Oxytocin enhances the triangular association among behavior, resting‐state, and task‐state functional connectivity

Abstract

Considerable advances in the role of oxytocin (OT) effect on behavior and the brain network have been made, but the effect of OT on the association between inter‐individual differences in functional connectivity (FC) and behavior is elusive. Here, by using a face‐perception task and multiple connectome‐based predictive models, we aimed to (1) determine whether OT could enhance the association among behavioral performance, resting‐state FC (rsFC), and task‐state FC (tsFC) and (2) if so, explore the role of OT in enhancing this triangular association. We found that in the OT group, the prediction performance of using rsFC or tsFC to predict task behavior was higher than that of the PL group. Additionally, the correlation coefficient between rsFC and tsFC was substantially higher in the OT group than in the PL group. The strength of these associations could be partly explained by OT altering the brain's FCs related to social cognition and face perception in both the resting and task states, mainly in brain regions such as the limbic system, prefrontal cortex, temporal poles, and temporoparietal junction. Taken together, these results provide novel evidence and a corresponding mechanism for how neuropeptides cause increased associations among inter‐individual differences across different levels (e.g., behavior and large‐scale brain networks in both resting and task‐state), and may inspire future research on the role of neuropeptides in the cross levels association of both clinical and nonclinical use.

Using a face perception task and multiple connectome‐based predictive (CPM) models, we provide novel evidence and a corresponding mechanism for how neuropeptides cause increased consistency between behavior, resting‐state, and task‐state functional connectivity.

1. INTRODUCTION

Oxytocin (OT), a neuropeptide has been proven to be closely linked to social adaptation and prosocial behaviors (Bartz et al., 2011; Churchland & Winkielman, 2012; Kirsch et al., 2005; Kosfeld et al., 2005; Liu et al., 2019; Ma et al., 2016), as well as social cognition (Baumgartner et al., 2008; Keech et al., 2018; Ross & Young, 2009). Previous evidence indicates that OT may increase attention to the eye region of human faces (Guastella et al., 2008) and improve the ability to infer the mental state of others from social cues of the eye region (Domes et al., 2007). In further studies, a more general enhancement effect of OT on motivation or sensitivity to social cues (Bartz et al., 2010; Hurlemann et al., 2010; Neumann, 2008) manifests in face‐ or emotion‐related tasks. OT can also modulate self‐recognition and face recognition, which have been studied extensively and have yielded interesting results and are hallmarks of human social cognition (Shamay‐Tsoory & Abu‐Akel, 2016; van IJzendoorn & Bakermans‐Kranenburg, 2012). For example, OT boosts recognition memory for faces but not for nonsocial stimuli (Rimmele et al., 2009). Such enhancement was not due to improved recognition memory for faces but to participants' liberal response bias toward familiar faces. This may further support the social salience hypothesis of oxytocin (Bate et al., 2015). These findings, suggesting that OT increases sensitivity to social stimuli (Shamay‐Tsoory & Abu‐Akel, 2016; Tillman et al., 2019), lead to further exploration of OT's broader impact on the neural mechanisms underlying social cognition.

The term “social brain” is commonly used to describe a network or collection of brain regions that consistently show association with sociocognitive tasks. The modulation effect of OT has been widely demonstrated in many social brain regions (Fineberg & Ross, 2017; Frith, 2007), including the amygdala, anterior cingulate cortex (ACC), prefrontal cortex (PFC), insula (Barrett & Satpute, 2013; Stanley & Adolphs, 2013), temporal pole and temporoparietal junction (TPJ) (Zink & Meyer‐Lindenberg, 2012). However, the effects of OT on specific brain areas or functional connectivity are mostly task‐dependent and inconsistent across studies (Bethlehem et al., 2013; Grace et al., 2018; Seeley et al., 2018). For instance, several studies have shown that OT increases amygdala responses to emotional faces or aversive stimuli (Domes et al., 2013; Frijling et al., 2016), whereas others have observed an attenuation effect in this area (Eckstein et al., 2015; Quintana et al., 2016). Additional studies have shown that OT changes functional connectivity (FC) in brain regions that belong to the social brain network (Bethlehem et al., 2013; Jiang et al., 2021; Sripada et al., 2013). For example, OT increases the effective flow from the midline default network, including the posterior cingulate cortex and precuneus, to the salience network, including the ACC and insula (Jiang et al., 2021), as well as the brain connectivity within the frontal network during the resting state (Zheng et al., 2022). OT has also been found to alter brain connectivity strength during different tasks. For instance, OT could enhance the FC between the left amygdala, left anterior insula, and left inferior frontal gyrus in emotional perception and memory tasks (Lou et al., 2004). Therefore, the specific modulatory effects of OT on connectivity among social subnetworks have been well‐documented.

Inconsistent results across studies indicate individual differences in the effect of OT (Bartz et al., 2011; Olff et al., 2013), which exist not only at the behavioral level but also at the brain level (Liu et al., 2019; Strathearn et al., 2009). Establishing the association between inter‐individual differences in different data levels may extend the understanding of the individual differences caused by OT. Moreover, in a previous study with the same data as ours, oxytocin improved the prediction of task‐evoked activation from resting state activity (Wu et al., 2023). Nonetheless, whether the OT administration could enhance the association between the inter‐individual difference in different levels (e.g., behavioral and brain levels) has not been systematically explored. The connectome‐based predictive model (CPM), which predicts the cognitive behavior or mental health phenotypes index based on FCs, has been widely used in brain‐wide association studies (BWAS) (Finn et al., 2015; Sui et al., 2020). It predicts individual variability in behavior or psychiatric symptoms by extracting and summarizing the most relevant features from FC using full cross‐validation (Shen et al., 2017). Many prior studies have demonstrated the robustness of CPM (Dadi et al., 2019; Yoo et al., 2019) in predicting individual differences in fluid intelligence (Tong et al., 2022), attention (Yoo et al., 2018), creative ability (Beaty et al., 2018), cheating behavior (Pang et al., 2022). While the resting state still dominates BWAS, data from tasks have empirically demonstrated benefits (Finn, 2021). For example, several recent studies have shown that FC during narrative movie watching (Finn & Bandettini, 2021) or specific tasks (Gbadeyan et al., 2022; Greene et al., 2018; Ovando‐Tellez et al., 2022) provides additional information for CPM prediction. However, although many brain‐wide associations have been established, few studies have investigated whether this association could be affected by neuropeptides.

