Top-down influence of the inferior frontal gyrus on motor cortex excitability during pain empathy Authors: Pengju Zhang, Yitong Lin, Yuting Li, Jianing Wei, Xinfan Wu, Xiaoying Tan, Jian Zhang Published: 2026 ABSTRACT Empathy for pain can reduce activity in the observer’s motor system, but the neural processes that shape this response remain unclear. Here we show that observing another person in pain reduces corticospinal excitability and increases perceived pain intensity when viewed from a first-person perspective, but not from a third-person perspective. This motor suppression is absent when noxious events are applied to a non-biological object. Using single-pulse, paired-pulse, and dual-site transcranial magnetic stimulation, we find no detectable change in local motor cortical inhibitory or facilitatory circuits, but observe a pain-specific reduction in the connectivity index between the inferior frontal gyrus and primary motor cortex. This effect covaries with perceived pain intensity and motor suppression and is site-specific relative to a Vertex control condition. Bayesian network modeling further suggests that perceived pain intensity predicts both unpleasantness and motor excitability. Together, these findings suggest that pain appraisal may be associated with frontal–motor interactions during pain empathy. Introduction When witnessing another person sustain a sports-related muscle tear, observers not only recognize the intensity of the other’s pain but may also vicariously experience a corresponding bodily sensation. Across both direct physical injury and psychologically driven experiences, such as social exclusion, pain empathy is not merely passive emotional resonance, but a foundation for interpersonal relationships and social bonds1. It contributes to mutual understanding and cooperation, promotes prosocial behavior, inhibits aggression, and shapes social morality2. Recent evidence indicates that pain empathy is multi-component, comprising affective, cognitive, and motor components3. A well-calibrated somatic resonance system, shaped by prior nociceptive experience, appears necessary for accurately estimating others’ pain4. Neurophysiological studies show that observing bodily injury reduces corticospinal excitability (CSE) in the observer’s primary motor cortex (M1) corresponding to the injured body part, suggesting motor-system resonance with others’ pain5–7. This motor component also shows atypical patterns in specific populations, including attenuated motor representations in autism spectrum disorder8 and atypically enhanced motor responses in individuals with high psychopathic traits9. These findings suggest that abnormalities in the motor component may be closely associated with social dysfunction. However, the mechanisms by which the motor component of pain empathy arises remain poorly understood. The perception–action model (PAM) proposes that understanding another’s emotional state involves mapping the observed state onto one’s own representations, generating analogous motor representations through observation–execution matching, and thereby accessing others’ goals and emotions via bottom-up processing10–13. Nociceptive signals from muscle are conveyed by group III and IV afferent fibers to the spinal cord, ascend to the thalamus, and project to the motor cortex14–16, where they modulate inhibitory and excitatory circuits to limit activity in pain-related muscles and prevent further damage17–19. If the motor component of pain empathy reflects such bottom-up automatic simulation, observing another’s pain should induce comparable inhibitory/excitatory changes in the observer’s motor cortex, thereby reducing CSE. Alternatively, cognitive models propose that affective and behavioral responses in pain empathy depend on cognitive appraisal of another’s pain20. Neuroimaging studies implicate the inferior frontal gyrus (IFG), particularly the right IFG (rIFG), in processing pain conveyed by facial expressions and limb-based cues, highlighting its role in extracting the meaning of observed pain21,22. Causal evidence from high-frequency repetitive transcranial magnetic stimulation (TMS) and high-definition transcranial direct current stimulation further indicates that rIFG activity supports cognitive processing in pain empathy and relates to individuals’ ability to evaluate others’ pain23,24. Thus, if the motor component is primarily governed by top-down cognitive control, IFG–M1 interactions should emerge during pain empathy and modulate CSE. The critical question is therefore whether the motor component of pain empathy depends mainly on bottom-up perception–action simulation, top-down modulation from higher-order regions, such as the IFG, or their interaction. TMS has been extensively applied to investigate the neurophysiological underpinnings of the motor component in pain empathy. Single-pulse TMS (spTMS) delivers a solitary magnetic pulse to the motor cortex, eliciting motor evoked potentials (MEPs) in target muscles and thus measuring changes in CSE. Although spTMS studies have documented alterations in CSE when participants observe others in pain7,8,25–31, spTMS cannot probe intracortical inhibitory/excitatory circuits or interactions between the motor cortex and other brain areas, limiting insights into the specific neural mechanisms of the motor component. Paired-pulse TMS (ppTMS), by varying interstimulus intervals (ISIs) between two pulses, can quantify short-interval intracortical inhibition (SICI) and intracortical facilitation (ICF), thereby elucidating intracortical circuit dynamics. Dual-site paired-pulse TMS (dsTMS), which simultaneously stimulates two regions (e.g., IFG and M1), can directly reveal functional interactions between these areas, providing physiological evidence of information transfer32,33. Together, ppTMS and dsTMS offer useful means to dissect the neurophysiological mechanisms underlying the motor component of pain empathy. To address these questions, we first used a perspective manipulation. Prior work shows that, relative to a first-person perspective, adopting a third-person perspective reduces perceived pain intensity without altering the physical nociceptive attributes of the stimulus34. Experiment 1 used spTMS to test whether CSE suppression during pain observation is driven primarily by bottom-up stimulus processing or by subjective pain appraisal by comparing painful and non-painful videos presented from first-person and third-person perspectives. A control experiment replaced the human hand with a non-biological object to exclude effects of visual salience or threat. Experiment 2 used ppTMS to examine whether pain observation alters SICI and ICF circuits in M1, and Experiment 3 used dsTMS to assess rIFG–M1 connectivity, with Vertex serving as a control site. We hypothesized that greater perceived pain would be associated with reduced CSE, absent for non-biological objects, and accompanied by changes in SICI/ICF and IFG–M1 connectivity. Data from all experiments were integrated using a Bayesian network (BN), with CSE representing the motor component and questionnaire measures indexing cognitive and affective components. The BN modeled dependencies among these components and enabled bidirectional inference between motor, cognitive, and affective states. Across experiments, painful events selectively increased perceived pain and suppressed CSE under the first-person perspective, whereas identical noxious stimuli applied to a non-biological object did not affect motor excitability. Although no reliable changes were observed in SICI or ICF, pain observation reduced IFG–M1 connectivity, and this reduction covaried with perceived pain intensity and CSE suppression. The BN further identified perceived pain intensity as a key predictor of both unpleasantness and motor excitability. Together, these findings suggest that empathic motor inhibition is linked to pain appraisal and that pain-related changes in IFG–M1 connectivity may reflect a potential top-down mechanism through which pain appraisal influences motor cortex excitability. Results Specific reduction of CSE in the primary motor cortex during pain empathy To examine changes in CSE of M1 during pain empathy, we applied spTMS and recorded MEPs from the right first dorsal interosseous (FDI) muscle (Fig. 1a). Video stimuli were presented in a 2 (Condition: painful, non-painful) × 2 (Perspective: first-person, third-person) within-subjects design. MEPs recorded under the baseline condition were used for normalization; their analysis is reported in the Supplementary Materials (see Supplementary Results for Experiment 1 and Supplementary Fig. 1).Fig. 1Experimental procedure and stimuli for Experiments 1–3.a Schematic of the pain empathy induction task in Experiment 1. Each trial began with a 1000 ms fixation cross, followed by an 1800 ms presentation of a painful, non-painful, or baseline condition, and a subsequent 7200 ms blank screen. Stimuli were presented either from a first-person or a third-person perspective. A spTMS was delivered 200–600 ms before the offset of the stimulus. b Schematic of the procedures for Experiments 2 and 3. Trials followed the same structure, but all stimuli were presented from the first-person perspective. Experiment 2 employed ppTMS to assess intracortical excitability, and Experiment 3 employed dsTMS to assess IFG–M1 connectivity. Experimental procedure and stimuli for Experiments 1–3. a Schematic of the pain empathy induction task in Experiment 1. Each trial began with a 1000 ms fixation cross, followed by an 1800 ms presentation of a painful, non-painful, or baseline condition, and a subsequent 7200 ms blank screen. Stimuli were presented either from a first-person or a third-person perspective. A spTMS was delivered 200–600 ms before the offset of the stimulus. b Schematic of the procedures