Neural synchrony links sensorimotor cortices in a network for facial motor control Authors: Yuriria Vázquez, Geena R. Ianni, Elie Rassi, Adam G. Rouse, Marc H. Schieber, Faraz Yazdani, Yifat Prut, Winrich A. Freiwald Published: 2025 ABSTRACT SignificancePrimates communicate through facial expressions. How the brain generates facial expressions remains poorly understood. To uncover how facial motor-related cortical brain regions interact to produce facial gestures, we combined fMRI-targeted electrophysiology and intracortical microstimulation while monkeys produced qualitatively different facial movements. Our two-pronged experimental approach revealed that facial motor-related cortical areas form an interconnected network characterized by synchronized neural activity demonstrating dynamic expression-selective activity states that are coordinated across the network nodes. Thus, the multiple facial motor-related cortical areas sending axons directly into the facial nucleus operate as a single network in which the overall complex, behavior-specific interareal interactions dictate the relevant motor output. In primate societies, individuals thrive with their ability to interpret and produce social signals (1). Primates communicate socially through vocalizations, speech, and facial expressions, all of which require precise control of facial musculature (2, 3). Facial expressions convey information about internal states or intentions, and thereby function as social signals among conspecifics. Despite the significance of facial expressions in social interactions, little is understood about the neural mechanisms underlying these social motor behaviors when compared to other motor actions such as reaching or grasping (4–7). The primate brain harbors several cortical regions dedicated to facial motor control within the frontal cortex—specifically, the primary motor (M1) and the ventrolateral premotor (PMv) cortices on the lateral aspect, as well as the supplementary (M2) and the cingulate (rostral M3 and caudal M4) motor cortices on the medial aspect (8–10). Each of these cortical areas project directly to the facial motor nucleus in the brainstem (11), suggesting that motor cortical areas exert direct control over facial motoneurons in a seemingly parallel, independent control strategy, where each cortical area directly controls a subset of facial muscles. Studies in human patients have shed light on the functional circuitry of these areas during facial expressions. Two facial areas (M1 and PMv) reside laterally, in the territory of the middle cerebral artery, and three areas (SMA, rostral, and caudal cingulate cortex) reside medially, in the territory of the anterior cerebral artery (8). Patients with a middle cerebral artery stroke display asymmetry in voluntary or goal-directed facial movements (performed in response to verbal commands), yet their ability to perform spontaneous emotional gestures like smiling remains intact (12). In contrast, patients with an anterior cerebral artery stroke (affecting medial motor areas) can execute normal voluntary facial movements, but struggle with emotional facial gestures (12). These findings led to the neuropsychological schema that facial movements are controlled by two distinct pathways: lateral and medial motor pathways (Fig. 1A) (13, 14), each controlling a qualitatively different type of facial movement. This is further supported by the unique input and output connectivity patterns of the medial and lateral motor areas of the face. On the other hand, anatomical studies in monkeys have identified corticocortical connectivity between medial and lateral motor areas of the face (11, 15–19), including the facial representation in the somatosensory cortex (8, 18). Brain-wide imaging studies in nonhuman primates using network connectivity analysis revealed a distributed face motor network whose activation during a specific communicative facial expression extends beyond the medial cingulate cortex to include lateral motor areas (20). Cortical facial motor representations in the primate brain and facial motor behaviors. Medial (green) and lateral (pink) motor systems are hypothesized to support emotional and voluntary facial movements, respectively. Two possible neural mechanisms for facial movements are proposed. (A) Null hypothesis: medial and lateral motor areas operate independently, with medial regions specialized in emotional movements- facial expressions, and lateral areas for voluntary movements such as chewing. (B) Alternative hypothesis: medial and lateral motor areas functionally interact to support both emotional and voluntary movements. (C) Structural MRI (T1, coronal plane) showing electrode array placement in one monkey. Floating microarrays targeted primary somatosensory (S1, purple), motor (M1, light and dark red) cortex, ventrolateral premotor (PMv, blue and yellow), and cingulate motor cortex (M3, green). Array locations were based on fMRI conjunction maps of facial expressions and mouth movement (Materials and Methods). (D) Examples of naturalistic facial movements: threat (T, red), a nonaffiliative emotional display with sustained open mouth; lip-smack (LS, yellow), a rhythmic affiliative emotional movement of jaw and lips; and chewing (C, blue), an ingestive voluntary movement. Corresponding cartoons below each photo are used as references through the text. Neural activity was recorded simultaneously across all areas during these behaviors, with emotional movements elicited through videos, an avatar, and primate interactions (Materials and Methods). Thus, there are two opposing views on the organization of cortical face-motor control: the two parallel pathway hypothesis and the interconnected network hypothesis. The first predicts that there are no functional interactions between medial and lateral facial motor areas. The second hypothesis predicts that such interactions should exist, as medial and lateral cortical areas jointly determine facial movements (Fig. 1B). To test these predictions, we implanted recording arrays in fMRI-identified cortical areas of the facial motor and somatosensory areas of nonhuman primates. We first determine the capacity of an area to exert an influence on activity in another area by delivering intracortical microstimulation (ICMS) to one brain area while simultaneously recording the evoked local field potentials (LFPs) in the other areas, and estimating the average