Role of GABA and NMDA receptors in shaping cortical timescales and large-scale network dynamics

Reviewer #1 (Public review):

Summary:

The authors investigate whether systematic manipulation of GABA-A and NMDA receptors influences large-scale cortical neuronal timescales and transient network dynamics. 57 healthy male participants completed placebo, lorazepam and D-cycloserine sessions in a double-blind, within-participant, cross-over design. The authors estimated timescales from the knee frequency of the aperiodic component of MEG power spectra, and identified transient large-scale cortical networks using a time-delay embedded hidden Markov model (TDE-HMM) analysis. The authors report that lorazepam increases estimated timescale across multiple cortical areas, with particularly pronounced changes in states interpreted as frontal default-mode network (DMN) and dorsal attention network (DAN). These changes include increased fractional occupancy of the DMN and decreased for DAN. D-cycloserine did not significantly affect neuronal timescales.

Strengths:

The study has several notable strengths:

(1) The pharma-MEG study is well designed and executed (e.g., a crossover design, collecting subjective, cardiovascular and additional control measurements).

(2) Neuronal-timescales maps are validated against previously published cortical timescale and hierarchy maps.

(3) TDE-HMM findings are examined using alternative numbers of hidden states and different preprocessing pipelines.

(4) The manuscript is generally clear and well written.

Weaknesses:

Nevertheless, several aspects of the primary neuronal timescale measure and the TDE-HMM analysis require further validation, and several points should be clarified:

(1) A central concern is whether the reported measure of neuronal timescale, on its own, is sufficient to support the interpretation assigned to it. Lorazepam has been shown to alter several spectral parameters, including oscillatory peaks and the aperiodic component of power spectra, in ways that could influence knee-frequency fitting. The manuscript, however, does not provide sufficient information about fit quality (across participants, parcels, and conditions), parcels within each participant/condition with identifiable knees, or the sensitivity of the results to the fitting range and peak settings (i.e., FOOOF parameters). Importantly, Figure S3 indicates that neither lorazepam nor D-cycloserine significantly affects knee frequency, although the neuronal timescale measure is mathematically derived from that knee frequency (i.e., tau = 1/(2*pi*f_knee)). Although such a result is mathematically possible, the pattern is unusual and requires explanation because timescales are not measured independently, but rather derived from the knee frequency. Furthermore, the neuronal timescale maps show a relatively homogenous increase across the cortex under lorazepam (Figure 2A), whereas the knee-frequency maps show a heterogeneous spatial pattern, with increased knee frequency under lorazepam in frontal regions, and decreased in occipital and temporal regions (Figure S3). A widespread significant effect appearing only after inversion could reflect a genuine effect, particularly in parcels with low knees to begin with. However, it could also arise if a small number of extremely low fitted knee frequencies (below 1Hz?) produce very long timescale estimates and disproportionately affect the statistical analysis.
The authors should address this discrepancy by presenting the distributions of knee-frequencies and neuronal timescales, and their ranges. Additional information about outliers and the robustness of the findings to extreme fitted values would also be valuable.

(2) A second concern relates to the TDE-HMM analysis. The model identifies states from the covariance matrix of time-delayed time-courses. Consequently, states are differentiated partly on the basis of their spectral and temporal characteristics. Knee frequencies for each state are then estimated from the power spectra of the same states used to define them. Differences in neuronal timescales across states may therefore be expected, at least in part, simply by the way the states were inferred. This point weakens the claim that cortical states are independent, and "operate on distinct timescales". A clearer separation between state definition and neuronal timescale estimation would be needed to establish that the reported state-specific differences are not partly an expected consequence of the fitted model. This could be addressed using, e.g., cross-validation or simulations. This concern is less substantial for the drug-related changes in neuronal time scale within individual states.

(3) The result and methods sections provide insufficient detail and statistical reporting for the TDE-HMM analysis. In the result section (page 10), the authors report only the *range* of fractional occupancies across all states. A range of 1% to 48% is substantial. It is therefore important to determine whether some states occupied only 1% (corresponding to ~3sec) while others occupied a much larger proportion. The authors should report the mean fractional occupancy of each state, together with its SD and range across participants.

(4) The authors should explain the apparent discrepancy between the relatively homogeneous increase in neuronal timescales under lorazepam across the complete recording (Figure 2) and the heterogeneous state-specific changes (both increases and decreases; Figure 4B). Could this pattern be explained, at least in part, by the differences in fractional occupancy of individual states? This provides an additional reason to report state-specific occupancy values in greater detail.