Social cognitive tasks, such as face recognition and mental‐state attribution, play a vital role in our ability to navigate social interactions and fulfill social functions (Adolphs, 2009). Face recognition is an intricate process through which we identify and distinguish individual faces. To effectively engage in higher‐order social cognition, it is crucial to differentiate oneself from others, which can be investigated through face perception processing, such as self‐other face recognition task (Platek et al., 2008). For this task, the subject needs to judge whether the morphed face presented is similar to their own. Successful self‐recognition needs knowledge of their own, and the brain regions involved in self‐face recognition largely overlap with the brain regions in the social brain network (Adolphs, 2009). Previous evidence has indicated that many brain regions such as frontal lobe, limbic (cingulate), fusiform gyrus, and near TPJ (Inferior parietal lobule, Precuneus, Superior temporal gyrus) activated in self‐other face recognition (Platek et al., 2008; Sugiura et al., 2000). In these brain regions, frontal, limbic, and TPJ are parts of the social brain network (Frith, 2007). Given the widely demonstrated modulation effect of OT on social behavior and underlying brain activity (Fineberg & Ross, 2017), the use of self‐face recognition tasks could evaluate a part of social cognitive abilities and activate the social brain networks. More importantly, it gives us the opportunity to study whether OT affects the association between the social brain and behavior.

Based on previous progress and shortcomings in OT effects on behavior, FC, and their individual difference, we ask whether and how OT modulates the association between inter‐individual differences in brain and behavior. Besides, although previous studies have found that task states were highly similar to the resting‐state network (Cole et al., 2014, 2016), there is still a lack of research on whether this association will be affected by OT. In the present study, we systematically explored the relationship between a type of social behavior (self‐face recognition), resting‐state FC (rsFC), and task‐state FC (tsFC) and investigated the following questions: (1) whether the association between behavioral performance and rsFC could be enhanced by OT administration; (2) whether the association between behavioral performance and tsFC could be enhanced by OT; and (3) whether OT could enhance the similarity between rsFC and tsFC at the whole‐brain level. For Questions 1–3, we hypothesized that OT could enhance the triangular association between self‐face recognition performance, rsFC, and tsFC. If this hypothesis holds, we then assessed Question 4): How does OT enhance this triangular association? Our previous study revealed that OT did not change the participants' behavioral performance (Y. Wang et al., 2023). Hence, we hypothesized that OT could enhance the triangular association by altering rsFC or tsFC. Since the modulation effect of OT has been widely demonstrated in social brain regions (Fineberg & Ross, 2017; Zink & Meyer‐Lindenberg, 2012), we further hypothesized that the altered FC may mainly be related to the brain regions that are activated in self‐face recognition and overlapped with social brain networks (Frith, 2007), including frontal cortex, limbic, and TPJ.

To investigate the OT effect on associated social brain and behavior, we first used the self‐other face recognition paradigm mentioned earlier with a between‐subject design for task fMRI (Figure 1a). We recorded the behavioral performance index, resting‐state fMRI signals, and task‐state fMRI signals. Combining resting state fMRI and task state could provide valuable insights into the stability of the brain‐behavior association over time and whether it depends on the state (Gratton et al., 2018). To answer Questions 1 and 2, we used CPM to examine whether FC could predict their corresponding behavioral performance and whether OT could enhance the CPM prediction accuracy of FC on behavioral performance. To answer Question 3, we conducted a correlation analysis at the whole‐brain level to evaluate the resting‐task similarity. For Question 4, we used the CPM to separate the FCs in the OT group from those in the placebo (PL) group and to identify which FC features contributed more to group differentiation.

FIGURE 1.

The experimental paradigm, main framework, and behavioral results. (a) The experimental paradigm. Participants need to judge whether the morphed face is similar to their own face. There are four conditions of faces: self‐child condition was the face morphed using the participant's own face and a stranger child's face; faces of the other‐child condition were morphed using an adult stranger's face and a stranger child's face; faces in the self‐adult condition were morphed using the participant's own face and a stranger adult's face; faces in the other‐adult condition were morphed using two adult strangers' faces. (b) The main structure of the present study (c) The behavioral results of the PL and OT groups. From upper to down, we presented the behavioral results of averaged accuracy (Acc mean), and average reaction time (RT mean). There were no significant behavioral differences between groups.

2. METHODS

2.1. Participants

The CONSORT flow chart is shown in Figure S1. Sixty‐six male participants were first recruited via online advertisement. All participants completed a self‐report screening form to evaluate various aspects such as their mental and physical health, medication use, and lifestyle habits. Only those free from significant medical or psychiatric illness, not using any medication, and abstaining from daily alcohol consumption or smoking were included. After excluding 6 participants with medical drug use or neurological disorders, 60 were enrolled in this double‐blind, randomized controlled study. All participants were randomly divided into two groups: the OT group with oxytocin administration and the PL group with placebo administration. All participants were right‐handed with normal or corrected‐to‐normal vision. The experimental protocol was approved by the local ethics committee of Beijing Normal University, and all provided a signed informed consent form before the formal experiment. We strictly follow the guidelines of OT administration (Guastella et al., 2013).

