Modelling the temporal evolution of plasma p-tau217, amyloid PET, tau PET and cognition

Abstract

Associations of Alzheimer’s disease biomarker progression with cognitive decline are important to inform patient prognosis. Of particular interest is how newly available plasma biomarkers evolve relative to cognitive decline. The goals of this work are to measure how much earlier versus later an individual’s progression on plasma and PET Alzheimer’s disease biomarkers is associated with earlier versus later cognitive progression and to estimate the average timeline of progression of these processes in the population.

In this cohort study of 2369 Mayo Clinic Study of Aging (MCSA) and 1591 Alzheimer’s Disease Neuroimaging Initiative (ADNI) participants, we fit non-linear mixed-effects models to estimate how much earlier versus later each individual progresses on plasma phosphorylated tau (p-tau)217, amyloid PET, tau PET and auditory verbal learning test (AVLT) sum of trials relative to the population mean (individual adjustment), the associations of these individual adjustments among biomarker pairs and how covariates affect the timing of biomarker progression. The association of individual adjustments implies mechanistic associations and the amount of variability in cognitive decline accounted for by each biomarker. By applying cut-off points, we also estimated the relative timing that these biomarkers become abnormal in the population.

Associations of individual adjustments were moderate between all biomarkers and AVLT (R = 0.38–0.47) in the MCSA and stronger (R = 0.74–0.81) in ADNI; plasma p-tau217 accounted for 16% of the variability in timing of AVLT decline in the MCSA and 64% in ADNI. APOE ɛ4 carriership was associated with earlier biomarker progression. AVLT became abnormal after the biomarkers up to age 90, after which AVLT was estimated to become abnormal prior to tau biomarkers.

The association of the timing of plasma and PET Alzheimer’s disease biomarker progression with cognitive decline was modest in the MCSA population-based sample and stronger in the Alzheimer’s disease-enriched ADNI cohort. The timing of plasma p-tau217 progression explained a similar degree of variability in AVLT progression as amyloid PET, supporting its utility as a marker of disease progression. The estimated temporal ordering of biomarkers and cognitive abnormality was as anticipated (amyloid, tau, cognition) up to the age of 90, beyond which AVLT was estimated to become abnormal prior to tau biomarkers, likely related to the effects of non-Alzheimer’s disease co-pathologies.

Cogswell et al. show that plasma and PET biomarkers for Alzheimer’s disease provide useful information about when cognitive decline is likely to begin and typically become abnormal before cognitive symptoms appear. However, at advanced ages cognitive decline may occur more quickly due to co-pathologies.

See Pelkmans et al. (https://doi.org/10.1093/brain/awag240) for a scientific commentary on this article.

See Pelkmans et al. (https://doi.org/10.1093/brain/awag240) for a scientific commentary on this article.

Introduction

It is important to understand the longitudinal relationships of Alzheimer’s disease biomarkers with cognition to inform patient prognosis and define targeted intervention timelines. Of particular recent interest in the field is plasma Alzheimer’s disease biomarkers^1-4 and how they correlate with cognitive metrics and compare with PET biomarkers in tracking disease status. Studies of the associations of plasma biomarkers with cognition have primarily evaluated the association of baseline, cross-sectional plasma Alzheimer’s disease biomarker levels with current cognitive status or future cognitive decline. These studies have shown that plasma phosphorylated tau (p-tau) analytes are associated with current cognitive status⁵ and future cognitive decline on the group level for cognitively unimpaired and impaired individuals along the Alzheimer’s disease spectrum.^6-14 However, fewer studies have examined the temporal evolution of plasma biomarkers along with cognitive decline. Some recent studies have shown that longitudinal increases in plasma p-tau217 and p-tau181 are associated with rates of cognitive decline.^15-17 Our study adds to the literature by modelling individual-level plasma and PET Alzheimer’s disease biomarker temporal progression and its relationship to longitudinal cognitive decline.

Relationships of interest that are not clearly understood and we directly address include the degree to which the timing of biomarker change correlates with timing of cognitive decline for an individual and the average timeline of biomarker and cognitive changes. These relationships may be evaluated with temporal modelling techniques, as have been applied to evaluate the temporal evolution of plasma Alzheimer’s disease biomarkers and in relationship to amyloid and tau PET biomarkers.^18-20 These studies have used different approaches to model longitudinal biomarker changes, although all similarly assume that there is a common trajectory for each biomarker.

