Dynamic laterality index

We analyzed resting-state fMRI data from 991 participants in the HCP cohort and extracted BOLD time series of all 360 cortical regions using a group-level parcellation scheme (HCP MMP1.0) [32]. To map the time-varying lateralization architecture, we developed a measure termed dynamic laterality index (DLI) by calculating the laterality index in each sliding window for each region of interest (ROI) (Fig 1A). Specifically, we adopted a global signal (GS)-based laterality index that can effectively capture brain lateralization characteristics underlying higher-order cognition [14], defined as the difference between an ROI’s BOLD correlation with the GS of left brain and its correlation with the right brain at each time window (see Methods). GS-based laterality index of an ROI reflects whether the activity of the ROI is more synchronized with the left or the right hemisphere. The correlation between BOLD signal of an ROI and the left/right hemispheric global signal (GS_L/GS_R) represents its synchronization with the left/right hemisphere, respectively. Higher correlation of an ROI with GS_L compared to GS_R indicates left-hemispheric lateralization and vice versa. DLI of an ROI (a time series of laterality index) is then obtained by calculating the laterality index for each time window; see Fig 1A. Positive DLI of a region indicates stronger interaction with the left hemisphere (i.e., leftward laterality), while a negative one indicates rightward laterality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The workflow of dynamic laterality analysis.

(a) Definition of the DLI. The DLI of ROI_i within time window t is defined as the correlation coefficient (z-transformed) between GS_L and the ROI minus the correlation coefficient between GS_R and the ROI. Using a sliding window approach, we obtained a time series of DLI for each ROI. (b) Laterality correlation matrix, which is obtained by correlating laterality time series across all ROIs. Spatial clustering is then performed to identify spatial clusters of brain regions demonstrating covariation in laterality time series. (c) Temporal clustering of whole-brain laterality patterns, which identifies potential recurring laterality patterns. (d) Illustration of DLI and relevant dynamic laterality measures using an ROI that is left-lateralized. The red curve represents the time series of DLI. The green dotted line is the MLI (0.26), and the blue double arrow denotes the standard deviation of the laterality time series, which measures the level of LFs. The black arrow represents the LR (the change of the sign of lateralization across 2 consecutive time windows). Large magnitude of laterality index indicates segregation at the hemispheric level, while small magnitude of laterality index (near 0) indicates integration across 2 hemispheres. DLI, dynamic laterality index; GS, global signal; GS_L, global signal of the left hemisphere; GS_R, global signal of the right hemisphere; LF, laterality fluctuation; LR, laterality reversal; MLI, mean laterality index; ROI, region of interest.

https://doi.org/10.1371/journal.pbio.3001560.g001

We demonstrate that the GS-based laterality index we adopted is highly similar to a conventional laterality measure autonomy index (AI) [12,33] that is widely used. Using resting-state fMRI data from HCP, we found high correlation coefficient between the whole-brain laterality profile obtained by the mean laterality index (MLI, average across all time windows) and that obtained by AI (Pearson r = 0.64 ± 0.12, all participants have p < 0.05). The MLI has good replicability over 4 sessions involving left and right scans (correlation across sessions: 0.66 ± 0.018); see S1 Fig. Furthermore, GS-based laterality index is efficient in that it only takes 2*n time, compared to the traditional ROI-based method that takes n² time (n being the number of regions), therefore is economic for multiple time windows.

Using this DLI, we characterized the time-averaged laterality of large-scale brain networks in resting state. The left and right hemispheres largely illustrated positive and negative mean laterality, respectively (Fig 2A). In the left hemisphere, regions in the language (MLI = 0.17 ± 0.08) and default mode (MLI = 0.14 ± 0.07) networks showed strong leftward laterality (networks defined by Cole-Anticevic Brain-wide Network Partition (CAB-NP) [34]). In the right hemisphere, the cingulo-opercular (MLI = −0.11 ± 0.07), dorsal attention (MLI = −0.10 ± 0.06), and visual networks (MLI = −0.08 ± 0.08; Fig 2A) showed strong rightward laterality. Frontoparietal network illustrated strong laterality within both hemispheres (left: MLI = 0.14 ± 0.07, right: MLI = −0.15 ± 0.07). Comparatively, bilateral sensorimotor regions (left: MLI = 0.036 ± 0.04, right: MLI = −0.022 ± 0.05) and left visual areas (MLI = −0.025 ± 0.06) depicted relatively weak mean laterality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Dynamic architecture of whole-brain functional lateralization.

