Head-to-head comparison of leading blood tests for Alzheimer’s disease pathology

2.1 Participants

The over-arching goal of this FNIH Biomarker Consortium effort is to evaluate the longitudinal trajectories of p-tau217, p-tau181 and Aβ42/Aβ40 in relation to amyloid PET. Therefore, ADNI participants were selected for inclusion who had plasma samples collected within 6 months of an amyloid PET scan for three distinct timepoints. For individuals with more than three time points, plasma samples were selected that represented the earliest, latest, and an intermediate time point. Based on these selection criteria, 393 ADNI participants had at least six plasma aliquots (1,179 samples total) that underwent analysis with the C2N Diagnostics PrecivityAD2 (C2N), Fujirebio Diagnostics Lumipulse (Fujirebio), ALZpath Quanterix (ALZpath), Janssen LucentAD Quanterix (Janssen), and Roche Diagnostics NeuroToolKit (Roche) assays. A seventh aliquot was available for 355 participants (1,063 participant time points, two participants had a seventh plasma aliquot for only two time points), and this aliquot was used for the remaining Quanterix Neurology 4-Plex (Quanterix) assays, which were given a lower priority due to lower accuracy reported by prior studies [9, 10].

For this current head-to-head study, samples were considered for inclusion if 1) no data for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays were missing (e.g., due to quality control failures or low volume) and if clinical data was available within one year of the sample collection. For samples that met these criteria, the most recently collected sample was chosen from each individual in order to increase the availability of associated tau PET data. One individual was excluded due to missing data from each of the three plasma samples, yielding 392 individuals in this study. Quanterix Neurology 4-Plex data was available for 342 of these individuals.

The current analyses utilized data from the ADNI database (adni.loni.usc.edu), which studies memory and aging in older adults, including both cognitively unimpaired and cognitively impaired individuals. ADNI was launched in 2003 as a public-private partnership led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). Written informed consent was obtained from each participant or their legally authorized representative. Race and sex were self-identified.

2.2 Diagnostic and CDR variables

Participants underwent a clinical assessment within one year of blood collection that included a detailed interview of a collateral source, a neurological examination of the participant, the Clinical Dementia Rating® (CDR®) and the Clinical Dementia Rating Sum of Boxes (CDR-SB) [16]. Individuals with a CDR of 0 were categorized as “cognitively unimpaired.” Individuals with a CDR>0 were categorized as “cognitively impaired;” this group includes individuals with mild cognitive impairment and dementia.

2.3 Plasma biomarker assays

The FNIH Biomarker Consortium project team selected assays based on recently published [17–24] as well as currently unpublished data presented by diagnostics companies. This included information on the analytical performance of each assay, such as the measurement range, dilutional linearity, and repeatability. Ultimately, four Aβ42/Aβ40, two p-tau181, four p-tau217, two GFAP, and two NfL assays were selected for this project.

Plasma was collected, processed, and stored according to ADNI protocols. Sample aliquots were pulled from the ADNI biorepository and shipped to four laboratories for analysis on dry ice. Pooled plasma controls with known higher or lower plasma p-tau217 values as measured by the ALZpath Quanterix assay were supplied by the University of Gothenburg to be run with each lot of samples. All laboratories were blinded to information about the participants.

The C2N PrecivityAD2 liquid chromatography tandem mass spectrometry (LC-MS) based assays for Aβ42, Aβ40, p-tau217, and non-phosphorylated tau217 (np-tau217) were run in singlicate at the C2N Diagnostics commercial laboratory in St. Louis, Missouri, USA. The %p-tau217 measure was calculated as the p-tau217 concentration divided by the np-tau217 concentration times 100, which is also described as the percent phosphorylation occupancy [20, 25]. The Fujirebio Lumipulse assays for Aβ42, Aβ40, and p-tau217 were run in singlicate with commercially available kits on a Fujirebio Lumipulse G1200 analyzer at the Indiana University National Centralized Repository for Alzheimer’s Disease and Related Dementias Biomarker Assay Laboratory (NCRAD-BAL). The Roche NeuroToolKit assays for Aβ42, Aβ40, p-tau181, NfL, and GFAP were run in singlicate on a Cobas® e 801 analyzer (Roche Diagnostics International Ltd, Rotkreuz, Switzerland) at the University of Gothenburg. The ALZpath Quanterix and Janssen LucentAD Quanterix assays for p-tau217, as well as the Quanterix Neurology 4-Plex E (N4PE) assays for Aβ42, Aβ40, p-tau181 (v2.1), NfL, and GFAP, were run in duplicate on a Quanterix Simoa-HD-X analyzer at the Quanterix Accelerator Laboratory. Further details are included in the study methodology report, which may be accessed from the ADNI database (adni.loni.usc.edu).

2.4 CSF collection and analysis and APOE genotyping

Data on CSF biomarkers and APOE genotype were obtained from ADNI. CSF Aβ42, total tau, and p-tau181 were measured using the Elecsys® β-amyloid (1–42) CSF, Elecsys Total-Tau CSF, and Elecsys Phospho-Tau (181P) CSF immunoassays, respectively, on a Cobas e 601 analyzer (Roche Diagnostics International Ltd, Rotkreuz, Switzerland). APOE genotyping was performed as part of the ADNI protocol.

