Action kinematics as an organising principle in the cortical control of human hand movement

Abstract

Hand movements are controlled by neuronal networks in primary motor cortex (M1). The organising principle encoding hand movements in M1 does not follow an anatomical body map, but rather a distributed representational structure in which motor primitives are combined to produce motor outputs. Electrophysiological recordings in primates suggest that M1 neurons encode kinematic features of movements, such as joint position and velocity. Human imaging data concur: relative differences in movement kinematics are mirrored by differences in the associated patterns of M1 activity. However, M1 exhibits well-documented sensory responses to cutaneous and proprioceptive stimuli, raising questions regarding the origins of kinematic motor representations: are they relevant in top-down motor control, or are they an epiphenomenon of bottom-up sensory feedback during movement? Here we show that the kinematic signature of a wide variety of naturalistic hand movements is encoded in human M1 prior to the point of movement initiation. Using a powerful combination of high-field fMRI and MEG, a spatial and temporal multivariate representational similarity analysis revealed that patterns of M1 activity mirrored kinematic, but not muscle-based features of naturalistic hand movements prior to movement onset. Comparable M1 activity was not observed for an ethological action model based functional mappings proposed in M1. Our spatial and temporal analyses provide firm evidence that the top-down control of dexterous movements activates cortical networks in M1 encoding hand kinematics.

Full text

Action kinematics as an organising principle in the
cortical control of human hand movement
James Kolasinski∗
, Diana C. Dima, David M. A. Mehler, Alice Stevenson,
Sara Valadan, Slawomir Kusmia, Holly E. Rossiter
Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff
University, Maindy Road, CF24 4HQ
Abstract
Hand movements are controlled by neuronal networks in primary motor cortex (M1)1,2,3.
The organising principle encoding hand movements in M1 does not follow an anatomical
body map, but rather a distributed representational structure in which motor primitives
are combined to produce motor outputs4,5. Electrophysiological recordings in primates
suggest that M1 neurons encode kinematic features of movements, such as joint position
and velocity6,7,8,9,10 . Human imaging data concur: relative differences in movement kine-
matics are mirrored by differences in the associated patterns of M1 activity3,11 . However,
M1 exhibits well-documented sensory responses to cutaneous and proprioceptive stim-
uli12, raising questions regarding the origins of kinematic motor representations: are they
relevant in top-down motor control, or are they an epiphenomenon of bottom-up sensory
feedback during movement? Here we show that the kinematic signature of a wide variety
of naturalistic hand movements is encoded in human M1 prior to the point of movement
initiation. Using a powerful combination of high-field fMRI and MEG, a spatial and tem-
poral multivariate representational similarity analysis revealed that patterns of M1 activity
mirrored kinematic, but not muscle-based features of naturalistic hand movements prior
to movement onset. Comparable M1 activity was not observed for an ethological action
model based functional mappings proposed in M1 13 . Our spatial and temporal analyses
provide firm evidence that the top-down control of dexterous movements activates cortical
networks in M1 encoding hand kinematics.
1. Main text
Mounting evidence supports the encoding of movements in M1 based on
kinematics and synergistic muscle activation, rather than the anatomy of
∗Corresponding author
Email address: kolasinskij@cardiff.ac.uk (James Kolasinski)
Preprint submitted to eLife April 9, 2019
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

the peripheral musculature4,5. Measurements from individual M1 neurons in
non-human primates reveal the encoding of multiple kinematic features, such
as speed, direction, and position in the same cells in a time-varying manner6.
The same neuronal populations have been shown to encode instantaneous
features during motor execution, as well as the target kinematic end point
and upcoming movement trajectory7,8,9,10.
In the human brain, evidence of neuronal tuning to multiple kinematic fea-
tures has been reported during the production of intended movements from
M1 microelectrode recordings made in tetraplegic patients14. The encoding
of kinematic features of hand movements in M1 has also been supported by
human imaging studies. Patterns of fMRI activity in sensorimotor cortex
have been shown to mirror the relative differences in the final joint config-
uration across a range of prehensile movements11. Similarly, the represen-
tational structure of fMRI activity in M1 during finger flexion is consistent
with patterns of finger co-use during naturalistic hand movements3.
However, the functional relevance of kinematic encoding in M1 to human
motor control remains a fundamental unknown. As well as their role in
motor output, M1 neurons exhibit rapid and integrative responses to so-
matosensory signals12,15. Kinematic information is inextricably linked to
proprioceptive and tactile signals: specific patterns of movement are as-
sociated with specific patterns of sensory feedback. Are kinematic motor
representations reported in human M1 functionally relevant in the process
of top-down motor control, or an epiphenomenon generated by bottom-up
sensory feedback during human movement production?
We addressed this question using a spatiotemporal multivariate representa-
tional similarity analysis to ask where in the human brain and when during
movement production are the kinematics of human hand movements en-
coded? This approach combined high-field fMRI and MEG data with kine-
2
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Hand movements
Abduct fingers Pinch: thumb and little finger
Cylinder Grip Pinch: thumb and index finger
Hook Grip Pinch: thumb and middle finger
Spherical Grip Pinch: thumb and ring finger
Index finger flexion (45°) Ring finger flexion (45°)
Index finger flexion (90°) Ring finger flexion (90°)
Index and middle finger flexion (90°) Ring and little finger flexion (90°)
Index finger and thumb roll Rock fingers
Little finger flexion (45°) Squeeze: thumb and fingers
Little finger flexion (90°) Abduct thumb
Middle finger flexion (45°) Extend thumb
Middle finger flexion (90°) Flex thumb
Middle and ring finger flexion (90°) Twiddle: thumb and index finger
Table 1: Outline of the 26 hand movements used in the motor task. Instructional videos presented
in Video S1.
matic data glove recordings made during a broad repertoire of prehensile
and non-prehensile hand movements. Probing recordings of human brain
activity with high spatial resolution from fMRI and high temporal resolu-
tion from MEG offers a powerful means to identify the location and tim-
ing of kinematic information encoding. Together this information was used
to dissociate the relevance of kinematic information in M1 to top-down or
bottom-up processes in motor control, as well as a relevance of alternative
muscle-based or ethological action based models.
Ten right-handed participants performed a range of 26 prehensile and non-
prehensile hand movements16,17 (Table 1, Video S1) in two fMRI sessions
(1.5 hours total fMRI data per participant), two MEG sessions (1.5 hours to-
tal MEG data per participant), and a behavioural testing session (35 minutes
kinematic data recording). In each session participants wore a right-handed
14-channel fibre optic data glove; kinematic data were recorded through-
out all sessions. Electromyography (EMG) data were acquired during MEG
sessions to validate the movement onset measures calculated from the data
3
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure 1: Spatial and temporal evidence for the encoding of hand kinematics in pri-
mary motor cortex prior to movement production. Kinematic, muscle, and ethological
action models of hand movement were used in a spatiotemporal representational similarity anal-
ysis. Top row: fMRI data show that kinematic information was encoded consistently in primary
motor cortex across all ten participants; complementary MEG data revealed temporal encoding of
kinematic information (blue box) prior to movement onset (green line). The muscle model (mid-
dle row) and ethological action model (bottom row) showed very limited evidence of encoding,
outside of M1 in the post-central gyrus and offered no evidence of significant temporal encoding
during movement production. MEG temporal searchlight plots: data presented are from beta
band analysis; full analysis presented in Figure 4, green line - movement onset defined by the
data glove; blue regions - significant peaks in representational similarity between MEG data and
the motor model; dashed line - correlation noise ceiling. EMG onset violin plots based on data
presented in Figure S10. Model matrices reproduced in a larger format in Figure S2.
4
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure 2: Single participant fMRI representational similarity analysis cortical search-
lights using individual kinematic models of hand movement. Cortical heatmaps of the left
(A) and right (B) hemisphere, show consistent encoding of kinematic information in the left mo-
tor cortex, contralateral to movement. Heatmaps were constructed from individually thresholded
cortical searchlights for each participant, derived using their own kinematic model (C) (Omnibus
threshold, α= 0.01, maximum accuracy distribution calculated from peak correlation value across
10,000 searchlight permutations with label-switching). Supra-threshold range of Spearman’s ρfor
each participant presented in Table S2.
5
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure 3: Comparison of movement models in spatial searchlight analysis reveals
a significantly greater model fit for the kinematic model over the muscle model.
Wilcoxon signed-rank test (one-sided) test applied to the difference between the kinematic and
muscle model rho-value maps. Significant region of kinematic model fit aligned with Brodmann
Area 4. Statistical maps sub ject to FDR correction (α= 0.05).
glove.
To probe the spatial and temporal correspondence between patterns of brain
activity and hand kinematics, data glove recordings were used to construct a
kinematic model quantifying the similarity of the kinematic signals measured
during each of the 26 movements (Figure 1: Top row, Figures S2 and S3).
The kinematic model quantified the distance between the displacement mea-
sures for each movement pair across the 14 channels (Pearson correlation),
subject to a Fisher Z-transformation and averaged across the 14 recording
channels. The resulting kinematic model exhibits strong split-half and inter-
session consistency within participant (Figure S1). A grand average of the
kinematic model across sessions and participants was subject to non-classical
multidimensional scaling for visualisation of the relative dissimilarity of each
movement across two dimensions (Video S2). In both the spatial and tempo-
ral representational similarity analysis, the kinematic model was investigated
6
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure 4: MEG temporal representational similarity analysis searchlight in motor
cortex reveals encoding of kinematic information prior to movement onset. Temporal
MEG searchlight analysis reveals evidence of the kinematic encoding of hand movements prior
to movement onset (green bar). Distinct significant peaks in the correspondence between the
kinematic model and MEG data (blue) were observed in the beta band (-210 ms to -85 ms) and
the alpha band (-175 ms to -115 ms). An additional significant peak in the beta band analysis
was observed after movement onset (1260 ms - 1350 ms). No such significant peaks were observed
for the muscle model or ethological action model. Green line - movement onset defined by the
data glove; blue regions - significant peaks in representational similarity between MEG data and
the model; dashed line - correlation noise ceiling.
7
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

