Separating dopamine D₂ and D₃ receptor sources of [¹¹C]-(+)-PHNO binding potential: independent component analysis of competitive binding

Abstract

Development of medications selective for dopamine D₂ or D₃ receptors is an active area of research in numerous neuropsychiatric disorders including addiction and Parkinson’s disease. The positron emission tomography (PET) radiotracer [¹¹C]-(+)-PHNO, an agonist that binds with high affinity to both D₂ and D₃ receptors, has been used to estimate relative receptor subtype occupancy by drugs based on a priori knowledge of regional variation in the expression of D₂ and D₃ receptors. The objective of this work was to use a data-driven independent component analysis (ICA) of receptor blocking scans to separate D₂- and D₃-related signal in [¹¹C]-(+)-PHNO binding data in order to improve the precision of subtype specific measurements of binding and occupancy. Eight healthy volunteers underwent [¹¹C](+)-PHNO PET scans at baseline and at two time points following administration of the D₃-preferring antagonist ABT-728 (150–1000 mg). Parametric binding potential (BP_ND) images were analyzed as four-dimensional image series using ICA to extract two independent sources of variation in [¹¹C]-(+)-PHNO BP_ND. Spatial source maps for each component were consistent with respective regional patterns of D₂- and D₃-related binding. ICA-derived occupancy estimates from each component were similar to D₂- and D₃-specific occupancy estimated from a region-based approach (intraclass correlation coefficients > 0.95). ICA-derived estimates of D₃ receptor occupancy improved quality of fit to a single site binding model. Furthermore, ICA-derived estimates of the regional fraction of [¹¹C]-(+)-PHNO binding related to D₃ receptors was generated for each subject and values showed good agreement with region-based model estimates and prior literature values. In summary, ICA successfully separated D₂- and D₃-related components of the [¹¹C]-(+)-PHNO binding signal, establishing this approach as a powerful data-driven method to quantify distinct biological features from PET data composed of mixed data sources.

1. Introduction

Dopamine transmission is a critical component of neural circuits controlling motor function, motivation, and reward. Impaired dopamine function is linked to disorders including Parkinson’s disease, addiction, and schizophrenia, and altered expression or activity of dopamine D₂- and D₃-type receptors is a common feature in these disorders. D₂ receptors are densely and widely expressed in the striatum. Expression of D₃ receptors is lower and more spatially restricted, with concentrations highest in midbrain nuclei, moderate in ventral striatum and pallidum, and low in dorsal striatum (Gurevich and Joyce, 1999; Murray et al., 1994; Piggott et al., 1999). Along with differential expression, growing evidence indicates that D₂ and D₃ receptors play distinct roles in disease. In substance use disorders, pre-clinical and clinical studies suggest that D₂ binding is typically reduced but D₃ binding is increased (Boileau et al., 2015, 2012; Le Foll et al., 2005; Martinez et al., 2004; Matuskey et al., 2015; Neisewander et al., 2004; Payer et al., 2014; Staley and Mash, 1996; Volkow et al., 2007, 1997, 1990; Worhunsky et al., 2017). The opposite pattern has been observed in Parkinson’s disease (Boileau et al., 2009; Payer et al., 2016). In schizophrenia, where D₂ antagonists have long been a mainstay for treatment of positive symptoms, D₃ receptors are emerging as a potential target for treatment of the negative and cognitive symptoms (Girgis et al., 2016; Sokoloff and Le Foll, 2017). Research in these areas would be enhanced by the ability to measure D₂ and D₃ receptors separately in vivo.

The positron emission tomography (PET) radiotracer [¹¹C]-(+)-4-propyl-3,4,4a,5, 6, 10 b-hexahydro-2Hnaphtho[1,2-b][1,4]oxazine-9-ol ([¹¹C]-(+)-PHNO) (Willeit et al., 2006) allows quantification of D₂/D₃ receptor binding availability in vivo. In contrast to the other commonly used PET tracers for dopamine receptors such as [¹¹C]raclopride and [¹⁸F]fallypride, [¹¹C]-(+)-PHNO binds with markedly higher affinity to D₃ than D₂ receptors (Gallezot et al., 2014; Narendran et al., 2006; Rabiner et al., 2009) and provides a strong signal in both D₃- and D₂-rich regions. In regions of mixed D₂ and D₃ expression such as ventral striatum (VS), [¹¹C]-(+)-PHNO specific binding represents a sum of D₂ and D₃ receptor availability weighted by its relative affinity (Searle et al., 2013; Tziortzi et al., 2011). Given these properties, [¹¹C]-(+)-PHNO PET has been a valuable tool to assess receptor occupancy and subtype selectivity of existing and emerging dopaminergic drugs (Di Ciano et al., 2019; Girgis et al., 2016, 2015; Graff-Guerrero et al., 2010, 2009; Le Foll et al., 2016; Mizrahi et al., 2011; Mugnaini et al., 2013; Searle et al., 2013; Tateno et al., 2018; Tziortzi et al., 2011). These estimates typically involve comparing displacement of tracer binding in D₂- or D₃-rich regions, with signals in the caudate or putamen commonly taken as representative of D₂ binding while those in D₃-expressing regions including substantia nigra (SN) or globus pallidus (GP) are generally interpreted as D₃ binding. While this approach has been successful in assessing receptor selectivity broadly, regional variability can present a challenge for detailed interpretation of subtype-specific effects (Le Foll et al., 2016.; Mizrahi et al., 2011).