2.5 Amyloid PET, tau PET and MRI variables

Amyloid PET imaging was conducted at each ADNI site following standardized protocols for florbetapir and florbetaben [26]. A global standardized uptake value ratio (SUVR) was estimated across cortical summary regions, specifically the frontal, cingulate, parietal, and lateral temporal cortices, as defined by FreeSurfer v7.1. This ratio was normalized to the whole cerebellum. To estimate the standardized global cortical Aβ burden in Centiloid units, transformations provided by the ADNI PET Core were applied to florbetaben and florbetapir. The threshold for amyloid PET positivity was set at >20 Centiloids, which corresponds to the recommended thresholds for florbetapir (whole cerebellum-normalized SUVR of 1.11 [27]) and florbetaben (whole cerebellum-normalized SUVR of 1.08 [28]).

Tau PET imaging was conducted at each ADNI site following standardized protocols for flortaucipir. The four 5-minute flortaucipir PET scans were corrected for motion, averaged, smoothed to a uniform resolution of 6 mm, and co-registered to the individual participant’s accompanying T1-weighted MRI scan. Using the CenTauR standardized masks [29], SUVRs were estimated for regions of interest (ROIs) as defined by FreeSurfer v7.1. The mesial temporal meta-ROI included regions with early tau PET signal (entorhinal, parahippocampus and amygdala), and the temporal-parietal meta-ROI included regions with both early and later tau PET signal (the entorhinal, parahippocampus, amygdala, fusiform, inferior and middle temporal gyri). These SUVRs were normalized to the inferior cerebellar gray matter reference region, as defined by the SUIT template [30]. The derivation of cut-offs for early and late tau PET positivity are described in Appendix A.

Structural brain MRI data included a 3D MP-RAGE or IR-SPGR T1-weighted 3T MRI with sagittal slices. The full cortex of each subject brain was parcellated into 41 bilateral ROIs using a volumetric Desikan-Killiany-Tourville atlas, using FreeSurfer v5.1 for ADNI-GO/2 data and v6.0 for ADNI-3 data. A composite meta-ROI cortical thickness measurement was calculated that included the entorhinal, fusiform, parahippocampal, mid-temporal, inferior-temporal, and angular gyrus [31]. To address differences in image acquisition protocol and image processing, the meta-ROI cortical thickness measure was harmonized using the ComBat-GAM method [32]. Harmonization parameters specific to each batch were estimated using data from the cognitively unimpaired individuals, while considering normal variance due to age, and sex. The age effect was modeled as a non-linear effect over the study age range. The estimated harmonization parameters were then applied to the full dataset. Finally, normal aging effects on the meta-ROI cortical thickness (harmonized) was estimated within cognitively unimpaired and Aβ-negative individuals and regressed out from the corresponding measures from all individuals (using age and quadratic age as predictors). The steps for ComBAT-GAM harmonization and adjusting for normal effects of age utilized all available ADNI data for unbiased modeling. The derivation of cut-offs for cortical thickness abnormality are described in Appendix A.

2.6 Statistical methods

The specific ADNI datasets used by this study are listed and the annotated code used for study analyses is provided in full so that other investigators can easily replicate and extended our analyses (see Appendix B). For participant characteristics tables, the significance of differences by amyloid PET status were evaluated with Wilcoxon ranked sum tests for continuous variables and Chi-Square or Fisher exact tests for categorical variables. The percent difference by amyloid PET status in the median plasma biomarker level was calculated as the difference between the median of the amyloid PET positive and amyloid PET negative groups divided by the median for the amyloid PET negative group times 100.

Unadjusted logistic regression models for a binary outcome (amyloid PET status, tau PET status, cortical thickness status, or cognitive impairment status [CDR=0 or CDR>0]) were used to assess the classification accuracy of plasma biomarker analytes separately or combined within a company/assay panel. Brier scores were used to quantify the prediction accuracy of the logistic regression models [33]. Logistic regression models adjusted for the covariates of age, sex, and APOE genotype were also examined. Differences between receiver operating characteristic areas under the curves (AUCs) were evaluated using DeLong tests [34]. When multiple AUCs were compared, the significance was adjusted using the Benjamini-Hochberg procedure [35]. Cut-off values for separate analytes that yielded the highest combined sensitivity and specificity (Youden Index) for an outcome were found. For each cut-off, the following metrics were calculated: sensitivity, specificity, overall accuracy, positive predictive value, and negative predictive value.

Spearman correlations were used to evaluate the continuous relationships of plasma biomarkers with amyloid PET Centiloid, tau PET measures, cortical thickness measure, and dementia severity as measured by the CDR-SB. For partial Spearman correlations, analyses included covariates of age, sex and APOE genotype. Comparisons between Spearman correlations were performed by bootstrapping. When multiple measures were compared, the significance was adjusted using the Benjamini-Hochberg procedure [35].