stimulus triggered response. We then determined the functional interactions taking place during naturalistic behavior, by measuring synchronous patterns in the local field potential (LFPs) among the multiple facial motor areas and the primary somatosensory cortex (S1) to assess functional connectivity. Our findings reveal that cortical motor areas of the face function as an interconnected network where medial and lateral areas can exert significant influence on each other, as shown through our ICMS approach. During emotional facial expressions, we observed alpha and beta synchronization between medial motor area M3, lateral areas (M1 and PMv), and the face representation in the primary somatosensory cortex, suggesting the presence of dynamic, expression-selective activity states coordinated across network nodes. In contrast, the coupling between medial and lateral motor areas during voluntary chewing occurred at lower frequencies, highlighting the frequency-specific nature of cortical coordination patterns for different facial movements. The interareal dynamics of this sensorimotor network displayed different patterns of functional connectivity that correlated with specific facial gestures. This supports our hypothesis that facial motor areas and S1 form an interconnected network with flexible interactions based on synchronized neural activity, generating distinct neural patterns for different types of facial motor behavior. Results We recorded LFPs from multiple channels (n = 32 channels per area) located in the sensorimotor areas of the face, including the primary somatosensory cortex (S1), primary motor cortex (M1), ventrolateral premotor cortex (PMv), and cingulate motor cortex (M3), as head-fixed monkeys produced ethologically relevant facial movements in response to various stimuli (Materials and Methods and Fig. 1 C and D and SI Appendix, Fig. S1). Estimating Interareal Connectivity in the Facial Motor System With Intracortical Microstimulation. To identify connectivity between facial motor areas, we employed intracortical microstimulation (ICMS) while the animal was sitting quietly, making no facial movements. For this purpose, we injected single pulse stimuli (250 µA, a train of biphasic pulses, each pulse duration=0.4 ms at a frequency of 3 Hz (333 ms interpulse interval) delivered for ~2 min) through dedicated low-impedance stimulation electrodes embedded in the recording arrays and estimated the average stimulus triggered response recorded in neighboring arrays (Fig. 2A). Mapping Facial Sensory and Motor Connectivity through Intracortical Microstimulation. (A) Illustration depicting a monkey’s brain with electrode arrays in the recorded areas. The intracortical microstimulation (ICMS) current protocol was applied to the sender area, in this case the M3 cingulate motor area. The protocol uses a train of biphasic pulses (250 µA, each pulse duration=0.4 ms at a frequency of 3 Hz (333 ms interpulse interval) delivered for ~2 min). The orange trace represents the PMv’s mean evoked local field potential due to current propagation, highlighting the first peak and trough after the stimulation onset. (B) Averaged evoked local field potentials (LFPs) across cortical areas in response to microstimulation of M3 (B1), M1(B2), PMv(B3), and S1 (B4) cortices. The stimulation onset is marked at zero on the x-axis. A stimulation artifact with rapid onset and large amplitude is visible across all averages during the first few milliseconds following stimulation. To exclude this artifact from the analysis, peaks and troughs were searched for starting 6 ms after stimulation onset in all receiving areas. (C) Confusion matrix depicting connectivity indexes measured by the absolute difference between the first peak or trough, and the second trough or peak, respectively (Δ|Peak-Trough|). Current sender areas are shown on the X-axis and receiver areas on the Y-axis. Gray diagonal elements represent untested within-area connectivity, while off-diagonal elements show the measured connectivity strength. Net signal propagation was assessed by analyzing the eLFP measured in the receiving areas. A key measurement used here was the response magnitude, calculated as the difference between either the first peak and trough or the first trough and peak following the stimulation pulse. To avoid measuring stimulation artifacts, we searched for peaks and troughs starting 6 ms after stimulation onset in all areas, when artifactual deflections had ended. Fig. 2B shows the average area-specific eLFP obtained in response to microstimulation applied in M3 (B1), M1 (B2), PMv (B3), and S1 (B4) obtained across several recording sessions (n ≥ 4,200 stimulation pulses). We found evidence for signal propagating from the medial (M3) to the lateral (M1 and PMv) motor areas, and the primary somatosensory cortex (S1). The eLFP responses obtained in S1 (purple line) were generally smaller in amplitude and delayed by ~10 ms relative to the ones in the lateral motor areas (red and orange lines) suggesting a slower signal propagation to this region (Fig. 2 B1). Altogether, we found that ICMS produced significant eLFP responses in all combinations of areas, though with varying magnitudes and waveform patterns. We quantified the evoked connectivity between brain regions using a connectivity matrix (Fig. 2C). This matrix revealed several key patterns, notably that M1 and S1 showed reciprocal connectivity (21). M3 received inputs from both lateral motor areas (M1 and PMv), with M1 eliciting a larger response than PMv. The response to primary somatosensory (S1) stimulation was comparatively smaller. When stimulation was applied to M3, both lateral motor areas responded (M1 and PMv), aligning with previously described anatomically connectivity patterns (8). However, an asymmetry was observed: while M3 stimulation produced relatively modest responses in all other areas, stimulation of M1 evoked strong responses in M3. This asymmetry could potentially be explained by orthodromic activation of the source area during local stimulation, supporting M3 receiving inputs from M1, as previously reported (19). This set of results validates the use of eLFP for identifying anatomically relevant functional connectivity across different cortical areas and suggests that the observed medial-to-lateral connectivity reveals physiologically relevant pathways between these network components. LFP Synchrony Links Sensorimotor Cortices