Independent component analysis (ICA) is a data-driven method used to identify multiple sources of variance in a dataset, ‘unmixing’ observed signals into maximally independent components. We recently performed a study using ICA to investigate [¹¹C]-(+)-PHNO binding in individuals with cocaine use disorder and healthy controls (Worhunsky et al., 2017). Without a priori information, ICA identified components of [¹¹C]-(+)-PHNO binding with spatial sources that were similar to dopaminergic pathways and closely followed known patterns of D₂ and D₃ receptor distribution. This analysis concurrently replicated the opposing changes in D₂- and D₃-related binding previously observed only in separate regions in studies of substance use disorders. While this initial finding suggests that ICA-derived components represent distinct estimates of D₂- and D₃-related binding, [¹¹C]-(+)-PHNO blocking studies with selective compounds would provide stronger evidence for subtype specificity of ICA components.

In brain imaging, ICA has been used to identify components of within-subjects spatiotemporal variance in functional magnetic resonance imaging (e.g., resting-state functional brain networks) (Calhoun et al 2001; Allen et al 2011) or components of between-subject spatial variance in cross-sectional PET study designs (Di et al., 2012; Pagani et al., 2016). The objective of the present study was to assess the utility of applying four-dimensional ICA to a competition binding PET study for the estimation of subtype-specific receptor occupancy. ICA results from a [¹¹C]-(+)-PHNO PET study with the D₃ antagonist ABT-728 were compared to traditional region-of-interest (ROI)-based estimates of local D₂ and D₃ binding concentrations, receptor occupancies, regional fractions of D₃ receptor binding, and concentration-occupancy curves and EC₅₀. We hypothesized that such an approach would improve the precision of these estimates by using data from the entire brain rather than a few selected regions.

2. Methods

2.1. Study design

Eight healthy male volunteers (mean age 33±11 years) were scanned with [¹¹C]-(+)-PHNO at baseline. Participants were then administered a single oral dose of the D₃-specific antagonist ABT-728 (150 mg, n=1; 500 mg, n=3; 1000 mg, n=4), which has 134-fold greater affinity for D₃ receptors over D₂ and >600-fold higher affinity over D₁ and D₄ receptors in vitro. Post-drug scans were performed at approximately 2 hours post-dose (mean 2.2±0.14 h) and again at approximately 24 hours post-dose (mean 22.5 ± 2.2 h), for a total of three scans per subject. For one subject the second blocking scan was performed at 8.1 h post-dose. Plasma concentration of ABT-728 was measured in venous blood samples drawn at 60, 90, and 120 minutes after tracer injection and mean plasma drug concentration was determined for each scan.

2.2. [¹¹C]-(+)-PHNO PET scans

[¹¹C]-(+)-PHNO was prepared as previously described (Gallezot et al., 2014; Wilson et al., 2005). Dynamic PET data were acquired for 120 minutes on a Siemens high-resolution research tomograph (HRRT; Siemens/CTI, Knoxville, TN, USA). Head motion was tracked using an optical detector (Vicra, NDI Systems, Waterloo, Ontario, Canada). A T1-weighted anatomical MR image was collected for each participant using a Siemens 3T Trio system (Siemens Medical Solutions, Malvern, PA) and standard magnetization prepared rapid acquisition gradient echo (MPRAGE) sequence (repeat time (TR)/echo time (TE)=2530/3.34, flip angle=7°, in-plane resolution=0.98x0.98 mm, matrix=256x256, slice thickness=1 mm, slices=176).

PET data were reconstructed with corrections for attenuation, normalization, scatter, randoms, deadtime and motion using the MOLAR algorithm (Carson et al., 2003), resulting in reconstructed image resolution of ~3 mm. Parametric whole-brain images of [¹¹C]-(+)-PHNO BP_ND were computed using a simplified reference tissue model (SRTM2, Wu and Carson, 2002) with the cerebellum as reference (Gallezot et al., 2014). Spatial processing of parametric images was performed using SPM12 (Wellcome Trust Centre for Neuroimaging, London, UK). A summed [¹¹C]-(+)-PHNO uptake (0–10 min post-injection) image was created from the motion-corrected PET data for each scan session and rigidbody registered to the subject’s MR image. Nonlinear registration of the MR image to MNI152 space was determined using optimized unified segmentation (Ashburner and Friston, 2005). The combined linear and nonlinear registrations were applied to the parametric BP_ND images and smoothed with a 4 mm FWHM Gaussian kernel.

2.3. Independent Component Analysis

Four-dimensional ICA was performed on the parametric images of [¹¹C]-(+)-PHNO BP_ND using the Group ICA of fMRI Toolbox (GIFT; http://trendscenter.orgsoftware/gift/) (Calhoun et al., 2001). The three parametric images for each subject were entered into ICA as the fourth dimension (usually the time dimension for fMRI analysis). The average BP_ND values across all 24 scans was used to define a mask at BP_ND > 0.3 to constrain the analysis to voxels of interest. The average BP_ND values across all 24 scans was used to define a mask at BP_ND > 0.3 to constrain the analysis to voxels of interest. The global mean within the mask was removed from each scan to facilitate component extraction while preserving original BP_ND scaled units in the identified solution.. The mean number of maximally independent components present in each subject’s time series was determined to be two using a minimum description length criterion (Li et al., 2007). Data from all subjects were concatenated into a single group and reduced to three components using principal component analysis prior to extracting the two independent components using the Infomax algorithm (Bell and Sejnowski, 1995). ICA was iterated 25 times using ICASSO to assess stability of extracted components and determine the most reliable extraction for the final output (Himberg et al., 2004).