for Experiments 2 and 3. Trials followed the same structure, but all stimuli were presented from the first-person perspective. Experiment 2 employed ppTMS to assess intracortical excitability, and Experiment 3 employed dsTMS to assess IFG–M1 connectivity. A repeated-measures analysis of variance (ANOVA) on normalized CSE revealed a significant main effect of condition (F1, 28 = 9.13, p = 0.005, ηp² = 0.25), indicating overall lower CSE during the painful than the non-painful condition. The main effect of perspective did not reach significance (F1, 28 = 1.37, p = 0.251, ηp² = 0.05). Importantly, the condition × perspective interaction was significant (F1, 28 = 6.30, p = 0.018, ηp² = 0.18). Simple-effects analyses showed that, in the painful condition, first-person perspective elicited significantly lower CSE than third-person perspective (difference: 0.29 ± 0.11, t28 = 2.63, p = 0.014, Cohen’s d = 0.49, 95% CI: −0.52, −0.06, Fig. 2a). By contrast, within the non-painful condition, no difference between perspectives was observed (p = 0.437). Conversely, when comparing conditions within each perspective, the painful condition elicited significantly lower CSE than the non-painful condition under the first-person perspective (difference: 0.42 ± 0.10, t28 = 4.15, p = 2.77 × 10−4, Cohen’s d = 0.77, 95% CI: −0.63, −0.21, Fig. 2a), whereas no condition difference emerged under the third-person perspective (p = 0.753).Fig. 2Results of Experiment 1, testing the influence of pain empathy on corticospinal excitability.a CSE in the right FDI muscle during observation of painful and non-painful conditions from the first-person and third-person perspectives. CSE suppression was evident only under the first-person painful condition, but not under the third-person painful condition. b Perceived pain intensity ratings. Painful conditions were consistently rated as more intense than non-painful conditions, with higher ratings under the first-person perspective than the third-person perspective. c Perceived unpleasantness ratings. The painful condition was judged more unpleasant than the non-painful condition regardless of perspective, with no significant modulation by perspective. d Correlation between CSE suppression and perceived pain intensity under the first-person painful condition. Stronger CSE reduction was associated with higher pain intensity ratings. n = 29 biologically independent participants. The split-half violin plots show the probability density of the data, and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. Horizontal dashed line indicates the null value of 1. Asterisks indicate significant post hoc comparisons: *p < 0.05, ***p < 0.001. Daggers indicate significant deviations from the null value of 1 based on one-sample tests: †††p < 0.001. Source data are provided as Supplementary Data 1. Results of Experiment 1, testing the influence of pain empathy on corticospinal excitability. a CSE in the right FDI muscle during observation of painful and non-painful conditions from the first-person and third-person perspectives. CSE suppression was evident only under the first-person painful condition, but not under the third-person painful condition. b Perceived pain intensity ratings. Painful conditions were consistently rated as more intense than non-painful conditions, with higher ratings under the first-person perspective than the third-person perspective. c Perceived unpleasantness ratings. The painful condition was judged more unpleasant than the non-painful condition regardless of perspective, with no significant modulation by perspective. d Correlation between CSE suppression and perceived pain intensity under the first-person painful condition. Stronger CSE reduction was associated with higher pain intensity ratings. n = 29 biologically independent participants. The split-half violin plots show the probability density of the data, and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. Horizontal dashed line indicates the null value of 1. Asterisks indicate significant post hoc comparisons: *p < 0.05, ***p < 0.001. Daggers indicate significant deviations from the null value of 1 based on one-sample tests: †††p < 0.001. Source data are provided as Supplementary Data 1. One-sample t-tests against the null value of 1 showed that only the CSE index in the painful condition under the first-person perspective was significantly below baseline (difference: −0.34 ± 0.28, t28 = −6.55, p = 4.26 × 10⁻7, Cohen’s d = −1.22, 95% CI: −0.45, −0.24), whereas all other conditions did not differ from 1 (all p > 0.05). Together, these results demonstrate that absolute CSE suppression during pain empathy occurred selectively when participants viewed the painful event from a first-person perspective. For perceived pain intensity, repeated-measures ANOVA revealed a significant main effect of condition (F1, 28 = 216.87, p = 1.033 × 10−14, ηp² = 0.89) and a significant condition × perspective interaction (F1, 28 = 6.35, p = 0.018, ηp² = 0.18), whereas the main effect of perspective was not significant. Simple-effects analyses showed that in the painful condition, first-person perspective yielded higher ratings than third-person perspective (difference: 0.79 ± 0.34, t28 = 2.31, p = 0.028, Cohen’s d = 0.43, 95% CI: 0.09, 1.50, Fig. 2b), whereas no perspective effect emerged in the non-painful condition. Comparing conditions within each perspective, the painful condition was rated as substantially more intense than the non-painful condition under both the first-person (difference: 4.66 ± 0.36, t28 = 13.11, p = 1.807 × 10−13, Cohen’s d = 2.43, 95% CI: 3.93, 5.38, Fig. 2b) and third-person perspectives (difference: 3.85 ± 0.30, t28 = 12.75, p = 3.528 × 10−13, Cohen’s d = 2.37, 95% CI: 3.23, 4.47, Fig. 2b). These findings demonstrate that while painful condition was consistently judged as more intense than non-painful condition, this effect was amplified under the first-person perspective and diminished under the third-person perspective, indicating that visual perspective effectively modulated the appraisal of others’ pain intensity. For perceived unpleasantness, repeated-measures ANOVA revealed a significant main effect of condition (F1, 28 = 220.79, p = 8.263 × 10−15, ηp² = 0.89) and a significant condition × perspective interaction (F1, 28 = 5.05, p = 0.033, ηp² = 0.15), whereas the main effect of perspective was not significant. Simple-effect analyses showed that within the painful condition, unpleasantness ratings did not significantly differ between first-person and third-person perspectives (p = 0.117). Similarly, no perspective difference was observed in the non-painful condition (p = 0.786). By contrast, when comparing conditions within each perspective, the painful condition was rated as substantially more unpleasant than non-painful condition under both the first-person perspective (difference: 4.59 ± 0.30, t28 = 13.91, p = 4.208 × 10−14, Cohen’s d = 2.58, 95% CI: 3.92, 5.27, Fig. 2c) and the third-person perspective (difference: 4.00 ± 0.31, t28 = 13.10, p = 1.844 × 10−13, Cohen’s d = 2.43, 95% CI: 3.37, 4.63, Fig. 2c). These results indicate that unpleasantness was robustly elevated for painful condition irrespective of perspective, and was not significantly modulated by visual perspective. Correlation analyses conducted under the first-person perspective revealed that, in the painful condition, the CSE obtained with spTMS was significantly negatively correlated with perceived pain intensity ratings (Pearson r(27) = −0.51, p = 4.730 × 10−3, n = 29, 95% CI: -0.738, -0.176, Fig. 2d). No significant correlations were found between CSE and the perceived unpleasantness, the total Interpersonal Reactivity Index (IRI) score, or any of its four subscales. In sum, these results showed that perceived pain intensity was higher and CSE was significantly suppressed when a painful event was viewed from a first-person perspective, whereas perceived pain intensity was reduced and CSE suppression disappeared under the third-person perspective. Moreover, CSE was significantly correlated with perceived pain intensity, suggesting that motor system responses may be related to observers’ appraisal of others’ pain in pain empathy. Pain empathy did not induce significant modulation of intracortical excitability To clarify the inhibitory and excitatory intracortical mechanisms underlying CSE modulations during pain empathy, paired-pulse stimulation of M1 was administered under the first-person perspective condition to assess SICI and ICF (Fig. 1b). The single-pulse MEP amplitudes and the baseline SICI and ICF values underlying the standardization procedure are shown in the Supplementary Information (see Supplementary Results for Experiment 2 and Supplementary Fig. 2). Regarding the paired-pulse measures, the analysis revealed no significant change in SICI values under the painful condition compared to the non-painful condition, with Bayesian analysis indicating moderate evidence for the null hypothesis (difference: −0.04 ± 0.75, Wilcoxon signed-rank: V = 198, p = 0.909, effect size r = −0.02, 95% CI: -0.23, 0.22, BF01 = 4.80, Fig. 3a). Similarly, the painful condition did not significantly affect ICF, which also provided moderate evidence in favor of the null hypothesis (difference: 0.03 ± 0.85, t27 = 0.22, p = 0.830, Cohen’s d = 0.04, 95% CI: -0.37, 0.30, BF01 = 4.88, Fig. 3b). Taken together, these converging analyses indicate that pain empathy did not significantly modulate the specific intracortical inhibitory and facilitatory circuits indexed by SICI and ICF, suggesting that the observed reduction in CSE was not accompanied by detectable changes in the particular intracortical mechanisms captured by these two measures.Fig. 3Results of Experiment 2, testing the influence of pain empathy on intracortical excitability indices.a No changes were observed