We enforced motion criteria of no more than 3 mm translation or 3 degrees rotation. One participant from the PL group who exceeded these parameters was excluded. This led to a final analytical sample of 59 participants, comprised of 30 in the OT group (age mean ± SD: 22.86 ± 1.57 years) and 29 in the PL group (age mean ± SD: 22.79 ± 2.27 years), without significant age differences (t ₅₇ = 0.134, p = .894). We gathered demographic and psychometric data from all participants, primarily before the administration (see more details in Supplementary Section 1; Tables S1 and S2). It was used to confirm uniformity between the OT and PL groups. After comparing measurements across several questionnaires between the treatment groups, our results showed no significant differences, validating the effectiveness of our randomized design.

2.2. Drug administration

We used a double‐blind placebo‐controlled group design to investigate the effects of a single dose of 24 international unit (IU) intranasal OT (40.08 μg) on FC and corresponding behavioral performance in a face‐perception task. The participants (all males) were randomly assigned to either the OT group (24 IU oxytocin) or PL group (same volume of saline water) in a double‐blind, placebo‐controlled, between‐subjects design. OT was administrated by a double‐blinded experimenter and underwent fMRI scanning 45 min later after spray administration. The choice of 24 IU oxytocin followed the instruction of previous research (Spengler et al., 2017) and was consistent with previous studies (Guastella et al., 2008; Liu et al., 2019). Participants were given explicit instructions on how to use the nasal spray, following the recommended procedures outlined by Guastella et al. (2013). The initial dose was administered under the experimenter's supervision, with participants instructed to pump the spray until a fine mist was visible. To prevent loss of the substance due to gravity, participants were advised to close one nostril, inhale deeply through the nose, and slightly tilt their heads backward during administration. Following the first use of the nasal spray, all participants were monitored onsite for approximately 45 min to assess any immediate reactions or side effects. As for the placebo, participants in this group received the same volume of saline water via intranasal administration as those in the oxytocin group, ensuring consistency in the dosing regimen across both groups. More details about the treatment can also be found in our previous work (Wu et al., 2023; Zheng et al., 2022; Y. Wang et al., 2023).

2.3. Experimental paradigm

The experimental paradigm was a face‐perception task with morphed faces, following previous studies (Alvergne et al., 2007) (see Figure 1a). Specifically, we used a face‐perception task with face stimuli that morphed the photos of an adult or child onto the participant (self) or another stranger (other) (Y. Wang et al., 2023). For each trial, the target‐morphed face was presented for 1.5 s. Participants were then asked to judge whether the face resembled their own face in the following response window within 0.5–2.5 s. The task was divided into two runs for each participant, with each run comprising six blocks. These blocks were randomly presented and included three blocks of adult faces and three blocks of child's faces. Consequently, the task consisted of a total of 12 blocks, with six blocks featuring morphed children's faces and another six blocks featuring morphed adult faces. Each block contained 10 trials, five of which presented a self‐resembling face, and five of which presented other‐resembling faces. This resulted in four distinct types of facial stimuli. The self‐resembling faces were created by morphing the participant's face with a face of a 23‐year‐old adult (self‐adult) or a face of a 1.5‐year‐old child (self‐child). The other‐resembling faces were created by morphing a stranger's face with the face of a 23‐year‐old adult (other‐adult) or a 1.5‐year‐old child (other‐child). All facial expressions were neutral.

We selected the face task for two reasons: Firstly, several previous studies have highlighted that hormonal fluctuations and OT treatment can impact attention and perception of faces, albeit with mixed outcomes and varying individual responses (Dalmaso et al., 2020; Harari‐Dahan & Bernstein, 2014; Kemp & Guastella, 2011; Marsh et al., 2021). Secondly, our implementation of a self‐other facial discrimination task is known to be modulated by OT at the neural level, but its effect on behavior was not significant (Y. Wang et al., 2023).

2.4. MRI acquisition

All MRI data were acquired using a 3.0‐T Siemens Tim Trio scanner equipped with a 12‐channel head coil. First, high‐resolution T1‐weighted images were acquired for each participant (TR = 2.53 s, TE = 3.39 ms, flip = 9°, FOV = 256 mm, 176 sagittal slices, voxel size = 1 × 1 × 1.33 mm³). Then, we collected resting‐state and task‐state fMRI data successively. All fMRI data were collected using an echo‐planar imaging sequence (TR = 2 s, TE = 30 ms, flip = 90°, FOV = 224 mm; 64 × 64 matrix, slice thickness = 3.5 mm, slice gap = 0.7 mm, with 33 slices parallel to the hippocampus and interleaved). For the resting state fMRI, each participant was scanned for a fixed duration of 5 minutes, which corresponds to 150 volumes. For the task fMRI, each run for each participant was scanned for a fixed duration of 4 min, corresponding to 120 volumes. This protocol was consistent for all participants to ensure that the order of the scans did not influence our results.

2.5. fMRI preprocessing

The fMRI data preprocessing was performed using SPM12 (Statistical Parametric Mapping; https://www.fil.ion.ucl.ac.uk/spm/software/spm12 ). The functional image time series were preprocessed to compensate for motion correction, slice‐timing correction, and linear detrending. Thereafter, they were co‐registered to the T1‐weighted anatomical image, normalized to a 3 × 3 × 3 mm³ Montreal Neurological Institute space, and smoothed with an isotropic Gaussian kernel of 6 mm full width at half maximum. Finally, the fMRI data were high‐pass filtered with a cutoff of 0.01 Hz. White matter, cerebrospinal fluid (CSF), global, and six head motion parameters, as well as their squares, derivatives, and squares of derivatives, were regressed (Friston et al., 1996). The resulting residuals were then lowpass filtered with a cutoff of 0.1 Hz. Both resting and task‐state fMRI scans were preprocessed using the same pipeline to ensure consistency across scans.