in a Network for Facial Motor Control. We assessed functional connectivity by calculating the synchronization of LFP signals during different facial gestures in monkeys. This approach allowed us to examine how brain areas interact and whether these interactions change depending on the specific behavioral context. LFP signals can provide insights into the neural processes underpinning various cognitive functions and actions (22–24). We used long stretches of LFP recordings (7 to 10 min) and identified facial movement epochs using manual scoring, and an automatic algorithm based on optic flow measurements (Materials and Methods). Functional connectivity was assessed using pairwise phase consistency (PPC) between the LFPs of the recorded areas, quantifying the distribution of phase differences (25). The advantage of using PPC versus other connectivity measurements is that it remains reliable even with few trials, and it is less prone to errors that occur when the consistency of phase differences correlates with the strength of the signal (25). This method assumes that neural synchronization between two areas will show a distribution of phase differences centered around a mean value. The analysis first examines differences in functional connectivity patterns between periods of facial stillness (rest) and movement. We found clear differences in the functional connectivity between the moving and the still face. Overall, during emotional behaviors, beta-range (13 to 40 Hz) connectivity between lateral and medial motor areas (M1 and M3) increased during movement. The M3-PMv connectivity in the beta range, however, showed behavior-specific patterns: It increased during lipsmacks (LS) but decreased during threats (T). For voluntary movements (chews, C), the medial-to-lateral connectivity shifted toward lower frequencies, showing increased theta (5 to 7 Hz) and alpha (8 to 12 Hz) connectivity while beta connectivity diminished during movement execution. Regarding lateral motor areas connectivity, both emotional and voluntary movements showed increased M1-PMv beta band connectivity during movement compared to rest, with voluntary movements additionally exhibiting enhanced theta and alpha connectivity (for details, see SI Appendix, Table S1, cluster-based nonparametric statistics with the dependent-samples t test, P ≤ 0.05). While the above analysis revealed general patterns of functional connectivity that point to movement-related modulation of interactions between these areas, our main analysis focused on understanding the movement-specific functional interactions between the sensorimotor face areas, as well as characterizing the directional flow of information within the network during different types of facial movements, specifically targeting the movement execution period, which enhances our ability to accurately define distinct actions. We first calculated functional interactions between medial and lateral motor areas during threats (T—red) and lipsmacks (LS—yellow), representing emotionally based gestures with opposite valence (Fig. 3 A–C). In general, synchrony appears to be frequency-dependent and consistently stronger during threats compared to lipsmacks. Statistical analyses revealed that threat-related interareal synchrony was significantly stronger than lipsmack-related synchrony, mostly in the low frequency range (thick shades in Fig. 3 A–C; cluster-based nonparametric statistics with the dependent-samples t test, P ≤ 0.05; see SI Appendix, Table S2) but not in the high-frequency range (i.e., beta range). Interestingly, the functional interactions between the medial (M3) and lateral (M1 or PMv) motor systems during emotional facial movements (depicted in Fig. 3 A and B) were comparable in properties but weaker in magnitude compared to the interactions within the lateral subnetwork (M1 and PMv in Fig. 3C). In both cases, the synchrony during threats was stronger than during lipsmacks in the low-frequency range (<20 Hz), but not in the beta range. Functional interactions between facial motor and sensorimotor cortices during the production of emotional facial movements. (A–F) Pairwise phase consistency (PPC) between M3 and M1 (medial and lateral motor areas), M3 and PMv (medial and lateral motor areas), M1 and PMv (lateral motor areas), M3 and S1 (medial motor and primary somatosensory area), PMv and S1 (lateral and primary somatosensory area), M1 and S1 (lateral motor and primary somatosensory area) during the production of threats (T: red solid lines, n = 1,200) and lipsmacks (LS: yellow solid lines, n = 1,200). In Panels A–F thicker, semitransparent lines atop the solid lines denote regions of statistical differences between both facial expressions (P ≤ 0.05). (G–L) Granger causality for pairs of areas: M3-M1, M3-PMv, M1-PMv, M3-S1, PMv-S1, M1-S1 respectively. Solid lines depict a particular directionality of Granger causality (e.g., M3 to M1), while dashed lines depict the opposite direction (e.g. M1 to M3). Thicker semitransparent lines indicate Granger causality values statistically different between the facial movements. Beyond motor system interactions, we investigated how primary somatosensory cortex (S1) functionally connects with motor areas during different emotional gestures, providing insight into sensorimotor integration. We observed beta synchrony between S1 and both medial (M3) and lateral motor areas, with a clear broad peak in the beta frequency range for functional connectivity between S1 and PMv (Fig. 3 D–E) during threats and lipsmacks. Similar to patterns within the motor network, coupling between S1 and motor areas was consistently stronger during threats than during lipsmacks (Fig. 3 D–F). This gesture-specific synchrony pattern suggests functional differences beyond the distinct kinematics characterizing these two gestures. Next, we compared the neural activity during emotional facial movements with that observed during voluntary facial movements (specifically, chewing movements). We found increased synchronization both between the lateral motor cortices (M1 and PMv), and between the medial and lateral motor cortices (M3 and M1, M3, and PMv) during chewing compared to rest (SI Appendix, Table S1). Further analysis revealed consistent coupling within the lateral system (M1 and PMv) during both voluntary and emotional movements. While no significant differences were detected at high frequencies (>20 Hz) both threats and chews