ICA identifies components of coherent variance, consisting of loading values for each subject at each scan for each of the M components ($A_{M,i}^t$; M = 1,2; i = 1, …,8; t =1,2,3), where A denotes the estimate of these loading values, the unknown mixing matrix (i.e. the inverse of ICA-derived un-mixing matrix). Note that the scan loadings also incorporate the PCA-reduction factor and are more accurately characterized as (R × A), but here we use to represent the final ICA-derived loading parameters. ICA also produces maps of the spatial source intensities of each component (y_M). Scan loading values and spatial source maps are related to the original data in such a way that an approximation ($\widetilde{BP}$) of the original BP_ND data (x) can be reconstructed as the sum of their products (e.g., ${\widetilde{BP}}_{M,i}^t=y_M\times A_{M,i}^t$) and the global mean value for each scan (${\overline{x}}_i^t$):

$${\widetilde{BP}}_i^t=y_1\times A_{1,i}^t+y_2\times A_{2,i}^t+{\overline{x}}_i^t$$ #(1)

2.4. ROI analyses

Seven bilateral ROIs frequently used in [¹¹C]-(+)-PHNO PET studies were examined: dorsal caudate (DCA), dorsal putamen (DPU), ventral pallidum (VP), GP, VS, hypothalamus (HY), and SN. ROIs were defined as in earlier work (Worhunsky et al., 2017) according to previously described guidelines (Mawlawi et al., 2001; Tziortzi et al., 2011): DCA, DPU and VS ROIs from FSL (Oxford Centre for Functional MRI of the Brain, Oxford, UK); GP, HY, and VP manually defined on the ICBM-MNI52 template (Montreal Neurological Institute, Montreal, Canada); and SN ROI manually drawn on a groupaverage parametric [¹¹C] -(+)-PHNO BP_ND image from an independent sample (Gallezot et al., 2014).

ROI masks were applied to individual smoothed parametric BP_ND images in standard space to determine mean [¹¹C]-(+)-PHNO BP_ND in each region. To examine the contribution of each identified component to the total estimated $\widetilde{BP}$ in a given brain region, mean regional intensity of each spatial source map (y_M) was extracted from each of the seven ROIs. These values were multiplied by scan loading values to generate component-specific partial $\widetilde{BP}$ values for each source, scan and ROI (i.e., regional ${\widetilde{BP}}_{M,i}^t$).

The regional partial $\widetilde{BP}$ values of each ICA component were compared with regional D₂- and D₃-related BP_ND calculated using literature proportions of [¹¹C]-(+)-PHNO BP_ND as previously (Worhunsky et al., 2017). Briefly, regional BP_ND values were multiplied by the average published estimates of the fraction of [¹¹C]-(+)-PHNO BP_ND reflecting D₂- and D₃-related specific binding respectively (Searle et al., 2013; Tziortzi et al., 2011) to produce $B{P}_{ND}^{D2}\left(f^{D2}\times B{P}_{ND}\right)$ and $B{P}_{ND}^{D3}\left(f^{D3}\times B{P}_{ND}\right)$. Pearson correlations were assessed between $B{P}_{ND}^{D2}$ and $B{P}_{ND}^{D3}$ and the mean baseline partial $\widetilde{BP}$ values for each component.

2.5.1. Region-based model for receptor occupancy and f^D3

For comparison with ICA results (see below), D₂- and D₃-specific occupancy (r^D2 and r^D3) were estimated using a model-based approach and ROI BP_ND values, assuming constant occupancy across the brain.

Apparent overall occupancy by a D₃ antagonist will vary with the proportion of [¹¹C]-(+)-PHNO binding attributable to D₃ receptors. To estimate receptor-specific occupancy, r, total apparent occupancy was computed from BP_ND in each region:

$$r^{app}=\frac{B{P}_{ND}^{\mathit{\text{baseline}}}-B{P}_{ND}^{\mathit{\text{block}}}}{B{P}_{ND}^{\mathit{\text{baseline}}}}$$ #(2)

Total BP_ND during blocking scans comprises available (unblocked) D₂ and D₃ receptors:

$$B{P}_{ND}^{\mathit{\text{block}}}=f^{D2}B{P}_{ND}^{\mathit{\text{baseline}}}\left(1-r^{D2}\right)+f^{D3}B{P}_{ND}^{\mathit{\text{baseline}}}\left(1-r^{D3}\right)$$ #(3)

where f^D2 and f^D3 are the fractions of [¹¹C]-(+)-PHNO binding attributable to D₂ and D₃ receptors in a given region (f^D2 + f^D3 = 1). D₂ and D₃ receptor occupancy during each blocking scan are therefore related to regional apparent occupancy according to:

$$r^{app}=\left(1-f^{D3}\right)r^{D2}+f^{D3}{r}^{D3}$$ #(4)

Based on prior [¹¹C]-(+)-PHNO blocking studies (Searle et al., 2013; Tziortzi et al., 2011), it was assumed that f^D3 = 1 in SN and f^D3 = 0 in DPU. This model was fitted to the seven ROIs simultaneously to produce estimates of r^D2, r^D3, and f^D3in the five remaining regions. Models were fitted separately for each subject using data from the early blocking scan and the constraints 0 ≤ r ≤ 0100% and 0 ≤ f^D3 ≤ 1. In late scans, limited range of occupancies (<25% at D₂ receptors) made the full model estimation unstable and so regional f^D3 values were fixed to the value from each subject’s early scan and only r^D2 and r^D3 were estimated.

