A mosaic of whole-body representations on the human precentral gyrus

Main

Nearly a century ago, Penfield and colleagues pioneered the motor homunculus, a topographical map of the human PCG, by electrically stimulating the cortical surface and observing which body parts moved in response17,18. Since that time, the motor cortex has been mapped extensively in non-human primate models using single-neuron recordings6,7,8, microstimulation of both short and long durations1,2,3,4,5 and lesion and inactivation studies4,9,10. This body of work has provided strong evidence that the motor cortex is more fractionated and intermixed than originally implied by the motor homunculus, with related body parts overlapping in the primary motor cortex2,22,23 (for example, wrist and fingers) and whole-body regions overlapping in the premotor cortex6,7,24 (for example, arm and face). Long stimulation durations have more recently revealed that the motor cortex is also organized by behaviour in ethological ‘action maps’ as opposed to a pure body part organization3,4,21,25.

In humans, however, exploration of the motor cortex has largely been limited to lower-resolution recording techniques—such as electrocorticography11,12,13 and functional magnetic resonance imaging (fMRI)14,15,16,26—as well as gross stimulation mapping in the operating room17,18,19,20. Studies reporting the effects of lesions27,28 and stimulation17,18,19,20 have found results that suggest a separation of leg, arm and face movement regions in the human PCG (that is, stimulation at any single location does not typically cause simultaneous arm, leg and face movement, with some exceptions19,29). On the other hand, fMRI and electrophysiological studies have emphasized a greater degree of intermixing13,14,15,16, including beta power modulation and blood-oxygenation-level-dependent deactivation spreading far from the somatotopic hotspot16,30, overlapping receptive fields14 and intereffector regions with whole-body intermixing15,31. However, these experiments cannot resolve detailed representations at the single-neuron level, and thus the representation of the body in the human motor cortex remains relatively unknown at a single-neuron scale. The crown (exposed surface) of the human PCG is thought to be anatomically composed largely of Brodmann area 6 (a premotor area), with the primary motor cortex lying within the central sulcus32,33,34. If the crown of the human PCG is functionally homologous to the macaque premotor cortex, we might expect to find dense intermixing of arm and leg movements dorsally3,7,24,35, and arm and orofacial movements ventrally3,6,24.

In previous work using microelectrode arrays capable of recording brain activity at single-neuron resolution, we have shown that a small, anatomically distinct area of the cortex in the dorsal PCG (referred to as the ‘hand knob’36) contained intermixed representations of the entire body37,38 (including all four limbs and head and face movements), where the limbs were interrelated with a compositional neural code. However, it is unknown whether a similar organization exists in the middle and ventral PCG; in addition to providing fundamental insight into the neural representation of movement in humans, an answer to this question is needed to inform the design of brain–computer interfaces (BCIs) that restore motor function to people with paralysis, including speech.

Here we revisit the motor representation of the whole body across a wide span of the PCG using microelectrode recordings from 8 human participants (Fig. 1a) implanted with 20 microelectrode arrays as part of BCI clinical trials. These participants had either spinal cord injury, amyotrophic lateral sclerosis (ALS) or brainstem stroke, resulting in severe motor impairment with varying levels of residual movement across individuals. All clinical trial procedures were conducted with oversight from ethics, institutional and regulatory bodies; before enrolment, an extensive consent process engaged potential participants and reinforced that no direct benefit from study participation was expected (see Methods). These neural recordings collectively sampled the length of the PCG spanning from the canonical arm area (near the superior frontal sulcus) to the canonical tongue and throat area (near the Sylvian fissure), and together constitute, to our knowledge, the first comprehensive motor map of the crown of the human PCG at single-neuron resolution.

Neural tuning to the whole body in the PCG

We assessed neural tuning to speech and movements of the face, head, arms and legs in eight participants in a visually cued movement task (Fig. 1b and Supplementary Tables 1 and 2). If participants could not physically complete a movement due to paralysis, they were instructed to attempt to complete the movement in a light, controlled and isolated manner, to the best of their ability (see Supplementary Video 1 for example movements from T12, who retained the greatest residual arm and leg movement among the participants). Example electrodes demonstrating strong neural tuning to various movements are shown in Fig. 1c.

To quantify neural tuning at a coarse scale, we measured the degree to which the neural population activity evoked by each movement differed from a ‘do nothing’ baseline condition (Fig. 2a). The majority of electrode arrays (15 of 20) recorded robust neural modulation in response to all tested movements, with the remainder showing modulation to almost all movements. We also analysed neural tuning at an individual electrode level and found that intermixed tuning to all body regions was frequently present within single electrodes (Extended Data Fig. 1) and spike-sorted single neurons (Extended Data Fig. 2).

Although all regions of the PCG were tuned to the whole body, sorting the arrays by anatomical location revealed that larger neuronal firing rate modulation aligned with the canonical motor homunculus, with arrays in the dorsal PCG showing a preference for arm movements, and arrays in the ventral PCG preferring orofacial movements (Fig. 2b). However, two findings also diverged from the classical homunculus. First, a majority of the arrays in the canonical orofacial region preferred speech to orofacial movement, particularly the most inferior ventral arrays (for example, T16-v1) and one middle PCG array (T15-m1). Second, the three arrays in between these speech-preferring arrays were more broadly tuned to the whole body. Arrays T12-v1 and T15-v1 in particular had substantial limb tuning and were also most strongly activated by orofacial movements as opposed to speech.

Whole-body decoding from PCG sites

Next, to investigate the extent to which the whole body is represented in a movement-specific way throughout the PCG, we tested whether a recurrent neural network decoder could classify between movements belonging to the same category using neural activity from single trials (Fig. 3). With the exception of speech, where four similar words could not be classified above chance in dorsal arrays, classification accuracy within each movement category was above chance in all arrays (Fig. 3a and Supplementary Table 3). This indicates that a differentiable whole-body representation exists at each point in the human motor cortex, but with a different emphasis depending on the location (with the dorsal PCG showing greater decodability of limb movements, and the ventral PCG showing greater orofacial and speech decodability).

Of note, the two most superior ventral arrays (T12-v1 and T15-v1) had some of the highest classification accuracies across all movement types. Figure 3b shows an example confusion matrix between all movements using activity from array T12-v1, which had an average 86% decoding accuracy. This result further supports the idea of a superior ventral orofacial zone with broad tuning to the whole body, but relatively weaker tuning to speech (than the speech zones in the middle PCG and inferior ventral PCG). Note that the neural representation of non-canonical movements (for example, orofacial movements in the dorsal PCG) was both multidimensional and relatively large (40% as large as that of canonical movements; Extended Data Fig. 3), suggesting that it may not simply reflect generic deactivation16 or minor spillover from nearby areas.

When interpreting the classification results above, it is important to keep in mind three potential confounds: (1) some arrays simply record more tuned neurons than others, increasing their classification accuracy in general; (2) potential variability in how precisely individual movements are executed or attempted; and (3) some categories may be harder to classify than others due to the particular selection of movement conditions (for example, speech, in which the four tested words are very similar to each other). The third confound could explain why classification accuracy for speech is generally lower than other categories of movement, in contrast to the normalized firing-rate modulation results above, in which many arrays were speech dominant.

Four functional zones in the PCG

Next, we applied principal component analysis (PCA) to the normalized firing-rate modulation profiles and classification accuracies of each array to visualize array-tuning characteristics in a low-dimensional space (Fig. 4a). PCA identified two major axes that explained 73% of the variance in tuning profiles: a ‘speech–face versus limb’ axis that discriminated the inferior ventral arrays (and the middle PCG array T15-m1) from the dorsal arrays, and a ‘breadth of tuning’ axis that discriminated the superior ventral arrays from the others. The coefficients for these two PCA axes (Fig. 4b) were similar for both the firing-rate modulation profiles (solid lines) and the classification accuracies (dashed lines), indicating general agreement between these two complementary methods of assessing whole-body tuning.

We coloured each array according to its location along the top two PCA axes, and displayed this information topographically to reveal regional patterns in tuning characteristics (Fig. 4c). On the basis of these results, as well as what is known from previous work (particularly studies identifying the middle PCG as a speech–language region32,39), we hypothesize that there are four main functional zones in the sampled area of the PCG (Fig. 4c): one dorsal zone emphasizing arm movements, one inferior ventral zone emphasizing speech, one middle PCG zone also emphasizing speech (array T15-m1), and one superior ventral orofacial zone between the speech zones that is most broadly tuned to the whole body (arrays T12-v1 and T15-v1). We also analysed fMRI resting state data from the Human Connectome Project32 and found evidence of two resting-state networks that parcellate the PCG into four zones (Extended Data Fig. 4): one language-related resting-state network with hotspots in the inferior ventral PCG and middle PCG (accounting for the speech zones), and one hand–arm-related resting-state network with hotspots in the dorsal PCG and superior ventral PCG (which could explain the strong limb tuning in the superior ventral PCG).

