Assessing molecular, cellular and transcriptomic bases of laminar perfusion and cytoarchitecture coupling in the human cortex Authors: Fanhua Guo, Chenyang Zhao, Ravi R Bhatt, Zixuan Liu, Andy Jeesu Kim, Zidong Yang, Siyi Xu, Kay Jann, Xingfeng Shao, Mara Mather, Neda Jahanshad, Danny JJ Wang Published: 2026 ABSTRACT AbstractUnderstanding how cellular architecture shapes cortical function requires mesoscopic approaches that resolve structure-function relationships in vivo. Here we introduce the cerebral blood flow (CBF) – cell-body staining intensity (CSI) similarity index (CCSI), a quantitative measure of laminar perfusion-cytoarchitecture coupling derived from whole-brain 1-mm isotropic 7 T arterial spin labeling and cell-body staining intensity profiles from the BigBrain atlas. Across 30 participants, CCSI revealed reproducible, region-specific alignment between laminar perfusion and cellular density distributions. CCSI was selectively associated with mitochondrial respiratory capacity per mitochondrion and colocalized with capillary endothelial and end-state oligodendrocyte populations. Gene ontology enrichment converged on metabolic regulation, neurovascular organization, and mitochondrial homeostasis. A parallel control analysis using quantitative T1 yielded no significant associations, supporting the specificity of these findings. At the systems level, CCSI improved structure-function gradient correspondence in higher-hierarchy association cortices. These results support CCSI as a reproducible, non-invasive mesoscopic index linking cortical perfusion, microstructure, and metabolism.Subject terms: Neuro-vascular interactions, Magnetic resonance imaging Introduction Structural connectivity provides the anatomical scaffold for functional interactions in the human brain1–3, yet how this scaffold gives rise to the rich repertoire of functional states remains incompletely understood. Over the past decade, macroscale frameworks combining diffusion MRI4, functional MRI5, and graph theory6 have delineated organizing principles that link cortical architecture to network dynamics2. A persistent and increasingly recognized exception, however, complicates this view: in higher-order transmodal networks, structural and functional gradients progressively diverge7–9, indicating that the anatomical scaffold alone cannot explain the topography of cortical function. This divergence points to a missing layer of organization, one that operates below the macroscale and is grounded in the metabolic and neurovascular substrate through which structure is translated into function. Among candidate organizers, brain metabolism has emerged as more than a passive energetic substrate10–14. Because neural computation imposes continuous and spatially uneven, energetic demands, the brain’s metabolic landscape is itself spatially structured rather than uniform10,11. Consistent with this view, recent work has shown that regional variations in energy utilization15, lipid metabolism16, and cerebral blood flow17 (CBF) align with macroscale patterns of functional organization, including the principal gradient of functional connectivity. These spatial correspondences position metabolism as a candidate organizing factor that may complement the anatomical scaffold in shaping cortical function. Yet metabolism itself is not free-floating: any organizing role it plays must be realized through the physical substrates that deliver and distribute energy across the cortex. Whatever organizing role metabolism plays must be realized through a physical delivery system. Unlike most organs, the brain maintains negligible endogenous energy reserves, with only limited glycogen stored primarily within astrocytes18,19, and neural activity depends on substrates supplied by the cerebral vasculature, with astrocytes20 and oligodendrocytes21–23 providing additional routes when capillary access is limited. Because blood vessels24 and neurons25 are non-uniformly distributed across the cortex, their relative spatial arrangement varies systematically across regions, and the local balance between metabolic supply and demand is likely to be spatially structured. This constraint is especially stringent for oxygen, whose effective diffusion distance from capillaries is limited to tens of micrometers in cerebral cortex26–29 (approximately 50 μm), so that the geometric relationship between microvasculature and surrounding cells directly shapes how adequately each neuron is supplied. Regions where vascular and cellular density distribution are co-localized may represent configurations in which substrates reach neurons via short, direct diffusion paths, supporting efficient use by local cellular machinery, including the mitochondria responsible for energy production. Where alignment is poorer, direct supply is less available, and sustained activity may rely more on indirect, glia-mediated routes, consistent with distinct patterns of metabolic organization. Quantifying how closely vascular supply aligns with cellular distribution, therefore, offers an indirect but informative window into cortical metabolic organization, reflecting both the geometry of substrate delivery and the efficiency with which delivered substrates are used. In the human neocortex, the structural architecture of cellular density is predominantly organized across layers30,31. Paralleling this cytoarchitectonic logic, the cortical microvasculature enters from the pial surface and ramifies through the cortical thickness, producing vascular density profiles that likewise vary across layers32–34. Given this shared laminar organization, the spatial alignment between vascular supply and cellular organization can be quantified by comparing their respective profiles along the cortical depth. To establish a structural reference, we drew on the BigBrain atlas35, a high-resolution 3D histological reconstruction that currently provides the most comprehensive cytoarchitectonic reference for the human cortex. Because laminar patterns of cell-body distribution follow consistent organizational principles31,36–38 across the human cortex, region-specific profiles of cell-body staining intensity (CSI) derived from BigBrain can serve as a representative reference for canonical cytoarchitectonic organization. The remaining challenge is therefore to obtain matching layer-resolved perfusion profiles in the living human brain. The development of ultra-high-field (UHF) MRI was driven by the goal of imaging the brain at the mesoscale (sub-mm), bridging the gap between microscale synapses and neurons and macroscale regions and networks39. CBF measured by arterial spin labeling (ASL) is a key in vivo metric of metabolic supply, as labeled arterial water exchanges with tissue water in capillaries near sites of neural activation40. Animal optical imaging shows that CBF can be rapidly regulated at the capillary level (e.g., via pericytes)41,42. At 7 T, ASL benefits from increased SNR (∝B01.65) and longer tracer half-life, yielding a nearly fivefold signal to noise (SNR) gain over 3T43,44. Pilot laminar ASL at 7 T has revealed higher resting perfusion in the middle cortical layers45–47, consistent with capillary density distributions derived from histology32,33. Because ASL is quantitative and originates from capillary/tissue compartments, high-resolution 7 T ASL provides the essential mesoscale tool to obtain the layer-specific vascular information required to quantify the spatial alignment between perfusion and cytoarchitecture. Here, building on these considerations, we introduce the CBF-CSI similarity index (CCSI), a quantitative measure of laminar perfusion-cytoarchitecture coupling derived from a novel turbo-FLASH (TFL) based rotated golden-angle stack-of-spirals (rGA-SoS) pseudocontinuous ASL (pCASL) sequence47 and cell-body staining intensity profiles from BigBrain35. We hypothesize that regions where laminar CBF and CSI profiles are closely aligned represent configurations in which oxygen is delivered in close proximity to metabolically active cells, supporting shorter diffusion distances and more efficient aerobic energy production; conversely, regions with poor alignment may rely to a greater extent on indirect metabolic supply routes (Fig. 1). Using a 1-mm isotropic pCASL sequence optimized for 7 T, we acquired laminar CBF maps across the cortex in 30 healthy participants, with 14 returning for a retest session to assess reproducibility. We first characterized the regional and laminar heterogeneity of cortical perfusion and CCSI, and then examined their biological underpinnings along three complementary axs: mitochondrial respiratory capacity from MitoBrainMap48, cell-type, vascular-subtype and oligodendrocyte-subtype distributions inferred from Allen Human Brain Atlas transcriptomics48,49, and gene ontology enrichment of CCSI-associated transcripts. We carried out a parallel control analysis substituting CBF with quantitative T1 (CCSI-control) to test the specificity of the CCSI-related findings. Finally, we examined whether CCSI contributes to the macroscale structure-function relationship by incorporating CCSI into gradient analyses of microstructural and functional organization, directly addressing the structural-functional divergence that motivated this study. Schematic illustration of the laminar perfusion-cytoarchitecture coupling hypothesis. Cortical regions differ in the degree of spatial alignment between laminar CBF profiles, measured by 7 T pCASL, and CSI profiles derived from the BigBrain atlas. In aligned regions (left), laminar CBF and CSI profiles co-vary across cortical depth, such that vascular supply is concentrated in layers with higher cellular density. This geometric correspondence minimizes oxygen diffusion distances from capillaries to metabolically active cells, supporting efficient aerobic energy production, a configuration quantified by a high CCSI. In misaligned regions (right), CBF and CSI profiles diverge across layers, indicating that peak perfusion does not correspond to layers of highest cellular density. This mismatch increases effective oxygen diffusion distances, potentially reducing aerobic metabolic efficiency. Background shading within each cortical column represents the local oxygen pressure gradient, which is shaped by the microvascular supply. The cortical column is divided into six cytoarchitectonic layers (I–VI), with cell bodies (blue circles) and penetrating vessels with capillary branches (red) depicted schematically. Brain silhouettes were created in BioRender. All other graphical elements were created by F.G. using Adobe Illustrator. Created in BioRender. Liu, Z. (2026) https://BioRender.com/1fr3w7o. Results Multiscale heterogeneity of CBF across cortical areas and lamina We first estimated cerebral blood flow (CBF), defined as the volume of