A local source for a network-level switch
Working memory requires activity across many brain areas, and the frontoparietal (FP) and default-mode (DM) networks are known to change their engagement between rest and task. What has been missing is a link between that large-scale reorganisation and the excitatory and inhibitory neurons inside one small region. This study asks whether the balance of excitation and inhibition within the left and right DLPFC could be such a link.
The authors frame the FP and DM networks as competing during working memory: the FP network supports processing of external stimuli and executive control, while the DM network integrates internal and external information. Their hypothesis is that a differential balance of excitatory and inhibitory populations in the right versus left DLPFC is associated with the engagement and disengagement of these two networks through structural connections.
A model fitted only to resting-state data
The core tool is an excitation and inhibition-balanced mean-field model (EI-MFM). Each cortical area is described by ordinary differential equations that include local recurrent excitation, local inhibition, long-range excitatory input arriving through structural connections, and a model of how neuronal activity becomes a blood oxygen level-dependent (BOLD) signal. Because the model separates excitatory from inhibitory populations, the authors can raise or lower one population's activity in a chosen area and observe the simulated consequence.
The model was fitted to resting-state functional connectivity and functional connectivity dynamics from Human Connectome Project data, using 1052 participants split into training (351), validation (350) and test (351) sets. Fitting was done separately for the Desikan-Killiany and Schaefer 100 parcellations, with a Glasser parcellation used later as a robustness check. Parameters were not set freely for every region; they were expressed as linear combinations of the principal resting-state functional connectivity gradient, a T1w/T2w myelin estimate, and a constant, which keeps the parameter space lower-dimensional.
One modelling choice matters for interpretation. The authors compared versions with and without long-range feedforward inhibition and found greater disagreement with empirical functional connectivity and dynamics when that component was included, so they retained the version without it. The fitted parameters still varied across cortex, which the authors describe as regional heterogeneity.
Left excitation with right inhibition moves the network
The central simulation result comes from adding a constant offset to excitatory or inhibitory populations in the DLPFC and asking whether the simulated functional connectivity becomes less like rest and more like working memory. Perturbing a single area did not produce a condition that was consistently significant across parcellations. Perturbing both DLPFC areas did.
Among the tested combinations, adding 0.1 to left excitatory activity together with 0.1 to right inhibitory activity (L-E +0.1 & R-I +0.1) produced the robust shift. Under this condition the simulated connectivity became less similar to resting-state connectivity (Desikan-Killiany: Cohen's d = −0.65 for whole brain and −0.43 for FP–DM; Schaefer 100: d = −0.11 for whole brain and −0.36 for FP–DM) and more similar to working-memory connectivity (Desikan-Killiany: d = 1.24 for whole brain and 1.20 for FP–DM; Schaefer 100: d = 3.36 for whole brain and 0.31 for FP–DM). A smaller right-side offset (L-E +0.1 & R-I +0.05) produced a similar direction of change, and the Glasser parcellation reproduced the pattern.
These are simulated effect sizes from a model, not measured effects in people. The asymmetry is the paper's specific claim: the shift depended on perturbing the left and right DLPFC in opposite directions, not on raising or lowering the same population in both hemispheres.
DLPFC changes covary with both networks
The authors then tested the prediction in the empirical fMRI data. They measured regional BOLD fluctuation amplitude in the DLPFC and in the FP and DM networks, took the working-memory-minus-rest difference, excluded DLPFC voxels from the two networks to avoid circularity, and regressed the network changes on the DLPFC changes across subjects. They report a significant correlation between functional changes in the DLPFC and those in both networks.
A second analysis used functional connectivity rather than amplitude. For each subject the authors computed the connectivity between each DLPFC area and each network, then subtracted resting-state values from working-memory values, and separately subtracted error-trial values from success-trial values in the 2-back condition. These connectivity differences also correlated with the DLPFC changes, and the authors report that the connectivity changes between DLPFC and the two networks correlated with their structural connections, which they interpret as a structural origin for the functional changes.
The distinction between the two measures matters. Regional BOLD fluctuation amplitude indexes how much a region's signal fluctuates, while functional connectivity indexes the correlation between regions; the paper uses both, and neither alone establishes that DLPFC activity causes the network change.
Receptor maps give convergent, not identical, support
To add a molecular angle, the authors used positron emission tomography (PET) density maps of GABA-A and NMDA receptors together with fMRI time series to build receptor-enriched functional maps. These maps reflect the spatial correspondence between fMRI activity and receptor-density maps; the authors explicitly state they are not direct measures of receptor activity or neurotransmitter release.
The analysis compared working memory with resting state and reported differential NMDA- and GABA-A-enriched functional signals within the DLPFC. Spatial permutation tests that rotated the receptor maps on the cortical surface 1000 times were used to check that the findings were not explained by nonspecific spatial smoothness or macroscale cortical gradients.
The authors are careful about what this adds. They describe the model perturbation and the receptor-enriched results as complementary but distinct: the model concerns dynamical excitatory and inhibitory population activity, whereas the receptor analysis reflects receptor-density-enriched functional signals rather than fitted structural excitation/inhibition parameters. The two lines of evidence point in the same direction without measuring the same quantity.
Where the causal language stops
The structural connections in the model are undirected, and the empirical analyses are correlational. The sustained additive offset used in the perturbation is described by the authors as a simplified, model-based manipulation of a local excitatory/inhibitory state shift, not as a transient or physiological input. They suggest future work with directed, weighted and asymmetric connections, and with alternative perturbation forms such as transient inputs or gain modulation, to test directional hypotheses more directly.
The behavioural link is also incomplete. The authors analyse correlations between functional interaction patterns and working-memory performance, but they state that how local imbalance within the DLPFC influences behaviour needs animal work that manipulates excitatory or inhibitory neurons directly. They also note that the persistent-activity account omits detailed synaptic mechanisms, and that matched verbal and spatial working-memory paradigms would be needed to test whether the hemispheric asymmetry depends on task content.
The sample is a specific one: right-handed HCP participants aged 22 to 36 for the connectivity analyses, and the HCP S1200 release more broadly for model fitting. The findings therefore describe this young-adult cohort and these parcellations, not clinical populations or development.
The asymmetry is the result worth testing
The paper's distinctive claim is not simply that the DLPFC matters for working memory, nor that excitation and inhibition matter. It is that the direction of the local shift differs between hemispheres: in the simulations, left excitatory elevation combined with right inhibitory elevation was the condition that moved connectivity toward the working-memory pattern, while single-area perturbations did not hold up across parcellations.
That asymmetry, together with the reported correlation between DLPFC-to-network connectivity changes and structural connections, is the specific prediction a follow-up study should target. The receptor-enriched results support the general involvement of NMDA- and GABA-A-related signals but do not test the hemispheric asymmetry itself, and the correlational design cannot decide direction. Readers should treat the model perturbation as a hypothesis-generating result whose value lies in naming a testable left-right dissociation, not as evidence that changing DLPFC inhibition in people will alter working-memory network dynamics.
Sources used in this answer
Differential excitatory-inhibitory balance within dorsolateral prefrontal cortex links inter-network interactions in working memory
A resting-state-fitted excitation/inhibition mean-field model shows that raising left DLPFC excitation while raising right DLPFC inhibition shifts simulated functional connectivity away from rest and toward working memory, and fMRI analyses show DLPFC changes covary with frontoparietal and default-mode network changes.