2.5.2. ICA-derived receptor occupancy and f^D3

ICA loading values are estimates of the contribution from each component to the total $\widetilde{BP}$ in a given scan. Source maps provide voxel-wise spatial information that is constant across scans and subjects. Scan loading values can therefore be compared directly between baseline and blocking scans to determine how each independent source of [¹¹C]-(+)-PHNO binding is influenced by the blocking drug. Therefore, component-specific occupancy measures were calculated from the loading values as:

$$r_i^M=\frac{A_{M,i}^{\mathit{\text{baseline}}}-A_{M,i}^{\mathit{\text{block}}}}{A_{M,i}^{\mathit{\text{baseline}}}}$$ #(5)

Concordance between region-based and ICA-derived occupancy estimates was assessed by computing intra-class correlation coefficients (ICC) using a two-way mixed effects model for absolute agreement of individual values (Koo and Li, 2016; Shrout and Fleiss, 1979).

From ICA results, the fraction of [¹¹C]-(+)-PHNO signal attributable to D₃ receptor binding in a given region or voxel can be calculated if the fraction of D₃ binding in each component is known, i.e., for each subject, this is the sum of D₃-related portion of $\widetilde{BP}$ information in each component and in the global mean BP_ND value (i.e., ${\widetilde{BP}}^{D3}=f_1^{D3}{y}_1{A}_1+f_2^{D3}{y}_2{A}_2+f_{\overline{x}}^{D3}\overline{x}$), relative to overall ICA-reconstructed $\widetilde{BP}$:

$$f_i^{D3}=\frac{f_{1,i}^{D3}{y}_1{A}_{1,i}+f_{2,i}^{D3}{y}_2{A}_{2,i}+f_{\overline{x},i}^{D3}\overline{x}}{y_1{A}_{1,i}+y_2{A}_{2,i}+\overline{x}}$$ #(6)

Based on results from the blocking studies (see below), it was determined that one component (IC1) represented only D₂-related binding and the other (IC2) only D₃-related binding: $f_1^{D3}=0$ and $f_2^{D3}=1$. To estimate $f_{\overline{x}}^{D3}$, the fraction of global mean BP_ND representing D₃ receptor binding, occupancy estimates for each component were taken to represent respective D₂- and D₃-receptor occupancy. As with region-based analyses described above (eq. 4), global mean BP_ND comprises a mixture of D₂ and D₃ receptor binding, and so percent change during blocking represents a weighted sum of D₂ and D₃ receptor occupancy. Therefore, $f_{\overline{x}}^{D3}$ was estimated by fitting the following equation to ICA occupancy estimates from the early blocking scan and percent difference in global mean during blocking ($r^{\overline{x}}$):

$$r^{\overline{x}}=\left(1-f_{\overline{x}}^{D3}\right)r^{IC1}+f_{\overline{x}}^{D3}{r}^{IC2}$$ #(7)

Regional f^D3 was then calculated for each subject as outlined above (eq. 6) using mean regional spatial intensity of each source, baseline scan loading values, global mean BP_ND, and $f_{\overline{x}}^{D3}$, estimated from equation 7. ICA-derived regional f^D3 values were compared to values estimated simultaneously with receptor occupancy in region-based analyses.

2.5.3. Concentration-occupancy curves and EC₅₀

Across subjects, concentration-occupancy curves were constructed using measured plasma drug concentrations for region-based D₂- and D₃-specific occupancy estimates and for each ICA component. Each curve was fitted to a single-site binding model to produce estimates of EC₅₀, the plasma drug concentration at which 50% of each binding site is occupied. For D₂ receptors, due to the limited range of measured occupancy values, models were fitted assuming maximum occupancy of 100% and estimating only EC₅₀. For D₃ receptors, EC₅₀ and maximum percent occupancy (≤100%) were estimated. Root mean square error (RMSE) of each curve fit was used to assess goodness of fit. Data from early and late blocking scans were analyzed together assuming a direct relationship between measured plasma drug concentration and brain occupancy.

3. Results

3.1. PET scan parameters

There were no differences between scans in [¹¹C]-(+)-PHNO molar activity (438 ± 111 MBq at baseline, 498 ± 132 MBq at early blocking scan, 492 ± 137 MBq at late blocking scan; F_2,20=0.54, p=0.59) or injected mass (0.028 ± 0.0042 μg/kg at scan 1, 0.029 ± 0.00067 μg/kg at early blocking scan, 0.027 ± 0.0056 μg/kg at late blocking scan; F_2,14=1.2, p=0.32; within-subjects comparisons between scans p’s>0.14, uncorrected).

3.2. ICA-identified components of [¹¹C]-(+)-PHNO binding

Two independent components of coherent variation in [¹¹C]-(+)-PHNO binding across the three time points were identified (Figure 1). The source of component 1 (IC1) comprised binding in the DCA, DPU, VS, and GP, while the source of component 2 (IC2) included binding in GP, VP, SN, and HY. The source intensity patterns and partial $\widetilde{BP}$ values were consistent with the expected distribution pattern of D₂- and D₃-related [¹¹C]-(+)-PHNO binding (Searle et al., 2013; Tziortzi et al., 2011), with mean regional contributions for IC1 (${\widetilde{BP}}_1$) correlating with estimated $B{P}_{ND}^{D2}$ (r = 0.94, p=0.0018) and contributions for IC2 (${\widetilde{BP}}_2$) correlating with $B{P}_{ND}^{D3}$ (r = 0.97, p=0.00035) across the seven ROIs. No such relationship was observed in the converse pairs, i.e. ${\widetilde{BP}}_1$ with $B{P}_{ND}^{D3}$ or ${\widetilde{BP}}_2$ with $B{P}_{ND}^{D2}$.