showed significant higher synchronization compared to lipsmacks at low frequencies (SI Appendix, Fig. S2 C and F and Table S2). Moreover, during voluntary chewing, the M3 area showed synchronization with M1(S2A, S2D) and PMv (S2B, S2E) mainly at low frequencies (4 to 8 Hz). In contrast, beta range synchronization was not prominent. In general, emotional movements elicited stronger alpha and beta coupling across the medial and lateral systems, while voluntary chewing movements relied more on low-frequency coupling, illustrating frequency-specific functional interactions during movement execution (SI Appendix, Fig. S2 A–F). In summary, the lateral motor system (M1 and PMv) maintains consistent beta frequency coupling during all movement types, with evident movement-specific differences occurring in the low frequency range (theta and alpha range), where both threats and chews showed stronger coupling than lipsmacks. The medial-to-lateral connectivity (M3 to M1 and PMv) shows broader frequency range variations that depend on the ongoing movement type: emotional movements engage primarily alpha and beta range, while voluntary chews show reduced synchronization in these bands and predominantly operate at lower frequencies (<10 Hz, Table 1). Table 1.Peak coupling frequency among motor circuitsMovement TypeGestureLateral-Lateral Circuit (M1-PMv)Medial-Lateral Circuit (M3-M1)Medial-Lateral Circuit (M3-PMv)EmotionalLip-Smackingθ = 6.5, 7.5 Hz,θ = 6 Hzα = 9 Hzα = 8.5 Hzβ = 14, 25 Hzβ = 17.5, 23.5 Hzβ = 17.5, 19.5, 27.5, 30, 31.5 Hzγ = 46 Hzγ = 45 HzThreatθ = 7 Hzθ = 7.5 Hzθ = 4 to 7 Hzα = 9, 13.5 Hz,α = 13.5 Hzα = 9,8 Hzβ = 15.5, 24 Hzβ = 21, 33.5 Hzβ = 12.5, 15, 17, 20 Hzγ = 41 Hzγ = 40 HzVoluntaryChewθ = 4 to 7 Hzθ = 5 Hz,θ = 4 to 8 Hzα = 10, 11 Hzα = 9.5 Hzα = 11.5 Hzβ = 15.5, 26, 23.5 Hzβ = 15 Hzβ = 25, 28 Hzγ = 44 Hzγ = 43.5 Hz Peak coupling frequency among motor circuits Directional Information Within the Face Sensorimotor Network. Our findings have revealed a dynamic sensorimotor network where frequency coupling varies with the type of facial movement. However, the synchrony measurement we used (PPC) cannot identify directionality of interactions. For this purpose, we employed Granger Causality analysis, which measures the extent to which activity in one area (e.g., variable X) can significantly predict subsequent activity in another area (e.g., variable Y) thereby hinting at causal relations (or “information flow”) between the two regions (26, 27). We used Granger causality to reveal differences in interareal information flow during different facial movements, with a particular emphasis on information flow between medial and lateral areas of the facial motor system (M3 and M1 and PMv). Using this approach, we found bidirectional information flow between medial and lateral motor areas, with the magnitude of the flow varying according to the facial expression. When comparing threats to lipsmacks, information flow from medial-to-lateral motor areas was stronger during threats, while flow in the opposite direction remained similar between both behaviors (Fig. 3 G and H). Interestingly, the interactions with S1 were substantially more asymmetric, showing a consistently stronger flow of information from motor regions to S1 than in the reverse direction (Fig. 3 J–L). During voluntary chewing movements, we observed slightly lower information flow from M3 to the lateral motor areas (M1 and PMv) when compared with threat-related emotional movements (SI Appendix, Fig. S2 J and K). Comparison between chewing and lipsmack movements revealed target-specific patterns. While M3-to-PMv information flow was similar between both movement types (SI Appendix, Fig. S2H), M3 exerted a significantly stronger influence on M1 during chewing than during lipsmacks (SI Appendix, Fig. S2G). Within the lateral system, higher information flow from M1 to PMv was detected during threats (9 to 15.5 Hz) compared to voluntary chewing. In contrast, information flow from PMv to M1 was similar between both movement types (SI Appendix, Fig. S2L). As shown in Fig. 4, the network analysis reveals that motor areas show largely symmetric information flow (except for M3-to-M1 during threats Fig. 4A, or chews SI Appendix, Fig. S2J), while motor-to-somatosensory information flow is predominantly unidirectional (Fig. 4 A and B). While interareal synchrony was stronger within the lateral motor areas and weaker for medial-lateral motor interactions, information flow remains reliable across both medial and lateral regions, indicating continuous information exchange through the facial motor system (Fig. 4B). The pattern of differential synchrony strength coupled with reliable information exchange demonstrates how this sensorimotor network maintains flexible functional interactions. These interactions can be dynamically modulated to coordinate neural activity patterns underlying different facial gestures. Summary of Network analysis during facial expressions, and local oscillatory activity within motor face and S1 areas. (A and B) Summary network diagram depicting the Granger’s Causality maximum peak between each pair of motor areas and S1 during threats (Left panel) and lipsmacks (Right panel). Color and thickness of the line indicate the flow’s maximum magnitude for each direction. (A) Medial-Lateral Directional Information Flow: The upper panel depicts the flow from medial (M3, orange node) to lateral motor and S1 areas, while the lower panel shows flow from lateral areas (M1-PMv-S1, blue nodes) to medial motor area (M3). (B) Lateral–Lateral Directional Information Flow: Upper panel depicts flow from M1 to lateral motor areas (PMv and S1), while the lower panel shows flow from area PMv (blue node) to M1 and S1. (C) Raw power around the movement onset (±1 s) for threats (T: red) and lip-smacks (LS: yellow) in M1, PMv, M3, and S1. Alpha (α = 8 to 12 Hz) and beta (β = 13 to 40 Hz) activity was modulated across all behaviors. Asterisks indicate significant power differences in α and β ranges between threats and LS (P ≤ 0.01, Kruskal–Wallis test corrected for multiple comparisons, n = 2 monkeys, T = 432, LS=292). (D) Time frequency representations (TFR, 5 to 40 Hz) normalized to premovement epoch for threats and lipsmacks in M1, PMv, M3, and S1. Time zero indicates movement onset. Asterisks show significant clusters (P ≤ 0.05, cluster-corrected nonparametric randomization test) comparing premovement (–0.75 to –0.25 s) versus movement epochs (0.0 to 0.5 s). Red asterisks denote activity increases, blue asterisks indicate decreases relative to baseline. Facial Movement-Specific Local Power Dynamics. After identifying distinct connectivity and directionality patterns during emotional and voluntary facial movements, we investigated how local neural activity in each network area relates to facial gestures, focusing on oscillatory power modulations that reflect the local neural population activity (22). To investigate this, we analyzed the region-specific LFP dynamics. Fig. 4C and SI Appendix, Fig. S4 A–C, illustrate modulation in the raw local power across the different frequency bands (alpha: 8 to 12 Hz, beta:13 to 40 Hz, and gamma 40 to 120 Hz) in lateral (M1 and PMv), medial (M3), and somatosensory (S1) areas during different facial expressions, analyzed from 1 s pre- to 1 s postmovement onset. Significant differences in the power between behaviors were confirmed by the Kruskal–Wallis test (P ≤ 0.01, SI Appendix, Tables S4 and S5), with each behavior showing distinct patterns of power modulation across regions. To assess the temporal pattern of the LFP signals in different areas and varying facial movements we normalized the power of the ongoing oscillations during the movement period against the preparatory period (Materials and Methods). In motor areas, we found a ubiquitous pattern of beta-suppression as has been identified during voluntary movements of the upper limb (5, 28). Nonetheless the details of the suppression dynamics were movement and area-dependent (Fig. 4D). For instance, in M1 the beta suppression was locked to movement onset and lasted longer during threats than lipsmacks. In PMv, beta suppression showed higher amplitude, ramped up a bit before the movement onset, and reached a maximum power during movement. In M3, suppression was prominent, more pronounced, and longer for lipsmacks than for threats. Voluntary chew movements also evoked beta suppressions with dynamics slightly different from emotional movements (for details, see SI Appendix, Fig. S3). In addition, we found in both M1 and M3 an increase in gamma power (>40 Hz), during emotional and voluntary movements (SI Appendix, Fig. S4 E–G). In S1, beta suppression was modest, reaching significance only during chewing movements (SI Appendix, Fig. S3), while movement-related increases in alpha, high beta (>20 Hz), and gamma power, were dominant across all movement types. In summary, our findings reveal area-specific oscillatory patterns reflecting most likely distinct neural computations within each region. Activity in both medial and lateral facial motor areas showed only subtle differentiation between emotional and voluntary movements. In contrast, we observed robust interareal synchronization based on consistent phase relationships predominantly in alpha/beta bands that varied systematically across different facial gestures. These synchronized interactions between medial area M3, lateral areas (M1, PMv), and S1 occurred with frequency-specific coupling patterns unique to each facial gesture. The magnitude and directionality of information flow within this network dynamically reconfigures depending on the type of facial movement. Discussion In this study, we identified functional interactions between medial and lateral facial motor areas and the primary somatosensory cortex during facial expressions and voluntary movements in nonhuman primates, with facial expressions representing behaviors that are fundamental for social communication. Our results demonstrate a distributed cortical sensorimotor network with functional overlap between the previously proposed separate emotional (medial) and voluntary (lateral) circuits at the cortical level. This finding, at the cortical level, challenges the classical two-pathway model based on stroke patient studies (12, 14) and brain imaging research (29, 30) which proposes separate control circuits for emotional and voluntary facial movements. We instead found modulations in the local activity of both medial and lateral facial motor areas for both types of movements. While there was only subtle differentiation in local activity (measured as site-specific LFP power) between movement types, we identified specific frequency-based coupling among these regions. This suggests that facial expressions could emerge from a broad sensorimotor network with dynamic and flexible connectivity patterns that adapt based on the movement type. The prominence of these functional interactions was particularly evident during emotional facial expressions, where medial and lateral motor cortices showed synchronized activity despite their distinct local oscillatory signatures. This integrated sensorimotor network perspective aligns with recent human studies (31) showing that lateral motor areas receive emotional movement-related information, as evidenced by M1 activation during both voluntary speech production and emotional movements like smiling. Crucially, our ICMS evoked interareal connectivity results provide evidence for this network architecture, by revealing connectivity between medial and lateral facial motor areas. The consistency of interareal interactions across both methodologies – neural synchronization and electrically evoked responses – validates our results about functional interactions within this motor network. Neural communication among populations is mediated, at least in part, by their synchronous activity (23, 32), which is fundamental for the dynamic coordination of distributed neural activity in local and extended networks underlying sensorimotor processes (33–36). Information processing between distant neural assemblies depends on the strength of coherent activity through synchronous oscillations (36), with long range interactions involving a broader spectrum of frequency bands including theta, alpha, and beta frequencies (37). Our results align with and extend these fundamental principles of neural communication, revealing specific patterns of synchrony between the neural populations of medial (M3) and lateral motor areas (M1 and PMv), as well as with the somatosensory cortex (S1) in the alpha and beta range, during facial expressions, extending into the theta range for voluntary chewing. The synchrony between areas, measured as phase relationships between the oscillations, was independent of changes in local power, and both the degree of synchrony and