alongside two other models. A muscle based model was constructed from
high-density EMG recordings (15 channels) made in an independent cohort
of 10 participants performing the same range of hand movements (Figure 1:
Middle row). An additional ethological action model classified movements
into precision prehensile, power prehensile, and non-prehensile, based on the
notion of ethological maps in primate M113 (Figure 1: Bottom row).
We first used high-resolution fMRI data to perform a cross-validated corti-
cal surface-based searchlight representational similarity analysis to find ev-
idence for the spatial encoding of kinematic information during movement.
In each participant and each cortical searchlight, the unsmoothed pattern
of fMRI activity during movement was used to construct a representational
dissimilarity matrix (RDM)18 . The RDM was compared to the participant’s
individual kinematic model, resulting in representational similarity surface
maps of Spearman’s ρvalues for each participant, which were subject to
an omnibus threshold (α= 0.01; suprathreshold range for each participant
outlined in table S2) and used to construct a cross-participant heatmap.
This analysis assessed where the relative dissimilarities in the kinematic
recordings across the different hand movements were mirrored by the rela-
tive differences in the pattern of fMRI activity elicited by performing the
same movements. The searchlight revealed a strong and consistent represen-
tational similarity in the contralateral pre-central region of the anatomical
hand-knob19 across participants (Figure 1). Specifically, the fMRI search-
light results revealed the consistent encoding of the kinematic information
in Brodmann Area 4 during the production of hand movements across par-
ticipants (Table 2)20 .
Inspection of the single-subject cortical searchlight results for the kinematic
model highlights the consistent and spatially limited correspondence of the
kinematic model and fMRI data at the level of individual participants and
8
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

models in contralateral M1 (Figure 2A). A highly comparable result was
also observed using the kinematic model constructed from the data glove
recordings made in the behavioural testing session (Figure S4), highlighting
the applicability of this result to real-world hand use in an upright sitting
position. No such consistent representational similarity was observed in
the corresponding searchlight of movement-related activity in the ipsilateral
hemisphere at the group level (Figure 2B and Figure S4B).
Equivalent spatial searchlight analyses for the muscle model and the etho-
logical action models revealed more limited evidence of consistent cortical
encoding across participants, centred on somatosensory cortex in the post-
central gyrus; specifically Brodmann Area 3b (Figure 1). Inspection of the
single-subject cortical searchlights for both of the muscle and ethological
action models again revealed more limited evidence of representational sim-
ilarity in pre-central and parietal regions (Figures S6 and S7). In light of
the interest in contrasting the kinematic and muscle models11, a Wilcoxon
signed-rank test (one-sided) was used to demonstrate the superior fit of the
kinematic model in comparison to the muscle model in a localised region
principally corresponding to Brodmann Area 4 18 (Figure 3).
Ultra high field fMRI data analysed at the level of individual subjects offered
detailed spatial resolution, revealing the encoding of kinematic information
in the hand knob region of M1. However, fMRI offers relatively poor tempo-
ral resolution to understand the point in time at which the kinematic model
matches the pattern of brain activity. The boundary between motor and
somatosensory cortex is increasingly blurred by evidence of sensory process-
ing in M112 and motor modulation of sensory afferents21. The observed
representational similarity of fMRI activity and hand kinematics may result
from top-down control of motor function or bottom-up proprioceptive infor-
mation passed back to M1 and S1. In order to dissociate the driving force
9
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

behind the kinematic model fit observed in the fMRI data, a temporal rep-
resentational similarity analysis of MEG data was used to identify the point
during movement preparation or execution at which kinematic information
is encoded in the M1.
A cross-validated fixed-effects representational similarity analysis was ap-
plied, comparing a group average kinematic model derived from data glove
recordings made in the MEG scanner to the pattern of alpha (7-14 Hz),
beta (15-30 Hz), and gamma (30-100 Hz) band MEG brain activity in M1
(Figure S9) in 20 ms sliding windows during movement preparation and ex-
ecution. The muscle model and ethological action model were assessed in
equivalent analyses. In light of the interest in contrasting the kinematic and
muscle models, the kinematic and muscle models were each assessed in a
partial correlation to discount the contribution of the other.
In both the alpha and beta band analysis, there was significant correspon-
dence between the MEG data and the kinematic model both preceding
movement onset, and in the case of the beta band, after movement on-
set (Figure 4). In the beta band, the kinematic model mirrored the pattern
of brain activity in a significant peak from -210 ms to -85 ms relative to
movement onset (peak Spearman’s ρ: 0.32). There was also a significant
peak in the correspondence between the kinematic model and MEG data in
the beta band after movement onset in the beta band (1260 - 1350 ms; peak
Spearman’s ρ: 0.35).
In the alpha band a correspondence between the kinematic model and MEG
data was observed prior to movement onset from -175 ms to -115 ms (peak
Spearman’s ρ: 0.37). No significant peaks were observed in the correspon-
dence between the M1 MEG signal in the alpha, beta, or gamma band and
either the muscle model or the ethological action model (Figure 4).
10
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Model
Peak heatmap
overlap
(Participants)
Peak Vertex Anatomical
location
Kinematic 10 8052 Area 4
Muscle 4
1904
4781
7266
11766
Area 2
VIP
SCEF
Area 23c
Ethological 8 8070 Area 3b
Table 2: Outline of peak anatomical correspondence between movement models and fMRI cal-
culated using across participant cortical heatmaps. Peak regions calculated as centre of gravity of
areas of peak overlap; peaks separated by a minimum of 20mm. Vertex positions and anatomical
definitions are based on HCP S1200 32k release 20.
An analogous MEG temporal searchlight analysis during action observa-
tion revealed limited evidence of a correspondence between the kinematic
model and brain activity in the alpha band in action observation during the
movement videos preceding each movement block (Figure S5). During ac-
tion observation a correspondence between the MEG signal and kinematic
model was observed from 315 ms - 380 ms in the beta band, relative to
stimulus onset (peak Spearman’s ρ: 0.32). No peaks in any frequency band
were observed for the muscle model or the ethological action model during
the period of video observation.
Taken together, the MEG and fMRI results presented here strongly implicate
the encoding of kinematic information in M1 as an organising feature in the
top-down control of movement, rather than as a result of bottom-up sensory
signals elicited by motor activity.
Using 7T fMRI we pinpointed a consistent encoding of kinematic informa-
tion firmly in a localised region of Brodmann area 4 in M1. A temporal
multivariate analysis of MEG data allowed us to further unpack this result,
delving into the encoding of hand kinematics during the production of an
individual movement. Using MEG, we observed that the encoding of this
11
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

kinematic information occurs prior to the onset of a movement. In other
words, the relative differences in the kinematic structure of a range of dif-
ferent hand movements is encoded in M1 up to 210 ms before the onset of
movement can be detected in the hand.
Information contained in the kinematic model showed temporally distinct
correspondence to information contained in the alpha and beta bands of
the MEG data. From 210 ms to 90 ms before movement is detected, the
representational structure in the M1 beta band corresponds significantly to
the representational similarity of the kinematics of the upcoming movement.
The correspondence between the kinematic model and the information con-
tained in the beta frequency band is consistent with the broad literature
concerning the role of this oscillatory frequency in motor control. Beta os-
cillations are observed at rest; it is well established that beta activity is
suppressed immediately prior to and during movement: movement-related
beta desynchronisation (MRBD), and then rebounds following movement
cessation: post-movement beta rebound (PMBR)22. The magnitude of the
reduction in beta-band power observed prior to movement onset in motor
cortex has been shown previously to relate to the degree of uncertainty in
the upcoming movement23 or action anticipation24. Previous comparisons of
beta desychronisation made across kinematic and kinetic tasks concur: the
strength of MRBD is correlated with the physical kinematic displacement of
a given hand movement rather than the magnitude of muscle contraction25.
Similar patterns of desynchronisation are observed in alpha band activity,
where ERD in M1 corresponds to increased activation in the region22 , with
post-motion event related synchronisation in M126. Here we demonstrate
that there is a link between information contained in the alpha and beta
frequencies in M1 before movement onset and the subsequent kinematics of
hand movements (Figures 1 and 4), suggesting that important information
12
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

about the upcoming motor command may be encoded within these oscilla-
tions25,27. A significant peak was also observed post-movement beta-band
analysis; it be could be speculated that this peak represents information en-
coding of kinematic information around the time PMBR is known to occur
after movement, or could reflect afferent inputs to M128 ; these two possibil-
ities cannnot be dissociated by the current experiment.
In contrast to alpha and beta frequencies, we observed no concurrence be-
tween the information contained in the gamma-frequency and the kinematic
model. An increase in the amplitude of gamma oscillations has previously
been reported during motor execution: movement-related gamma synchro-
nisation (MRGS)29,30 . In contrast to alpha and beta frequencies, evidence
from studies of gamma oscillations report changes only after movement on-
set, and therefore would not be implicated in the encoding of information
in M1 prior to movement, consistent with the data herein31,32,33 .
Hand kinematics have previously been investigated in the context of hu-
man fMRI. Relative differences in target joint position at the end of a hand
movement have been shown previously to mirror the relative differences in
the fMRI signal in a broad region sensorimotor cortex11. Additional work
considering unidigit and multidigit flexion has demonstrated that patterns
of M1 fMRI activity associated with such movements are better explained
by kinematic models of digit co-use than by competing muscle-based mod-
els3. In the present study we have used MEG to fundamentally extend on
these findings, demonstrating a top-down role for kinematic encoding prior
to movement onset. Furthermore by using a kinematic model that compared
the displacement trajectory of each movement rather than a single joint po-
sition, it was possible to contrast the kinematic features of both prehensile
and non-prehensile movements to explore cortical encoding relevant to a full
range of naturalistic hand use. Moreover, by capitalising on the gains in spa-
13
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