Figure 1.

ICA-identified components of [¹¹C]-(+)-PHNO BP_ND. A and D, Source maps of IC1 (A) and IC2 (D) thresholded at ICA-estimated $\widetilde{BP}>0.3$; B and E, mean (± S.E.M.) scan loading values for each source in baseline, early blocking, and late blocking scans; C and F, correlation between literature-based estimates (Searle et al., 2013; Tziortzi et al., 2011; see text) of D₂-related BP_ND and IC1 $\widetilde{BP}$ (C) and D₃-related BP_ND and IC2 $\widetilde{BP}$ (F) across regions.

3.3. ABT-728 occupancy

3.3.1. Region-based estimates.

Using an occupancy model and change in ROI BP_ND values, overall occupancy of D₃ receptors at two hours post-drug was 26% with a dose of 150 mg (n=1), 63 ± 7.4% with 500 mg, and 69 ± 3.3% with 1000 mg. Corresponding occupancy at D₂ receptors was 0%, 20 ± 3.4%, and 31 ± 8.6%, respectively (Table 1).

Table 1.

Estimated occupancy in [¹¹C]-(+)-PHNO blocking scans using ICA and regional analyses.

Subject	Dose (mg)	Blocking scan	Plasma concentration ABT-728 (μg/mL)	Occupancy
				Regional BPND		ICA
				D2	D3	IC1	IC2
1	150	1	3.4	0.0%	26%	−1.6%	32%
		2	1.0	0.0%	2.7%	−3.8%	6.9%
2	500	1	11.0	17%	64%	15%	60%
		2	8.1	23%	66%	21%	59%
3	500	1	12.0	22%	69%	21%	66%
		2	4.5	3.8%	39%	3%	36%
4	500	1	11.8	23%	55%	21%	57%
		2	3.7	9.7%	38%	8%	34%
5	1000	1	19.7	28%	65%	30%	79%
		2	7.4	15%	49%	14%	56%
6	1000	1	25.2	25%	68%	25%	71%
		2	8.2	12%	54%	9%	55%
7	1000	1	25.2	44%	73%	45%	85%
		2	8.5	15%	50%	16%	57%
8	1000	1	20.1	29%	70%	31%	76%
		2	8.7	12%	56%	13%	58%

Region-based occupancy values were estimated by fitting a model (eq. 4) to regional BP_ND values in baseline and blocking scans. ICA-based occupancy values are percent difference from baseline to blocking scans in loading value for each component, IC1 and IC2. See text for details.

3.3.2. ICA-based estimates.

Scan loading values for IC2 were markedly reduced during blocking, with a smaller magnitude of change in IC1. At two hours post-dose, occupancy in IC2 was 32% with 150 mg, 61 ± 4.5% with 500 mg, and 78 ± 5.8% with 1000 mg; corresponding values for IC1 were −2%, 19 ± 3.7%, and 33 ± 5.3% (Table 1). Across subjects, IC1 occupancy values were closely matched to region-based estimates of occupancy at D₂ receptors (ICC=0.98) and IC2 occupancy to occupancy at D₃ receptors to (ICC=0.95) (Figure 2). Altogether, spatial source patterns of each component and occupancy analyses strongly support the interpretation that IC1 corresponds to D₂-related binding in overall [¹¹C]-(+)-PHNO BP_ND and IC2 to D₃-related binding. Note that ICA extracted this pattern with no a priori information.

Figure 2.

Relationship between ROI-based estimates of D₂ receptor occupancy and IC1 occupancy from ICA (A) and between ROI-based D₃ receptor occupancy and IC2 occupancy (B).

3.4. Regional proportion of D₃ binding

3.4.1. Region-based estimates.

Regional f^D3 values estimated simultaneously with receptor occupancy in region-based analyses are shown in Table 2. From this method, mean f^D3 was 0.74 ± 0.17 in HY, 0.59 ± 0.087 in GP, 0.56 ± 0.14 in VP, 0.24 ± 0.057 in VS, and 0.068 ± 0.073 in DCA.

Table 2.

Regional f^D3 estimated from ROI BP_ND values and from ICA.

	SN		HY		VP		GP		VS		DCA		DPU
subj	region	ICA	region	ICA	region	ICA	region	ICA	region	ICA	region	ICA	region	ICA
1	Fixed to 1	0.91	1.00	0.64	0.43	0.47	0.62	0.51	0.23	0.19	<10−8	0.04	Fixed to 0	−0.010
2		0.89	0.62	0.72	0.45	0.55	0.53	0.59	0.24	0.26	<10−8	0.13		0.047
3		0.96	0.57	0.71	0.55	0.53	0.57	0.57	0.15	0.24	0.02	0.08		0.012
4		0.92	0.83	0.68	0.39	0.50	0.51	0.54	0.18	0.22	0.04	0.09		0.022
5		0.81	0.82	0.61	0.80	0.47	0.70	0.50	0.32	0.18	0.10	0.02		−0.028
6		0.87	0.57	0.62	0.54	0.47	0.45	0.50	0.24	0.18	0.03	0.02		−0.024
7		0.91	0.90	0.63	0.72	0.46	0.67	0.49	0.29	0.18	0.16	0.04		−0.0089
8		0.91	0.61	0.65	0.61	0.49	0.65	0.52	0.29	0.20	0.18	0.03		−0.022
mean		0.90	0.74	0.66	0.56	0.49	0.59	0.53	0.24	0.21	0.068	0.056		-0.0018
S.D.		0.044	0.17	0.043	0.14	0.033	0.087	0.037	0.057	0.031	0.073	0.039		0.026

SN, substantia nigra; HY, hypothalamus; VP, ventral pallidum; GP, globus pallidus; VS, ventral striatum; DCA, dorsal caudate; DPU, dorsal putamen.