It is possible that anatomical variability across individuals could cause other zones to sometimes appear on the crown of the PCG. Array T16-m1 in particular appears to be an outlier in exhibiting strong head-related modulation and classification accuracy. Using the Human Connectome Project cortical parcellation procedure applied to the preoperative neuroimaging data for T16, we identified array T16-m1 as being on the border of the premotor eye field, an area that may be highly tuned to movements of the eyes and head and that typically lies within the precentral sulcus (Extended Data Fig. 5).

Finally, using preoperative neuroimaging data obtained from four participants (T12, T15, T16 and T17), we were able to assess array locations relative to recently identified inter-effector regions in the somatocognitive action network15, as well as the hand–face border40 (Extended Data Fig. 6). We found that all arrays appeared to lie outside of the inter-effector regions. In addition, the superior ventral arrays with broad whole-body tuning (T12-v1 and T15-v1) were placed below the hand–face border and within the canonical face region of the motor cortex. These results show that strong whole-body tuning is not limited to previously identified inter-effector regions or to the hand–face border.

A compositional whole-body neural code

Finally, we sought to elucidate the representational structure of the whole-body code found throughout the PCG. In previous work investigating the hand-knob area37, we found a ‘compositional’ neural code linking all four limbs together that had two main features: (1) correlated representations of homologous movements across limbs (for example, wrist flexion and ankle flexion), and (2) representation of the limb itself independent of the movement.

Here we also found strong correlations between all four limbs in multiple regions of the PCG. Figure 5a shows example arm–leg correlations for the dorsal PCG and superior ventral PCG (Extended Data Fig. 7 shows the correlation matrices for all regions and limb pairs). Representational similarity between limbs was present for most limb pairs and most arrays (Fig. 5b), although the arms appeared to be most strongly correlated with one another (significant for 17 of 20 arrays). These results show that inter-limb neural correlations linking all four limbs are a general feature of the human motor cortex, appearing even in orofacial and speech regions. Using video data from participant T12, who could still make visible arm and leg movements, we confirmed that this effect is not the result of inadvertent movement of uninstructed effectors (Extended Data Fig. 8 and Supplementary Video 1).

Next, we tested for the existence of limb-coding activity in the form of ‘laterality’ neural dimensions that would separate the left from right arms independently of the movement itself, as found previously in the hand-knob area37,41. Consistent with previous results, we found that a laterality dimension is a dominant feature of the dorsal PCG, appearing clearly in the top two principal components of the neural activity (Fig. 5c; Extended Data Fig. 9 shows scatterplots for all arrays). Laterality information was less prevalent in other areas of the PCG (Fig. 5c,d), indicating that this feature is probably unique to the dorsal PCG.

Discussion

Here we created a comprehensive motor map of the crown of the human PCG at single-neuron resolution, leveraging microelectrode array recordings from 20 arrays across 8 individuals with paralysis (Extended Data Fig. 10). This map revealed that distinguishable, multi-dimensional whole-body representations exist at all sampled points of the PCG. We also found that these whole-body representations are abstract, linking together the legs and arms in a partially limb-independent representation of motor action. A limb-independent movement code could facilitate the transfer of learned actions from one limb to another42,43, and was present throughout all sampled areas of the PCG. Our results also shed new light on the regional arrangement of the human motor cortex. We found evidence that the sampled extent of the PCG between the Sylvian fissure and superior frontal sulcus appears to contain four functional zones, each with their own whole-body representations: (1) a dorsal arm–hand zone, (2) a middle speech zone (also highlighted in recent work32,39), (3) a broadly tuned, superior ventral orofacial zone, and (4) an inferior ventral speech zone.

Our results support a growing body of evidence that the motor cortex is more integrative than implied by the classical homunculus. Recent results from lower-resolution recording methods in humans (for example, fMRI and electrocorticography) have demonstrated overlapping body part representations in the PCG13,14,15,31. In animal models, the idea of ethological action maps as an organizational principle for the motor cortex has come to prominence in the past few decades3,4,21,25. Long-train stimulation delivered to premotor areas creates complex, multi-joint action patterns that can span body regions and appear to be pieces of ethological behaviours3. This theory could explain the overlapping whole-body representations that we observed in the PCG. For instance, the dorsal PCG could implement behaviours that coordinate the limbs, whereas the ventral PCG may implement behaviours involving combined orofacial and limb movement (for example, self-feeding); more broadly, the entire body may be involved in many behaviours due to the need to stabilize the limbs, trunk and head. Our findings of speech-specific areas in the PCG also support the idea of a behavioural organization in the motor cortex.

Although the crown of the human PCG is sometimes referred to as the primary motor cortex19,20, anatomical evidence has shown that the crown of the PCG in humans is composed largely of Brodmann area 6 (a premotor area), whereas Brodmann area 4 (primary motor cortex) lies within the central sulcus32,33,34. Our results support a homology between the crown of the human PCG and the macaque premotor cortex, which is known to contain more intermixed6,7,24 and abstract44,45 movement representations than the primary motor cortex. Specifically, the superior ventral region of the PCG was strongly tuned to both orofacial and arm–hand movement, and may be homologous to the macaque ventral premotor cortex, which is known to contain grasping, reaching and orofacial-related information6,46,47 (see also ref. 48); likewise, the dorsal PCG may be homologous to the macaque dorsal premotor cortex, which contains both forelimb and hindlimb movement representations5,7,24,35,47. A recent cortical atlas based on neuroimaging data from the Human Connectome Project32 has also divided the crown of the human PCG into a series of premotor areas that are distinct from the central sulcus (Extended Data Fig. 5).

The degree of intermixing observed here appears to be at odds with electrical stimulation studies in people, which generally have revealed separate regions for leg, arm and face movements that do not often intermix at the same location, consistent with Penfield’s original homuncular diagram (despite some exceptions found in one-fifth of patients19). Lesions to the human motor cortex have also been reported to cause focal deficits49,50 (for example, lesions to the hand-knob area have been reported to affect only arm and hand function). This could indicate that the whole-body representations observed here, although relatively large (approximately 40% the size of canonical movement representations), may be ancillary to the main function of the areas, for instance, reflecting inputs from neighbouring regions that support coordination of movements involving multiple effectors. However, other interpretations are possible. For instance, it may simply be the case that electrical stimulation activates mainly the strongest representation (especially when stimulating for short durations). In addition, lesions are not often contained only in the grey matter on the crown of the PCG, and may cause focal deficits due to white matter damage or damage to adjacent primary motor areas in the central sulcus.

One important question is whether the results shown here would generalize to able-bodied individuals without spinal cord injury, brainstem stroke or ALS. The dense intermixing of the whole body found at each sampled point and within each participant, as well as the limb-independent movement code found across all participants, suggest that these features are not specific to a particular disease or level of paralysis (see Supplementary Table 4 for motor strength scores). Although our results from the ventral PCG are limited only to individuals with brainstem stroke or ALS, results still reproduced for different stages and levels of paralysis (for example, T15 and T12 both showed broad orofacial and limb tuning in the superior ventral PCG, despite T15 being much more paralysed than T12). In addition, a large body of work has shown that the extent of motor reorganization in the adult brain after amputation is quite limited, suggesting that loss of limb function does not greatly alter motor topography51,52 (see also recent neuroimaging results in spinal cord injury53). Furthermore, although myelin and motor cortical function is known to degrade in ALS54,55, it is not clear why this would cause body parts to intermix more than they otherwise would normally. Finally, it is possible that people with severe motor impairment may unintentionally co-activate other effectors to compensate for their weakness when attempting isolated movements, potentially recruiting neural representations of other body parts. We minimized this possibility by coaching participants to attempt movements in a light, controlled manner and confirmed in T12—who retained visible arm and leg movements—that tuning to all four limbs was not due to overt co-movement. However, most other participants had more limited residual movement, so comparable behavioural verification was not feasible to rule out subtle co-activation. Nonetheless, only recordings from able-bodied people can provide a definitive answer as to whether these results generalize to healthy individuals.