We conducted independent t‐tests to compare the six head motion parameters (three translation parameters and three rotation parameters) between the two groups in both the resting and the task state. We did not observe significant group differences in head motion parameters in either state (p > .05 for all comparisons; see more details in Tables S3 and S4). These results suggest that head motion did not differ between the groups during the experiment.

2.6. CPM predictor

The main analysis utilized CPM to predict behavioral indices or participant groups based on rsFC or tsFC. The CPM built a bridge between FC and behavior for each group. The workflow of the CPM predictor is shown in Figure 2. First, we extracted the averaged blood‐oxygen‐level‐dependent (BOLD) time series of the 90 brain regions based on the AAL atlas (Tzourio‐Mazoyer et al., 2002) as ROIs. A 90 × 90 FC matrix was obtained by Pearson's correlation between the averaged BOLD time series of each pair of ROIs. We retained only the lower‐triangle matrix (4005 FC features) to remove diagonal and repetitive features for further analysis. The Nilearn toolbox (https://nilearn.github.io/ ) (Abraham et al., 2014) was used to construct the FC matrix.

FIGURE 2.

The workflow of the CPM predictor analysis for the present study. (a) FC matrix construction and feature selection. (b) Model validation with LOOCV. (c) Model performance measurement by Spearman correlation and calculate statistics by Permutation test.

The first part of the CPM predictor uses Pearson's correlation to select significant features from all FC features. We selected the features that had significant positive or negative Pearson's correlation between FC and the behavioral index (Figure 2a). This approach was used to select the features with potential information for prediction and was consistent with previous studies (Beaty et al., 2018; Pang et al., 2022; Shen et al., 2017). Only FCs whose p‐value of Pearson's correlation was lower than the threshold were maintained. By changing the threshold value, different FC quantities could be retained.

Precious studies have not clearly defined the threshold value (Pang et al., 2022; Speer et al., 2022). In our study, we used two different thresholds (p < .05 and p < .01) to select features separately and conduct the next step of analysis to ensure the stability of the results.

We then implemented a support vector machine (SVM) regressor with a linear kernel to predict participants' behavioral index based on their FC and validated it with leave‐one‐out cross‐validation (LOOCV) (see Figure 2b) (Beaty et al., 2018; Dosenbach et al., 2010; Finn et al., 2015). The prediction model was fitted for each validation based on n − 1 (n is the number of participants in each group) participants' selected FCs and their corresponding behavioral index. The model was then tested on the leave‐out participants' data to obtain a predicted behavior value. Thus LOOCV resulted in n behavior index predictions for n participants. We used the Spearman correlation between the predicted and actual behavioral value as the model accuracy measurement. Additional details of CPM predictor parameters are provided in Table S5.

After obtaining accuracy by LOOCV, we examined whether the model performance was significantly higher than chance. We used a permutation test to evaluate the statistical significance of the CPM prediction (Figure 2c). We randomly shuffled the behavioral index values 10,000 times and used LOOCV to obtain the corresponding accuracy for each shuffle. Finally, we sorted the accuracy from 10,000 permutations and counted the position of the true model accuracy to obtain the p‐value.

2.7. rsFC‐tsFC similarity analysis

Building on prior research that highlighted OT's role in enhancing the association between individuals' rsFC and tsFC (Wu et al., 2023), our study replicated this effect using a simplified approach. We conducted correlation analyses for both OT and placebo (PL) groups (Figure 4a), extracting the lower triangle FC matrix for each state and group. We then computed the Pearson correlation between the resting and task states for each FC edge. This yielded pairs of FC edge correlation coefficients, signifying the similarity between rsFCs and tsFCs for each group. A paired t‐test was then used to discern any significant differences in resting‐task FC similarity between the OT and PL groups.

FIGURE 4.

The Pearson correlation between each rsFC and tsFC edge for both groups. (a) The workflow to calculate and compare the rsFC‐tsFC similarity of both groups. (b) The rsFC‐tsFC similarity in both groups. The correlation in the OT group was significantly higher than in the PL group (p < .001). Fisher z‐transformation was performed to FC value before the statistical test.

2.8. CPM classifier

The CPM classifier was used to classify OT/PL groups based on FC patterns and identify FCs with significant differences between the OT and PL groups in both rsFC and tsFC. The CPM classifier followed a workflow similar to that of the CPM predictor but with one classifier across both groups (Figure S3). The same FC extraction approach was used to obtain a lower‐triangle FC matrix between the AAL ROIs based on Pearson's correlation. Feature selection was slightly different from that in the predictor workflow. For each FC in the lower‐triangle matrix, we conducted logistic regressions between each FC value and their group label (OT/PL) to obtain the regression coefficient. Only FCs whose p‐value of the regression coefficient was lower than the threshold were retained. By changing the threshold value, different FC quantities could be retained. In the present study, we used three different thresholds—with p < .05, p < .02, and p < .01 (see Table S6).

Then, we implemented a linear kernel SVM classifier based on the maintained FCs to classify the corresponding groups (OT/PL) using LOOCV (Beaty et al., 2018; Dosenbach et al., 2010; Finn et al., 2015). More specifically, the model was fitted based on n − 1 (n is the total number of participants in the two groups) participants' FC matrixes and corresponding group labels. The fitted model was tested on one left‐out participant. After n cycles, each participant was tested once and eventually achieved accuracy, which was used to quantify the performance of the model. Finally, we used the same permutation procedure but shuffled the group labels rather than the behavior indices.

