The laminar pattern of proprioceptive activation in human primary motor cortex

Abstract

The primary motor cortex (M1) is increasingly being recognized for its vital role in proprioceptive somatosensation. However, our current understanding of proprioceptive processing at the laminar scale is limited. Empirical findings in primates and rodents suggest a pronounced role of superficial cortical layers, but the involvement of deep layers has yet to be examined in humans. Submillimeter resolution functional magnetic resonance imaging (fMRI) has emerged in recent years, paving the way for studying layer-dependent activity in humans (laminar fMRI). In the present study, laminar fMRI was employed to investigate the influence of proprioceptive somatosensation on M1 deep layer activation using passive finger movements. Significant M1 deep layer activation was observed in response to proprioceptive stimulation across 10 healthy subjects using a vascular space occupancy (VASO)-sequence at 7 T. For further validation, two additional datasets were included which were obtained using a balanced steady-state free precession sequence with ultrahigh (0.3 mm) in-plane resolution, yielding converging results. These results were interpreted in the light of previous laminar fMRI studies and the active inference account of motor control. We propose that a considerable proportion of M1 deep layer activation is due to proprioceptive influence and that deep layers of M1 constitute a key component in proprioceptive circuits.

balanced-SSFP, cortical layers, laminar fMRI, primary motor cortex, VASO

Introduction

The primary motor cortex (M1) plays a key role in generating corticospinal output to facilitate voluntary movement, relying heavily on contextual information from other brain regions associated with limb position, limb kinematics, motor planning, etc. Beyond its well-known motor functions, M1 is increasingly recognized as a critical hub for processing proprioceptive somatosensation—the sense of body position and movement (Naito 2004; Hatsopoulos and Suminski 2011). Proprioceptive inputs to M1 originate from muscle spindles and joint receptors, providing essential feedback for seamless coordination of movement and spatial awareness (Proske and Gandevia 2012). As such, understanding how proprioception is processed in M1 has significant implications for motor control theories and clinical contexts, including neurorehabilitation and neurodegenerative diseases.

The present understanding of the circuitry underlying information passing in M1 is primarily derived from tracer studies in rodents and nonhuman primates (Fig. 1A). Briefly, somatosensory input (including proprioception and tactile sources) has consistently been reported to terminate in superficial depths of M1 (layers II to III and layer Va) (Shipp 2005; Mao et al. 2011; Adams et al. 2013; Petrof et al. 2015; Hooks 2017). The reciprocal connection to the primary somatosensory cortex (S1) also originates from these layers (Mao et al. 2011; Hooks 2017). Motor-related inputs from frontal and secondary motor areas may terminate more widely across the full cortical depth but appear to favor superficial layers (Barbas and Pandya 1987; Watanabe-Sawaguchi et al. 1991; Adams et al. 2013; Hooks 2017; Ninomiya et al. 2019). Importantly, these inputs are transmitted from superficial depths to pyramidal cells in layer V via strong interlaminar connections where they influence corticospinal and corticothalamic outputs (Weiler et al. 2008; Bastos et al. 2012; Shipp 2016). The majority of corticofugal output cells are located in deep layers Vb (corticospinal/upper motor neurons) and VI (corticothalamic). It is currently unknown whether a similar input/output structure is present in human M1, primarily due to a lack of human-relevant methodology operating at the scale of cortical layers.

Fig. 1

Study overview. A) Model of possible M1 microcircuitry, based on animal literature (references listed in main text). Most input to M1 arrives in superficial layers, and the majority of output (corticofugal) originates from deep layers. The strong intracolumnar projection from superficial to deep layers is consistent with the view that preprocessed information from superficial layers is integrated by deep layers before the resulting output is directed to, eg the spinal cord (Bastos et al. 2012). Superficial and deep layers can thus be roughly segregated into input and output layers, respectively. B) Illustration of the research questions of the present study. Previous laminar fMRI studies found robust M1 deep layer activation during active finger movements, interpreted to primarily reflect motor output (see main text). Here, we test if passive movement (proprioceptive stimulation, no motor output) also activates deep layers of M1. This would suggest that proprioception, alongside motor output, contributes to the deep layer signal observed during active movements. Please note that we do not make predictions about the relative laminar activation strength between active and passive conditions, and the response amplitude was thus deliberately made equal in this model. The depth-dependent signal was modeled as two peaks based on prior M1 laminar fMRI studies (see main text). C) Illustration of the methodological setup used to assess this question. Two different fMRI sequences with a complementary set of strengths and weaknesses were used in separate sessions, ie VASO and 2D-bSSFP. All sessions included active and passive finger movements. The finger was moved voluntarily by self-generated force during active finger movements (red device, top panel). For passive movements, subjects would relax while their finger was moved by an fMRI-compatible pneumatic muscle driven by air pressure (blue device, top panel, Lolli et al. 2019). These conditions alternated with rest in a block-designed manner and were performed in separate runs in counterbalanced order (middle panel). The bottom panel shows the perpendicular slice placement with respect to the cortical surface of M1 due to anisotropic voxels.