for the SICI index when participants viewed pain-related videos compared to non-pain-related videos (a, p = 0.909). b No changes were observed for the ICF index (b, p = 0.830). n = 28 biologically independent participants. The split-half violin plots show the probability density of the data and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. Source data are provided as Supplementary Data 1. Results of Experiment 2, testing the influence of pain empathy on intracortical excitability indices. a No changes were observed for the SICI index when participants viewed pain-related videos compared to non-pain-related videos (a, p = 0.909). b No changes were observed for the ICF index (b, p = 0.830). n = 28 biologically independent participants. The split-half violin plots show the probability density of the data and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. Source data are provided as Supplementary Data 1. Testing specific changes in intercortical connectivity between the IFG and M1 during pain empathy To investigate intercortical connectivity between the right IFG (Montreal Neurological Institute [MNI] coordinates: 48, 36, 6) and M1 during pain empathy, a dsTMS paradigm was employed under the first-person perspective condition to assess the effective IFG–M1 pathway (Fig. 1b). The single-pulse MEP amplitudes and connectivity values obtained under the baseline condition are shown in the Supplementary Information (Supplementary Fig. 3). We conducted a 2 (Condition: painful, non-painful) × 2 (ISI: 6 ms, 12 ms) within-subjects repeated-measures ANOVA to examine the effects of condition and time on IFG–M1 intercortical connectivity. The results revealed a significant main effect of ISI (F1, 22 = 5.25, p = 0.032, ηp² = 0.19), while the main effect of Condition was not significant (F1, 22 = 1.16, p = 0.292, ηp² = 0.05). Notably, the Condition × ISI interaction was significant (F1, 22 = 5.20, p = 0.033, ηp² = 0.19). Comparisons of ISI within each Condition showed that under the painful condition, connectivity at an ISI of 12 ms was significantly lower than at 6 ms (difference: 0.32 ± 0.09, t22 = 3.60, p = 0.002, Cohen’s d = 0.75, 95% CI: 0.13, 0.50, Fig. 4b), whereas under the non-painful condition, the difference in connectivity between 6 ms and 12 ms was not significant (p = 0.369). Comparisons of condition within each ISI indicated that at an ISI of 6 ms, there was no significant difference in connectivity between the painful and non-painful conditions (p = 0.503). In contrast, at 12 ms, connectivity was significantly reduced under the painful condition compared to the non-painful condition (difference: 0.16 ± 0.06, t22 = 2.53, p = 0.019, Cohen’s d = 0.53, 95% CI: −0.29, −0.03, Fig. 4b). The simple-effects analysis indicated that the time effect of ISI was moderated by Condition: under the painful condition, the difference between ISIs was significant, whereas under the non-painful condition it was not. Furthermore, at an ISI of 12 ms, IFG–M1 connectivity in the painful condition was significantly lower than in the non-painful condition.Fig. 4Results of Experiment 3, investigating the influence of pain empathy on IFG–M1 connectivity.a Anatomical localization of the stimulation site over the rIFG. The stimulation target is projected onto a 3D cortical surface template, displayed from a right lateral view (top) and a superior view (bottom). b Changes in interhemispheric connectivity between the IFG and M1 during pain empathy. The results indicated that the IFG exerted inhibitory connectivity to M1 only when the ISI was 12 ms. c Pearson correlation analysis showed a significant positive correlation between the magnitude of CSE reduction and the degree of inhibitory connectivity between the IFG and M1 during the viewing of pain-related video stimuli (r = 0.507, p = 0.013). d Pearson correlation analysis revealed a significant negative correlation between the perceived pain intensity of others and the degree of inhibitory connectivity between the IFG and M1 (r = -0.437, p = 0.037). n = 23 biologically independent participants. The split-half violin plots show the probability density of the data, and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. A horizontal dashed line indicates the null value of 1. Asterisks indicate significant post hoc comparison: *p < 0.05, **p < 0.01. Daggers indicate significant deviations from the null value of 1 based on one-sample tests: ††p < 0.01. Source data are provided as Supplementary Data 1. Results of Experiment 3, investigating the influence of pain empathy on IFG–M1 connectivity. a Anatomical localization of the stimulation site over the rIFG. The stimulation target is projected onto a 3D cortical surface template, displayed from a right lateral view (top) and a superior view (bottom). b Changes in interhemispheric connectivity between the IFG and M1 during pain empathy. The results indicated that the IFG exerted inhibitory connectivity to M1 only when the ISI was 12 ms. c Pearson correlation analysis showed a significant positive correlation between the magnitude of CSE reduction and the degree of inhibitory connectivity between the IFG and M1 during the viewing of pain-related video stimuli (r = 0.507, p = 0.013). d Pearson correlation analysis revealed a significant negative correlation between the perceived pain intensity of others and the degree of inhibitory connectivity between the IFG and M1 (r = -0.437, p = 0.037). n = 23 biologically independent participants. The split-half violin plots show the probability density of the data, and the boxplots show the median and first and third quartiles with whiskers extended to the most extreme data points that are no more than 1.50 times the interquartile range. Dots and lines denote individual data points. A horizontal dashed line indicates the null value of 1. Asterisks indicate significant post hoc comparison: *p < 0.05, **p < 0.01. Daggers indicate significant deviations from the null value of 1 based on one-sample tests: ††p < 0.01. Source data are provided as Supplementary Data 1. To assess whether these intercortical connectivity differences represented absolute changes, we conducted one-sample tests against the null value of 1. At the 12 ms ISI, the connectivity index under the painful condition was significantly lower than 1 (difference: -0.13 ± 0.21, t22 = −3.01, p = 0.006, Cohen’s d = −0.63, 95% CI: −0.22, −0.04), indicating an absolute shift toward increased inhibitory IFG–M1 connectivity. In contrast, none of the other conditions (the non-painful condition at 12 ms, and both conditions at 6 ms) showed significant deviation from 1 (all p > 0.05). Collectively, these findings indicate that, even after standardization relative to the corresponding single-pulse MEPs and connectivity value obtained under the baseline condition, a pain-specific reduction in the IFG–M1 connectivity index remained evident at the 12-ms interval. This pattern was not observed at the 6-ms interval or in the non-painful condition, and the absolute decrease of the standardized index below 1 under the painful 12-ms condition is consistent with a selective increase in inhibitory interactions indexed by the IFG–M1 dsTMS protocol during pain empathy. Correlation analyses showed that, under the painful video condition and at an ISI of 12 ms, the degree of inhibitory IFG–M1 connectivity was significantly positively associated with the magnitude of CSE reduction (Pearson r(21) = 0.507, p = 0.013, n = 23, 95% CI: 0.12, 0.76, Fig. 4c) and significantly negatively associated with perceived pain intensity (Pearson r(21) = −0.437, p = 0.037, n = 23, 95% CI: -0.72, -0.03, Fig. 4d), but was not significantly related to perceived unpleasantness, the total IRI score, or any of its four subscales. In addition, the magnitude of CSE reduction was itself significantly negatively correlated with perceived pain intensity (Pearson r(21) = −0.443, p = 0.034, n = 23, 95% CI: -0.72, -0.04). In sum, these findings indicate that viewing painful videos selectively enhanced inhibitory IFG–M1 intercortical connectivity at the longer latency (12 ms), and that this change was significantly associated with greater suppression of CSE as well as higher perceived pain intensity. This pattern supports a context-dependent, pain-specific modulation of the frontal–motor pathway, which may help explain the motor inhibition observed during pain empathy. Absence of motor resonance for non-biological objects To address whether the observed CSE reduction was specifically driven by body-related pain empathy rather than the visual salience or aversiveness of the needle, we conducted Experiment 4 (Supplementary Fig. 4a). In this experiment, the human hand was replaced with a non-biological object (an apple), while all other experimental parameters remained identical to Experiment 1. Similar to Experiment 1, a paired t-test first confirmed that MEPs obtained under the baseline condition did not significantly differ between the first-person and third-person perspectives (p = 0.229, Supplementary Fig. 4b). This confirms the stability of MEPs obtained under the baseline condition across perspectives, ensuring the reliability of the normalization procedure. A repeated-measures ANOVA on normalized CSE revealed no significant main effect of Condition (F1, 20 = 0.01, p = 0.914, ηp² = 0.00, BF01 = 4.37) or Perspective (F1, 20 = 0.71, p = 0.408, ηp² = 0.03, BF01 = 3.20). Crucially, the Condition × Perspective interaction was also non-significant (F1, 20 = 0.38, p = 0.544, ηp² = 0.02, BF01 = 3.70, Supplementary Fig. 4c). The Bayesian analysis yielded Bayes factor in favor of the null hypothesis (BF01) values greater than 3 for all effects, providing substantial evidence in favor of the null hypothesis. These results demonstrate that the needle-related stimuli did not modulate M1 excitability when applied to