blood delivered to a unit of tissue per unit time, at the voxel level in the cortex for each participant using our newly developed rGA-SoS pCASL technique47. Regional mean CBF values within the cortical gray matter were then parcellated for each participant using segmentation tools in FreeSurfer and the HCP-MMP1.0 atlas50 (Fig. S1a), yielding a 360 (regions) × 30 (participants) matrix. Figure 2a shows the average CBF parcel distribution across all participants, which appears to exhibit slight asymmetry. Previous studies have suggested that inhomogeneities in the B0 magnetic field at the cervical labeling plane51 and anatomical variations in the Circle of Willis52 can contribute to the slight asymmetry of CBF. However, we found no significant difference in CBF measurements between the left and right hemispheres (Fig. 2b) and the spatial distributions of CBF between the left and right hemispheres were highly similar (Fig. 2c), indicating that the asymmetry in our data was weak and did not fundamentally affect the results. To minimize the impact of individual high perfusions on the distribution, CBF values were normalized (z-scored) within each participant. Principal component analysis (PCA) was then applied to this matrix, with the first principal component explaining 52.75% of the total variance. The principal component captures the dominant spatial distribution pattern of CBF shared by all participants (Fig. 2d), and the spatial distribution of the principal component is mainly derived from the average distribution of all participants (Pearson r = 0.9989 between group average CBF and CBF PC1-score). Among these regions, early sensory cortices (including primary visual, somatosensory, and auditory areas), the dorsolateral frontal cortex, posterior cingulate cortex, and the heavily myelinated parieto-occipital sulcus area 2 exhibit high scores, whereas regions in the temporal lobe, such as the inferior temporal gyrus and portions of the middle temporal gyrus, show low scores. This spatial pattern aligns with observations reported in previous studies involving large cohorts using PET53 (Fig. S2a) and low-res ASL-MRI54 (Fig. S2b). Spatial distribution heterogeneity of whole and laminar CBF. a Parcellation pattern of group average CBF value across all participants (n = 30 independent biological replicates). b Comparison of mean CBF in the left and right hemispheres for all participants. Gray circles and lines represent the CBF distribution of each individual participant in the left and right hemispheres, and black circles and lines represent the mean CBF of all participants. P values were calculated using the two-sided paired t-test. c Correlation analysis of the spatial distribution of CBF in the left and right hemispheres. The line is the regression line, and the shading is the 95% confidence interval. Dots represent the CBF value from the left and right of the same brain region. We used Pearson correlation (r) and the variogram-null model to calculate the two-sided P value considering spatial autocorrelation. Exact r and P value are marked in the figure. d Parcellation pattern of CBF principal component (PC1) scores (CBF-score) across all participants. e Parcellation pattern of 3 laminar CBF PC1-scores across all participants. Three layers of gray matter from the superficial to the deep layer were calculated by the equal volume algorithm. We next segmented the cortical gray matter into 3 equi-volume layers based on T1-weighted structural MRI data. Using the same approach described above, we also computed the spatial distribution scores of CBF for each of the 3 cortical layers separately (Fig. 2e). The global CBF-score maps were broadly consistent across layers, yet a minority of regions showed clear layer-specific deviations. Among these regions, the spatial distribution score of the dorsolateral prefrontal cortex (dlPFC) and the dorsomedial prefrontal cortex (dmPFC) gradually decreases from the superficial to the deep layers. In contrast, early sensory cortices (including primary visual, somatosensory, and auditory areas) show a progressive increase in CBF-score from the superficial layer to deep layer. Laminar-specific coupling between perfusion and cytoarchitectonic architecture Using the same methodology, we segmented the cortical gray matter into 12 equi-volume layers. This allowed us to derive laminar profiles of CBF within each ROI. To quantify the similarity in the laminar distribution of cell density and perfusion, we utilized histological cell-body staining intensity (CSI) data provided by the BigBrain template to obtain layer-specific cytoarchitectonic properties for each ROI (Fig. 3a). To ensure technical alignment with the CBF laminar profiles, we segmented the BigBrain cortical gray matter into 12 equi-volume layers (Fig. S1c) and mapped it onto the same 360 brain regions using the same method described above. Figure 3b illustrates the laminar profiles of CBF and CSI within V1 for Participant 01. Then, we employed cosine similarity to quantify the resemblance between the laminar profiles of CBF and cell-body staining intensity (CSI) within each ROI for each participant. This measure, referred to as the CBF-CSI similarity index (CCSI, see “Methods” for calculation details), ranges from −π/2 to π/2, with positive values indicating aligned laminar profiles. Higher CCSI values reflect greater similarity between the laminar profiles of CBF and CSI. Given that the CSI, CBF, and subsequent analyses rely primarily on correlations across independent datasets and that such cross-dataset analyses are susceptible to confounding factors such as potential registration errors, differences in reference dataset properties, and inter-subject variability, we conducted a parallel control analysis to guard against results being misled by such biases. Specifically, we substituted CBF with quantitative T1 values (qT1, shown in Fig. 3b), estimated using the MP2RAGE sequence, and performed the exact same set of analyses (yielding the qT1-CSI similarity index, or CCSI-control). Spatial distribution heterogeneity of whole, laminar CBF-score and CCSI. a Schematic diagram of the distribution of laminar cell-body staining intensity (CSI) in brain histology of BigBrain data donors and the calculation of laminar profile. b Laminar profiles of CBF, CSI and qT1 within V1 in participant 01. c Average parcellation pattern of the CBF-CSI similarity index (CCSI) calculated from the laminar profiles of CBF and CSI across all participants after data bias correction and quality control by SNR. d Bar plot of CCSI distribution in each unilateral brain region across all participants (n = 30 independent biological replicates) after data bias correction and quality control by SNR. Statistical significance was assessed using one-way RM-ANOVA and two-sided T-test (BH-FDR correction). The horizontal lines and asterisks below the bars indicate brain regions that are significantly divergent from 0 after multiple comparison correction (FDR < 0.05). Data are presented as mean values ± SEM. FL frontal lobe; Ins Insula; OL occipital lobe; PL parietal lobe; TL temporal lobe. e The average parcellation pattern of the qT1-CSI similarity index (CCSI-control) was calculated by the same method as CCSI. f Laminar profiles of CBF, CSI and qT1 within 6 different CCSI rank percentile regions. Exact group average CCSI (n = 30 independent biological replicates) and FDR values after BH-FDR correction are marked in the figure. The shaded area represents mean values ± SD among participants. V2 second visual cortex; V1 primary visual cortex; MIP medial intraparietal area; TPOJ1 temporo-parieto-occipital junction area 1; TA2 temporal area 2. After data bias correction and quality control by SNR (Fig. S3, see details in “Methods”), Fig. 3c presents the average bilateral CCSI map across participants, and Fig. 3d presents the unilateral regional distribution of CCSI. Laminar profiles of CBF and CSI remained significantly correlated in nearly all cortical regions across all participants (T-test with Benjamini-Hochberg false discovery rate correction; 59/60 in the frontal lobe, 5/6 in the insula, 23/23 in the occipital lobe, 46/46 in the parietal lobe, and 16/20 in the temporal lobe). Moreover, CCSI demonstrated significant region-specific heterogeneity (one-way repeated-measures ANOVA: main effect of brain region, F(154, 4466) = 17.89, P = 1.5 × 10−30, partial η2 = 0.38, Greenhouse-Geisser corrected). The group-level CCSI map exhibited broadly positive values across most cortical regions, with the strongest similarity along the central sulcus in primary sensorimotor areas and in occipital visual cortex. Moderate positive values extended into widespread association cortices, while lower or negative CCSI values were restricted to circumscribed regions of the inferior temporal and occipitotemporal cortex. Figure S4 shows the results based on the retest data of the 14 participants with two measurement sessions, which demonstrates the reliability and repeatability of CCSI and CBF-score maps obtained by our method. Figure 3e illustrates the distribution of the CCSI-control map. It reveals that the layer-specific correlation between qT1 and CSI is relatively weak, with many regions exhibiting a negative correlation. To gain a more intuitive understanding of the physiological significance represented by varying levels of CCSI, Fig. 3f presents the laminar profiles of CSI, CBF, and qT1 within ROIs corresponding to six distinct CCSI percentile ranks. We observe that in regions with high CCSI, the laminar distribution patterns of perfusion and cell density are nearly identical. In regions with moderate CCSI, a certain degree of misalignment exists between the two. However, in regions with extremely low CCSI, a severe misalignment is evident. For instance, in the posterior insula 1 (Fig. 3f), cell density is concentrated in the deep layers, whereas perfusion is concentrated in the superficial layers. Overall, CBF and CSI exhibit a high degree of similarity along the laminar depth direction, with substantial regional variation. Conceptually, CCSI may serve as a quantitative marker of the coupling between laminar scale cytoarchitecture and perfusion (i.e., blood flow and energy supply). These findings suggest that CCSI (a mesoscopic anatomic indicator) captures biologically meaningful laminar coupling between cytoarchitecture and perfusion. To investigate whether variations in the degree of this coupling across different brain regions reflect differences in metabolic capacity or metabolic patterns, we next examined the analysis of the CCSI distribution map at the molecular, transcriptomic, and cell-type levels. Laminar perfusion-cytoarchitecture coupling reflects mitochondrial metabolic capacity Consistent with previous findings, the spatial distribution