GABA_A receptor α subunit distribution is congruent with functional anatomical areas in the human cortex

Abstract

Background:

We recently published putative in vivo maps of GABA_A receptor subtypes—α₁/α₃, α₂, and α₅. Here we tested how these maps align with functional anatomical cortical areas and large-scale functional networks.

Methods:

Cortical α subunit maps were projected to the symmetric fs_LR template and compared with the Human Connectome Project multi‑modal parcellation. Congruence was assessed by visual inspection, within‑parcel variability, and the distance‑controlled boundary coefficient (DCBC) with spin‑permutation null models. Subunit profiles were summarized per cortical area and related to functional networks.

Results:

Inspection revealed conspicuous congruence between α -subunit distributions and areal boundaries. α₁/α₃ subunits predominated in primary sensory/motor areas, α₂ peaked in neighboring secondary areas and α₅ was concentrated in cingulo-opercular and insular areas. Within‑parcel variability was low, and congruence was supported by a positive, significant DCBC (0.1065, p-spin = 0.0006) exceeding the random parcellation. Large‑scale networks showed characteristic subunit combinations: functionally “narrow” networks were subunit-predominant, while associative networks clustered around mixed patterns away from extremes.

Conclusion:

Cortical GABA_A receptor α subunit distributions align with multimodal areas and display network‑specific profiles, linking inhibitory chemiarchitectonics to cortical organization. The present analysis supports integration of molecular with structural and functional imaging for extended understanding of human brain function.

ClinicalTrials.gov: NCT00681720 and NCT00681746

Introduction

The human brain, especially the cortex, is composed of areas with regional specificity in histological tissue composition.^1,2 The functional relevance of such histologically and anatomically defined areas has since long been a matter of intense debate. ² Early research relied primarily on observational lesion data related to loss and re-acquisition of brain function in patients with brain injury or stroke, as well as to sensory and motor responses evoked following direct electrical stimulation of the cortex.^2,3

Over time, the extreme views claiming either precise functional localization, or, conversely, functional holism across the cortex have both been rejected. ² Phenomena such as neural plasticity, remodeling, dynamic network adaptation as well as the common existence of interindividual differences are now well recognized. Importantly, a major aim in current functional brain research is to discover predominant regional patterns in the organization and cooperation between cortical foci in relation to cognitive tasks, that is, uncovering functional connectivity or neural networks.

Current brain research relies heavily on noninvasive neuroimaging using structural and functional MRI. These two MRI-based imaging modalities enable parcellation of the cortex into areas with homogeneous within-area anatomical composition and dominant functional specialization. This approach has recently been taken in a seminal work using the Human Connectome Project’s (HCP) neuroimaging data set. ⁴ In particular, based on a number of structural and functional properties HCP multi-modal parcellation (HCP-MMP) of the cerebral cortex designated areas with a remarkable degree of reproducibility across subjects. ⁴ Subsequent work has then utilized the HCP-MMP as a basis to identify large-scale networks of functional specialization in a highly reproducible manner. ⁵

Another potential source of information is neuroreceptor mapping using molecular imaging to enquire into the chemiarchitectonics of the brain. This approach enables synergistic, molecular-enriched multimodal analysis of functional imaging data. ⁶ The pentameric GABA_A receptor (GABA_AR) complex has been one of the most studied neurotransmitter targets using PET. ⁷ Of particular interest is its benzodiazepine (BZD) sensitive subtypes containing one of the α₁, α₂, α₃, or α₅ subunits at the BZD binding site. ⁸ In a recent work, we published three high-resolution putative cortical and subcortical maps visualizing the distributions of BZD-sensitive GABA_AR α subunit variants, namely maps of GABA_AR’s containing the α₁/α₃ (denoted below simply as α_1 + 3), α₂ and α₅ subunits at the BZD binding site, respectively. ⁹ That work was based on [¹¹C]flumazenil PET images at baseline and under heterologous competition conditions using two GABA_AR partial agonist drug candidates, AZD7325 and AZD6280, with differing subtype selectivity profiles to derive the subunit maps. The analysis relied on the fact that flumazenil has equipotent affinity to all benzodiazepine-sensitive GABA_ARs. In contrast, AZD7325 has high affinity to α₁, α₂, and α₃ subunits (K_i of 0.3 to 1.3 nM in vitro), and much lower affinity to α₅ subunits (K_i = 230 nM). ¹⁰ AZD6280 in turn has high affinity to α₁ subunits (K_i = 0.5 nM), lower affinity to α₂ (K_i = 21 nM) and α₃ (K_i = 31 nM) subunits, and very low affinity to α₅ subunits (K_i = 1680 nM). ¹¹ Importantly, the three α subunit maps showed unique regional patterns with sharp transitions in the levels of α subunits between neighboring points on the cortical surface as well as in subcortical grey matter. ⁹