interarea influence varied with the type of facial movement. Alpha synchrony typically reflects the inhibition of task-irrelevant information (38, 39), crucial for executive functions (36). The patterns we observed in the face sensorimotor network align with this gating role. We found differential alpha coupling between lipsmacks and threats in key sensorimotor areas (medial-lateral: M3-PMv, M3-M1, M3-S1; lateral: M1-PMV, PMv-S1; SI Appendix, Table S2) suggesting expression-specific synchrony modulations. This supports alpha’s role in gating information based on context and attention (39). The modulations in alpha synchrony across facial expressions likely regulate information flow during facial movements by suppressing irrelevant inputs while maintaining attention on motor behavior. Complementing the role of alpha synchrony, beta synchrony plays a crucial role in the large-scale coupling of sensorimotor information (40). Beta band coherency is ideally suited for flexibly and dynamically forming neural assemblies (41), and long-distance interarea communication (37). Beta oscillations can also be context-specific, reflecting current task rules or decisions (42, 43). In our case, shifts in the beta frequency peaks across different facial expressions (Fig. 3 A and B medial-lateral connectivity M3-M1, M3-PMv; Table 1) may contain information about the decision to perform specific facial movements in particular social contexts (43, 44). The significance of beta synchrony extends beyond our specific findings. Previous research has demonstrated beta synchrony’s crucial role in synchronizing large-scale cortical networks in sensorimotor (45) and social decision-making tasks (46). The occurrence of beta oscillations in the face motor system, with expression-specific frequency shifts, reinforces beta’s role as a rhythm for sensorimotor network synchronization across different effector systems, while suggesting these variations may reflect the selection process underlying specific facial movements. In summary, our data confirm that phase coupling is highly specific to gesture type, frequency, and the areas engaged in the coupling. The distinction between emotional and voluntary facial movements is frequency dependent: Coupling at low frequencies (theta, alpha) shows larger differences within the emotional movements category (threat > lipsmack), while coupling in the beta range between medial and lateral areas (M3-PMv) and between these areas and S1 reveals a clear categorical distinction where both emotional expressions exhibit stronger coupling than voluntary chewing. This suggests low frequencies encode information specific to each gesture, whereas beta coupling may specifically encode aspects related to the production of emotional movements. To understand how this synchrony-mediated coordination occurs, we examined information flow using Granger Causality analysis, which quantifies the extent to which activity in one area significantly predicts subsequent activity in another area. Results showed continuous bidirectional exchange between medial and lateral motor regions, as well as among lateral regions, indicating dynamic information flow during facial expressions. We observed behavior-specific asymmetries, such as enhanced information flow from medial area M3 to lateral motor area M1 during threats. This asymmetry might reflect the unique demands of threat displays, which require sustained, nonaffiliative movements that may depend more heavily on M3 control, unlike the affiliative and rhythmic patterns of lipsmacks in which the M3-to-M1 information flow was attenuated. Importantly, while we categorized lipsmacks as emotional movements, other studies suggest that they may be evolutionary precursors of human speech, potentially incorporating both emotional and voluntary control components (47). This dual nature, along with the opposing valence between threats and lipsmacks, may contribute to the observed asymmetric information flow pattern from medial-to-lateral areas. In contrast, PMv, showed symmetric information flow patterns between medial and lateral areas during both emotional movements, suggesting it could function as a central hub facilitating bidirectional information exchange (Fig. 3 H and I) (48). The primary somatosensory cortex exhibited asymmetric communication patterns, receiving more information from motor areas than it sent back. While this incoming flow might represent information about planned or ongoing movement (49), further studies are needed to test this hypothesis. Our results demonstrate that information flow within the face sensorimotor network exhibits varying patterns of functional symmetries and asymmetries, resulting in different degrees of influence between the interconnected areas. One could speculate about the existence of a flexible functional hierarchy within this network shaped by the kind of facial movement, similar to what has been proposed for the lateral forelimb motor cortices in mice (50). Such a flexible, context-dependent, switch in hierarchy may explain the discrepancy between the extensive medial-to-lateral interactions found in this study and the specific emotional vs. voluntary impairments reported in patients suffering from medial vs. lateral lesions (12). In summary, our findings reveal a dynamic cortical sensorimotor network with flexible connectivity that adapts based on the behavioral context and facial movement type. Intracortical stimulation showed that medial and lateral areas exert significant functional impact on each other, while simultaneous LFP recordings confirmed these regions interact primarily in alpha and beta frequency ranges during facial expressions. Rather than separate motor cortical circuits for emotional and voluntary movements, our data suggest that facial control relies on coordinated interareal interactions within an extensive sensorimotor network. This finding challenges the hypothesis of independent cortical medial/lateral streams for facial motor control. This network likely enhances signal coordination to the facial nucleus by integrating information across brain regions. Importantly, cortical integration does not exclude the possibility of downstream segregation. Beyond the sensorimotor network itself, a key question for future research is how this network for facial expression production interfaces with a broader social network comprising the STS and limbic structures (1, 51). The