Finally, our study charts the first map of how well movements from each category can be decoded at single-neuron resolution throughout the PCG. This map can inform the design of intracortical BCIs aiming to restore speech, arm, hand and leg movement to people with paralysis56,57,58. To achieve the highest performance with the minimal number of electrodes, our results show that speech BCIs may benefit from targeting the inferior ventral and middle PCG while avoiding the superior ventral region. By contrast, the superior ventral region may be an underappreciated complement to the dorsal PCG for arm–hand BCIs48,59. Overall, the widespread decodability of the whole body throughout the PCG is advantageous for BCIs seeking to restore multiple functions with a small footprint, although not all body parts are equally decodable from all locations60.

Methods

Experimental procedures

Study participants

This study includes data from eight participants who each gave informed consent before any experimental procedures. Participants T5, T11, T12, T15, T16 and T17 were enrolled in the BrainGate2 Neural Interface System clinical trial (ClinicalTrials.gov Identifier: NCT00912041, registered 3 June 2009), and represent all BrainGate2 participants who were enrolled at the time the data were collected and analysed. This pilot clinical trial was approved under an Investigational Device Exemption (IDE) by the US Food and Drug Administration (FDA; Investigational Device Exemption #G090003). Permission was also granted by the Stanford University Institutional Review Board (IRB; protocol #20804), the Mass General Brigham IRB (protocol #2009P000505), the University of California, Davis IRB (protocol #1843264), the Emory University IRB (protocol #00003070) and the Providence VA Healthcare IRB. Participants C1 and C2 were enrolled under a separate multi-site clinical trial (ClinicalTrials.gov Identifier: NCT01894802, registered 10 July 2013), which was also conducted under an IDE from the US FDA and approved by the IRBs at the University of Pittsburgh and the University of Chicago. C1 and C2 represent all participants enrolled under NCT01894802 who were available at the time of data collection to participate in this study, given competing demands on participant time for other research goals. All research was performed in accordance with relevant guidelines and regulations.

Recognizing that cortical mapping and neural activity recording in humans introduces distinctive ethical and philosophical considerations, ethics oversight was embedded from the outset. All risks (including surgical risks) arise solely from the parent safety and feasibility trials (and not the present study), each approved and monitored by local IRBs. Additional safeguards were provided by multiple and in part redundant additional layers of protection, including continuous oversight from local Clinical Oversight Committees including an independent medical monitor, a Data Safety and Monitoring Board, strict adherence to FDA guidance on conduct and safety under our IDE, clinical trial offices from the US National Institutes of Health (NIH), and engagement with the Neuroethics Working Group (NEWG) of the NIH BRAIN Initiative, to ensure adherence to the highest standards of safety and ethical rigour. All study procedures complied with the Declaration of Helsinki, the Belmont Report, CIOMS guidance and the NIH BRAIN Neuroethics Principles. In the parent clinical trials, a months-long, dialogue-driven consent process safeguards participant autonomy and adapts to evolving communication needs, protecting participant privacy. During this process, we explained that there would be no direct benefits from the implanted device. Within the trials, proactively embedded neuroethics expertise encourages continual reflection on risk–benefit balance, participant welfare and societal impact.

T5 is a right-handed man, 69 years of age at the time of this study, with tetraplegia due to cervical spinal cord injury (classified as C4 AIS-C), which occurred approximately 9 years before enrolment in the clinical trial. T5 has two 96-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the hand-knob area of the left PCG. The hand-knob area was identified anatomically by preoperative MRI. T5 has full movement of the face and head and the ability to shrug his shoulders. Below the level of the spinal cord injury, T5 has very limited voluntary motion of the legs and arms.

T11 is a right-handed man, 38 years of age at the time of this study, with tetraplegia due to a cervical spinal cord injury (classified as C4 AIS-B), which occurred approximately 14 years before. T11 has two 96-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left dorsal PCG, targeting the hand-knob area as identified anatomically by preoperative MRI. T11 has full movement of the face and head with very limited voluntary motion of the arms.

T12 is a left-handed woman, 67 years of age at the time of this study, with slowly progressive bulbar-onset ALS. T12 was diagnosed at age 59 with an ALS-Functional Rating Scale score of 26 at the time of study enrolment. T12 has four 64-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left (language dominant, motor non-dominant) hemisphere, based on preoperative anatomical MRI, fMRI and cortical parcellation using the Human Connectome Project (HCP)32 multi-modal parcellation pipeline (see Extended Data Fig. 5 for cortical parcellation results). Two arrays were placed in the inferior frontal gyrus (not included in this study), and two arrays were placed in the ventral PCG, targeting area 6v; see Willett et al. for more details56. T12 remains functionally independent with 3–4 out of 5 strength (assessed using the Medical Research Council (MRC) scale) in all limbs, but is anarthric (able to vocalize, but unable to produce intelligible speech).

T15 is a left-handed man, 45 years of age at the time of this study, with ALS (ALS-Functional Rating Scale score of 23 at the time of study enrolment). T15 has four 64-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left (language dominant, motor non-dominant) PCG, based on preoperative anatomical MRI and HCP cortical parcellation (see Extended Data Fig. 5 for cortical parcellation results). One array was placed targeting area 55b, two arrays targeting area 6v and one array targeting area 4; see Card et al. for more details61. T15 has limited orofacial movement with the capacity for vocalization, but is unable to produce intelligible speech. T15 has very limited voluntary motion of the rest of the body.

T16 is a right-handed woman, 52 years of age at the time of this study, with tetraplegia and dysarthria due to a pontine stroke approximately 19 years before enrolment in the BrainGate2 clinical trial. T16 has four 64-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in her left PCG: two in the hand-knob area (targeting area 6d), one in the speech-related ventral premotor cortex (targeting ventral area 6v) and one in the middle PCG (targeting area 55b). Implant area targets were identified by the HCP multimodal cortical parcellation procedure (see Extended Data Fig. 5 for the cortical parcellation results). Examination of post-implant array locations indicated that the middle PCG array appears to be on the border between area PEF and 55b. T16 is able to speak slowly and quietly, but speech cadence is reduced due to poor diaphragm voluntary control. She has limited voluntary control of her upper extremities, with some shoulder motion and some slow and contractured wrist and finger movements. She has limited-to-no voluntary control of her lower extremities. Sensation for T16 is fully intact.

T17 is a right-handed man, 33 years of age with a history of rapidly progressive ALS. T17 has six 64-channel intracortical microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left hemisphere, based on preoperative anatomical MRI, task-based fMRI and cortical parcellation using the HCP multimodal parcellation pipeline. Two arrays were placed in the dorsal PCG (targeting area 6d), two arrays were placed in the ventral PCG (targeting area 6v) and two arrays were placed in area 55b. At the time of this study, T17 is quadriplegic, anarthric and ventilator dependent. His only remaining volitional motor control is over his extra-ocular movements.

C1 is a right-handed man, 57 years of age at the time of implant, who presented with a C4-level ASIA-D spinal cord injury that occurred 35 years before implant. C1 has four microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left hemisphere. Two 96-channel microelectrode arrays were implanted in the arm and hand area of the motor cortex and two other 32-channel arrays (not included in this study) were implanted in the somatosensory cortex. Targeted array placement was based on functional neuroimaging (fMRI) of the participant attempting to make movements of the hand and arm, within the constraints of anatomical features such as blood vessels and cortical topography; see Greenspon et al. for more details62. C1 had no control of the intrinsic or extrinsic muscles of the right hand but retained the ability to move his arm with noted weakness in many upper limb muscles. He retained impaired, but largely functional, movement of the other limbs, full control of head and face movement, and could speak fluently. The data included here were collected 3.25 years post-implant.

C2 is a right-handed man, 60 years of age at the time of implant, who presented with a C4-level ASIA-D spinal cord injury and right brachial plexus injury that occurred 4 years before implant. C2 has four microelectrode arrays (Blackrock Microsystems; 1.5-mm electrode length) placed in the left hemisphere. Two 96-channel microelectrode arrays were implanted in the arm and hand area of the motor cortex and two other 32-channel arrays (not included in this study) were implanted in the somatosensory cortex. Targeted array placement was based on functional neuroimaging (fMRI and magnetoencephalography) of the participant attempting to make movements of the hand and arm, within the constraints of anatomical features such as blood vessels and cortical topography62. C2 retained full control of his entire body except for right hand and arm movement. He could speak fluently. The data included here were collected 0.75 years post-implant.