of CBF was highly correlated with cerebral metabolic rates of oxygen and glucose (Fig. S2c, d), supporting its role as a proxy for local energy supply and metabolic activity53. Mitochondrial respiratory chain complexes, the core machinery of oxidative phosphorylation, are regionally heterogeneous across the human cortex48. To assess whether regional perfusion patterns and their laminar coupling with cytoarchitecture reflect underlying metabolic enzyme activity, we examined the spatial correspondence between CBF-score, CCSI and CCSI-control maps and cortical distributions of mitochondrial Complexes I (NADH dehydrogenase, CI), II (succinate dehydrogenase, CII) and IV (cytochrome c oxidase, CIV). Figure 4a shows the spatial distribution maps of mitochondrial Complexes I, II, and IV (MCs), derived from MitoBrainMap48, which constructed whole brain maps by voxelizing human brain tissue, quantifying mitochondrial enzyme activity, and projecting to the cortex using multimodal imaging models. While mitochondrial complexes exhibit similar global gradient patterns across the cortex, they display distinct regional distributions, particularly in the temporoparietal junction, visual cortex, and parietal lobule. Figure 4c shows the Pearson correlation results between the spatial distributions of mitochondrial complexes and CCSI. It indicates that CCSI exhibited a significant positive correlation with CI. Figure 4e presents the distributions of mitochondrial density (MitoD), along with its correlation with CCSI. It indicates that there is no significant relationship between CCSI and MitoD. Figures S5a, b and 4f show the distributions of tissue respiratory capacity (TRC) and mitochondrial respiratory capacity (MRC, i.e., the respiratory capacity per mitochondrion), along with their respective correlations with CCSI. It indicates that CCSI exhibited significant positive correlations with both TRC and MRC. These results indicate that the distribution of CCSI is indeed significantly correlated with mitochondrial metabolic capacity, specifically, with the metabolic capacity per mitochondrion, rather than with high metabolic states resulting from mitochondrial abundance. Figures 4b and S5c–e present the results of correlation analyses examining the spatial distribution of CBF-scores and various mitochondrial metrics. The CBF-scores exhibited no significant correlation with any of these distributions. Figures 4d and S5f–h present the results of correlation analyses examining the spatial distribution of CCSI-control and various mitochondrial metrics. The CCSI-control exhibited no significant correlation with any of these distributions, suggesting that the observed correlation between CCSI and mitochondrial metabolic capacity is not attributable to systematic biases across different datasets. Spatial correlation analysis of CBF-score, CCSI or CCSI-control map with mitochondrial complex and respiratory capacity related maps. a Parcellation pattern maps of 3 mitochondrial complexes. These include complex I (CI, NADH–ubiquinone oxidoreductase), complex II (CII, succinate dehydrogenase), and complex IV (CIV, cytochrome c oxidase). b Relationship between CBF-score distribution and mitochondrial complexes. We used Pearson correlation (r) and the variogram-null model to calculate the two-sided P value considering spatial autocorrelation. Exact r and FDR values after BH-FDR correction are marked in the figure. The black line represents the regression line, the shading is the 95% confidence interval, and dots represent individual brain region data (likewise in subsequent panels). c Relationship between CCSI distribution and mitochondrial complexes. d Relationship between CCSI-control distribution and mitochondrial complexes. e The Parcellation pattern maps of mitochondrial density (MitoD) and the relationship between CCSI and MitoD. Exact r and variogram-null two-sided P value are marked in the figure. f Parcellation pattern maps of mitochondrial respiratory capacity (MRC) and the relationship between CCSI and TRC. Exact r and variogram-null two-sided P value are marked in the figure. Taken together, these results suggest that the spatial heterogeneity of CCSI is more closely related to mitochondrial metabolic capacity than to mitochondrial abundance. This observation is in line with our hypothesis that regional variations in perfusion-cytoarchitecture coupling may reflect differences in underlying metabolic organization. CCSI is selectively associated with oligodendrocyte and capillary organization Distinct brain cell types exhibit different metabolic demands and energy burdens55, and the laminar coupling between perfusion and cytoarchitecture may reflect underlying cellular composition or function. Recent studies have leveraged microarray data from six donors of Allen Human Brain Atlas49 (AHBA) to infer the spatial distribution of major brain cell types and their associations with MRI-derived maps56,57. Here, we applied gene-category enrichment analysis57 (GCEA) to evaluate the spatial correspondence between our CBF-score and CCSI maps and the distribution of canonical cell types, and cross-validated the results using cell-class gene expression mapping56 (CGEM). Based on the marker genes of cell-class in the brain by previous studies58, Fig. 5a, b presents the results of GCEA for the CBF-score, CCSI and CCSI-control maps, while Figure S6a, b shows the corresponding results from CGEM. Both CBF and CCSI exhibited consistent and significant positive enrichments with endothelial cells (Endo). Figure 5c further illustrates the spatial correspondence between the CGEM-estimated Endo-map and both perfusion-derived measures. Given that both CBF and CCSI are inherently vascular in nature, these results support the validity of the enrichment analysis in capturing cellular features. Notably, CCSI also demonstrated a robust and significant enrichment with oligodendrocytes (Oligo), whereas this relationship was absent for the CBF. Figure 5d shows the spatial correspondence between the CGEM-derived Oligo-map and the CCSI distribution. This finding suggests that laminar-scale perfusion-cytoarchitecture coupling may be particularly linked to oligodendrocyte-related structure or function. Finally, CBF also showed a correlated distribution with inhibitory neuron subtypes In6a and In6b (Fig. S6c, d), consistent with prior studies17. In6a and In6b cell types are fast-spiking parvalbumin-positive GABAergic neurons59, and this fast-spiking action is often considered to require higher energy support60,61. Cell-class enrichment analysis of CBF-score, CCSI and CCSI-control. Whole cell-class enrichment analysis of CBF-score (a) and CCSI/CCSI-control (b) by gene category enrichment analysis (GCEA). We used the variogram-null model to calculate the two-sided P value considering spatial autocorrelation. The exact FDR values that were significant after BH-FDR correction are marked in red (likewise in subsequent panels). Astro Astrocytes; Oligo Oligodendrocytes; OPC Oligodendrocyte precursor cells; Endo Endothelial cells; Per Pericytes; Exa Excitatory neuron subtype a; Ina, Inhibitory neuron subtype a. c Spatial distribution of endothelial cells estimated using gene expression mapping and its relationship with CBF-score and CCSI. Exact r and two-sided P values calculated by the variogram-null model are marked in the figure. The black line represents the regression line, the shading is the 95% confidence interval, and dots represent individual brain region data (same as (d)). d Spatial distribution of Oligodendrocyte estimated using gene expression mapping and its relationship with CCSI. e Vascular-related subtype cell-class enrichment analysis of CBF-score, CCSI and CCSI-control by GCEA. These include various subtypes of three types of vascular cells: endothelial, pericyte, and smooth muscle cells. The vascular cells at the top come from arteries, the ones at the bottom come from veins, and the y-axis position of the capillaries in the middle also reflects the distance of each capillary cell from the artery or vein. f Oligodendrocyte’s subtype cell-class enrichment analysis of CBF-score, CCSI and CCSI-control by GCEA. COPs, Committed oligodendrocyte precursors; ImOlGs, Immature oligodendrocytes. g Distribution of gene sets and overlap between end-states oligodendrocytes and other states oligodendrocytes. Each circle represents an oligodendrocyte in a certain state. The number in the circle indicates the number of genes in the marker gene set of that state. The number in the brackets indicates the number of genes remaining in the gene set after removing the overlapping genes. The colored shading is based on the report of Jakel et al. 63, which summarizes these six states into four functional states, including actively myelinating state, intermediate state, end state and oxidative stress state. h Oligodendrocyte’s subtype cell-class enrichment analysis of CBF-score by redefined state-specific gene sets in (g). The bar colors are consistent with the state classification in (g). i Oligodendrocyte’s subtype cell-class enrichment analysis of CCSI and CCSI-control by redefined state-specific gene sets in (g). Significant FDR values calculated by the variogram-null model (two-sided) and BH-FDR correction are marked on each bar. We next performed enrichment analysis targeting vascular-associated62 and oligodendrocyte63 subtypes previously defined in single-cell transcriptome studies. Figure 5e and Figure S6e show the enrichment results for vascular subpopulations derived from both analytical frameworks. Both CBF and CCSI were consistently and significantly enriched in vascular-associated cell types. And both CCSI and CBF demonstrate a tendency to be enriched within vascular cell types associated with capillaries. Given that ASL-based CBF measurements predominantly reflect capillary perfusion, these results further support the validity of the enrichment analysis. Figures 5f and S6f display the enrichment results for oligodendrocyte subtypes. CCSI showed robust and consistent enrichment across multiple mature oligodendrocyte populations, but not in precursor, developing, or differentiating subtypes. In contrast, no significant associations were observed for CBF. This finding suggests that mature differentiation of oligodendrocytes may be closely associated with laminar-scale coupling between perfusion and cytoarchitectonic organization, and that this relationship is not directly captured by total blood flow. However, as noted by Jakel et al. 63, the six oligodendrocyte categories (Oligo1–6) represent distinct transcriptional states rather than discrete cell subtypes. Figure S7a illustrates these six oligodendrocyte states63, along with the expression profiles of their respective marker gene sets and the degree of gene overlap among them. Thus, the