The approach we have implemented is non-linear mixed effects (NLME) models, which are an extension of our earlier work.^18,21,22 In these models each biomarker is assumed to progress along a common temporal evolution curve that is shifted left or right in time for each individual (referred to as an individual adjustment in years). The outcomes are jointly modelled such that the model output includes correlations for the individual adjustments of outcomes pairs, indicating the degree of variation in timing of progression of one outcome that can be explained by another—for example, how much variation in timing of cognitive decline can be explained by timing of a biomarker’s change. Covariate effects are also expressed in years and can be interpreted as an average shift earlier/later in time for a biomarker. This approach works well for biomarkers that have a consistent baseline (e.g. amyloid PET centiloid = 0). Modelling cognition poses challenges due to variability in the baseline level from multiple factors, such as sex and education level,^23-26 leading to vertical (up/down) shifts unrelated to the timing (left/right shift) of disease progression. In this work we extended prior models to account for this variation in baseline cognitive performance.

The goals of this work are to measure how much earlier versus later an individual’s progression on plasma p-tau217, amyloid PET and tau PET biomarkers is associated with earlier versus later cognitive progression relative to the population mean; estimate the average timeline of progression of these biomarkers and cognition in the population; and determine the effects of common covariates on these processes. We evaluate these relationships in the Mayo Clinic Study of Aging (MCSA) and validate in Alzheimer’s Disease Neuroimaging Initiative (ADNI).

Materials and methods

Participants

This study included participants in the MCSA, a longitudinal population-based study of individuals residing in Olmsted County, Minnesota. We included participants with at least one visit with PET imaging performed in 2009 or after or at least two visits with cognitive scores. Participants were required to have a diagnosis of cognitively unimpaired (CU), mild cognitive impairment (MCI) or Alzheimer's clinical syndrome (AlzCS) dementia and have sex, education and APOE ε4 genotype available. Visits before age 50 were not used due to the very low rate of cognitive change before that age. Clinical diagnoses were determined by an expert panel based on established criteria.^27-29

Standard protocols and patient consents

The study was approved by the Mayo Clinic and Olmsted Medical Center institutional review boards and was performed in accordance with the ethical standards of the Declaration of Helsinki and its later amendments. All participants or a legally authorized representative provided informed written consent.

Plasma processing and metrics

Plasma samples were collected after an overnight fast, centrifuged and 500 µl plasma aliquoted into polypropylene tubes and stored at −80°C; see prior publications for details on plasma handling.³⁰ Subsets of participants had plasma samples analysed via immunoprecipitation followed by liquid chromatography–tandem mass spectrometry (assay V1, Supplementary material, ‘Methods’ section) at C2N Diagnostics (St. Louis, MO, USA) and/or plasma phosphorylated tau at position 217 (p-tau217) immunoassay on the Meso Scale Diagnostics (MSD) platform (Lilly Research Laboratories, IN, USA). The mass spectrometry plasma assays were licensed by Washington University to C2N Diagnostics and included Aβ₄₂, Aβ₄₀, p-tau217 and non-phosphorylated tau at position 217 (np-tau217), which are generally applied as the amyloid-β (Aβ)_42/40 ratio and %p-tau217 (ratio of p-tau217 to np-tau217 × 100). We used the p-tau217 concentration (pg/ml) rather than %p-tau217 as the C2N plasma p-tau217 measure due to lower noise as compared to the ratio in longitudinal analyses. Additionally, the p-tau217 V1 concentrations were converted to the V2 assay equivalent,³¹ as that is the currently available version of the assay and used in the replication data set. See Supplementary material, ‘Methods’ section for additional details.