Neural signal processing

For each participant, neural signals were recorded from the microelectrode arrays using the NeuroPort system (Blackrock Microsystems). The signals were then analogue filtered (fourth-order Butterworth with corners at 0.3 Hz to 7.5 kHz) and digitized at 30 kHz (250-nV resolution). The subsequent digital filtering and neural feature extraction methods differed between participants due to variations in the systems at different sites.

For T5, T11 and T12 the signals were decimated to 15 kHz and band-pass filtered between 250 Hz and 4,900 Hz using a fourth-order zero-phase non-causal Butterworth filter. Linear regression referencing (LRR) was then applied to further reduce ambient noise artefacts63 before spike detection. Spike threshold crossing detection was implemented using a −4.5 × RMS threshold applied to each electrode, where RMS is the electrode-specific root mean square of the time series voltage recorded on that electrode.

For T15, the signals were band-pass filtered between 250 Hz and 5 kHz using a fourth-order zero-phase non-causal Butterworth filter, and LRR was then used to reduce noise artefacts. Spike threshold crossing detection was implemented using a −4.5 × RMS threshold.

For T16, each electrode was high-pass filtered with a 250 Hz cut-off using a fourth-order zero-phase non-causal Butterworth filter. LRR was used for noise reduction and artefact removal with parameters computed post-filtering from a dedicated reference block at the beginning of the session. A −3.5 × RMS threshold was applied to each electrode for spike threshold crossing detection.

For T17, the signals were decimated to 15 kHz and band-pass filtered between 250 Hz and 5,000 Hz using a fourth-order zero-phase non-causal Butterworth filter. LRR was then applied to further reduce ambient noise artefacts before spike detection. Spike threshold crossing detection was implemented using a −3.5 × RMS threshold applied to each electrode.

For participants C1 and C2, a high-pass filter (250 Hz) was applied to each electrode before spike detection. Spike threshold crossing detection was implemented using a −4.5 × RMS threshold applied to each electrode.

The resulting spiking data from each participant mentioned above was binned in 20-ms bins for offline analyses and decoding as presented throughout this study.

Overview of data collection sessions and cued movement task

For each participant, neural data were recorded in a single ‘session’ on a scheduled day. During the session, the participant was seated in a chair in front of a computer monitor at an idle and relaxed position (with the exception of participants T5 and T17, who were lying in a bed at an incline). Each participant completed a series of 5–10 min ‘blocks’ of the cued movement task, consisting of an uninterrupted series of trials. Supplementary Table 1 lists all data collection sessions reported in this work. Variation in the number of trials and/or blocks collected for each participant is due to differences in session durations for each participant and their respective comfort and/or fatigue levels.

The cued movement task followed a simple instructed delay paradigm (Fig. 1b). During the instructed delay period, a red square and text appeared in the centre of the screen indicating to the participant that they should prepare to make the specified movement. The instructed delay period varied randomly (except for participants C1 and C2), with the range tailored to preference of each participant for sufficient time to read the prompt and prepare the movement. After the delay, the square turned green and the text indicating the movement changed to ‘go’, at which point the participant executed the movement immediately. The participant was instructed to make the movement if they were able to overtly move that body part, otherwise they were instructed to attempt the movement. They were then directed to continue attempting the movement, or holding the posture of the completed movement until the text changed to ‘return’, at which point the participant relaxed and returned to a neutral posture. Typically, the movement period lasted 1.5 s and the return period lasted 1 s, although these durations were adjusted if participants required more time to complete the task. See Supplementary Table 2 for task timing parameters.

A standardized document describing task details of the whole-body movement sweep was shared with each site before data collection. Participants were instructed to prioritize isolated and consistent attempted movements. The following text summarizes the key task elements and instructions provided in that document.

The session design details included: “This session is to collect open-loop attempted movements for various effectors across the entire body. We ask the participant to attempt or make a single isolated movement which is cued via text on a computer monitor in a delayed movement task. As the participant attempts the cued movement, neural data is recorded”.

Key notes included:

Make sure the idle or neutral body position of the participant is comfortable and that they are not inherently attempting or making any movements while in an idle position

Please try not to move until the go cue occurs

Only move the body part that is being cued, please try not to move any other body part (isolated movements)

Please keep all movements as consistent as possible (for example, move the right ankle up the same way each time)

Sustain the movement throughout the go period (for example, flex the left wrist throughout the duration of the go period until the return period is cued at which point return to idle)

Ask the participant to refrain from speaking during the block (as speaking is one of the tested conditions)

Before session start, please go over the cued movements with the participant (show them demonstrations of the movements).

An excerpt of the script included: “You will be performing a series of single movements. The movements will be cued in text on the screen accompanied by a coloured square. Each trial starts with the cued movement being presented with a red square. The red square means ‘prepare’, so please prepare to make the movement, but try your best not to move until the ‘go’ cue. The ‘go’ cue occurs shortly after when the square turns green and is accompanied by a beep. At this time, please attempt only the movement that is cued. Another beep will then follow with a ‘return’ prompt, indicating that you should return to an idle or neutral state. When presented with a ‘do nothing’ cue, please try not to make any movements and remain at idle”.

In addition to these standardized instructions, participants were coached during a practice block at the beginning of the session to attempt each cued movement in a light and controlled manner, as opposed to straining against their paralysis (which could induce unintended co-movement of other joints). Researchers visually monitored for non-isolated movements and provided feedback as needed. If participants appeared to recruit other joints in an attempt to generate overt motion of the cued joint, they were instructed to instead isolate their movement attempts to the cued joint only, even if doing so would result in no overt motion being generated. Researchers either read the script verbatim or paraphrased in an extended discussion with the participant.

Neural representations of whole-body movements in the motor cortex

Peristimulus time histograms

To generate the peristimulus time histograms shown in Fig. 1c, we first started with binned threshold crossing (TX) spike counts (20-ms bins). For visualization purposes, we denoised the data by convolving the TX counts of each electrode with a Gaussian smoothing kernel (120 ms s.d.). Next, for each electrode, we extracted TX counts for each trial in a −2-s to 2.5-s time window relative to the ‘go’ cue of each trial. For each of the movement conditions (see Fig. 1c legend), we computed the mean and 95% CIs (estimated using MATLAB’s normfit function) for the TX counts in each time bin across all trials. We then scaled the resulting means and CIs (multiplying by \(\frac{1}{0.02{s}}\)) to convert them into units of Hz.

Neural-tuning strength

Note that the neural population analyses discussed in this section rely heavily on the cvVectorStats library of functions for using cross-validation to estimate Euclidean distance between the means of multivariate distributions (and for estimating other statistics that require Euclidean distance, such as Pearson’s correlation). This library has been extensively used in our previous reports37,41,56,64 (see the code repository on GitHub (https://github.com/fwillett/cvVectorStats)).

Tuning strength heat maps. The neural-tuning strength heat maps shown in Fig. 2a were generated using cross-validated estimates of Euclidean distances between the distribution of population-level neural firing rates for a control ‘do nothing’ condition and the distribution of firing rates for a specific movement condition (see Fig. 2a column labels for each movement condition).

We first started with binned TX counts (20-ms bins) for each array. To account for drifts in mean firing rates across the session, the TX rates were mean subtracted within each block (that is, for each electrode, its mean firing rate within each block was subtracted from the binned spike count for each time step). Next, for each trial and electrode, TX counts were averaged in the entire time window following the ‘go’ cue, accounting for a 300-ms reaction time to cue onset (see Supplementary Table 2 for analysis windows for each participant). This yields an N × 1 neural firing rate vector for each trial, where N is the number of electrodes. For each unique movement condition, including the ‘do nothing’ condition, we stacked the firing vectors for each trial to create a T × N matrix, where T is the number of trials for that specific movement condition. Next, we estimated the Euclidean distance between the ‘do nothing’ firing vectors and the firing vectors for each unique movement condition using the cvDistance function in the cvVectorStats library. cvDistance also returns 95% CIs via jackknife resampling (see Supplementary Table 3 for statistical details). If the confidence interval contained 0, the tuning strength was considered insignificant relative to the ‘do nothing’ condition (denoted by a white ‘X’). To compare tuning strengths across arrays and participants whose neural data may differ in signal quality, the Euclidean distances calculated for each array (that is, each row of the heat map) were normalized by the maximum Euclidean distance in that row.

Movement set-wise tuning strength across spatially organized arrays. For Fig. 2b, we plotted the microelectrode arrays as they appeared on the PCG to better interpret how they are tuned as a function of their spatial location. We used anatomical landmarks to position the arrays relative to each other across participants in a generic space.

