Fronto-parietal homotopy in resting-state functional connectivity predicts task-switching performance

Forty-seven young healthy individuals (26 females; mean age = 24.7 years, SD = 3.5) voluntarily took part in the study. All participants gave informed consent prior to their recruitment. For their time, they were reimbursed for both the fMRI and the behavioral experimental sessions (see details on the General procedure section). All of them were right-handed as assessed with the Edinburgh Handedness Inventory (Oldfield 1971) with an average score of 87.8 (SD = 12.2), reported no history of neurological or psychiatric disorders, had a normal color vision and normal or corrected-to-normal visual acuity. The procedures involved in this study were approved by the Bioethical Committee of the Azienda Ospedaliera di Padova–AOP (Prot # 2758P).

Each participant was tested in two separated experimental sessions (mean inter-session interval = 21.5 days, SD = 17.4). In the first session, structural and resting-state functional MRI (rs-fMRI) data were acquired along with some task-related fMRI data collected for different aims and published elsewhere (e.g., Visalli et al. 2019). The order of acquisition was: (i) rs-fMRI, (ii) task-related fMRI, and (iii) structural MRI. In the second session, participants performed the color-shape task-switching paradigm, which is the focus of the present study.

Participants performed a color-shape task-switching paradigm (Rubin and Meiran 2005; Ambrosini and Vallesi 2016; Vallesi et al. 2016), which was implemented in MATLAB (The MathWorks, Inc., Natick, Massachusetts, United States) using the Psychophysics Toolbox 3 (Kleiner et al. 2007). A graphical representation of the paradigm is presented in Fig. 1. Participants sat in a dimly lit soundproof cabin at a viewing distance of approximately 60 cm from the computer screen. Each trial started with a black fixation cross (visual angle: 0.28° by 0.28°) displayed at the center of the screen. After 1500 ms, a cue stimulus (visual angle: 5.5° by 1.7°) signaling the task to be performed was presented 4.6° above the fixation cross. The color task cue consisted of a row of three colored rectangles (purple, orange, and yellow), while the shape task cue consisted of a row of three small black shapes (a triangle, a circle, and a square). Graphic cues were chosen to limit the use of linguistic information (Ambrosini and Vallesi 2016). After a cue-to-target interval (CTI) of either 300 ms (short CTI) or 1200 ms (long CTI), the fixation cross was replaced by the target stimulus consisting of a heart or a star shape (visual angle: 3.5° by 3.5°) in either red or blue color. The short CTI was long enough for task-cue encoding but it was probably too short for task reconfiguration (cf., Meiran 1996; see also: Luria and Meiran 2006; Schneider 2016; Lange et al. 2018). The long CTI was instead long enough for the completion of task reconfiguration since its duration was considerably higher than average RTs in slow conditions (Rogers and Monsell 1995; Meiran 1996). The target was displayed until a response was produced by the participant (maximum response time allowed: 2500 ms). Participants were instructed to respond as fast and accurately as possible to either the shape or the color of the target on the basis of the target cue. Responses were provided by pressing one of the two lower buttons of the CEDRUS RB-840 response pad with the left and right index fingers. The assignment of the two shapes and the two colors to response keys (2-by-2 possible combinations) was counterbalanced across participants. The four possible stimuli (two colors by two shapes combination) and the two CTIs were pseudo-randomly interleaved within each block of trials ensuring that each possible stimulus-by-CTI combination was presented an equal number of times across trial types (see below).

Participants completed eight blocks of trials of two different types: (i) single-task blocks, during which participants performed only one task for the entire block (either color or shape), (ii) and task-switching blocks, in which participants were required to switch or repeat the task performed at the previous trial on the basis of the task cue. In the single-task blocks, the cue was always the same (i.e., the color task cue for the pure color block, and the shape task cue for the pure shape block). The task procedure was structured as follows: (i) two single-task blocks of 40 trials each (one for each task; task order counterbalanced across participants); (ii) two task-switching blocks of 65 trials each (32 switch trials + 32 repeat trials + the initial trial); two 40-trial single-task blocks (task order inverted with respect to the two initial single-task blocks). The CTI manipulation was implemented in each block type. This trial-wise (instead of block-wise) CTI manipulation was used to overcome the possible confounding effects due to changes in task strategy or differences in memory load (Meiran 1996). The first three blocks were preceded by a short practice phase (8 trials for the single-task blocks, 16 trials for the switching block). During practice only, feedback about accuracy and speed was provided after each response (the Italian translations for “Well done”, “Correct, but try to be faster”, or “Wrong” were displayed at the center of the screen after a correct response given within the 2 s maximum RT allowed in test trials, a correct response > 2 s, or an incorrect response, respectively). Self-paced breaks were available between blocks.