3.4.2. ICA-based estimates.

Regional f^D3 was also computed from ICA results using the apparent evidence that IC1 and IC2 represent [¹¹C]-(+)-PHNO binding to D₂ and D₃ receptors, respectively (i.e., $f_1^{D3}=0$ and $f_2^{D3}=1$). Thus, D₃-related binding was calculated as the sum of regional ${\widetilde{BP}}_2$ (i.e., from IC2) and D₃-specific binding in the global mean. Regional f^D3 values for each subject are presented in Table 2. Mean f^D3 was 0.92 ± 0.079 in SN, 0.66 ± 0.043 in HY, 0.53 ± 0.037 in GP, 0.49 ± 0.033 in VP, 0.21 ± 0.031 in VS, 0.56 ± 0.039 in DCA, and −0.002 ± 0.026 in DPU (Figure 3, top row). In all regions, ICA-derived f^D3 values were in good agreement with region-based estimates, with a tendency to be slightly lower (Table 2). ICA estimates were markedly less variable across subjects compared to region-based f^D3, with a coefficient of variation ranging from 3.5% in HY to 8.4% in DCA for ICA and from 13.7% in GP to 108% in DCA for region-based values. Mean values from both methods were consistent with group-level regional f^D3 estimates from previous studies (Searle et al., 2013; Tziortzi et al., 2011).

Figure 3.

ICA-derived fraction of [¹¹C]-(+)-PHNO signal representing D₃ binding (f^D3), mean of eight subjects. Top row, f^D3 estimates displayed uniformly in each ROI (see Table 2). Bottom row, voxel-wise f^D3 maps showing voxel-level variation in D₃ binding proportion within masked region of mean BP_ND>0.3: a, varying f^D3 within GP following a decreasing internal-external gradient; b, low f^D3 in DCA and DPU increasing into VS; c, overall decreasing rostral-caudal gradient within DCA and DPU; d, increasing rostral-caudal gradient within VS; e, high f^D3 within SN. DCA, dorsal caudate; DPU, dorsal putamen; GP, globus pallidus; VP, ventral pallidum; HY, hypothalamus; VS, ventral striatum; SN, substantia nigra.

A voxel-wise map of mean f^D3 across subjects is shown in Figure 3 (bottom row), demonstrating varied proportion of D₃-related binding within GP, low f^D3 in DCA and DPU with values increasing into the VS, and high D₃ binding proportion in SN. Variation within regions followed patterns similar to those identified in prior [¹¹C]-(+)-PHNO PET and post-mortem studies (Gurevich and Joyce, 1999; Murray et al., 1994; Tziortzi et al., 2011). For example, the proportion of D₃ binding appeared to be higher in the internal compared to external segment of the GP (bottom left and center panels, a) and followed an increasing rostral to caudal gradient in GP and VS but an overall decreasing rostral-caudal gradient in dorsal striatum (bottom center and right, b-d).

3.5. Concentration-occupancy relationships

Concentration-occupancy curves were produced for both region-based D₂ and D₃ estimates (Figure 4A) and ICA-identified components of [¹¹C]-(+)-PHNO binding (Figure 4B). ABT-728 EC₅₀ estimates (Table 3) were similar between region-based D₂ occupancy and IC1 (D₂, 48 μg/mL SE 8.5; IC1, 48 μg/mL SE 9.0) and between D₃ occupancy and IC2 (D₃, 5.6 μg/mL SE 2.6; IC2, 6.7 μg/mL SE 2.0). Thus, EC₅₀ was approximately 7-fold lower for D₃ compared to D₂ receptors. This is in line with in vitro studies showing higher affinity at D₃ receptors, though markedly smaller than the 140-fold difference in Ki values observed in those experiments (Investigator’s Brochure, Abbott Laboratories). The phenomenon of lower selectivity in vivo has previously been noted for other D₃-targeting drugs (Searle et al., 2013). For D₃ receptors, estimated maximum occupancy was 89% from region-based values and 100% from IC2 (Table 3). Concentration-occupancy curve fits were qualitatively better for IC2 occupancy values than for region-based D₃ occupancy estimates (RMSE 4.2 compared to 6.1; Figure 4). For D₂-like binding, quality of fit was comparable for IC1 and region-based estimates (RMSE 5.3 and 5.0, respectively).

Figure 4.

ABT-728 concentration-occupancy curves. Percent occupancy from [¹¹C]-(+)-PHNO PET scans at each plasma drug concentration for region-based D₂- and D₃-specific occupancy estimates (A) and for ICA-derived IC1 and IC2 occupancy (B).

Table 3.

[¹¹C]-(+)-PHNO dose-occupancy results.

	Region-based		ICA
	D2	D3	IC1	IC2
EC50 (S.E.), μg/mL	48 (8.5)	5.6 (2.6)	48 (9.0)	6.7 (2.0)
Maximum occupancy, %	fixed to 100	89 (S.E. 9.0)	fixed to 100	~100
RMSE	5.0	6.1	5.3	4.2

RMSE, root mean square error.