The brain surface depicted in the left-most panel of Fig. 2b is the HCP’s group-averaged pial surface (HCP S1200 (refs. 32,65)). The superior frontal sulcus (SFS) and inferior frontal sulcus (IFS) are illustrated by hand-drawn black lines and are easily identifiable landmarks despite variability in brain folding across individuals. We then defined the extremes of the crown of the PCG as ‘superior’ and ‘inferior’. Using the MRI-derived brain anatomies for each participant, we approximated the location for each array relative to these four landmarks for plotting purposes (right panels of Fig. 2b).

For each microelectrode array, tuning strength to each movement set (speech, face, head, right arm, right leg, left arm and left leg) was computed by averaging the tuning strengths across the sub-movements within each set (that is, averaging the relevant entries in Fig. 2a before row-wise normalization). Sub-movements with no statistically significant tuning (that is, cells with white ‘X’) were still included in the average. Computing set-wise averages yields a tuning strength matrix of size A × S, where A is the number of arrays and S is the number of movement sets. The set-wise tuning strength matrix was then row-wise normalized by the maximum value. Next, each microelectrode array (square) in Fig. 2b was coloured according to the corresponding value in this normalized matrix.

Recurrent neural network classifier

For Fig. 3, we used the recurrent neural network (RNN) decoding architecture described in our previous work56 to classify movements using single-trial neural recordings. For each trial, binned TX spike count time series from a 4-s window after the go period were used as input to the RNN. Binned TX spike counts were block-wise mean subtracted to remove firing-rate drifts across the session. The RNN was trained on all movement classes and arrays simultaneously, with a separate input layer for each array (using the same methods as in ref. 56, in which separate input layers were used for each day of data). Training separate input layers helps to place the data into a common space before processing by the RNN. To account for different array sizes (64 versus 96 electrodes), data from 64-electrode arrays were zero padded to increase the dimensionality to 96. All hyperparameters were the same as those used in our previous work56, except instead of using the connectionist temporal classification (CTC) loss, we used a cross-entropy loss applied to a linear readout from the last RNN layer (using the PyTorch function torch.nn.CrossEntropyLoss).

We used fivefold cross-validation to estimate classification accuracy (5 RNNs were trained on 80% of the trials each and tested on the remaining 20%). Only a single RNN was trained for each fold to classify all movements across all arrays (as opposed to training RNNs separately for each movement category and array). When computing classification accuracy for a particular movement category, the decoder output was constrained to be one of the relevant movements from that category only (the maximum probability output within the allowable set was chosen as the final output). Classification accuracies were considered statistically significant if their 95% CIs did not intersect chance level. The 95% confidence intervals for the classification accuracies were calculated using binofit in MATLAB. The chance level was calculated as \(\frac{1}{C}\), where C is the number of conditions within each set.

PCA of array-tuning properties

To make Fig. 4a, normalized modulation magnitudes for each category of movement (as depicted in Fig. 2a) and classification accuracies (as depicted in Fig. 3a) were concatenated together for each array to produce a 20 × 14 matrix (20 arrays, 7 + 7 accuracies and normalized modulation magnitudes). Classification accuracies were expressed as success rates varying from 0 to 1, and normalized modulation magnitudes also varied from 0 to 1. PCA was then applied to the rows of this matrix to find the dimensions that explained the most variance in array-tuning properties (these two principal components are depicted in Fig. 4b). Array-tuning properties were then visualized in this two-dimensional space (Fig. 4a), and each array was coloured using an HSV colour scheme. The hue of each point (x, y) in this space was determined by the angle atan2(y, x), the saturation was determined by the magnitude \(\sqrt{{x}^{2}+{y}^{2}}\,\), and the value was set to 0.85. Saturation values were clipped at 1.

Correlation between homologous limb movement representations

Figure 5a and Extended Data Fig. 7 show average pairwise correlations (Pearson’s r) between the neural representations of homologous limb movements.

First, following the same pre-processing steps as in the ‘Tuning strength heat maps’ section, binned TX spike counts for each microelectrode array were block-wise mean removed, and neural firing rate vectors were computed by averaging the TX counts within the go period window (Supplementary Table 2) for each trial. Firing-rate vectors were then concatenated into T × N matrices for each movement condition m, where T is the number of trials for movement m and N is the number of electrodes within an array. Before computing the correlations, we subtracted the average firing rate within each movement set to remove any effector-dependent neural dimensions such as the potentially large laterality and arm versus leg dimensions reported in previous work37,41. For example, for arm movements, we averaged across all neural vectors for movements within the arm set, creating a 1 × N average vector, which was then subtracted from the neural vectors within this arm set. Next, for each pair of movements (mi and mj), the corresponding concatenated firing-rate matrices were used to compute the cross-validated correlation between the mean firing-rate vectors of each movement (cvCorr function from the cvVectorStats library). The correlation matrices were then averaged across arrays within the following sets: the dorsal PCG, the middle PCG, the superior ventral PCG and the inferior ventral PCG. See the row labels of Fig. 2a for how the arrays were grouped into these sets.

For Extended Data Fig. 7, note that the row–column entries of the correlation matrices for ‘right arm–left arm’ and ‘right leg–left leg’ were re-ordered to align movements that are the same in joint-angle space as opposed to extrinsic, Cartesian space. For example, ‘right arm–raise right’ is homologous to ‘left arm–raise left’. This aligns with our previous reports that showed that directional movements were correlated across effectors in intrinsic space37,41.

The heat map in Fig. 5b indicates the average correlation between homologous movements for limb pairs (column entries) for each microelectrode array (row entries). Each cell represents the mean of the diagonal values of the corresponding correlation matrix for each limb pair. A statistical test for significance was performed by a shuffle-control test (the black ‘X’ indicates non-significant correlation between homologous movements). See Supplementary Table 3 for statistical details.

Movement-independent neural coding of laterality

We used PCA to visualize the neural activity, in select microelectrode arrays, in a lower-dimensional space as illustrated in Fig. 5c. First, binned TX spike counts were block-wise mean removed and z-scored (mean subtracted and divided by the standard deviation) for visualization purposes. Next, firing-rate vectors were computed for each trial by averaging the counts within the ‘go’ period windows (Supplementary Table 2). We then concatenated all of the firing-rate vectors for right and left arm movements into a matrix of size T × N, where T is the total number of trials and N is the number of electrodes. PCA was performed on this monolithic matrix and each z-scored firing-rate vector was subsequently projected onto the top two principal components (PCs). The single-trial projections were coloured by the corresponding movement set (right arm trials in red, and left arm trials in blue). See Supplementary Table 3 for details on the number of trials.

The heat map in Fig. 5d summarizes the size of laterality-related tuning using a variation of demixed PCA66 (dPCA; https://github.com/machenslab/dPCA). A core concept of dPCA involves marginalizing neural data across different experimentally manipulated factors or variables. Each marginalization averages across variables not in the set, creating a data tensor that captures the effect of those factors on the neural activity. Using the dPCA library, we applied a cross-validated variance computation to estimate the amount of variance in the neural activity due to each factor while reducing bias41. To reduce bias when estimating variance, we split the trials into twofolds and computed the marginalizations separately for each fold. We then estimated the covariance matrix of each marginalization as \({X}_{1}{X}_{2}^{T}\), where X1 is the marginalization computed on fold 1 and X2 is the marginalization computed on fold 2. We then took the real part of the eigenvalues of \({X}_{1}{X}_{2}^{T}\) to estimate the variance of each component. The data were marginalized into the following four factors: laterality (left or right arm), movement type (arm raise left, arm raise right, hand close, hand open, wrist up, wrist down, wrist left and wrist right), laterality–movement type interaction and time. We computed the cross-validated variance in the aforementioned factors for each microelectrode array. Each cell of the heat map in Fig. 5d represents the cross-validated marginalized variance for each factor (labelled along the x axis) across each microelectrode array.

Intermixed tuning within electrodes

Extended Data Fig. 1a shows matrices that indicate whether an electrode was significantly tuned to an individual movement. To assess significant tuning, we first started with binned TX spike counts that were block-wise mean removed. Next, for each trial and electrode, we averaged the TX counts within the ‘go’ period window. Significance of tuning was then assessed via a two-sample, two-sided Student’s t-test (using the ttest2 function in MATLAB) applied per electrode, where group 1 comprised the single-trial average firing rates for movement m, and group 2 comprised the single-trial average firing rates for the ‘do nothing’ condition (see Supplementary Table 1 for the number of trials per condition). This two-sample t-test was performed for each movement (P < 0.00001 defined significance). Electrodes that were significantly tuned to a movement appear as a filled in circle in Extended Data Fig. 1a.