tial and temporal resolution afforded by using 7T fMRI and a limited field
of view, we have been able to firmly pinpoint the spatial location at which
kinematic information is encoded to the motor region of the anatomical hand
knob, corresponding principally to Brodmann Area 4 19. The fMRI spatial
searchlight analysis did not reveal evidence of consistent encoding of kine-
matic information in ipsilateral M1 across participants (Figure 2). Previous
fMRI studies provide evidence for the activation of ipsilateral M1 during
the production of individual uni-digit movements34,35 but not multi-digit
sequences of uni-digit movements36. This study considered a broad array of
naturalistic hand movements, engaging a wide variety of hand kinematics,
involving simultaneous and/or sequential movement of different digits. It is
possible that unlike sequences of uni-digit movement, these more complex
movements do not drive the circuits of ipsilateral M1 as uni-digit movements
do34,35.
Previous studies have made direct comparisons between muscle-based mod-
els and kinematic models, arguing for the latter as an organising principle
in the encoding of hand movements3,11. Here, a model constructed from an
independently acquired set of high-density EMG recordings did not reveal
any evidence for the spatial or temporal encoding of information on the ba-
sis of differences in muscle activity across the range of 26 hand movements
under study. In addition, the kinematic model showed a superior represen-
tational similarity to the fMRI dataset than the muscle model in primary
motor cortex (Figure 3). As with previous studies, these findings do not
rule out the existence of muscle representations in M1, but rather support
the existence of highly organised muscle representations structured around
movement kinematics rather than anatomy. The assertion perhaps explains
the fractures and repetitions observed in muscle representations during the
search for an M1 body map2.
14
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

The ethological action model also reported less consistent patterns of fMRI
encoding, centred on the postcentral gyrus, consistent with activation in S1
(Figure 1). The ethological action model also did not reveal any significant
peak in the temporal representational analysis. It is possible that while
at a coarse level, ethological maps exist in the primate cortex, the concept
of ethological organisation does not extend down to the fine-grain level of
individual encoding of human hand movements; in other words, the broad
motor reportoire of the human hand may not be encoded on the basis of
the functional role of each movement. However, in the case of the primate,
the coarser division of movements based on the functional role of the entire
upper limb, including the hand (e.g. feeding, reaching), may play a role in
the way the cortex is organised37. The observed patterns of post-central
activity may alternatively result from selective disinhibition of S1 by M1
during motor activity, though such direct cortico-cortical signalling remains
speculative in the human brain21,38,39 .
Analysis of the action observation period of the MEG data preceding each
movement block also provided some support for the kinematic encoding of
information in M1 (Figure S5). Previous MEG data acquired during action
observation have demonstrated characteristic changes in M1 activity com-
parable to action execution40. Analyses of event related desynchronisation
(ERD) in M1 during action observation suggest a peak change in the mu
frequency as the observed movement evolves41 . These observations are po-
tentially consistent with the pattern of kinematic model fit observed in the
beta band MEG data early during action observation (315 - 380 ms after
stimulus onset), when the trajectory of movement has become clear (Fig-
ure S5). Additional work considering the encoding of kinematic information
in oscillatory alpha band activity in M1 suggests that the observation of
stimuli consistent with biological motion is sufficient to induce ERD in this
15
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

frequency band42, potentially consistent with the notion that during obser-
vation of biological motion, M1 may encode kinematic information.
The data presented in this study rely on complementary information ac-
quired from BOLD fMRI and MEG, though the remit of this work does
not extend to fusion of the two modalities. BOLD fMRI provides only an
indirect measure of neuronal activity based on haemodynamic changes as-
sociated with the execution of a given task43, which can be resolved with
a relatively high degree of spatial specificity with 7T imaging. In contrast,
MEG reflects a more direct, temporally-rich, measure of neuronal activity.
While the origins of the measured signals differ, compelling recent evidence
provides non-coincidental data to support the notion of shared information
across MEG and fMRI measures of brain activity across a wide range of
frequency bands44; similar correspondences have been reported from inva-
sive electrocorticography data45 . However, the spatial component of MEG
data must be inferred from mathematical modelling. Despite advances in
the context of MEG source localisation, this feature of MEG analysis limits
the spatial specificity of the measured signals, which integrate information
across relatively large tissue volumes in comparison with fMRI46. We have
harnessed the spatial and temporal strengths of fMRI and MEG, which in
combination provide greater insight regarding the encoding of movements
in M1 than the sum of their individual parts.
Here we apply a rich multi-modal design with multivariate analysis to demon-
strate that the encoding of kinematic information in human M1 occurs prior
to the onset of a wide range of naturalistic hand movements, in contrast to
competing muscle and ethological action models. Mounting evidence for
the encoding of complex kinematic information in M1 from this and other
work continues to blur the boundary between primary somatosensory and
primary motor cortex: even M1 neurons have been shown to rapidly con-
16
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

solidate sensory torque information across multiple joints15. The notion of
kinematic representation in M1 is compatible with recent evidence of the
tight integration of information across the central sulcus47, whereby S1 en-
codes the current body state, while M1 encodes the kinematics necessary
to achieve the intended body state. Such a system of motor control would
see kinematic information encoded prior to movement onset as a prediction
for the future sensory inputs expected by S1 when a movement has been
achieved48.
Acknowledgments
J.K. holds a Wellcome Trust Sir Henry Wellcome Postdoctoral Fellowship
(204696/Z/16/Z). CUBRIC is supported by a Strategic Award from the
Wellcome Trust (104943/Z/14/Z). This study was supported by the UK
MEG Partnership Grant (MRC/EPSRC, MR/K005464/1). The authors
are grateful to Krish Singh for his advice regarding the MEG analysis and
for his comments on the manuscript, and to Yi-Jhong Han for his technical
assistance with EMG data acquisition.
References
[1] W. Penfield, E. Boldrey, Somatic motor and sensory representation in
the cerebral cortex of man as studied by electrical stimulation, Brain
60 (1937) 389–443.
[2] R. Lemon, The output map of the primate motor cortex, Trends in
Neurosciences 11 (1988) 501–506.
[3] N. Ejaz, M. Hamada, J. Diedrichsen, Hand use predicts the structure
of representations in sensorimotor cortex., Nat. Neurosci. 18 (2015)
1034–1040.
17
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[4] S. A. Overduin, A. dAvella, J. M. Carmena, E. Bizzi, Microstimulation
Activates a Handful of Muscle Synergies, Neuron 76 (2012) 1071–1077.
[5] S. A. Overduin, A. d’Avella, J. Roh, J. M. Carmena, E. Bizzi, Repre-
sentation of Muscle Synergies in the Primate Brain, Journal of Neuro-
science 35 (2015) 12615–12624.
[6] Q. G. Fu, D. Flament, J. D. Coltz, T. J. Ebner, Temporal encoding of
movement kinematics in the discharge of primate primary motor and
premotor neurons, Journal of Neurophysiology 73 (1995) 836–854.
[7] M. M. Churchland, K. V. Shenoy, Temporal complexity and heterogene-
ity of single-neuron activity in premotor and motor cortex, Journal of
Neurophysiology 97 (2007) 4235–4257.
[8] N. G. Hatsopoulos, Q. Xu, Y. Amit, Encoding of movement fragments
in the motor cortex, Journal of Neuroscience 27 (2007) 5105–5114.
[9] T. N. Aflalo, M. S. A. Graziano, Possible origins of the complex topo-
graphic organization of motor cortex: reduction of a multidimensional
space onto a two-dimensional array., Journal of Neuroscience 26 (2006)
6288–6297.
[10] M. Saleh, K. Takahashi, N. G. Hatsopoulos, Encoding of coordinated
reach and grasp trajectories in primary motor cortex, Journal of Neu-
roscience 32 (2012) 1220–1232.
[11] A. Leo, G. Handjaras, M. Bianchi, H. Marino, M. Gabiccini, A. Guidi,
E. P. Scilingo, P. Pietrini, A. Bicchi, M. Santello, E. Ricciardi, A
synergy-based hand control is encoded in human motor cortical areas,
eLife 5 (2016) e13420.
18
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[12] N. Hatsopoulos, A. Suminski, Sensing with the Motor Cortex, Neuron
72 (2011) 477–487.
[13] M. S. A. Graziano, Ethological Action Maps: A Paradigm Shift for the
Motor Cortex, Trends in Cognitive Sciences 20 (2016) 121–132.
[14] W. Truccolo, G. M. Friehs, J. P. Donoghue, L. R. Hochberg, Primary
Motor Cortex Tuning to Intended Movement Kinematics in Humans
with Tetraplegia, Journal of Neuroscience 28 (2008) 1163–1178.
[15] J. A. Pruszynski, I. Kurtzer, J. Y. Nashed, M. Omrani, B. Brouwer,
S. H. Scott, Primary motor cortex underlies multi-joint integration for
fast feedback control, Nature 478 (2011) 387–390.
[16] Elliott J. M., Connolly K. J., A classification of manipulative hand
movements, Developmental Medicine & Child Neurology 26 (1984)
283–296.
[17] L. A. Jones, S. J. Lederman, Human Hand Function, Oxford University
Press, 2006.
[18] H. Nili, C. Wingfield, A. Walther, L. Su, W. Marslen-Wilson,
N. Kriegeskorte, A toolbox for representational similarity analysis.,
PLoS Comput. Biol. 10 (2014) e1003553.
[19] T. A. Yousry, U. D. Schmid, H. Alkadhi, D. Schmidt, A. Peraud,
A. Buettner, P. Winkler, Localization of the motor hand area to a
knob on the precentral gyrus. A new landmark., Brain 120 ( Pt 1)
(1997) 141–157.
[20] M. F. Glasser, T. S. Coalson, E. C. Robinson, C. D. Hacker, J. Harwell,
E. Yacoub, K. Ugurbil, J. Andersson, C. F. Beckmann, M. Jenkinson,
19
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