Participants

The full cohort included data from 392 individuals with a median age of 78.1 years (interquartile range 72.9-83.3 years); 193 (49.2%) were female; 132 (33.7%) carried an APOE ε4 allele; 366 (93.4%) self-identified as their race as white; 191 (48.7%) were amyloid PET positive; and 192 (49.0%) were cognitively impaired (CDR>0), (Table 1). Of the participants with cognitive impairment, 148 had mild cognitive impairment or very mild dementia (CDR 0.5) and 44 had dementia (CDR 1 or higher), (see Tables S1-S2 for characteristics of the cognitively impaired and unimpaired sub-cohorts). Within one year of plasma collection, all individuals in the full cohort had amyloid PET, 188 had tau PET, 266 had structural brain MRI, and 122 had CSF collection with Elecsys data for CSF p-tau181 and Aβ42 (see Tables S3-5 for characteristics of the tau PET, brain MRI, and CSF sub-cohorts).

Sensitivity and variance of assays

Of the 392 samples included, 110 samples had p-tau217 values that were below the level of detection for the C2N PrecivityAD2 assay and were assigned values of 0.65 pg/ml (half the value of the limit of detection). There were 12 samples with p-tau217 values that were below the level of detection for the Fujirebio Lumipulse assay and were assigned values of 0.030 pg/ml. Analyte values were within range for the other assays.

The variance in measurements was assessed using aliquots of two pooled control plasma samples with known higher or lower p-tau217 values that were run with each batch of samples. For most assays, the coefficient of variance (CV) ranged from 3-10% (see Appendix C). The exception was for Fujirebio Lumipulse p-tau217, which had CVs of 11% and 34% for the higher and lower controls, respectively. The low control was near the lower range of the Fujirebio Lumipulse p-tau217 assay, which is an explanation for the higher imprecision.

Correlations between plasma biomarkers

Correlations between plasma biomarker levels across platforms were evaluated with samples from the 392 individuals (Figure 1). C2N PrecivityAD2 %p-tau217 and C2N PrecivityAD2 p-tau217 values were highly correlated (unadjusted Spearman ρ=0.94 [0.93-0.95]); notably, the numerator in %p-tau217 is p-tau217, so a high correlation was expected. ALZpath Quanterix and Janssen LucentAD Quanterix p-tau217 concentrations were also highly correlated (ρ=0.94 [0.92-0.95]). Compared to C2N PrecivityAD2 p-tau217, C2N PrecivityAD2 %p-tau217 had lower correlations with p-tau217 concentrations as measured by the three immunoassays: ALZpath Quanterix, ρ=0.89 versus 0.81, p=0.0002; Fujirebio Lumipulse, ρ=0.88 versus 0.81, p=0.001; Janssen LucentAD Quanterix ρ=0.88 versus 0.80, p=0.0002. C2N PrecivityAD2 %p-tau217 had much lower correlations with Roche NeuroToolKit p-tau181 (ρ=0.69 [0.63-0.74]) and Quanterix Neurology 4-Plex p-tau181 (ρ=0.54 [0.45-0.62]). The four Aβ42/Aβ40 measures had pairwise correlations that ranged from 0.63 to 0.43; the Fujirebio Lumipulse and Quanterix Neurology 4-Plex had the highest correlation (ρ=0.63 [0.55-0.70]). Roche NeuroToolKit and Quanterix Neurology 4-Plex concentrations of GFAP and NfL were highly correlated (ρ=0.94 [0.91-0.97] and ρ=0.96 [0.94-0.97], respectively).

• Download figure
• Open in new tab

Figure 1. Correlation matrix of plasma biomarker measures in the full cohort.

All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. The unadjusted Spearman correlation of plasma biomarker measures with one another is shown. Redder colors are used for higher absolute correlations and bluer colors are used for lower absolute correlations. Black boxes represent identity.

Differences in plasma biomarkers by amyloid PET status

Differences in plasma biomarker levels associated with positive amyloid PET status were examined. Because plasma biomarker values tend to be highly skewed and therefore mean levels may overestimate the degree of abnormality for the typical amyloid PET positive individual, percent differences were calculated for the median plasma biomarker levels (Table 2). Compared to the amyloid PET negative group, median plasma p-tau217 was higher in the amyloid PET positive group by 255% (Fujirebio Lumipulse), 245% (C2N PrecivityAD2 %p-tau217), 228% (C2N PrecivityAD2 p-tau217), 170% (ALZpath Quanterix) and 139% (Janssen LucentAD Quanterix). The percent difference in median p-tau181 was 66% (Roche NeuroToolKit) and 55% (Quanterix Neurology 4-Plex). Differences in median plasma Aβ42/Aβ40 levels were much smaller in magnitude, ranging from 13% to 10% lower in the amyloid PET positive group, but were still highly significant due to low variance in values. Median plasma GFAP and NfL were higher in the amyloid PET positive group (64 to 61% for GFAP, 28% to 27% for NfL).