significant enrichment of CCSI across all six oligodendrocyte categories observed in Fig. 5f may partly reflect shared gene expression, rather than true state-specific associations. To address this, we removed all overlapping genes to generate non-redundant, state-specific marker sets (Fig. S7a) and repeated the enrichment analysis. Figure S7b, c show the results for CBF and CCSI, respectively. CCSI showed stronger enrichment in end-state oligodendrocytes, whereas results for other states were less reliable due to the small size of their gene sets. To address this, we only removed overlaps with the end-state gene set and redefined state-specific gene sets (Fig. 5g). Subsequent analyses revealed that only CCSI remained significantly enriched in end-state (FDRvariogram = 0.001/0.002) and intermediate-state (FDRvariogram = 0.002) oligodendrocytes (Fig. 5h, i). These findings suggest that laminar perfusion–cytoarchitecture coupling captured by CCSI is particularly associated with end-state oligodendrocytes and those transitioning toward end-states. Furthermore, none of the results associated with the CCSI-control reached statistical significance, thereby reinforcing the validity of the CCSI findings. Gene ontology signatures of CCSI To further interpret the spatial heterogeneity of CCSI, we next performed gene ontology (GO) enrichment analysis to identify associated biological processes using the aforementioned GCEA method (details in “Methods”). A substantial fraction of the genes most strongly correlated with CCSI are involved in metabolic regulation (Fig. 6a), including IGFBP264, a modulator of insulin sensitivity and lipid metabolism; PRMT765, a regulator of muscle oxidative metabolism via the p38MAPK/PGC-1α axis; and VMP166, an ER-resident protein essential for lipoprotein secretion and mitochondrial fatty acid oxidation. GO enrichment analysis identified 63 significantly enriched biological process terms (Fig. 6b), which converge on three interconnected functional themes. The most prominent theme is metabolic regulation: 16 terms are directly metabolic, including glycogen metabolism, lipid and phospholipid transport, sphingolipid biosynthesis, and phosphatidylinositol metabolism, with 5 additional terms indirectly related, including protein catabolic process and cellular response to amino acid starvation. Several of these processes are functionally integrated along the insulin signaling cascade, suggesting that CCSI may capture spatial variation in the coupling between vascular nutrient supply and local cellular energy demands. A second group of enriched terms (including blood vessel development, regulation of cell adhesion, cell polarity maintenance, and cell shape regulation, together with ion transport processes such as bicarbonate, chloride, and potassium transmembrane transport) points to the structural and functional organization of the neurovascular unit (NVU) and blood-brain barrier. This may reflect that the perfusion-cytoarchitecture coupling quantified by CCSI reflects, in part, the spatial integrity of the NVU. Third, enrichment of mitochondrial organization, autophagosome assembly, protein K63-linked ubiquitination, and protein catabolic processes indicates that regions with distinct CCSI values differ in intracellular energy production and quality control capacity. In contrast, the GO analysis for CBF yielded only six relatively scattered significant terms, while CCSI-control showed no significant terms, not even a single gene. GO enrichment analysis of CCSI. a Volcano plot of CCSI correlation with all genes. We used the variogram-null model to calculate the two-sided P value considering spatial autocorrelation. Significant genes after BH-FDR correction are marked with colored (blue or red) dots, while insignificant ones are shown in gray. b CCSI GO enrichment analysis of significant items. The figure shows the GO items that are significant after the variogram-null model (two-sided) and BH-FDR correction. Collectively, these results suggest that the spatial heterogeneity of CCSI is not merely a geometric property of vascular-cellular alignment but is associated with a biologically meaningful landscape in which vascular architecture is co-organized with local metabolic demands, NVU integrity, and intracellular homeostatic mechanisms. Reliability verification of the main findings We examined the molecular, cellular, and transcriptomic correlates of CCSI and then assessed the reproducibility of our primary findings. Using the 14 participants’ retest data (Fig. 7a–c), we found that CCSI derived from the independent retest session also significantly colocalized with mitochondrial respiratory capacity and end-state oligodendrocyte distributions, confirming the reproducibility of our main results. Because our analysis assumes that the molecular (MitoBrainMap)48, cytoarchitectonic (BigBrain)35, and transcriptomic (AHBA)49 data from donors match the common distribution across multiple samples, we sought to determine the approximate sample size required to achieve robust and reliable results. To this end, we randomly selected subsets of 5, 10, 15, and 20 participants (20 iterations each) from the 30-participant dataset to replicate our primary findings. Robust and reproducible associations were consistently observed with sample sizes of ≥10 (Fig. 7d–g), underscoring the robustness and generalizability of our findings. To validate the internal consistency of our data, we also conducted a split-half analysis (20 iterations each). Figure 7h illustrates the spatial correlation analysis of the CCSI results from the two halves, demonstrating extremely strong internal consistency (r = 0.92). Figure 7i displays the replication of our main findings across the 20 split-half iterations, further substantiating the reliability of our measurements. Main findings are reproducible by the retest of 14 participants, randomized samples, and split-half analysis. a–c Main findings are reproduced in the retest session of 14 participants. Relationship between CCSI distribution and mitochondrial CI (a) and MRC (b). We used Pearson correlation and the variogram-null model to calculate the two-sided P value considering spatial autocorrelation. Exact r and P values are marked in the figure. The black line represents the regression line, the shading is the 95% confidence interval, and the dots represent individual brain regions. c Reproduce the enrichment analysis results of CCSI and different oligodendrocyte states in Fig. 4i using 14 retest data. The bar colors are consistent with the state classification in Fig. 4g. Exact FDR values calculated by the variogram-null model (two-sided) and BH-FDR correction are marked on each bar. d–g Main findings reproduction in randomly selected subsets of 5, 10, 15, and 20 participants (20 iterations each) from the 30-participant dataset. Each dot represents the main finding calculated after repeated sampling. The color of the dot represents the two-sided P value calculated by the variogram-null model (corresponding to the color bar in the figure). h The correlation between the spatial distributions of CCSI in the two parts after a split-half. The black line represents the regression line, the shading is the 95% confidence interval, and the dots represent individual brain regions. i The main findings were reproduced in a split-half analysis (20 iterations) from the 30-participant dataset. Triangles represent the spatial correlation between the two parts in each split. Circles and squares represent the analysis results of the first and second parts in each split, respectively. The color represents the two-sided P value calculated by the variogram-null model (corresponding to the color bar in the figure). CCSI strengthens the interpretation of functional connectivity gradient in transmodal regions We next examined whether laminar perfusion-cytoarchitecture coupling, as quantified by CCSI, contributes to understanding macro-scale patterns of cortical structure-function coupling. While studies have demonstrated a close correspondence between cortical microstructure and function67,68, recent human research has revealed a growing dissociation between structural and functional gradients in the transmodal networks7–9. Building on our hypothesis that metabolic constraints provide an additional organizational principle, we examined whether CCSI, interpreted as a mesoscale metric associated with metabolic constraints, can help account for variability in structure-function coupling across the cortex. Figure 8a shows the parcellation map of the microstructure profile covariance gradient (GMPC) derived from BigBrain histological images7, and Fig. 8b presents the principal functional gradient69 (Gfunc) computed from resting-state fMRI data from the Human Connectome Project70 (HCP). At the whole-cortex level (Fig. 8c), GMPC alone explained 23.4% of the variance in Gfunc, consistent with prior studies7. Adding CCSI to the model yielded a notable change in the GMPC coefficient (ΔβG-MPC = 0.086, FDR = 0.032), though the overall model improvement and CCSI coefficient did not survive correction. We then stratified the cortex into lower-hierarchy (sensorimotor) and higher-hierarchy (association) regions based on GMPC (Fig. 8d). In lower-hierarchy regions (Fig. 8e), CCSI contributed negligibly to model fit. In contrast, in higher-hierarchy regions (Fig. 8f), incorporating CCSI significantly improved model performance (ΔR2 = 0.137, FDR = 0.038), substantially altered the structure-function relationship (ΔβG-MPC, FDR = 0.002), and independently predicted functional organization (βCCSI, FDR = 0.032). Although both CCSI and GMPC draw on BigBrain-derived cytoarchitectonic data, they capture fundamentally different properties: GMPC reflects inter-regional similarity in laminar cytoarchitectonic profiles, whereas CCSI quantifies intra-regional correspondence between perfusion and cell density and critically depends on in vivo CBF measurements entirely independent of BigBrain. A variance inflation factor (VIF) of 1.26 confirmed negligible multicollinearity between the two predictors, supporting their inclusion as independent regressors. In contrast, substituting CCSI with CBF-score or CCSI-control yielded no significant improvement in model fit, no significant change in the GMPC coefficient, and no significant independent contribution across any cortical subdivision (Fig. S8), further underscoring the specificity of laminar perfusion-cytoarchitecture coupling as captured by CCSI. These results indicate that perfusion-cytoarchitecture coupling plays a significant role in shaping the functional organization of higher-order association cortex, beyond what microstructural similarity alone can explain. Fitting changes of the structure-function regression model with CCSI. a Parcellation