Region- and voxel-based quantification in human brain of [¹⁸F]LSN3316612, a radioligand for O-GlcNAcase

Neurofibrillary tangles—aggregates of hyperphosphorylated tau protein—are a typical feature of Alzheimer's disease (AD) and other tauopathies [1]. The process of O-GlcNAcylation [2] involves attaching O-linked β-N-acetylglucosamine (O-GlcNAc) at the serine and threonine residue of tau protein [3]. Two enzymes modulate this process: O-GlcNAc transferase and O-GlcNAc hydrolase (O-GlcNAcase, OGA), which catalyze the attachment and detachment of O-GlcNAc, respectively. Because this process modifies the key protein involved in tauopathies, and because OGA is located on chromosome 10q24.1 (a locus associated with late-onset AD [4, 5]), interest has grown in O-GlcNAcylation as a therapeutic target. Notably, O-GlcNAcylation hinders abnormal phosphorylation and aggregation of tau, thus stabilizing the microtubule-associated tau protein [6, 7]. Indeed, tau-specific and overall O-GlcNAc are decreased in the brains of individuals with AD [8, 9]. Preclinical studies have also demonstrated that upregulation of O-GlcNAcylation by OGA inhibitors reduces pathologic tau phosphorylation and aggregation and prevents neurodegeneration [10‐12]. These findings suggest that OGA inhibition may be a potential strategy for treating tauopathies [13, 14].

Our laboratory recently characterized [¹⁸F]LSN3316612 (fluorine-18-N-(5-(((2S,4S)-2-methyl-4-(6-fluoropyridin-2-yloxy) piperidin-1-yl)methyl)thiazol-2-yl)acetamide), a positron emission tomography (PET) radioligand capable of measuring OGA in brain. An initial in vitro study demonstrated that [¹⁸F]LSN3316612 was a highly specific and potent OGA ligand [15] with good imaging properties in nonhuman primates [16]. Human studies subsequently confirmed that [¹⁸F]LSN3316612 was an excellent PET radioligand for imaging OGA in vivo. Specifically, brain uptake was well quantified as total distribution volume (V_T) using the two-tissue compartment model (2TCM); V_T measures showed low variability and moderate reliability under test and retest conditions, and V_T was stable about two hours after injection, suggesting no significant radiometabolite accumulation in brain [17]. However, only V_T derived from compartment modeling was examined at the region level, and 2TCM is not suitable for voxel-wise analysis due to its high computational demand [18]. Thus, the feasibility of other kinetic models remains unknown, particularly with regard to generating high-quality parametric images of [¹⁸F]LSN3316612 binding. The present study used the previously obtained test–retest dataset [17] to identify which kinetic model(s) yield binding parameters that are close to 2TCM-derived V_T, which is considered the gold standard for quantifying [¹⁸F]LSN3316612 binding.

The present study—which used previous data to explore different kinetic model(s) for quantifying [¹⁸F]LSN3316612 binding—found that Logan and MA1 were the models whose V_T estimations were the closest to those obtained using the gold standard of 2TCM, particularly with regard to generating good-quality parametric images. Compared to SA, Logan and MA1 showed the smallest bias of V_T compared to 2TCM. Although V_T values from all techniques were strongly correlated with 2TCM-derived V_T at both the region and voxel levels, the biases for Logan and MA1 were generally within − 8% of the 2TCM-derived V_T values, while SA showed a larger bias (~ 16%). In contrast, while IRF₁₂₀ was strongly correlated with 2TCM-derived V_T in both region- and voxel-wise analyses, its test–retest variability was significantly higher than for the other techniques. Therefore, IRF₁₂₀ should not be considered a viable alternative for [¹⁸F]LSN3316612 quantification. Finally, the lack of correlation between SUVR and 2TCM-derived V_T values and the absence of arteriovenous equilibrium 180 min post-injection suggest that arterial sampling is needed to quantify [¹⁸F]LSN3316612 scans.