(a) MLI, i.e., time average of DLI. (b) LFs, i.e., standard deviation of laterality time series. (c) LR, i.e., the number of switches in the sign of laterality (from positive to negative or vice versa) of all 360 brain regions (averaged across 991 participants), which are rearranged into 12 subnetworks by the CAB-NP. Vis1, Visual1; vis2, Visual2; smn, Somatomotor; con, Cingulo-Opercular network; dan, Dorsal-Attention network; lan, Language network; fpn, Frontoparietal network; aud, Auditory network; dmn, Default Mode network; pmm, Posterior-Multimodal; vmm, Ventral-Multimodal; ora, Orbito-Affective. L, left hemisphere; R, right hemisphere. The underlying data for Fig 2A can be found in S1 Data. The underlying data for Fig 2B can be found in S2 Data. The underlying data for Fig 2C can be found in S3 Data. CAB-NP, Cole-Anticevic Brain-wide Network Partition; DLI, dynamic laterality index; LF, laterality fluctuation; LR, laterality reversal; MLI, mean laterality index.

https://doi.org/10.1371/journal.pbio.3001560.g002

We further characterized the dynamic changes in laterality of a region from 2 different perspectives: the magnitude and the sign of laterality, by LF and LR, respectively; see Fig 1D. LF is defined as the standard deviation of laterality time series. Standard deviation of a time-varying measure is commonly used in dynamic brain network analysis (e.g., DFC), which reflects its variability [24,35–37]. LR specifically refers to the number of zero crossings of laterality (switch between left and right laterality) in 2 consecutive windows. It is also inspired by the dynamic brain network analysis, i.e., a region may change the module it belongs to (named “nodal flexibility” proposed by Bassett and colleagues), which correlated closely with cognitive ability [38,39].

We illustrate laterality time series and these 2 measures using a brain region with high level of mean laterality (0.26, left laterality; Fig 1D). The standard deviation of time-varying laterality index corresponds to moderate changes, or deviations, with respect to the mean laterality. By contrast, LR reflects larger changes in laterality, corresponding to sufficiently extreme deviations from the mean, which results in sign changes in laterality. Moreover, a large magnitude of the laterality index corresponds to segregation at the level of hemispheres, and a small magnitude (near 0) indicates integration across 2 hemispheres; see Fig 1D and “Network and structural basis of laterality dynamics” (Results) for explanation.

Based on the above 2 dynamic laterality characteristics, we found that the laterality of most brain regions varied considerably over time. Specifically, primary visual (LF = 0.46 ± 0.09; LR = 46.8 ± 1.7) and sensorimotor regions (LF = 0.44 ± 0.09; LR = 46.6 ± 1.9) generally illustrated stronger levels of variation in laterality (both LF and LR) than higher-order association regions—including the frontoparietal (LF = 0.43 ± 0.11; LR = 43.4 ± 1.97), language (LF = 0.43 ± 0.10; LR = 43.5 ± 2.1), and default mode networks (LF = 0.41 ± 0.10; LR = 44.7 ± 1.7); see Fig 2B and 2C.

Spatial clustering of the laterality dynamics

After characterizing the temporal LF across the whole brain, we investigated how spatially distributed regions that support different functional specializations correlate with each other in laterality (Fig 1B). Through spatial clustering of the laterality time series across all brain regions, we identified 4 major clusters (Fig 3A): Cluster 1 consisted mainly of regions from the bilateral visual network; Cluster 2 consisted of regions from the bilateral sensorimotor and cingulo-opercular networks; Cluster 3 mainly covered the frontoparietal network (more from the right hemisphere) and attention network; and Cluster 4 mainly consisted of regions from bilateral default mode network, part of the frontoparietal network (mainly the left hemisphere) and regions from the language network (Fig 3A). Clusters 1 and 2 showed higher levels of LF (reflected by higher standard deviation of DLI and LR) when compared to Clusters 3 and 4 (repeated-measures ANOVA, p < 0.001, pairwise t test with p < 0.01; see S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Spatial clustering of laterality dynamics.

(a) Four spatial clusters revealed by clustering of the laterality time series of all 360 brain regions. The word cloud for each cluster indicates the functional networks being involved (based on the CAB-NP), with the font size representing the proportion of each network in the cluster. L, left-lateralized; R, right-lateralized; MLI, mean laterality index; LF, laterality fluctuations; LR, laterality reversal. The mean value of the 3 dynamic laterality measures of the 4 spatial clusters are shown in the inset. (b) The laterality correlation matrix across 360 brain regions (averaged over all 991 participants). (c) The correlation between the 3 dynamic laterality measures (MLI, LF, LR) of the 4 spatial clusters and cognitive performance of 3 tasks. CardSort, the performance of Dimensional Change Card Sort Test; ProcSpeed, the performance of Pattern Comparison Processing Speed Test; LanDiff, mean difficulty of stories for each participant in HCP language task. (d) The association between the laterality correlation within/between clusters and the processing speed (Pattern Comparison Processing Speed Test). Only significant results (Pearson correlation, FDR corrected, FDR q < 0.05, marked by one asterisk *) are shown. The underlying data for this figure can be found in S4 Data. CAB-NP, Cole-Anticevic Brain-wide Network Partition; FDR, false discovery rate; HCP, Human Connectome Project.

