Dynamic causal brain circuits during working memory and their functional controllability

Abstract

Control processes associated with working memory play a central role in human cognition, but their underlying dynamic brain circuit mechanisms are poorly understood. Here we use system identification, network science, stability analysis, and control theory to probe functional circuit dynamics during working memory task performance. Our results show that dynamic signaling between distributed brain areas encompassing the salience (SN), fronto-parietal (FPN), and default mode networks can distinguish between working memory load and predict performance. Network analysis of directed causal influences suggests the anterior insula node of the SN and dorsolateral prefrontal cortex node of the FPN are causal outflow and inflow hubs, respectively. Network controllability decreases with working memory load and SN nodes show the highest functional controllability. Our findings reveal dissociable roles of the SN and FPN in systems control and provide novel insights into dynamic circuit mechanisms by which cognitive control circuits operate asymmetrically during cognition.

Fig. 1. Schematic illustration of data analysis strategy and procedures.

A Salience Network (SN), Frontal-Parietal Network (FPN), and Default Mode network (DMN) ROIs: (1) left anterior insula (lAI); (2) right anterior insula (rAI); (3) dorsomedial prefrontal cortex (DMPFC); (4) left middle frontal gyrus (lMFG); (5) right middle frontal gyrus (rMFG); (6) left frontal eye field (lFEF); (7) right frontal eye field (rFEF); (8) left intraparietal lobule (lIPL); (9) right intraparietal lobule (rIPL); (10) posterior cingulate cortex (PCC), and (11) ventromedial prefrontal cortex (VMPFC). B General linear model analysis revealed significant working-memory-load-related activation in SN and FPN (task-positive nodes), and deactivation in the DMN (task-negative nodes) (p < 0.01, FDR-corrected, two-sided t-test). N = 737 participants. Data are presented as mean ± SEM. C Overview of data analysis pipeline. We first screened the HCP n-back working-memory dataset based on head motion, behavioral performance, and participant handedness. We then extracted time series from each of the 11 ROIs and applied MDSI to determine working-memory load-specific dynamic causal interactions from each participant in the 2-back and 0-back task conditions. MDSI-derived causal influences were then used to investigate (1) whether multivariate dynamic causal interaction patterns distinguished 2-back versus 0-back task conditions, (2) task-dependent causal outflow from each ROI, (3) network controllability as a function of working-memory load, and (4) the relationship between the strength of dynamic causal interactions and behavioral performance. HCP human connectome project, MDSI multivariate dynamic state-space systems identification, ROI region of interest. Source data are provided as a Source data file.

Fig. 2. Working-memory load-specific dynamic causal influences.

A MDSI and graph-theoretic analyses identified four communities associated with directed causal influences in both the 2-back and 0-back working-memory task conditions: (i) SN consisting of lAI, rAI, and DMPFC nodes, (ii) left FPN consisting of lMFG, lFEF, and lIPL nodes, (iii) right FPN consisting of rMFG, rFEF and rIPL nodes, and (iv) DMN consisting of PCC and VMPFC nodes (p < 0.01, FDR-corrected, two-sided t-test). N = 737 participants. Yellow cells indicate that a pair of ROIs are grouped into the same community and cyan cells indicate that a pair of ROIs belong to different communities. B Significant directed causal influences between SN, FPN, and DMN ROIs in the 2-back and 0-back working-memory task conditions (p < 0.01, FDR-corrected, two-sided t-test). N = 737 participants. Red cells indicate significant positive influences and blue indicates significant negative influences. C Stability analyses revealed highly stable multivariate patterns of causal influences among SN, FPN, and DMN nodes in 2-back and 0-back task conditions (r > 0.8 for sample size >25). X-axis shows sample sizes ranging from 20 to 600. Y-axis shows stability, computed as the correlation of multivariate causal influence patterns between the original sample and random subsamples drawn from N = 737 participants. lAI left anterior insula, rAI right anterior insula, DMPFC dorsomedial prefrontal cortex, lMFG left middle frontal gyrus, rMFG right middle frontal gyrus, lFEF left frontal eye field, rFEF right frontal eye field, lIPL left intraparietal lobule, rIPL right intraparietal lobule, PCC posterior cingulate cortex, VMPFC ventromedial prefrontal cortex. Source data are provided as a Source data file.

Fig. 3. Working-memory load-specific dynamic causal outflow.

A AI showed the highest directed causal outflow between SN, FPN, and DMN nodes in both the 2-back and 0-back working-memory task conditions (p < 0.05, FDR-corrected, two-sided t-test). n = 737 participants. Data are presented as mean ± SEM. In contrast, the rMFG showed the highest directed causal inflow among all nodes in both task conditions (p < 0.05, FDR-corrected, two-sided t-test). n = 737 participants. Data are presented as mean ± SEM. B Stability analyses revealed highly stable directed causal outflow from the rAI and directed causal inflow into the rMFG in both the 2-back and 0-back working-memory task conditions. X-axis shows sample size, ranging from 20 to 600. Y-axis shows stability, computed as the probability that the rAI shows the highest positive directed causal outflow among SN, FPN, and DMN nodes, and the probability that the rMFG shows the highest causal inflow in random subsamples drawn from N = 737 participants. lAI left anterior insula, rAI right anterior insula, DMPFC dorsomedial prefrontal cortex, lMFG left middle frontal gyrus, rMFG right middle frontal gyrus, lFEF left frontal eye field, rFEF right frontal eye field, lIPL left intraparietal lobule, rIPL right intraparietal lobule, PCC posterior cingulate cortex, VMPFC ventromedial prefrontal cortex. Source data are provided as a Source data file.