Data from three participants were discarded since they were excluded for imaging issues (see below). The final sample for all the reported analyses, hence, included 44 participants. A sensitivity power analysis (G*Power 3 software; Faul et al. 2007) revealed that our sample size was large enough to detect significant (α = 0.05) mean differences between two dependent means (e.g., the main effect or the interaction effect of a 2-by-2 repeated measures ANOVA) with a medium effect size d = 0.43 (Cohen 1988) with a statistical power of 0.80. Data from the practice phase, first trial in a block, error (incorrect or no response) and post-error trials were excluded from analysis. For each trial type (i.e., single-task, repeat and switch trials) at each CTI, response time (RT) values more extreme than one and a half times the interquartile range (i.e., the difference between the upper and lower quartile) above the upper quartile or below the lower quartile were identified as outliers (Borcard et al. 2011) and excluded (mean excluded trials = 5.1%, SD = 1.9%). Mean RTs were, then, computed for each trial type at each CTI and used to calculate two behavioral cost measures: (i) switching costs (Monsell 2003), computed as the difference between the mean RT from switch trials and that from repeat trials; (ii) and mixing costs (Rubin and Meiran 2005), computed as the difference between the mean RT from repeat trials and that from pure trials. The resulting four scores (i.e., mixing and switching costs at short and long CTI; Fig. 2) were used for the brain-behavior correlations with the homotopy scores. The correlation between switching and mixing costs was r = 0.03 (p = 0.858) at short CTI, and r = − 0.40 (p = 0.005) at long CTI.

Structural and functional images were acquired with a 3 T Ingenia Philips scanner (Philips Medical Systems, Best, The Netherlands) at the Neuroradiology Unit of the Padova University Hospital. The system was equipped with a 32-channel head-coil. Functional data consisted of 250 T2*-weighted echo-planar image (EPI) volumes (repetition time, TR: 2000 ms; echo time, TE: 30 ms; 39 axial-slices with ascending acquisition; voxel size: 3 × 3 × 3 mm; flip angle, FA: 76°; acquisition matrix: 84 × 84). After the functional session, high resolution T1-weighted anatomical images were acquired (TR/TE: 8.10/3.72 ms; 180 sagittal slices; voxel size: 1 × 1 × 1 mm; FA: 8; acquisition matrix: 256 × 256). To reduce head movements during data acquisition, small foam cushions and sponge pads were placed around the participant’s head.

Two participants were excluded after visual quality check, reporting enlarged ventricles which could affect normalization steps, while one participant was excluded due to an incomplete rs-fMRI exam, thus leaving a sample of 44 participants for the analysis. Functional data were preprocessed using the FMRIB Software Library, version 6.0.0) (Smith et al. 2004), according to a previously described pipeline (Pini et al. 2020). Functional data were motion corrected, brain extracted and registered to the corresponding structural image with rigid-body transformation and nonlinearly registered to a symmetric brain template (ICBM 2009a Nonlinear symmetric template). Functional data were high-pass filtered (100 s) and for each functional volume, the framewise displacement (FD) was estimated and volumes with more than 0.25 mm FD were regressed out from the time series. Confounding variables also included the average signal of white matter and cerebro-spinal fluid, 6 head motion parameters, the derivatives of these 8 regressors, and the square of these 16 regressors, as suggested by Satterthwaite et al. (2013). In addition, the time-courses of the first 5 principal components calculated over white matter and cerebro-spinal fluid voxels were included as confounding variables to minimize further the effect of physiological noise, a method referred as aCompCor, widely used in rs-fMRI pre-processing (Behzadi et al. 2007; Chai et al. 2012). Finally, images were spatially smoothed using a Gaussian kernel with a full-width at half-maximum of 6 mm.