https://doi.org/10.1371/journal.pbio.3001560.g003

We then explored the association between laterality dynamics and out-of-scanner cognitive performance. Specifically, we used multiple tasks adopted by HCP that are generally associated with either lateralized (e.g., language: story comprehension task, and attention: Flanker task) or bilateral (e.g., cognitive flexibility: Card Sort task, and working memory: List Sorting task) brain function. First, we found that the LF and LR of the identified spatial clusters correlated in a positive and negative manner, respectively, with cognitive performance. Second, significant correlation was only found for language function (story comprehension tasks), cognitive flexibility (Card Sort Test and pattern comparison task), and processing speed (Pattern Comparison Test) out of all 13 cognitive measures; see S1 Table. Correction for multiple comparisons were performed using the false discovery rate (FDR) method for all 29 laterality measures employed in our analyses (see Methods section for details of these laterality measures). Furthermore, we also adopted a more conservative multiple-comparisons correction scheme for the 13 behavioral measures. That is, based on the FDR correction of the 29 dynamic laterality measures (FDR q < 0.05), we further employed Bonferroni correction for the 13 behavioral measures, i.e., FDR q < 0.05/13 = 0.0038. All the following results were based on the FDR correction for all 29 laterality measures (FDR q < 0.05). Those correlations that passed the stricter correction method (FDR q < 0.05/13) were also marked.

Specifically, LF of Clusters 3 (FPN) and 4 (DMN and language network) correlated positively with the difficulty of stories a participant could understand (language task difficulty, LanDiff, Cluster 4: r = 0.11, p = 0.018; Cluster 3: t = 0.11, p = 0.014; Fig 3C). A cognitive flexibility measure (Dimensional Change Card Sort Test, CardSort) also correlated positively with LF across all clusters (Cluster 1, r = 0.12, p = 0.001; Cluster 2, r = 0.14, p < 0.001; Cluster 3, r = 0.14, p < 0.001; Cluster 4, r = 0.14, p < 0.001; all FDR q < 0.05/13). In contrast, LR of these clusters correlated negatively with cognitive performance: for language task difficulty, r = −0.2 (p < 0.001, FDR q < 0.05/13) for Clusters 3, and r = −0.14 (p = 0.002) for Clusters 4. For Card Sort task, r = −0.15 (p < 0.001, FDR q < 0.05/13) for Cluster 3 and r = −0.15 (p < 0.001, FDR q < 0.05/13) for Cluster 4. In addition, the performance of Pattern Comparison Processing Speed Test (ProcSpeed, the speed of completing the task) also showed positive association with LF of Cluster 3 (r = 0.09, p = 0.009) and negative association with LR of Clusters 3 (r = −0.14, p = 0.001) and 4 (r = −0.12, p = 0.003).

We furthermore explored how time-varying laterality time series of different clusters correlate with each other. Importantly, Cluster 4 showed negative laterality correlations with all other 3 clusters (Cluster 1: t = −81.7; Cluster 2: t = −84.37; Cluster 3: t = −62.1, all p < 0.001; Fig 3B), while Clusters 1, 2, and 3 showed positive correlation among themselves (Cluster 1 and 2, t = 20.4; p < 0.001; Clusters 2 and 3, t = 15.2; p < 0.001), indicating that Cluster 4 shows a tendency to lateralize in the hemisphere opposite to those of the other 3 clusters over time. The negative laterality correlation between Cluster 4 and Clusters 1, 2, and 3 also bear functional significance. We found that individuals with higher Card Sort score showed stronger negative correlation in laterality between Cluster 4 and Cluster 2 (r = −0.09, p = 0.012), and Cluster 4 between Cluster 3 (r = −0.12, p = 0.001, FDR q < 0.05/13); see Fig 3D and S2 Table.

Temporal clustering of laterality dynamics

We further explored the temporal organization of laterality dynamics by clustering whole-brain laterality state of multiple time windows (see Fig 1C and Methods). Three recurring laterality states (Fig 4A) were identified, each showing distinct laterality patterns, different dwelling times, and transition probabilities (Fig 4B). Specifically, State 1 showed a typical leftward laterality in Cluster 4 (t = 77.9, p < 1e-20) and a rightward laterality in Cluster 2 (t = −109.3, p < 1e-20) (Fig 4A and 4B). State 1 is the primary laterality state with the largest fraction of 41% and a mean dwelling time of 49.9 ± 3.58 windows (about 35 s). State 2 showed a typical rightward laterality in Cluster 1 (t = −108, p < 1e-20), and a leftward laterality in Cluster 2 (t = 60.1, p < 1e-20) and Cluster 4 (t = 53.2, p < 1e-20) (Fig 4A and 4B). State 3 showed a leftward laterality in Cluster 1 and 2, and rightward laterality of Cluster 4 (t = −64.7, p < 1e-20) and Cluster 3 (t = −41.4, p < 1e-20). These 2 states showed similar fractions (State 2: 28%; State 3: 31%) and dwelling time (State 2: 10.5 ± 2 s; State 3: 12.1 ± 2.3 s).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Temporal clustering of laterality dynamics.