Classification of amyloid PET status by plasma biomarkers

The classification accuracies of individual plasma biomarkers for amyloid PET status as defined by Centiloid >20 or ≤20 were evaluated using logistic regression models in the full cohort (Figure 2 and Figure 3). With the expectation that companies may provide a panel of plasma biomarkers, models that combined plasma biomarkers from the same company were also examined. The highest classification accuracies of amyloid PET status were achieved by C2N PrecivityAD2 and Fuijirebio Lumipulse models that included measures of both p-tau217 and Aβ42/Aβ40 (Table 3). However, the C2N PrecivityAD2 %p-tau217 and Aβ42/Aβ40 combined model (AUC 0.929 [0.902-0.956]) was not superior to the model with %p-tau217 alone (AUC 0.927 [0.900-0.955]), p=0.54. For C2N PrecivityAD2 %p-tau217, a cut-off of 4.06% yielded a positive percent agreement (PPA) of 88.5%, a negative percent agreement (NPA) of 87.1%, an overall percent accuracy of 87.8%, a positive predictive value (PPV) of 86.7%, and a negative predictive value (NPV) of 88.8% for amyloid PET status (Table 4). The Fujirebio Lumipulse p-tau217 and Aβ42/Aβ40 combined model (AUC 0.911 [0.882-0.940]) did not have significantly different performance from the C2N PrecivityAD2 %p-tau217 and Aβ42/Aβ40 combined model (p=0.16), or the model with Fujirebio Lumipulse p-tau217 alone (AUC 0.896 [0.864-0.928], p=0.12), (Table 3). For Fujirebio Lumipulse p-tau217, a cut-off of 0.158 yielded a PPA of 84.3%, a NPA of 82.1%, an accuracy of 83.2%, a PPV of 81.7%, and a NPV of 84.6% (Table 4).

• Download figure
• Open in new tab

Figure 2. Receiver operating characteristics area under the curve plots for amyloid PET status in the full cohort.

All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. The best performing plasma biomarker measure or combination of measures from each company is shown for unadjusted models of amyloid PET > or ≤20 Centiloids. The solid black line represents a model of covariates alone. The dotted black line represents chance classification.

• Download figure
• Open in new tab

Figure 3. Classification accuracies of individual and combined plasma biomarker measures for key outcomes in the full cohort.

All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. The receiver operating characteristics area under the curve (AUC) point estimate (midpoint) and 95% confidence intervals are shown for classification of amyloid PET status (> or ≤20 Centiloids), early tau PET status, cortical thickness status, and cognitive impairment status (CDR>0 or =0) by individual or combined plasma biomarker measures. The dashed vertical reference lines represent the AUCs for models including only covariates (age, sex, APOE genotype) as predictors. Please see methods and Appendix A for definitions of early tau PET status and cortical thickness status.

Compared to the C2N PrecivityAD2 combined model, the ALZpath Quanterix p-tau217 and Janssen LucentAD Quanterix p-tau217 models had significantly lower performance (ALZpath Quanterix AUC 0.885 [0.851-0.920]; Janssen LucentAD Quanterix AUC 0.882 [0.848-0.916]), p=0.0006 for both versus the C2N PrecivityAD2 combined model (Table 3). For the Roche NeuroToolKit and Quanterix Neurology 4-Plex assays, models including p-tau181, Aβ42/Aβ40, NfL, and GFAP were most accurate (Roche NeuroToolKit AUC 0.873 [0.838-0.909]; Quanterix Neurology 4-Plex AUC 0.808 [0.763-0.854]), but were not significantly better than models including only p-tau181 and Aβ42/Aβ40 (Roche NeuroToolKit AUC 0.871 [0.834-0.907], p=0.67 for the two compared to the four analyte model; Quanterix Neurology 4-Plex AUC 0.782 [0.733-0.830], p=0.083 for the two compared to the four analyte model), (Table 3).

In a sub-cohort of 122 individuals with CSF data available, the classification accuracies of the plasma biomarker models for amyloid PET status were directly compared to CSF Elecsys p-tau181/Aβ42. In this smaller sample, the classification accuracy of the CSF Elecsys p-tau181/Aβ42 model (AUC 0.915 [0.864-0.967]) was not significantly different from the classification accuracies of the top performing plasma models for all six companies: C2N PrecivityAD2 %p-tau217 + Aβ42/Aβ40 (0.907 [0.851-0.963], p=0.67); Fujirebio Lumipulse p-tau217 + Aβ42/Aβ40 (AUC 0.897 [0.840-0.955], p=0.58); ALZpath Quanterix p-tau217 (AUC 0.883 [0.822-0.945], p=0.22); Janssen LucentAD Quanterix p-tau217 (AUC 0.858 [0.792-0.924], p=0.12); Roche NeuroToolKit p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.834 [0.760-0.908], p=0.098); Quanterix Neurology 4-Plex p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.822 [0.740-0.905], p=0.13), (Tables S6 and S7).

The classification accuracies of plasma biomarker models of amyloid PET status as defined by Centiloid >20 were evaluated in the sub-cohort of 192 individuals with cognitive impairment (CDR>0), which is relevant to clinical practice in which biomarker testing is limited to patients with cognitive impairment (Table S8). Classification accuracies were slightly higher than in the full cohort: the C2N PrecivityAD2 %p-tau217 and Aβ42/Aβ40 combined model had an AUC of 0.960 [0.934-0.986] and the Fujirebio Lumipulse p-tau217 and Aβ42/Aβ40 combined model had an AUC of 0.952 [0.925-0.980]. The PPA, NPA, and overall percent accuracies for the C2N PrecivityAD2 and Fujirebio Lumipulse p-tau217 measures were approximately 90% (Table S9). Conversely, classification accuracies were lower in the sub-cohort of 200 individuals who were cognitively unimpaired (CDR=0), which is relevant to clinical trials of preventative treatments (Tables S10 and S11). The C2N PrecivityAD2 combined model had an AUC of 0.890 [0.839-0.942] and was statistically superior to the best performing models from other companies including the Fujirebio Lumipulse combined model, which had an AUC of 0.840 [0.782-0.897], p=0.024. If a more abnormal cut-off value were chosen for amyloid PET positivity (Centiloid >37), the classification accuracies were similar or slightly higher for p-tau217 and p-tau181 models and were similar or slightly lower for Aβ42/Aβ40 models in the full cohort (Tables S12 and S13), cognitively impaired sub-cohort (Tables S14 and S15), and cognitively unimpaired sub-cohort (Tables S16 and S17).