a non-biological object, confirming that the CSE inhibition observed in Experiment 1 is specifically tied to pain empathy rather than the mere visual presence of a sharp object or threat-related cues. Testing the site specificity of rIFG-conditioned modulation of M1 during pain empathy To further examine whether the modulation observed in Experiment 3 was specific to rIFG conditioning, a dsTMS paradigm was employed under the first-person perspective condition using Vertex (MNI coordinates: 0, 0, 75) as a control stimulation site (Supplementary Fig. 5b). The single-pulse MEP amplitudes and connectivity values underlying this standardization procedure are shown in the Supplementary Information (Supplementary Fig. 5). We conducted a 2 (Condition: painful, non-painful) × 2 (Stimulation Site: rIFG, Vertex) within-subjects repeated-measures ANOVA to examine the effects of condition and stimulation site on site-conditioned connectivity. The results revealed that the main effect of Stimulation Site was not significant (F1, 14 = 3.33, p = 0.089, ηp² = 0.19), and the main effect of Condition was also not significant (F1, 14 = 0.11, p = 0.742, ηp² = 0.01). Notably, the Condition × Stimulation Site interaction was significant (F1, 14 = 16.38, p = 0.001, ηp² = 0.54). Comparisons of condition within each stimulation site showed that under rIFG stimulation, connectivity was significantly reduced in the painful condition relative to the non-painful condition (difference: -0.25 ± 0.41, t14 = −2.38, p = 0.032, Cohen’s d = −0.81, 95% CI: −0.48, −0.02, Supplementary Fig. 5f), whereas under Vertex stimulation, the difference between the painful and non-painful conditions was not significant (p = 0.079). Comparisons of stimulation site within each condition indicated that in the painful condition, connectivity under rIFG stimulation was significantly lower than under Vertex stimulation (difference: -0.50 ± 0.61, t14 = −3.17, p = 0.007, Cohen’s d = −1.07, 95% CI: -0.84, -0.16, Supplementary Fig. 5f), whereas in the non-painful condition, connectivity did not differ significantly between rIFG and Vertex stimulation (p = 0.424). One-sample t-tests were performed to compare the connectivity index against a baseline of 1 for all conditions. The results indicated that only in the rIFG-painful condition was the connectivity index significantly lower than 1 (difference: -0.20 ± 0.23, t14 = −3.45, p = 0.004, Cohen’s d = −0.89, 95% CI: −0.33, −0.08). For all other experimental conditions, the connectivity indices did not deviate significantly from 1 (all p > 0.05). These results indicated that the effect of condition was moderated by stimulation site: painful versus non-painful condition differed under rIFG stimulation but not under Vertex stimulation. Furthermore, the difference between stimulation sites emerged only in the painful condition. These findings support the site specificity of the rIFG-conditioned modulation observed in Experiment 3. Bayesian network During the structure-learning phase, we applied a score-based hill-climbing algorithm to infer the Bayesian network topology. The learned structure (Fig. 5a) identified Pain as the sole parent of both Unpleasantness and Excitability, with directed edges Pain → Unpleasantness and Pain → Excitability. Here, P(Pain) denotes the marginal distribution of the perceived pain intensity; P(Unpleasantness | Pain) denotes the conditional distribution of the perceived unpleasantness given perceived pain intensity; and P(Excitability | Pain) denotes the conditional distribution of CSE given perceived pain intensity. Furthermore, a supplementary Bayesian network analysis replacing cortical excitability with the IFG–M1 connectivity yielded a structurally identical topology (Supplementary Fig. 6 and Supplementary Table 1).Fig. 5Structure learning, moral graph, and probability distributions of the Bayesian network model.a Structure learning result of the BN showing directed edges from Pain to Excitability and Unpleasantness, indicating that the perceived pain intensity predicts both motor cortex excitability and perceived unpleasantness during pain empathy. b Moral graph of the BN, in which undirected edges depict the Markov blanket relationships among variables, highlighting the direct connections between Pain, Excitability, and Unpleasantness; this undirected representation was used for subsequent probabilistic inference in the pain empathy context. c Probability distribution of each node (Pain, Unpleasantness, and Excitability) in the BN. n = 80 participant-level observations. Light blue bars represent the actual experimental data, and dark blue bars represent the model-predicted probabilities. The close correspondence between empirical and model-calculated distributions supports the adequacy of the BN structure in capturing the probabilistic relationships among variables in pain empathy. Source data are provided as Supplementary Data 1. Structure learning, moral graph, and probability distributions of the Bayesian network model. a Structure learning result of the BN showing directed edges from Pain to Excitability and Unpleasantness, indicating that the perceived pain intensity predicts both motor cortex excitability and perceived unpleasantness during pain empathy. b Moral graph of the BN, in which undirected edges depict the Markov blanket relationships among variables, highlighting the direct connections between Pain, Excitability, and Unpleasantness; this undirected representation was used for subsequent probabilistic inference in the pain empathy context. c Probability distribution of each node (Pain, Unpleasantness, and Excitability) in the BN. n = 80 participant-level observations. Light blue bars represent the actual experimental data, and dark blue bars represent the model-predicted probabilities. The close correspondence between empirical and model-calculated distributions supports the adequacy of the BN structure in capturing the probabilistic relationships among variables in pain empathy. Source data are provided as Supplementary Data 1. Bayesian networks (BNs) provide a suitable framework that not only reveals the dependency structure among variables but also enables probabilistic inference to estimate the probability distribution of each node. In this study, to assess the accuracy of our model, we compared the node-wise probability distributions computed by the BN with the corresponding empirical statistics (Fig. 5c). A core capability of BNs is to quantify dependencies among nodes by designating evidence and query variables. We set the Pain parent node as the evidence variable and specified excitability and unpleasantness as query variables to compute their probabilities under different states of the parent node; we then compared these results with the experimental summary statistics (Supplementary Table 2). The validation showed close agreement between the model-predicted distributions and the empirical data, indicating that the BN constructed here successfully captures the probability distributions of the variables. Predictive inference is a procedure for deducing effects from causes. In this study, Unpleasantness and Excitability were treated as outcome variables, whereas Pain was regarded as the causal variable. Predictive inference was used to examine the impact of perceived pain intensity on perceived unpleasantness and cortical excitability. We set Pain (low/medium/high) as the evidence variable and Unpleasantness and Excitability as the query variables, then computed P(Unpleasantness | Pain) and P(Excitability | Pain) to characterize the probability distribution of Unpleasantness and Excitability across levels of perceived pain intensity. The results showed a monotonic, level-dependent increase in perceived unpleasantness with escalating perceived pain intensity (Fig. 6a): under low pain, low unpleasantness predominated (88.6%); under medium pain, the probability mass shifted toward medium unpleasantness (42.2%); under high pain, high unpleasantness was dominant (68.8%).In contrast, excitability exhibited a decreasing, level-dependent pattern (Fig. 6b): under low pain, high excitability predominated (56.9%); under medium pain, medium excitability became the most probable state (46.0%); under high pain, low excitability was most frequent (65.0%). These predictive inferences suggest that increases in perceived pain intensity may function as a unifying driving signal that monotonically amplifies the observer’s perceived unpleasantness while concurrently suppressing cortical excitability in the M1.Fig. 6Results of predictive and diagnostic inference in the Bayesian network.a Predictive inference, P(Unpleasantness | Pain): posterior mass shifts from low to medium and high perceived unpleasantness as perceived pain intensity increases. b Predictive inference, P(Excitability | Pain): posterior mass shifts from high to medium and low excitability as perceived pain intensity increases. c Diagnostic inference, P(Pain | Unpleasantness): lower perceived unpleasantness is associated with a higher probability of low perceived pain intensity, whereas higher perceived unpleasantness concentrates the probability on high perceived pain intensity. d Diagnostic inference, P(Pain | Excitability): lower CSE is associated with a higher probability of high perceived pain intensity, whereas higher CSE favors low perceived pain intensity. e Joint diagnostic inference, P(Pain | Unpleasantness, Excitability): the combination of high perceived unpleasantness with low CSE concentrates probability on high perceived pain intensity, whereas low perceived unpleasantness with high CSE concentrates probability on low perceived pain intensity; intermediate combinations yield intermediate probability distributions. Source data are provided as Supplementary Data 1. Results of predictive and diagnostic inference in the