S. M. Smith, D. C. V. Essen, A multi-modal parcellation of human
cerebral cortex, Nature 536 (2016) 171–178.
[21] S. Lee, G. E. Carvell, D. J. Simons, Motor modulation of afferent
somatosensory circuits, Nat. Neurosci. 11 (2008) 1430–1438.
[22] G. Pfurtscheller, F. H. Lopes da Silva, Event-related EEG/MEG syn-
chronization and desynchronization: basic principles, Clinical Neuro-
physiology 110 (1999) 1842–1857.
[23] C. Tzagarakis, N. F. Ince, A. C. Leuthold, G. Pellizzer, Beta-Band
Activity during Motor Planning Reflects Response Uncertainty, Journal
of Neuroscience 30 (2010) 11270–11277.
[24] D. Denis, R. Rowe, A. M. Williams, E. Milne, The role of cortical
sensorimotor oscillations in action anticipation, NeuroImage 146 (2017)
1102–1114.
[25] K. Nakayashiki, M. Saeki, Y. Takata, Y. Hayashi, T. Kondo, Modu-
lation of event-related desynchronization during kinematic and kinetic
hand movements, Journal of NeuroEngineering and Rehabilitation 11
(2014) 90.
[26] S. Ohara, A. Ikeda, T. Kunieda, S. Yazawa, K. Baba, T. Nagamine,
W. Taki, N. Hashimoto, T. Mihara, H. Shibasaki, Movement-related
change of electrocorticographic activity in human supplementary motor
area proper, Brain: A Journal of Neurology 123 ( Pt 6) (2000) 1203–
1215.
[27] A. K. Engel, P. Fries, Beta-band oscillationssignalling the status quo?,
Current Opinion in Neurobiology 20 (2010) 156–165.
20
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[28] H. Tan, C. Wade, P. Brown, Post-Movement Beta Activity in Senso-
rimotor Cortex Indexes Confidence in the Estimations from Internal
Models, The Journal of Neuroscience: The Official Journal of the So-
ciety for Neuroscience 36 (2016) 1516–1528.
[29] D. Cheyne, S. Bells, P. Ferrari, W. Gaetz, A. C. Bostan, Self-paced
movements induce high-frequency gamma oscillations in primary motor
cortex, NeuroImage 42 (2008) 332–342.
[30] M. Nowak, C. Zich, C. J. Stagg, Motor Cortical Gamma Oscillations:
What Have We Learnt and Where Are We Headed?, Current Behavioral
Neuroscience Reports 5 (2018) 136–142.
[31] S. D. Muthukumaraswamy, Functional Properties of Human Primary
Motor Cortex Gamma Oscillations, Journal of Neurophysiology 104
(2010) 2873–2885.
[32] D. Cheyne, P. Ferrari, MEG studies of motor cortex gamma oscillations:
evidence for a gamma ”fingerprint” in the brain?, Frontiers in Human
Neuroscience 7 (2013) 575.
[33] K. J. Miller, G. Schalk, E. E. Fetz, M. den Nijs, J. G. Ojemann, R. P. N.
Rao, Cortical activity during motor execution, motor imagery, and
imagery-based online feedback, Proceedings of the National Academy
of Sciences of the United States of America 107 (2010) 4430–4435.
[34] J. Diedrichsen, T. Wiestler, J. W. Krakauer, Two Distinct Ipsilateral
Cortical Representations for Individuated Finger Movements, Cereb.
Cortex 23 (2013) 1362–1377.
[35] E. Berlot, G. Prichard, J. O’Reilly, N. Ejaz, J. Diedrichsen, Ipsilateral
finger representations are engaged in active movement, but not sensory
processing, bioRxiv (2018) 285809.
21
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[36] A. Yokoi, S. A. Arbuckle, J. Diedrichsen, The Role of Human Primary
Motor Cortex in the Production of Skilled Finger Sequences, Journal
of Neuroscience 38 (2018) 1430–1442.
[37] M. S. A. Graziano, C. S. R. Taylor, T. Moore, Complex movements
evoked by microstimulation of precentral cortex, Neuron 34 (2002)
841–851.
[38] S. Lee, I. Kruglikov, Z. J. Huang, G. Fishell, B. Rudy, A disinhibitory
circuit mediates motor integration in the somatosensory cortex., Nat.
Neurosci. 16 (2013) 1662–1670.
[39] I. Choi, J.-Y. Lee, S.-H. Lee, Bottom-up and top-down modulation of
multisensory integration, Current Opinion in Neurobiology 52 (2018)
115–122.
[40] R. Hari, N. Forss, S. Avikainen, E. Kirveskari, S. Salenius, G. Rizzolatti,
Activation of human primary motor cortex during action observation: A
neuromagnetic study, Proceedings of the National Academy of Sciences
95 (1998) 15061–15065.
[41] M. Tani, Y. Ono, M. Matsubara, S. Ohmatsu, Y. Yukawa, M. Kohno,
T. Tominaga, Action observation facilitates motor cortical activity in
patients with stroke and hemiplegia, Neuroscience Research 133 (2018)
7–14.
[42] Y. Meirovitch, H. Harris, E. Dayan, A. Arieli, T. Flash, Alpha and
Beta Band Event-Related Desynchronization Reflects Kinematic Reg-
ularities, Journal of Neuroscience 35 (2015) 1627–1637.
[43] P. Jezzard, P. M. Matthews, S. M. Smith, Functional Magnetic Reso-
nance Imaging, An Introduction to Methods, Oxford University Press,
Oxford, 2001.
22
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[44] J. F. Hipp, M. Siegel, BOLD fMRI Correlation Reflects Frequency-
Specific Neuronal Correlation, Current biology: CB 25 (2015) 1368–
1374.
[45] J. C. W. Siero, D. Hermes, H. Hoogduin, P. R. Luijten, N. F. Ramsey,
N. Petridou, BOLD matches neuronal activity at the mm scale: a
combined 7t fMRI and ECoG study in human sensorimotor cortex.,
NeuroImage 101 (2014) 177–184.
[46] E. L. Hall, S. E. Robson, P. G. Morris, M. J. Brookes, The relationship
between MEG and fMRI, NeuroImage 102 Pt 1 (2014) 80–91.
[47] F. I. Arce-McShane, C. F. Ross, K. Takahashi, B. J. Sessle, N. G. Hat-
sopoulos, Primary motor and sensory cortical areas communicate via
spatiotemporally coordinated networks at multiple frequencies, Pro-
ceedings of the National Academy of Sciences 113 (2016) 5083–5088.
[48] R. A. Adams, S. Shipp, K. J. Friston, Predictions not commands: active
inference in the motor system., Brain Struct Funct 218 (2013) 611–643.
[49] R. C. Oldfield, The assessment and analysis of handedness: the Edin-
burgh inventory., Neuropsychologia 9 (1971) 97–113.
[50] J. Gross, S. Baillet, G. R. Barnes, R. N. Henson, A. Hillebrand,
O. Jensen, K. Jerbi, V. Litvak, B. Maess, R. Oostenveld, L. Parkkonen,
J. R. Taylor, V. van Wassenhove, M. Wibral, J.-M. Schoffelen, Good
practice for conducting and reporting MEG research, Neuroimage 65
(2013) 349–363.
[51] J. W. Peirce, PsychoPy: Psychophysics software in Python, Journal of
Neuroscience Methods 162 (2007) 8–13.
23
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[52] J. W. Peirce, Generating stimuli for neuroscience using PsychoPy,
Frontiers in Neuroinformatics 2 (2009).
[53] C. Blais, Random without replacement is not random: Caveat emptor,
Behavior Research Methods 40 (2008) 961–968.
[54] E. Freud, S. N. Macdonald, J. Chen, D. J. Quinlan, M. A. Goodale,
J. C. Culham, Getting a grip on reality: Grasping movements directed
to real objects and images rely on dissociable neural representations,
Cortex 98 (2018) 34–48.
[55] R. Oostenveld, P. Fries, E. Maris, J.-M. Schoffelen, FieldTrip: Open
source software for advanced analysis of MEG, EEG, and invasive elec-
trophysiological data, Computational Intelligence and Neuroscience
2011 (2011) 156869.
[56] Y. Zhang, M. Brady, S. Smith, Segmentation of brain MR images
through a hidden Markov random field model and the expectation-
maximization algorithm., IEEE Trans Med Imaging 20 (2001) 45–57.
[57] S. M. Smith, Fast robust automated brain extraction., Hum. Brain
Mapp. 17 (2002) 143–155.
[58] M. Jenkinson, C. F. Beckmann, T. E. J. Behrens, M. W. Woolrich,
S. M. Smith, FSL., NeuroImage 62 (2012) 782–790.
[59] A. M. Dale, B. Fischl, M. I. Sereno, Cortical surface-based analysis. I.
Segmentation and surface reconstruction., NeuroImage 9 (1999) 179–
194.
[60] B. Fischl, A. Liu, A. M. Dale, Automated manifold surgery: con-
structing geometrically accurate and topologically correct models of
the human cerebral cortex., IEEE Trans Med Imaging 20 (2001) 70–80.
24
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[61] M. Jenkinson, P. R. Bannister, M. Brady, S. Smith, Improved Opti-
mization for the Robust and Accurate Linear Registration and Motion
Correction of Brain Images, NeuroImage 17 (2002) 825–841.
[62] L. Griffanti, G. Douaud, J. Bijsterbosch, S. Evangelisti, F. Alfaro-
Almagro, M. F. Glasser, E. P. Duff, S. Fitzgibbon, R. Westphal,
D. Carone, C. F. Beckmann, S. M. Smith, Hand classification of fMRI
ICA noise components, NeuroImage 154 (2017) 188–205.
[63] M. F. Glasser, S. N. Sotiropoulos, J. A. Wilson, T. S. Coalson, B. Fis-
chl, J. L. Andersson, J. Xu, S. Jbabdi, M. Webster, J. R. Polimeni,
D. C. Van Essen, M. Jenkinson, WU-Minn HCP Consortium, The
minimal preprocessing pipelines for the Human Connectome Project,
NeuroImage 80 (2013) 105–124.
[64] M. Jenkinson, S. Smith, A global optimisation method for robust affine
registration of brain images., Med Image Anal 5 (2001) 143–156.
[65] D. N. Greve, B. Fischl, Accurate and robust brain image alignment
using boundary-based registration, NeuroImage 48 (2009) 63–72.
[66] A. Walther, H. Nili, N. Ejaz, A. Alink, N. Kriegeskorte, J. Diedrichsen,
Reliability of dissimilarity measures for multi-voxel pattern analysis,
NeuroImage 137 (2016) 188–200.
[67] N. N. Oosterhof, T. Wiestler, P. E. Downing, J. Diedrichsen, A com-
parison of volume-based and surface-based multi-voxel pattern analysis,
NeuroImage 56 (2011) 593–600.
[68] J. Diedrichsen, N. Kriegeskorte, Representational models: A com-
mon framework for understanding encoding, pattern-component, and
representational-similarity analysis, PLoS Computational Biology 13
(2017).
25
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[69] J.-D. Haynes, A Primer on Pattern-Based Approaches to fMRI: Prin-
ciples, Pitfalls, and Perspectives, Neuron 87 (2015) 257–270.
[70] J. D. Power, K. A. Barnes, A. Z. Snyder, B. L. Schlaggar, S. E. Petersen,
Spurious but systematic correlations in functional connectivity MRI
networks arise from subject motion, NeuroImage 59 (2012) 2142–2154.
[71] J. Vrba, S. E. Robinson, Signal Processing in Magnetoencephalography,
Methods 25 (2001) 249–271.
[72] G. Nolte, The magnetic lead field theorem in the quasi-static approxi-
mation and its use for magnetoencephalography forward calculation in
realistic volume conductors, Physics in Medicine and Biology 48 (2003)
3637–3652.
[73] B. D. V. Veen, W. V. Drongelen, M. Yuchtman, A. Suzuki, Localization
of brain electrical activity via linearly constrained minimum variance
spatial filtering, IEEE Trans. Biomed. Eng (1997) 867–880.
[74] M. Guggenmos, P. Sterzer, R. M. Cichy, Multivariate pattern analysis
for MEG: A comparison of dissimilarity measures, NeuroImage 173
(2018) 434–447.
[75] O. Ledoit, M. Wolf, A well-conditioned estimator for large-dimensional
covariance matrices, Journal of Multivariate Analysis 88 (2004) 365–
411.
[76] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello,
O. Etard, N. Delcroix, B. Mazoyer, M. Joliot, Automated anatomical
labeling of activations in SPM using a macroscopic anatomical parcel-
lation of the MNI MRI single-subject brain, NeuroImage 15 (2002)
273–289.
26
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