Fig. 4. Dynamic causal influences distinguish working-memory conditions.

Linear SVM analyses with 10-fold cross-validation revealed that dynamic causal influences between SN, FPN, and DMN nodes distinguished the 2-back and 0-back working-memory task conditions (ps < 0.01, permutation test).

Fig. 5. Working-memory load-dependent dynamic causal influence and net outflow.

A MDSI analysis revealed links with significantly greater directed causal influences between SN, FPN, and DMN nodes in the 2-back, compared to the 0-back, working-memory task condition (p < 0.05, FDR-corrected, two-sided t-test). n = 737 participants. B The DMPFC and lIPL showed significantly higher directed causal outflow in the 2-back, compared to the 0-back, task condition. In contrast, the PCC, rFEF and rIPL showed significantly higher directed causal inflow in the 2-back, compared to the 0-back, task condition (p < 0.05, FDR-corrected, two-sided t-test). n = 737 participants. Data are presented as mean ± SEM. C Stability analyses revealed highly stable multivariate patterns of causal influences between SN, FPN, and DMN nodes in 2-back versus 0-back (r > 0.8 with samples >200). X-axis is the subsample sizes, ranging from 20 to 600. Y-axis is the stability measures, which is the correlation of multivariate causal interaction patterns between subsamples and original dataset. lAI left anterior insula, rAI right anterior insula, DMPFC dorsomedial prefrontal cortex, lMFG left middle frontal gyrus, rMFG right middle frontal gyrus, lFEF left frontal eye field, rFEF right frontal eye field, lIPL left intraparietal lobule, rIPL right intraparietal lobule, PCC posterior cingulate cortex, VMPFC ventromedial prefrontal cortex. Source data are provided as a Source data file.

Fig. 6. Dynamic causal influence relation to behavioral performance.

A Canonical correlation analysis revealed a significant relationship between directed SN, FPN, and DMN causal influences and behavioral performance in the 2-back working-memory task condition (r = 0.46, p < 0.001, Pearson’s correlation). B Correlation coefficients contributing to brain-behavior relations highlights positive influences between SN and FPN nodes and negative influences of SN and FPN nodes on PCC and VMPFC nodes of the DMN. lAI left anterior insula, rAI right anterior insula, DMPFC dorsomedial prefrontal cortex, lMFG left middle frontal gyrus, rMFG right middle frontal gyrus, lFEF left frontal eye field, rFEF right frontal eye field, lIPL left intraparietal lobule, rIPL right intraparietal lobule, PCC posterior cingulate cortex, VMPFC ventromedial prefrontal cortex. Source data are provided as a Source data file.

Fig. 7. Working-memory load-dependent functional controllability in SN, FPN, and DMN.

A Functional network controllability, assessed in each brain node, was significantly lower in the 2-back, compared to the 0-back, working-memory task condition (p < 0.001, two-sided t-test). n = 737 participants. Data are presented as mean ± SEM. AI and DMPFC nodes in the SN have significantly higher controllability than FPN and DMN nodes, except for lFEF and lMFG nodes of the FPN in the 2-back task condition and lFEF in the 0-back task condition (p < 0.001, two-sided t-test). n = 737 participants. Data are presented as mean ± SEM. B Functional network controllability, assessed across SN, FPN, and DMN nodes, was significantly lower in the 2-back, compared to the 0-back, working-memory task condition (p < 0.001, two-sided t-test). N = 737 participants. Data are presented as mean ± SEM. The SN shows significantly higher controllability than the FPN (p = 4.61e−06, two-side paired t-test) and DMN (p = 9.84e−05, two-side paired t-test). N = 737 participants. Data are presented as mean ± SEM. C Stability analyses revealed stable load effect (0-back > 2-back) and network difference (SN > FPN and SN > DMN). X-axis shows sample size, ranging from 20 to 600. Y-axis shows stability, computed as the probability that the load effect of controllability is significantly different between 2-back and 0-back working-memory conditions, and the probability that the SN shows greater controllability than the FPN and DMN in both 2-back and 0-back working-memory conditions, in random subsamples drawn from N = 737 participants. lAI left anterior insula, rAI right anterior insula, DMPFC dorsomedial prefrontal cortex, lMFG left middle frontal gyrus, rMFG right middle frontal gyrus, lFEF left frontal eye field, rFEF right frontal eye field, lIPL left intraparietal lobule, rIPL right intraparietal lobule, PCC posterior cingulate cortex, VMPFC ventromedial prefrontal cortex. Source data are provided as a Source data file.