The heat map in Extended Data Fig. 1b summarizes the fraction of electrodes that exhibited significant tuning to at least one movement within each movement set as computed in Extended Data Fig. 1a. The heat map in Extended Data Fig. 1c summarizes the fraction of electrodes that had statistically significant tuning to each possible number of movement sets (from 0 to 7). The colour bar is clipped at 0.5 to better visualize the range of values. Extended Data Fig. 2a,c,d repeat the corresponding analyses in Extended Data Fig. 1a–c for spike-sorted single units, restricted to units with mean firing rates of 2–200 Hz.

Neural dimensionality of canonical and non-canonical movements

Extended Data Fig. 3 summarizes the dimensionality of neural activity evoked by movement subsets classified as ‘canonical’ (classically homuncular) or non-canonical in the dorsal and ventral PCG regions. In the dorsal PCG, canonical movements were defined as arm and hand movements, whereas non-canonical movements were defined as speech and face movements. In the ventral PCG, speech and face movements were defined as canonical, whereas arm and leg movements were defined as non-canonical.

To estimate the spatial dimensionality of non-canonical movements, we applied cross-validated PCA (see Stringer et al. for further details67). For each array, binned threshold crossing counts were block-wise mean removed and reshaped into a four-dimensional data tensor of shape N × C × T × R, where N is the number of electrodes, C is the number of non-canonical movement conditions, T is the number of time bins in the ‘go’ window, and R is the number of trial repetitions.

The cross-validated PCA procedure was applied by averaging neural activity within the analysis window (that is, over the T dimension), yielding a matrix of shape N × C × R. This matrix was then split randomly into two halves across repetitions to create independent datasets X1 and X2. Each dataset was then trial averaged, yielding matrices of shape N × C. X1 and X2 were then projected onto the PCs from X1. Shared variance was computed via dot products of matching PC projections across splits. This process was repeated 100 times with different trial splits, and the average cumulative shared variance with 95% CIs was plotted (Extended Data Fig. 3a).

Extended Data Fig. 3b shows the ratio of non-canonical-to-canonical movement-related neural activity, computed using cross-validated neural distances (cvDistance; see the ‘Tuning strength heat maps’ section). For each array, we calculated pairwise neural distances within the non-canonical and canonical movement sets. The modulation ratio was defined as the average distance within the non-canonical set divided by the average distance within the canonical set. A similar approach was used for Extended Data Fig. 2b, except that all single-sorted units were pooled across dorsal arrays and separately across ventral arrays (only including units with a 2–200-Hz mean firing rate). This pooling was necessary due to the limited number of sorted units per array, which made within-array neural distance estimates unreliable. To match trial counts across participants in the pooled sorted-unit analysis, we used, for each array set, the minimum number of trial repetitions available across participants (12 trial repetitions per movement condition for the dorsal set, 7 trial repetitions per movement condition for the ventral set); for participants with more trials than this minimum, trials were randomly selected without replacement. Cross-validated pairwise neural distances were then computed for the dorsal and ventral pools, independently, and modulation ratios were calculated. Jackknife resampling was used to estimate 95% CIs.

Data exclusion

Microelectrode array recordings may not be representative of the neural population tuning in a cortical area if they fail to record a sufficient amount of tuned spiking activity. To test whether an array contained any movement-related information, we used a cross-validated (tenfold) Gaussian naive Bayes classifier (following the methods described in ref. 37) applied to a 100-ms sliding window of neural activity. We found that four microelectrode arrays (C2-d1, T11-d2, T17-m1 and T17-m2) failed to demonstrate consistently above-chance classification performance for any time epoch, when classifying from among all 46 movements (Extended Data Fig. 10), and were thus excluded from this study and all other main results.

Also of note is that the neural recordings of participant C1 exhibited large noise when they attempted movements of the head moving up and down, which may have mechanically disturbed the pedestals fixed to the head or the cables. These two head conditions were removed for participant C1 for all population-level tuning analyses.

HCP parcellation, resting-state networks and motor task activations

For all HCP results presented in this study, we analysed the WU-Minn HCP 1200 Subjects Group Average Data Release32,65 and the original 210 subject release32. This data includes group-averaged structural data, functional connectivity data and task fMRI data. The Connectome Workbench (v1.5.0; www.humanconnectome.org), an open-source visualization and discovery tool, was used to explore and analyse the data generated by the HCP.

Group cortical parcellation

Extended Data Fig. 5 (first panel) displays the group-averaged pial surface of the left hemisphere with 180 areas delineated and identified by the HCP’s multi-modal parcellation32 (HCP_MMP1.0). To generate this panel, we loaded the HCP_S1200_GroupAvg_v1.scene file into the Connectome Workbench and selected the ‘Cortical Parcellations’ scene (scene 5). Next, we viewed the pial surface of the left hemisphere (S1200.L.pial_MSMAll.32k_fs_LR.surf.gii) and overlaid the cortical parcellations (Q1-Q6_RelatedValidation210.CorticalAreas_dil_Final_Final_Areas_Group_Colors.32k_fs_LR.dlabel.nii) with modifications to the colour scheme. We enabled the borders and selected only the following regions to display: L_4_ROI, L_6mp_ROI, L_6d_ROI, L_6v_ROI, L_PEF_ROI, L_FEF_ROI and L_55b_ROI.

Resting-state fMRI networks for language and arm movement

The resting-state networks shown in Extended Data Fig. 4 were generated using the same scene file in the section ‘Group cortical parcellation’ and overlaying resting-state networks 25 and 28 (Q1-Q6_RelatedParcellation210.individual_RSNs_d40_WR_norm_MSMAll_2_d41_WRN_DeDrift.32k_fs_LR.dscalar.nii; file can be found at https://balsa.wustl.edu/study/RVVG). Networks 25 and 28 were highlighted as a language network and upper limb network, respectively, by Glasser et al.32.

Task fMRI activations

Extended Data Fig. 6b shows task fMRI activations for the group-averaged HCP S1200 dataset. The same main scene file from section ‘Group cortical parcellation’ was used with the ‘S1200 task fMRI Cohen’s D effect-size maps’ scene loaded (scene 3). The task activation file HCP_S1200_997_tfMRI_ALLTASKS_level2_cohensd_hp200_s2_MSMAll.dscalar.nii was used for overlay visualization. Only task fMRI activation maps for right-hand (map tfMRI_MOTOR_RH) and tongue (map tfMRI_MOTOR_T) movements were displayed. For visualization, the following display settings were applied to each task activation file: ‘source’ was set to ‘self’, ‘high’ to 1.308, ‘low’ to −3.057, ‘Abs Pct’ was toggled on, with ‘Pos Max’ set to 99.5 and ‘Pos Min’ to 2, and the colour palette set to ‘black-blue-positive’ and ‘black-red-positive’ for the right hand and tongue, respectively. The opacity was set to 0.55 for both maps.

Functional connectivity map for the somato-cognitive action network

Extended Data Fig. 6a shows a functional connectivity map for the group-averaged HCP S1200 dataset. The same main scene file from section ‘Group cortical parcellation’ was used with the ‘S1200 fcMRI, full correlation’ scene loaded (scene 4). The correlation file HCP_S1200_1003_rfMRI_MSMAll_groupPCA_d4500ROW_zcorr was used for overlay visualization only. The seed point was manually selected to best match the location reported by Gordon et al.15 to identify the somato-cognitive action network (SCAN).

To enhance visualization of the SCAN, the following display settings were adjusted in the settings panel of the correlation file: the ‘Source’ was set to ‘self’, ‘high’ threshold to 0.9, ‘low’ threshold to −1, ‘Abs Pct’ was toggled on, with ‘Pos Max’ set to 99.5 and ‘Pos Min’ to 2. The resulting SCAN visualization is shown in Extended Data Fig. 6b.

For Extended Data Fig. 6b, array locations were manually selected on the individual HCP pial surface of each participant (in MSMAll space), with locations estimated using surgical photos, and then projected onto the HCP group-averaged pial surface (also in MSMAll space) within the Connectome Workbench. As a result, the spatial ordering of arrays may differ from the anatomically aligned organization shown in Fig. 2b, which is based on the sulcal landmarks IFS and SFS.

For Extended Data Fig. 6c, we used the HCP-style multimodal surface registration for each participant (MSMAll) to project the HCP group-averaged SCAN network and hand–tongue task-fMRI activation maps onto the pial surface of that participant (all in MSMAll space).