[77] T. E. Nichols, A. P. Holmes, Nonparametric permutation tests for
functional neuroimaging: a primer with examples., Hum. Brain Mapp.
15 (2002) 1–25.
[78] J. D. Power, A. Mitra, T. O. Laumann, A. Z. Snyder, B. L. Schlaggar,
S. E. Petersen, Methods to detect, characterize, and remove motion
artifact in resting state fMRI, NeuroImage 84 (2014).
[79] S. Afyouni, T. E. Nichols, Insight and inference for DVARS, NeuroIm-
age 172 (2018) 291–312.
27
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

2. Materials and Methods
2.1. Participants and Experimental Design
All data were acquired according to the local university research ethics com-
mittee approval in line with the Declaration of Helsinki (Cardiff University
School of Psychology Research Ethics Committee: EC.17.03.14.4874 and
EC.17.04.11.4885) All participants provided written informed consent and
met local MRI and MEG safety criteria.
Ten right-handed participants were recruited in the main study (Age range:
22-30; Mean age: 24.0; Age SD: 2.8; 5 Female). Participants were not
currently taking any psychoactive medications, and were right-handed ac-
cording to the Edinburgh Handedness Inventory49. No participants had a
history of any disorder affecting tactile sensory or motor function or any
history of neurological illness. Each participant undertook five experimen-
tal sessions: two MRI scan sessions, two MEG recording sessions, and one
behavioural testing session. All participants undertook the behavioural test-
ing session first; the subsequent order of the fMRI and MEG sessions was
counterbalanced, leaving a minimum of two weeks between any one MRI
and MEG session to minimise the effects of magnetic noise on the MEG
signal50. The datasets generated and analysed during the current study are
available from the corresponding author on reasonable request.
2.2. Motor task and kinematic data acquisition
During all sessions participants were engaged in a motor task involving the
production of a range of 26 hand movements (Table 1) with the right hand
while wearing a fibre-optic kinematic data glove (Data Glove 14 Ultra; Fifth
Dimension Technologies: 5DT, Orlando, FL, USA). Kinematic data were
28
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

acquired across 14 independent fibre optic channels (one proximal and one
distal sensor per digit, plus one sensor between each digit pair) at 60 Hz.
The behavioural task was implemented in PsychoPy (Version 1.84.20)51,52
using the Python Computer Graphics Kit (CGkit: cgkit.sourceforge.net)
SDK wrapper for the 5DT data glove.
Each recording session was divided into task runs; each task run was com-
posed of blocks of a specific movement; each block comprised individual
movement trials; details of the number runs, blocks, and trials are speci-
fied for MEG and fMRI sessions respectively below. Instructions were pre-
sented on a screen in the testing environment. Each task run contained one
block of each of the 26 movement types, ordered using a random-without-
replacement selection method. Progressive determination effects were min-
imised by maximising the range of different conditions in each run; present-
ing all 26 movements once per run53. At the beginning of each movement
block, participants were shown a 3 second video of the movement to be
produced (Video S1). Participants were cued to produce the movement in
question in each subsequent movement trial of the block by an expanding
and contracting horizontal bar. In each movement trial the bar began at
a fully contracted width, coloured red, indicating that the hand should be
static and in a resting flat position. The bar subsequently turned green and
began to expand symmetrically at its left and right flanks. Once it reached
its maximal width, the bar began to contract back to its original width.
Once the bar reached its original contracted width, it turned red, signifying
the end of the movement trial. Participants were instructed to pace their
movements to coincide with the period of expansion and contraction of the
green bar, such that their hand assumed a flat position at the beginning
and end of each trial, corresponding to the time that the static red bar was
presented. The motor task was conducted in a behavioural testing lab, in
29
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

the MRI scanner, and in the MEG scanner, as detailed below.
None of the grasping tasks in this study engaged participants with real
objects; previous work has differentiated motor activity with or without
real objects in anterior intraparietal sulcus, but not primary motor cortex:
as such an object-free study design seemed appropriate for a study focusing
on M154.
2.3. Kinematic recording session
During the behavioural testing session participants performed five runs of
the motor task. Participants were seated at a desk with their right forearm
supported on a memory foam mount, while wearing the data glove. Partici-
pants viewed instructions presented on a 14 inch laptop display. Each move-
ment block comprised a 3 second video of the movement to be produced, a 1
second preparation period and 8 subsequent movement trials; each compris-
ing 1.6 seconds of movement (green expanding/contracting bar), followed
by a 0.8 second rest period (red static bar). The transition of the bar from
red to green was defined as the go signal. A break period of up to 15 sec-
onds was permitted between each movement block; participants advanced
the task with a key-press using their left hand. Excluding break periods
each task run was 10 minutes and 3.2 seconds in duration. The four task
runs yielded 33 minutes and 16.8 seconds of kinematic data recording per
participant.
2.4. Kinematic movement model
For each participant kinematic data from the behavioural, MRI, and MEG
sessions were each processed in parallel. This yielded a separate kinematic
model from each session type for each participant. These models were used
30
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

in subsequent multivariate fMRI and MEG analysis; they captured the kine-
matic similarities and differences of the 26 distinct movements under study.
Initially the kinematic data from each session and each movement block
were epoched into individual movement trials using the time of onset of the
green bars and averaged. The resulting 14 channels of data represented the
average pattern of displacement of the hand during a movement trial for
a given movement, termed the kinematics of the movement: the motion of
the hand without reference to the forces that produce this motion. In order
to compare this signature of kinematic activity for each possible pairing of
the 26 movements the activity pattern of each of the 14 recording channels
was correlated using Pearson’s correlation coefficient, subject to the Fisher
Z-transformation, and averaged to yield a single measure of the similarity of
kinematics across each movement pair. The resulting value was transformed
back into a Pearson’s r-value and used to construct a 1-r dissimilarity matrix
for each movement pair.
The kinematic dissimilarity matrices were averaged across task runs within-
participant to yield an fMRI, MEG, and behavioural kinematic model for
each participant. The split-half consistency and inter-session consistency
of these models is outlined in Figure S1. A grand average across partic-
ipant and session kinematic model (Figures 1 and S2) was computed and
subject to hierarchical clustering; this resulting clustering was applied to all
visualisations of the kinematic model.
2.5. Muscle model
An independent EMG dataset was acquired in order to construct a model
of hand movement dissimilarity on the basis of muscle activity in the hand.
An independent cohort of ten participants (Age range: 20-30; mean age:
31
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