Amyloid-β and tau PET

Aβ PET was performed with Pittsburgh compound B³² and tau PET with ¹⁸F-flortaucipir (Avid Radiopharmaceuticals) on GE (models Discovery 690XT, Discovery RX and Discovery MI) or Siemens (Biograph Vision 600) scanners. PET was processed using in-house pipelines^33,34 using corresponding T1-weighted MRIs. Harmonization was performed via the methods of Joshi et al.³⁵ The Aβ PET meta-region of interest (ROI) was derived via the voxel number weighted average of the median uptake in each of the prefrontal, orbitofrontal, parietal, temporal, anterior and posterior cingulate and precuneus regions normalized to the cerebellar crus grey matter. Amyloid PET standardized uptake value ratio (SUVR) was converted to centiloids.³⁶ The tau PET temporal meta-ROI was derived via the voxel number weighted average of the median uptake in each of the amygdala, fusiform, middle/inferior temporal, entorhinal and parahippocampal regions normalized by the cerebellar crus grey matter. PET data were not partial volume corrected. All available amyloid and tau PET scans obtained in 2009 and after for qualifying participants were included in the analyses.

Cognitive assessments

Participants underwent neuropsychological exams including a standardized set of cognitive evaluations covering the domains of memory, language, executive and visuospatial function at on average 15-month intervals.²⁹ The primary cognitive outcome for this study was the auditory verbal learning test (AVLT) sum of trials (trials 1–5 total + short delay recall + long delay recall) raw score that ranges from 0 to 105 as each trial has a range of 0–15.^23,37 The AVLT sum of trials is referred to as the AVLT going forward. This cognitive metric was chosen as it is a robust metric available in both the MCSA and ADNI. Sensitivity analyses used a composite global z-score as the outcome. The global z-score was generated using the four-domain z-scores with weights based on the age and sex distribution of the Olmsted County population. All available longitudinal cognitive assessments were used for qualifying participants.

Noise properties of biomarker and cognitive metrics

To inform potential challenges in our longitudinal modelling approach, we first evaluated the reproducibility of the biomarkers and cognitive metrics modelled in this study along with a subset of additional biomarkers for reference in the MCSA. Using a subset of longitudinal biomarker data (first three measurements in participants with at least three serial measurements of that biomarker), we calculated an indicator of the noise-to-signal ratio where lower values imply better reproducibility or stability. The ratio is based on within individual variability (25th, 50th and 75th percentiles of the per-subject residual standard deviation across repeated measures) relative to between-individual differences (interquartile range of individual-level predicted means) or the interquartile range (IQR)-scaled quantiles of per-subject residual standard deviation.

Primary model and individual adjustments

We implemented NLME models, referred to as an accelerated failure time (AFT) model in prior work,^18,21,22 to estimate biomarker and cognitive progression by age. In these models it is assumed that a given biomarker progresses along a similar trajectory for all individuals but is shifted left or right depending on whether the individual progresses earlier or later in age relative to others in the population. Because the NLME framework estimates per-participant horizontal time shifts on shared trajectories, outcomes measured on different visit dates can be fit jointly without requiring same-day measurements. In this work we fit an NLME model with amyloid PET centiloid, tau PET temporal meta-ROI SUVR, plasma p-tau217 concentration (Lilly), plasma p-tau217 concentration (C2N) and AVLT as the outcomes. Sex, APOE ε4 carrier status and education were included as covariates; biomarker trajectories are modelled versus age and therefore age is not listed as a separate covariate effect. See Supplementary material, ‘Methods’ section—NLME for more details. Although available in a subset of the cohort, Aβ_42/40 was not included in the models as its high measurement variability (Fig. 1) did not allow for reliable longitudinal modelling.

Figure 1.

Estimated noise-to-signal ratios for each biomarker and cognitive outcome. Longitudinal variability was assessed by comparing inter-individual versus intra-individual variability in MCSA participants with three serial measurements of each biomarker or cognitive outcome. For each measure, we calculated the percentile of each subject’s residual standard deviation, scaled by the IQR of individual-level predicted means. AVLT = auditory verbal learning test; C2N = C2N Diagnostics; IQR = interquartile range; Lilly = Lilly Research Laboratories; MCSA = Mayo Clinic Study of Aging; SUVR = standardized uptake value ratio.

As in our prior work,^18,21,22 the model included a smooth progression function for each outcome (simplest possible non-linear function), regression coefficients for each covariate and outcome pair and per participant left–right age adjustments for each outcome (individual-level adjustment). We extended models applied in prior work to include a second random effect (an up–down shift) to accommodate baseline variability in cognition based on known factors such as sex and education; we will refer to this as a baseline effect. Covariate effects were as follows: APOE ε4 carrier status was estimated as a left–right effect on each outcome; sex and education estimated as a left–right effect on PET and plasma biomarkers and as a baseline effect on AVLT, as informed by prior work.^23,25,26 The model output included a correlation matrix of the random effects or the correlation (R) of the individual-level adjustment between biomarker pairs. See prior publications^18,21,22 for additional information on the model fits.