To gain further insight into the biology of the GABA_AR system a step forward is to investigate whether and how these nuanced regional patterns may relate to brain areas identified based on structure and function as well as to large-scale functional networks. Accordingly, our primary aim in the present work was to build on the findings of the HCP multi-modal parcellation by employing it as a scaffold and assess the alignment of our previously identified α subunit distributional patterns with neuroanatomy and function. As part of preparing the α subunit maps for this aim, the symmetric template space used by the HCP also enabled a more fine-grained evaluation of the symmetricity of the α subunit patterns, which in our prior work was considered at the cortical parcel level of the Desikan–Killiany atlas. ⁹ Finally, the α subunit patterns were positioned in terms of previously published functional network assignments based on the HCP-MMP atlas.

Methods

PET images used in the present analysis had been obtained in two PET studies sponsored by AstraZeneca (ClinicalTrials.gov: NCT00681720 and NCT00681746). Participants, experimental procedures and the image analysis methodology used to obtain the overall [¹¹C]flumazenil binding as well as the fractional and absolute α subunit specific cortical maps have previously been described in detail.^9,12 Importantly, both studies were approved by the Regional Ethical Review Board in Stockholm, and the Radiation Safety Committee at Karolinska University Hospital (KUH), Stockholm, Sweden. The studies were performed in accordance with the Declaration of Helsinki and International Conference on Harmonization/Good Clinical Practice Guidelines. Written informed consent was obtained from participants before inclusion in the studies.

High-resolution research tomograph [¹¹C]flumazenil PET images at baseline and following administration of various doses of one of two drugs with differing partial α subunit selectivity (AZD7325 and AZD6280) had been analyzed (N = 22). Parametric images of [¹¹C]flumazenil binding had been obtained and cortical surface and subcortical volume maps of binding had been extracted and projected to the cvs_avg35_inMNI152 template space using FreeSurfer’s combined surface and volume (CVS) registration pipeline. ¹³ The [¹¹C]flumazenil binding maps had then been entered into an analysis to identify GABA_AR occupancy components of the two drugs. The major outcome parameters of the fitted models had been maps of the fractional and absolute (i.e. in terms of the non-displaceable binding potential, BP_ND) contributions of these components to the overall [¹¹C]flumazenil binding. Three independent components had been detected and had then been putatively identified using GABA_AR gene expression data as corresponding to α subunits 1 + 3, 2 and 3, respectively. ⁹

Extended analysis

Template transformation of α subunit maps and symmetricity assessment

To explore α subunit maps in the context of functional anatomical areas of the brain, we applied the Human Connectome Project Multi-Modal Parcellation (version 1.0) of cortex (HCP-MMP). ⁴ The HCP-MMP (version 1.0) cortical labels are available for 180 areas per hemisphere surface in the symmetric fs_LR template space (upper panel in Figure 1). ⁴ However, our α subunit maps were originally obtained in the cvs_avg32_inMNI152 template space. ⁹ Accordingly, the cortical α subunit maps had to be transferred to the fs_LR template space. This was executed in two steps. First, the maps were transformed to FreeSurfer’s fsaverage space using the spherical registration of the cvs_avg32_inMNI152 surface to the space that was readily available in the FreeSurfer reconstruction output for the cvs_avg32_inMNI152 template. This step was executed using FreeSurfer. Then, the maps were further transformed into fs_LR template space (both to the 164k and the 32k resolution variants) using HCP workbench software suit and readily available spherical registration files in the HCP pipelines’ standard mesh dataset online (https://github.com/Washington-University/HCPpipelines/tree/master/global/templates/standard_mesh_atlases). The analysis was based on the absolute α subunit contributions to [¹¹C]flumazenil binding, that is, the α subunit binding (BP_ND) maps. ⁹ Moreover, the fs_LR template enabled visualization and vertex-wise numerical assessment of the symmetricity of the α subunit density maps. Accordingly, the cortical surface asymmetry index (AI %) was calculated for each left-right symmetric vertex pair using the following expression:

Figure 1.

Areas of the cortical surface according to Human Connectome Project Multi-Modal Parcellation (HCP-MMP, version 1.0) displayed on the flattened fs_LR cortical surface on top.