To assess the contribution of each FC in the CPM classifier model, we implemented a feature importance calculation approach utilized in previous studies (Pang et al., 2022; Speer et al., 2022). Specifically, we established the full CPM classifier accuracy as a baseline performance using all selected FCs. Then we removed one FC at a time and calculated the lesion model performance with the remaining FCs based on the same LOOCV process. If the model performance of the lesion model is lower than the full model, it means that the corresponding FC is important. The performance difference between the full and lesion models indicates the importance of the removed FC. Consequently, the higher the difference, the more important the FC. Among the important FCs, we also evaluate which FCs are critical in the full model. We calculated the permutation significance of the lesion model without one important FC. Suppose the performance of the lesion model was not significant. In that case, it indicates that the CPM could not significantly classify group labels without this removed important FC, which means that this FC is critical in group classification.

2.9. Regression analysis

After finding the critical FCs in the classification, we speculated whether these FCs may only be significantly associated with behavioral performance in the OT group. To achieve this goal, we conducted a linear regression analysis on the OT and PL groups. The behavioral performance index was linearly regressed onto FC (Equation (1)).

$$\text{Behavior}={\beta }_0+{\beta }_1\mathit{FC}+ϵ$$ (1)

where FC denotes the selected important functional connectivity. β ₀ as the regression intercept. β ₁ as the regression slop. Behavior denotes the behavioral performance value. The model parameters were estimated using ordinary least squares. After each regression process, we obtained the regression slope and corresponding p‐value coefficients.

2.10. Statistical analysis

Independent sample t‐tests were utilized for continuous variables including head motion, age, questionnaire, task behavior performance, and FC between the OT and PL groups. Paired t‐tests were used to access the rsFC‐tsFC similarities of all FCs between the OT and PL groups. Spearman's correlation was utilized to calculate the prediction performance of the CPM predictor. The performance correlation coefficients of two groups were compared using the cocor package (Diedenhofen & Musch, 2015) in R, which offers two statistical difference tests: p‐value based on Fisher's z transformation (Fisher, 1925) or confidence interval (Zou, 2007). A one‐tailed test was conducted as it was hypothesized that OT would outperform PL. Permutation tests were utilized to evaluate the statistical significance of the CPM predictor and classifier performance. One‐dimensional linear regression analysis was utilized to evaluate the association between specific FC and task behavior performance.

3. RESULTS

3.1. Behavioral results

In the face‐perception task, we calculated two indices for each participant—namely, mean decision accuracy (Acc mean) and mean reaction time (RT mean). Figure 1c presents the two behavioral response indices in the face‐perception task in the OT and PL groups. We performed independent two‐sample t‐tests between the OT and PL groups for these behavioral indices and found no significant difference between the two groups in both Acc mean (t ₅₇ = −0.552, p = .583) and RT mean (t ₅₇ = −0.257, p = .798).

3.2. OT enhances the prediction effect from FC to behavior

First, we assessed whether OT enhanced the association between rsFC and behavior. We used a CPM to predict the behavioral index based on rsFCs for the OT and PL groups and examined whether the prediction performance of the CPM predictor in the OT and PL groups differed. We trained two models with the same hyperparameters for each behavioral index to compare the predictive abilities of the two groups. Because there were two groups (OT/PL) and two behavioral performance indices (Acc mean/RT mean), we trained a total of four CPMs. By setting the threshold to 0.05, we selected 349 features in the OT group and 176 features in the PL group to predict the Acc mean, and 35 features in the OT group and 80 features in the PL group to predict RT mean.

Based on the selected FCs, we used a linear SVM regressor to predict task performance (Figure 3). We found that a significant predictive effect existed only in the OT group. Figure 3a presents the results of using the CPM to predict the ACC mean; the prediction effect was insignificant in the PL group (ρ = −0.164, p = .655). However, we observed significant prediction accuracy in the OT group (ρ = 0.498, p = .0025**). We also found the performance in the OT group was significantly higher than that in the PL group for (z = 2.59, p = 0.005**, 95% CI [0.16,1.07]). Figure 3b presents the results of using the CPM to predict the RT mean; the prediction effect was not significant in the PL group (ρ = −0.411, p = .943), but significant in OT group (ρ = 0.677, p < .001***). The performance in the OT group was significantly higher than the PL group (z = 4.59, p < 0.001***, 95% CI [0.65,1.40]). We also repeated the same prediction analysis using the feature selected by threshold = 0.01. We found the significant prediction effect is only in the OT group but not in the PL group (see more details in Table S5 and Figure S2 AB).

FIGURE 3.

Results of using the CPM predictor to predict task performance by resting state FC and task‐state FC. We present Spearman's correlation between the actual behavioral value and predicted behavioral value alongside the permutation result of the correlation coefficient. The significant features were selected by threshold = 0.05. (a) The result of using the resting‐state FC to predict Acc mean in the PL (left) and OT (right) groups. (b) The result of using the resting‐state FC to predict RT mean in the PL (left) and OT (right) groups. (c) The result of using the task‐state FC to predict the Acc mean in the PL (left) and OT (right) groups. (d) The result of using the task‐state FC to predict the RT mean in the PL (left) and OT (right) groups.

Next, we tested whether there was a similar pattern between tsFC and behavior. By setting the threshold to 0.05, we selected 9 features in the OT group and 146 features in the PL group to predict the.

Acc mean. Besides, 884 features in the OT group and 362 features in the PL group predict RT mean.