Cut-offs for amyloid status that were previously published or reported by the company were examined in the full cohort (Table S18). Notably, these cut-offs were developed in different cohorts and may have been determined using different standards for amyloid positivity than in the current study (amyloid PET Centiloid >20). For C2N PrecivityAD2 %p-tau217, the cut-off of 4.2% for distinguishing amyloid status that was previously reported [20] was similar to the cut-off determined in the current study (4.06%), with similar accuracy for classifying amyloid PET status (86.7% for Meyer cut-off, 87.8% for current study cut-off). For ALZpath Quanterix p-tau217, the cut-off of 0.42 pg/ml that was previously reported [21] was also similar to the cut-off determined in the current study (0.444 pg/ml), with similar accuracy (82.4% for Ashton cut-off, 83.7% for current study cut-off).

Correlations between amyloid PET Centiloid and plasma biomarkers

Correlations between amyloid PET Centiloid and individual plasma biomarkers were examined with Spearman correlations because of expected non-linearity in these relationships (Figure 4). In the full cohort, amyloid PET Centiloid had the strongest correlation with the C2N PrecivityAD2 %p-tau217 (unadjusted Spearman ρ=0.771 [95% confidence intervals 0.735 to 0.804]), which was superior to correlations of analytes from other companies including with Fujirebio Lumipulse p-tau217 (ρ=0.715 [0.663 to 0.760], p=0.011), ALZpath Quanterix p-tau217 (ρ=0.702 [0.649 to 0.748], p=0.0040), and Janssen LucentAD Quanterix p-tau217 (ρ=0.699 [0.647 to 0.748], p=0.0029) (Figure 5, Table 5). Correlations between amyloid PET Centiloid and p-tau181 were much lower as compared to C2N PrecivityAD2 %p-tau217: Roche NeuroToolKit p-tau181 (ρ=0.547 [0.472 to 0.618], p<0.0001) and Quanterix Neurology 4-Plex p-tau181 (ρ=0.438 (0.351 to 0.512), p<0.0001). Correlations between amyloid PET Centiloid and Aβ42/Aβ40 were also lower, ranging from a correlation of ρ=-0.531 (−0.595 to −0.455) for the Roche NeuroToolKit to ρ=-0.404 (−0.486 to −0.316) for the Quanterix Neurology 4-Plex.

• Download figure
• Open in new tab

Figure 4. Scatterplots of plasma biomarker measures and amyloid PET Centiloid.

The best performing plasma biomarker measure from each company is shown. All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. Point colors represent the following: blue, cognitively unimpaired and amyloid PET negative; green, cognitively impaired and amyloid PET negative; orange, cognitively unimpaired and amyloid PET positive; red, cognitively impaired and amyloid PET positive.

• Download figure
• Open in new tab

Figure 5. Correlations between individual plasma biomarker measures and key outcomes in the full cohort.

All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. The unadjusted Spearman correlation point estimate (midpoint) and 95% confidence intervals with plasma biomarker measures is shown for amyloid PET Centiloid, the early tau PET measure, cortical thickness, and dementia severity by the Clinical Dementia Rating Sum of Boxes. The dashed vertical reference lines are arbitrary and presented to aide in visual interpretation.

In the cognitively impaired sub-cohort, the correlation of C2N PrecivityAD2 %p-tau217 with amyloid PET Centiloid was ρ=0.792 [0.744 to 0.832], which was not significantly different from Fujirebio Lumipulse p-tau217 (ρ=0.792 [0.746 to 0.829], p=0.97), ALZpath Quanterix p-tau217 (ρ=0.768 [0.714 to 0.807], p=0.35), or Janssen LucentAD Quanterix p-tau217 (ρ=0.744 [0.680 to 0.792], 0.087), (Figure 6, Table S19). Correlations between amyloid PET Centiloid and p-tau181 remained lower as compared to C2N PrecivityAD2 %p-tau217: Roche NeuroToolKit p-tau181 (ρ=0.613 [0.519 to 0.685], p<0.0001) and Quanterix Neurology 4-Plex p-tau181 (ρ=0.512 [0.401 to 0.606], p<0.0001). Correlations between amyloid PET Centiloid and Aβ42/Aβ40 were also low, ranging from a correlation of ρ=-0.521 (−0.636 to −0.412) for the Roche NeuroToolKit to ρ=-0.384 (−0.506 to −0.248) for the Quanterix Neurology 4-Plex.

• Download figure
• Open in new tab

Figure 6. Correlations between individual plasma biomarker measures and amyloid PET Centiloid in sub-cohorts.