4. Discussion

This study presents a novel data-driven method for assessing D₂- and D₃-specific binding and occupancy using four-dimensional ICA of [¹¹C]-(+)-PHNO PET data. ICA identified two distinct components of change in binding on the basis of spatiotemporally coherent variance across subjects and time points. The respective spatial sources of these components were highly consistent with D₂- and D₃-related [¹¹C]-(+)-PHNO binding distributions in the brain (Figure 1), suggesting that this analysis successfully separated binding signal specific to each receptor subtype without a priori assumptions. This interpretation is supported by relative changes in intensity of each source during blockade with the D₃-selective antagonist ABT-728, which were closely matched to region-based occupancy estimates. Component loading was used to determine the proportion of D₃-specific binding in the [¹¹C]-(+)-PHNO signal across the brain, producing values consistent with those from prior literature but allowing for individual variation in D₃ receptor binding contribution and visualization of this measure at the voxel level. ICA-based values of D₃ occupancy improved quality of fit to a single-site concentration-occupancy model compared to region-based values, suggesting that ICA may improve precision in D₃ BP_ND estimates by using information from a greater number of voxels.

4.1. Sources of [¹¹C]-(+)-PHNO BP_ND

In a previous analysis of a separate sample of healthy volunteers and individuals with cocaine use disorder, cross-sectional spatial ICA was used to estimate three independent sources of [¹¹C]-(+)PHNO, two of which were associated with estimated D₂- or D₃-specific binding (Worhunsky et al., 2017). IC1 in the present results follows a similar spatial pattern to the striatopallidal source in the earlier analysis. In both cases, regional source-specific $\widetilde{BP}$ (here, ${\widetilde{BP}}_1$) was correlated with estimated D₂-related BP_ND. Similarly, IC2 here resembles the pallidonigral source in the previous study, and source-specific $\widetilde{BP}$ contributions for both were associated with D₃-related BP_ND. Thus, qualitatively similar sources linked to the same biological features were extracted in two independent samples composed of different data types (previously, patients and healthy controls; here, baseline and blocking scans in healthy volunteers) and using different ICA approaches (previously spatial ICA; here, spatiotemporal ICA). These data strongly support the utility of ICA to identify distinct D₂- and D₃-related binding in [¹¹C]-(+)-PHNO data.

4.2. Determination of D₂- and D₃-specific occupancy and f^D3

ICA produced similar estimates of D₂- and D₃-specific occupancy compared to region-based analyses. Previous studies have estimated the relative D₂ and D₃ occupancy of new or existing medications using [¹¹C]-(+)-PHNO PET by either comparing occupancy in regions presumed to represent D₂ or D₃ binding only or by fitting occupancy models as here (Di Ciano et al., 2019; Girgis et al., 2016; Graff-Guerrero et al., 2010; Mizrahi et al., 2011; Tateno et al., 2018). While these studies have produced valuable results, data-driven ICA of [¹¹C]-(+)-PHNO PET results may improve precision by avoiding some of the challenges and confounds associated with ROI- and model-based analysis. For example, the relatively greater difficulty of delineating D₃-rich regions such as the SN compared to D₂-rich dorsal striatum regions may introduce bias in comparing occupancy between receptor subtypes (Le Foll et al., 2016). Indeed, concentration-occupancy curves for D₃ receptors constructed with ICA-estimated values yielded qualitatively superior fits compared to region-based estimates. Thus, we propose that ICA offers a data-driven approach to estimating D₃-specific binding that provides improved quantification compared to current model-based methods.

Possible limitations associated with using literature estimates of regional D₃ receptor distribution were evident in the present work. For region-based analyses, it was assumed that DPU BP_ND reflects only D₂-related binding and SN BP_ND only D₃-related binding, as in previous [¹¹C]-(+)-PHNO studies. However, while ICA-derived estimates of f^D3 were not significantly different from zero (t=0.19, p=0.85), in SN f^D3 was only approximately 0.9, i.e., not 1.0. This is remarkably similar to the estimated maximum D₃ occupancy of 89% in the region-based concentration-occupancy curve, compared to 100% in the ICA-derived curve. While other biological factors may contribute to sub-maximal D₃ occupancy, this observation suggests that assuming complete D₃ binding in SN may introduce bias into the region-based estimates. This would account for the tendency towards slightly higher f^D3 estimates across regions with the region-based approach (Table 2).

To determine if region-based results could be improved by avoiding assumptions of pure subtype binding in SN and DPU, we also fitted the occupancy model (eq. 4) to the seven ROIs estimating f^D3 for all ROIs (as well as r^D2 and r^D3), rather than fixing f^D3 in SN and DPU and estimating this parameter in only the remaining five regions. This produced f^D3 estimates in mixed regions that were frequently inconsistent with literature estimates, particularly in GP (lower than expected) and in DCA and VS (higher than expected) (Supplemental Table 1), and worse fits to a concentration-occupancy model for D₂ receptors. We therefore conclude that results from the unconstrained model are less likely to be accurate, despite possible bias in the primary analysis. By comparison, ICA occupancy estimates, in addition to providing greater precision, do not rely on prior knowledge or assumptions about receptor distribution patterns and so avoid this issue. Notably, this potential source of bias may be exaggerated in disease populations with altered D₂ or D₃ receptor expression.

4.3. Heterogeneity in f^D3

Beyond occupancy, independent assessment of D₂- and D₃-related binding may have applications in studies aiming to understand disease and treatment mechanisms. Recent work has identified patterns of [¹¹C]-(+)-PHNO binding in Parkinson’s disease and schizophrenia that suggest differential D₂/D₃ changes, spatially restricted disease/treatment effects, or both (Mizrahi et al., 2011; Payer et al., 2016). Such overlapping effects are difficult to characterize conclusively with ROI-based analyses but could be determined by examining independent sources of binding. Further research is necessary to explore these possibilities and assess the sensitivity of ICA-derived measures to differences in the relative levels and spatial patterns of D₂ and D₃ receptor binding in patient populations, though our previous work is encouraging in this respect (Worhunsky et al., 2017).