pattern map of principal gradient of microstructure profile covariance (GMPC). b Parcellation pattern map of principal gradient of functional connections (Gfunc). c The spatial distribution relationship between GMPC and Gfunc and the fit changes of the structure-function regression model with CCSI. The black line represents the regression line, the shading is the 95% confidence interval, and dots represent individual brain regions (same as (e, f)). Exact fitting parameter changes, and two-sided P values calculated by the variogram-null model are marked in the figure. The gray dots represent the changes in the fitted parameters distribution calculated using the variogram-null model, and the red dots represent the positions of the true values in the null distribution (same as (e, f)). d Parcellation map of lower and higher hierarchy regions from GMPC. Blue regions are in the lower hierarchy, and red regions are in the higher hierarchy. e The spatial distribution relationship between GMPC and Gfunc and the fit changes of the structure-function regression model with CCSI in the lower hierarchy. f The spatial distribution relationship between GMPC and Gfunc and the fit changes of the structure-function regression model with CCSI in higher hierarchy. Discussion Our study provides a multiscale characterization of cortical perfusion and its coupling with cytoarchitectonic structure, integrating ultra-high-field 7 T ASL with histological, metabolic, and transcriptomic references. We demonstrate that CBF exhibits marked regional and laminar heterogeneity, reflecting spatial variations in metabolic demand and vascular organization. Beyond global perfusion patterns, we introduce the CBF-CSI similarity index (CCSI) as a quantitative and physiologically grounded index of laminar perfusion-cytoarchitecture coupling. CCSI revealed robust region-specific alignment between perfusion and cellular density profiles, showing strong reproducibility across sessions and participants. Importantly, CCSI showed selective associations with mitochondrial respiratory capacity and with the spatial distributions of endothelial and mature oligodendrocyte populations that form neurovascular interfaces. Gene enrichment analysis further highlighted pathways converging on metabolic regulation, neurovascular organization, and mitochondrial homeostasis, suggesting that the spatial heterogeneity of CCSI may reflect integrated metabolic-structural specialization. At the systems level, CCSI enhanced structure-function correspondence in association cortices, where conventional structural gradients show reduced explanatory power. Furthermore, the parallel control analysis using qT1 (CCSI-control) did not replicate any of these associations, reinforcing the specificity of the CCSI-related findings. Together, these results support the interpretation of CCSI as a reproducible, non-invasive mesoscopic index that captures biologically meaningful aspects of laminar perfusion–cytoarchitecture alignment in the human cortex. The development of CCSI addresses a key challenge in cortical neuroscience: quantifying the laminar-scale alignment between vascular supply and cellular organization in vivo. Although the laminar organization of both cell density and microvasculature has long been recognized from histological studies32–34, the extent to which their spatial profiles co-vary across cortical regions has remained largely uncharacterized in the living brain. By combining whole-brain 1-mm isotropic 7 T ASL with cytoarchitectonic profiles from the BigBrain atlas35, we quantified this correspondence across the entire cortex. To guard against confounds arising from cross-dataset analyses, a parallel control analysis substituting CBF with quantitative T1 values (CCSI-control) yielded no significant associations in any downstream analysis, supporting the specificity of the CCSI-related findings. In what follows, we discuss the potential physiological mechanisms reflected by the spatial heterogeneity of CCSI. CCSI is broadly positive across the cortex, indicating that laminar perfusion and cytoarchitectonic profiles are generally aligned in most regions. However, the substantial regional variation in CCSI suggests that this alignment is not uniform but instead shaped by local biological properties. Regions with high CCSI, such as primary sensorimotor and visual cortices, may represent areas where capillary supply is tightly matched to cellular density at each cortical depth, whereas regions with low CCSI, such as portions of the inferior temporal cortex, exhibit notable misalignment, possibly reflecting a greater reliance on alternative metabolic pathways such as glia-mediated substrate shuttling20–23. Understanding which molecular and cellular features covary with this heterogeneity may therefore offer insight into the metabolic and structural principles governing local cortical organization. Our cell-class enrichment analysis revealed that CCSI was significantly associated with both capillary endothelial cells and mature oligodendrocyte populations. The enrichment with endothelial cells is expected, given the inherently vascular nature of ASL-derived CBF measurements, and serves as an internal validation of the analytical framework. More notably, CCSI showed robust and consistent enrichment across mature, end-state oligodendrocyte populations, whereas this relationship was absent for the CBF-score map. This dissociation suggests that laminar perfusion-cytoarchitecture coupling reflects cellular organizational features that are not captured by total blood flow alone, and points to a specific role of mature oligodendrocytes in this coupling. Our enrichment analysis identified end-state oligodendrocytes (es-OLGs) as the cell population associated with CCSI. es-OLGs are mature, metabolically stable oligodendrocytes whose marker gene profiles overlap with those of both perineuronal and perivascular oligodendrocytes63,71,72. Histological studies have shown that these two subtypes frequently coexist within the same microenvironment, bridging neuronal and vascular domains71–74, and recent connectomic reconstruction of the human cortex75 confirmed that oligodendrocytes are the most abundant cell type and exhibit prominent perivascular alignment across cortical layers. This anatomical positioning is consistent with developmental evidence that oligodendrocyte precursor cells migrate along the vasculature under endothelial CXCL12 signaling76 and mature with pericyte support77,78, a trajectory that aligns with our observation that CCSI is also enriched in capillary-associated endothelial and pericyte populations. Functionally, es-OLGs are known to shuttle glycolytic substrates to axons via monocarboxylate transporters21–23 and to buffer oxidative stress through iron sequestration79, thereby supporting neuronal metabolic homeostasis80. These roles suggest that regions enriched in es-OLGs may be better equipped to sustain efficient mitochondrial function, which is consistent with our observation that CCSI correlates with mitochondrial respiratory capacity per mitochondrion (MRC) rather than mitochondrial density. The GO enrichment results further support this interpretation, as the significantly enriched pathways (including glycogen and lipid metabolism, blood vessel development, and mitochondrial organization) largely overlap with the known functional repertoire of es-OLGs and their associated neurovascular microenvironment63,71,72. Taken together, the convergence of cellular, metabolic, and transcriptomic evidence suggests that regions with stronger laminar perfusion–cytoarchitecture coupling are characterized by a higher presence of es-OLGs, although whether this association reflects a direct mechanistic link or a shared organizational principle remains to be determined. At the systems level, we examined whether CCSI contributes to understanding macroscale structure-function coupling. In higher-hierarchy association cortices, incorporating CCSI as an additional predictor significantly improved the explanatory power of the microstructural gradient (GMPC) for the functional gradient (Gfunc). Although both CCSI and GMPC are partially derived from the BigBrain atlas35, they capture fundamentally different properties: GMPC reflects inter-regional similarity in laminar cytoarchitectonic profiles, whereas CCSI quantifies the intra-regional correspondence between perfusion and cell density. Moreover, CCSI critically depends on in vivo CBF data, which is entirely independent of the BigBrain dataset. A VIF of 1.26 between CCSI and GMPC further confirms that multicollinearity is negligible, supporting their inclusion as independent predictors. The significant improvement in model fit and the independent contribution of CCSI in transmodal cortices suggest that laminar perfusion-cytoarchitecture coupling captures variance in functional organization beyond what structural similarity alone explains, consistent with the hypothesis that metabolic constraints provide an additional organizational principle complementing the static anatomical scaffold7–9. Notably, CCSI contributed negligibly in lower-hierarchy sensorimotor cortices where structure–function coupling is already strong, indicating that its contribution is most relevant where conventional structural gradients are least predictive. CCSI offers a physiologically interpretable index of laminar perfusion-structure coupling with broad potential applicability. Because it captures both vascular and oligodendrocyte-related organizational features, CCSI may be particularly informative for disorders in which perfusion–structure coupling is disrupted, such as multiple sclerosis, Alzheimer’s disease, schizophrenia, and epilepsy, where alterations in perineuronal and perivascular oligodendrocyte populations have been documented71,72,74,80. However, CCSI in its current form is a group-level metric referenced to a single histological template and therefore does not capture inter-individual variability. Future efforts to derive individualized laminar microstructural maps, for instance, through high-resolution diffusion MRI in long-time81 or next-generation 7 T scanner82–84 and soma-sensitive diffusion modeling85, could enable individual-specific CCSI computation, potentially extending its utility toward detecting microvascular or metabolic alterations and evaluating neurovascular-targeted interventions in clinical cohorts. Several limitations should be acknowledged. First, the laminar resolution of in vivo perfusion MRI remains limited compared to histological methods, and partial volume effects are inherent at 1-mm isotropic resolution. However, ROI-based laminar analysis, which aggregates signals across a large number of voxels at different cortical depths within each parcel, substantially mitigates these