All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. The unadjusted Spearman correlation point estimate (midpoint) and 95% confidence intervals are shown for amyloid PET Centiloid, the early tau PET measure, cortical thickness, and dementia severity by the Clinical Dementia Rating Sum of Boxes. The dashed vertical reference lines are arbitrary and presented to aide in visual interpretation.

In the cognitively unimpaired sub-cohort, C2N PrecivityAD2 %p-tau217 had a significantly higher correlation with amyloid PET Centiloid (ρ=0.679 [0.592 to 0.756]) compared to the other p-tau217 measures: Janssen LucentAD Quanterix (ρ=0.580 [0.477 to 0.683), p=0.030), Fujirebio Lumipulse (ρ=0.563 [0.447 to 0.661], p=0.019) and ALZpath Quanterix (ρ=0.546 [0.422 to 0.658], p=0.0078), (Figure 6, Table S20). Correlations between amyloid PET Centiloid and p-tau181 dropped much lower in this sub-cohort as compared to C2N PrecivityAD2 %p-tau217: Roche NeuroToolKit p-tau181 (ρ=0.389 [0.253 to 0.525], p<0.0001) and Quanterix Neurology 4-Plex p-tau181 (ρ=0.293 [0.132 to 0.440], p<0.0001). In contrast, correlations between amyloid PET Centiloid and Aβ42/Aβ40 remained relatively similar, ranging from a correlation of ρ=-0.484 (−0.584 to −0.371) for the Roche NeuroToolKit to ρ=-0.352 (−0.470 to −0.221) for the Quanterix Neurology 4-Plex.

In a sub-cohort with amyloid PET >20 Centiloids, Fujirebio Lumipulse p-tau217 had a correlation of ρ=0.518 [0.401 to 0.615] with amyloid PET Centiloid, which was not statistically different than correlations for the C2N PrecivityAD2 %p-tau217 (ρ=0.494 [0.366 to 0.593], p=0.70), Janssen LucentAD Quanterix p-tau217 (ρ=0.509 [0.393 to 0.606], p=0.83), or ALZpath Quanterix p-tau217 (ρ=0.496 [0.384 to 0.596], p=0.62) (Figure 6, Table S21). In a sub-cohort with amyloid PET ≤20 Centiloids, very low levels of amyloid burden had the highest correlations with Roche NeuroToolKit Aβ42/Aβ40 (ρ=-0.349 [-0.472 to −0.211]), Fujirebio Lumipulse Aβ42/Aβ40 (ρ=-0.248 [-0.372 to −0.109]), and C2N PrecivityAD2 %p-tau217 (ρ=0.245 [0.102 to 0.379]), (Figure 6, Table S22).

Classification of tau PET status by plasma biomarkers

The classification accuracies of individual and combined plasma biomarkers for early tau PET status as defined by flortaucipir signal in the mesial temporal meta-ROI (entorhinal, parahippocampus and amygdala) were evaluated using logistic regression models in the full cohort (Figure 3, Tables S23 and S24). The highest performing models from each company included either p-tau217 or p-tau181 and were not significantly different from one another: C2N PrecivityAD2 %p-tau217 (AUC 0.888 [0.836-0.940]), Fujirebio Lumipulse p-tau217 (AUC 0.859 [0.802-0.916]), ALZpath Quanterix p-tau217 (AUC 0.850 [0.788-0.912]), Janssen LucentAD Quanterix p-tau217 (AUC 0.846 [0.780-0.912]), Roche NeuroToolKit p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.874 [0.823-0.925]), and Quanterix Neurology 4-Plex p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.874 [0.819-0.930]). Notably, the Roche NeuroToolKit p-tau181 (0.823 [0.755-0.891]) and Quanterix Neurology 4-Plex p-tau181 (0.803 [0.724-0.883]) models were not significantly different from the models combining all four analytes (p=0.095 for difference in Roche NeuroToolKit 2 versus 4 analyte models; p=0.10 for difference in Quanterix Neurology 4-Plex 2 versus 4 analyte models).

In the cognitively impaired sub-cohort, classification accuracies for early tau PET status were slightly higher, including for the C2N PrecivityAD2 %p-tau217 (AUC 0.921 [0.857-0.985]) and Janssen LucentAD Quanterix p-tau217 (AUC 0.921 [0.855-0.986]) models, and again there were no significant differences between the top performing models from each company (Tables S25 and S26). In the cognitively unimpaired sub-cohort, classification accuracies were lower for the C2N PrecivityAD2 %p-tau217 (AUC 0.805 [0.707-0.903]) and Janssen LucentAD Quanterix p-tau217 (AUC 0.677 [0.522-0.833]) models (Table S27). In the sub-cohort with amyloid PET >20 Centiloids, classification accuracies for models of early tau PET status were lower than in the full cohort, including for the C2N PrecivityAD2 %p-tau217 (AUC 0.792 [0.694-0.891]) and Janssen LucentAD Quanterix p-tau217 (AUC 0.744 [0.638-0.850]) models, (Tables S28 and S29).