To this end, the advent of laminar fMRI (Lawrence et al. 2019; Norris and Polimeni 2019; Bandettini et al. 2021) has enabled assessment of the layer-dependent functional organization in humans, underpinning prominent theories of brain function, such as predictive coding (Rao and Ballard 1999; Stephan et al. 2019). Laminar fMRI was employed by Huber et al. (2017) to compare laminar activation of the M1 hand representation in response to tasks with varying degrees of motor and somatosensory involvement. Active finger movements elicited a double peak response pattern with signal increases in both superficial and deep layers, which has been replicated in several subsequent studies (Huber et al. 2018; Beckett et al. 2020; Chai et al. 2020; Guidi et al. 2020; Persichetti et al. 2020; Shao et al. 2021; Knudsen et al. 2023; Pais-Roldán et al. 2023). Moreover, tactile stimulation (Huber et al. 2017; Shao et al. 2021) and imagined finger movements (Turner 2016; Persichetti et al. 2020), both of which presumably lack corticospinal motor output, were found to predominantly activate superficial layers. Informed by the animal literature (see Fig. 1A), the general interpretation across these studies has therefore been that superficial peak activation mainly reflects somatosensory and premotor input, while the deep peak reflects motor output (see also Dumoulin 2017; Larkum et al. 2018; McColgan et al. 2020).

Due to a paucity of studies employing a direct proprioceptive control condition, the pattern of laminar activation associated with proprioceptive somatosensation is currently unknown. Although tactile somatosensory processing appears to predominantly engage the superficial layers of M1, proprioceptive somatosensory processing may additionally engage deep layers. For example, the lack of robust deep layer activation during tactile somatosensory stimulation might simply be due to the fact that signals were extracted from the phylogenetically older and upper part of M1 (BA4a) (Huber et al. 2017; Shao et al. 2021), which has few tactile responsive cells (Rathelot and Strick 2009). In contrast, proprioceptive responsive cells are particularly concentrated in that region (Tanji and Wise 1981; Strick and Preston 1982; Stepniewska et al. 1993). Also, electrophysiological measurements obtained during proprioceptive stimulation in nonhuman primates demonstrated activation in pyramidal tract cells, which are selectively located in deep layers (Wiesendanger 1973; Lemon and Porter 1976). However, it remains unknown whether such activation extends to humans. Although human positron emission tomography (PET) and fMRI studies have consistently reported robust M1 activation in response to proprioceptive stimulation induced by passive movements, these studies lacked the spatial resolution necessary to decipher the laminar origin of the measured signals (Weiller et al. 1996; Ciccarelli et al. 2005; Guzzetta et al. 2007; Blatow et al. 2011; Mehta et al. 2012; Michels et al. 2014).

In the present study, we aimed to determine the laminar response of human M1 to proprioceptive stimulation. Specifically, we used 7 T laminar fMRI with vascular space occupancy (VASO) (Huber et al. 2014) and 2D passband balanced steady-state free precession (2D-bSSFP) sequences (Miller 2012; Scheffler et al. 2018) to examine whether passive finger movements (proprioceptive stimulation) elicit deep layer responses in the hand knob of contralateral M1 (Fig. 1B). Addressing this basic neuroscience question may advance our understanding of laminar function in human M1 and contribute to the development and validation of theories concerning motor control and somatosensory perception.

Materials and methods

Subjects

A total of 12 healthy volunteers (23 to 30 years, 6 females) participated in the study. The 2D-bSSFP dataset of one subject, which was also scanned with VASO, was excluded due to excessive motion and nonperpendicular slice placement (Imaging Protocol). Analysis was thereby performed on 10 VASO subjects and 2 2D-bSSFP subjects (one of which was scanned twice across different days to display the reliability of 2D-bSSFP activation maps, Supplementary Fig. S1) (Fig. 1C). Before participation, all subjects provided written consent and were carefully informed about the procedures, approved by the institutional review board (2012-IRB-011) of the Institute of Biophysics, Chinese Academy of Sciences.

Active and passive movement devices

Passive finger tapping was performed using a fMRI-compatible device inspired by (Lolli et al. 2019) (Fig. 1C). It was driven by a fluidic muscle (DMSP type, diameter 10 mm, length of the contracting part 150 mm; Festo AG & Co, Esslingen, Germany), which shortens upon increased air pressure and extends toward its initial position upon pressure release. Air supply was controlled via a preprogrammed electrical valve to obtain the desired frequency and amplitude. An identical device was used for active tapping except it did not contain any muscle (Fig. 1C). The design goal was to make active and passive movements as similar as possible to standardize sensory input. A custom-made handprint was positioned around each device for each subject to standardize the amplitude of vertical movement across conditions while enabling subjects to switch devices within the scanner. A soft pad was inserted inside the active device to roughly match impact forces between the movable part (to which the finger is attached) and the outer tube across conditions.