Bayesian network. a Predictive inference, P(Unpleasantness | Pain): posterior mass shifts from low to medium and high perceived unpleasantness as perceived pain intensity increases. b Predictive inference, P(Excitability | Pain): posterior mass shifts from high to medium and low excitability as perceived pain intensity increases. c Diagnostic inference, P(Pain | Unpleasantness): lower perceived unpleasantness is associated with a higher probability of low perceived pain intensity, whereas higher perceived unpleasantness concentrates the probability on high perceived pain intensity. d Diagnostic inference, P(Pain | Excitability): lower CSE is associated with a higher probability of high perceived pain intensity, whereas higher CSE favors low perceived pain intensity. e Joint diagnostic inference, P(Pain | Unpleasantness, Excitability): the combination of high perceived unpleasantness with low CSE concentrates probability on high perceived pain intensity, whereas low perceived unpleasantness with high CSE concentrates probability on low perceived pain intensity; intermediate combinations yield intermediate probability distributions. Source data are provided as Supplementary Data 1. Diagnostic inference refers to identifying the potential causes of an observed outcome and computing their posterior probabilities. As a complement to the preceding forward (cause-to-effect) predictions, we conducted diagnostic inference by integrating evidence from unpleasantness and cortical excitability to estimate the posterior distribution of perceived pain intensity. Setting Unpleasantness as the evidence node (Low/Medium/High) and Pain as the query variable, we computed P(Pain | Unpleasantness). The posterior distribution exhibited a clear rank-consistent pattern (Fig. 6c): when Unpleasantness = Low, P(Pain = Low | Unpleasantness) = 0.691, exceeding Medium (0.223) and High (0.086); when Unpleasantness = Medium, P(Pain = Medium | Unpleasantness) = 0.606 was maximal; when Unpleasantness = High, P(Pain = High ∣ Unpleasantness) = 0.688, exceeding Medium (0.270) and Low (0.042). Setting Excitability as the evidence node (Low/Medium/High) and Pain as the query variable, we computed P(Pain | Excitability). The posterior distribution showed an inverse rank-consistent pattern (Fig. 6d): when Excitability = Low, P(Pain = High | Excitability) = 0.626 was highest, exceeding Medium (0.150) and Low (0.224); when Excitability = Medium, P(Pain = Medium | Excitability) = 0.460 became highest, exceeding High (0.308) and Low (0.232); when Excitability = High, the distribution shifted toward low pain, with P(Pain = Low | Excitability) = 0.589, exceeding Medium (0.370) and High (0.041). In the dual-evidence diagnostic inference, we simultaneously set Unpleasantness (Low/Medium/High) and Excitability (Low/Medium/High) as evidence nodes and treated perceived pain intensity as the query variable, computing P(Pain | Unpleasantness, Excitability). The results exhibited a clear crossed gradient (Fig. 6e): when Unpleasantness = Low, low pain predominated across Excitability levels, dipping slightly at medium excitability before peaking at high excitability (0.622, 0.536, and 0.815 for Excitability = Low/Medium/High, respectively); when Unpleasantness = Medium, the inference shifted dynamically: high pain was most probable at low Excitability (0.603), whereas medium pain became the dominant mode at medium (0.716) and high (0.751) Excitability; Finally, when Unpleasantness = High, high pain strongly predominated but progressively weakened as Excitability increased (0.897, 0.614, and 0.185 for Excitability = Low/Medium/High, respectively). Notably, at Excitability = High, the distribution completely reversed, with medium pain becoming the dominant inference (0.662). These Bayesian network inferences reveal a level-dependent modulatory pattern of the perceived unpleasantness and CSE with respect to perceived pain intensity: as perceived pain intensity increases, the individual’s unpleasantness correspondingly intensifies, whereas CSE decreases markedly; conversely, weaker perceived pain intensity is characterized by lower unpleasantness and higher CSE. Discussion The spTMS experiment showed that M1 CSE was selectively reduced when participants viewed painful events from a first-person perspective, and this suppression scaled with perceived pain intensity. A control experiment confirmed that this effect was not attributable to the visual salience or aversiveness of the noxious instrument. The ppTMS experiment found no reliable pain-related changes in SICI or ICF, suggesting that the observed CSE suppression was not driven by detectable modulation of local M1 inhibitory or excitatory circuits. In contrast, dsTMS revealed pain-specific inhibitory connectivity from rIFG to left M1 under the first-person perspective; this effect covaried with both perceived pain intensity and CSE suppression and was absent at the Vertex control site. Bayesian network modeling further identified perceived pain intensity as a central predictor of both unpleasantness and motor excitability. Together, these findings suggest that empathic motor inhibition is closely linked to pain appraisal and may involve top-down IFG–M1 modulation of motor cortical excitability. The results of spTMS in Experiment 1 suggest that motor resonance during pain empathy is modulated by visual perspective and is closely associated with the observer’s appraisal of pain intensity. Furthermore, our control experiment (Experiment 4) demonstrated that applying identical noxious visual stimuli to a non-biological object (an apple) yielded no motor modulation, thereby ruling out the possibility that the observed CSE reduction merely reflects a generalized response driven by the visual salience or aversiveness of the stimuli. Prior work suggests that pain empathy involves somatomotor representations of others’ pain6,7,24,27,28, and that the accompanying reduction in CSE is commonly understood as arising from a bottom-up, automatic simulation of observed painful stimulus attributes3,12,13,35–38. However, this pattern is difficult to explain by low-level stimulus features alone. The visual nociceptive features were matched across the first-person and third-person perspectives, yet absolute CSE suppression was observed selectively in the first-person painful condition and was correlated with perceived pain intensity. First-person viewing may enhance body ownership, self–other sensorimotor alignment, and motor resonance, thereby providing a condition under which pain-specific corticospinal suppression is more likely to emerge. This view is consistent with previous work showing that pain-related MEP suppression can emerge selectively under a first-person perspective and may reflect ownership-related mechanisms rather than empathy alone6. The present findings extend this interpretation by showing that, within this first-person embodied context, the magnitude of CSE suppression covaried with the observer’s perceived pain intensity. Thus, while enhanced embodiment or ownership may contribute to the emergence of the CSE effect, the strength of this motor inhibition also appears to be linked to pain appraisal. Accordingly, our data suggest that empathic motor inhibition is unlikely to be fully accounted for by stimulus-driven embodiment alone. Instead, the present pattern is more consistent with a model in which embodiment-related sensorimotor alignment and appraisal-related modulatory processes jointly shape motor cortical responses during pain empathy. Nevertheless, because Experiment 1 with spTMS only revealed overall CSE changes, it cannot establish whether the observed association reflects bottom-up mirroring within M1 or top-down modulation associated with cognitive appraisal. This question was therefore further examined in Experiments 2 and 3 using ppTMS and dsTMS and interpreted in conjunction with the Bayesian network analysis. During self-experienced pain, nociceptive afferents ascend via the spinothalamic pathway to the cerebral cortex and suppress activity in muscles associated with the noxious stimulus15–19. This process has a characteristic M1 electrophysiological signature, with actual nociception associated with enhanced SICI and reduced ICF19. According to the PAM, pain empathy might therefore be expected to involve a partially isomorphic bottom-up simulation of first-hand pain, accompanied by similar changes in local M1 intracortical excitability. However, our ppTMS experiment showed no significant modulation of SICI or ICF, and Bayesian analyses provided moderate evidence for the null hypothesis. Although this does not rule out all local M1 involvement, it indicates that pain-empathy-related CSE suppression was not accompanied by detectable changes in the specific intracortical mechanisms indexed by SICI and ICF. Given that first-hand pain typically enhances SICI and reduces ICF19, the absence of such modulation suggests that empathic motor suppression was not clearly mediated by these local paired-pulse mechanisms. This interpretation is consistent with evidence that spinal dorsal horn processing of one’s own nociceptive signals is suppressed, rather than activated, when observing another’s pain39, potentially attenuating ascending modulation of cortical circuits. Converging evidence further suggests that pain empathy may depend more strongly on higher-order cortical systems than on lower-level nociceptive-spinal processing40. Thus, downstream somatomotor responses during pain empathy are unlikely to be driven solely by bottom-up isomorphic simulation; instead, CSE reduction may also reflect top-down modulation from distal cortical regions that influence motor representations and inhibitory control over M141–43. Convergent neuroimaging evidence has shown robust IFG engagement when observers decode pain from bodily