For an overview of the methodology see Fig. 3. First, we identify brain regions organized within the FPN through an independent component analysis (ICA). In this step, we further included 15 participants with the same rs-fMRI data who did not come back to complete the second (behavioral) session and an independent dataset of 21 healthy young individuals with MRI sequences with equivalent acquisition parameters collected at a different site (see Supplementary Methods), for a total of 83 healthy participants (45 females; mean age = 24.4 years, SD = 3.2; EHI = 84.5, SD = 27.2). The inclusion of a larger sample is known to ensure a more robust output in fMRI analyses (Desmond and Glover 2002; Mumford and Nichols 2008; Lindquist et al. 2013). Preprocessed fMRI data were fed into the Group ICA Toolbox (GIFT version 3.0a http://​mialab.​mrn.​org/​software/​gift/​) (Calhoun et al. 2001). The number of independent components extracted (n = 76) was chosen according to the minimum description length criteria (Li et al. 2007). The resulting FPN group map was identified through a template matching spatial correlation procedure with standard templates (Shirer et al. 2012), and visually inspected to exclude components underlying artifacts, according to Griffanti et al. (2017) criteria. The FPN map was binarized at z > 2 and used as a whole region of interest (ROI) in the homotopy analysis, according to previous procedures applied by our group to investigate this network (Pini et al. 2020). For simplicity, results are reported in the left hemisphere. Finally, we included the language network, identified through a spatial correlation procedure with the fslcc utility from FSL compared with the Shirer’s language template (Shirer et al. 2012). This network was included as a control network. We assumed that homotopy of regions overlapping with language network hubs would not exhibit a significant association with executive tasks, since these regions converge on a left-lateralized network active during speech reception and production (Braga et al. 2020).

For each participant, we computed homotopy brain property as the voxel-mirrored homotopic functional connectivity between every pair of mirror voxels through in-house scripts based on R using unsmoothed data as previously reported (Gracia-Tabuenca et al. 2018; Zuo et al. 2010). Briefly, in the symmetrical brain space we computed the Pearson’s correlation coefficient between each voxel time-series residuals (after pre-processing steps) with the hemispheric counterpart. Voxels belonging to the three sagittal midline slices (9 mm) were excluded, to avoid inter-hemispheric partial volume effect. Then correlation values were Fisher Z-transformed. Finally, homotopy maps of each individual were smoothed with a gaussian kernel of 6 mm.

We included in the analysis both whole-brain homotopy maps (whole voxel values) and homotopy connectivity within the FPN, computed masking the whole-brain homotopy maps with the group ROI-FPN binarized map (see previous section).

We investigated the association between brain functional properties (i.e., connectivity strength and homotopy) and mixing/switching costs in different CTI (long, short) in the sample of participants who completed both fMRI and behavioral sessions.

This analysis was performed at FPN ROI levels: (i) for the homotopy analysis, we first averaged homotopy values of voxels within the FPN map. Then we computed correlational analysis between behavioral measures and mean homotopy; (ii) for the functional connectivity strength analysis, the FPN maps from GIFT were threshold at z > 2 and an average score for each participant was calculated. The ROI analysis was repeated considering whole (left) brain mean homotopy values to assess whether the association between homotopy and mixing/switching costs was stronger within FPN regions than for the whole brain. Due to the non-linear distribution of homotopy values, we performed a nonparametric Spearman correlation to assess the association between connectivity organization and cognition. p values were corrected for multiple comparisons (n = 4 behavioral measures; p < 0.0125) to control for type I error with Bonferroni correction.

At the voxel-wise level, we investigated brain-behavior correlation between cognitive performance with both FPN connectivity and homotopy strength. We implemented a nonparametric inference based on permutations (n = 5000) using FSL randomise (https://​fsl.​fmrib.​ox.​ac.​uk/​fsl/​fslwiki/​Randomise). Multiple comparisons were corrected across space using familywise error (FWE) based on permutation testing at a threshold-free cluster enhancement (TFCE; Smith and Nichols 2009). Positive and negative associations between homotopy and behavioral measures were tested at a TFCE level of p < 0.025 corresponding to a two-tailed p < 0.05 (randomise correction is only performed on one-tailed tests). Each contrast was restricted to the FPN map, computed as described above.