Source: Data for interactive viewing at https://doi.org/10.12751/g-node.6ldy1i.

The legend inserted under the images in the middle shows the 3D color space used for coloring areas. ⁴ Functional network assignment of cortical areas according to the Cole-Anticevic Brain-wide Network Partition (CAB-NP, version 1.0) are shown in bottom. ⁵ The text color of the network labels matches the color used to indicate network membership for each cortical area (note: gray label background was used for the lightest colors to improve readability, but gray itself is not part of the color code). 3D views of the same type of information projected on the inflated cortical surfaces of both hemispheres can be found in Supplemental Figure S1.

$$AI(\%)=\frac{\mathrm{Left}-\mathrm{Right}}{\mathrm{Left}+\mathrm{Right}}\times 200$$ (1)

The median AI value across all vertices was used as the overall AI of a given α subunit.

Assessment of the congruence between α subunit maps and HCP-MMP areas

We assessed the congruence between the cortical α subunit maps and the HCP-MMP areas both visually and quantitatively. The quantitative congruence assessment included both univariate and multivariate approaches.

Visual inspection was based on displaying the α subunit maps on the flattened or inflated cortical surface of the fs_LR template with an overlay of the HCP-MMP areal boundaries. Besides inspecting the individual α subunit maps separately, a comprehensive comparison and inspection was enabled by a composite view of all 3 maps using a 3D color look-up table (LUT). In detail, each of the red, green and blue color channels was assigned to display one of the 3 α subunit maps (i.e. α_1 + 3, α₂ and α₅), respectively, with the color intensity range from 0 to 1 corresponding to 0 to quasi-maximum (99th percentile to avoid outlier effects) cortical density for each given α subunit map.

The univariate quantitative assessment of the congruence included calculation of the vertex-wise coefficient of variation (CV) within each cortical area and for each α subunit map. Then, for each α subunit map, the histogram of the CVs was plotted to show the within-area homogeneity of α subunit densities.

Finally, the multivariate evaluation enabled an overall quantitative assessment of the congruence between the α subunit maps and the HCP-MMP atlas. In detail, the assessment was carried out by calculating the previously-described distance-controlled boundary coefficient (DCBC), normalized by the average within-parcel vertex distance. ¹⁴ At its core the DCBC algorithm calculates the correlation between the α subunit densities of pairs of cortical vertices. The DCBC criterion is a measure of the difference between the correlation of within-parcel vertex pairs and between-parcel vertex pairs. As such, a parcellation that “lines up” with, that is, is able to predict, the transitions, boundaries present in the density patterns is expected to have a positive value. ¹⁴ A random parcellation with similar parcel-wise surface areas, however, is expected to yield a near-zero DCBC criterion. Note that even a negative DCBC criterion is possible, indicating a parcellation that is worse than a random one when it comes to predicting boundaries.

However, interpretation of the DCBC criterion requires further consideration (or “attention”) as the value does not simply fall within pre-definable limits, unlike, for example, a simple linear regression coefficient. Accordingly, to understand the scale and statistical significance of the calculated DCBC criterion, the value was compared to DCBC values obtained using random parcellation and so-called null models of the α subunit maps. Here, a null model of the α subunit maps is a transformed set of maps obtained with a spin permutation: that is, a shift of all vertex-wise α subunit density values (a set of 3 values) along the spherical cortical surface according to a random 2-axis rotation. 10,000 null models (spin permutations) were obtained. The random parcellation used was equivalent to the one published by Zhi et al. ¹⁴ with a similar number of per hemisphere parcels as the HCP-MMP, labelled “Icosahedron” (162 parcels with similar parcel-wise surface areas in the spherical projection).

The DCBC criterion was calculated for each of the null models of the α subunit maps for both parcellations. Furthermore, to remove any potential residual effect of the difference between parcellations regarding average parcel area, the criterion was normalized by the average distance between within-parcel vertex pairs of the given parcellation. Then, using the obtained DCBC criteria, the scale of the true (non-permuted) DCBC for the HCP-MMP was established by comparing it to the mean and, respectively, maximum DCBCs calculated across all null-model-specific values for the Icosahedron parcellation. Finally, the statistical significance of the true DCBC criterion was established by calculating a permutation-based p-value (a.k.a. p-spin): that is, dividing the number of spin-permutations with a DCBC for the HCP-MMP equal to or larger than the true DCBC, with the total number of null models (10,000).