Imaging protocol

Rationale behind choice of functional sequences

We intended to assess the laminar pattern of passive movements within the same functional region as in previous laminar M1 studies, ie the part of the hand knob in Brodmann area 4a (BA4a) expressing a double peak response pattern during active movements (Huber et al. 2017; Beckett et al. 2020; Chai et al. 2020; Guidi et al. 2020; Persichetti et al. 2020; Han et al. 2021; Shao et al. 2021; Knudsen et al. 2023; Pais-Roldán et al. 2023). Observing distinct peaks at certain depths not only facilitates comparison with existing literature but also enhances confidence in the activation of the underlying layer. That is, it helps rule out the possibility that the observed signal is merely an artifact of blurring or partial voluming from adjacent activated layers. Accordingly, 3D-echo planar imaging (EPI) SS-SI-VASO (Huber et al. 2014) was chosen as the main sequence due to its cerebral blood volume (CBV)-weighted contrast associated with high laminar specificity, which has proven successful for reliable identification of the double peak feature (Huber et al. 2017; Beckett et al. 2020; Guidi et al. 2020; Persichetti et al. 2020). The strong microvascular weighting of VASO, however, results in inherently reduced contrast-to-noise ratio (CNR), which was amplified by a relatively weak proprioceptive stimulus (passive movement of a single finger with a small range of motion), yielding noisy activation maps and laminar profiles. To further support that deep layer signal is a robust feature of proprioceptive stimulation, we included additional data from two subjects scanned with 2D-bSSFP. This sequence was also expected capable of resolving laminar activation peaks due to its ultrahigh in-plane resolution (0.3 mm), minimal T2*-based blurring, no spatial distortions, and spin echo-like contrast mechanism yielding blood-oxygen-level-dependent (BOLD) contrast with reduced sensitivity toward spatially unspecific large veins. Combined with relatively high sensitivity, it was expected to facilitate further support for our main conclusion at the level of individual subjects (Miller 2012; Báez-Yánez et al. 2017; Scheffler et al. 2018; Liang et al. 2022).

Sequence parameters

Imaging was performed using a MAGNETOM 7 T MRI scanner (Siemens Healthineers, Erlangen, Germany) equipped with a standard 32Ch-receive head coil (NOVA Medical, Wilmington, MA, USA) and a SC72 body gradient coil for spatial encoding. Anatomical T1-weighted images were acquired with an MP2RAGE sequence (Marques et al. 2010) and parameters: TR = 4,000 ms, TE = 3.05 ms, voxel size = 0.7 mm isotropic, field of view (FOV) = 224 |$\times$| 224 |$\times$| 179.2 mm³, GRAPPA acceleration factor = 3, TI1 = 750 ms with flip angle = 4 deg and TI2 = 2,500 ms with flip angle = 5 deg, phase and slice partial Fourier factor = 7/8, bandwidth = 240 Hz/px).

The parameters of the functional sequences were:

• VASO: voxel size = 0.82|$\times$|0.82|$\times$|1.5 |${mm}^3$|⁠, TI1/TI2 = 1,593/3,780 ms, volume acquisition time (nulled and not-nulled combined) = 4,872 ms, TE = 25 ms, GRAPPA acceleration factor = 3, phase partial Fourier = 6/8, bandwidth 1064 Hz/px, and matrix size = 162|$\times$|216|$\times$|26. A variable flip angle scheme was used to minimize T1-related blurring (see Huber et al. 2017).

• 2D-bSSFP: voxel size = 0.3|$\times$|0.3|$\times$|3 |${mm}^3$|⁠, slice acquisition time = 3 s, TE = 4.79 ms, TR = 9.58 ms, no parallel imaging, no partial Fourier, nominal flip angle = 35 deg, bandwidth 174 Hz/px, and matrix size 302|$\times$|320|$\times$|1. Prior to acquiring data for each slice, a series of radio frequency (RF) pulses were applied to ensure that the MRI signal reached a steady state. To avoid the banding artifact of bSSFP images, we carefully adjusted the RF frequency to keep the region of interest (ROI) in the middle of the passband off-resonance profile. A series of images were collected with an RF frequency shift (Δf) from −70 to 70 Hz with a step of 10 Hz. The optimal Δf was determined from the off-resonance profile from the location of interest in M1 gray matter.

Both functional sequences had partial brain coverage. Imaging slabs were placed in the left hemisphere to cover the hand knob (Yousry et al. 1997) in the upper part of M1, ie BA4a, in line with previous M1 laminar fMRI studies. Due to the anisotropic voxels in VASO and 2D-bSSFP, slabs were placed perpendicularly to the cortical surface of the hand knob (Fig. 1C).

Experimental paradigm and procedures

Prior to scanning, subjects practiced reproducing the passive movement pattern (~1.5–2 cm amplitude, 2 Hz movement frequency) during active movements of the right index finger and were instructed to completely relax during the passive condition. Movement frequency was guided visually by blinking text saying either “you tap” for the active condition, “machine tap” for the passive condition, or “relax,” presented with an MRI-compatible projector onto a semitransparent screen (1,024 × 768, 60 Hz). The visual stimuli were programmed in MATLAB (Mathworks Inc.) using the Psychophysics Toolbox (Brainard 1997). Subjects were asked to keep their fixations on the visual stimuli during the whole experiment, except in resting periods between runs. Each session consisted of an anatomical scan (T1-weighted MP2RAGE) followed by 2 functional runs per condition for VASO sessions and 3–4 runs per condition for 2D-bSSFP. Each run had a duration of 12 min for VASO and 6 min for 2D-bSSFP. In active runs, ~ 30 s of rest alternated with ~30 s of active movement with the right index finger. Passive runs alternated between ~30 s of rest and ~30 s of passive movement where the subject would relax while the finger was moved by the device (Fig. 1C). The order of conditions was counterbalanced across subjects, and, after finishing the runs of one condition, subjects were asked to slowly shift from the active to the passive device (or vice versa) while trying to minimize head motion. Head motion was restricted by custom-made bite bars.