(a) The 3 recurring laterality states obtained by temporal clustering of the time-varying, whole-brain laterality states (360 regions) with colors indicating the network arrangement of the CAB-NP. L, left hemisphere; R, right hemisphere. (b) The mean fraction of 3 states and the mean probability of switching between them. The bar plot next to each state represents the averaged laterality of the 4 spatial clusters in each state. L, left-lateralized; R, right-lateralized. (c) The correlation between the fraction/mean dwelling time of 3 states and memory ability (mem)/language ability. (d) The correlation between the probability of switching between/within the 3 states and language ability. Only significant results (Pearson correlation, FDR corrected, FDR q < 0.05, marked by one asterisk *) are shown. The underlying data for this figure can be found in S4 Data. CAB-NP, Cole-Anticevic Brain-wide Network Partition; FDR, false discovery rate.

https://doi.org/10.1371/journal.pbio.3001560.g004

We further investigated the behavioral correlates of individual variations in the state transition properties. Results showed a link between State 2 and cognitive functions: Individuals with less State 2 showed higher language task difficulty (fraction, r = −0.1, p = 0.007; dwelling time, r = −0.09, p = 0.015), higher CardSort performance (fraction, r = −0.08, p = 0.014) and higher ProcSpeed score (fraction, r = −0.09, p = 0.01). In addition, CardSort was positively correlated with the fraction (r = 0.1, p = 0.006) and dwelling time (r = 0.1, p = 0.014) of State 1. See Fig 4C and 4D and S3 Table. Multiple comparisons correction was performed using the FDR method for all laterality measures used in the analyses; see Methods section.

Network and structural basis of laterality dynamics

Next, we investigated how dynamic laterality is related to functional and structural brain network properties. First, we investigated the functional correlates of dynamic laterality, i.e., we investigated how dynamic laterality of a cluster is related to the time-varying functional connectivity of the brain. We then calculated the correlation between time-varying laterality index and time-varying network features (including degree, participation coefficient, and modularity, reflecting brain integration/segregation) and BOLD activity (amplitude of low-frequency fluctuation (ALFF)).

We found that the functional connectivity of the high-order brain networks, especially those in the left hemisphere (e.g., FPN, DMN, and language network, mainly covered by Cluster 4) play an essential role in dynamic changes in lateralization. For the 3 right-lateralized clusters (Clusters 1/2/3), higher level of right lateralization was associated mainly with weaker functional connectivity of these clusters with the higher-order networks on the left hemisphere (Fig 5B). For Cluster 4 that is intrinsically left-lateralized, its increased left lateralization was accompanied by reduced functional connectivity with multiple lower-order networks at the right hemisphere, including many of the networks covered by Clusters 1/2/3 (the right subfigure of Fig 5B). Stronger functional connectivity of a cluster with its ipsilateral higher-order networks also plays a role in the increased lateralization of the 4 clusters but has less contribution (Fig 5B; details can be found in S4 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Association between dynamic laterality time series of the 4 identified clusters and the DFC within/between 24 subnetworks.

(a) Analysis pipeline. We established linear regression models of cluster-level DLI time series on subnetwork-level DFC and performed one-sample t test on the regression coefficient (β) to investigate whether the influence of DFC on DLI was significantly greater than 0 (positive coupling) or less than 0 (negative coupling). (b) The association between dynamic laterality time series of Clusters 1/2/3/4 and the DFC within/between 24 subnetworks (according to CAB-NP). Only functional connectivity with very significant t-value of regression coefficient (with p < 1e-200 or |t| > 40, one-sample t test) is being shown since there are too many significant connections. Here, a positive laterality index indicates left lateralization. Colors of points indicate the network arrangement of the CAB-NP. L, left hemisphere; R, right hemisphere. The underlying data for this figure can be found in S4 Data. CAB-NP, Cole-Anticevic Brain-wide Network Partition; DFC, dynamic functional connectivity; DLI, dynamic laterality index; aud, Auditory network; con, Cingulo-Opercular network; dan, Dorsal-Attention network; dmn, Default Mode network; fpn, Frontoparietal network; lan, Language network; ora, Orbito-Affective; pmm, Posterior-Multimodal; smn, Somatomotor; Vis1, Visual1; vis2, Visual2; vmm, Ventral-Multimodal.

https://doi.org/10.1371/journal.pbio.3001560.g005

For functional network features and ALFF, we found that the whole-brain lateralization (the mean absolute value of DLI across all 360 regions) correlated negatively with the averaged degree (r = −0.43 ± 0.1, 91% p < 0.05), participation coefficient (r = −0.2 ± 0.12, 81% p < 0.05), interhemispheric connection (r = −0.52 ± 0.11, 92% p < 0.05), and ALFF (r = −0.22 ± 0.12, 85% p < 0.05), while showing positive correlations with modularity of the brain network (r = 0.47 ± 0.06, 91% p <0 .05) across time windows; see S5 Fig. This indicates that when the whole brain is highly segregated (modular), it also demonstrates high laterality, while an integrated state of the whole brain (low modularity) are accompanied by low laterality. At the local level, we found that these correlation patterns are most prominent for the high-order brain regions (most significantly in default mode network, the language network, and the frontoparietal network; S5 Fig).