Figure 3c,d present the results of using the CPM to predict the ACC mean and RT mean, respectively. Similar to the previous result of rsFC, a significant prediction effect only existed in the OT group. The prediction effect from tsFC to Acc mean was not significant in the PL group (ρ = −0.432, p = .347) (Figure 3c). Furthermore, there is a positive relationship between the actual behavioral value and prediction in the OT group, although the effect was not significant but close to (ρ = 0.366, p = .052). We also found the performance in the OT group was significantly higher than that in the PL group (z = 3.08, p < .001***, 95% CI [0.29,1.18]). A pattern similar to the rsFC CPM was also found in the prediction from task FC to RT mean (Figure 3d. The prediction effect was also insignificant in the PL group (ρ = 0.107, p = .200) but significant in the OT group (ρ = 0.313, p = .029*). But the performance in the OT group was not significantly higher than the PL group (z = 0.79, p < .22, 95% CI [−0.30,0.68]). We also repeated the same prediction analysis using the feature selected by threshold = 0.01, and found the significant prediction effect is only in the OT group but not in the PL group (see more details in Table S5 and Figure S2 CD). Results demonstrated that the significant prediction effect from FCs to behavior indices could only exist after OT administration in both the resting‐ and task‐state fMRI.

3.3. OT enhances similarity between rsFC and tsFC

In the present study, we used a simple correlation analysis to verify that OT enhances resting‐task association at the whole‐brain level (Figure 4a). As shown in Figure 4b, the mean correlation coefficient between rsFC and tsFC in the OT group was significantly higher than that in the PL group (t ₄₀₀₄ = 28.98, p < .001).

3.4. OT alters resting‐state FC

After finding that OT could enhance the association between behavioral performance, rsFC, and tsFC, we were curious about how OT could enhance this triangular association. OT might alter one or more vertices of the triangle. Because the behavioral performance was not significantly different (Figure 1c), we hypothesized that OT might directly alter the participants' FC. For this conjecture, we developed the CPM classifier to test whether it could discriminate the FCs in the OT group from those in the PL group and to determine which FC features were significantly different between the two groups. We speculate that these features may only have a significant correlation with behavioral indicators in the OT group.

After the feature selection, 25 significant FCs values were retained by setting the threshold value to p < .02. To verify the stability of the results, we also used two different thresholds to extract the FC features and conducted the following analysis separately (Supplementary Section 2; Figure S4). We used an SVM classifier based on the selected FCs to classify the groups (OT/PL). We found that the classification accuracy was significantly higher than the chance level: the classification accuracy resulted in Acc = 0.678 with significance level p = .0047** (see Figure 5a).

FIGURE 5.

Results of the CPM classifier in resting‐state FC classification. (a) The permutation result of the model performance. (b) The location of the top 7 important FC features (c) The feature importance of the top 7 FC features. (d) Association between Acc mean and TPOmid.R‐ANG.L. The brain map above shows the location of the two ROIs of FC, and the correlogram below shows the association between FC and Acc mean in the PL group (Green line) and OT group (Orange line). (e) Association between Acc mean and FC (Cau.R‐ORBmid.R. The brain map above shows the location of the two ROIs of FC, and the correlogram below shows the association between FC and Acc mean in the PL group (Green line) and OT group (Orange line).

Next, we explored which FCs contribute the most to classification accuracy and found seven important FCs that had a positive impact on classification performance. We revealed that the FC between the right temporal pole (TP) middle temporal gyrus (TPOmid.R) and the left angular gyrus (ANG.L) was crucial for classifying whether FC belonged to participants treated by OT or PL. Additionally, the connectivities between the right caudate nucleus (CAU.R) and right middle frontal gyrus orbital part (ORBmid.R), right thalamus (THA.R) and left hippocampus (HIP.L), right lenticular nucleus putamen (PUT.R), right middle frontal gyrus orbital part (ORBmid.R), left insula (INS.L) and left middle frontal gyrus (MFG.L), left TP superior temporal gyrus (TPOsup.L), left angular gyrus (ANG.L), right and left paracentral lobule (PCL.L), and posterior cingulate gyrus (PCG.R) were found to be important FCs (see Figure 5c). The location of all the important FCs can be found in Figure 5b. Among all seven important FCs, we found only the lesion model that removed the first important FC (TPOmid.R—ANG.L, ρ = 0.1036) or the second important FC (Cau.R—ORBmid.R, ρ = 0.0729) could not significantly classify group label, which means these two features are critical in OT/PL classification. The significance level of the model without other important FCs can be found in Table S7. According to an independent sample t‐test, we found that the strength of all important rsFCs in the OT group was significantly lower than that of the PL group (see Table S9). This suggested that OT administration significantly reduced FC strength.

Because OT affects rsFC, and some FCs are critical in OT/PL groups classification, we speculated whether these rsFCs that are significantly altered by OT would show different associations with behavioral performance. For each pair of rsFC and behavior, we performed a regression analysis using rsFC as the regressor and task behavior performance as the data to be regressed. Regression analysis was performed separately for the OT and PL groups to obtain the respective regression coefficients in the two groups. We only focused on rsFC, which showed significant differences between the OT and PL groups; hence, we chose the first two critical rsFC features that contributed the most to the CPM (TPOmid.R —ANG.L, p = .1036; and CAU.R—ORBmid.R, p = .0729).

In the PL group, we found no significant association between these rsFC features and task behavior. However, in the OT group, we found that the first two rsFCs were associated with the behavioral index. The FC between the right TP of the middle temporal gyrus (TPOmid.R) and the left angular gyrus (ANG.L) was significantly associated with the Acc mean (β = 0.199, p = .03*) in the OT group (Figure 5d), but not in the PL group (β = 0.133, p = .429). The FC between the right caudate nucleus (CAU.R) and the orbital part of the right middle frontal gyrus (ORBmid.R) was significantly negatively associated with the Acc mean (β = −0.214, p = .039*) in the OT group (Figure 5e), but not in the PL group (β = 0.133, p = .553). We further used a statistical model to examine whether OT has a moderating effect on the association between rsFC and behavior (Aiken et al., 1991), and found the moderating effect was nonsignificant (p = .081, see Supplementary Section 4).