Timing of biomarker progression on the population level

We estimated the relative timing that biomarkers and cognition become abnormal in the population by applying established, published cut-off points to the model output. For amyloid PET centiloid, we used an accepted value of 25, which corresponds with 1.52 SUVR from our in-house pipelines.³³ For tau PET SUVR, we used a cut-off point of 1.29, based on a neuropathological correlate.³⁸ For Lilly p-tau-217 concentration, we used a cut-off point of 0.26 pg/ml, previously determined in the MCSA.³⁹ For C2N p-tau217 concentration, we used a Deming regression to convert the published V2%p-tau217 cut-off point of 4.2% to the p-tau217 scale, giving a cut-off point of 2.4 pg/ml (Supplementary material, ‘Methods’ section and Supplementary Fig. 1). For AVLT, cognition was considered abnormal if the AVLT sum of trials raw score was less than or equal to 35; this corresponds to an unadjusted scaled score of <6 in the AVLT normative sample.^23,33,40-43

Sensitivity analyses

We fit an additional NLME model in the MCSA with global z-score as the cognitive outcome in place of the AVLT. Prior to modelling, we subtracted the practice effect according to prior publications, (0.23, 0.07, 0.04) for visits 1 to 2, 2 to 3, 3 to 4.^44,45

Validation cohort

External validation was performed in an ADNI cohort (ADNIGO, ADNI2 and ADNI3). Data used in the preparation of this article were obtained from the ADNI database (adni.loni.usc.edu). The 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 MRI, PET, other biological markers and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early Alzheimer’s disease.

The inclusion criteria for the ADNI validation cohort were comparable to the criteria used for the MCSA sample. We included participants with at least one amyloid PET scan and one AVLT sum of trials score, a diagnosis of CU, MCI due to Alzheimer’s disease, or dementia due to Alzheimer’s disease and age 55 years or older at their earliest biomarker measurement. All available C2N plasma measurements, amyloid PET scans (various tracers converted to centiloid scale), flortaucipir tau PET scans and AVLT scores were used in model fits for the participants meeting inclusion criteria. C2N plasma p-tau217 was performed with assay V2 in ADNI. Flortaucipir tau PET was processed using the same in-house pipeline as the MCSA data,^33,34 while the other data were downloaded directly from LONI.

An NLME model was fit with p-tau217 (C2N), amyloid PET centiloid, tau PET temporal meta-ROI SUVR and AVLT as the outcomes. The published cut-off points described above were similarly applied to the ADNI cohort to estimate the timing that the markers become abnormal in the population.

Results

Participants

The study included 2369 MCSA participants, mean (SD) age of 77 (11) years, 47% female, 28% APOE ε4 carriers and 82% CU (Table 1). Most participants had more than one cognitive assessment with an average of 7.2 (4.3) years of cognitive follow-up. A total of 503 participants had plasma analysed at C2N and 1037 had Lilly p-tau217. The demographics of the C2N and Lilly cohorts are shown in Supplementary Table 1 and, for reference, Supplementary Fig. 2 shows the correlation of Lilly versus C2N p-tau217 concentrations for the subset of 325 individuals with a same visit sample analysed on both assays. Nearly all (96%) participants had at least one amyloid PET scan and 1407 (59%) of participants had at least one tau PET.

Table 1.

Participant characteristics at the most recent visit with PET or plasma, unless otherwise specified

