BOLD signal changes can oppose oxygen metabolism across the human cortex

The paper is internally highly consistent. The central claim—that ~40% of voxels with significant BOLD responses show discordant CMRO2 changes—is supported through multiple complementary analyses: PLS statistics, GLM validation, conjunction voxel analysis, Davis model comparison, and a replication study. The mechanistic explanation (discordant voxels rely on OEF rather than CBF changes) follows logically from Fick's principle and is empirically supported. Claims about baseline OEF predicting hemodynamic response type are coherently connected to the observed task-related findings. The 'mixed voxel' analysis elegantly rules out voxel-specific confounds by demonstrating that the same voxels can show concordant responses in one condition and discordant in another.

The core equations are well-established and correctly applied: Fick's principle for CMRO2, the mqBOLD relation for OEF (R2'/c×sCBV), the BOLD-informed R2' estimation, and the Davis model. The BOLD-informed surrogate for R2' (Equation 4) is a reasonable approximation, and the authors appropriately compare it against fully quantitative estimates in a sensitivity analysis (Supplementary Fig. 5). The Davis model parameter derivation from their own data (α=0.38, M=11.2%) is a meaningful cross-validation. One minor concern is the 25% upscaling correction applied to CBF for background suppression pulses—while referenced, the propagation of this correction into CMRO2 uncertainty is not explicitly quantified in the main text. The CBV normalization to a WM reference value of 2.5% is standard but introduces assumptions not fully discussed. These are minor issues in an otherwise mathematically sound framework.

The work is strongly falsifiable and in fact directly falsifies aspects of the canonical neurovascular coupling assumption. Specific quantitative predictions are tested: the Davis model's prediction of discordant BOLD is empirically confirmed. The finding that baseline OEF predicts hemodynamic response type is a testable hypothesis that could be examined with independent datasets. The use of external QSM datasets to validate lower susceptibility and higher venous density in discordant voxels represents prospective validation. The claim that discordant voxels are distributed uniformly across BOLD amplitude quartiles (no amplitude bias) is a specific, checkable prediction. The entire framework could be falsified if quantitative measurements with PET-MRI showed concordant CMRO2 in these voxels, or if higher-resolution CBF measurements revealed CBF decreases in the negative BOLD regions.

The paper is well-written and logically structured. The introduction clearly motivates the departure from canonical assumptions. Figure 1's explanation of the hemodynamic model and the mqBOLD pipeline is pedagogically effective. The terminology of 'concordant' and 'discordant' voxels is consistently applied throughout. The distinction between BOLD-informed and fully quantitative CMRO2 estimation could be more prominently flagged in the main text (it is relegated to Methods and a supplementary figure). The description of the conjunction voxel analysis is clear and the logic for ruling out confounds is well-articulated. Some figure captions are dense with statistical notation that may require careful reading, but this is appropriate for the technical audience.

This paper makes a genuinely important empirical contribution. While the theoretical possibility of discordant BOLD responses was predicted by the Davis model and suggested by prior animal and limited human studies, no prior work has systematically quantified the prevalence (~40%) and spatial distribution of discordant voxels across the whole human cortex using in-session multimodal qfMRI. The identification of two mechanistically distinct hemodynamic response modes—CBF-dominated versus OEF-dominated—with baseline OEF as a predictor, is novel and provides a physiological framework for understanding regional heterogeneity. The validation of decades-old Davis model parameters empirically from voxel-level data is itself a noteworthy contribution. The 'mixed voxel' finding (same voxel concordant for one task, discordant for another) is a particularly elegant and original demonstration.

The paper is substantially complete within its stated scope. Multiple control analyses address the major alternative explanations: partial volume effects, signal drift, sensitivity limitations, and voxel-specific confounds. The replication study with harmonized voxel sizes is important and strengthens the claims. Areas that could be more fully developed: (1) The neurophysiological interpretation of OEF-mediated CMRO2 regulation is discussed speculatively (astrocytic activity, E/I balance, neuromodulation) but not empirically tested—this is acknowledged as future work; (2) The restriction to healthy young adults (mean age ~32y) and the question of generalizability to clinical populations is acknowledged but not empirically addressed; (3) White matter is appropriately excluded but subcortical gray matter receives limited discussion; (4) The temporal resolution mismatch between BOLD (30-s blocks) and qfMRI (~6-min blocks) is a design limitation that is acknowledged but whose potential impact on the concordance/discordance classification warrants more explicit discussion.