3.5. OT alters task‐state FC

Next, we performed the same CPM classification analysis for task‐state FC and obtained a classification accuracy significantly higher than chance (Acc = 0.661, p = .0381*, see Figure 6a). Similar to the classification of rsFC, we used two different thresholds to perform the same analysis (Supplementary Section 2; Figure S5). The results indicated that OT altered FC in both the resting and task states. Similarly, we identified four important FCs that positively contributed to the task‐state OT/PL classification (Figure 6b). We showed FC between the right TP superior temporal gyrus (TPOsup.R) and left superior frontal gyrus orbital part (ORBsup.L), TPOsup.R and left supramarginal gyrus (SMG.L), left TP middle temporal gyrus (TPOmid.L) and inferior frontal gyrus orbital part (ORBinf.R), right TP middle temporal gyrus (TPOmid.R) and right superior parietal gyrus (SPG.R) exhibited a significant difference between the OT and PL groups in the task state (Figure 6c). Among four important FCs, we found only the lesion model that removed the first important FC (TPOsup.R—ORBsup.L, p = .3178), the second important FC (TPOsup.R—SMG.L, p = .1084), and the third important FC (TPOmid.L—ORBinf.R, p = .0834) could not significantly classify group label, which means these three features are critical in OT/PL classification. The significance level of the model without other important FC can be found in Table S8. We also found that the strength of all important tsFCs in the OT group was significantly lower than that of the PL group (see Table S10). For tsFC, we found that only one of the three critical FCs showed a different association with behavioral performance between the OT and PL groups. The FC between the right TP superior temporal gyrus (TPOsup.R) and left supramarginal gyrus (SMG.L) was significantly associated with the RT mean (β = 0.204, p = .04*) in OT group, but the association was not significant in PL group (β = −0.127, p = .414; Figure S6). The modulating effect was nonsignificant (p = .064, Supplementary Section 4).

FIGURE 6.

Result of the CPM classifier in task‐state FC classification. (a) The permutation result of the model performance. (b) The location of the top four important FC features. (c) The feature importance of all 8 FC features.

4. DISCUSSION

Linking individual differences in cognitive traits and their brain activity is a long‐lasting research direction (Dadi et al., 2019). Previous researchers have tried to better establish this association by using a new brain imaging technique (Shahsavarani et al., 2023) or a better predicting model (Gal et al., 2022). However, few studies have investigated the role of neuropeptides in this association. In this study, we conducted comprehensive whole‐brain analyses to investigate whether OT modulates the relationship between behavioral performance, rsFC, and tsFC. Using CPM‐based methods, our findings demonstrate that OT could enhance the prediction association between behavioral performance and FC in both the resting and task states. OT was also found to enhance the similarity between rsFC and tsFC. Furthermore, our results indicated that OT enhanced this triangular association (Figure 1b) by altering the FC in social‐related areas in both the resting and task states. Overall, our work validated the triangular association‐enhancing effect of OT and provided the first evidence that OT could enhance the association between FCs and behavior performance by altering the FC in both the resting and task states.

By utilizing CPMs, we demonstrated that FCs in both the resting and task states could significantly predict the Acc mean and RT mean in the OT group but not in the PL group (Questions 1 and 2). This finding suggests that OT enhances the brain‐behavior association across different brain states. The brain‐behavior association is generally considered the association between individual differences in brain activity and behavior (Rosenberg & Finn, 2022). Prior research has established that OT modulates brain activity based on individual factors such as age (Horta et al., 2019), gender (Dumais & Veenema, 2016), and OT receptor genetics (Zhao et al., 2020). Consequently, OT may amplify individual differences in social brain activity related to self‐referential or face recognition tasks, leading to a stronger brain‐behavior association in our study. Interestingly, rsFC demonstrated superior predictive accuracy than tsFC for both behavioral indices in the OT group. This suggests that face‐perception accuracy and reaction time may reflect inherent individual traits, which are more readily captured by rsFC (Tavor et al., 2016).

Although many studies have explored the consistency between individual differences at different levels (Beaty et al., 2018; Finn et al., 2015; Ovando‐Tellez et al., 2022) including our previous studies (R. Wang et al., 2022; Li et al., 2023; Zhang et al., 2023), few studies have investigated the influence of neuropeptides, including OT, on this consistency. Our analysis, however, revealed a significant increase in the consistency between rsFC and tsFC under the influence of OT, suggesting heightened stability of FC across different states (Question 3). This aligns with prior research (Gratton et al., 2018; Stevens, 2016) indicating the relative stability of functional brain networks across resting and task states. Furthermore, it has been demonstrated that task activity can be predicted from resting state data in healthy adults (Cole et al., 2016), and that OT can modulate the relationship between task and resting‐state social networks (Wu et al., 2023). These findings collectively underscore the potential of OT in enhancing the consistency between rsFC and tsFC.