Second, we explored the structural correlates of laterality dynamics using diffusion MRI data. We used diffusion imaging data from a subset of HCP S1200 (HCP Unrelated 100) to constructed 2 kinds of structural connectivity matrices, fractional anisotropy (FA) matrix and fiber number (FN) matrix for each participant. We found significant positive correlations between the laterality correlation matrix and structural connectivity matrix, including FA matrix (Spearman’s rho = 0.14 ± 0.02, significant in 99% participants) and FN matrix (Spearman’s rho = 0.14 ± 0.02, significant in 99% participants); see Fig 6A. In line with this, homotopic regions generally showed positive correlations in their laterality time series (mean r = 0.39 to 0.79; S2 Fig) due to the critical role of corpus callosum [40]. These results suggested that brain regions with similar laterality dynamics tend to have stronger structural connection. Furthermore, LF of a region correlated positively with its FA/FN-degree, i.e., the sum of FA/FN of fibers between this region and all other regions (LF-FN: rho = 0.32 ± 0.09, significant in 99% participants; LF-FA: rho = 0.28 ± 0.08, significant in 99% participants; see Fig 6B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Structural basis of dynamic laterality.

(a) Laterality correlation matrix (left half-matrix) and structural connection matrix (right half-matrix) demonstrates high level of correlation. These 2 matrices are obtained by averaging across all 99 participants. L, left hemisphere; R, right hemisphere. The color ribbons indicate the network arrangement of CAB-NP detailed in Fig 2. (b) Upper: the relationship between a structural measure (FN-degree) and the LF across all brain regions. Lower: the relationship between another structural measure (FA-degree) and the LF. The large scatter maps show the association between the averaged map of dynamic laterality measures and the averaged FA-/FN-degree map (both across all 99 participants), and the inset shows the distribution of the individual correlation coefficients between the FA-/FN-degree map and laterality fluctuation map of each participant. The underlying data for this figure can be found in S4 Data. CAB-NP, Cole-Anticevic Brain-wide Network Partition; FA, fractional anisotropy; FN, fiber number; LF, laterality fluctuations.

https://doi.org/10.1371/journal.pbio.3001560.g006

Heritability of laterality dynamics

Finally, we investigated heritability of lateralization dynamics by twin analyses (using the kinship information). First, we estimated heritability through the ACE model [additive heritability (A), common (C), and specific (E) environmental factors model]. We found that LF showed the highest heritability (h²) among all dynamic laterality measures (h² = 0.09 to 0.48 for LF; h² = 0.002 to 0.41 for MLI; h² = 0.001 to 0.31 for LR; see Fig 7A–7C). Of all 4 spatial clusters, Cluster 4 (e.g., the default mode network, part of the right frontoparietal network, and the language network) showed the highest heritability (MLI h² = 0.19; LF h² = 0.4; LR h² = 0.06), followed by Cluster 3 (MLI h² = 0.19; LF h² = 0.36; LR h² = 0.06. The heritability of Cluster 2 (MLI h² = 0.13; LF h² = 0.35; LR h² = 0.03) and Cluster 1 (MLI h² = 0.12; LF h² = 0.33; LR h² = 0.03) were relatively low.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Heritability of dynamic laterality.

(a-c) Heritability analysis of MLI, LF, and LR. Upper: Heritability (h²) of each brain region. Only the regions with h² of p < 0.05 are retained. Lower left: The h² and common environmental factors (c²) for each cluster, and each data point represents a brain region. Lower right: The cosine distance between whole brain map for each pair of MZ twins, DZ twins, SI, and UN. Asterisk indicates the significant difference between the groups (one-way ANOVA). (d) Heritability of lateralization correlation matrix. Left: Average h² and c² in 4 clusters. Upper right: Differences in heritability between the positive and the negative links, i.e., links with positive or negative correlation in laterality. Lower right: The cosine distance of laterality correlation matrix of each pair of MZ, DZ, SI, and UN. The error bar represents the mean value ±1 SD. One asterisk (*), p < 0.05; 2 asterisks (**), p < 0.01; 3 asterisks (***), p < 0.001; ns, nonsignificant. Pos, positive correlation; neg, negative correlation. The underlying data for this figure can be found in S4 Data. DZ, dizygotic; LF, laterality fluctuations; LR, laterality reversal; MLI, mean laterality index; MZ, monozygotic; SI, sibling; UN, unrelated group.

https://doi.org/10.1371/journal.pbio.3001560.g007

We then calculated the cosine distance between each pair of monozygotic (MZ) twins, dizygotic (DZ) twins, sibling (SI), and unrelated group participants using various dynamic laterality measures across the whole brain. As expected, the laterality correlation matrix showed greater similarity within MZ twins than those within DZ or non-twins (one-way ANOVA, F = 152.7, p < 0.0001; Fig 7D). Similarly, the similarity of MLI and LF (of all regions) in twins, SIs, and unrelated participants showed a decreasing trend, although the difference between MZ and DZ was not significant (for MLI, F = 23.8, p < 0.0001, MZ versus DZ, p > 0.99; for LF, F = 10.8, p < 0.0001, MZ versus DZ, p = 0.99; Fig 7A and 7B). There was no significant difference in LR between the 4 groups (F = 1.48, p = 0.22; Fig 7C). All one-way ANOVA were done using GraphPad Prism 9.3.1 (http://www.graphpad.com). In summary, dynamic laterality measures were generally heritable.