STS, recognized as the third visual pathway (52), provides dynamic social cues such as facial expressions, eye gaze, and body movements—while limbic structures (insula, orbitofrontal cortex, and amygdala) encode emotional and reward information. Understanding how these systems interface through the cingulate and premotor areas (20, 53–56) is crucial for explaining how value, emotion, and social context influence facial motor output decisions. Based on known anatomical connections, we propose that social and emotional information flows from these regions—with the STS detecting dynamic social signals and limbic structures computing emotional valence and reward value—converging in cingulate and premotor areas where these inputs are integrated into the motor network to guide the decision about which facial expression to produce. Understanding this integrated network architecture has important clinical implications. Facial palsy affects over 60% of stroke patients, with deficits often persisting beyond 40 d(57), causing significant functional and aesthetic impairments (58). While traditional predictive models focus on lesion characteristics, these have not translated into effective therapeutic interventions. Our findings offer a different contribution: We identify the cortical facial motor network (M1-PMv, M3-PMv, M3-M1 extending to S1) as a potential therapeutic target for neuromodulation-based rehabilitation strategies. Specifically, PMv showed symmetric information flow patterns between both medial and lateral areas, identifying it as a potential hub facilitating bidirectional information exchange (Fig. 3 H and I and SI Appendix, Fig. S2 H and K). These findings suggest specific application for poststroke facial palsy rehabilitation using network-targeted neuromodulation (TMS, tDCS), which has shown promise in motor recovery by inducing spike-timing-dependent plasticity (59–63). While bihemispheric approaches may yield different results for facial versus limb motor control given the automaticity of symmetric facial muscle activation, our results particularly highlight PMv as a target for intrahemispheric neuromodulation therapy. Targeting PMv with neuromodulation, with the aim of enhancing both medial (PMv-M3) and lateral (PMv-M1) cortical connectivity upstream to the frequently injured M1, would be both a technically feasible and well-motivated hypothesis. Finally, we acknowledge the putative overlap between motor and premotor cortical representations and descending outputs to laryngeal motor neurons controlling vocalizations (64), highlighting the need for future experiments to elucidate the neural mechanisms underlying co-occurring facial expressions and vocal output during social communication. Understanding both the dynamics of this cortical network during facial expressions alone versus when accompanied by vocalizations, and the temporal dynamics between these behaviors, is fundamental to understanding how human speech could have evolved. Materials and Methods Full methodological details are available in the SI Appendix. All animal housing, care, and experimental procedures complied with the NIH Guide for Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committees of the Rockefeller University (protocol numbers 21104-H USDA) and Weill Cornell Medical College (protocol number 2010-0029). Subjects. We studied two male monkeys [Macaca fascicularis (M1) and Macaca mulatta (M2)] that were pair housed before and during the experiments. fMRI Behavioral Task and Acquisition. Monkeys performed a social interaction free-viewing task as described in Shepherd & Freiwald (20), and an oro-facial motor localizer in which sparse, intermittent fluid reward was delivered during a passive fixation task. Facial movements were tracked using an MRI-compatible infrared video camera. The onset and offset of the facial emotional (threats, lipsmacks), voluntary (chew, drinking), and other (yawn) movements was determined offline by two independent observers, manually annotated and cross-validated. Detailed information on the fMRI task, acquisition protocol, and analysis can be found in the SI Appendix. fMRI Analysis and Face Sensorimotor Cortex Localization. We used GLM models to identify facial sensorimotor cortical areas. Facial expression onset times were convolved with the MION HRF to generate a primary regressor of interest that will result in a “face expression activity map”. A second GLM model using juice delivery times created an “orofacial motor activity map” (SI Appendix). Significant voxels in both maps were identified, with a conjunction analysis using logical ANDs, taking the least significant P-value from each contrast entered in the conjunction. The identified voxels primarily correlated with oro-facial movement, and not body movements during the social interaction free-viewing task. The conjunction facial expression-orofacial motor map was used for surgical targeting. fMRI-Guided Electrophysiological Recording sites. We used CORTEXPLORER SCI, an integrated functional/structural MRI-based stereotaxic planning and intrasurgical tracking approach (cortEXplore GmbH, Linz Austria) to target the 6 to 8 FMAs (floating microarrays, Microprobes) based on the fMRI facial expression-orofacial motor map (SI Appendix). We targeted the face representation in the following areas based on the fMRI map: primary somatosensory cortex (S1), primary motor cortex (M1), ventro-lateral premotor cortex (PMv), and rostral cingulate motor area (M3). All arrays were implanted in the right hemisphere of both animals (Fig. 1C and SI Appendix, Fig. S1) Neural Recording Behavioral Task. We used a randomized block design where each recording day was divided into 7 to 10-minute blocks, with behavioral and electrophysiological data recorded from head-fixed monkeys in different behavioral conditions. Block order was randomized to control for time and attentional effects, with a minimum of 3 different block types recorded daily: rest (baseline/control condition), voluntary chewing/ingestive movements, and emotional facial expressions for social communication including threats (aggressive gestures) and lipsmacks (affiliative gestures). During rest blocks, animals remained calm with minimal movement. During chew blocks, animals were intermittently fed fruit pieces. Facial expressions were elicited using a battery of stimuli including conspecific face stimuli from social interaction task (20), conspecific movie stimuli, monkey avatar stimuli, and live interactions with conspecifics and humans (SI Appendix, Visual Stimuli). Animals were previously trained on fixation tasks to ensure attention to stimuli. During recordings, the reward tube was removed to prevent mouth occlusion and disambiguate social expressions from ingestive movements. While facial expression blocks were generally dominated by one expression type, some datasets contained both threats and lipsmacks within the same block. Every trial began with a variable period of required visual fixation, after which a video stimulus appeared. Subjects were able to freely move their eyes to explore the videos as long as their gaze stayed within a virtual window the size of the frame of the video frame plus an additional 2 degrees surrounding it. No reward was given during these runs, and the reward tube was removed. Facial movements were recorded using a FLIR camera (70 Hz, monkey 1) and Optitrack system (120 Hz, monkey 2). Eye movements were tracked with ISCAN. All systems were synchronized with a TDT system (details in SI Appendix). Electrophysiological Recordings. Neural activity was recorded simultaneously using FMA arrays with 36 channels per array (FMA, Microprobes Inc) with spacing of 400 µm, and 4 × 1.8 mm dimensions. We recorded a total of 256 channels (monkey1) and 192 (monkey2) of 6 to 8 individual arrays implanted in the face representation of the following areas: S1, M1, PMv, and the rostral motor cingulate (M3). Recording electrode impedances were ~0.5 MOhm (n = 28), microstimulation electrode impedance ~10 K and less than 10 K for ground (n = 2) and reference (n = 2) electrodes (SI Appendix). Data Acquisition. Neural activity was recorded at full bandwidth with a sampling frequency of 24 kHz, and a resolution of 16 bits using a Tucker-Davis Technologies (TDT) system RZ2 Biosignal Processor. Neural data and all behaviors were synchronously stored to disk together with the face videos and eye movements. Raw recordings were filtered offline to obtain LFP (1 kHz). Functional Connectivity Paradigm. In monkey 1, the paradigm consisted in applying a (250 µA, a train of single biphasic pulses, each pulse duration = 0.4 ms at a frequency of 3 Hz (333 ms interpulse interval) delivered for ~2 min) through each of the stimulation electrodes. Current acquisition was done at 24 kHz. Neural activity was stored as described above. Each array contained four low impedance (~10 K) electrodes evenly spaced across the array, and 28 recording electrodes. Crucially, stimulation was delivered while the animal sat quietly without making facial or body movements. Using this current protocol, we did not evoke any facial or body movements. These parameters were specifically chosen to evoke neural responses in the receiver area rather than overt movements. Detailed information on functional connectivity analysis is available in the SI Appendix. Behavioral Analysis. Facial expressions, voluntary chewing movements, and other facial movements were manually scored by 3 human observers for one monkey. The onset of the facial movements was determined by the time at which the subject’s mouth started to move preceded by one second of mouth stillness. For the other monkey, we used for half of the sessions a customized python based automatic detector for mouth movements, based on the optic flow of the recorded video. Details and scoring validation are provided in SI Appendix. Spectral Analysis. LFP preprocessing is detailed in the SI Appendix. Using a fast Fourier transform approach, we computed time frequency representations (TFR) of the power (5 to 120 Hz). For the range of 5 to 40 Hz, we used an adaptive sliding time window 4 cycles long (Dt = 4/f) multiplied with a Hanning taper. For higher frequencies (40 to 120 Hz), we used multitapers Slepian sequences in a 0.2 s window with ±10 Hz of smoothing. The windows moved in .05 s steps, from 1 s before the movement onset to 1 s after the movement execution. Per type of facial movement condition, i.e., threat, lipsmack, or chew we averaged over trials (SI Appendix, Table S6) within each recording session and then normalized by an absolute baseline subtraction using a 0.75 s window before movement onset. For visualization purposes, a grand average was computed over the recording sites (per area). In the case of the raw power, a similar analysis was performed, except for the normalization. Statistical analyses were computed using the raw power. Connectivity Analysis. We measured functional connectivity between the pairs of recorded areas by computing the PPC measurement as implemented in Fieldtrip (25, 65). The PPC is an unbiased and consistent estimate of how well two signals generated by two sources show a consistent phase relationship in a particular frequency band (5 to 50 Hz). The advantage of using PPC versus other connectivity measurements is 1) it remains reliable even with few trials, and 2) it is less prone to errors that occur when the consistency of phase differences correlates with the strength of the signal (25). In the case of the connectivity analysis, we focused on the movement execution period, to maximize trial numbers and ensure certainty about the nature of the ongoing facial movement type. We divided each full movement period into 0.5 s epochs, with each epoch serving as an individual trial for that behavior (SI Appendix, Table S7). As a first approach, we calculated the connectivity (PPC) of the network when the face was at rest, i.e. not moving or planning to move and compared it with the connectivity when the face was moving. Statistical differences within this contrast were assessed by a cluster-based nonparametric randomization (65), with a dependent-samples t test as statistic. We used bivariate, nonparametric Granger Causality as implemented in Fieldtrip during the movement period to investigate the directionality of the connectivity (25) during the movement epoch. Detailed statistical methods are described in the SI Appendix, Statistical Analysis. Editorial Assistance. ChatGPT v4 (OpenAI) and Claude 4.5 Sonnet were used for English language editing and to improve manuscript clarity and readability. Neither of them contributed to scientific content or data interpretation. The final version of this manuscript was thoroughly reviewed and revised by all the authors, who are responsible for all content. Supplementary Material Appendix 01 (PDF)