To ensure that the mean homotopy map used in the correlational analysis exhibits a reliable profile, we computed, through the same procedure, voxel-mirrored homotopic functional connectivity from a larger open-source available dataset. To this aim, we retrieved participants from the Population Imaging of Psychology dataset (PIOP2), which is released within the Amsterdam Open MRI Collection (AOMIC). This dataset was selected according to the number of scans (250 scans) and TR (2000 ms), which were comparable with the dataset used in the present study (240 scans and TR = 2000 ms). For further details about PIOP2 dataset, see Snoek et al. (2021). From the PIOP2 dataset, we retrieved 224 participants with both structural MRI and rs-fMRI data available. The list of PIOP2 participants included in this analysis is reported in Supplementary Table S1. From the PIOP2 dataset, we computed both mean Z-Fisher homotopy maps and one-sample homotopy t-maps. We then ran spatial cross-correlation between both mean and t-maps from PIOP2 and the study dataset through the FSL utility fslcc (FSL v.6.0.0; https://​fsl.​fmrib.​ox.​ac.​uk/​fsl/​).

Reliability of behavioral measures—i.e., mean RT for each trial type (pure, switch, and repeat trials) at each CTI (300 and 1200 ms) and the resulting mixing and switching costs—was evaluated by means of split-half correlations corrected with Spearman-Brown formula. Reliability coefficients were calculated as follows: (i) trials of each participant for each trial type-by-CTI combination were randomly splitted into two subsets of equal size; (ii) mean RTs was computed separately for the two subsets; (iii) switching and mixing costs were also calculated for the two subsets; (iv) correlations of participant’s mean RTs and costs between the two subsets were calculated and corrected. The procedure was repeated 2000 times and the median correlation taken as an index of reliability.

Descriptive statistics about switching and mixing costs are shown in Fig. 4 and Table 1. Concerning switching costs, 2-by-2 repeated-measures ANOVAs with within-subject factors “Trial type” (levels: repeat and switch trials) and CTI (levels: short and long) were conducted on (arcsine transformed) accuracy rates and mean RTs. The ANOVA on accuracy (see Fig. 4, Panel A) showed significant effects for Trial type (F(1,43) = 26.2, p < 0.001, partial η² = 0.38), CTI (F(1,43) = 25.5, p < 0.001, partial η² = 0.37), and their interaction (F(1,43) = 9.46, p = 0.004, partial η² = 0.18). Accuracy was higher for repeat trials at long compared to short CTI, although this CTI effect was greater for switch trials. Post-hoc comparisons (Holm correction method) showed a significant trial-type difference at short CTI (mean difference = 0.14, SE = 0.02, t = 5.92, p < 0.001, 95% CI for mean difference = 0.07–0.20, Cohen’s d = 0.89), but not at long CTI (mean difference = 0.05, SE = 0.02, t = 2.10, p = 0.118, 95% CI = − 0.01–0.11, Cohen’s d = 0.32). They also showed a significant CTI difference for switch trials (mean difference = − 0.12, SE = 0.02, t = − 5.76, p < 0.001, 95% CI = − 0.17 to − 0.06, Cohen’s d = -− 0.87), but not for repeat trials (mean difference = − 0.03, SE = 0.02, t = − 1.46, p = 0.296, 95% CI = − 0.08–0.02, Cohen’s d = − 0.22). The ANOVA on RTs revealed significant effects for Trial type (F(1,43) = 108.42, p < 0.001, partial η² = 0.72), CTI (F(1,43) = 291.61, p < 0.001, partial η² = 0.87), and their interaction (F(1,43) = 52.30, p < 0.001, partial η² = 0.55). Figure 4, Panel B shows that RTs were longer in switch trials. They were longer also at the short CTI and this CTI effect was greater for switch trials. Post-hoc comparisons revealed that CTI differences were significant in both repeat trials (mean difference = 0.12 s, SE = 0.02, t = 7.13, p < 0.001, 95% CI = 0.08–0.17, Cohen’s d = 1.08) and switch trials (mean difference = 0.30 s, SE = 0.02, t = 17.25, p < 0.001, 95% CI = 0.25–0.35, Cohen’s d = 2.60). Trial type differences were significant at both short (mean difference = − 0.21 s, SE = 0.02, t = − 12.45, p < 0.001, 95% CI = − 0.25 to – 0.17, Cohen’s d = − 1.88) and long CTI (mean difference = − 0.03 s, SE = 0.02, t = − 2.11, p = 0.037, 95% CI = − 0.08–0.01, Cohen’s d = − 0.32). A paired samples t test comparing switching costs at short vs long CTI revealed a significant difference (mean difference = 0.18, SE = 0.03, t(43) = 7.23, p < 0.001, 95% CI = 0.13–0.23, Cohen’s d = 1.09).