A functional localizer run (12 min of active finger movements) using a 2D gradient echo EPI BOLD sequence (volume acquisition time: 2,000 ms, TE: 25.6 ms, flip angle: 80 deg, GRAPPA acceleration factor: 3, phase partial Fourier: 6/8, voxel size: 0.75|$\times$|0.75|$\times$|0.86 |${mm}^3$|⁠, and matrix size: 170|$\times$|170|$\times$|28) was included at the end of VASO sessions for the purpose of ROI definition (see ROI Definition).

Data analysis

Preprocessing

The VASO data underwent standard preprocessing steps in line with previous laminar VASO studies (Huber et al. 2017; Finn et al. 2019). SS-SI-VASO produces two time series in an alternating fashion, one where the blood-signal is nulled and another without blood-nulling (BOLD time series). Nulled and not-nulled images were motion-corrected separately with the application of a spatial weighting mask to optimize the realignment around M1 using SPM12 (Functional Imaging Laboratory, University College London, UK). To correct for the influence of positive BOLD effects, which counteract the negative CBV response, dynamic division of nulled and not-nulled images was implemented after trial averaging using the LN_BOCO function in LAYNII (version 2.4.0) (Huber et al. 2014; Huber et al. 2021). Realigned nulled and not-nulled images were further used to compute an image with T1-weighted anatomical contrast (Huber et al. 2017), which was used as a “high-quality reference image” for nonlinear registration of the MP2RAGE image to EPI space. This coregistration step was performed by combining rigid, affine and nonlinear (SyN-algorithm) transformations in ANTs (Avants et al. 2011).

The bSSFP data were preprocessed using Analysis of Functional NeuroImages (AFNI) (Cox 1996), including removal of outliers using 3dToutcount (the definition of outliers used here is described in the AFNI documentation, https://afni.nimh.nih.gov/pub/dist/doc/program_help/3dToutcount.html), motion correction (in-plane only, ie two translation and one rotation parameters using 3dAllineate), brain masking, and coregistration of the T1 weighted anatomical volume to the mean functional image with localized Pearson correlation as the cost function (align_epi_anat.py). Functional bSSFP images have minimal distortions, analogous to MP2RAGE, and high-quality coregistration was thus possible using linear transformations only (Liang et al. 2022).

For all sequences, activation maps were computed voxel-wise as the average of condition-block timepoints minus the average of resting-block timepoints. This difference was converted to percent signal change by dividing by the average of resting-block timepoints. The first two volumes of the nulled and not-nulled data, as well as the first four volumes of the 2D-bSSFP data, were discarded to limit the influence of transition periods. Also, activation-induced CBV increases result in negative VASO signal changes, and the corresponding activation maps were thus inverted. To reduce noise in the activation maps, within-layer smoothed versions were also generated. This was done using LAYNII’s LN2_LAYER_SMOOTH (full width at half maximum (FWHM) = 0.75 mm for VASO, FWHM = 0.5 mm for 2D-bSSFP), including the option to mask out activation outside the segmentation boundaries (Huber et al. 2021). These maps were used for visualization of the activation maps only.

ROI definition

ROIs were defined by first identifying the hand knob based on its omega-shaped anatomical landmark (Yousry et al. 1997). Then, to identify the double peak region, VASO activation maps of the active condition were used to localize the subregion of the hand knob in area BA4a that most clearly expressed a double peak response pattern, inspired by Persichetti et al. (2020). Since VASO activation maps had relatively low CNR, independent localizer maps with BOLD contrast were used to confirm robust responses to index finger movements in the selected area. Connected ROIs spanning 1–2 slices and without holes were manually drawn in this area, with columnar widths roughly corresponding to that expected from individual finger representations in M1 (Huber et al. 2020). The same procedure was followed for 2D-bSSFP, except it was done on the single available slice. For the subject that was scanned twice with 2D-bSSFP, we selected the session with strongest activation for the active condition to evaluate the response to passive movements. The other session was used to assess the robustness of activation maps across sessions (Supplementary Fig. S1). Note that since data from the active condition were used to extract ROIs, we only statistically analyzed data from the passive condition to avoid “double dipping” (Kriegeskorte et al. 2009). Laminar profiles corresponding to the active condition are still shown for reference but are likely biased and should be interpreted as such.

Laminar profiles and assignment of cortical layers to relative depths

For VASO, white matter (WM)/gray matter (GM) and cerebrospinal fluid (CSF)/GM boundaries were drawn manually around ROIs based on MP2RAGE images and used to compute equidistant depth maps with LAYNII’s LN2_LAYERS (Huber et al. 2021). This was done on an upsampled grid with 0.2 mm in-plane resolution. The same procedure was followed for 2D-bSSFP, except that upsampling was not required due to the inherently high in-plane resolution of this sequence. Laminar profiles with 18 bins were generated for each condition by computing the average percent signal change across all ROI voxels within each bin (no thresholding).

Layer Va’s approximate depth and width were inferred from laminar thickness estimates outlined in Zilles and Amunts (2012) and Palomero-Gallagher and Zilles (2019). This depth co-localized with the dip between deep and superficial peaks in the active condition’s group-level VASO profile and the T1-profile plateau, which have previously been used as M1 layer Va landmarks (Huber et al. 2017) (see Supplementary Fig. S2 for details).