At the voxel level, a previous study (Tziortzi et al., 2011) generated images of D₂- and D₃-related specific binding by [¹¹C]-(+)-PHNO using two-stage estimation of an ROI-based occupancy model to produce group-level f^D3 maps used to scale BP_ND images. The resulting $B{P}_{ND}^{D2}$ and $B{P}_{ND}^{D3}$ images are very similar to the respective $\widetilde{BP}$-scaled source maps for IC1 and IC2 (Figure 1) in spatial pattern and magnitude (accounting for global mean correction). Here, maps of f^D3 (Figure 3, bottom row) were also produced from ICA results to visualize heterogeneity in receptor binding proportion within regions. This illustrates the variation in relative D₂/D₃ binding within structures, unobscured by variation in total BP_ND. These images reproduce some of the patterns previously observed in post-mortem studies, including opposing rostral-caudal gradients in DCA/DPU compared to VS and the appearance of higher relative D₃ content in the internal compared to external segment of GP (Gurevitch and Joyce, 1999; Murray et al., 1994). These patterns of relative receptor binding in vivo can inform the interpretation of future [¹¹C]-(+)-PHNO PET studies.

4.4. Use of ICA in multi-target PET

In principle, similar methods could be applied to other systems and PET tracers. It is not uncommon for PET signals to comprise binding to two closely related target sites, as with [¹¹C]-(+)-PHNO, or to a primary target and one or more off-target sites, such as tau-binding radiotracers like [¹⁸F]AV-1451 that also have affinity for monoamine oxidase B (Murugan et al., 2019). While this seems an appealing and intuitive application of ICA, several important factors may limit the generalizability of this method. [¹¹C](+)-PHNO is particularly well-suited to this approach since its signal represents a linear sum of binding at two sites. Radiotracers with low specific binding or target sites with highly correlated expression may not be successfully or reliably separated. The present approach using blocking studies with selective drugs, and thus different levels of occupancy at each site, might be particularly powerful in such cases. However, the maximum number of independent components that can be extracted from this method is dimensionally limited to the number of scans performed on each subject. Future studies could explore the utility of spatial (i.e., 3-dimensional) ICA in occupancy studies where more binding targets might be anticipated, fewer time points are available, or the number of time points varies across subjects. Under favorable conditions ICA could be a useful method for more precise quantification of biological features in mixed-target PET data across systems.

4.5. Limitations

Several limitations of the present study should be noted. First, a gold standard comparator for ICA-based occupancy estimates was not available for these analyses. Accuracy of ICA results was assessed based on agreement with region-based occupancy estimates, which have limitations as discussed above, and with literature estimates of f^D3. We therefore conclude that ICA performs similarly to or better than existing methods for estimating D₂ and D₃ receptor occupancy. Both sets of analyses used SRTM estimates of BP_ND with cerebellum as reference region rather than arterial sampling with full quantification. This may not be optimal since reductions in V_T in cerebellum during blocking experiments has been observed in previous studies (Rabiner et al., 2009; Searle et al., 2013), consistent with some specific binding in the reference region which would bias BP_ND estimates. However, SRTM-derived BP_ND using cerebellum reference is by far the most commonly used outcome measure in [¹¹C]-(+)-PHNO studies, and any bias associated with this method would not be expected to affect our primary conclusion about the utility of ICA for this application. Concentration-occupancy analyses pooled scans from early and late time points post-drug, which assumes a linear relationship between drug concentration in brain and blood over time. To assess the validity of this, binding models were fit to occupancies from early and late scans separately. In each case, EC₅₀ estimates were very similar to those from the pooled analysis in Table 3. Injected mass of PHNO could also influence occupancy estimates due to binding competition by unlabeled PHNO, particularly at higher-affinity D₃ receptors (Searle et al., 2013). Applying a correction for injected mass (Gallezot et al., 2017; see Supplemental Information) did not improve variability or goodness of fits to the single-site binding model (Supplemental Table 2 and Supplemental Figure). Notably, PHNO mass influences both ROI-based and ICA-based methods, and so this does not affect our main conclusions. Given that injected mass did not differ between scans, this correction was not pursued further, but may represent a source of bias in occupancy estimates using either analysis method. Finally, as in previous analyses of D₃-specific binding with [¹¹C]-(+)-PHNO, only male subjects were included in this study. The present methods, and in vivo estimates of D₃ occupancy and regional binding patterns more generally, require validation in women.

5. Conclusions.

This analysis presents successful separation of D₂ and D₃ receptor specific binding using ICA in a [¹¹C]-(+)-PHNO PET study with a D₃-selective drug. As development continues for drugs specifically targeting D₃ receptors in neuropsychiatric disease, precise quantification of D₃ receptor availability and occupancy will become increasingly important. Given the structural similarities between the two receptors, achieving high specificity for D₃ over D₂ receptors is a challenge in drug design and precise estimates of subtype-specific occupancy in vivo are crucial. The approach presented here may offer a powerful tool to dissect the role of D₂ and D₃ receptors in human brain function, disease, and treatment. More broadly, this work demonstrates the potential for ICA methods to disentangle PET signals from tracers that have mixed binding, which would facilitate a truly quantitative understanding of in vivo neurochemistry in a range of systems and clinical applications.