Relating α subunit density patterns to functional network assignments

As a secondary aim we sought to evaluate how the α subunit patterns may relate to large-scale functional brain networks. While the color codes associated with the HCP-MMP cortical areas in the original publication by Glasser et al. ⁴ already hint at functional specialization according to association with five functional groupings (upper panel in Figure 1), the color encoding therein is meant primarily as a visual aid and not as a rigorous assignment to functional networks. In the current analysis we instead relied on the Cole-Anticevic Brain-wide Network Partition (CAB-NP, version 1.0), which made use of the HCP-MMP atlas to identify large-scale functional cortical (and subcortical) network communities in a statistically and neurobiologically driven manner. ⁵ In detail, the CAB-NP identified 12 networks, and the assignment of cortical areas to these networks was shown to be reproducible and largely symmetrical (lower panel in Figure 1).

To facilitate the evaluation of relations of the α subunit maps to functional networks assignments, the maps were parcellated, that is, the average scaled α subunit density was calculated for each cortical area (0–1). In addition, the composite parcellated α subunit maps were visualized using the above-described 3D LUT on the cortical surface overlaid with color-coded boundaries of the functional network assignments. Finally, 3D histograms of scaled α subunit densities were obtained and visualized for each network in an attempt to reveal their characteristic α subunit patterns.

Results

α subunit maps in fs_LR template space and symmetricity assessment

The maps of α subunit density distribution in the cerebral cortex were scaled relative to the quasi-maximum (i.e. the 99th percentile) cortical value for the given subunit and shown on the flattened fs_LR template cortical surface (Figure 2). The 3D inflated cortical surfaces of both hemispheres are shown in Supplemental Figure S2. Furthermore, interactive viewing of the maps on the flattened or on the 3D cortical surface is possible by accessing the data and figure scene files, in Connectome Workbench viewer format, online at https://doi.org/10.12751/g-node.6ldy1i. Each subunit had a distinct pattern of distribution, reaching their respective cortical maximums in non-overlapping regions. In detail, the α_1 + 3 subunit had its cortical maximum in the occipital cortex, around the calcarine sulcus and was at a low or negligible level in limbic regions. The α₂ subunit was at its maximum in a broad band of regions throughout the opercular and parietal cortex and was low in limbic regions as well as around the central sulcus and superior frontal cortex. The α₅ subunit, in contrast, was at highest in the limbic regions with low or negligible levels where the other subunits were at their highest levels, that is, in the pre- postcentral, opercular, parietal, and occipital cortex.

Figure 2.

α subunit specific absolute (BP_ND) contribution to total [¹¹C]flumazenil binding in cerebral cortex displayed on the flattened cortical surface.

Source: Data at https://doi.org/10.12751/g-node.6ldy1i.

The color-scale encompasses the subunit-specific cortical range of BP_ND with the quasi-maximum (99th percentile) value shown above each pair of maps. Areas of the cortical surface according to the HCP-MMP atlas are indicated with boundary lines. (3D views of the same type of information projected on the inflated cortical surfaces of both hemispheres can be found in Supplemental Figure S2).

Projecting the subunit maps to the symmetric fs_LR template enabled visual confirmation of the symmetric nature of α subunit distributions when considering the overall large-scale patterns and gradients. However, closer inspection revealed left-right differences and pattern variations at a finer scale. The histograms of vertex-wise AIs and the median AI values for the whole cortex are shown in Figure 3. In summary, α_1 + 3 and α₅ subunits display a left-sided asymmetry while α₂ subunit has a similar-sized right-sided asymmetry.

Figure 3.

Histograms of vertex-wise asymmetry index (AI) for cortical α subunit maps with median values shown in the legend.

Congruence between α subunit maps and HCP-MMP cortical areas

The areal borders of the HCP-MMP atlas were overlaid on the cortical surface maps to support visual comparison of the congruence between the GABA_AR system and the cortical areas (Figure 2, Supplemental Figure S2). To further facilitate visual assessment of the congruence, a composite view of α subunit distribution using a 3D color LUT is shown in Figure 4 (inflated cortical view in Supplemental Figure S3). For all three α subunit maps, the areal borders tended to line up with transitions between zones of markedly different α subunit levels.

Figure 4.

Composite α subunit distribution map displayed on the flattened cortical surface using a 3D color lookup table (LUT) obtained by assigning the scaled density of each of the 3 α subunits (α_1 + 3, α₂ and α₅) to the primary red, green, and blue color channels, respectively.

Source: Data at https://doi.org/10.12751/g-node.6ldy1i.

The boundaries between cortical areas according to the HCP-MMP atlas are overlaid using black border lines. The charts at the bottom show the 3D LUT entries, relating scaled subunit-specific densities to colors. Only those entries are visible which are present in the composite map. (3D views of the same type of information projected on the inflated cortical surfaces of both hemispheres can be found in Supplemental Figure S3).

The correspondence between α subunit maps and cortical parcellation appeared to be particularly prominent at and around the primary sensory and motor areas, and was discernible even in cortical regions distant to these areas (Figures 2 and 4, Supplemental Figures S2 and S3). In detail, α_1 + 3 subunit density was highest in cortex precisely in the primary visual cortex (V1) where there is a remarkable congruence of the HCP-MMP areal boundary and the outer contour of the zone of the most intense α_1 + 3 subunit density (Figures 2 and 4). Furthermore, α_1 + 3 subunit density reached high levels (⩾80% of the cortical maximum) focally in the primary auditory area (A1) in such a way that even the slight left-right asymmetry of the topology and areal boundaries of A1 was closely mirrored in the pattern of the corresponding “hot spots” in the subunit map (Figures 2 and 4 with light reddish orange color in the 3D LUT). Finally, while α_1 + 3 subunit density was only moderately high (60%–70% of the cortical maximum) in the primary sensorimotor regions (such as areas 3b, 3a, 1) this subunit was anyhow predominant in these areas since the densities of the other subunits (α₂, α₅) were low (⩽20% of their cortical maximum) (Figures 2 and 4). In addition, this observation was visually corroborated by the deep red or plum color of the primary sensorimotor cortex in the composite view of the α subunit maps (Figure 4).

The congruence between the α subunit maps and the HCP-MMP atlas was also evaluated quantitatively. The vertex-wise coefficient of variation (CV) for each α subunit map was calculated within each cortical area. Figure 5 shows the histograms of within-area CV, which indicated that in most areas the CV was below 30% with the typical value at ~5% for the α₁₊₃ subunit, and at ~15% for the other subunits (Figure 5).

Figure 5.

Histogram of vertex-wise within-area coefficient of variation (CV%) for the different α subunit maps.

Beyond this univariate evaluation, however, the overall congruence between the HCP-MMP atlas and the α subunit maps was calculated using the distance-controlled boundary coefficient (DCBC). The DCBC criterion was 0.1065 with a positive value indicating that the pairwise correlation of α subunit densities between cortical surface vertices that are within the same cortical area tended to surpass those between vertices that are in different areas.

To further understand the scale and statistical significance of the calculated DCBC criterion, the value was compared to DCBC values obtained using the random parcellation (Icosahedron) and DCBCs calculated for null models (random spin permutations) of the α subunit maps. First, the mean DCBC for the Icosahedron parcellation was close to zero (0.0001) with a maximum at 0.026 (Figure 6), indicating that the calculated HCP-MMP criterion was four-fold higher than this maximum value. Second, the permutation-based p-value (p-spin) for the HCP-MMP criterion was 0.0006, that is, significant at the 5% level (Figure 6). The density distribution of DCBC criteria obtained for the spin permutation null-models showed that most null models had a positive DCBC criterion with the most common value close to the maximum of the Icosahedron-based values (Figure 6). Importantly, there was a rapid drop in the number of null models with higher-than-typical values. However, the lower part of the distribution was more tapered as it had a long negative tail with the lowest DCBC value close to −0.5. As such, the density distribution of the null models’ DCBC values supported the uniqueness of the observed value.

Figure 6.

Distance controlled boundary coefficient (DCBC) criterion calculated on GABA_A receptor α subunit cortical maps for random parcellation (Icosahedron) and the HCP-MMP (version 1.0) areas. ⁴

The violin plots show the histograms of null models obtained using random-spin permutations (10,000 permutations) of the α subunit maps. The magenta line and number show the average DCBC across permutations for Icosahedron and the true DCBC for HCP-MMP, respectively. The spin-based p-value for HCP-MMP is shown on top of the chart.

Relating α subunit density patterns to functional network assignments

The average scaled α subunit density was calculated for each cortical area of the HCP-MMP atlas. Figure 7 shows a composite view of the parcellated scaled densities for the 3 α subunit maps using the same 3D color LUT as the one employed in Figure 4. Visual inspection showed that there is no area in which all 3 α subunit would have near-zero density, that is, if one of the α subunits is near zero then at least one other α subunit has a substantial scaled density (>50% of its maximum). Similarly, there is no area where all 3 α subunits would be near their cortical maximums, that is, if one of the α subunits is near its cortical maximum then at least one other α subunit is below 50% of its maximum.

Figure 7.

Composite α subunit distribution values summarized for each functional network.

Source: Data at https://doi.org/10.12751/g-node.6ldy1i.

CX: Cortex.

The α subunit distribution for each HCP-MMP cortical area is shown on the flattened cortical surface using a 3D color lookup table (LUT) obtained by assigning the scaled density of each of the 3 α subunit maps (α_1 + 3, α₂ and α₅) to the primary red, green, and blue color channels, respectively. The 3D LUT is shown below the cortical map and relates scaled subunit-specific densities to colors. Only those entries present in the parcellated composite map are visible. Functional network assignment of cortical areas is indicated using colored boundary lines matching the color of the network labels used in the summary 3D histogram charts below (Note: Same network colors as in Figure 1; grey label background used in case of the lightest colors to improve readability, but gray itself is not part of the color code). When multiple neighboring HCP-MMP cortical areas belong to the same functional network then the boundaries between them are shown using thin dark border lines. The 3D histograms in the lower panel summarize the α subunit distribution for each functional network (Note: The axis tick labels of the large 3D LUT can be used to read off scaled α subunit densities in the histograms). In detail, 3D LUT entries that are present in the cortical area members of the given functional network are indicated by colored circles and the size of the circle corresponds to the fraction of member areas containing the corresponding α subunit density combination. (3D views of the same type of information projected on the inflated cortical surfaces of both hemispheres can be found in Supplemental Figure S4).

We then explored the α subunit densities for the functional networks with the help of 3D histograms (Figure 7). Accordingly, the areas with highest α_1 + 3 subunit density or predominance belonged to the unimodal, sensori/motor networks (Visual1, Visual2, Somatomotor, Auditory). There was a gradual transition with decreasing α_1 + 3 subunit and increasing α₂ and/or α₅ subunit densities from the core primary area of each of the unimodal networks towards the boundaries with neighboring associative networks. The α₂ subunit, in turn, reached its highest cortical densities in secondary processing areas at the outskirts of networks such as the Visual2 or Somatomotor communities. Finally, the α₅ subunit density was considerable in associative functional networks with cingular, opercular or insular members. The α₅ subunit, however, was most exclusively expressed especially in areas of the Orbito-Affective network, such as Pir, and there was a step difference toward lower α₅ and/or higher α₂ subunit levels, in neighboring areas belonging to other functional networks.

What remains is to consider the α subunit density patterns of the largest, most scattered networks. While the spatially extensive and scattered high-level functional networks (Cingulo-Opercular, Fronto-Parietal and Default) were more heterogeneous in terms of their α subunit composition, they still had predominant combinations around which the α subunit patterns of areas were clustered. For example, the anatomically cingular and opercular areas of the Cingulo-Opercular network had the highest α₅ densities in cortex with only low densities for the other α subunits. Other areas functionally assigned to this network were clustered around a combination with roughly average levels for all α subunits with a slight dominance of α_1 + 3. The largest, that is, the Default network, had similar cluster centers with somewhat lower α₅ subunit density for its cingulo-opercular and somewhat higher α₅ density for its non-cingulo-opercular members.

Discussion

The present work is an extended analysis of our recently published high resolution cortical and sub-cortical maps of BZD-sensitive GABA_AR α subunit distribution, that is, maps of α₁ + α₃, α₂ and α₅ subunits. ⁹ Here we first aimed to explore how these patterns relate to previously described multimodal anatomical areas, and then to large-scale functional networks. Our initial assessment revealed a remarkable congruence between cortical areas and α subunit distribution patterns. The subsequent analysis showed that α subunits are clustered in relation to several functional networks. Worth noting is that even the largest networks had predominant α subunit density combinations. At high level, our results hint at a tight coupling between GABA_AR cytoarchitectonics and the functional anatomical organization of the human cortex.

The correspondence between putative α subunit distribution patterns and architectonic/functional areas was investigated according to the multi-modal parcellation (MMP) of the cortex obtained in the human connectome project (HCP), that is, the HCP-MMP atlas. ⁴ To our knowledge, our present analysis visualizes, for the first time, the potential distinct α subunit distribution in the primary visual, auditory, and sensory/motor areas in vivo. In particular, the most striking aspect of the noted congruence was the dominance of the α_1 + 3 subunit in the primary sensory and motor areas, that is, areas V1, A1, 3a, 3b, 4, with focally sharp agreement between parcel boundaries and marked transitions in α subunit densities toward the neighboring (secondary) areas. Furthermore, alignment between areal borders and transitions in α subunit distribution could be perceived not only in the primary, secondary sensory/motor regions but also in other parts of the cortex (Figures 2 and 4, Supplemental Figures S2 and S3). The remarkable congruence suggests that future refinements of the multimodal atlas could likely benefit from the inclusion of high-resolution cortical receptor maps, obtained using PET, in the derivation of areal boundaries.

As part of our primary analysis, we also performed a visual, and vertex-level numerical assessment of the left-right asymmetry of the α subunit maps by using the symmetric fs_LR template space. Our analysis showed that while the α subunit distribution patterns were symmetric at large scale, left-right asymmetry was anyhow detected at a finer scale. Furthermore, certain fine scale asymmetries proved to be congruent with topological asymmetries present in the HCP-MMP areas, such in case of the left-right difference in the layout of the primary auditory cortex which was closely followed by local peaks in α_1 + 3 subunit levels. Importantly, the observed degree of asymmetries were similar to asymmetry indices recently reported for measures of cortical neuroanatomy or brain function. ¹⁵ As such, the present results corroborate the methodology previously employed for derivation of α subunit maps and underscores that cytoarchitectonics falls in line with other aspects of brain composition and organization regarding laterality.

As a secondary aim we explored the parcellated α subunit maps in relation to the large-scale functional network membership of each cortical area according to the CAB-NP assignment. ⁵ This explorative step was an initial attempt to associate area-specific α subunit density combinations with each functional network. In particular, each of the α subunits tended to have their respective cortical maximums and/or the locus of their highest predominance over the other subunits in areas with a “narrow” functional orientation (e.g. primary sensory/motor, affective). Conversely, the high-level, associative functional networks covered the widest range of α subunit densities with their most typical area specific α subunit combination away from the extremes. This sensory/motor to affective organization of alfa-subunit densities may concur with the general principle of progressive molecular diversity along the sensory-association axis recently outlined using in vitro quantitative receptor autoradiography. ¹⁶

Moreover, the findings show that α subunits peak within specific functional networks. For example, though widely present in several networks, α₅ subunits clearly dominated in the cingulo-opercular and orbito-affective networks, with negligible presence of α_1–3 subunits therein. As noted in our earlier work, ⁹ and in PET studies using the partly selective radioligand [¹¹C]Ro15-4513, ¹⁷ α₅ subunit anatomical distribution may be allocated predominantly to the limbic system. These present observations may be of interest in relation to the “affective network” concept in neuroscience and the novel theory of constructed emotions assuming a central role of the insula-opercular network in the conceptualization of emotions. ¹⁸

Altogether, our preliminary observations encourage augmenting future molecular-informed fMRI studies with subject-level receptor imaging which should likely sharpen our understanding of associations between receptor expression patterns and functional specialization. ⁶ However, even the present population-level α subunit maps may be readily employed in, for example, whole brain computational modelling of cognitive states incorporating receptor information.^6,19 Incidentally, the strong congruence with the MMP atlas also seems to suggest that such modelling may be optimally configured when using nodes corresponding to the MMP areas.

Our findings pertain to one of the basic questions in neuroscience: the existence of a link between brain neurochemical systems and function of neural networks. We would like to suggest that our results support such a link, by elucidating a differentiated pattern of GABA_A receptor composition across cortical areas in humans. The remarkable distinction of α-subunit predominant brain regions may be of interest in association with the known tonic and phasic modes of GABA_A receptor activation as based on differential biophysical and pharmacological properties of GABA_AR subtypes/subunit compositions such as kinetic receptor gating, binding and sensitivity, etc. ²⁰ Thus, integrating the distinct modes of activation of GABA_AR α subunits with detailed cortical α-subunit maps may be of high interest in gaining further understanding of the function of GABA_AR system and potential novel applications for its pharmacological modulation.

Furthermore, our exploratory results, indicating a connection between α subunit distribution and functional networks, seem to align with recent publications, which demonstrated a strong alignment between regional multi-receptor fingerprints (by PET or autoradiography) or cell-type fingerprints (by mRNA expression) and the functional network organization of the cortex.^21,22 It is worth noting that the previous publications operated on multiple receptor systems and/or cell types but had to rely on methodology with relatively low spatial resolution and less certain cortical localization due to the need to harmonize data from multiple sources in case of PET (different studies, camera systems), or due to the use of disjoint/partial brain samples in case of mRNA or autoradiography. In contrast, our work focused on subtypes of a single receptor system (the BDZ-sensitive GABA_ARs), made use of α subunit maps with higher spatial resolution due to data collection on the HRRT PET system, and had sharper cortical localization thanks to the employed cortical surface-based processing workflow. As such it provides an extension and corroboration of prior literature. Importantly, our findings imply that it may be beneficial to base such multi-receptor analyses on the MMP area atlas, similarly to the case of whole brain modelling.

A limitation of our work was the choice of method for inter-subject registration of cortical data. To achieve inter-subject registration of cortical data and characterization of the detected areas the HCP-MMP area map is based on structural/functional features extracted from an extensive set of structural and functional MRI sequences. ⁴ However, in lack of a similar extensive set of MRI sequences in the participants of our data sets, our original work derived the α subunit maps only by employing cortical folding pattern matching for inter-subject registration. As noted by the authors, the HCP-MMP areas and folding patterns were imperfectly correlated outside primary sensory/motor areas, ⁴ thus potentially decreasing the degree by which α subunit density patterns (based only on unimodal matching of folding patterns) and areal borders aligned. However, despite the methodological constraints for subject-to-template registration in our data set, a cortex-wide alignment could anyhow be demonstrated. Even so, future studies employing multi-modal registration may even more precisely characterize alignments outside the primary areas.

Another limitation was the constraint on how the α subunit maps could be derived. Namely, no fully subtype-selective GABA_AR radioligands are available to date for the direct visualization of subtype distribution with only partial (α₅ subunit-) selectivity achieved in case of [¹¹C]Ro15-4513.^17,23 Accordingly, as described in our prior work, the PET images used for derivation of the maps were obtained using the non-selective radioligand [¹¹C]flumazenil, which is non-selective for the α subtypes. ⁹ However, as previously described, we could overcome this limitation to a large degree by employing heterologous competition experiments with two GABA_AR drug candidates with differing subtype selectivity profiles. ¹² This approach enabled the extraction of α subunit maps, albeit only at the population level. Our results underline the need to develop fully selective radioligands, which would allow for a direct, subject-level visualization of the α subunit distributions and thus be ideal in confirmatory studies of our findings as well as in studies aiming to link individual subtype fingerprints to other domains such as disease conditions or pharmacological effects.

On a related note, today’s knowledge about GABA_AR subunit distribution in human brain is at large based on immunohistochemistry studies of postmortem brain tissue. Regional changes in α-subunit distribution have been reported in several disorders, for example, deficit of α₂-subunits in intellectual disability syndrome, ²⁴ in auditory cortex in humans in chronic stroke, ²⁵ reductions in α₁ and increases in α₂ expression in schizophrenia, ²⁶ increase in α₁, α₂, α₅ in hippocampus and decrease in α₅ in superior temporal gyrus in AD patients. ²⁷ However, clinical PET imaging of the GABA_AR system has hitherto not supported postmortem observations.^{28 –30} One of the potential reasons for such discrepancy is, again, the use of the non-subtype-selective radioligand [¹¹C]flumazenil. The current putative detailed cortical maps of GABA_AR α-subunits suggest that an improved characterization of the role of GABA_ARs in neuropsychiatric disorders requires successful development of GABA_AR α-subunit selective PET radioligands. Furthermore, selective radioligands could then serve as new critical tools in the development of selective compounds targeting GABA_ARs, in search of new therapies for major psychiatric disorders.

Conclusions

In summary, the main analysis in our current work revealed a remarkable congruence between GABA_AR α subunit density patterns and multimodal area boundaries in cortex. Furthermore, the extended analysis hinted at a possible link between the α subunit density profiles and the large-scale functional network memberships of cortical areas. In conclusion, our novel findings provide further impetus to the view that a line of research which combines the strengths of molecular imaging (high quality cortical surface receptor maps) and structural/functional MRI holds promise for extended understanding of the human brain.

Data availability statement

All datasets, code and scene files necessary to reproduce result, figures in the main article or in the supplements are available at https://doi.org/10.12751/g-node.6ldy1i. In particular, all processing can be repeated, if desired, on a Linux system which has Apptainer (https://apptainer.org/) or Singularity (https://sylabs.io/) available as well as python with snakemake (https://snakemake.github.io/). Furthermore, the data for the cortical surface figures are in CIFTI file format, which, with the help of the accompanying scene files, can be opened in the Connectome Workbench viewer software (http://humanconnectome.org). Once downloaded and opened in Connectome Workbench, the user can interact with the scene to explore the data, such as rotating, zooming the 3D view or selecting a different overlay, etc.