Concerning mixing costs, 2-by-2 repeated-measures ANOVAs with within-subject factors “Trial type” (levels: pure and repeat trials) and CTI (levels: short and long) were conducted on (arcsine transformed) accuracy rates and mean RTs. The ANOVA on accuracy (see Fig. 4, Panel C) showed no significant effects for Trial type (F(1,43) = 0.35, p = 0.560, partial η² = 0.008), CTI (F(1,43) = 1.08, p = 0.305, partial η² = 0.02), and their interaction (F(1,43) = 2.34, p = 0.134, partial η² = 0.05). The ANOVA on RTs revealed significant effects for Trial type (F(1,43) = 90.08, p < 0.001, partial η² = 0.68), CTI (F(1,43) = 92.70, p < 0.001, partial η² = 0.68), and their interaction (F(1,43) = 56.82, p < 0.001, partial η² = 0.57). As shown in Fig. 4, Panel D, RTs were shorter in pure trials. Moreover, they were also shorter with long CTI, especially in repeat trials. Post-hoc comparisons showed that the CTI differences were significant for repeat trials (mean difference = 0.12 s, SE = 0.01, t = 12.18, p < 0.001, 95% CI = 0.10–0.15, Cohen’s d = 1.84), but not for pure trials (mean difference = 0.02 s, SE = 0.01, t = 1.88, p < 0.063, 95% CI = − 0.01–0.05, Cohen’s d = 0.28). Trial type differences were significant at both short (mean difference = − 0.19 s, SE = 0.02, t = − 11.85, p < 0.001, 95% CI = − 0.23 to – 0.14, Cohen’s d = − 1.79) and long CTIs (mean difference = − 0.08, SE = 0.02, t = − 5.16, p < 0.001, 95% CI = − 0.12 to − 0.04, Cohen’s d = − 0.78). A paired samples t test comparing mixing costs at short vs long CTI revealed a significant difference (mean difference = 0.11, SE = 0.01, t(43) = 7.54, p < 0.001, 95% CI = 0.08–0.13, Cohen’s d = 1.14).

Mean maps of the larger sample (n = 83) showed higher homotopy values within the motor and visual regions, and a decreased pattern in medial frontal, orbitofrontal and limbic regions (Fig. 5, Panel A). This pattern was echoed by the one-sample t map (Fig. 5, Panel B) in the whole sample (fsl-randomize, n = 5000 permutation). T values highlight consistent homotopy values across individuals, incorporating inter-subject variability. Moreover, the same pattern was reported in the subsample of participants who performed the cognitive tasks (n = 44) (Supplementary Figure S1), supporting the robustness of homotopy findings in our sample.

There was a positive correlation between homotopy values at the whole-brain level and mixing costs with long CTI duration (Fig. 6, Panel B), although not surviving multiple comparison correction (Spearman r, r_s = 0.332; p < 0.028). The association between whole-brain homotopy and mixing costs with short CTI and switching costs (both long and short CTI) showed no significant associations (p > 0.1; Fig. 6).

By contrast, there was a significant association between the FPN-ROI homotopic connectivity and mixing costs with both short and long CTI (Fig. 7, Panels A, B), with the latter surviving multiple comparison correction (r_s = 0.493; p = 0.001). A negative trend was observed between FPN-homotopy and switching costs with long CTI, although not surviving correction for multiple comparisons (r_s = 0.364; p = 0.015; Fig. 7, Panel D). No significant association was found for FPN-homotopy and switching costs with short CTI (Fig. 7, Panel C). Overall, this analysis showed that the correlation between homotopy and cognitive measures was significant and survived correction for multiple comparisons in FPN regions, especially for mixing costs.