effects46,86,87. Second, CCSI references a single postmortem BigBrain dataset that may not fully capture inter-individual variability in laminar cytoarchitecture. Although canonical laminar organization follows consistent principles across individuals31,36–38, individual variation in cortical folding and cytoarchitectonic boundaries may affect cross-dataset alignment accuracy. Group-level averaging across 30 participants, together with left-right symmetry analysis and test-retest reproducibility, mitigates but does not eliminate this constraint. Validation against additional histological datasets would strengthen generalizability. Third, the mitochondrial and transcriptomic analyses relied on atlases with limited donor samples. The MitoBrainMap48 was derived from a single brain, and the AHBA49 comprises bulk tissue microarray data from only six donors with complete bilateral sampling in only two, which may obscure cell-type-specific contributions. Cross-validation through both GCEA57 and CGEM56 improved robustness, but broader datasets and spatially resolved single-cell transcriptomics are needed to confirm the specificity of these associations. Fourth, all findings are correlational. The spatial co-localization of CCSI with mitochondrial, cellular, or transcriptomic maps does not establish causal relationships, and the enrichment results, while internally consistent, represent post-hoc interpretations. Future longitudinal or experimental studies will be needed to determine whether these associations reflect mechanistic links. Finally, regional SNR variations at 7 T88, particularly in inferior temporal and orbitofrontal regions, may influence CCSI estimates. Although SNR-based quality control was applied and CCSI-SNR correlations were absent in high-SNR regions, residual bias in low-quality regions cannot be fully excluded. In summary, we introduce the CBF-CSI similarity index (CCSI) as a quantitative measure of laminar perfusion-cytoarchitecture coupling in the living human brain. Integrating high-resolution 7T ASL with cytoarchitectonic, metabolic, and cellular references, CCSI reveals meaningful associations with mitochondrial respiratory capacity, oligodendrocyte organization, and macroscale structure-function coupling, extending beyond what conventional perfusion metrics capture. Despite the methodological constraints inherent in cross-dataset analyses, CCSI offers a promising framework for bridging vascular, metabolic, and cellular dimensions of cortical organization. Future work combining mesoscopic multimodal imaging with individualized structural references and clinical cohorts will be essential to refine this framework and realize its translational potential for neuroscience and neurological disease. Methods Ethics statement This study complied with all relevant ethical regulations. The study protocol was approved by the Institutional Review Board (IRB) of the University of Southern California. All participants provided written informed consent prior to participation. Participant sex was determined based on self-report. The study included 18 female and 12 male participants. Sex was not considered as an experimental factor because the objective of this study was to develop and validate an imaging methodology rather than investigate sex-related biological differences, and the sample size was not powered for sex-stratified analyses. Participants Thirty volunteers (mean age = 25.9 years, SD = 5.1 years, 18 females) participated in this study. All participants in this study were healthy without known neurological, psychiatric disorders or major systemic diseases and refrained from caffeine 3 h before the scan. All participants provided written informed consents according to a protocol approved by the IRB of the University of Southern California and were paid for attendance (40 US dollars). Head motion was minimized by placing cushions on the top and two sides of the head. Fourteen of the participants underwent repeated scans on 2 different days for the test and re-test. MRI data acquisition MRI data were acquired on the investigational pTx part of a 7 T Terra system (Siemens Medical Systems, Erlangen, Germany) with an 8Tx/32Rx head coil (Nova Medical, Cambridge, MA, USA). Anatomical images were acquired with a T1-weighted MP2RAGE sequence (0.7-mm isotropic voxels, 224 sagittal slices, FOV = 224 × 224 mm, TR = 4500 ms, TE = 3.43 ms, TI1 = 1000 ms, flip angle = 4°, TI2 = 3200 ms, flip angle = 5°, bandwidth = 200 Hz/Px, slice partial Fourier = 6/8, GRAPPA = 3). For the perfusion image, a whole-brain 3D pCASL imaging with 1 mm isotropic resolution was performed at 7 T using the FLASH-based rotated golden-angle stack of spirals (rGA-SoS)47 sampling scheme with following imaging parameters: FOV = 256 × 256 × 144 mm3, matrix size = 256 × 256 × 144, TR = 7 s, echo spacing (ES) = 19.5 ms, TE = 1.72 ms, ADC bandwidth = 400 k Hz, ADC duration = 10.5 ms, flip angle = 15°, and echo train length (ETL) = 72, acceleration factor (R) = 4. Each M0, label, and control acquisition was acquired over 12 TRs, which were divided into 3 dynamic frames and can be reconstructed into 3 individual images to avoid a lengthy temporal footprint, resulting in an effective temporal resolution for each image of 28 s. In each TR, 1 spiral interleave with variable density (Rspiral = 2) in the kxy plane was acquired across 72 kz samples (Rz = 2). GA rotation (137.5°) was applied to spiral interleaves across 3D spatial space and dynamic frames, enabling superior incoherence in 4D. pCASL labeling has a duration of 1500 ms, and post labeling delay was 1800 ms. To avoid the loss of labeling efficiency due to field inhomogeneity at 7 T. An optimized parameter set (Optimization of pseudo-continuous arterial spin labeling at 7 T with parallel transmission B1 shimming) and a phase adjustment according to a prescan at the beginning of each scan (optimization of pseudo-continuous arterial spin labeling using off-resonance compensation strategies at 7 T) was used. ASL reconstruction We used a novel approach for ASL by integrating compressed sensing (CS) reconstruction with motion-resolved self-navigation to minimize motion artifacts89. Dynamic CS reconstruction incorporating spatial regularization on total variation (TVr) and temporal regularization on finite difference (FDt) across dynamic frames is performed to exploit the sparsity in 4D enabled by the sequence, which is described by, where x and b represent 4D image and k-space data, respectively. The forward operator A incorporates coil sensitivity weighting, non-uniform fast Fourier transform (NUFFT), and k-space undersampling. The inverse problem was solved using the alternating direction method of multipliers (ADMM). Reconstruction was performed using 10 ADMM iterations with 3 nested conjugate gradient steps per iteration. Blurring caused by spiral sampling was mitigated by multi-frequency interpolation (MFI), and spiral trajectory errors were corrected by incorporating the gradient impulse response (GIRF) that was measured on the scanner. Reconstruction was performed using 10 ADMM iterations with 3 nested conjugate gradient steps per iteration, requiring ~90 s and 14 GB VRAM on a Nvidia RTX 5090 for each label/control image. MRI data preprocessing, parcellation and cortical layers definition MRI data preprocessing were analyzed using AFNI90, FreeSurfer91, ANTs92, the mripy package87 (https://github.com/herrlich10/mripy) and MATLAB (R2021a). The preprocessing of the ASL data follows linear motion correction for control and label images, and registration of the average control image to the T1w anatomical image. The perfusion image obtained by subtracting the label image from the control image can be used to obtain the CBF image according to the classic pCASL calculation method93,94. Subsequently, the HCP-MMP1 atlas50 was projected from the SUMA std.141 standard surface space to each participant’s native cortical surface using FreeSurfer-based surface reconstruction and SUMA surface registration. The resulting parcellation was then used to extract voxel-wise CBF signals within each parcel using AFNI. ASL images were upsampled to a finer grid of 0.5 mm3 isotropic spatial resolution to avoid singularities at the edges in angular voxel space using the ‘3dresample’ program in AFNI with the Nearest-Neighbor (NN) interpolation method. The T1w MP2RAGE anatomical volume was segmented into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) using FreeSurfer’s automated procedure and its high-resolution option. A boundary-based algorithm in ANTs was used to co-register ASL control images to structural MP2RAGE images95. 12 cortical layers in gray matter were calculated using AFNI/SUMA and custom Python codes in the mripy package with the equi-volume layering approach96. Although the cortical depths of different brain regions vary97 (From ~4 mm thickness of the prefrontal cortex to ~2 mm thickness of the primary sensory cortex) and the voxel size is not large enough to clearly distinguish the cortical layers, a large number of voxels randomly sampled at different cortical depths allowed us to derive a reliable continuous laminar profile of neural activity46,86,87 (ROI-based analysis). Finally, we can generate the laminar profile of CBF for 360 HCP-MMP1 brain regions in MATLAB. Histological data We utilized the open-access BigBrain repository35, an ultra-high-resolution 3D reconstruction of a postmortem human brain stained for cell bodies, as the source of histological data. Their preprocessing pipeline included the following steps: First, the brain was paraffin-embedded and coronally sectioned into 7400 slices, each 20 μm thick, which were then silver-stained to highlight cell bodies and digitized98. These images underwent manual inspection to identify artifacts such as rips, tears, shears, and stain crystallization. Subsequently, automatic correction procedures were applied, including nonlinear alignment to a postmortem MRI of the same individual acquired prior to sectioning, intensity normalization, and block averaging. Finally, the 3D reconstruction was achieved through a coarse-to-fine hierarchical procedure35. Computations were performed on inverted images, on which staining intensity reflects cellular density and soma size. We downloaded the 3D volume at 200-μm resolution BigBrain data (https://bigbrain.loris.ca/main.php). Geometric meshes approximating the outer and inner cortical interface (i.e., the GM/CSF boundary and the GM/WM boundary) were generated for computing the depth30. Then we used the same aforementioned equi-volume algorithm to calculate the relative depth value of each gray matter voxel via the mripy package (https://github.com/herrlich10/mripy) based on the AFNI and FreeSurfer. Considering that the spatial resolution of our original ASL image is 1 