Quantifying movement using optical flow

To collect video evidence demonstrating that neural correlations between limbs were not caused by inadvertent co-movement of uninstructed effectors, participant T12 performed an additional abridged version of the whole-body cued movement task, involving only movements of the four limbs (arms and legs), while video recordings captured her limb movements (trial day 1,064, blocks 5–16). T12 retains residual voluntary motion and demonstrates isolated movements of her limbs (Supplementary Video 1). To quantify limb motion, dense optical flow was computed using the Farneback method (calcOpticalFlowFarneback function from the cv2 package in Python with the following standard parameter settings: ‘flow’ to ‘none’, ‘pyr_scale’ to 0.5, ‘levels’ to 3, ‘winsize’ to 15, ‘iterations’ to 3, ‘poly_n’ to 5, ‘poly_sigma’ to 1.2, and ‘flags’ to 0) across consecutive frames of each video. For each frame, the optical flow vector within each region was calculated, and its magnitude was used to generate a time series of movement intensity for each limb. This analysis is summarized in Extended Data Fig. 8c.

Statistics

Supplementary Table 3 lists statistical details for each CI or hypothesis test reported in this work. In this study, uncertainty was quantified mainly with 95% CIs.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All data required to reproduce the findings in this study are publicly available on Dryad (https://doi.org/10.5061/dryad.mpg4f4rg5; data for T5, T11, T12, T15, T16 and T17) and DABI (https://dabi.loni.usc.edu/projects/K1KK0H4SRS11; data for C1 and C2). The datasets contain neural activity recorded during instructed delay tasks designed to investigate the neural representation of attempted movements spanning the entire body. As part of this study, we also analysed publicly available neuroimaging data from the HCP (1200 Subject Data Release; https://www.humanconnectome.org/study/hcp-young-adult).

Code availability

The codes used to reproduce the findings in this study (https://github.com/d-r-deo/pcg-mosaic), for cross-validated estimation of neural correlation and distance (https://github.com/fwillett/cvVectorStats) and for implementing the recurrent neural network decoder (https://github.com/cffan/neural_seq_decoder) are publicly available on GitHub.

References

Cheney, P. D. & Fetz, E. E. Comparable patterns of muscle facilitation evoked by individual corticomotoneuronal (CM) cells and by single intracortical microstimuli in primates: evidence for functional groups of CM cells. J. Neurophysiol. 53, 786–804 (1985).

Article

CAS

PubMed

Google Scholar

Donoghue, J. P., Leibovic, S. & Sanes, J. N. Organization of the forelimb area in squirrel monkey motor cortex: representation of digit, wrist, and elbow muscles. Exp. Brain Res. 89, 1–19 (1992).

Article

CAS

PubMed

Google Scholar

Graziano, M. S. A., Taylor, C. S. R. & Moore, T. Complex movements evoked by microstimulation of precentral cortex. Neuron 34, 841–851 (2002).

Article

CAS

PubMed

Google Scholar

Stepniewska, I., Gharbawie, O. A., Burish, M. J. & Kaas, J. H. Effects of muscimol inactivations of functional domains in motor, premotor, and posterior parietal cortex on complex movements evoked by electrical stimulation. J. Neurophysiol. 111, 1100–1119 (2014).

Article

CAS

PubMed

Google Scholar

Preuss, T. M., Stepniewska, I. & Kaas, J. H. Movement representation in the dorsal and ventral premotor areas of owl monkeys: a microstimulation study. J. Comp. Neurol. 371, 649–676 (1996).

Article

CAS

PubMed

Google Scholar

Maranesi, M. et al. Anatomo-functional organization of the ventral primary motor and premotor cortex in the macaque monkey. Eur. J. Neurosci. 36, 3376–3387 (2012).

Article

PubMed

Google Scholar

Kurata, K. Distribution of neurons with set- and movement-related activity before hand and foot movements in the premotor cortex of rhesus monkeys. Exp. Brain Res. 77, 245–256 (1989).

Article

CAS

PubMed

Google Scholar

Cheney, P. D., Fetz, E. E. & Palmer, S. S. Patterns of facilitation and suppression of antagonist forelimb muscles from motor cortex sites in the awake monkey. J. Neurophysiol. 53, 805–820 (1985).

Article

CAS

PubMed

Google Scholar

Schieber, M. H. & Poliakov, A. V. Partial inactivation of the primary motor cortex hand area: effects on individuated finger movements. J. Neurosci. 18, 9038–9054 (1998).

Article

CAS

PubMed

PubMed Central

Google Scholar

Friel, K. M. & Nudo, R. J. Recovery of motor function after focal cortical injury in primates: compensatory movement patterns used during rehabilitative training. Somatosens. Mot. Res. 15, 173–189 (1998).

Article

CAS

PubMed

Google Scholar

Miller, K. J. et al. Spectral changes in cortical surface potentials during motor movement. J. Neurosci. 27, 2424–2432 (2007).

Article

CAS

PubMed

PubMed Central

Google Scholar

Crone, N. E. et al. Functional mapping of human sensorimotor cortex with electrocorticographic spectral analysis. I. Alpha and beta event-related desynchronization. Brain 121, 2271–2299 (1998).

Article

PubMed

Google Scholar

Natraj, N. et al. Sampling representational plasticity of simple imagined movements across days enables long-term neuroprosthetic control. Cell 188, 1208–1225.e32 (2025).

Article

CAS

PubMed

PubMed Central

Google Scholar

Schellekens, W., Bakker, C., Ramsey, N. F. & Petridou, N. Moving in on human motor cortex. Characterizing the relationship between body parts with non-rigid population response fields. PLoS Comput. Biol. 18, e1009955 (2022).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Gordon, E. M. et al. A somato-cognitive action network alternates with effector regions in motor cortex. Nature 617, 351–359 (2023).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Zeharia, N., Hertz, U., Flash, T. & Amedi, A. Negative blood oxygenation level dependent homunculus and somatotopic information in primary motor cortex and supplementary motor area. Proc. Natl Acad. Sci. USA 109, 18565–18570 (2012).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Penfield, W. & Boldrey, E. Somatic motor and sensory representation in the cerebral cortex of man as studied by electrical stimulation. Brain 60, 389–443 (1937).

Article

Google Scholar

Penfield, W. & Rasmussen, T. The cerebral cortex of man: a clinical study of localization of function. JAMA https://doi.org/10.1001/jama.1950.02920160086033 (1950).

Branco, D. M. et al. Functional variability of the human cortical motor map: electrical stimulation findings in perirolandic epilepsy surgery. J. Clin. Neurophysiol. 20, 17–25 (2003).

Article

PubMed

Google Scholar

Roux, F.-E., Niare, M., Charni, S., Giussani, C. & Durand, J.-B. Functional architecture of the motor homunculus detected by electrostimulation. J. Physiol. 598, 5487–5504 (2020).

Article

CAS

PubMed

Google Scholar

Graziano, M. S. A. Ethological action maps: a paradigm shift for the motor cortex. Trends Cogn. Sci. 20, 121–132 (2016).

Article

PubMed

Google Scholar

Schieber, M. H. Constraints on somatotopic organization in the primary motor cortex. J. Neurophysiol. 86, 2125–2143 (2001).

Article

CAS

PubMed

Google Scholar

Park, M. C., Belhaj-Saïf, A., Gordon, M. & Cheney, P. D. Consistent features in the forelimb representation of primary motor cortex in rhesus macaques. J. Neurosci. 21, 2784–2792 (2001).

Article

CAS

PubMed

PubMed Central

Google Scholar

Godschalk, M., Mitz, A. R., van Duin, B. & van der Burg, H. Somatotopy of monkey premotor cortex examined with microstimulation. Neurosci. Res. 23, 269–279 (1995).

Article

CAS

PubMed

Google Scholar

Stepniewska, I., Fang, P.-C. & Kaas, J. H. Microstimulation reveals specialized subregions for different complex movements in posterior parietal cortex of prosimian galagos. Proc. Natl Acad. Sci. USA 102, 4878–4883 (2005).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Meier, J. D., Aflalo, T. N., Kastner, S. & Graziano, M. S. A. Complex organization of human primary motor cortex: a high-resolution fMRI study. J. Neurophysiol. 100, 1800–1812 (2008).

Article

PubMed

PubMed Central

Google Scholar

Peters, N. et al. Ischemic stroke of the cortical ‘hand knob’ area: stroke mechanisms and prognosis. J. Neurol. 256, 1146–1151 (2009).