Validation analysis

To rule out the possible influence of potential nuisance variables on our dynamic laterality results, we performed a series of validation analyses: (1) the effect of sample size: all 991 participants were split into two halves, and the main analyses were repeated separately on both halves of the data (S10 Fig); (2) head movements: we repeated the main analyses with multiple head movement parameters (Friston 24 parameters [41]) being regressed out from the BOLD signal (S12 Fig). We also used more stringent exclusion criteria (mean frame distance (FD) < 0.15 mm and FD < 0.1 mm) to repeat our analyses (S12 Fig); (3) GS: we repeated the main analyses, while regressing out the GS of the whole brain (S11 Fig). This resulted in the sum of the GS from the left and right hemispheres being 0, i.e., when the GS_L is positive, the GS_R is equal to its negative number (rather than a GS of 0 for one hemisphere). Therefore, the correlation coefficients could still be calculated between hemispheric GS and individual ROIs; (4) time window lengths: another 2 window lengths (60 TRs and 90 TRs) were used to investigate the reproducibility of DLI (S13 and S14 Figs); and (5) the potential contaminating effect of an ROI to its ipsilateral GS: when we calculated the correlation coefficient between an ROI and the GS of its ipsilateral hemisphere, the ROI was included in this calculation, which may have “contaminated” its ipsilateral GS. We therefore removed the ROI when extracting the ipsilateral hemisphere GS and repeated our analyses to examine whether our results are affected (S15 Fig). In all the above cases, our main results, including the distribution of dynamic laterality indices across the whole brain, and their correlation with cognitive performance, were largely replicated, which indicated the reliability of the DLIs. See more details in S1 Text.

Limitations

In this study, resting-state data were analyzed, so it was not possible to directly relate the results to specific cognitive processes. Future studies should investigate dynamic changes in brain lateralization under task modulation, particularly with multiple task states or various task loads, to better understand the specific cognitive advantages of dynamic brain lateralization. In addition, while the GS-based laterality index has the advantages of being computationally convenient and highly correlated with the traditional ROI-based laterality index, it has the problem of averaging and being too coarse to detect network interactions. Therefore, we further analyzed which specific functional connectivity or network drives the temporal changes of laterality index of the 4 clusters. Finally, age and handedness are important factors potentially affecting lateralization, and handedness is also partly controlled by genetic factors [60]. Future studies are warranted that explore how cognitive deterioration with aging may affect lateralization dynamics.

Ethics statement

This paper utilized data collected for the HCP. The scanning protocol, participant recruitment procedures, and informed written consent forms, including consent to share deidentified data, were approved by the Washington University institutional review board [31]. Data acquisition for the HCP was approved by the Institutional Review Board of The Washington University in St. Louis (IRB # 201204036), and all open access data were deidentified. The data collected by the HCP adhered to the principles of the Declaration of Helsinki. No experimental activity with any involvement of human participants took place at the author’s institutions. Our data analysis was performed in accordance with ethical guidelines of the Fudan University ethics committee.

Dataset

We analyzed multimodal brain imaging data and behavioral measures from the HCP 1200 Subjects Release (S1200) [31]. All brain imaging data were acquired using a multiband sequence on a 3-Tesla Siemens Skyra scanner. For each participant, 4 resting-state fMRI scans were acquired: 2 with right-to-left phase encoding and 2 with left-to-right phase encoding direction (1,200 volumes for each scanning session, TR = 0.72 s, voxel size = 2 × 2 × 2 mm). High-resolution T1-weighted MRI (voxel size = 0.7 × 0.7 × 0.7 mm) and diffusion MRI (voxel size = 1.25 mm, 3 shells: b-values = 1,000, 2,000, and 3,000 s/mm², 90 diffusion directions per shell) were also acquired for each participant. A total of 991 participants (28.7 ± 3.7 years old, 528 females) were included in the final cohort utilized in this study according to the following criteria: (1) completed all 4 resting-state fMRI scans; (2) limited in-scanner head motion (mean FD < 0.2 mm); (3) same number of sampling points (i.e., 1,200 volumes per scanning session); (4) without any missing data in regions included in the employed parcellation scheme.

Resting-state fMRI preprocessing

Resting-state fMRI data were preprocessed using the HCP minimal preprocessing pipeline (fMRIVolume) [61] and were denoised using the ICA-FIX method [62]. Then, a spatial smoothing (FHWM = 4 mm) and a low-pass filtering (0.01 Hz to 0.1 Hz) were performed. We did not employ GS regression, as the mean hemispheric time series were used for calculating the lateralization index [14]. The whole brain was parcellated into 360 regions (180 for each hemisphere), using the HCP’s multimodal parcellation (HCP MMP1.0) [32]. These regions were grouped into 12 functional networks based on the CAB-NP [34].

Dynamic laterality index

The pipeline for the DLI analysis is illustrated in Fig 1A. The DLI adopted a sliding window approach and a traditional laterality index [14] to capture the dynamics of laterality. DLI at the tth sliding time window is defined by: where ROIi indicates BOLD time series of ROI i, and GS_L and GS_R indicate the GS, i.e., averaged time series of the voxels within the left and the right hemispheres, respectively. Pearson correlation coefficient r is Fisher-z-transformed. There are 3 main reasons why we adopted the hemispheric GS-based laterality: (1) GS has been shown to be able to reflect and characterize laterality by previous studies [14], and the computational complexity is very low (n*2, n being the number of regions); (2) the whole-brain laterality pattern obtained using hemispheric GS is highly correlated to that obtained by ROI-based method like AI defined as follows [12,13,33,63]: where and indicate the averaged functional connectivity between BOLD time series of ROI_i and all ROIs in the left and right hemisphere, respectively (S1 Fig), and n and m are number of regions in left and right hemisphere.