25.1; age SD: 3.57; 5 female) undertook a more detailed EMG recording
than was feasible during the MEG session, while performing the same 26
hand movement. EMG data were acquired using a Biosemi Active 2 system
with a 32 channel headbox (Biosemi B.V. Amsterdam). Muscle activity
was recorded using touchproof flat active electrodes. Electrodes 1-15 were
placed as labelled in Figure S16, electrode 16 was used to rereference the
EMG data in subsequent analysis and was placed on the bony protrusion
of the elbow. There were also CMS and DRL electrodes, which served as a
ground/reference during recording in the Biosemi software; they were placed
on the palmar side of the wrist. The EMG data were recorded at 2048Hz.
The EMG recording sessions mirrored the design and setup of the kine-
matic recording session outlined above and were informed by previous fMRI
MVPA studies of digit flexion 3. Five runs were recorded in total, each con-
taining 26 trials (one for each of the movements). The EMG data were
processed using Fieldtrip55. EMG data were rereferenced to electrode 16,
rectified and low-pass filtered (fourth order Butterworth filter: 40 Hz), and
epoched relative to earliest measured muscle onset in any EMG channel us-
ing an adaptive threshold (activity duration threshold: 200ms) (Hooman
Sedghamiz: Matlab File Exchange: Automatic Activity Detection in Noisy
Signals using Hilbert Transform.) This results in individual trials of 2.0s in
duration. These trials were baselined using the fixation cross window at the
start of each trial. EMG trial data were then subject to multivariate noise
normalisation by weighting channels in trial by the error covariance across
the different channels in order to more accurate quantify the true differences
between the muscle activity across different movements. The normalised tri-
als were averaged into 5 folds. A Mahalanobis distance comparing each of
the muscle activity of each of 26 different movement types was calculated
using a cross-validated leave-one-out approach. In each iteration, the muscle
32
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

activity patterns from one fold were assigned to fold A and the muscle ac-
tivity data from the remaining four folds were assigned to fold B; distances
were calculated between all possible pairs of the 26 movement muscle ac-
tivity recordings across these two folds (Equation (3)). Distance measures
were calculated across all possible pairs of cross-validation folds. An average
muscle model across all ten participants’ data was generated and used to
probe the spatial and temporal encoding of muscle based dissimilarities in
the brain using fMRI and MEG (Figures 1 and S2).
2.6. Ethological action movement model
An alternative ethological action based model was constructed based on
more recent evidence of ethological maps in primate M113 , and therefore
categorises movements on the basis of their specific action, namely pre-
hensile movements, sub-categorised into precision grip and power grip, and
non-prehensile movements 17 (Figure 1). The ethological action model was
subject to hierarchical clustering for visualisation.
2.7. MRI data acquisition
MR data were acquired using a Siemens 7T Magnetom system (Siemens
Healthcare, Erlangen, Germany) with a 32-channel head coil. Blood oxy-
genation level dependent (BOLD) fMRI was acquired with a T2*-weighted
multi-slice gradient echo planar imaging (EPI). True axial slices were po-
sitioned for optimal coverage of the left and right anatomical hand knob 19
(TR/TE: 1500/25 ms, resolution: 1.2 mm isotropic, 22 axial slices, flip an-
gle:90; GRAPPA factor: 2; anterior-posterior phase-encoding direction; 391
measurements). Magnetization prepared rapid gradient echo (MPRAGE)
structural MRI data were acquired to facilitate BOLD EPI slice placement
33
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

and for cortical surface reconstruction (TR/TE: 2200/2.82 ms, isotropic res-
olution: 1.0 mm, GRAPPA factor = 2). An additional gradient echo BOLD
EPI acquisition of 4 volumes was acquired using posterior-anterior phase-
encoding direction for distortion correction.
2.8. fMRI behavioural task
During the fMRI acquisitions participants performed a total of ten runs of
the motor task (5 runs per MRI session). Participants were led supine with
their right forearm supported against their right hip and their elbow sup-
ported by a foam pad, while wearing the data glove. Participants viewed
instructions via a mirror mounted on the transmit coil and a projector screen
mounted at the end of the bore. Each movement block comprised of a 3 sec-
ond instruction screen (“Prepare to Move”), a 3 second video of the move-
ment to be produced, and a 1 second further instruction screen (“Move”),
followed by 5 movement trials, each comprising 1.6 seconds of movement
(green expanding/contracting bar), followed by a 0.4 second rest period
(red static bar). Each movement block was 17 seconds. In addition to the
movement blocks, 8 rest blocks were included in each task run; rest blocks
were of equivalent duration to movement blocks and comprised of a 3 second
instruction screen (“Rest”), a 3 second video of a static resting hand, and a
1 second further instruction screen (“Rest”), followed by the same period of
expanding and contracting bar visual stimuli as the fMRI movement blocks.
Rest blocks were positioned randomly in each run, excluding self-adjacency.
2.9. Structural MRI data preprocessing
MPRAGE data were subject to reorientation, bias-field correction and brain
extraction using the FMRIB Software Library (FSL) fsl anat tool 56,57,58
prior to cortical surface reconstruction using FreeSurfer Version 5.3.059,60.
34
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

2.10. fMRI data analysis
2.10.1. fMRI preprocessing and general linear modelling
fMRI data were subject to standard preprocessing, including motion cor-
rection with MCFLIRT 61, brain extraction using BET57, and high pass
temporal filtering (100 second threshold). fMRI data were not subject to
spatial smoothing. All fMRI data were subject to manual independent com-
ponents analysis denoising62. Distortion correction was undertaken using
FSL Topup to estimate a fieldmap image for use in FSL FUGUE 63. Undis-
torted BOLD EPI data were co-registered with structural MPRAGE data
using Boundary-Based-Registration from FMRIB’s Linear Registration Tool
(FLIRT) implemented in epi reg64,61,65. Example fMRI timeseries from a
single voxel located in the anatomical hand knob is presented for four par-
ticipants on a single session in Figure S15.
For each participant and each fMRI run, fMRI data were analysed using
a first-level general linear modelling (GLM) approach implemented in FSL
FEAT 58 using FMRIBs Improved Linear Model (FILM) to estimate time
series autocorrelation and pre-whiten each voxel. Each of the 26 movements
was modelled with a separate boxcar regressor with gamma-HRF convolu-
tion and its temporal derivative, giving a total of 52 regressors. Parameter
estimates were calculated, contrasting each movement type against the rest
condition; these voxel-wise maps and an estimate of the residuals from the
GLM were resampled into the respective participants’ structural space and
used in subsequent representational similarity analysis (RSA).
2.10.2. fMRI multivariate noise normalisation
In order to account for the spatial structure of the noise inherent to fMRI
data, spatial prewhitening of the parameter estimates from each participant
35
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

and each fMRI task run was conducted. The residuals (R) from the first-level
GLM analysis provided an estimate of data not fit by the model regressors
across voxels (V) and time (T), from which a V x V covariance matrix (b
Σ)
can estimate the noise structure across voxels (Equation (1))66.
b
Σ = 1
TRTR(1)
The noise covariance structure was combined with the voxel-wise parameter
estimates (P) for a given movement type (k) to generate a spatially pre-
whitened parameter estimate (P∗
k: Equation (2)):
P∗
k=Pkb
Σ−1
2(2)
2.10.3. fMRI surface-based searchlight representational similarity analysis
A surface-based representational similarity analysis searchlight approach
was used to identify regions in which the multivariate pattern of BOLD
activity mirrored the kinematic and categorical models. This surface-based
analysis constrained the voxels under consideration in each searchlight to the
grey matter and prevented the issue of sampling of voxels that span a sul-
cus in a single searchlight, which is inherent to volumetric approaches67. A
searchlight was constructed at the centre of each vertex within the individual
participants’ anatomical cortical surface region corresponding to the field of
view of their task fMRI data (Figure S8). Each searchlight had a diameter
of 10mm. The region of interest of each searchlight was projected from 2D
surface to 3D volumetric space using the Connectome Workbench Tool63,
masked by a FMRIB Automatic Segmentation Tool grey matter map56 and a
mask excluding voxels spanning across sulci in the FreeSurfer reconstruction
36
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

to improve spatial specificity. Spatially pre-whitened parameter estimates
were extracted from the resulting volumetric region corresponding to each
searchlight.
2.10.4. fMRI cross-validated distance measures
Within each searchlight the similarity between each of the spatially pre-
whitened voxel-wise parameter estimates corresponding to each of the 26
different movement types was calculated using a cross-validated leave-one-
out approach to avoid the possibility of over-fitting the data 68,69. In each
iteration, the parameter estimate maps from one fMRI task run was as-
signed to fold A and the parameter estimate maps from the remaining nine
task fMRI runs were assigned to fold B; squared Euclidean distances were
calculated between all possible pairs of the 26 movement parameter esti-
mate maps across these two folds (Equation (3)). Distance measures were
calculated across all possible pairs of cross-validation folds and averaged 66 .
The use of spatially pre-whitened parameter estimate combined with the
cross-validation approach yielded cross-validated Mahalanobis distance rep-
resentational dissimilarity matrices (RDMs) comparing each of the activa-
tion patterns across all possible pairings of the 26 movements. For example,
calculation of the distance between movement k and movement l in one
iteration:
d2
Cr ossvalidated Mahalanobis (P∗
k, P ∗
l)=(P∗
k−P∗
l)A(P∗
k−P∗
l)T
B(3)
The correspondence between the fMRI-derived RDM in each searchlight
and the candidate kinematic and theoretical models was assessed using a
Spearman’s rank correlation, with the resulting ρ(rho) value was plotted
in each searchlight’s central vertex on the cortical surface. For statistical
37
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

inference a fixed effects randomisation test18 was applied on the individual
participant level: correlations using 10,000 condition-label randomisations
were undertaken in each searchlight. From each of the permutations, the
spatial peak ρvalue (rho) was extracted from across the cortical surface,
forming a maximum accuracy distribution from which an omnibus threshold
(α= 0.01) was extracted. The resulting thresholded ρ-value surface maps
for each participant were resampled onto the Human Connectome Project
32k surface (S1200.L.pial.MSMAll.32k fs LR.surf.gii), binarised and used to
form a heatmap corresponding to the spatial distribution of the each model
fit across participants. In light of the interest in contrasting the kinematic
and muscle models, a comparison of the corresponding unthresholded Spear-
man’s ρcortical surface maps was undertaken using a Wilcoxon signed-rank
test (one-sided), subject to FDR correction (α= 0.05).
2.11. fMRI motion considerations
Variability in the magnitude of fMRI motion across different movement con-
ditions has the potential to influence the observed pattern of results. The
potential for noise induced by participant motion has been mitigated in a
number of ways. First, all data has been subject to ICA denoising to remove
any characteristic motion artefacts62. Second, the multivariate analysis of
fMRI data emoloyed herein used spatial prewhitening of the parameter esti-
mates to account for voxel-wise variability in noise to not downweight voxels
with high error variance and to account for noise covariance between vox-
els66. Finally DVAR values were calculated for each fMRI timeseries (D:
temporal derivative of time courses, VARS: root mean squares variance over
voxels). These values quantify for each frame of an fMRI acquisition the
magnitude of signal intensity change in comparison in volume N compared
with volume N-1, as per the following formula:
38
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