	MCSA (N = 2369)	ADNI (N = 1591)
Age, years; mean (SD) [range]	77 (11) [50–102]	76 (8) [53–97]
Female sex, n (%)	1124 (47%)	783 (49%)
APOE ε4 carrier, n (%)	657 (28%)	693 (44%)
Education, years; mean (SD) [range]	15 (3) [0–20]	16 (3) [6–20]
Diagnosis
CU	1937 (82%)	648 (41%)
MCI	360 (15%)	515 (32%)
Dementia	68 (3%)	428 (27%)
Total no. amyloid-β PETa
0	97 (4%)	0
1	1018 (43%)	643 (40%)
2	570 (24%)	443 (28%)
≥3	684 (29%)	505 (32%)
Total no. tau PET
0	961 (41%)	811 (51%)
1	685 (29%)	456 (29%)
2	428 (18%)	193 (12%)
≥3	295 (12%)	131 (8%)
Total no. p-tau217 (C2N)
0	1866 (79%)	1229 (77%)
1	22 (1%)	1 (0.1%)
2	144 (6%)	3 (0.2%)
≥3	337 (14%)	358 (23%)
Total no. p-tau217 (Lilly)
0	1332 (56%)	1591
1	60 (3%)	0
2	261 (11%)	0
≥3	716 (30%)	0
Total no. AVLT
0	14 (1%)	0 (0%)
1	142 (6%)	246 (15%)
2	191 (8%)	236 (15%)
≥3	2022 (85%)	1109 (70%)
Amyloid-β PET, centiloids; mean (SD) [range]	35 (39) [−15 to 206]	43 (5) [−37 to 264]
Tau PET (FTP), SUVR; mean (SD) [range]	1.21 (0.14) [0.86–2.46]	1.31 (0.27) [0.92–2.82]
p-Tau 217 (C2N), pg/mlb; mean (SD) [range]	2.45 (2.91) [0.08–24.22]	3.35 (3.22) [0.65–19.86]
p-Tau 217 (Lilly), pg/mlc; mean (SD) [range]	0.20 (0.16) [0.02–1.28]	NA
AVLT, mean (SD) [range]	57 (21) [3–105]	46 (24) [0–104]
Global z-score, mean (SD) [range]	−0.2 (1.4) [−5.6 to 3.0]	NA

AlzCS = Alzheimer's clinical syndrome; ADNI = Alzheimer’s Disease Neuroimaging Initiative; AVLT = auditory verbal learning test sum of trials; C2N = C2N Diagnostics; CU = cognitively unimpaired; Lilly = Lilly Research Laboratories; MCI = mild cognitive impairment; MCSA = Mayo Clinic Study of Aging; SD = standard deviation; SUVR = standardized uptake value ratio.

^aNumber of participants with 0, 1, 2, or ≥3 serial measurements of that marker.

^bC2N p-tau217 available as V1 in MCSA and V2 in ADNI. p-Tau217 concentrations were converted from version1 to version2 using a formula provided by C2N: version1 × 1.793791 − 0.372865.

^cLilly p-tau217 available only in MCSA participants.

The ADNI cohort was comprised of 1591 participants, age 76 (8) years, 49% female, 44% APOE ε4 carriers and 41% CU (Table 1). Participants had on average 3.5 (3.0) years of cognitive follow-up. All included ADNI participants had an amyloid PET. A total of 362 participants had plasma analysed at C2N and 780 (49%) had at least one tau PET scan.

Variability in metrics

Figure 1 presents a horizontal box plot summarizing the 25th, 50th and 75th percentiles of the noise-to-signal ratio for each biomarker and cognitive measure. The centre denotes the median (50th percentile) of this ratio and the box spans the 25th–75th percentiles across individuals. Lower values indicate that within-person variability is small relative to the range of between-individual signal, suggesting greater longitudinal stability and better reproducibility. In contrast, higher values—closer to 1—indicate that within-individual noise is comparable to or exceeds between-individual variability, making the measure more difficult to model reliably over time. Amyloid and tau PET biomarkers have the lowest noise-to-sign (highest reproducibility) with median SD/IQR of 0.11 and 0.12, respectively. C2N Plasma %p-tau217, p-tau217, Lilly p-tau217 and AVLT had values of 0.26, 0.23, 0.17 and 0.18, respectively. The SD/IQR was greatest for Aβ_42/40 (0.37) and %p-tau181 (0.37).

Model fits

The longitudinal trajectories of biomarker values versus age and adjusted age (the participant’s estimated age with respect to the biomarker of interest based on the covariate and random effects) for the MCSA and ADNI are shown in Supplementary Figs 3 and 4.

