free-speed conditions (highlighted with arrows).