Spatial clustering of laterality time series across the whole brain

To explore the spatial organization of dynamic lateralization across the whole brain, we performed spatial clustering on the laterality correlation matrix (obtained by calculating Pearson correlation between laterality time series of all pairs of brain regions). Specifically, we applied the Louvain community detection algorithm on the laterality correlation matrix averaged over all 991 participants and 4 runs. We set the γ parameter to 1, ran the algorithm for 100 times, and reported the cluster partitioning with the maximum modularity parameter (Q).

Temporal clustering of whole-brain laterality state

To identify the recurring laterality states of the whole brain from 991 participants and 1,171 time windows, we adopted a 3-stage temporal clustering approach. Firstly, the 1,171 × 4 time windows of the 4 runs of each participant were clustered into 10 states by k-means clustering. Second, all 9,910 states (991 participants × 10 states) were clustered into 1,000 groups by a second level k-means clustering. Finally, we used Arenas-Fernandez-Gomez (AFG) community detection to cluster the 1,000 groups obtained in the second step. AFG allows for multiple resolution screening of the modular structure and a data-driving selection of clustering number. We determined the optimal number of clusters using the modularity coefficients obtained by changing the resolution parameter from 0.1 to 1.5 in steps of 0.1 [64]. The code of K-means and AFG clustering were from MATLAB function kmeans (with the cosine distance metric) and MATLAB Community Detection Toolbox (CDTB v. 0.9, https://www.mathworks.com/matlabcentral/fileexchange/45867-community-detection-toolbox) [65], respectively. Based on the final clustering, we took the average pattern of each category as centroids and reclassified all the windows of each participant according to the cosine distance between each window and each centroid.

Cognitive performance in behavioral tasks

To understand the cognitive significance of the dynamic brain laterality, we utilized individual performance of a wide variety of cognitive tasks from the NIH Toolbox for Assessment of Neurological and Behavioral function and Penn computerized neurocognitive battery [66]. These involve story comprehension task that is more left lateralized, Flanker Task (inhibitory control and attention) that maybe more right lateralized, and Card Sort (cognitive flexibility) and List Sorting task (working memory) that tend to more bilateral.

The story comprehension task [67] includes a story condition that presents brief auditory stories adapted from Aesop’s fables followed by a binomial forced-choice question to check the participants’ understanding of the story topic (For example, after a story about an eagle that saves a man who had done him a favor, participants were asked, “Was that about revenge or reciprocity?”), and a math condition to answer addition or subtraction problems. To ensure similar level of difficulty across participants, math trials automatically adapted to the participants’ responses. The story and math trials were matched in terms of auditory, duration, attention demand, and phonological input. We used 3 relevant measures: “LanAcc” (“Language_Task_Story_Acc”, the accuracy in answering questions about the story), “LanRT” (“Language_Task_Story_Median_RT”, median response time to answer the questions), and “LanDiff” (“Language_Task_Story_Avg_Difficulty_Level”, the average difficulty of all stories a participant can understand).

Correlation between dynamic laterality index and cognitive performance

In calculating the correlation between dynamic laterality measures and cognitive performance (unadjusted scores), we used Permutation Analysis of Linear Models (PALM; http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PALM) to correct for the bias of significance estimation due to the kinship among participants. PALM used exchangeability blocks that is widely adopted to control the family structure–related bias in HCP data [68,69]. We calculated the Pearson correlation coefficients between dynamic lateralization attributes [MLI, LF, LR of 4 clusters; laterality correlation within and between 4 clusters; fraction and dwelling time of 3 states, state transitions] and cognitive performances, regressing out sex, age, education years, race, body mass index (BMI), handedness, gray matter volume, white matter volume, and head motion (mean FD). We obtained p-values for all correlation coefficients using 5,000 times permutation tests. FDR (q = 0.05) was used for multiple comparison correction (multiply comparison times = 4 MLI + 4 LF + 4 LR + 10 DLI correlation + 3 dwelling time +3 fraction + transition = 29). We show all the results of FDR q < 0.05. Moreover, we also highlight results that survive a more conservative multiple comparisons correction (FDR q < 0.05/13 = 0.0038, where 13 is the number or behavioral measurements used in this study).

Association between dynamic laterality and dynamic functional connectivity

To identify which functional interactions may play important roles in the dynamic change of laterality, we adopted a regression-based approach to investigate which subnetworks (12 pairs of symmetric networks in CAB-NP) may affect the dynamic laterality of the 4 identified clusters. We averaged the functional connectivity within/between the 24 subnetworks; therefore, there are 24 * 23 / 2 = 276 subnetwork-level functional connectivity. Specifically, we built linear regression models of cluster-level DLI time series on subnetwork-level DFC and performed one-sample t test on the regression coefficient (β), to investigate whether the influence of DFC on DLI was significantly greater than 0 (positive coupling) or less than 0 (negative coupling) (Fig 5A).