Previous studies from our laboratory showed that [¹⁸F]LSN3316612 is suitable for clinical research [17]. Specifically, brain uptake was high and could be well-quantified as V_T with 2TCM. Furthermore, V_T was stable over time and showed low variability and fair reliability under test–retest conditions. However, quantifying increased OGA density during disease development and progression requires not only favorable kinetic properties but also validation of an appropriate analytic methodology. In this context, voxel-based analyses may enable early detection compared to region-based analyses. Theoretically, region-based analyses may be appropriate if a study has an a priori hypothesis about a specific brain region. However, voxel-based analyses may identify changes in smaller and isolated areas that take place during early disease stages. Indeed, a change in a subset of voxels may be smoothed out when all voxel values are averaged together to a large ROI. Moreover, voxel-based analyses examine the whole brain without a priori spatial assumption and are more appropriate for explorative studies of radioligands that bind across all brain regions, such as [¹⁸F]LSN3316612. Voxel-based analyses are more affected by noise, and adjacent voxels are not statistically independent observations because they are spatially correlated due to the finite resolution of the scanner. In our study, however, the impact of noise at the voxel level was minimized by merging the individual voxels of the parametric images into relatively noiseless large regions.

Although 2TCM is considered to be the gold standard for quantifying [¹⁸F]LSN3316612 in the brain, it is unsuitable for voxel-level analyses. The conventional nonlinear least-squares fitting is susceptible to the high noise encountered at the voxel level [26]. As an example, Rizzo and colleagues [27] found that generating parametric images of [¹¹C](R)-rolipram with 2TCM took about 20 h of computing time per participant and resulted in a map where more than half of voxels had to be discarded because of outlying values. Parametric images may also be generated using the basis function derived from 2TCM [28], but the results may be inaccurate and are still computationally challenging [18]. Alternative methods were recently proposed for computing all rate constants of 2TCM within a reasonable time and improving robustness to measurement noise, but they require validation in other reversible radioligands [18, 29].

In this context, the present study found that Logan and MA1 were valid alternatives to 2TCM at the region level and suitable for generating parametric images. Logan and MA1 use a linear estimator, thus enabling the rapid estimation of macro-parameters, such as V_T [30, 31]. Moreover, these graphical approaches do not require a priori knowledge of compartmental configurations. One drawback of the graphical analyses is that they underestimate V_T due to noise in the PET data, especially for Logan [32]. In our study, however, the bias was both small (up to about 8%) and similar between region-based and voxel-based analyses.

Interestingly, SA similarly uses a linear estimator and does not depend on specific compartment model configuration, thus making it both computationally rapid and widely applicable to PET radioligand research. However, in the present study, SA overestimated V_T by about 16%. SA is well known to be sensitive to noise in the data [33], with a tendency to overestimate macro-parameters like V_T, especially for shorter acquisition times.

Reliable quantification of [¹⁸F]LSN3316612 requires arterial blood sampling. To eliminate the need for this invasive and burdensome procedure, the present study examined use of a pseudo-reference region and also measured simultaneous arterial and venous concentrations in four participants. Because pseudo-reference regions are relatively unaffected by the disease [32, 33], this approach can only be properly validated using groups of both patients and healthy controls [34]. However, given the very poor correlation with V_T already present in healthy participants, a pseudo-reference region may not be a viable option for [¹⁸F]LSN3316612. Not only was a suitable arteriovenous equilibrium never reached over 180 min, but considerable inter-subject variability was also observed. However, this lack of suitable equilibrium between the artery and vein is common for radioligands [35]. Even if an average equilibrium is obtained at a late timepoint, considerable inter-subject variability will result in V_T estimation errors in one or more participants. Given the often limited number of participants in PET studies, the presence of these outliers could easily compromise results.

Finally, and as previously reported [17], V_T was generally higher in the second scan compared to the first, regardless of the method used. If this were purely due to chance, the probability of obtaining nine higher retest scans from 10 subjects would be smaller than 1%. Curiously, Coughlin and colleagues similarly found a systematic higher V_T value in retest scans compared to test scans when scanning six subjects with the TSPO radioligand [¹¹C]DPA-713 [36]. In their study—and in contrast to ours—all retest scans were acquired on the same day as the test scan. The authors speculated that these findings might be due to factors such as hormone-mediated phasic changes in TSPO expression, tonic changes due to stress/anxiety related to the procedure, or alterations in blood cholesterol levels due to food intake between the scans. As yet unidentified physiological variables may also explain our results.

Logan plot and MA1 showed the smallest V_T bias when compared with 2TCM at the region and voxel levels and generated good-quality parametric images. Therefore, both models were better than SA or IRF for quantifying [¹⁸F]LSN3316612 uptake in human brain.