Association of biomarker and cognitive timing

Associations of individual adjustments among biomarker and AVLT pairs are shown in Figs 2 and 3, Table 2 and Supplementary Fig. 5; note, these are correlations of time-shifts in years not raw biomarker values. In the MCSA, each of the biomarker individual adjustments was modestly correlated with that of the AVLT, R (95% credible interval) = 0.47 (0.39, 0.54) for amyloid PET, 0.40 (0.27, 0.52) for C2N p-tau217, 0.44 (0.35, 0.52) for Lilly p-tau217 and 0.38 (0.29, 0.47) for tau PET (Fig. 2 and Table 2). These results may be interpreted as: how early versus late an individual progressed on C2N p-tau217 accounted for 16% (R² = 0.16) of the variability in how early versus late an individual progressed on the AVLT relative to the population mean. The associations of individual adjustments of biomarker pairs were greatest for the two p-tau217 assays, R (95% CI) = 0.76 (0.69, 0.82). Plasma p-tau217 (C2N and Lilly) individual adjustments were more strongly associated with amyloid than tau PET biomarkers: R (95% CI) = 0.64 (0.57, 0.71) and 0.56 (0.51, 0.61) for amyloid PET and 0.36 (0.27, 0.45) and 0.28 (0.20, 0.35) for tau PET.

Figure 2.

MCSA correlations of individual adjustments. Relationships of individual adjustments between amyloid PET (PiB, meta-ROI), tau PET (flortaucipir, temporal meta-ROI), p-tau217 (Lilly), p-tau217 (C2N) and the auditory verbal learning test (AVLT) sum of trials. These are correlations of time shifts in years, not correlations of raw biomarker values. Each dot represents one participant and the number of participants included in each comparison varies by data availability. The axes represent individual adjustments (time shifts) in years, indicating how much earlier or later an individual progresses on that marker relative to the population mean. A higher positive value or earlier onset relative to the population mean is shown to the right of the x-axis and top of the y-axis. An 80% ellipse is included with a perfect circle indicating no relationship between adjustments. The per cent variation explained (square of the correlation × 100) between individual-level adjustments is given in the top left. C2N = C2N Diagnostics; Lilly = Lilly Research Laboratories; MCSA = Mayo Clinic Study of Aging; PiB = Pittsburgh Compound-B; ROI = region of interest.

Figure 3.

ADNI correlation of individual adjustments. Relationships of individual adjustments between amyloid PET, tau PET (temporal meta-ROI), p-tau217 (C2N) and the auditory verbal learning test (AVLT) sum of trials. These are correlations of time shifts in years, not correlations of raw biomarker values. Each dot represents one participant and the number of participants included in each comparison varies by data availability. The axes represent individual adjustments (time shifts) in years, indicating how much earlier or later an individual progresses on that marker relative to the population mean. A higher positive value or earlier onset relative to the population mean is shown to the right of the x-axis and top of the y-axis. An 80% ellipse is included with a perfect circle indicating no relationship between adjustments. The per cent variation explained (square of the correlation × 100) between individual-level adjustments is given in the top left. ADNI = Alzheimer's Disease Neuroimaging Initiative; C2N = C2N Diagnostics; Lilly = Lilly Research Laboratories; ROI = region of interest.

Table 2.

MCSA and ADNI correlation coefficients, R (95% credible interval), between individual-level adjustments from a non-linear mixed-effects model

	Amyloid PET	Tau PET	p-Tau217 (C2N)	p-Tau217 (Lilly)
MCSA
Tau PET	0.45 (0.40, 0.50)	–	–	–
p-tau217 (C2N)	0.64 (0.57, 0.71)	0.36 (0.27, 0.45)	–	–
p-tau217 (Lilly)	0.56 (0.51, 0.61)	0.28 (0.20, 0.35)	0.76 (0.69, 0.82)	–
AVLT	0.47 (0.39, 0.54)	0.38 (0.29, 0.47)	0.40 (0.27, 0.52)	0.44 (0.35, 0.52)
ADNI				–
Tau PET	0.60 (0.54, 0.65)	–	–	–
p-Tau217 (C2N)	0.84 (0.80, 0.87)	0.63 (0.56, 0.70)	–	–
AVLT	0.74 (0.69, 0.79)	0.76 (0.69, 0.82)	0.81 (0.74, 0.86)	–

ADNI = Alzheimer's Disease Neuroimaging Initiative; AVLT = auditory verbal learning test sum of trials; C2N = C2N Diagnostics; Lilly = Lilly Research Laboratories; MCSA = Mayo Clinic Study of Aging.