cues22,44–47, and cognitive reappraisal further augments rIFG activation and its coupling with visual cortex, insula, and hippocampus, thereby reducing empathic bias and improving the accuracy of pain empathy judgments48. Causal evidence aligns with this role: perturbing rIFG with cathodal transcranial direct-current stimulation or repetitive TMS lowers cortical excitability and impairs cognitive empathy, and transient “virtual lesions” likewise disrupt pain-empathy judgments23,24. Collectively, these convergent findings implicate IFG as an important node for the cognitive appraisal of others’ pain21. In this context, our dsTMS experiment revealed a pain-specific reduction in the IFG–M1 connectivity index at the 12-ms interval. Moreover, the strength of this inhibitory IFG–M1 connectivity covaried with both greater CSE reduction and higher perceived pain intensity. A complementary site-control experiment further showed that connectivity was significantly lower during rIFG than Vertex stimulation in the painful condition, whereas no site difference emerged in the non-painful condition. This pattern supports the site specificity of the modulation observed in Experiment 3. Taken together, these findings suggest that the observed dsTMS effect may reflect one possible pathway through which pain-related appraisal is associated with changes in motor cortical excitability, although they do not by themselves establish the precise anatomical route of this influence. The temporal specificity of this effect may also be informative. Diffusion-tensor imaging has identified a direct white-matter pathway between the IFG and the subthalamic nucleus (STN)49, and converging evidence suggests that IFG–STN projections can influence basal ganglia–thalamocortical motor output, thereby modulating motor cortical excitability50–52. In addition, previous diffusion-weighted imaging and dsTMS studies indicate that IFG–M1 interactions may occur through direct cortico-cortical pathways at short ISI (6 ms), whereas at longer ISI (12 ms) it could additionally involve subcortical relays via the STN53,54. From this perspective, the inhibitory modulation observed selectively at the 12-ms ISI may be compatible with the engagement of a longer-latency frontal–subcortical–motor route; however, our data cannot directly test this pathway, and future work combining TMS with multimodal imaging will be necessary to clarify the contribution of subcortical circuits to motor excitability during pain empathy. Neurophysiological evidence indicates that pain empathy recruits partially overlapping networks. The anterior cingulate cortex (ACC) supports shared affective representations of self- and other-pain55–57; the amygdala contributes to recognizing others’ emotional states and negative affect58–60, and the anterior insula (AI) integrates affective signals through low-frequency coupling with the ACC and amygdala61, while also encoding perceived pain intensity via its own low-frequency oscillations62, similar to the IFG’s role in extracting pain-related information21,63,64. Intracranial electroencephalography (EEG) studies further suggest that perceived pain intensity is reflected in early IFG high-gamma activity and β-band phase–amplitude coupling with the AI and ACC, whereas later β-band coherence among the ACC, AI, and amygdala, together with amygdala β-power changes, may represent unpleasantness during pain empathy61,65. Early IFG high-gamma responses may also interact with basal ganglia–motor networks and influence M1 excitability66–68. Consistent with this evidence, our Bayesian network (BN) identified perceived pain intensity as a central predictor of both unpleasant affect and motor-cortex excitability. These findings suggest that intensity-related IFG/AI activity may contribute to downstream affective representations in the ACC/amygdala and to motor modulation via cortico–basal ganglia pathways, potentially explaining the association between perceived pain intensity and M1 excitability. Cortical–cortical paired associative stimulation (ccPAS) exploits spike-timing–dependent plasticity to bias interactions between distinct brain regions69,70. Building on our observation of a pain-specific increase in inhibitory connectivity from rIFG to M1, ccPAS provides an effective means to manipulate this linkage. Recent work shows that ccPAS targeting the IFG-STN-M1 pathway can alter balance control, indicating that elements of this circuit are amenable to plasticity-inducing protocols53. Accordingly, state-dependent ccPAS applied over IFG-M1 during pain observation tasks could test whether biasing IFG–M1 coupling modulates CSE and behavioral indices of pain empathy. Such experiments—ideally combined with TMS-EEG — could offer stronger causal support for a prefrontal contribution to empathic motor inhibition and, pending replication and safety considerations, may eventually inform approaches for conditions marked by empathy-related disturbances. Finally, several limitations should be noted. First, consistent with previous TMS studies on pain empathy, we used a blocked design rather than a fully randomized trial-by-trial design to reduce inter-trial fluctuations and maintain a stable task set3,8,71,72. Although block order was randomized across participants, blocked presentation may still induce expectancy or habituation effects within blocks, potentially affecting the magnitude of corticospinal modulation. Future event-related studies with trial-by-trial randomization are needed to further test the robustness of these findings. Second, each condition in Experiments 1–3 included 12 trials. Although this is consistent with prior TMS work and acceptable for MEP-based analyses, it remains relatively limited given the intrinsic variability of MEP amplitudes. A larger number of trials per condition would improve the stability of condition-level estimates and the sensitivity to subtle effects. Third, the null effects for SICI and ICF should be interpreted cautiously. The trial-to-trial variability and limited sensitivity of paired-pulse indices, together with the modest sample size, may have reduced the ability to detect subtle intracortical changes. Moreover, the 3 and 12 ms interstimulus intervals used here do not cover the full temporal range or conditioning-intensity dependence of SICI and ICF. Thus, although our data do not support a dominant role for the specific pathways indexed by SICI and ICF, they cannot rule out local M1 involvement. Future studies should map additional interstimulus intervals and stimulus intensities, ideally combined with TMS–EEG to assess circuit-level responses more directly. In sum, our findings suggest that, during pain empathy, perceived pain intensity may be a key predictor of variation in unpleasantness and motor cortex excitability; moreover, the pain-specific inhibitory IFG–M1 connectivity covaried with higher intensity ratings and lower CSE, suggesting an IFG-related top-down modulatory contribution to empathic motor inhibition. Methods Ethical approval was obtained from the Research Ethics Committee of Shanghai University of Sport before the commencement of the study (Approval No. 102772024RT115). All ethical regulations relevant to human research participants were followed. Participants One hundred and sixteen healthy university students took part in this study. Twenty-nine participants (17 women; mean age 21.10 ± 2.60 years) completed Experiment 1; twenty-eight participants (13 women; mean age 23.50 ± 2.50 years) completed Experiment 2; twenty-three participants (11 women; mean age 22.74 ± 2.60 years) completed Experiment 3; twenty-one participants (12 women; mean age 21.81 ± 1.80 years) completed Experiment 4; and fifteen participants (9 women; mean age 23.00 ± 2.53 years) completed Experiment 5. All participants were right-handed and had normal or corrected-to-normal vision. Each participant provided written informed consent in accordance with the Declaration of Helsinki. The TMS protocols were reviewed and approved by the local ethics committee. Materials This study adopted the stimulus preparation procedures of Avenanti et al3. to create two sets of 24 visual materials. For the hand stimuli, a set of 12 original materials was generated: four painful videos, four non-painful videos, and four static hand dorsum images. In the painful videos, the right FDI muscle was shown being pierced by a needle; in the non-painful videos, the same muscle was touched with a cotton swab. The static images portrayed the right hand dorsum at rest. For the object control stimuli, an apple was used in place of the human hand, yielding a corresponding set of 12 materials: four needle–apple videos (pierced by a needle), four Q-tip–apple videos (touched with a cotton swab), and four static apple images. All videos or images lasted 1.8 s, and all materials depicted the target object (hand or apple) resting on a white tabletop. Furthermore, all original videos and images were initially produced from a first-person perspective. To generate the matched third-person perspective stimuli, the corresponding first-person materials were rotated by 180°. Experimental paradigm The experiment employed a video-based induction task, programmed in MATLAB using the Psychophysics Toolbox to control stimulus sequence and duration and to synchronize TMS delivery with electromyographic (EMG) recording. Each trial proceeded as follows (Fig. 1a): a fixation cross (“+”) was displayed for 1000 ms, followed by an 1800 ms stimulus display consisting of either a video clip or a static image, and then a 7200 ms blank screen. All visual stimuli appeared centrally on the monitor; the initial fixation cross served to focus participants’ attention. To minimize any startle-related modulation of MEP amplitude, TMS was triggered at a random time between 200 and 600 ms before the end of the stimulus