Statistical analysis

Cluster-based permutation testing (Nichols and Holmes 2003) was employed to determine the significance of bins in passive laminar profiles. The permutation strategy used here accounts for multiple comparisons (ie controls Type I error, see Nichols and Holmes (2003) for details) by deriving statistical distributions and values based on the profiles as a whole rather than individual layer-bins. Testing was performed across subjects for VASO (group level) and across trials for 2D-bSSFP (subject level). For each permutation, passive laminar profiles were randomly inverted with a minus sign since, under the null hypothesis, all profiles could equally likely have come from the contrast “rest>passive” as from the contrast “passive>rest.” P-values were then calculated for each bin across resulting profiles using one-sample t-tests. The maximum t-value sum observed for a cluster (range of consecutive bins with uncorrected P-values < |$0.05$|⁠) was recorded for each permutation. This was repeated in 100,000 permutations, yielding a null distribution from which the P-value could be derived for the t-value sum of a given cluster in the measured data. Note that WM and CSF bins (bins 1:3, and 18, respectively) were excluded in these analyses as our goal was to test for significant activation in gray matter (bins 4:17). This method tests the significance of a cluster as a whole, but significance of individual bins within the cluster cannot be inferred (Nichols and Holmes 2003). Thus, to evaluate significance of deep layer bins in isolation (without contribution from bins in superficial layers), P-values were also derived for subclusters that only contained deep layer bins (between the GM/WM boundary and the layer Vb/Va boundary). We derived one-sided P-values to test our hypothesis that proprioception is a contributing factor to the positive laminar activation observed during active finger movements (see Fig. 1B); this hypothesis predicts a unidirectional effect where only an enhancement—not a decrease—in activation levels would be considered indicative of significant proprioceptive contribution.

The significance level for all statistical tests was set at |$\alpha$| = 0.05.

Data and code availability

The raw data will be available for download via figshare at the time of publication (https://doi.org/10.6084/m9.figshare.23834838). Analysis code will be made publicly available on GitHub (https://github.com/LasseKnudsen1/7TactiveVSpassiveStudy/tree/main).

Results

To test whether human M1 deep layer responses are evoked by proprioception, laminar profiles were extracted from ROIs containing the characteristic double peak response pattern for active movements (see ROI Definition). Figure 2A depicts the raw VASO activation maps of an example subject for each condition. The active map has robust activation in the expected part of the hand knob (BA4a), but it is relatively noisy, making the double peak feature less clearly visible (Fig. 2A, upper middle). This was alleviated in the within-layer smoothed activation maps, from which the double peak pattern more clearly emerged in the active tapping condition (Fig. 2A, upper right). The smoothed map for the passive condition also shows superficial and deep activations in similar locations (as indicated by arrows in Fig. 2A, lower right) but was weaker compared with the active tapping condition. Figure 2B shows the group-level VASO profiles associated with active and passive finger movements (for individual subject profiles, see Supplementary Fig. S3). Significant clusters of activation were found during passive movements in both the superficial (P = 0.039) and deep layers (P = 0.014) (shown as horizontal black lines in Fig. 2B; numbers denote a significant subcluster of deep layer bins, P = 0.022). Interestingly, the deep layer peak appears to be slightly displaced toward layer Va during passive movements compared to active movements (Fig. 2B).

Fig. 2

Evaluation of deep layer signal during passive movements using VASO A) The first column shows the depth map and ROI used to extract VASO and EPI-BOLD laminar profiles for this example subject. The second columns show raw percent signal change maps for each condition, and the third column shows corresponding maps after within-layer smoothing. Arrows point to indications of the double peak feature. Maps are thresholded at t = 1.5 (smoothed maps are further masked by GM segmentation). B) Laminar group-level profiles for each condition, both for VASO and BOLD. Asterisks (red) denote circular/biased profiles for the active condition. Solid horizontal lines illustrate significant clusters for the passive condition as evaluated by permutation tests. Dotted lines and corresponding numbers denote bins (counting from left to right) belonging to a significant cluster containing deep layer bins only (Statistical Analysis). Solid vertical lines illustrate WM/GM and GM/CSF boundaries. The blurry gray-shaded area denotes the approximate position of layer Va. Error bars refer to the standard error of the mean across subjects.

Activation maps (Supplementary Fig. S4A) and laminar profiles (Fig. 2B, Supplementary Fig. S3) were similarly generated from the not-nulled timeseries of the VASO sequence (referred to here as EPI-BOLD to avoid confusion with the BOLD-contrast of bSSFP). In line with VASO results, significant activation was found in both superficial and deep layers during passive movements (all GM bins were above threshold, resulting in a single large cluster shown with the horizontal black line in Fig. 2B, P = 0.002; numbers denote significant subcluster of deep-layer bins, P = 0.019). However, the maps and profiles were characterized by a bias toward superficial layers, and no clear double peak was observed at the group level in either condition. This is expected due to the sensitivity of gradient echo EPI-BOLD toward draining veins, which displace the signal from local neuronal activity and thus limit its laminar specificity (Turner 2002; Menon 2012). Consequently, the superficial layer signal becomes hard to interpret, but the deep layer signal, which is of main interest for our research question, is largely unaffected by large veins (Turner 2002; Menon 2012). Nevertheless, in an attempt to account for the signal dispersion originating from intracortical veins, spatial deconvolution (Markuerkiaga et al. 2016; Marquardt et al. 2018) was applied to the profiles in a supplementary analysis (Supplementary Fig. S4B). After deconvolution, the profiles began to resemble the corresponding VASO profiles (low superficial bias and indications of double peaks), known to be less influenced by macrovascular contamination. However, considering the limitations of the applied correction strategy (see Supplementary Fig. S4B), the corrected profiles should be interpreted with caution.

The activation maps for 2D-bSSFP are shown in Fig. 3A. Only one of the subjects had a clear double peak response within the same area in both conditions, but both subjects appeared to have robust deep layer activation in the passive tapping condition where activation of the active condition mostly resembled the double peak feature (denoted by arrows). This is supported by both subjects’ laminar profiles (Fig. 3B) showing significant subclusters of activation in the deep layers during passive movements (P = 0.004 and P = 0.001 for subject 1 and subject 2, respectively). Due to bSSFP being relatively insensitive to T2* contributions, it is expected to have less bias toward large veins than gradient echo EPI (Miller 2012; Báez-Yánez et al. 2017; Scheffler et al. 2018; Liang et al. 2022). Nevertheless, as visible in activation maps, we did observe strong signals toward CSF, which likely reflects high sensitivity of bSSFP-BOLD to intravascular signals and inflow effects (Liang et al. 2022). Inspired by Liu et al. (2020), voxels with percent signal change exceeding 4% in the across-condition averaged 2D-bSSFP map were excluded from the map of each condition before generating laminar profiles to minimize the effect of large vessels (uncorrected profiles are shown in Supplementary Fig. S5).

Fig. 3

Evaluation of deep layer signal during passive movements using 2D-bSSFP. A) The first column shows the depth map and ROI used to extract 2D-bSSFP laminar profiles for each subject. The subsequent columns show raw percent signal change maps for each condition, and the last columns show corresponding maps after within-layer smoothing. Arrows point to locations of robust deep layer activation. Raw maps are thresholded at t = 2 and smoothed maps at t = 3 (smoothed maps are further masked by the depth map). B) Subject-level laminar profiles for each condition. Asterisks (red) denote circular/biased profiles for the active condition. Solid horizontal lines illustrate significant clusters for the passive condition as evaluated by permutation tests. Dotted lines and corresponding numbers denote bins (counting from left to right) belonging to a significant cluster containing deep layer bins only (Statistical Analysis). Solid vertical lines illustrate WM/GM and GM/CSF boundaries. The blurry gray-shaded area denotes the approximate position of layer Va. Error bars refer to standard error of the mean across trials.

Discussion

In the present study, we used submillimeter resolution VASO and bSSFP at 7 T to demonstrate significant deep layer activation in human M1 in response to proprioceptive stimulation. Below, we discuss how this finding aligns with existing literature, interpret the M1 deep layer signal in the context of previous laminar fMRI studies and motor control theories, and address study limitations.

Proprioception explains signal in both superficial and deep layers of M1

Multiple studies have investigated the M1 laminar activation pattern associated with active finger movements. A consistent finding is the characteristic double peak response within the hand knob of BA4a (Huber et al. 2017; Beckett et al. 2020; Chai et al. 2020; Guidi et al. 2020; Persichetti et al. 2020; Shao et al. 2021; Knudsen et al. 2023; Pais-Roldán et al. 2023). This could similarly be observed in the present study, both in VASO (Fig. 2) and 2D-bSSFP (Fig. 3). Our novel finding is that deep layers are also activated in response to passive movement, ie a task in which voluntary motor aspects are presumably absent. Specifically, group-level VASO laminar profiles revealed significant activation in both superficial and deep layers during proprioceptive stimulation (Fig. 2). Significant deep layer activation was similarly observed at the group level for EPI-BOLD (Fig. 2B) and at the individual-subject level for the two subjects scanned with 2D-bSSFP (Fig. 3). These findings suggest a direct proprioceptive influence on M1 deep layer signals.

To our knowledge, M1 activation in response to proprioceptive stimulation has not been studied at the human laminar level previously, but, consistent with our findings, numerous studies have established strong M1 involvement in relation to proprioceptive perception. For instance, tendon vibration can induce proprioceptive illusions of limb movement that strongly activate M1 without actual or intended movement (Naito 2004). Similarly, proprioceptive stimulation significantly increases M1 activation (Weiller et al. 1996; Ciccarelli et al. 2005; Guzzetta et al. 2007; Blatow et al. 2011; Mehta et al. 2012; Michels et al. 2014) and modulates its excitability (Carel et al. 2000; Lewis and Byblow 2004; Onishi 2018). These findings suggest an important role of M1 in the proprioceptive hierarchy and that voluntary movement is not a prerequisite for its activation. Our study extends these findings by demonstrating involvement of both superficial and deep layers. This is consistent with nonhuman primate electrophysiological studies showing that a large proportion of M1 cells are responsive to proprioceptive stimulation (see Naito 2004, and references therein) and that such cells are found in both superficial and deep layers, including corticospinal pyramidal tract neurons (Wiesendanger 1973; Lemon and Porter 1976).

Interpretation of the deep layer signal in M1

Huber et al. (2017) suggested that the peak in superficial layers mainly reflects corticocortical input from somatosensory (tactile and proprioception) and premotor areas. This interpretation was informed by anatomical tracer studies in rodents (Mao et al. 2011; Weiler et al. 2008, see also Fig. 1A) and was supported by their resting state connectivity results in humans. Furthermore, tactile stimulation was shown to affect superficial layers only, as has been shown for imagined movements, believed to entail substantial input from premotor areas (Turner 2016; Persichetti et al. 2020). In other words, tasks devoid of corticospinal motor output appear to lack deep layer signal changes, whereas active movements, involving motor output, are linked to deep layer activation. When factoring in that corticospinal neurons are selectively located in layer Vb, it is reasonable that the double peak has commonly been interpreted as premotor/somatosensory input in superficial layers and voluntary-movement-related corticospinal output in deep layers (Dumoulin 2017; Huber et al. 2017; Larkum et al. 2018; Beckett et al. 2020; Chai et al. 2020; McColgan et al. 2020; Persichetti et al. 2020; Shao et al. 2021; Knudsen et al. 2023; Pais-Roldán et al. 2023). Our observation of significant deep layer activation during passive movements indicates a concomitant proprioceptive contribution to the deep layer signal.

Based on the model proposed in Fig. 1A, proprioceptive information is predominantly received by superficial layers for preprocessing before being relayed to deep neurons via strong superficial-to-deep intracolumnar connections (Weiler et al. 2008; Bastos et al. 2012). This circuitry largely aligns with a follow-up psychophysiological interaction (PPI) connectivity analysis (detailed in the “Supplementary PPI Analysis” section in the Supplementary Material), which showed a significant interaction between S1 and M1 superficial layers and between M1 superficial and deep layers during active movements. However, these interactions were not observed in the passive condition, possibly due to the inherent sensitivity limitations of PPI analysis. These results and the associated limitations are discussed in further detail in the Supplementary Material.

Considering that M1 relies on proprioceptive state information for proper generation of motor commands (Hatsopoulos and Suminski 2011; Proske and Gandevia 2012; Omrani et al. 2017), deep layer activation may partly reflect delivery of proprioceptive information to pyramidal tract cells in layer Vb to shape corticospinal output. Furthermore, M1 has been shown to contribute to the generation of proprioceptive perception (Nudo et al. 2000; Naito 2004; Gandevia et al. 2006), and deep output layers may be actively involved in shaping proprioceptive sensations, rather than solely using it to guide output.

In view of the increasing prominence of hierarchical Bayesian theories of brain function and their inherent link to laminar circuitry, we additionally propose an interpretation of the observed deep layer signal based on the active inference account of sensorimotor function (Friston 2010; Adams et al. 2013; Shipp et al. 2013). In this framework, proprioceptive stimulation should strongly activate the superficial layers of M1 since neuronal units at this depth are the targets of driving feedforward proprioceptive prediction errors, which continuously update expectations about the current proprioceptive state (Adams et al. 2013; Shipp et al. 2013). The potent superficial-to-deep intracolumnar pathway conveys these expectations onto deep layer prediction units (Weiler et al. 2008; Bastos et al. 2012), which is accompanied by feedback connections from the premotor cortex. Thus, deep layer units integrate extensive synaptic input, which in itself may give rise to deep layer fMRI signals (Logothetis 2008). Furthermore, based on this input, proprioceptive predictions are generated and transmitted to S1 and to thalamic somatosensory nuclei so to inform proprioceptive perceptions (“perceptual inference” component of active inference) (Adams et al. 2013; Shipp et al. 2013). Within this framework, corticospinal output constitutes these same proprioceptive predictions directed to lower motor neurons in the spinal cord (the distinction between motor output and proprioception, as in Fig. 1B, is thus rendered meaningless as motor output and proprioception become two sides of the same coin). Corticospinal predictions about the current proprioceptive state are compared with incoming data from the muscular Ia stretch receptors. Any disparity between prediction and data yields a proprioceptive prediction error that activates the motor neuron until the prediction is realized in the subsequent movement (“action” component of active inference) (Adams et al. 2013; Shipp et al. 2013). Consequently, another possible explanation for the observed deep layer signal is corticospinal drive being present even during passive movements; if the afferent data from the externally induced movement are not matched by a corticospinal proprioceptive prediction, a prediction error ensues leading to a countermovement (akin to the classical stretch reflex, Lidell and Sherrington 1924). Thus, for participants to be able to relax, M1 may continuously transmit corticospinal drive in the form of predictions about proprioceptive consequences of the passive movement. This interpretation is consistent with the observation that pyramidal tract neurons in layer Vb fire in response to passive movements (Wiesendanger 1973; Lemon and Porter 1976) and with clinical indications that upper motor neuron damage includes a failure to relax during passive movement (Menon and Vucic 2021). Nevertheless, we cannot exclude that the observed deep layer signal originated from nonpyramidal tract cells and more evidence for corticospinal drive during passive movements is necessary to support this explanation.