mm-isotropic, we also constructed 12 equivolumetric surfaces between the outer and inner cortical surfaces for BigBrain data. To map the HCP-MMP1 parcellations onto the BigBrain space, we used the BigBrainWarp toolbox99, which provides surface-based registration between BigBrain and standard MRI spaces. Subsequently, we generated the laminar profile of cell-body staining intensity (CSI) from Bigbrain data for 360 HCP-MMP1 brain regions. Laminar similarity calculation between CBF and CSI To investigate the relationship between layer-specific CBF and CSI, we conducted a correlation analysis of the laminar profiles of CBF and CSI across different ROIs. Layer 1 was excluded due to its low cellular density and the presence of large penetrating arteries, both of which may distort the CBF-CSI relationship. Since layer 1 accounts for approximately 10% of the cortical thickness30, we discard the 1 data point from the outermost layer during the calculation. We employed vector cosine similarity to assess the CBF–CSI relationship. Prior to calculating cosine similarity, we z-scored the 11 data points for both CBF and CSI profiles to eliminate baseline differences that could bias the vector angle. The similarity was then computed using the following formula: Here, CBF and CSI are 11-dimensional vectors representing cortical layers. Because cosine similarity values are bounded and skewed, Fisher's z-transformation was applied prior to group-level statistical testing. For interpretability, we converted cosine similarity to angular distance using the arccosine transform and defined the CBF-CSI Similarity Index (CCSI) as: This transformation ensures that higher values indicate greater similarity. Specifically, we first inverted the arccos-derived similarity values (which range from 0–π, with smaller values indicating higher similarity) by subtracting them from π, such that larger values represent greater similarity. We then centered the distribution by subtracting π/2, resulting in a final range of [–π/2, π/2], where 0 denotes no correlation and positive values indicate increasing similarity. Mean bias correction Figure S3a presents the original average bilateral CCSI map across participants. We observed a significant difference in mean CCSI between the left and right hemispheres (paired t-test, P < 0.05; Fig. S3b), whereas no significant difference was found in the standard deviation, indicating that the two hemispheres share a similar distribution shape but differ by a systematic additive offset. Furthermore, the spatial distributions of CCSI between hemispheres were highly correlated (Pearson r = 0.69, variogram-p < 0.001; Fig. S3c), confirming that the regional pattern is well preserved despite the intercept shift (The distributions of both the left and right cerebral hemispheres conform to a normal distribution: Kolmogorov-Smirnov test P = 0.37 and 0.28; and variance homogeneity is satisfied between the two hemispheres: Bartlett test P = 0.10). Since CBF itself showed no significant hemispheric difference (Fig. 2b), this offset is unlikely to reflect biological perfusion asymmetry and more likely arises from factors on the histological side, such as staining intensity heterogeneity or tissue processing artifacts in the BigBrain volume35, and/or B0/B1+ field asymmetries at the 7 T labeling plane51. To eliminate this systematic offset while preserving the original physical range and regional contrast of CCSI, we applied a mean bias correction: This method shifts the distribution between the left and right hemispheres, aligning their means with the whole-brain level, thereby eliminating intercept differences caused by potential staining heterogeneity in BigBrain or 7 T label field asymmetry. This process ensures the statistical rigor of the whole-brain spatial pattern while maintaining the original signal contrast. CBF SNR estimation To estimate the signal-to-noise ratio (SNR) of CBF images in a manner that accounts for both limited repetition and physiological variability, we employed a split-half method100 using interleaved repetitions. Specifically, we divided the CBF time series into odd and even subsets and calculated the mean CBF map for each. For each voxel, the noise standard deviation was estimated using the absolute difference between the odd and even means, scaled by their respective sample sizes: Where µodd and µeven denote the mean CBF images from odd and even repetitions, and Nodd, Neven are their respective counts. The median noise estimate across all voxels within each ROI was then taken as the representative noise level for that region. Finally, the regional SNR was computed as the ratio of the mean CBF value to the estimated noise standard deviation. CCSI quality control by SNR Figure S3d presents the average bilateral CCSI map across participants after bias correction. Overall, most cortical regions exhibit a high degree of similarity between CBF and CSI laminar profiles, with the exception of certain regions within the temporal lobe, where the similarity is relatively weaker. Figure S3e shows the distribution of group-level regional CCSI values after merging hemispheres. Except for the temporal lobe (27/37 regions), nearly all cortical areas showed significantly positive similarity after BH-FDR correction (66/67 in the frontal lobe, 5/6 in the insula, 24/24 in the occipital lobe, and 46/46 in the parietal lobe), broadly consistent with the CCSI map. Given the pronounced B0/B1+ inhomogeneities, increased susceptibility artifacts, and reduced coil sensitivity in inferior temporal regions at 7T88, the lower similarity observed in these areas may reflect diminished signal-to-noise ratio rather than true physiological differences. Figure S3f shows the average SNR map across participants. While the overall spatial distributions of CCSI and SNR differ across the cortex, low CCSI and low SNR co-localize in inferior temporal and orbitofrontal regions near the brain base. Across the entire cortical ROI, the correlation between CCSI and SNR was not statistically significant (Fig. S3g, Pearson’s r = 0.18, two-sided variogram-p = 0.39). However, a significant positive correlation was observed in low-SNR regions (Fig. S3h, SNR < 4, r = 0.34, variogram-p = 0.03), whereas no correlation was found in regions with higher SNR (Fig. S3i, SNR > 4, r = −0.06, variogram-p = 0.71). Therefore, in the subsequent analysis of CCSI, we selected ROIs with an SNR greater than 4 for analysis to exclude CCSI heterogeneity that may be caused by SNR. Mitochondria-related maps We utilized recently published maps of mitochondrial respiratory capacity and diversity from Mosharov et al. 48, in which a single postmortem brain from a neurotypical 54-year-old male donor was physically voxelized into 703 samples at MRI-comparable resolution (3 × 3 × 3 mm). Each voxel was biochemically profiled for mitochondrial density and oxidative phosphorylation capacity using assays of citrate synthase, mitochondrial DNA, and respiratory chain enzyme activities (Complex I, II, and IV), complemented by frozen tissue respirometry. These measures were integrated to generate indices of mitochondrial tissue density (MitoD), tissue respiratory capacity (TRC), and mitochondrial respiratory capacity (MRC = TRC/MitoD), thereby capturing both abundance and specialization of mitochondria across gray and white matter. The resulting voxel-wise mitochondrial features were registered to MNI stereotaxic space and upscaled to whole-brain maps using regression models linking biochemical measurements with multimodal MRI features. For the present study, we used the processed maps of CI, CII, CIV, MitoD, TRC, and MRC provided by the authors in MNI152 template space and resampled them to the HCP-MMP1.0 parcellation by AFNI to enable direct comparison with our imaging-derived metrics. Allen Human Brain Atlas (AHBA) gene expression data We incorporated transcriptomic data from the Allen Human Brain Atlas49 (AHBA), using the optimized preprocessing pipeline58 described by Dear et al. Microarray measurements from six postmortem adult donors were first mapped to the HCP-MMP1 cortical parcellation (180 left-hemisphere parcels). To maximize cross-donor reproducibility, only parcels with samples from at least three donors were retained, yielding 137 cortical regions. Genes were ranked by differential stability (DS) across donors, and the top 50% (7973 genes) were selected to reduce noise. In the present study, we directly utilized the donor-consistent gene × region matrix, which represents a spatially reliable transcriptomic dataset. This matrix was used in its original form for integration with our imaging-derived measures. Variogram-null model generation To assess the statistical significance of spatial correspondences while accounting for spatial autocorrelation, we generated null brain maps using the generative modeling framework proposed by Burt et al. 101 and implemented in the neuromaps Python library69. This method constructs surrogate maps that preserve the empirical degree of spatial autocorrelation by matching the variogram of the observed data, while randomizing its topography. Specifically, the empirical map is first randomly permuted to destroy spatial structure, then spatial autocorrelation is reintroduced via distance-dependent kernel smoothing, with parameters optimized to minimize the error between empirical and surrogate variograms. Repeating this procedure yields a set of surrogate maps with spatial autocorrelation statistically indistinguishable from the empirical data but randomized spatial arrangement. Unless otherwise specified, we used the default settings and generated n = 20000 surrogates for each empirical map, which were subsequently used to derive null distributions of test statistics. To maintain the repeatability of our statistical estimates, we fixed the random seed at 1234. This approach provides spatially informed null models that offer more conservative and interpretable benchmarks than conventional permutation-based tests, which assume independence among regions. Gene category enrichment analysis (GCEA) We performed gene category enrichment analysis (GCEA) following the ensemble-based framework proposed by Fulcher et al. 57,102. For each gene, we calculated the spatial correlation between its AHBA-derived expression profile and the phenotype map of interest. Within each predefined gene set (GO and cell class), these correlations were Fisher’s z-transformed and summarized using the median as a robust category score, reducing sensitivity to a small number of highly deviant genes within each set. Statistical significance was assessed by comparing the observed category scores to null distributions as described in the variogram-null model generation section of the methods, with empirical P values BH-FDR-corrected across