presentation3. Across all experiments, conditions were presented in separate blocks, and trial composition varied by experiment3,26,71. In Experiment 1, each condition included 12 spTMS trials. In Experiment 2, each condition included 12 spTMS trials, 12 SICI trials, and 12 ICF trials. In Experiment 3, each condition included 12 spTMS trials, 12 dsTMS trials at the 6-ms ISI, and 12 dsTMS trials at the 12-ms ISI. In Experiment 4, each condition included 16 spTMS trials. In Experiment 5, each condition included 16 spTMS trials and 16 dsTMS trials at the 12-ms ISI for each stimulation site (Vertex and rIFG). To control for potential order effects, the sequence of blocks was randomized across participants. Throughout all observation conditions, participants were instructed to attend closely to each stimulus and to attempt to understand the sensations experienced by the target depicted. Detailed procedures for outcome quantification and statistical analysis are described in the Statistics and Reproducibility section. Experiment 1: changes in CSE of M1 during pain empathy This experiment used a 2 × 2 within-subjects design with factors Condition (painful, non-painful) and Perspective (first-person, third-person). The dependent variables were MEPs, perceived pain intensity, and unpleasantness. Each of the experimental conditions comprised 12 trials. Twenty-nine right-handed university students from Shanghai University of Sport (normal or corrected-to-normal vision) were ultimately recruited to participate in the experiment. Motor-evoked potential (MEP) amplitude served as the primary index of CSE in M1. “1 mV intensity” was defined as the percentage of maximal stimulator output required to elicit MEPs exceeding 1 mV on at least five out of ten consecutive TMS pulses applied to the relaxed FDI muscle. This 1 mV intensity was then used to deliver single-pulse TMS over the motor cortex, with subsequent changes in MEP amplitude reflecting modulations in CSE. Experiment 2: intracortical circuits within M1 during pain empathy This experiment used a within-subjects single-factor design with the factor Condition (painful, non-painful), conducted under the first-person perspective to ensure robust empathic responses. The dependent variables were SICI, ICF, perceived pain intensity and unpleasantness. For each experimental condition, 12 trials were administered for each stimulation protocol (spTMS, SICI, and ICF). Twenty-eight right-handed university students from Shanghai University of Sport (normal or corrected-to-normal vision) were recruited. Intracortical inhibitory and excitatory function in left M1 was assessed with paired-pulse TMS: Resting motor threshold (RMT) was defined as the minimum stimulator intensity (expressed as a percentage of maximum output) required to evoke MEPs ≥ 50 μV in the relaxed FDI muscle on at least 5 of 10 consecutive trials (mean ± SD: 54.0 ± 8.7% of maximum stimulator output). SICI used a conditioning stimulus (CS) at 80% RMT and a test stimulus (TS) at the 1 mV intensity with a 3-ms interstimulus interval (ISI). ICF used the same CS/TS with a 12-ms ISI. These ISIs were selected in accordance with established protocols, whereby a 3-ms interval reliably isolates MEP inhibition reflecting the activation of cortical inhibitory interneurons72,73, and a 12-ms interval produces MEP facilitation reflecting excitatory cortical circuits, without affecting spinal excitability19,73–76. Experiment 3: changes in intercortical connectivity between the right IFG and left M1 during pain empathy This experiment used a 2 × 2 within-subjects design with factors Condition (painful, non-painful) and ISI (6 ms, 12 ms), conducted under the first-person perspective to ensure robust empathic responses. Each of the experimental conditions comprised 12 trials. Twenty-three right-handed university students from Shanghai University of Sport (normal or corrected-to-normal vision) were ultimately recruited for this experiment. The mean RMT across participants in this experiment was 52.6 ± 8.9% of the maximum stimulator output. The dsTMS quantified connectivity from the right IFG to the left M1 using two 5-cm figure-of-eight coils, each driven by a separate single-pulse stimulator (Magstim 200, Whitland, Dyfed, UK). On each paired-pulse trial, a conditioning stimulus (CS; right IFG; MNI coordinates: 48, 36, 6; 90% RMT; frameless stereotactic neuronavigation; with the coil held tangential to the scalp and oriented at 90° to the sagittal plane) preceded a test stimulus (TS; left M1 hotspot for right FDI, “1 mV” intensity; coil handle ~30°–45° to midline, posterior-to-anterior current). ISIs of 6 and 12 ms probed timing-specific influences of rIFG on M1. The ISIs of 6 and 12 ms were selected based on previous dsTMS and diffusion-weighted imaging studies showing that rIFG–M1 interactions may occur at two distinct temporal scales54. Short-latency effects at 6 ms have been linked to relatively direct cortico-cortical pathways, whereas longer-latency effects at 12 ms may additionally involve subcortical relays, particularly pathways passing near the STN. We therefore included both ISIs to examine whether pain-related IFG–M1 modulation would emerge at a short-latency cortical timescale, a longer-latency pathway potentially involving subcortical circuitry, or both. Experiment 4: control for biological specificity of CSE This experiment followed a 2 × 2 within-subjects design with the factors Condition (needle-apple, Q-tip-apple) and Perspective (first-person, third-person). The dependent variables were MEPs. The primary objective was to determine whether the CSE modulation observed in Experiment 1 was contingent upon the biological nature of the stimuli. Twenty-one right-handed university students with normal or corrected-to-normal vision participated in this experiment. The experimental paradigm and TMS protocols were identical to Experiment 1, with the following two modifications: (1) the human hand stimuli were replaced with a non-biological object (an apple), and (2) the number of trials for each experimental condition was increased to 16 trials to enhance statistical power. Experiment 5: site-control test of rIFG-conditioned modulation of M1 This experiment followed a 2 × 2 within-subjects design with the factors Condition (painful, non-painful) and Stimulation Site (rIFG, Vertex). The dependent variable was the standardized connectivity index. The mean RMT across participants in this experiment was 51.8 ± 12.5% of the maximum stimulator output. The primary objective was to test whether the modulation observed in Experiment 3 was specific to conditioning over the rIFG rather than reflecting nonspecific dual-site stimulation effects. Fifteen right-handed university students with normal or corrected-to-normal vision participated in this experiment. The experimental paradigm and TMS procedures were identical to those used in Experiment 3, with the following modifications: (1) only the 12-ms interstimulus interval was tested; (2) the number of trials for each experimental condition was increased to 16 trials to enhance statistical power; and (3) the conditioning stimulus was delivered either over the rIFG (MNI: 48, 36, 6) or over the control site Vertex (MNI: 0, 0, 75). Vertex was chosen as the control site because previous studies have shown that stimulation over Vertex does not significantly affect cortical activity77–81. Frameless stereotactic neuronavigation system For Experiments 3 and 5, we used a frameless stereotactic neuronavigation system (eemagine Visor2; Berlin, Germany) to localize both the rIFG (MNI coordinates: 48, 36, 6) and the Vertex (MNI coordinates: 0, 0, 75) in each participant23. First, high-resolution T1-weighted anatomical images were acquired on a 3.0 T magnetic resonance imaging (MRI) scanner (Siemens Magnetom Trio 3.0 T). During acquisition, participants’ heads were stabilized with standard foam padding to minimize motion artifacts; they were instructed to remain as still as possible and wore noise-reducing earplugs throughout. Imaging parameters for the magnetization-prepared rapid gradient-echo (MPRAGE) sequence were: repetition time (TR) = 2300 ms; echo time (TE) = 3.1 ms; echo spacing = 7.6 ms; flip angle = 7°; slice thickness = 0.8 mm; number of slices = 240; field of view (FOV) = 256 × 256 mm; voxel size = 0.8 × 0.8 × 0.8 mm3; parallel imaging was performed using generalized autocalibrating partially parallel acquisitions (GRAPPA) to shorten scan time. Next, individual T1 images were processed in Linux (Ubuntu 20.04.6) using FSL 6.0.7.14 to map the target coordinates onto the MNI standard brain, creating a single-point template for visualization. These coordinates were then imported into the frameless neuronavigation system to generate a subject-specific cortical model and visualize the rIFG and Vertex location on the reconstructed cortical surface. Finally, using cortical-surface–based navigation, we identified the optimal coil placement over each participant’s rIFG and Vertex and marked this location on the scalp with a skin-safe marker. During the experiment, the conditioning-stimulus coil was positioned over either the rIFG or the Vertex according to the stimulation-site condition and remained fixed at the marked location throughout each block. Electromyography recording Crescent-shaped Ag/AgCl surface electrodes (9 mm diameter) were used to record EMG activity from the right FDI muscle. During recording, the active electrode was placed over the muscle belly, the reference electrode over the metacarpophalangeal joint of the index finger, and the ground electrode on the dorsum of the right hand. EMG signals were band-pass filtered (20 Hz–2.5 kHz) and amplified 1000-fold using an Intronix Technologies Model 2024 F physiological signal amplifier. The conditioned signals were digitized at 1 kHz via a CED Micro1401 neural data acquisition interface and Signal 6.0 software, then stored for offline analysis. Questionnaire data collection Trait empathy was assessed using the IRI, a 22-item self-report questionnaire. The IRI comprises four subscales: Perspective Taking (PT), Fantasy Scale (FS), Empathic Concern (EC), and Personal Distress (PD). PT, FS, and EC reflect cognitive and other-oriented emotional components of empathy, whereas PD corresponds to self-oriented emotional responses. Higher scores on the PT, FS, and EC subscales indicate greater trait empathy82. In addition, participants completed a test session in which they provided two subjective ratings for each video clip: (i) the perceived pain intensity experienced by the target (“How much pain do you think the person experienced in this situation?”) and (ii) the intensity of their own unpleasant emotional state (“How unpleasant did you feel while viewing this stimulus?”). Both ratings were made on visual analog scales (VAS) ranging from 0 (“no sensation of pain or unpleasantness”) to 10 (“the maximum imaginable pain or unpleasantness”). Bayesian network model A Bayesian network provides a graphical representation of the conditional dependency structure among a set of random variables. The model is expressed as a directed acyclic graph (DAG), in which nodes correspond to variables and directed edges indicate probabilistic dependencies between variables. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Let}}}}$$\end{document}Let X \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=$$\end{document}= {X1, X2, …,Xn} denote the set of variables included in the network. The joint probability distribution over these variables can be decomposed into a product of local conditional distributions: P \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=$$\end{document}= {X1, X2, …,Xn} = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\prod }_{{{{\rm{i}}}}=1}}^{n}P({X}_{{{{\rm{i}}}}}|\pi ({X}_{{{{\rm{i}}}}}))$$\end{document}∏i=1nP(Xi∣π(Xi)) where π(Xi) denotes the parent set of Xi in the graph. This factorization allows complex multivariable relationships to be represented through a series of lower-dimensional conditional probability distributions. Before model building, to maximize statistical power and ensure a robust network structure, we pooled the participant-level averaged data from the first-person painful conditions across three experiments (Experiments 1, 2, and 3). This yielded a pooled sample of 80 participant-level observations. We preprocessed this dataset by labeling the CSE as Excitability, the perceived pain intensity as Pain, and the perceived unpleasantness as Unpleasantness, and then discretized each continuous variable into three categorical states (low, medium, and high) via equal-frequency binning83. Bayesian-network construction proceeds in two stages: structure learning and parameter learning84. Given a data set, structure learning identifies the DAG that best explains the observed dependencies. Once the topology is fixed, parameter learning estimates each node’s conditional-probability distribution. In this study, we used a score-based hill-climbing algorithm for structure learning and Bayesian estimation for parameter learning. To enhance robustness under a modest sample and reduce model-selection instability, we embedded hill-climbing in a bootstrap model-averaging scheme85,86 (R = 1000 resamples; m = n), computed arc strength and direction confidence, and retained only edges exceeding a conservative inclusion threshold ( ≥ 0.85) to form a consensus DAG. This procedure mitigates small-sample noise and idiosyncrasies of any single run, curbs spurious dependencies, and yields a structurally more stable and generalizable network. Each node’s prior distribution was specified as a Dirichlet distribution with uniform hyperparameters—i.e., equal weight for all states—reflecting an unbiased prior assumption and minimizing reliance on subjective priors87. For predictive inference and diagnostic inference, we converted the learned network into a junction tree and applied the clique-tree propagation algorithm (CTP) of Kim and Pearl to perform efficient belief-updating and compute posterior probabilities given evidence88. The moral graph of BN structured based on the CTP was shown in Fig. 5b. Statistics and reproducibility In the present study, outcome measures comprised both behavioral and physiological indices. The behavioral indices included scores on the IRI and participants’ ratings of perceived pain intensity and unpleasantness in response to video stimuli, while the physiological index was the peak-to-peak amplitude of MEPs elicited during video observation. IRI scores were computed according to the instrument’s scoring guidelines, with values from 0 to 4 corresponding to the five response options—‘Not at all characteristic,’ ‘Slightly characteristic,’ ‘Moderately characteristic,’ ‘Characteristic,’ and ‘Highly characteristic’—and with items 2, 5, 10, 11, and 14 reverse-scored. For perceived pain intensity and unpleasantness, we first computed each participant’s mean rating within each video category. To verify that the selected stimuli elicited empathic responses and to examine the effect of viewing perspective, ratings were analyzed using a two-way repeated-measures ANOVA with condition (painful vs. non-painful) and perspective (first-person vs. third-person) as within-subject factors. When a significant interaction emerged, Tukey-adjusted simple-effects comparisons were performed. Because each factor had two levels, sphericity correction was not required. MEP data were preprocessed as follows. Peak-to-peak MEP amplitude—defined as the voltage difference between the maximum positive and negative deflections occurring 20–40 ms after each TMS pulse—was extracted using Signal 6.02 software. To prevent the confounding effects of background muscle activation on motor excitability89,90, MEPs preceded by background EMG activity deviating from the individual mean of the block by more than 2 standard deviations (SDs) were discarded from the analysis; moreover, trials deviating from the mean amplitude of their test block by more than 2 SDs were also discarded. The detailed percentages of excluded trials for each experimental condition across all TMS experiments are presented in Supplementary Tables 3–7. To quantify M1 excitability, we pooled MEPs elicited by spTMS in Experiments 1 and 4. For each participant, task-related excitability was normalized by dividing the mean MEP amplitude measured during each task condition by the mean MEP amplitude recorded during the corresponding baseline condition, yielding a CSE index. We then examined the effects on these normalized MEP amplitudes (CSE indices) using separate two-way repeated-measures ANOVAs for each experiment. The within-subjects factors were Condition (painful vs. non-painful for Experiment 1; needle–apple vs. Q-tip–apple for Experiment 4) and Perspective (first-person vs. third-person). When a significant interaction emerged, Tukey-adjusted simple-effects comparisons were performed. Furthermore, to quantify the strength of evidence for non-significant results in Experiment 4 (i.e., the absence of motor resonance), Bayes factors (BF01) were additionally calculated. Because each factor had two levels, sphericity correction was not required. For ppTMS data indexing SICI and ICF, task-related ppTMS-evoked MEP amplitudes were normalized by dividing them by the mean spTMS amplitude obtained under the same task condition (mean MEPppTMS / mean MEPspTMS). To obtain standardized indices, each participant’s SICI and ICF values in the painful and non-painful conditions were then normalized to their corresponding SICI and ICF values recorded under the baseline condition. For the standardized SICI and ICF indices, we conducted two-tailed paired samples t-tests comparing the painful and non-painful conditions. Effect sizes (Cohen’s d) and 95% confidence intervals were reported. In addition to conventional null-hypothesis significance testing, Bayesian analyses were conducted for the normalized SICI and ICF comparisons to quantify the relative evidence for the null versus alternative hypothesis. For the dual-site paired-pulse TMS protocols assessing connectivity between the right IFG (MNI coordinates: 48, 36, 6; Fig. 4a) and the left M1, and in Experiment 5 also between the Vertex control site (MNI coordinates: 0, 0, 75) and the left M1, we computed a connectivity index as the ratio of the task-evoked MEP amplitude of dsTMS to the corresponding MEP amplitude of spTMS (mean MEPdsTMS / mean MEPspTMS). Subsequently, for each participant, the connectivity values obtained under each experimental condition were normalized to the connectivity values recorded under the corresponding baseline condition, yielding standardized connectivity indices. In Experiment 3, the effects of Condition (painful, non-painful) and ISI (6 ms, 12 ms) on the standardized connectivity index were tested using a two-way repeated-measures ANOVA. In Experiment 5, the effects of Condition (painful, non-painful) and Stimulation Site (rIFG, Vertex) on the standardized connectivity index were tested using a two-way repeated-measures ANOVA. When a significant interaction emerged, Tukey-adjusted simple-effects comparisons were performed. Because each factor had two levels, sphericity correction was not required. Correlation analyses. Normality of variables entering correlation analyses was first assessed using the Shapiro–Wilk test; Pearson’s correlation was used when normality was satisfied, and Spearman’s rank-order correlation when it was violated. For a priori correlations reported in the Results, we report the correlation coefficient r, two-tailed p values, and 95% CI (from Fisher’s z transformation). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary