Putative mapping of α-subunits in the human brain: A PET study of GABA_A receptor binding

The benzodiazepines (BZs) were introduced in the early 1960s and became widely used for their anxiolytic, sedative, anticonvulsant, and muscle relaxant effects. With time it became obvious that the BZs carried concomitant undesirable effects such as dependence and overdose risk. For that reason, numerous attempts have been made to develop safer BZs.

The BZs were originally developed purely from empirical observations (Wick, 2013), whereas the pharmacological target at the γ-aminobutyric acid (GABA) system was not identified until the late 70s (Braestrup & Squires, 1977; Haefely et al., 1975). Now we know that the GABA_A receptor (GABA_AR) is a pentameric ligand-gated ion channel assembled from the known 19 pore-forming subunits (subunit families α, β, γ, δ, ε, θ, and ρ) with a most typical stoichiometry of two α-, two β-, and one γ-subunit (Kim & Hibbs, 2021; Olsen & Sieghart, 2008; Sieghart, 1995). Importantly, BZs serve as allosteric modulators at the α-subunit. This new understanding raised the hypothesis whether GABA_AR-mediated therapy can be optimized by the development of α-subunit selective drugs (Rudolph et al., 1999). However, despite considerable industrial efforts, the development of such drugs has had limited success. In addition, there is lack of radioligands that are selective for distinct α-subunits, thus limiting the tools for direct in vitro/in vivo confirmation of subunit selectivity.

Flumazenil is a nonselective GABA_AR antagonist at the BZ site, with equipotent affinity for the benzodiazepine-sensitive GABA_ARs (i.e., those containing the α₁, α₂, α₃, or the α₅ GABA_AR subunits) and two orders of magnitude lower affinity for those with the α₄- and α₆-subunits (Sieghart, 1995). Flumazenil was early on labeled with carbon-11 for molecular imaging by positron emission tomography (PET) (Persson et al., 1985) and has been broadly applied for imaging of the GABA_AR and recently for detailed mapping of GABA_AR binding in the human brain (Nørgaard et al., 2021). An initial step toward selective visualization of α-subunits was taken by the development of [¹¹C]RO15-4513, a radioligand partially selective for the α₅- subunit (Halldin et al., 1992; Myers et al., 2017). However, due to the mentioned lack of selective radioligands, the distribution of the other α-subunits in the human brain in vivo remains to be fully understood.

A decade ago, AstraZeneca Pharmaceuticals developed two GABA_AR partial agonists at the BZ site intended for the treatment of anxiety (Chen et al., 2014, 2015). AZD7325 has high in vitro affinity for the α₁-, α₂-, and α₃-subunits (K_i values from 0.3 to 1.3 nM) while AZD6280 is more selective and has high affinity to the α₁-subunit (K_i of 0.5 nM) and lower affinity to the α₂- and α₃-subunits (K_i of 21 and 31 nM, respectively). Both drug candidates have low affinity for the α₅-subunit containing GABA_A receptors (K_i of 230 and 1680 nM, respectively). The development program included two PET studies on drug-induced occupancy at the [¹¹C]flumazenil binding sites in the human brain (Jucaite et al., 2017). Importantly, the pattern of occupancy was heterogeneous across brain regions and different between the two drugs (Fig. 1).

The demonstrated differences between the regional occupancy patterns encouraged us to extend the PET image analysis regarding the level of anatomical and explanatory detail. The first key condition for this analysis is that occupancy at a given α-subunit-containing GABA_AR subtype is expected to be the same across brain regions, though different for each drug and dose. This condition is supported by the literature indicating that the majority of benzodiazepine-sensitive GABA_AR subtypes in brain comprise α_1/2/3/5β_1–3γ₂ pentamers (Olsen & Sieghart, 2008; Sieghart & Sperk, 2002). Importantly, [¹¹C]flumazenil has comparable affinities for all these subtypes, which in turn suggests that its PET signal is likely dominated by these GABA_ARs (Sieghart, 1995). In other words, [¹¹C]flumazenil has lower affinities for other GABA_AR subtypes and such binding is thus not likely contributing to the PET signal. Furthermore, it was earlier reported that for benzodiazepines in general, there are no major affinity differences as a function of the constituent β subunit, only as a function of the α-subunit (Sieghart, 1995).

The second condition is that the expression of each α-subunit varies across the brain with detectable contribution to the PET signal in several regions and is not dependent on drug or drug dose and likely similar across subjects. An indirect support for this condition is the previously published data on GABA_AR subunit gene expression that showed that proportional gene expression of each of the 4 α-subunits found in BZ receptors reaches at least ~10% in at least 25% of brain regions relative to the summed expression of these 4 subunits (Sequeira et al., 2019). Taking advantage of these two conditions, the dataset, obtained on a high-resolution PET system, should allow for disentangling the observed [¹¹C]flumazenil binding and occupancy into constituents/components, which are differently bound and occupied by AZD7325 and AZD6280, with considerable anatomical precision across the brain. This analytical approach is reminiscent of prior published works using heterologous competition data: for example, where the binding of the purported mGluR5 radioligand [¹¹C]AZD9272 has been decomposed into MAO-specific and mGluR5-specific components (Varnäs et al., 2020), or where the α₅-subunit-containing GABA_AR selectivity of [¹¹C]Ro15-4513 was quantified (Myers et al., 2017).

The aim of the present image analysis was to capture and quantify the differences in drug-specific regional [¹¹C]flumazenil occupancy patterns in more detail. In total, 12 subjects had been examined and parametric images of [¹¹C]flumazenil binding had been obtained at baseline and after administration of different doses of AZD7325 or AZD6280 (Jucaite et al., 2017, Fig. 1). Parametric images were entered into a comprehensive analysis to obtain maps of the fractional contribution of components of [¹¹C]flumazenil binding across the brain. In addition, recently published GABA_AR subunit gene expression data (Hawrylycz et al., 2012; Sequeira et al., 2019) were used to explore the components and interpret them in terms of α-subunit identity.

The present image analysis was based on two previous PET studies sponsored by AstraZeneca (ClinicalTrials.gov: NCT00681720 and NCT00681746). The experimental procedure and results have been published in detail (Jucaite et al., 2017). Written informed consent was obtained from participants before the initiation of the studies. Both studies were approved by the Regional Ethical Review Board in Stockholm, and the Radiation Safety Committee at Karolinska University Hospital (KUH), Stockholm, Sweden.

Study participants were recruited at the AstraZeneca CPU, Karolinska University Hospital (KUH), Huddinge. Prior to PET measurements, 3D brain MRI examinations were acquired on a 1.5-T General Electric Signa system (GE, Milwaukee, WI, USA) at the MRI Center of KUH, Solna. Two examinations were made in one session. The T2-weighted images were acquired for clinical evaluation, and the T1-weighted images were used as anatomical reference for analyzing the PET images. The T1-weighted sequence was a 3D SPGR protocol in the axial plane with the following parameters: TR 23 ms, TE 4 ms, matrix 256 × 192 × 156, and voxel size 1.02 × 1.02 × 1.0 mm. PET imaging was performed at the PET Centre, Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden. The binding of the radioligand [¹¹C]flumazenil was estimated at baseline conditions and after drug administration. Single oral doses of AZD7325 (0.2, 1, 2, 5, 20, 30 mg) and AZD6280 (5, 12, 20, 30, 40 mg) were administered and PET data were obtained over 63 minutes with a HRRT PET system (Siemens/CTI, Knoxville, TN, USA) (for full details on subjects and doses, see Jucaite et al., 2017). The plasma concentration of tested drugs during PET acquisition was determined by solid-phase extraction using Waters HLB Elution plate followed by LCMS/MS (AstraZeneca, DE, USA).

In the previously reported analysis, parametric images showing the binding potential (BP_ND) were generated using wavelet-aided parametric imaging (WAPI) (Cselényi et al., 2006). This approach included the multi-linear, reference input version of Logan’s graphical analysis using the time–radioactivity curve (TAC) of pons as input. Pons was used as reference region of interest (ROI). The parametric images were used as “input data” in the present analysis (see sample images in Fig. 1). The analysis was implemented in MATLAB® R2022b (The Mathworks, Natick, MA, USA) with some steps carried out in FreeSurfer (version 7.2.0) as indicated below (Fischl, 2012).

MR images and BP_ND parametric images were entered into an extended analysis, which consisted of three major parts (Fig. 2):

I. Preparation. To enable a comprehensive analysis, individual imaging data were projected into a common neuroanatomical template space accounting for interindividual differences in cortical folding patterns and subcortical structures. The main output of this part was overall [¹¹C]flumazenil occupancy in template space.

II. Fitting of occupancy model. The aim of this part was to generate maps showing the distribution of drug-displaceable and non-drug-displaceable components of [¹¹C]flumazenil binding. A series of tailor-made occupancy models, having up to four binding components (C₁–C₄), were fitted to parametric maps of overall occupancy for both drugs at the same time. In particular, the brain region-specific overall measured occupancy was fitted as the weighted sum of the component-specific occupancies where the weighting factors were the fractional contributions of the model-derived components. As pointed out in the introduction, the component- and drug-specific occupancies are not dependent on location in brain, whereas the regional expression levels are location specific. The fitting procedure in part II yielded maps of the components’ fractional and absolute contributions. A preferred model variant was selected based on model performance.

III. Interpretation and validation of results. To summarize and explore the derived maps, gross anatomical brain atlases were applied. Furthermore, to explore the biological validity and underpinnings of the component maps, the patterns of contributions were compared with gene expression data of the relevant GABA_AR α-subunits.

The detailed steps of the extended analysis are illustrated in Figure 2 and described below where the label and numbering of each step match between description and figure:

• Preparation

  • 1.

    FreeSurfer:

    The individual reoriented MR images were processed in the FreeSurfer (version 7.2.0) analysis pipeline to obtain cortical surface extraction and anatomical parcellation as well as subcortical segmentation (Fischl, 2012).

  • 2.

    Combined Volumetric and Surface (CVS) registration:

    The results of cortical parcellation and subcortical segmentation were for each subject entered for atlas registration using the CVS registration algorithm in FreeSurfer (Postelnicu et al., 2009). The method accounts for interindividual overall and local volumetric differences in brain shape, nuclei, and other subcortical brain structures as well as cortical folding patterns. The CVS template in the Montreal Neurological Institute (MNI) atlas space (cvs_avg35_inMNI152) was used as target of the registration. Thereby, the CVS registration provided the parameters necessary for warping arbitrary volumetric and surface data from the individual anatomical to the MNI152 atlas space. Note that the FreeSurfer processing in step 1 excluded the hippocampus from the cortical surface. However, the cortical surface in the cvs_avg35_inMNI152 CVS template (originally derived in FreeSurfer version 6.0) cuts into the hippocampus and amygdala. Since this discrepancy would lead to erroneous hippocampal CVS registration, we performed cortical surface extraction, parcellation, and segmentation of the cvs_avg35_inMNI152 template image in FreeSurfer vesion 7.2.0 before step 2 (this version of the template can be found online at https://doi.org/10.12751/g-node.83wfme).

  • 3.

    Extraction:

    FreeSurfer provided the inner (toward white matter) and outer (toward pia mater) surfaces of the cortical gray matter, which were used to extract flumazenil binding potential (BP_ND) data for cortex. In detail, each surface vertex in the “white” surface mesh has a corresponding vertex in the pial surface with a line connecting the two that is exactly perpendicular to the tangential plane of the pial cortical surface at that point. Accordingly, for each pair of vertices, the parametric image was sampled along the connecting line at a fractional distance of 0.4 to 0.6 (in 0.02 increments) between the points with the mean value of these 11 samples taken as the corresponding surface BP_ND. The conservative sampling around the middle point across the thickness of the cortical ribbon was applied to minimize the impact of partial volume effects in the PET parametric images.

  • 4.

    CVS application:

    Using the CVS registration parameters, the individual volumetric and cortical surface BP_ND data were warped to the cvs_avg35_inMNI152 template space. A binary image mask of subcortical voxels in MNI152 space was used to retain subcortical voxels in the volumetric portion of the binding data. As noted above (step 2), hippocampus was assigned to the volumetric, “subcortical” data in FreeSurfer and not to the cortical ribbon. Both volumetric and surface data were then smoothed with a Gaussian filter (2D or 3D, respectively) using a kernel with a full width at half maximum (FWHM) of 5 mm. Several of the steps below operated on such data at a volumetric level for subcortical voxels or at a surface level for cortical vertices. The expression “elementwise” refers to this collection of volume and surface data points.

  • 5.

    Inter-subject statistics:

    For each baseline BP_ND element, the lowest value (minimum) and the mean (N = 12) were calculated.

  • 6.

    Thresholding & masking:

    A binary mask of “valid” data elements was obtained by thresholding the inter-individual minimum elementwise BP_ND data at 0.1 (inclusive). Furthermore, for the volumetric part excluding the cortex, only voxels within the following anatomical regions were retained: amygdala, hippocampus, striatum, pallidum, thalamus, ventral diencephalon, brain stem, and cerebellar cortex (the non-cortical volumetric mask can be found online at https://doi.org/10.12751/g-node.83wfme).

  • 7.

    Calculation of occupancy:

    Elementwise overall occupancy, that is, reduction in [¹¹C]flumazenil binding, was calculated for each pretreatment PET measurement according to the following equation:

    $Occi,jtot=1−BPNDPT_i,jBPNDBL_i,j|1≤i≤MVS;j∈{AZD7325,AZD6280},$

    (1)

    where i is indexing elements with M_VS equal to the number of valid volume and surface elements (507690), j refers to the drug (AZD7325 or AZD6280), $BPNDBL_i,j$ and $BPNDPT_i,j$ are the baseline and pretreatment binding potentials for drug j at element i. Note that indexing by occasion (drug dose) is omitted for clarity.

  • 8.

    Cluster definition:

    To reduce the computational burden of the subsequent non-linear modeling step, and to reduce the variance of the elementwise overall occupancy data, due to inter-individual variability, test–retest variation and statistical noise, the elements were assigned to geometrically defined clusters or neighborhoods. In detail, voxels in 4 × 4 × 4 blocks were assigned to the same cluster. This step corresponds to a 64-fold data reduction, that is, with a Euclidean distance less than 4 mm from the cluster centroid for each voxel. Analogously, neighboring vertices were assigned to clusters leading to a ~120-fold surface data reduction. Here, the surface clusters were obtained by a geometry-preserving reduction of the vertices of the inflated, spherical representation of the cortical data such that ~120 vertices within small hexagonal patches across the surface were assigned to the same cluster with an average Euclidean distance of ~4 mm of the vertices to the cluster centroid. Following this reduction, the total number of clusters was 6469, that is, far less than the original 507690 elements.

  • 9.

    Pooling of occupancy values:

    The mean elementwise overall occupancy data were calculated across cluster members for each cluster and drug pretreatment occasion (i.e., yielding cluster-wise $Occi,jtot$ with i now indexing clusters).

• Fitting of occupancy model

  • 10.

    Fit of two-drug, multi-component occupancy models:

    Initially, several multi-component model configurations were fitted and compared as described in Section 2.5.1 below. See also Supplementary Figure S1 for a schematic view of the tested model configurations. Based on model comparisons, a preferred model configuration was finally selected.

    Each model was fitted to the cluster-wise $Occi,jtot$ and drug plasma exposure (C_p) data and was based on the following set of equations:

    $Occj,k=Cp_jCp_j+Ki_j,k|j∈{AZD7325,AZD6280};1≤k≤Nj;Nj≤N−1,$

    (2)
    $Oc^ci,jtot=∑k=1Njfi,k×Occj,k|1≤i≤MCl;∑k=1Nfi,k=1,$

    (3)

    where i is indexing clusters with M_Cl equal to the number of clusters (6469), j refers to the drug (AZD7325 or AZD6280), N_j is the number of displaceable components of [¹¹C]flumazenil’s binding for drug j, N is the total number of components of [¹¹C]flumazenil binding including a non-drug-displaceable component, $Occj,k$ is the occupancy of drug j at displaceable component k of flumazenil binding, $Cp_j$ is the plasma exposure of drug j at each pretreatment occasion (indexing by occasion is omitted for clarity), $Ki_j,k$ is the dissociation constant of drug j at binding component k, $Oc^ci,jtot$ is the model-predicted overall occupancy (reduction in radioligand binding) for cluster i and drug j, and $fi,k$ is the fractional contribution of component k to [¹¹C]flumazenil binding in cluster i. As already pointed out, the component-specific occupancies are not location dependent, while the component-specific fractional contributions to [¹¹C]flumazenil binding are not drug dependent. Using this model, the best-fit values of $Ki_j,k$ and $fi,k$ were estimated using non-linear least squares optimization, that is, minimizing the sum of squared residuals: $∑i=1MCl∑j(Occi,jtot−Oc^ci,jtot)2$⁠.

    Instead of directly fitting the component-specific fractional contributions (⁠$fi,k$⁠) according to Equation 3, a change of variables $fi,k→vi,k$ was applied using the following equations to avoid generation of nonsensical values:

    $vi,k=∑q=1kfi,q∑q=1k+1fi,q|1≤k<N;0≤vi,k≤1,$

    (4)

    and its inverse:

    $fi,k=∏q=kNvi,q−∏q=k−1Nvi,q|1≤k≤N;v0≡0;vN≡1.$

    (5)

    Accordingly, with the fitted $vi,k$ parameters (⁠$1≤k<N$⁠) restricted to fall between 0 and 1 (inclusive), the corresponding $fi,k$ parameters were ensured to be non-negative and sum to unity (⁠$1≤k≤N)$ as required by the model Equation 3.

    The starting values for the component-specific affinity values (or rather $Ki_j,k$ parameters) were such that the higher affinity values (lower $Ki_j,k$⁠) were given to the lower order components. The $vi,k$ parameters were initialized by obtaining them according to Equation 4 from initial $fi,k$ set to 1/N.

  • 11.

    Interpolation:

    The cluster-wise fractional contributions from the components were projected to the level of elements in two ways. In the first way, supporting the comparison with regional gene expression (step 13b), the elementwise fractional contributions were obtained by simply assigning the corresponding cluster’s result. In the second way, appropriate for visualization and derivation of absolute contribution maps (step 12), the cluster-wise component fractional contributions were assigned to the level of elements using linear interpolation. Trilinear interpolation was used in the volumetric (subcortical) part, and bilinear interpolation was used in the surface (cortical) part obtained from FreeSurfer’s spherical representation of the cortical data.

  • 12.

    Elementwise multiplication:

    The inter-subject mean elementwise BP_ND data set (obtained in step 5) was multiplied with the linearly interpolated elementwise component-specific fractional contribution parameters to obtain the elementwise component-specific absolute contributions expressed as BP_NDs.

• Interpretation and validation of results

  • 13.

    Application of atlases and comparison with gene expression. The region (structure) definitions of two anatomical atlases were used for visualization and to calculate regional average values in the contribution maps. In detail:

    • For gross anatomical designation of the observed component-specific cortical and subcortical patterns, the Desikan–Killiany atlas’s surface parcellation and the FreeSurfer’s subcortical segmentation were used, respectively.

    • To obtain a more detailed anatomical designation and for comparison with GABA_AR gene expression data, the 3D Allen Human Reference Atlas (AHRA) was applied (Ding et al., 2020, http://download.alleninstitute.org/informatics-archive/allen_human_reference_atlas_3d_2020/version_1/). The brain regions of this atlas are available as a volumetric label of the MNI152NLin2009bSym template (ICBM 2009b Nonlinear Symmetric) brain image (Fonov et al., 2009) and comprise 141 substructures including gray and white matter substructures (129 left/right mirrored and 12 midline structures). While the subcortical structures are in close register between the MNI152NLin2009bSym and cvs_avg32_inMNI152 templates, the cortical folding patterns are slightly different. Accordingly, the AHRA volumetric labels could be used directly for subcortical structures, but the labeling of cortical structures in the cvs_avg32_inMNI152 template required additional steps. First, the MNI152NLin2009bSym T1-weighted MR image was processed in FreeSurfer (version 7.2.0) to obtain the cortical surface for this template. The cortical surface labels for the Allen Human Reference Atlas (AHRA) were then extracted by sampling the volumetric labels along the cortical ribbon of the MNI152NLin2009bSym template. Following spherical registration of the cortical surfaces of the two templates in FreeSurfer, the labels were finally transferred to the cortical surface of the cvs_avg32_inMNI152 template. The complete surface and volumetric label set contained 270 distinct regions of interest (ROIs) (2 × 129 side-specific and 12 midline ROIs), of which 221 labeled ROIs represented gray matter (2 × 108 side-specific and 5 midline ROIs, i.e., 113 ROIs with left-right pooling). The surface and volumetric label files and the list of ROI names can be found online at https://doi.org/10.12751/g-node.83wfme.

Regional microarray expression data were obtained from an analysis of 6 post-mortem brains (1 female, 5 male, ages 24.0–57.0, 42.50 ± 13.38 years) provided by the Allen Human Brain Atlas (AHBA, https://human.brain-map.org) (Hawrylycz et al., 2012). Microarray data were processed using an augmented variant of the abagen software toolbox (based on version 0.1.4 + 15.gdc4a007; https://github.com/rmarkello/abagen) (Arnatkevic̆iūtė et al., 2019; Markello et al., 2021). To extract regional gene expression data, a curated volumetric version of the AHRA atlas in MNI152NLin2009bSym space was obtained, with labels for 98 distinct substructures (covering all 221 gray matter ROIs considering left–right pooling and pooling 2 × 23 ROIs belonging to 8 substructures that allowed for better sample coverage in the gene expression data, see Supplementary File 2). We employed the microarray probe selection for GABA_AR genes using a previously published approach (Sequeira et al., 2019). The tissue samples metadata in the microarray data include the anatomical substructural designations provided by the AHBA ontology (see the coronal AHBA sections at https://atlas.brain-map.org/atlas?atlas=265297125), as well as MNI 3D coordinates. The MNI coordinates of tissue samples were updated to those generated via non-linear registration using the Advanced Normalization Tools (ANTs; https://github.com/chrisfilo/alleninf). To increase spatial coverage, tissue samples were mirrored bilaterally across the left and right hemispheres (Romero-Garcia et al., 2018). The original abagen software performs sample assignments to brain regions based on sample MNI coordinates by identifying the atlas parcel containing (or nearest to) the given sample. To reduce the potential for misassignment, the matching in the original version is constrained at the gross anatomical structural level (such as cortex and subcortex).

However, thanks to the 3D AHRA, which was made available following the generation of the original abagen software, we were able to implement a more fine-grained approach for constraining sample-to-region matching, further minimizing the possibility of sample misassignment. Importantly, there is a close correspondence of the AHRA and AHBA ontologies at the substructural level, in terms of both structure terminology and boundaries between corresponding substructures (confirmed by inspecting the 3D AHRA volume and the coronal sections for AHBA). Accordingly, for each of the 98 curated 3D AHRA substructures matching AHBA substructures could be identified. In most cases, an equivalent AHBA substructure or set of AHBA substructures were found. In a few cases, two or three AHRA substructures together were covering an equivalent AHBA substructure. The master table showing the mapping between curated AHRA substructures (with their constituent AHRA ROIs) and the corresponding AHBA substructures is found in Supplementary File 2. Based on this mapping, the augmented abagen sample assignment process restricted the sample MNI coordinate mapping to those atlas parcels which had a corresponding AHBA substructure containing the given sample according to their original anatomical designation. In other words, samples that had a one-to-one mapping between the AHRA and AHBA nomenclatures were directly assigned to the corresponding AHRA substructure. Samples with multiple AHRA correspondences, however, were assigned by abagen to the correct AHRA substructure that was closest based on their MNI coordinates. All tissue samples not assigned to an AHRA substructure were discarded. Samples assigned to the same AHRA substructure were averaged separately for each donor and then across donors yielding a single average value for each gene per substructure. Only regions containing gene expression data for all 6 donors and valid regional component contribution parameters were included in the comparison (62 substructures).

The regional patterns of component contributions were compared visually with the processed GABA_AR α-subunit gene expression data and by calculating Pearson product-moment correlation coefficients (R values). The degree (size) of positive association was categorized according to the following cutoffs for the R value: R ≤ 0.0—lack of positive correlation; 0.0 < R ≤ 0.3—negligible positive correlation; 0.3 < R ≤ 0.5—low positive correlation; 0.5 < R ≤ 0.7—moderate positive correlation; 0.7 < R ≤ 0.9—high positive correlation; 0.9 < R ≤ 1.0—very high positive correlation. Note that only positive correlations are of interest in exploring the α-subunit identity of the components. The p-values for the correlations were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure (Benjamini & Hochberg, 1995). In detail, the obtained regional gene expression data for GABRA1, GABRA2, GABRA3, and GABRA5 genes, that is, the genes corresponding to the four GABA_AR subtypes recognized by flumazenil, were first converted to fractional units by dividing the gene-specific expression with the summed expression of the four genes (non-log2-transformed). In this way, the gene expression of interest was quantified in a manner commensurate with the primary outcome parameters of the present PET analysis, the component-wise estimated regional fractional contributions. Note that while the selection of the preferred model was prior to, that is, did not rely on the comparison with gene expression data, yet the observed correlations for the selected model were subjected to sensitivity analysis by calculating the correlations for the cross-validation data sets first used in the model selection (described in Section 2.5.1).

As outlined in step 10 above, multiple occupancy model configurations were fitted to the data (see schematic view of model configurations in the top panel of Supplementary Fig. S1 with detailed description in the figure legend). The initial choice of model configurations was guided by the binding characteristics in vitro of the radioligand [¹¹C]flumazenil as well as the two drug candidates AZD7325 and AZD6280 (values shown in the introduction). In detail, the main guiding principles were that

• [¹¹C]flumazenil is expected to bind more subtypes than either drug, resulting in an overall displaceable and non-displaceable component of its binding.

• In the drug-displaceable portion, AZD7325 may be able to occupy more components than AZD6280.

• AZD6280 should have a maximum of one high affinity component and potentially one lower affinity one.

• Maximum three drug-displaceable components are possible with AZD7325 having high affinity for at least two and AZD6280 having high affinity for one.

Accordingly, a total of two to four model components were considered including one that was not displaced by either drug. The two-component model (M2a) assumed that only the lumped occupancy can be quantified, representing the high-affinity portions of drug binding. In all model configurations with three or four components, both drugs were assumed to have at least one high-affinity component with detectable occupancy. The four-component models assumed that specific binding at all three drug-recognizable α-subunits (α₁, α₂, α₃), and thus their respective occupancy components could be disentangled.

The models were compared with the aim of identifying a preferred configuration. However, while in typical PET quantification the preference is established through purely statistical means, the current work relied on multiple means of evaluating model performance. The present evaluation still included the statistical assessment of the goodness of the fit. However, it also included parameter sanity and model generalizability checks.

The goodness of fit statistical parameters included the final sum of squared residuals, mean squared error, Akaike’s Information Criterion, and pair-wise F-tests between all possible pairs of less and more complex model configurations.

The parameter sanity and model generalizability checks included (1) calculating the percent of zero fitted fractional contribution parameters, (2) cross-validation (sensitivity analysis), and (3) checking parameter asymmetry. The fraction of zero fitted fraction contribution parameters was calculated since a model with relatively more zeros may indicate an overfit to the data.

For cross-validation (CV), each model was repeatedly fitted to a subset of the data, in each CV subset excluding one subject’s data from the AZD7325 and one from the AZD6280 experiments, respectively. All possible subject pairings were considered for exclusion in turn, that is, in total 32 CV models were fitted. Then, for each CV model-, cluster-, and component-specific fractional contribution value, the difference from the full model’s respective value was calculated. The median absolute difference was used as an index of the stability (variability) of the given model configuration.

Asymmetry was assessed in two steps. First, asymmetry was calculated at the level of brain regions (according to the Desikan–Killiany cortical parcels and FreeSurfer subcortical segments) by taking the absolute difference between the left and right side regional fractional contribution values in percentage of the mean of the two values. Second, the overall asymmetry for the brain was obtained by taking the median asymmetry value across regions. Note that prior studies have shown that the overall [¹¹C]flumazenil binding is symmetric (Nørgaard et al., 2021), which was confirmed in the present data set (median absolute asymmetry was 4%). Thus, it appeared reasonable that subtype-specific components of [¹¹C]flumazenil binding should also be symmetric.

Initially, the PET parametric and MR images were processed to obtain elementwise occupancy maps in the cvs_avg35_inMNI152 template space. Several multi-component model configurations were then fitted to the prepared occupancy data (see Supplementary Fig. S1, supplementary figures and tables are found in Supplementary File 3). The model variant M3c was the overall best considering statistical performance, parameter sanity, and model generalizability checks. In detail, M3c was among the top 3 variants across all metrics of model performance. According to F-tests, model M3c was superior to the three simpler model variants (M2a, M3a, and M3e) and to most of the more complex model variants (Supplementary Table S1). The exceptions were model M3d and M4d of which M3d was superior in all F-tests. However, while M3d was preferred according to purely statistical metrics, it was behind M3c in sanity and model generalizability checks as it had a higher proportion of zero parameters, higher cross-validation error, and on average higher parameter asymmetry. Similar considerations applied to model M4d as well. Thus, in contrast to model M3c, models M3d and M4d appeared to end up overfitting the data with lower generalizability despite their more favorable statistical performance. Taken together, the M3c variant was selected as the preferred model.

The M3c model has three binding components (Fig. 3A). AZD7325 was able to occupy binding components 1 and 2, whereas AZD6280 occupied only component 1. For AZD7325, there was a 12-fold difference between the 2 component-specific affinity parameters (K_i) with the higher affinity corresponding to component 1 (Fig. 3B). The difference was statistically significant, at the α = 5% level, as indicated by the narrow, non-overlapping 95% confidence intervals of the 2 K_i estimates. Cortical surface and non-cortical volumetric maps of component-specific contributions for the selected model are available online at https://doi.org/10.12751/g-node.83wfme.

[¹¹C]flumazenil binding (BP_ND) across cortex, which is the foundational data for the presented analysis, was visualized on a left cortical flat map with demarcations of the Desikan–Killiany atlas cortical parcels (left panels in Fig. 4A, B; see also Supplementary Figs. S2 with demarcations according to the AHRA cortical parcels). Cortical binding is also shown in more detail for both hemispheres on 3D inflated cortical surfaces in the left panels of Supplementary Figures S3 and S4. Subcortical binding is shown in 3D volume rendering in the right panel of Figure 4A (subcortical segmentation shown in the right panel of Fig. 4B) and on a montage of 2D cross sections with segment borders according to FreeSurfer’s subcortical segmentation in the right panels of Supplementary Figures S3 and S4, respectively. A tabulation of regional mean [¹¹C]flumazenil binding values is available in Supplementary File 4.

As seen on the surface maps, the cortical binding potential (BP_ND) was typically above 4.0 with highest values for the occipital cortex (Fig. 4A, Supplementary Figs. S2 and S3). Subcortical values were generally lower with voxel-wise values reaching 4.0 or higher only in hippocampus and cerebellar cortex (Fig. 4A, Supplementary Fig. S3).

Furthermore, [¹¹C]flumazenil BP_ND values were plotted for the 62 anatomical regions defined by the anatomically more detailed AHRA that were used for correlation with gene expression (Fig. 5, for a plot for all 95 AHRA substructures with binding data, see Supplementary Fig. S5). The highest region-specific average cortical value was 6.0 BP_ND and the cortical binding was below 4.0 only in the hippocampal formation (HiF) and limen insula (LI, see Supplemental Fig. S5). Regional subcortical [¹¹C]flumazenil BP_ND was at or below 2.5 (Fig. 5).

The component-specific fractional (%) contributions to binding across the brain were visualized in a similar manner to overall [¹¹C]flumazenil binding (Figs. 6 and 7; Supplementary Figs. S6 to S10). Taken as a whole, the component-specific patterns were distinct and did not line up with each other or with the pattern for overall [¹¹C]flumazenil binding. Regional mean component-specific fractional contributions are available in Supplementary File 4.

According to the cortical maps, the fractional contribution of component 1, which could be occupied by both drug candidates, was typically above 60% across the cortex with values up to ~80% in the primary somatosensory/motor cortex and in the pericalcarine cortex (upper panel in Fig. 6A, Supplementary Fig. S6A, left panel in Supplementary Fig. S7). On the contrary, the contribution of component 1 was low in the limbic, cingulate, and insular cortices. In most subcortical cerebral nuclei, the regional distribution of component 1 was low (<20%) to moderate (20–50%), whereas it was high (~80%) in globus pallidus (lower panel in Fig. 6A, right panel in Supplementary Fig. S7, regional plot in Fig. 7, Supplementary Fig. S10). The contribution in thalamus was similar to cortical levels (~60%), but reached above 80% in sub- and hypothalamus, mesencephalon, cerebellar cortex, and myelencephalon.

As visualized in the cortical surface maps, the fractional contribution of component 2, which was only displaceable by AZD7325, and not by AZD6280, was overall lower than component 1, and displayed a gradient of increasingly higher levels across cortical regions in an anterior-to-posterior and superior-to-inferior direction (upper panel in Fig. 6B, Supplementary Fig. S6B, left panel in Supplementary Fig. S8). The highest contributions, with up to ~40%, were observed in the temporal, insular, and parietal cortices. The contribution of component 2 was slightly higher in most cerebral nuclei than in cortex, where its contribution was at or below 40% at the voxel level (Fig. 6B, Supplementary Fig. S8), below 20% regionally (Fig. 7, Supplemental Fig. S10). In the diencephalon, mesencephalon, and cerebellum, the contribution was also typically low (below 20% voxel-wise, below 10% regionally) with higher level only in the dorsal lateral geniculate nucleus of thalamus (lower panel in Figs. 6B and 7, right panel in Supplementary Figs. S8 and S10).

Component 3, which was not displaceable by either of the two drugs, had high fractional contribution in sharply demarcated areas in the limbic, cingulate, and insular cortices (upper panel in Fig. 6C, Supplementary Fig. S6C, left panel in Supplementary Fig. S9). Specifically, the contribution was up to ~65% in these areas, whereas it was below 40% in the other parts of the cortex. The contribution of component 3 in subcortical nuclei was generally similar to the highest cortical levels with the regional average reaching almost 70% (voxel-wise approaching 100%) in amygdalohippocampal area and nucleus accumbens. However, values in globus pallidus and caudate body were low (below 50% at the voxel level, below 25% regionally) (lower panel in Fig. 6C, Fig. 7, right panel in Supplementary Figs. S9 and S10). The contribution in the thalamus was about 35%, whereas it was negligible (below 15% voxel-wise, below 10% regionally) in subthalamus, mesencephalon, and below.

The fractional contributions were multiplied by the intersubject mean [¹¹C]flumazenil binding to obtain the absolute component-specific contributions expressed as BP_ND. The cortical surface view, subcortical volume and section views, and regional summary of the absolute contributions are shown in Figs. 8 and 9, Supplementary Figs. S11 to S15). Regional mean absolute contributions are available in Supplementary File 4.

As seen on the cortical maps, the absolute cortical contribution of component 1 was highest in the pericalcarine cortex, followed by the transverse temporal cortex (upper panel in Fig. 8A, Supplementary Fig. S11A, left panel in Supplementary Fig. S12). The absolute contribution of component 1 was lowest in the limbic, cingulate, and insular cortices. The absolute contribution of component 1 in most subcortical regions was below typical cortical values (BP_ND values of 0.0–2.0 vs. ~3.0, respectively), despite the previously indicated high fractional contribution in several subcortical structures (Fig. 9, Supplementary Fig. S15). However, higher BP_ND values (up to ~3.8 on the voxel level) were observed in parts of the cerebellar cortex (CBL) and the dentate nucleus (CbDN) (lower panel in Figs. 8A and 9, right panel in Supplementary Figs. S12 and S15).

The cortical values for the absolute contribution of component 2 covered a range from near 0 to ~1.8 on the vertex level with gradually higher values in an anterior-posterior and superior-inferior direction (lower panel in Fig. 8B, Supplementary Fig. S11B, left panel in Supplementary Fig. S13). As seen on the plot of regional values, cortical surface and subcortical views, the absolute (BP_ND) contribution for component 2 was generally different between telencephalon (cortex with values up to ~1.8 in distinct patches; cerebral nuclei) and the remaining brain regions (thalamus and below), with voxel-wise values below 0.5, regional values below 0.2 (Figs. 8B and 9, Supplementary Figs. S11, S13, and S15).

The absolute contribution of component 3 in cortex ranged between 0.0 and ~3.5 in such a way that values above ~2.0 were confined to sharply demarked cortical areas in the limbic, cingulate, and insular cortices (upper panel in Fig. 8C, Supplementary Fig. S11C, left panel in Supplementary Fig. S14). The plot of regional values and subcortical views showed that the absolute contribution of component 3 was detectable in prosencephalon (thalamus and above) and negligible in the rest of the brain (lower panel in Figs. 8C and 9, right panel in Supplementary Figs. S14 and S15).

For biological interpretation and validation, the fractional component contributions for the AHRA substructures were correlated with published gene expression data (Hawrylycz et al., 2012) expressed in fractional units (see Supplementary File 5 with merged gene expression, component contribution, and overall [¹¹C]flumazenil binding data). Correlation coefficients are shown in Figure 3C. For visual comparison, the gene expression and component contribution data were plotted for selected gene and component pairs (Fig. 10). To aid visual evaluation, the regional values were normalized to the average value across brain regions, whereas un-normalized values are shown in Supplementary Figure S16.

The contribution of component 1, which was displaced by both drugs, agreed best with the expression of the α₁-subunit gene (statistically significant positive correlation, R = 0.80 [cross-validation set R = 0.80 ± 0.011, 0.78–0.83], FDR-corrected p-value = 1 × 10^-14) as compared with the gene expression of the other flumazenil recognized α-subunits (Fig. 3C). Visual inspection of the corresponding plots also suggested agreement (topmost chart in Fig. 10 and Supplementary Fig. S16). Based on further considerations, the comparisons between component 1 and GABRA3 (R = 0.303 [cross-validation set R = 0.304 ± 0.011, 0.28–0.32], FDR-corrected p-value = 0.017) and component 1 and the sum of GABRA1 and GABRA3 (R = 0.87 [cross-validation set R = 0.87 ± 0.009, 0.85–0.89], FDR-corrected p-value = 5 × 10^-19) were also plotted (Fig. 10, Supplementary Fig. S16).

The regional pattern of the contribution of component 2 correlated most strongly with α₂- subunit gene expression (R = 0.71 [cross-validation set R = 0.66 ± 0.082, 0.43–0.77], FDR-corrected p-value = 2 × 10^-10) and the contribution of component 3 correlated with α₅-subunit gene expression (R = 0.81 [cross-validation set R = 0.80 ± 0.011, 0.77–0.81], FDR-corrected p-value = 1 × 10^-14). In both cases, there was visual similarity of the compared patterns (Fig. 10, Supplementary Fig. S16).

In the present image analysis, we took advantage of two previous PET studies in which the occupancy of [¹¹C]flumazenil binding to GABA_AR had been estimated in the human brain after administration of two drug candidates with different binding characteristics at α-subunits (Jucaite et al., 2017). An important observation was that the drug candidates induced different regional patterns of drug occupancy in the brain. Taking advantage of the different occupancy patterns, the present extended image analysis revealed three components of [¹¹C]flumazenil binding. The α-subunits (α₁, α₂, α₃, α₅) corresponding to each component were putatively identified using GABA_AR subunit gene expression data (Hawrylycz et al., 2012). The analysis confirms that the pharmaceutical industry has been able to generate GABA_AR modulators that are partly selective at the α-subunit and encourages the development of even more selective compounds with potential for improved drug treatments of neurological and psychiatric disorders.

The comparison of model fitting results with gene expression data focused on correlating the component-specific contributions with gene expression of the α-subunit genes GABRA1, GABRA2, GABRA3, and GABRA5, respectively.

The contribution of component 1 was positively correlated with α₁- and α₃-subunit gene expression with a high-degree positive correlation with the summed gene expression of these two subunits (Fig. 10, Supplementary Fig. S16). Given the lack of a positive, significant correlation between the gene expression of these subunits (R = -0.12, FDR-corrected p-value = 0.59), a likely interpretation is that component 1 incorporates both α₁- and α₃-subunit distribution. At the same time, among the flumazenil detected α-subunits, the gene expression of the α₃-subunit is generally low (Supplementary Fig. S16). Furthermore, the low α₃ expression levels may explain why disentangling a purely α₃-related component of [¹¹C]flumazenil binding was not successful.

The contribution of component 2 was positively correlated with α₂ and, to a lower degree, with α₅-subunit gene expression (Figs. 3C and 10, Supplementary Fig. S16). Importantly, the gene expression of these two subunits has previously been shown to correlate (Sequeira et al., 2019), which is consistent with our analysis (R = 0.43, FDR-corrected p-value = 6 × 10^-4). Taken together with the fact that previous in vitro binding assays of the two drugs indicated a lack of substantial affinity to α₅-subunit-containing GABA_ARs, the most plausible interpretation is that component 2 preferentially represents α₂-subunit expression.

Conversely, component 3’s contribution was positively correlated with α₅ and, to a lower degree, with α₂-subunit gene expression (Figs. 3C and 10, Supplementary Fig. S16). Considering that component 3 represents flumazenil binding not displaced by either drug, component 3 appears to be more directly related to α₅-subunit expression. As a further confirmation, the pattern here described for component 3 shows a strong similarity to prior published results obtained using the α₅-selective radioligand [¹¹C]Ro15-4513 (Myers et al., 2017).

The interpretations discussed above were based on correlations of fractional expression values. It is worth noting that correlations performed on absolute component contributions (BP_ND) versus absolute gene expression data overall concurred with the findings based on fractional values, albeit with numerically lower degree, statistically less significant associations (data in Supplementary File 5).

Previous publications comparing [¹¹C]flumazenil binding profiles in brain with gene expression of various GABA_AR α- (and other) subunits tended to find only weak-to-moderate associations, the latter in case of GABRA1 (Hansen et al., 2022; Nørgaard et al., 2021). This is expected since [¹¹C]flumazenil comprises the summed signal of several subtypes of which the α₁-dependent signal is the dominant one. The relatively smaller contributions of the other subunits are likely to influence the sum only to a lesser degree. The correlations between [¹¹C]flumazenil binding and the expression of relevant subunit genes (i.e., GABRA1, GABRA2, GABRA3, GABRA5) in our data concur with previous findings (Supplementary Table S2, data in Supplementary File 5). However, our analysis went further. By dissecting the [¹¹C]flumazenil signal, we were able to reveal strong correlations between its components and α-subunit gene expression.

Furthermore, in addition to the component-specific comparisons, the overall [¹¹C]flumazenil binding was also compared with the summed absolute expression of the relevant α-subunit genes (i.e., GABRA1 + GABRA2 + GABRA3 + GABRA5, Supplementary Fig. S17, data in Supplementary File 5). While visual inspection suggests some similarity and the correlation was of high positive degree (R = 0.72, p-value = 2.5 × 10^-10), the aforementioned level difference between cortical and non-cortical regions in [¹¹C]flumazenil binding (Figs. 4 and 5) was less pronounced in the case of gene expression (Supplementary Fig. S17). Of interest is that a more pronounced cortical-to-non-cortical difference in the case of radioligand binding versus gene expression has been demonstrated previously (see Fig. 3A in Nørgaard et al., 2021). It cannot be excluded that differences in the processing of cortical and non-cortical structures may contribute to this discrepancy. Yet the striking regional differences in [¹¹C]flumazenil binding are conspicuous even in unprocessed PET radioactivity images or parametric images (see Fig. 1), as well as shown by in vitro autoradiography (Braestrup et al., 1977; Nørgaard et al., 2021; Persson et al., 1991). Future work is needed to elucidate the nature of this discrepancy between gene and protein expression.

The present assessment relied on the application of combined volume and surface (CVS) registration of individual imaging data to a common template space (MNI152). The traditional approach is to apply the CVS-registration parameters to acquired dynamic PET measurements. The novel approach used in the present analysis was to apply the parameters directly to parametric images of [¹¹C]flumazenil binding. The parametric images were obtained using a previously described wavelet-aided approach which allows handling of noise in an optimal manner and yields images suitable for assessing binding at high resolution (Cselényi et al., 2006; Turkheimer et al., 2003). Accordingly, the resulting component maps provided detailed information about putative GABA_AR α-subunit distribution throughout the brain.

A fundamental assumption of the analysis was that [¹¹C]flumazenil binding and the expression patterns of the various GABA_AR subunits are stable across subjects. The baseline [¹¹C]flumazenil binding was similar across subjects: the coefficient of variation for the 12 subjects was 23% on average for each volume and surface element. Moreover, the average correlation coefficient in elementwise baseline binding across every pair of subjects (66 pairs for N = 12), that is, the so-called differential stability, was 0.91 (Hansen et al., 2022). Furthermore, strong positive correlation has previously been reported between post mortem receptor densities with autoradiography, and regional [¹¹C]flumazenil binding in data obtained in different subjects (Nørgaard et al., 2021). In the same report, the corresponding regression was then used to estimate benzodiazepine receptor (BZR) density (B_max) values within Desikan–Killiany cortical parcels as well as FreeSurfer subcortical segments (Nørgaard et al., 2021). We compared the inter-subject mean regional [¹¹C]flumazenil BP_NDs in the present study with these published B_max values. The observed correlation had a very high positive degree and was highly significant (R = 0.95, p-value = 3 × 10^-24, see also the regression plot in Supplementary Fig. S18, tabulated data in Supplementary File 6). Finally, strong, significant correlations between component contribution and gene expression were observed in the validation part of the present study, as indicated above. Taken together, the strong agreements across multiple domains and studies lend support to the remarkable inter-subject stability of the GABA_AR system regarding subunit gene and protein expression.

Worth noting is that the strong correlations between component contributions and gene expression also support, post hoc, the validity of the first assumed key condition for our analysis, that is, that occupancy at a given α-subunit-containing GABA_AR subtype is the same across brain regions for a given drug and dose. This conclusion follows, since a major violation of this condition would by itself preclude the possibility to achieve high correlations of the regional component contribution patterns with an independent measure on regional subunit levels such as gene expression.

After selection and application of model M3c for the derivation of components, all model variants were reevaluated by calculation of the correlation between component contribution and gene expression. Interestingly, besides the two-component M2a model, the selected M3c model was the only variant that had highly positive correlation (R > 0.70) with gene expression data for each of its components (see insert table in Supplementary Fig. S1). Note that the two models, M3d and M4d, which were deemed to likely overfit the data despite favorable statistical performance, had no high degree positive correlation with gene expression for any of their displaceable components. This reevaluation further strengthens the selection of model M3c as the most appropriate one.

Limitations of the study were the low number of subjects and the lack of head-to-head data for the two drugs. However, relatedly, the strength of the present study was that the same PET system and experimental protocol were employed in all cases. A further limitation could be that pons was used as reference region in the quantification of BP_ND estimates, instead of using a metabolite-corrected arterial plasma input function (Nørgaard et al., 2021; Odano et al., 2009). However, the original two PET studies (Jucaite et al., 2017) did not include arterial sampling. Earlier studies with [¹¹C]flumazenil suggested that pons may have specific binding albeit much lower than gray matter (Braestrup et al., 1977; Persson et al., 1991). Importantly, however, a subsequent study, including a thorough cross-validation of quantification methods using plasma or reference region radioactivity as input, found no detectable specific binding in pons, concluding that using it as the reference regions yields accurate binding estimates (Odano et al., 2009).

The aforementioned strong correlation between our data and previously published results on estimated regional BZR density, obtained from using a measured plasma input function, supports the presently applied methodology (Supplementary Fig. S18). It is important to note that voxels in the pons were not explicitly removed from the analysis. Instead, when identifying valid voxels as described in step 6 of the image processing workflow, most pontine voxels were automatically eliminated due to their below-threshold baseline binding. As shown in the results, the AHRA substructures for the remaining valid basal part of pons and the pontine tegmentum had low average baseline binding. Moreover, encouraged by the strong observed correlation with previously estimated BZR density, we used the equation for the corresponding linear regression (Supplementary Fig. S18) to cast the overall [¹¹C]flumazenil binding map to BZR density units from which tentative absolute BZR density (B_max) contribution maps were also calculated for each component (note that all supplementary tabulated data contain both [¹¹C]flumazenil binding and projected BZR density data).

GABA_ARs containing different α-subunits are known to have differential distribution in rodent brain and have been associated with different functional roles (Fritschy & Mohler, 1995; Rudolph et al., 2001). Knowledge on GABA_AR subunit distribution in the human brain has primarily emerged from immunohistochemistry studies. Though very detailed, those studies have been limited to selective regions, such as hippocampus, amygdala, subthalamic nucleus, or certain cortical areas (Sacco et al., 2009; Stefanits et al., 2018; Wu et al., 2018). In this perspective, the main contribution of the present study is a neuroanatomically more exhaustive map of putative GABA_AR α-subunit distribution across all brain regions in the human brain in vivo. To facilitate use of these maps for further research, the component contributions, overall flumazenil binding (BP_ND), and BZR density (B_max) maps, as well as the AHRA atlas re-formatted as described in Section 2, are made available online at https://doi.org/10.12751/g-node.83wfme. We see potential application of fractional maps for drug development, informing on the distribution of and balance between α-subunits as specific targets for selective drugs. In addition, the absolute GABA_AR subunit maps may be of interest for development of selective radioligands and consequently studying brain biochemistry and pathophysiology of CNS disorders.

The present study extends knowledge on the GABA_AR system by demonstrating the feasibility of disentangling specific patterns to the GABA_ARs complexes containing differentiated α-subtypes in vivo in the human brain. Further development of the molecular imaging toolbox could in turn invigorate the search for more optimized subtype-selective GABA_AR compounds. The success of these efforts would be of benefit in basic research as well as in developing treatment for neurological and psychiatric disorders.

Besides the comprehensive online data and results depository described in Section 4.2, other data used in the manuscript will be made available on request by email to the corresponding author pending approval.

Zsolt Cselényi: Conceptualization, Methodology, Software, Validation, Formal Analysis, Data Curation, Writing—Original Draft, Writing—Review & Editing, Visualization. Aurelija Jucaite: Conceptualization, Investigation, Writing—Original Draft, Writing—Review & Editing. Lars Farde: Conceptualization, Resources, Writing—Review & Editing.

The PET data were acquired in the clinical trials sponsored by AstraZeneca.

Zsolt Cselényi, Aurelija Jucaite are AstraZeneca employees. Lars Farde was AstraZeneca employee during the time of the clinical studies.

This research has been conducted using the data from AZ clinical trials (ClinicalTrials.gov identifier D1140C00007, NCT00681720 and D0850C00011, NCT00681746). Lars Farde acknowledges support from the Swedish Scientific Council (Grant 2015-02398).

Supplementary figures, tables, and tabulated data can be found in the online version of the article, at https://doi.org/10.1162/imag_a_00464. The first supplementary file contains a list and short description of all the supplements.