DV ARS(∆I)i=rD[Ii(−→
x)−Ii−1−→
x]2E(4)
Where Iiis image intensity at locus −→
xon frame i; angle brackets denote
the spatial average over the whole brain 70 . DVARs are able to quantify
corruption of fMRI acqusition due to head motion. DVAR values were ex-
tracted for volumes corresponding to each of the 26 hand movements for
all participants; the resulting distribution of DVAR values is presented in
Figure S13. The profiles of very limited motion across participants during
each session of around 10 minutes in duration also demonstrate high quality
data acquisition (Figure S14).
2.12. MEG data acquisition
MEG signals were measured continuously at 1200Hz during the motor task
using a whole-head 275-channel axial gradiometer CTF MEG system (CTF,
Vancouver, Canada) located inside a magnetically shielded room. An ad-
ditional 29 reference channels were recorded for noise cancellation purposes
and the primary sensors were analysed as synthetic third-order gradiome-
ters71. Three electromagnetic coils were placed on three fiduciary locations
(nasion, left and right pre-auricular) and their position relative to the MEG
sensors were recorded continuously during each experimental block. The
head surface and fiducial locations were digitized using an ANT Xensor
digitizer (ANT Neuro, Enschede, Netherlands) prior to the MEG recording.
2.13. MEG behavioural task
During the MEG data acquisitions participants performed a total of ten runs
of the motor task (5 runs per MEG session). Participants were sitting up-
right with their right forearm and elbow supported on a foam armrest, while
39
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

wearing the data glove. Participants viewed instruction on a back-projected
screen in front of them from a projector mounted outside the shielded room.
Each movement block comprised of a 2 second period with a central fixation
cross, a 3 second video of the movement to be produced, and a 1 second in-
struction screen (“Prepare to Move”) followed by five movement trials, each
comprising 1.6 seconds of movement (green expanding/contracting bar), fol-
lowed by a 0.8 second rest period (red static bar). Each movement block
was 18 seconds. The order of movement blocks was randomised within each
task run; each movement was presented once per task run.
2.14. Data glove movement onset detection: MEG sessions
The 14 channels of data glove recordings collected during the MEG sessions
were aligned with the MEG acquisitions using the onset of the green bar and
were epoched alongside the MEG data. Epoched data glove recordings were
subject to onset segmentation using an adaptive threshold (activity duration
threshold: 200ms) (Hooman Sedghamiz: Matlab File Exchange: Automatic
Activity Detection in Noisy Signals using Hilbert Transform.). A conserva-
tive estimate of movement onset was derived by taking the earliest signal
onset detected across the fourteen data glove channels for each movement
trial. The resulting movement onset time was used to epoch MEG data in
further analysis.
2.15. MEG data analysis
2.15.1. MEG preprocessing
Each participant’s head shape was digitized using Xensor digitizer soft-
ware (ANT software BV, Enschede, The Netherlands). All MEG anal-
ysis was conducted using the Fieldtrip toolbox for EEG/MEG-analysis55
40
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

(Donders Institute for Brain, Cognition and Behaviour, Radboud University
Nijmegen, the Netherlands. See http://www.ru.nl/neuroimaging/fieldtrip).
Co-registration was performed in a two stage process: first the fiducial lo-
cations were marked on the T1 structural for that participant; the head
digitization data was then used to align the data with the MRI, subject
to manual adjustment. Alignment was undertaken independently for data
from the two MEG sessions.
Data from each movement type were epoched from the 10 task runs and
concatenated into a new dataset containing 10 blocks, each containing 5
movement trials. The fixation cross and movement trials were epoched from
the overall block. The movement trials were defined relative to the data
glove defined movement onset time (movement trial time: 2 s; pre-onset
time: 0.5s, post-onset time: 1.5s). The fixation cross period was used as
a baseline for the 5 movement trials within each movement block. A high
pass filter of 1 Hz and a low pass filter of 100 Hz were applied. MEG
analyses were conducted across three frequency bands: alpha (7-14 Hz),
beta (15-30 Hz) and gamma (30-100 Hz). All of the movements trials for
a given movement type were concatenated across the 10 task runs, creating
a dataset comprising 50 repeats of a movement. At this point the data
was visually inspected and those trials containing artefacts were removed
from further analysis up to a maximum of 10 trials, such that the minimum
number of movements trials per movement included in further analysis was
40.
2.15.2. MEG source reconstruction
In order to reconstruct oscillatory activity at brain locations directly com-
parable across participants, the individual anatomical MRI was non-linearly
warped to the MNI MRI template. The MNI template was divided into a
41
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

10 mm isotropic grid and the inverse of the previously calculated non-linear
warp was used to warp the template grid into the anatomical space of each
participant. Sensor leadfields were calculated using a semi-realistic volume
conduction model based on the individual anatomy72. The temporal evolu-
tion of source activation at each location in the brain was estimated using
a linearly constrained minimum variance (LCMV) beamformer algorithm73
with the optimal dipole orientation at each voxel estimated using singular
value decomposition (SVD). Virtual sensors were then reconstructed from
all 3294 voxels by multiplying the sensor level data by the corresponding set
of optimised weights. At this stage data were subject to multivariate noise
normalization74,75: we calculated the error covariance matrix at sensor level
and then used this combined with the filters from the LCMV to create the
virtual sensor data. This means that sensors with more noise would be
down-weighted compared to those with less noise. At this stage the data
was also down-sampled to 600 Hz to reduce computational cost.
2.15.3. MEG temporal representational similarity analysis
The MEG data were split to produce 10 partitions and then averaged within
each partition to perform a cross-validated representational similarity anal-
ysis to avoid the possibility of over-fitting the data 68,69. RSA was performed
across time using a sliding time window with a width of 20 ms and a time
step of 5 ms creating 396 time windows across 2 seconds of the movement
trial (0.5s rest; 1.5s movement). After selecting virtual sensors within the left
hemisphere motor region of the AAL atlas76 (Precentral L, 31 sources; Fig-
ure S9), the frequency-filtered MEG signal measured during each movement
type was compared using a cross-validated leave-one-out approach within
each time width. In each iteration, the signals from one MEG data par-
tition were assigned to fold A, and the signals from the remaining nine
42
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

partitions were assigned to fold B; squared Euclidean distances were calcu-
lated between all possible pairs of the 26 signals across the two folds and
averaged66. The use of multivariate noise normalisation to account for spa-
tial autocorrelation in the MEG signal yielded subject-wise cross-validated
Mahalanobis distance RDMs comparing the alpha, beta, or gamma-band
signal in the motor ROI across all possible pairings of the 26 movements 74.
Participant-level motor ROI RDMs were averaged in order to perform a
fixed-effects analysis. The correspondence between the MEG-derived RDMs
and the candidate kinematic and theoretical models across time was assessed
using a Spearman’s rank correlation, with the resulting ρ(rho) values plotted
for each time window. In light of the interest in contrasting the kinematic
and muscle models, these were each assessed in a partial correlation to dis-
count the contribution of the other. Randomization testing was used for
statistical inference77, whereby candidate model RDMs were shuffled 1000
times and time-resolved correlation coefficients were recomputed in order
to estimate an empirical null distribution. P-values were calculated using
a cluster thresholding approach across time. To correct for multiple com-
parisons, the cluster-forming threshold was set to P<0.01 and clusters
in the correlation time-courses corresponding to each candidate model were
thresholded against the maximal cluster distribution (α= 0.001).
To assess the maximal correlation possible with our data, each participant’s
RDM was correlated with the average cross-subject RDM; the correlations
were then averaged to obtain an upper bound of the noise ceiling 18.
2.15.4. MEG: action observation analysis
MEG from the period of action observation during the instruction video
preceding each movement block were epoched using the same approach as
43
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

the MEG data recorded during movement. The fixation cross and action ob-
servation trials were epoched from the overall block. The action observation
trial was defined relative to the video stimulus onset time (pre-onset time:
0.5s, post-onset time: 3.0s). The fixation cross period was used as a base-
line for the action observation period. Temporal representational similarity
analyses were conducted using the same apporoach as the MEG movement
data, as described above.
2.15.5. MEG motion considerations
MEG analysis included multivariate noise normalisation to account par-
tially for the effects of motion, where each channels are normalised by an
estimate of error covariance across different sensors; this process has been
demonstrated to substantially improve multivariate analyses of MEG data74.
Motion parameters for all MEG acquisitions were extracted and analysed
to rule out the possibility of excessive head motion as a potential driving
force behind any observed patterns of brain activity. Rotational and trans-
lational displacement for each participant and each experimental session are
presented in Figure S11. In addition, the motion parameters during each
movement block were extracted and the resulting distribution is presented
across the 26 different movement types (Figure S12). The profiles of motion
across participants demonstrate a high quality data acquisition.
2.16. Electromyography with MEG
Electromyography (EMG) data were acquired simultaneously with MEG
data. Three surface EMG electrodes were attached to the right hand un-
derneath the data glove, positioned on abductor pollicis brevis (APB), first
44
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