The classification accuracies of individual and combined plasma biomarkers for late tau PET status as defined by flortaucipir signal in the temporal-parietal meta-ROI (entorhinal, parahippocampus, amygdala, fusiform, inferior and middle temporal gyri) were evaluated using logistic regression models in the full cohort (Table 30). The highest performing models were C2N PrecivityAD2 %p-tau217 + Aβ42/Aβ40 (AUC 0.901 [0.823-0.980]), Janssen LucentAD Quanterix p-tau217 (AUC 0.879 [0.796-0.961]), and Fujirebio Lumipulse p-tau217 (AUC 0.878 [0.797-0.958]). As with early tau PET, classification accuracies for late tau PET status were slightly higher in the cognitively impaired sub-cohort (Table 31), and lower in the cognitively unimpaired sub-cohort (Table S32) and sub-cohort with amyloid PET > 20 Centiloids (Table 33).

Correlations between tau PET and plasma biomarkers

Correlations between the early tau PET measure and individual plasma biomarkers were examined with Spearman correlations because of expected non-linearity in these relationships (Figure 7). In the full cohort, the early tau PET measure had the highest correlation with C2N PrecivityAD2 p-tau217 (ρ=0.579 [0.468 to 0.666]), which was not different from correlations with Fujirebio Lumipulse p-tau217 (ρ=0.526 [0.416 to 0.620], p=0.11), but was stronger than correlations with Janssen LucentAD Quanterix p-tau217 (ρ=0.499 [0.384 to 0.594], p=0.022), ALZpath Quanterix p-tau217 (ρ=0.490 [0.371 to 0.586], p=0.011), Roche NeuroToolKit p-tau181 (ρ=0.459 [0.341 to 0.568], p=0.0061) and Quanterix Neurology 4-Plex p-tau181 (0.411 [0.266 to 0.535], p=0.013), (Figure 5, Table S34).

• Download figure
• Open in new tab

Figure 7. Scatterplots of plasma biomarker measures and the early tau PET measure.

The best performing plasma biomarker measure from each company is shown. All individuals (n=392) had plasma biomarker measures for the C2N PrecivityAD2, Fujirebio Lumipulse, ALZpath Quanterix, Janssen LucentAD Quanterix, and Roche NeuroToolKit assays. A sub-cohort (n=342) additionally had plasma biomarker measures for the Quanterix Neurology 4-Plex assays. Point colors represent the following: blue, cognitively unimpaired and amyloid PET negative; green, cognitively impaired and amyloid PET negative; orange, cognitively unimpaired and amyloid PET positive; red, cognitively impaired and amyloid PET positive.

In the cognitively impaired sub-cohort, correlations with the early tau PET measure were stronger, including for C2N PrecivityAD2 p-tau217 (ρ=0.755 [0.626 to 0.833]) and Fujirebio Lumipulse p-tau217 (ρ=0.726 [0.591 to 0.818]), (Table S35). In the cognitively unimpaired cohort, correlations with the early tau PET measure were much weaker including for C2N PrecivityAD2 p-tau217 (ρ=0.401 [0.249 to 0.548]) and Fujirebio Lumipulse p-tau217 (ρ=0.324 [0.154 to 0.473]), (Table S36). In the sub-cohort with amyloid PET >20 Centiloids, correlations with the early tau PET measure were lower, but still strong for C2N PrecivityAD2 p-tau217 (ρ=0.600 [0.450 to 0.719]), Fujirebio Lumipulse p-tau217 (ρ=0.543 [0.395 to 0.663]), and Janssen LucentAD Quanterix p-tau217 (ρ=0.545 [0.391 to 0.676]), (Table S37).

Correlations were also examined for the late tau PET measure. In the full cohort, the late tau PET measure had similar patterns of correlations as the early tau PET measure but weaker correlations: the highest correlations were with C2N PrecivityAD2 p-tau217 (ρ=0.377 [0.243 to 0.507]) and Fujirebio Lumipulse p-tau217 (ρ=0.327 [0.187 to 0.456]), (Table S38). Patterns of correlations for the late tau PET measure were also similar but weaker than with the early tau PET measure in the cognitively impaired sub-cohort (Table S39) and the sub-cohort with amyloid PET >20 Centiloids (Table S40). Correlations between the late tau PET measures and plasma biomarkers in the cognitively unimpaired sub-cohort were not significant.

Classification of cortical thickness status by plasma biomarkers

Cortical thickness status was evaluated using logistic regression models in the full cohort (Figure 3, Tables S41 and S42). As with tau PET, the highest performing models from each company included either p-tau217 or p-tau181 and did not vary significantly from one another in classification accuracy: Janssen LucentAD Quanterix p-tau217 (AUC 0.849 [0.765-0.933]), Fujirebio Lumipulse p-tau217 (AUC 0.841 [0.765-0.916]), C2N PrecivityAD2 p-tau217 + Aβ42/Aβ40 (AUC 0.840 [0.765-0.914]), Roche NeuroToolKit p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.838 [0.774-0.902]), ALZpath Quanterix p-tau217 (AUC 0.833 [0.755-0.912]), and Quanterix Neurology 4-Plex p-tau181 + Aβ42/Aβ40 + NfL (AUC 0.783 [0.707-0.860]). In contrast, classification accuracy of cortical thickness status was lower for plasma NfL than covariates alone (AUC 0.758 [0.664-0.852] for covariates; AUC 0.743 [0.663-0.823] for Roche NeuroToolKit NfL; AUC 0.745 [0.662-0.829] for Quanterix Neurology 4-Plex NfL). Classification accuracies were very similar in the cognitively impaired sub-cohort (Tables S43 and S44) and similar but slightly lower in the sub-cohort with amyloid PET >20 Centiloids (Tables S45 and S46), with no significant differences between the top performing models for each company.

Correlations between cortical thickness and plasma biomarkers

Correlations between the cortical thickness measure and individual plasma biomarkers were examined with Spearman correlations. In the full cohort, the cortical thickness measure had the highest correlation with Janssen LucentAD Quanterix p-tau217 (ρ=-0.327 [-0.436 to −0.217]), which was not different from correlations with Roche NeuroToolKit p-tau181 (ρ=-0.275 [-0.386 to −0.165], p=0.10) or C2N PrecivityAD2 p-tau217 (ρ=-0.274 [-0.388 to −0.159], p=0.79), but was higher than correlations with ALZpath Quanterix p-tau217 (ρ=-0.273 [-0.383 to −0.166], p=0.016), Fujirebio Lumipulse p-tau217 (ρ=-0.256 [-0.369 to −0.138], p=0.018), or Quanterix Neurology 4-Plex p-tau181 (ρ=-0.194 [-0.317 to −0.066], p=0.010), (Table S47). Correlations between the cortical thickness measure and NfL were inferior to Janssen LucentAD Quanterix p-tau217 (ρ=-0.198 [-0.314 to −0.085] for Roche NeuroToolKit NfL, p=0.041; ρ=-0.114 (−0.234 to 0.000) for Quanterix Neurology 4-Plex NfL, p=0.0008).

In the cognitively impaired sub-cohort, correlations with cortical thickness measure were higher, including for Janssen LucentAD Quanterix p-tau217 (ρ=-0.425 [-0.568 to −0.277]), Fujirebio Lumipulse p-tau217 (ρ=-0.384 [-0.520 to −0.222], and C2N PrecivityAD2 p-tau217 (ρ=-0.356 [-0.497 to −0.196]), (Figure 5, Table S48). Correlations were even higher in the sub-cohort with amyloid PET >20 Centiloids, with the strongest correlations for Janssen LucentAD Quanterix p-tau217 (ρ=-0.527 [-0.650 to −0.386]), Fujirebio Lumipulse p-tau217 (ρ=-0.477 [-0.608 to −0.334]), and C2N PrecivityAD2 p-tau217 (ρ=-0.458 [-0.585 to −0.317]), (Table S49).

Classification of cognitive status by plasma biomarkers

Cognitive status (CDR>0, cognitively impaired; CDR=0, cognitively normal) was evaluated using logistic regression models in the full cohort (Figure 3, Tables S50 and S51). The classification accuracy of all models was relatively low. The highest performing model was C2N PrecivityAD2 p-tau217 + Aβ42/Aβ40 (AUC 0.680 [0.627-0.733]), but it was not significantly different from the top performing models from the other companies: Janssen LucentAD Quanterix p-tau217 (AUC 0.670 [0.615-0.724], p=0.46); Fujirebio Lumipulse p-tau217 (AUC 0.667 [0.613-0.721], p=0.40); ALZpath Quanterix p-tau217 (AUC 0.664 [0.610-0.718], p=0.40); Roche NeuroToolKit p-tau181 + Aβ42/Aβ40 + NfL (AUC 0.648 [0.593-0.703], p=0.40); and Quanterix Neurology 4-Plex p-tau181 + Aβ42/Aβ40 + GFAP + NfL (AUC 0.635 [0.576-0.695], p=0.40). Classification accuracies were slightly higher in the sub-cohort with amyloid PET >20 Centiloids, but there was still no distinction between the top performing models (Tables S52 and S53).

Correlations between dementia symptoms and plasma biomarkers

Correlations between dementia symptoms as measured by the CDR-SB and individual plasma biomarkers were examined with Spearman correlations. In the full cohort, dementia symptoms had the highest correlation with C2N PrecivityAD2 p-tau217 (ρ=0.414 [0.325 to 0.505]), which was not different from correlations with Fujirebio Lumipulse p-tau217 (ρ=0.403 [0.311 to 0.490], p=0.65), Janssen LucentAD Quanterix p-tau217 (ρ=0.391 [0.295 to 0.482], p=0.34), or ALZpath Quanterix p-tau217 (ρ=0.374 [0.280 to 0.461], p=0.074), but was stronger than correlations with Roche NeuroToolKit p-tau181 (ρ=0.331 [0.237 to 0.415], p=0.0038), Roche NeuroToolKit NfL (ρ=0.261 [0.164 to 0.351], p=0.0022), Quanterix Neurology 4-Plex p-tau181 (ρ=0.234 [0.131 to 0.332], p<0.0001), and Quanterix Neurology 4-Plex NfL (ρ=0.224 [0.118 to 0.321], p=0.0006), (Table S54). In the sub-cohort with amyloid PET >20 Centiloids, correlations were slightly higher and the findings were similar, with stronger correlations between dementia symptoms and p-tau217 as compared to p-tau181, (Table S55).