categories. Cell-class gene expression mapping (CGEM) When conducting cell category enrichment analysis, we also used the cell-class gene expression mapping56 (CGEM) method for cross-validation. For each cell category, a spatial gene-expression map was generated by computing the median expression value across all genes assigned to that category for each cortical region. The resulting category-level expression maps were then correlated with CBF or CCSI maps using Pearson’s correlation, and the correlation coefficients were Fisher z-transformed to yield enrichment scores. To assess statistical significance, null distributions of enrichment scores were constructed by repeating the same correlation using 20,000 variogram-null surrogate maps of CBF or CCSI that preserved spatial autocorrelation but randomized topography. p values were computed as the proportion of null scores exceeding the empirical enrichment score for each cell class. Gene set selection of GO items To derive functional gene sets for enrichment analysis, we mapped the 7,973 reliably expressed genes retained from the AHBA preprocessing to Gene Ontology103 (GO) annotations using the R package biomaRt, interfacing with the Ensembl database104. Gene symbols were first standardized (e.g., resolving ambiguous aliases such as MARCH and SEPT families) before querying. For each gene, associated GO identifiers, term names, and namespaces were retrieved via the getBM function. Only terms belonging to the Biological Process ontology were retained, while Molecular Function and Cellular Component categories were excluded. To ensure statistical robustness, we further restricted the analysis to GO categories containing at least 20 AHBA-mapped genes. The resulting curated collection of GO gene sets was used for downstream enrichment analyses. Gene set selection of cell class Gene sets for major cortical cell types were obtained from Lake et al. 59, who profiled over 60,000 single nuclei from human cortex and identified cell-type-specific marker genes distinguishing excitatory and inhibitory neurons, astrocytes, oligodendrocytes, OPCs, microglia, and endothelial cells. Vascular-associated cell subtypes, including endothelial and mural (pericyte and smooth-muscle), were derived from Garcia et al. 62. Oligodendrocyte lineage subclusters, including mature and end-state oligodendrocytes, were obtained from Jäkel et al. 63. Because the gene sets representing different mature oligodendrocyte states showed substantial overlap, direct enrichment analyses risked confounding effects driven by shared genes rather than state-specific expression. To enhance the specificity of oligodendrocyte subtype comparisons, we implemented a two-step gene-set refinement procedure. First, we removed duplicate genes shared between any two of the six mature oligodendrocyte state gene sets to obtain non-overlapping subtype-specific gene lists, as shown in Fig. S7a. This stringent filtering substantially reduced gene-set sizes for certain subtypes, occasionally leading to unstable or unreliable enrichment estimates. Next, given that CCSI showed stronger associations with mature oligodendrocytes biased toward end-state phenotypes, we further refined specificity by removing genes shared between the end-state oligodendrocyte gene set and all other oligodendrocyte states, yielding the final gene-set distribution shown in Fig. 5g. Other cortical maps The PET-CBF, ASL-CBF, CMRO2, and CMRglc maps used in Fig. S2 were obtained from previously published datasets53,54 compiled within the neuromaps toolbox69. The functional gradient map shown in Fig. 8b was also derived from the neuromaps toolbox69. The cell microarchitecture and structural gradient map presented in Fig. 8a were generated from BigBrain cell body staining intensity35 profiles processed using the ENIGMA toolbox105. Structural-functional coupling analysis To assess whether CBF, CCSI, or CCSI-control contributes to large-scale functional organization, we examined their effects on the relationship between cortical microstructure and function. Specifically, we quantified structure-function coupling by relating the principal functional gradient (Gfunc) to the principal microstructural gradient (GMPC) across cortical regions defined by the HCP-MMP1.0 atlas. We first established a baseline model using a simple linear regression: which captures the canonical correspondence between microstructural and functional gradients. We then extended this model by including either CBF, CCSI, or CCSI-control as an additional predictor: where X denotes the regional CBF, CCSI, or CCSI-control. Model performance was evaluated using the coefficient of determination (R2). The improvement in model fit (ΔR2) was computed as the difference between the extended and baseline models, quantifying the additional proportion of variance in Gfunc explained by X beyond microstructural gradients. In addition, changes in the microstructural coefficient (ΔβMPC = β’MPC - βMPC) were used to assess whether inclusion of X modulates the strength of structure–function coupling. The regression coefficient of the added predictor (βX) was also examined to evaluate its direct association with functional organization. To account for potential hierarchical differences, cortical regions were further stratified into lower- and higher-order groups based on the principal microstructural gradient (GMPC). The above analyses were repeated within each subgroup to assess whether the contributions of CBF, CCSI or CCSI-control vary across the cortical hierarchy. Statistical significance of ΔR2, ΔβMPC, and βX was assessed using 20,000 spatial permutations generated from variogram-preserving null models. Statistical analysis All statistical analyses were performed in MATLAB (R2021a). Group-level statistical significance of regional CCSI values was assessed using one-sample t-tests against zero, with P values corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate (BH-FDR). A one-way repeated-measures ANOVA was applied to evaluate region-specific differences in CCSI, with Greenhouse-Geisser correction applied where sphericity was violated. Prior to conducting the ANOVA test, each dataset was evaluated for normality using the Shapiro-Wilk test and for homogeneity of variance using Levene’s test. Spatial correlations between imaging-derived maps (e.g., CBF, CCSI) and histological, metabolic, or transcriptomic references were evaluated using variogram-based permutation tests (n = 20,000) implemented in the neuromaps using the variogram framework proposed by Burt et al. 101, which preserves spatial autocorrelation of cortical topology. For gene and cell-class enrichment analyses, empirical correlations were Fisher z-transformed and compared against null distributions generated from the same variogram-based null permutations, with significance determined after BH-FDR correction across categories. In the structure-function coupling analysis, model improvements (ΔR2) and coefficient changes (ΔβMPC and β1) were evaluated via identical permutation tests. Unless otherwise specified, all tests were two-tailed with a significance threshold of P < 0.05 and corrected for multiple comparisons using the BH-FDR. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting summary Transparent Peer Review file Source data Source Data of all Acknowledgments The authors sincerely thank all study participants for their time and commitment to this research. And the authors acknowledge the developers and contributors of the BigBrain, Allen Human Brain Atlas, MitoBrainMap, ENIGMA Toolbox, and neuromaps projects for making these valuable resources publicly available. Author contributions F.G. and D.J.W. conceived the study. F.G. and D.J.W. designed experiments. F.G., C.Z., Z.L., and A.J.K. executed experiments. F.G., C.Z., and R.R.B. analyzed experiments. C.Z. provided support to ASL. F.G. and D.J.W. drafted the manuscript. F.G., C.Z., R.R.B., Z.L., A.J.K., Z.Y., S.X., K.J., X.S., M.M., N.J., and D.J.W. interpreted the results, reviewed the paper, and approved the decision to submit the paper. Peer review Peer review information Nature Communications thanks Boris Bernhardt and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by US National Institute of Health (NIH) grants UF1-NS100614, S10-OD025312, R01-EB032169, RF1-AG084072, R01-MH134004, R01-AG082073, and R01-NS134712. Data availability The MP2RAGE and ASL datasets generated in this study are publicly available in the OpenNeuro repository under accession ds006871 (https://doi.org/10.18112/openneuro.ds006871.v1.0.0). Source Data are provided with this paper. Publicly available datasets and resources used in this study include the BigBrain atlas35 (https://bigbrain.loris.ca/main.php), the Allen Human Brain Atlas49 (https://human.brain-map.org/), MitoBrainMap48 (http://humanmitobrainmap.bcblab.com), cortical maps distributed through the neuromaps69 (https://netneurolab.github.io/neuromaps/) and ENIGMA toolboxes105 (https://enigma-toolbox.readthedocs.io/), and published cell-type marker gene sets, including major cell classes59, vascular cell subtypes62, and oligodendrocyte-state marker genes63, as described in the Methods. Source data are provided with this paper. Code availability The custom analysis and plotting code developed in this study is publicly available at GitHub (https://github.com/FanhuaGuo/ASL-Bigbrain-CCSI) and archived in Zenodo (https://doi.org/10.5281/zenodo.21363179). Competing interests All authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Fanhua Guo, Chenyang Zhao. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76812-w. References Associated Data Supplementary Materials Supplementary Information Reporting summary Transparent Peer Review file Source Data of all Data Availability Statement The MP2RAGE and ASL datasets generated in this study are publicly available in the OpenNeuro repository under accession ds006871 (https://doi.org/10.18112/openneuro.ds006871.v1.0.0). Source Data are provided with this paper. Publicly available datasets and resources used in this study include the BigBrain atlas35 (https://bigbrain.loris.ca/main.php), the Allen Human Brain Atlas49 (https://human.brain-map.org/), MitoBrainMap48 (http://humanmitobrainmap.bcblab.com), cortical maps distributed through the neuromaps69 (https://netneurolab.github.io/neuromaps/) and ENIGMA toolboxes105 (https://enigma-toolbox.readthedocs.io/), and published cell-type marker gene sets, including major cell classes59, vascular cell subtypes62, and oligodendrocyte-state marker genes63, as described in the Methods. Source data are provided with this paper. The custom analysis and plotting code developed in this study is publicly available at GitHub (https://github.com/FanhuaGuo/ASL-Bigbrain-CCSI) and archived in Zenodo (https://doi.org/10.5281/zenodo.21363179).