Our findings underscore that OT can modify FCs within the social network, aligning with previous research (Bethlehem et al., 2013; Jiang et al., 2021; Sripada et al., 2013). The OT effect on human brain, and the brain regions involved in self‐face recognition largely overlap with the brain regions in the social brain network (Adolphs, 2009). For instance, OT has been shown to alter FCs between key areas such as the limbic system, orbitofrontal cortex, precuneus, and middle temporal sulcus (Riem et al., 2012). Furthermore, OT may modulate connectivity in regions including the thalamus and superior temporal sulcus (Bos et al., 2012). Relatedly, for rsFC, we demonstrated that the FC of brain regions in TP (TPOmid.R; TPOsup.L), PFC (ORBmid.R; MFG.L), TPJ (ANG.L), insula (INS.L), and limbic system (CAU.R; THA.R; HIP.L) made significant contributions to the classification of the OT and PL groups. For tsFC, we demonstrated that the FC of brain regions in the TP (TPOmid; TPOsup.R), PFC (ORBsup.L; ORBinf.R), and TPJ (SMG.L; SPG.R) is important for classification. These findings not only align with previous studies (Bos et al., 2012; Riem et al., 2012) but also provide comprehensive, whole‐brain evidence that OT can regulate the FC between these socially relevant brain regions. OT system is highly conserved in evolution, from its molecular structure to the location of its receptors. Due to the difficulty in developing specific reporters for OT receptors. Alternative methods have been applied to infer the distribution of oxytocin receptors in human brain (Boccia et al., 2013; Grinevich et al., 2016). In human brain samples, OT receptors were observed in the basal nucleus of Meynert, the nucleus of the vertical limb of the diagonal band of Broca, the ventral part of the lateral septal nucleus, the hypothalamic, pallidum, amygdala, striatum, caudate, dorsal ACC, and MPFC (Loup et al., 1991; Skuse & Gallagher, 2009). Our result showed that the FC related to the hippocampus and prefrontal cortex could be altered by OT, likely through the OT receptors expressed in these brain regions.

Further analysis substantiates our hypothesis that OT may alter FC to enhance the triangular association (Question 4). Several key FCs in the classification were significantly associated with behavioral performance only in the OT group, mirroring the results from the CPM predictor. For instance, during the resting state, the FCs between the TP and TPJ, and between the caudate and vmPFC, were significantly associated with mean accuracy (Acc mean) following OT administration. These regions have been implicated in social cognition and face recognition (Adolphs, 2009; Olson et al., 2007), and disruptions in the connection between the caudate and vmPFC have been linked to impaired social cognition (Alexander et al., 1986; Graff‐Radford et al., 2017). In the task state, the FC between the TP and TPJ was significantly associated with mean reaction time (RT mean) after OT administration, with the SMG being related to awareness and attention control (Wilterson et al., 2021). Our results indicate that OT primarily influences individuals' baseline activity (resting) rather than task‐related activity, which aligns with previous findings suggesting that rsFC embodies more stable individual traits (Bethlehem et al., 2013; Tavor et al., 2016). However, the differing acquisition times for our resting‐state and task‐based scans warrant careful consideration. While our 5‐minute resting scan could potentially affect the reliability of rsFC estimates (Noble et al., 2019; Shah et al., 2016), other studies have found this duration to be sufficient for reliable rsFC (Airan et al., 2016; van Dijk et al., 2010; Zuo et al., 2013). The literature is less conclusive regarding the impact of scan duration on tsFC. This gap in understanding suggests that the influence of acquisition time on task‐based scans remains an area for future research and should be interpreted with caution (Shah et al., 2016). Given our focus on the predictive effects of rsFC and tsFC on behavior between the OT and PL groups, these acquisition time differences may exert limited influence on our primary findings, as each scanning condition serves as its own baseline for predictive modeling.

While our study provides valuable insights, it has certain limitations and presents opportunities for future research. First, our study, employing connectome‐based predictive modeling (CPM), was an exploratory analysis partly based on our previous results about the oxytocin effect. As such, we did not have an explicit preregistration plan, which may limit the explanatory power of our findings regarding the mechanisms of OT's effects on brain‐behavior relationships (Nosek et al., 2018). While this approach allowed us to predict these relationships effectively, it does not provide a detailed explanation of the underlying mechanisms. Further, the lack of significant brain‐behavior associations in the placebo group could be a sensitivity issue and should be addressed in future studies. Future research could adopt a theory‐driven approach to investigate the neural mechanism behind our findings, potentially utilizing other techniques such as electroencephalography (EEG) or magnetoencephalography (MEG). These techniques could provide high‐temporal resolution and direct measures of neural activity, which may offer insights into the immediate and dynamic effects of OT on brain‐behavior relationships (Korisky et al., 2022; Schiller et al., 2019). Second, prior research suggests that task‐state functional connectivity (tsFC) may be more suitable for dynamic FC calculation, given the high time‐varying property of the task‐state BOLD signal (Feifel et al., 2016; Hutchison et al., 2013). Therefore, future research could benefit from employing dynamic FC calculations to validate our findings' robustness and provide a more nuanced understanding of the effects of OT on tsFC (Li'egeois et al., 2019; Preti et al., 2017). Lastly, our results primarily reflect the short‐term effects of OT on brain‐behavior prediction. Future studies could explore the effects of OT on other social tasks, with both short and long‐term predictions in the OT and placebo (PL) groups. Such research could further refine our understanding of OT's impact on brain‐behavior relationships and could provide insights into the temporal dynamics of OT's effects (Bales et al., 2013; Knight & Lindsay, 1970).

5. CONCLUSION

Our results provide a novel insight that OT may enhance the triangular association between face‐perception behavior, rsFC, and tsFC. This enhancement could be partly explained by OT altering the FC between social brain regions in both the resting and task states. This result provides evidence that neuropeptides can increase the consistency between different modalities. Future studies should be conducted to validate this effect of OT using different social tasks, neuroimaging signals, or patients with mental disorders.

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no competing financial interests.

Supporting information

Data S1: Supporting Information.

Zhang, H. , Chen, K. , Bao, J. , & Wu, H. (2023). Oxytocin enhances the triangular association among behavior, resting‐state, and task‐state functional connectivity. Human Brain Mapping, 44(17), 6074–6089. 10.1002/hbm.26498

DATA AVAILABILITY STATEMENT

The fMRI data used in this study cannot be made publicly available but can be requested from the corresponding author [Haiyan Wu]. Code for the present data analyses is available at the GitHub repository: https://github.com/andlab-um/OT-cpm.