The deep peak for the passive condition appears slightly shifted toward layer Va compared to the active condition (Fig. 2B). Layer Va links M1 and S1 (Mao et al. 2011; Hooks 2017), whereas layer Vb contains all corticospinal cells. In line with the active inference framework, one possible explanation is that the proprioceptive predictions transmitted to S1—proposed to mediate “perceptual inference”—dominate during passive movements, whereas corticospinal drive may exert more influence during “action” (Friston 2010; Adams et al. 2013; Shipp et al. 2013). However, this remains speculative. Given the small sample size and intersubject variability observed in laminar profiles (Supplementary Fig. S3), further studies are needed to establish whether the shift is a robust feature distinguishing laminar profiles of active and passive finger movements using rigorous statistical testing.

Limitations

Our conclusion that proprioceptive stimulation activates deep layers of M1 is partly based on the significant deep activation cluster in the group-level laminar VASO profile (Fig. 2B). However, large intersubject variability was observed in the profiles (Supplementary Fig. S3). This limitation is alleviated to some extent by corroborating BOLD-EPI and bSSFP data. Nevertheless, our findings would provide stronger evidence if robust laminar patterns of VASO activation could be demonstrated across subjects for passive movements, as Huber et al. (2017) did for active movements. Part of the variability is likely explained by relatively weak stimuli in our movement tasks where only one finger was moved, and the range of motion was small due to the passive movement device’s constraints. To increase response magnitudes, it would be interesting to replicate the experiment of the present study using a passive movement device that tolerates multifinger movements and generates a wider range of motion.

Additionally, passive movement may involve attentional effects, tactile input, or premotor contributions, which were not controlled for in this study. That is, other neural sources than proprioceptive processing could potentially explain the observed deep layer signal. However, the effect of attention on the fMRI signal has been reported to primarily modulate BA4p (Binkofski et al. 2002) but not the ROI in the present study, ie BA4a. Furthermore, the lack of robust M1 deep layer activation during tactile and imagined movement conditions (Turner 2016; Huber et al. 2017; Persichetti et al. 2020; Shao et al. 2021), makes tactile and premotor input unlikely to confound our interpretation of proprioceptive processing being the main source of deep layer activation during passive movements. That being said, future studies controlling directly for such nonproprioceptive sources would be needed to definitively settle this.

Finally, our findings are only valid if subjects relaxed their fingers during passive movements, which could have been confirmed through electromyography (EMG) recordings. Although we did not obtain such recordings, previous research has demonstrated that subjects generally have no difficulty relaxing their muscles during passive movements (Mima et al. 1999; Thickbroom et al. 2003; Ciccarelli et al. 2005).

Conclusion

We used 7 T laminar fMRI to investigate the laminar pattern of activation in M1 in response to passive finger movements. Significant activation was observed in both superficial and deep layers. Our results align with the M1 laminar input/output structure proposed by animal studies and human laminar fMRI research. They further detail this model by suggesting that M1 deep layers constitute an important component in proprioceptive circuits and that a non-negligible portion of the M1 deep layer activation likely reflects proprioceptive influence. We believe this study enhances our understanding of the laminar circuitry in human M1, thereby facilitating interpretations of future laminar research and contributing to the development of theoretical frameworks of motor control and somatosensory perception.

Acknowledgments

The authors would like to thank all subjects for their participation. We would further like to thank Wu Chen and Yanhui Fu (Institute of Biophysics, Beijing, China) for technical assistance, as well as Chengwen Liu and Chencan Qian (Institute of Biophysics, Beijing, China) for guidance on analysis. We thank Christopher J. Bailey (Aarhus University Hospital) for guidance throughout the project and development of the passive movement device. We thank Alessandra Pizzuti, Ana Arsenovic, David Norris, Maria Guidi, Omer Faruk Gulban, Renzo Huber, and Sebastian Dresbach for their valuable feedback on the manuscript.

Author contributions

Lasse Knudsen (Conceptualization, Methodology, Formal analysis, Investigation, Writing—Original Draft, Writing—Review & Editing, Visualization, Funding acquisition), Fanhua Guo (Methodology, Formal analysis, Investigation, Writing—Review & Editing, Visualization), Daniel Sharoh (Conceptualization, Formal analysis, Writing—Original Draft, Writing—Review & Editing), Jiepin Huang (Methodology, Investigation, Writing—Review & Editing), Jakob U. Blicher (Conceptualization, Methodology, Writing—Review & Editing, Supervision, Project administration, Funding acquisition), Torben E. Lund (Conceptualization, Methodology, Writing—Review & Editing), Yan Zhou (Writing—Review & Editing, Project administration, Funding acquisition), Peng Zhang (Conceptualization, Methodology, Writing—Review & Editing, Supervision, Project administration, Funding acquisition), and Yan Yang (Conceptualization, Methodology, Writing—Review & Editing, Supervision, Project administration, Funding acquisition).

Funding

This work was supported by Beijing Natural Science Foundation (Z210009), STI2030-Major Projects (2022ZD0204800, 2022ZD0211900, and 2021ZD0204200), National Key R&D Program of China (2022YFB4700101), the National Natural Science Foundation of China (32070987, 31722025, and 31930053), the Strategic Priority Research Program of Chinese Academy of Sciences (XDB37030303), and a grant from the Sino-Danish Center (SDC).

Conflict of interest statement: None declared.

Author notes

Lasse Knudsen and Fanhua Guo contributed equally to this research.