Association between dynamic laterality and dynamic network properties and amplitude of low-frequency fluctuation

To understand the neural and network basis of dynamic brain lateralization, we correlated the laterality time series of each brain region with time-varying brain network measures (which reflect the organization of the brain like modular, or integration property) and the ALFF averaged over the whole brain for each participant. We constructed functional brain network using Pearson correlation within each sliding time window and computed the following topological measures:

1. Degree centrality (DC) [70,71], measuring the centrality of each node in the network: where G is a graph with the node i and j connected by edge A_ij.
2. Participation coefficient (B_T) measuring the contribution of each brain region to whole-brain integration: where κ_isT is the strength of the positive connections of region i to regions belong to the module s at time T; κ_iT is the sum of strengths of all positive connections of region i at time T. The participation coefficient of a region is close to 1 if its connections are distributed among all of the modules and 0 if all of its links are within its own module. Community division was obtained by community detection in each time window, and then the participation coefficient was calculated [25].
3. Modularity (Q) measuring the tendency of the network to segregate into several independent modules [72,73]: where m is total number of edges in the network; A is adjacent matrix; k_i is degree of node i; C_i is community assignment of node i; when and only when C_i = C_j, δ(C_i,C_j) = 1, otherwise, δ(C_i,C_j) = 0.
4. Interhemispheric connection [74], measuring the averaged interhemispheric connection strength.

Among these indicators, the averaged DC is measured at the global network level, B_T and Q are measured at the module level, and interhemispheric connection is measured at the hemispheric level DC, B_T and Q were calculated using the Brain Connectivity Toolbox (BCT; http://www.brain-connectivity-toolbox.net/) [75]. We removed all negative connections in line with the general assumptions of graph theory [71].

ALFF measures spontaneous energetic activity within each time window [76] calculated as follows: where fft is Fourier transform; abs is operator of taking absolute value; BOLD is BOLD signal within a time window in any region; and N is the length of time window. At global level, we calculated the Pearson r between the ALFF/network time series (averaged across the brain) and the absolute laterality time series (averaged across the brain). At local level, we correlated the ALFF/network time series (averaged across the brain) with the absolute laterality time series of each brain region.

Structural network analysis

Because the probabilistic fiber tracking is very time-consuming, diffusion tensor imaging (DTI) data of a subset of HCP S1200 release, the “100 Unrelated Subjects” (n = 100, 54 females, mean age = 29 years) was used for the structural analysis (1 participant was removed due to head motion). Data underwent motion, susceptibility distortion, and eddy current distortion correction [61,77]. We used FMRIB Software Library v6.0 (FSL; https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) and MRtrix3, a toolkit for diffusion-weighted MRI analysis (https://mrtrix.readthedocs.io/en/dev/) [78], to construct the SC matrix. The processing steps were as follows: (1) tissue segmentation based on T1-weighted structural image; (2) calculation of 4D images with gray matter, white matter, and cerebrospinal fluid using multi-shell, multi-tissue constrained spherical deconvolution [79]; (3) generation of white matter–constrained tractography using second-order integration over fiber orientation distributions (iFOD2), a probabilistic tracking algorithm (1,000,000 streamlines for each participant, max length = 250 mm, cutoff = 0.06) [80]; (4) use of FSL’s FNIRT to reverse-register the MMP parcellation to individual space; (5) calculation of FA maps using dtifit of FSL; and (6) measurement of the average FA on all the streamlines connecting any 2 parcels. This process finally resulted in 360 × 360 FA matrices and FN matrices for 99 participants. We calculated the FA-degree and FN-degree based on this structural connectivity (SC) matrix, i.e., the sum of FA and FN of each row of the SC matrix. The Spearman correlation between laterality correlation matrix and SC matrix and the correlation between MLI/LF/LR and FN-/FA-degree were calculated for each participant.

Heritability analysis

We used twin information in the HCP data to evaluate the heritability of the dynamic laterality measures. In the 991 participants, 103 pairs of MZ twins, 98 pairs of DZ twins, 195 pairs of SI were included, and their zygosity information was used to fit ACE model, which is able to divide total variance in a phenotype (σ²) into 3 additive parts: (A) additive common genetic factor (heritability), (C) common environment, and (E) unshared environment, or other source of measurement error [81]:

The ratio of the variance explained by A, C, and E to the total variance was used as estimates of heritability (h²), environmental factors (c²), and error (e²), all of which sum to 1. All 4 models (for MLI, LF, LR, and laterality correlation matrix) were fitted and estimated using APACE (the Advanced Permutation inference for ACE models; www.warwick.ac.uk/tenichols/apace) [82]. Before model fitting, sex, age, education years, race, BMI, handedness, gray matter volume, white matter volume, and mean FD were regressed out from phenotype as covariables. The significance of heritability for each brain region was obtained using permutation test (1,000 permutations per region).

Code availability

The code used in this study to calculate DLI is available on https://github.com/XinRanWu/Dynamic_Laterality.