Article

CAS

PubMed

Google Scholar

Nah, H.-U., Park, H. K. & Kang, D.-W. Isolated shoulder weakness due to a small cortical infarction. J. Clin. Neurol. 2, 209–211 (2006).

Article

PubMed

PubMed Central

Google Scholar

Desmurget, M. & Sirigu, A. Revealing humans’ sensorimotor functions with electrical cortical stimulation. Philos. Trans. R. Soc. Lond. B Biol. Sci. 370, 20140207 (2015).

Article

PubMed

PubMed Central

Google Scholar

Miller, K. J. et al. Human motor cortical activity is selectively phase-entrained on underlying rhythms. PLoS Comput. Biol. 8, e1002655 (2012).

Article

CAS

PubMed

PubMed Central

Google Scholar

Jensen, M. A. et al. A motor association area in the depths of the central sulcus. Nat. Neurosci. https://doi.org/10.1038/s41593-023-01346-z (2023).

Glasser, M. F. et al. A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178 (2016).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Rademacher, J. et al. Variability and asymmetry in the human precentral motor system. A cytoarchitectonic and myeloarchitectonic brain mapping study. Brain 124, 2232–2258 (2001).

Article

CAS

PubMed

Google Scholar

White, L. E. et al. Structure of the human sensorimotor system. I: morphology and cytoarchitecture of the central sulcus. Cereb. Cortex 7, 18–30 (1997).

Article

CAS

PubMed

Google Scholar

Raos, V., Franchi, G., Gallese, V. & Fogassi, L. Somatotopic organization of the lateral part of area F2 (dorsal premotor cortex) of the macaque monkey. J. Neurophysiol. 89, 1503–1518 (2003).

Article

PubMed

Google Scholar

Yousry, T. A. et al. Localization of the motor hand area to a knob on the precentral gyrus. A new landmark. Brain 120, 141–157 (1997).

Article

PubMed

Google Scholar

Willett, F. R. et al. Hand knob area of premotor cortex represents the whole body in a compositional way. Cell https://doi.org/10.1016/j.cell.2020.02.043 (2020).

Article

PubMed

PubMed Central

Google Scholar

Stavisky, S. D. et al. Neural ensemble dynamics in dorsal motor cortex during speech in people with paralysis. eLife 8, e46015 (2019).

Article

PubMed

PubMed Central

Google Scholar

Silva, A. B. et al. A neurosurgical functional dissection of the middle precentral gyrus during speech production. J. Neurosci. 42, 8416–8426 (2022).

Article

CAS

PubMed

PubMed Central

Google Scholar

Kuehn, E. et al. Body topography parcellates human sensory and motor cortex. Cereb. Cortex 27, 3790–3805 (2017).

Article

PubMed

PubMed Central

Google Scholar

Deo, D. R. et al. Brain control of bimanual movement enabled by recurrent neural networks. Sci. Rep. 14, 1598 (2024).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Criscimagna-Hemminger, S. E., Donchin, O., Gazzaniga, M. S. & Shadmehr, R. Learned dynamics of reaching movements generalize from dominant to nondominant arm. J. Neurophysiol. 89, 168–176 (2003).

Article

PubMed

Google Scholar

Kelso, J. A. S. & Zanone, P. G. Coordination dynamics of learning and transfer across different effector systems. J. Exp. Psychol. Hum. Percept. Perform. 28, 776–797 (2002).

Article

CAS

PubMed

Google Scholar

Gallese, V., Fadiga, L., Fogassi, L. & Rizzolatti, G. Action recognition in the premotor cortex. Brain 119, 593–609 (1996).

Article

ADS

PubMed

Google Scholar

Cisek, P., Crammond, D. J. & Kalaska, J. F. Neural activity in primary motor and dorsal premotor cortex in reaching tasks with the contralateral versus ipsilateral arm. J. Neurophysiol. 89, 922–942 (2003).

Article

PubMed

Google Scholar

Gentilucci, M. et al. Functional organization of inferior area 6 in the macaque monkey. I. Somatotopy and the control of proximal movements. Exp. Brain Res. 71, 475–490 (1988).

Article

CAS

PubMed

Google Scholar

Graziano, M. The organization of behavioral repertoire in motor cortex. Annu. Rev. Neurosci. 29, 105–134 (2006).

Article

CAS

PubMed

Google Scholar

Wandelt, S. K. et al. Decoding grasp and speech signals from the cortical grasp circuit in a tetraplegic human. Neuron 110, 1777–1787.e3 (2022).

Article

CAS

PubMed

PubMed Central

Google Scholar

Finkelsteyn, A. M. et al. Ischemic stroke of the ‘hand knob area’: a case series and literature review. J. Clin. Neurosci. 65, 100–105 (2019).

Article

PubMed

Google Scholar

Celebisoy, M., Ozdemirkiran, T., Tokucoglu, F., Kaplangi, D. N. & Arici, S. Isolated hand palsy due to cortical infarction: localization of the motor hand area. Neurologist 13, 376–379 (2007).

Article

PubMed

Google Scholar

Makin, T. R. et al. Phantom pain is associated with preserved structure and function in the former hand area. Nat. Commun. 4, 1570 (2013).

Article

ADS

PubMed

PubMed Central

Google Scholar

Schone, H. R. et al. Stable cortical body maps before and after arm amputation. Nat. Neurosci. 28, 2015–2021 (2025).

Article

CAS

PubMed

Google Scholar

Howell, P. et al. Revisiting face-to-hand area remapping in the human primary somatosensory cortex after a cervical spinal cord injury. Preprint at bioRxiv https://doi.org/10.1101/2025.03.11.642631 (2025).

Northall, A. et al. Multimodal layer modelling reveals in vivo pathology in amyotrophic lateral sclerosis. Brain 147, 1087–1099 (2023).

Article

Google Scholar

Al-Chalabi, A. et al. Amyotrophic lateral sclerosis: moving towards a new classification system. Lancet Neurol. 15, 1182–1194 (2016).

Article

PubMed

Google Scholar

Willett, F. R. et al. A high-performance speech neuroprosthesis. Nature 620, 1031–1036 (2023).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Flesher, S. N. et al. A brain-computer interface that evokes tactile sensations improves robotic arm control. Science 372, 831–836 (2021).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Lorach, H. et al. Walking naturally after spinal cord injury using a brain-spine interface. Nature 618, 126–133 (2023).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Singer-Clark, T. et al. Speech motor cortex enables BCI cursor control and click. J. Neural Eng. 22, 036015 (2025).

Article

ADS

Google Scholar

Kunigk, N. G. et al. Motor somatotopy impacts imagery strategy success in human intracortical brain-computer interfaces. J. Neural Eng. https://doi.org/10.1088/1741-2552/adb995 (2025).

Card, N. S. et al. An accurate and rapidly calibrating speech neuroprosthesis. N. Engl. J. Med. 391, 609–618 (2024).

Article

PubMed

PubMed Central

Google Scholar

Greenspon, C. M. et al. Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex. Nat. Biomed. Eng. 9, 935–951 (2025).

Young, D. et al. Signal processing methods for reducing artifacts in microelectrode brain recordings caused by functional electrical stimulation. J. Neural Eng. 15, 026014 (2018).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M. & Shenoy, K. V. High-performance brain-to-text communication via handwriting. Nature 593, 249–254 (2021).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Van Essen, D. C. et al. The WU-Minn Human Connectome Project: an overview. Neuroimage 80, 62–79 (2013).

Article

PubMed

PubMed Central

Google Scholar

Kobak, D. et al. Demixed principal component analysis of neural population data. eLife 5, e10989 (2016).

Article

PubMed

PubMed Central

Google Scholar

Stringer, C., Pachitariu, M., Steinmetz, N., Carandini, M. & Harris, K. D. High-dimensional geometry of population responses in visual cortex. Nature 571, 361–365 (2019).

Article

ADS

CAS

PubMed

PubMed Central

Google Scholar

Download references

Acknowledgements

We thank participants T5, T11, T12, T15, T16, T17, C1 and C2 and their caregivers for their generously volunteered time and dedicated contributions to this research; M. Young for contributions to the neuroethical considerations of this research; B. Davis, K. Tsou and S. Kosasih for administrative support; A. Paulk and P. Hadar for help with T17 pre-surgical imaging; and M. Boninger for serving as sponsor-investigator on the IDE for the University of Chicago. This project made use of Connectome DB and Connectome Workbench, developed under the auspices of the Human Connectome Project at the Washington Universit