We further investigated the association between homotopy and FPN-ROI with a voxel-wise level analysis. Clusters showing a significant positive association between homotopy and mixing costs were reported for each CTI. Specifically, for the long CTI, mixing costs were positively related to homotopy in a FPN cluster mapping to the supramarginal gyrus, while for the short CTI, the mixing costs were linked with homotopy values of the superior frontal gyrus (Fig. 8; Table 2). The plots of the distribution of correlation values between behavioral performance and averaged homotopy of these clusters (assessed with Pearson’s correlation) could be appreciated in Fig. 8. In a post-hoc analysis, we investigated whether other frontal regions showed a significant correlation with the switching task performance. To this aim, the group FPN map from GIFT was thresholded at a threshold of z = 1, which allowed the inclusion of larger clusters mapping to the dorsolateral prefrontal cortex and in the temporal gyrus. We applied the same statistical model with a more lenient threshold (TFCE level p < 0.001, uncorrected). This analysis showed positive homotopy-behavior correlations in the same clusters reported in the main analysis (Supplementary Figure S2). Finally, homotopy functional connectivity within the language network (control analysis) did not show significant associations at the voxel-wise level with either mixing or switching costs, as expected.

In contrast to the homotopy results, we did not find a significant association between functional connectivity strength and task-switching performance. At the ROI level, the connectivity strength of the FPN was not linked with either mixing costs (CTI long: r_s = 0.117; p = 0.449; CTI short: r_s = 0.216; p = 0.158; see Fig. 9), or switching-costs (CTI long: r_s = − 0.146; p = 0.343; CTI short: r_s = − 0.292; p = 0.054). This result was in line with the voxel-wise analysis, where we did not observe voxels expressing a significant relationship between connectivity strength and task performance, for any of the contrasts investigated. To ensure that these results were not threshold-dependent, we performed this correlational analysis without applying a threshold to the FPN maps. The same null results were confirmed for both ROIs (mixing costs: CTI long: r_s = − 0.098; p = 0.526; CTI short: r_s = − 0.294; p = 0.053; switching costs: CTI long: r_s = 0.109; p = 0.483; CTI short: r_s = 0.232; p = 0.130) and voxel-wise approaches.

As shown in Fig. 10 (panel A), the homotopy maps from the PIOP2 cohort and our study dataset were similar. The spatial cross-correlation between mean (Z-Fisher) maps was r = 0.89 (Fig. 10; panel B). A similar cross-correlation value was reported when comparing one-sample t-maps, with r = 0.92, suggesting that homotopy features are highly reproducible among independent datasets.

For the behavioral measures, reliability of mean RTs for the six trial type-by-CTI combinations was: 0.96 (range of Pearson’s correlation coefficients: 0.91–98) for pure trials at short CTI; 0.96 (range: 0.92–0.99) for pure trials at long CTI; 0.92 (range: 0.80–0.97) for repeat trials at short CTI; 0.93 (range: 0.79–0.98) for repeat trials at long CTI; 0.91 (range: 0.81–0.96) for switch trials at short CTI; 0.91 (range: 0.75–0.97) for switch trials at long CTI. Median correlations for switching costs were 0.89 (range: 0.71–0.96) at short CTI and 0.88 (range: 0.63-0.97) at long CTI. Median correlations for mixing costs were 0.81 (range: 0.59–0.94) at short CTI and 0.87 (range: 0.61–0.96) at long CTI.

This study has both strengths and limitations. The main merit is that the still poorly understood relationship between executive functions and brain homotopy was investigated using state-of-the-art tools for assessing both cognitive functions and brain connectivity. Second, a data-driven approach (ICA) was implemented to determine FPN hubs, in which the relationship between functional homotopy and task-switching performance was assessed. Data-driven connectivity-based parcellation might increase functional connectivity representation compared to existing atlases (Ren et al. 2019).

Among the limitations, we acknowledge that, since our focus was on young adults with a homogeneous age range, our results cannot be generalized to the whole lifespan. In this regard, our work could be considered as an ideal starting point to further investigate whether the relationship between task-switching performance and functional homotopy is also evident in other age groups.