(5) On page 13, 2nd paragraph, the authors state that the findings indicate that the frontal DMN is an important driver of the global prolongation of neuronal timescales, based on the strong similarities to the time-averaged timescales. However, Figure S5 appears to show that the spatial correlation between state-specific and time-averaged timescales is as high for State 8 and is similar for State 2. The current statement therefore appears to be an overinterpretation. Demonstrating that the correlation for State 3 is significantly larger than the correlations for the other states would provide stronger support for this claim.

(6) Spatial maps are presented inconsistently throughout the main manuscript and in the supplementary. Specifically, some figures only show thresholded maps (at p<0.05 or p<0.001; e.g., Figure 4b), whereas others show unthresholded maps. Reporting only the number of parcels exhibiting a significant effect, without presenting the corresponding thresholded maps, makes it difficult to evaluate the spatial distribution of the reported changes. At least for the main findings (e.g., lorazepam effects on neuronal timescales), I recommend presenting both thresholded and unthresholded maps within the same figure.

(7) Figure 3: The rationale for presenting the mean power between 3-30 Hz is unclear. The authors should explain why this metric was selected to represent each state. Visual inspection suggests that the state-specific power spectra differ across several dimensions, including the aperiodic exponent, offset, alpha power, etc. Characterizing states using these parameters may be more informative. Each of these features could potentially also influence the estimated knee frequency.

(8) The conclusion on page 17 (and similar statement in the intro): "...these findings provide causal evidence that microscale synaptic inhibition directly shapes both local neuronal timescale organization... ") appears to be overstated based on the evidence presented. Although the lorazepam manipulation supports a causal effect of the drug on MEG-derived cortical timescales and network dynamics, the study does not directly measure microscale synaptic inhibition or establish a direct mechanistic link across scales. The phrasing should therefore distinguish the observed pharmacological effects from the inferred role of GABAergic inhibition and be phrased more cautiously.

(9) The method section lacks several critical details, including: criteria for excluding MEG channels and noisy segments (see comment below), details on source reconstruction (e.g., number of vertices used to project the sensor-level data), procedures used to generate the null distributions for the spatial autocorrelation preserving permutation test, transition probability analysis and more. Relatedly, the preprocessing pipeline of the MEG data appears to be based on manual inspection for the removal of channels, ICA components and noisy segments. This procedure is inherently subjective. The authors should provide details on the exact criteria used to make these decisions. Critically, the authors should also report the duration of usable resting-state data remaining after segment rejection. The Methods section should provide sufficient detail to allow readers to evaluate the validity of the study and reproduce the analysis as closely as possible. In its current form, it does not do so.

Reviewer #2 (Public review):

This work provides empirical data on how GABA and NMDA agonists globally affect timescales as measured through MEG. The authors reproduce the previously observed gradient of intrinsic timescales in the placebo condition, as well as its relationship to cortical hierarchy in T1/T2w maps. Timescales were not fixed, but dynamic, as revealed by large-scale network analysis separating into discrete network states. Pharmacologically, GABA agonist Lorazepam produced a brain-wide increase in timescales while NMDA agonist D-cycloserine did not. Furthermore, the GABA-mediated increase in timescale was area- and state-dependent, and more detailed analyses show changes in state occupancy mainly for DMN and DAN.

Overall, the paper contributes valuable data on a relevant topic of research in understanding the timescales of network dynamics at the local and global level. The question is well-motivated, and the analyses are technically sound and described in a straightforward manner. The network-level analysis in TDE-HMM is interesting and provides a complementary and more fine-grained perspective to the global timescale gradient, both in terms of space and time. The hypotheses were straightforward since both GABA and NMDA have relatively long timescales (of the dominant synaptic currents), though the lack of effect from NMDA-agonist is quite surprising but reasonably explained by the voltage-dependence of NMDA receptors in such a task-free setting. The paper overall is clearly written, and the figures are of high-quality, though some things could be presented in slightly more informative ways (see below). I have some questions and minor suggestions, but don't have too much to criticize as a whole.

My biggest question is the following: the mixed effects model shows that lorazepam additionally mediates timescale over and above the hierarchy (myelination map). This leaves a very clear gap. What the authors also probably want to show is that greater GABA_A receptor expression (of any or all the subunits) results in greater change under lorazepam, not against the myelination map only, or that the residue can be explained by the GABA_A maps. This, of course, would not explain the non-stationary nature of the dynamics (and state-dependent timescale maps), but would give a more direct explanation of the spatial effect. Since you already compared to the prior MEG map in Shafiei et al., 2023, I guess it's not a huge technical effort to grab the gene maps from neuromaps (https://github.com/netneurolab/neuromaps). I think this could strengthen the current manuscript.