dorsal interosseus (FDI) and abductor digiti minimi (ADM). The area un-
der the electrodes was exfoliated and cleaned with alcohol prior to data
acquisition. EMG signals were recorded at 1200Hz.
EMG data were initially subject to a bandpass filter (20-500Hz) and a notch
filter (50 Hz). EMG data were epoched and baselined alongside the MEG
data. Epoched EMG data were subject to manual artefact rejection. Sig-
nals from the three electrodes during each epoch were independently subject
to a Hilbert transform and smoothing (5 ms window) prior to activity on-
set segmentation using an adaptive threshold (activity duration threshold:
200ms) (Hooman Sedghamiz: Matlab File Exchange: Automatic Activity
Detection in Noisy Signals using Hilbert Transform.). A conservative esti-
mate of muscle activity onset was derived by taking the earliest signal onset
detected across the three EMG channels for each movement trial. Due to
constraints of electrode placement alongside the kinematic data glove, mea-
sures of activity onset were not robustly measured in all participants. EMG
onset data are presented in order to validate the data glove measures of
movement onset, which have been used to epoch the MEG data (Figures 1
and S10).
Video S1: Compilation of instructional videos used at the begin-
ning of movement blocks in all testing sessions. Movement labels
are provided for reference only; labels were not included during the task
(VideoS1.mov).
Video S2: Visualisation of multidimensional scaling of grand aver-
age kinematic model constructed across participants and sessions.
(VideoS2.mov).
45
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Average St. Dev.
Kinematic model: fMRI 0.6318 (0.6010-0.6627) 0.1027 (0.0850 - 0.1297)
Kinematic model: behavioural 0.5812 (0.5506-0.6119) 0.1021 (0.0845 - 0.1289)
Kinematic model: MEG 0.6577 (0.6328 -0.6826) 0.0828 (0.0686 - 0.1046)
Muscle model 0.5286 (0.4822-0.5749) 0.1542 (0.1276-0.1948)
Table S1: Inter-subject consistency of kinematic and muscle models. Spearman’s ρ
(rho) vales. Figures in brackets represent confidence intervals.
.
Figure S1: Data-driven kinematic models constructed for each participant and each
session type exhibit strong split-half and inter-session consistency. Muscle model
reproducibility data also presented.
46
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Kinematic (fMRI) Kinematic (Behav) Ethological Action Muscle
Participant Left-contra Right-ipsi Left-contra Right-ipsi Left-contra Right-ipsi Left-contra Right-ipsi
10.265 - 0.474 - 0.262 - 0.382 - 0.259 - 0.410 0.272 - 0.305 0.283 - 0.420 0.286 - 0.467
20.250 - 0.563 0.251 - 0.500 0.249 - 0.384 0.255 - 0.326 0.249 - 0.470 0.250 - 0.441 0.276 - 0.525 0.277 - 0.379
3 0.247 - 0.430 0.252 - 0.405 0.252 - 0.382 0.263 - 0.308 0.249 - 0.301 0.249 - 0.296 0.279 - 0.543 0.285 - 0.432
4 0.256 - 0.411 0.257 - 0.407 0.252 - 0.334 0.277 - 0.327 0.264 - 0.483 - 0.304 - 0.418 0.276 - 0.376
5 0.254 - 0.453 0.251 - 0.272 0.252 - 0.378 0.246 - 0.363 0.248 - 0.349 - 0.270 - 0.542 0.281 - 0.467
6 0.258 - 0.412 0.274 - 0.369 0.260 - 0.344 0.266 - 0.327 0.257 - 0.421 0.274 - 0.281 0.285 - 0.383 0.300 - 0.374
7 0.250 - 0.478 0.251 - 0.435 0.241 - 0.548 0.240 - 0.509 - - 0.271 - 0.577 0.275 - 0.467
8 0.269 - 0.416 0.263 - 0.333 0.255 - 0.453 0.279 - 0.345 0.252 - 0.291 - 0.284 - 0.397 0.289 - 0.439
9 0.250 - 0.386 0.251 - 0.284 0.243 - 0.360 - 0.253 - 0.450 0.296 - 0.305 0.270 - 0.515 0.275 - 0.467
10 0.259 - 0.525 0.254 - 0.479 0.261 - 0.549 0.263 - 0.451 0.249 - 0.481 - 0.276 - 0.0.587 0.274 - 0.467
Table S2: Summary of supra-threshold Spearman’s ρ(rho) values for each participant across each cortical surface searchlight.
Data presented for left and right hemisphere. Kinematic behavioural session data were used to construct an additional model based on movement
data outside of the scanner also used in an equivalent searchlight analysis to the fMRI kinematic model. Individual participant surfaces searchlights
are presented in Figures 2, S4, S6 and S7).
47
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S2: The average kinematic model across participants for each session type, the
muscle model derived from an independent cohort of participants, and the categorical
ethological action model.
48
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S3: The average kinematic model constructed from the data glove measures
acquired during the (A) behavioural session, (B) fMRI sessions, and (C) MEG ses-
sions; presented alongside the individual subject models for each participant and
session.
49
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S4: fMRI searchlight analysis conducted using the kinematic model con-
structed from data glove recordings made during the behavioural testing session.
Evidence of the encoding of kinematic data in contralateral primary motor cortex persists using
independent data glove recordings while participants were sitting upright in a more naturalistic po-
sition. Comparison with data presented in Figures 1 and 2. Supra-threshold range of Spearman’s
ρfor each participant presented in Table S2.
50
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S5: MEG searchlight analysis during action observation Evidence of a significant
peak in the correspondence between the kinematic model and beta band MEG data in the period of
action observation (315 ms - 380 ms). The green line indicates the onset of the stimulus video; the
blue regions indicate significant peaks in representational similarity between MEG data and the
motor model; the dashed line indicates noise ceiling. Comparison with MEG temporal searchlight
results presented in Figures 1 and 4.
51
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S6: Individual participant cortical searchlight results using an muscle model
of movement encoding. Cortical heatmaps of the left (A) and right (B) hemisphere, show
limited but consistent encoding of an action model in the left post-central gyrus, contralateral
to movement. Heatmaps were constructed from individually thresholded cortical searchlights
for each participant using a single categorical action model (C) (Omnibus threshold, α= 0.01,
maximum accuracy distribution calculated from peak correlation value across 10,000 searchlight
permutations with label-switching.) Supra-threshold range of Spearman’s ρfor each participant
presented in Table S2.
52
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S7: Individual participant cortical searchlight results using an ethological
action model of movement encoding. Cortical heatmaps of the left (A) and right (B) hemi-
sphere, show limited but consistent encoding of an action model in the left post-central gyrus,
contralateral to movement. Heatmaps were constructed from individually thresholded cortical
searchlights for each participant using a single categorical action model (C) (Omnibus threshold,
α= 0.01, maximum accuracy distribution calculated from peak correlation value across 10,000
searchlight permutations with label-switching.) Supra-threshold range of Spearman’s ρfor each
participant presented in Table S2.
53
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S8: Heatmap surface visualisation of fMRI data coverage across participants
on inflated (A) and midthickness (B) surfaces.
54
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S9: Surface visualisation of the left hemisphere motor region of the AAL atlas
used in MEG temporal searchlight analysis: precentral L.
55
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S10: EMG muscle activity onset time. A. Example of EMG onset detection using adaptive threshold. B. Smoothed envelope of signal
(Hilbert transform) used for automatic onset detection. C. Violin plots visualising the distribution of EMG onset times for trials from which onset
could be detected across all participants. Bars represent interquartile range, white dots represent median values which are plotted in Figure 1.
Where available, EMG data were used to validate the data glove derived measures of movement onset used to epoch the MEG data. Time point
0 corresponds to the data-glove movement onset time.
56
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S11: MEG motion timeseries across participants and sessions. Blue: X-axis,
orange: Y-axis, yellow: Z-axis
57
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S12: MEG motion comparison across movement conditions.
58
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S13: fMRI motion comparison across movement conditions using DVARS
presented in arbitrary units70,78,79 .
59
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S14: fMRI motion displacement plots. Plots of absolute (blue) and relative (orange)
motion calculated using FSL MCFLIRT for each participant and each fMRI task run; motion
correction was undertaken prior to ICA denoising.
60
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S15: Example fMRI axial slices and single voxel timeseries from 4 participants. fMRI timeseries data extracted from a single
voxel (red crosshairs) for four participants. Data presented were subject to high-pass filter (100 seconds). fMRI data were not subject to spatial
or temporal smoothing.
61
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S16: EMG recordings from an independent cohort used to generate a muscle
model of hand movement. A) Schematic demonstrating electrode placement for EMG record-
ings on both the palmar and dorsal surface of the hand and forearm covering these muscles: 1
first dorsal interosseus (FDI), 2-3 dorsal interosseus muscles, 4 abductor digiti minimi, 5 ab-
ductor pollicis brevis (APB), 6 adductor pollicis, 7-9 lumbrical muscles, 10 flexor carpi ulnaris,
11 flexor carpi radialis, 12-14 flexor digitorum superficialis and flexor digitorum profundus, 15
flexor pollicis longus. B) Example EMG trace from one participant for three different movements
(squeeze, grip sphere and finger abduct) showing electrodes 1-15 across the 2s movement time
window.
62
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;

Figure S17: Noise ceiling calculation for spatial searchlight using fMRI data. To assess
the spatial consistency in RDMs calculated from fMRI data at each vertex, each participant’s RDM
was correlated with the average cross-subject RDM; the correlations were then averaged to obtain
a vertex-wise upper bound of the noise ceiling.
63
.CC-BY 4.0 International licenseIt is made available under a
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint. http://dx.doi.org/10.1101/613323doi: bioRxiv preprint first posted online Apr. 18, 2019;