In ADNI, the trends were similar though the correlations were stronger for all pairwise comparisons (Fig. 3, Table 2 and Supplementary Fig. 5). Each of the biomarker individual adjustments showed a modest to strong association with the AVLT, R (95% CI) = 0.74 (0.69, 0.79) for amyloid PET to R (95% CI) = 0.81 (0.74, 0.86) for C2N p-tau217. The timing of p-tau217 progression was more strongly correlated with that of amyloid PET, R (95% CI) = 0.84 (0.80, 0.87), than tau PET, R (95% CI) = 0.63 (0.56, 0.70).

Covariate effects

The left–right covariate effects are shown in Fig. 4 and Supplementary Tables 2 and 3. In the MCSA, APOE ε4 carriership showed the strongest covariate effect on amyloid PET, plasma p-tau217 and AVLT biomarkers (Fig. 4 and Supplementary Table 2). For example, APOE ε4 carriers progressed on amyloid PET on average 8.4 (7.3, 10.0) years earlier than non-carriers. Sex and education showed small, variable effects on the timing of biomarker progression. For example, male sex was associated with slightly later progression on amyloid PET, −1.3 (−2.4, −0.3) years, but slightly early progression on p-tau217 (Lilly), 2.2 (0.2, 4.0) years. Education did not have a statistically significant effect on any of the biomarkers.

Figure 4.

MCSA and ADNI estimated covariate effects with 95% credible interval. A higher positive value or earlier onset relative to the population mean is shown to the right. For example, APOE ε4 carriership is associated earlier progression on all markers. ADNI = Alzheimer’s Disease Neuroimaging Initiative; AVLT = auditory verbal learning test sum of trials; C2N = C2N Diagnostics; Lilly = Lilly Research Laboratories; MCSA = Mayo Clinic Study of Aging.

For the baseline covariate effects (Table 3), male sex and lower education were associated with lower baseline AVLT scores, −12.3 (−13.4, −11.1) for male versus female sex and −8.2 (−9.7, −6.8) for high school versus college education.

Table 3.

Baseline covariate effects (95% credible intervals) on the AVLT for MCSA and ADNI from non-linear mixed effect models

	MCSA	ADNI
Male versus female	−12.3 (−13.4, −11.1)	−10.6(−12.5, −8.7)
High school versus college	−8.2 (−9.7, −6.8)	−8.0 (−11.1, −4.8)
Graduate school versus college	4.2 (2.8, 5.6)	5.3 (3.4, 7.4)

The covariates shift curves along the y-axis (Supplementary Figs 2 and 3), accounting for baseline effects. ADNI = Alzheimer’s Disease Neuroimaging Initiative; AVLT = auditory verbal learning test sum of trials; MCSA = Mayo Clinic Study of Aging.

In ADNI, APOE ε4 carriership similarly had a strong effect on the timing of progression of all biomarkers and on average a larger effect than in the MCSA sample (Fig. 4 and Supplementary Table 3). For example, amyloid PET progressed on average 18.2 (16, 20) years earlier in APOE ε4 carriers versus non-carriers. Sex and education effects remained small, close to 0 and not statistically significant with the exception of the male sex effect on p-tau217, R = −2.0 (−3.9, −0.1).

Timing of biomarker progression on the population level

The estimated timing of biomarker and cognitive progression are shown in Fig. 5 via the population mean curves of the percentage of participants above the employed cut-off points versus the adjusted age. The adjusted age accounts for covariates and random effects and corresponds to the right column in Supplementary Figs 3 and 4. Using established cut-off points, amyloid PET becomes abnormal first, followed by p-tau217, tau PET and AVLT. For example, at age 80 years, 49% of MCSA participants are estimated to have abnormal amyloid PET, 36% C2N p-tau217, 23% tau PET and 9% AVLT (Supplementary Table 4). Above the age of 90, the slope of the AVLT curve increases and crosses (becomes abnormal before) the Lilly plasma p-tau217 and tau PET curves. Therefore, at age 90 years, 77% MCSA participants are abnormal on amyloid PET, 69% on C2N p-tau217 and 39% on tau PET and 36% on the AVLT. In ADNI, at age 80 years, 59% of participants are abnormal on amyloid PET, 57% on C2N p-tau217, 45% on tau PET and 35% on AVLT. In the parallel portions of the amyloid PET and AVLT curves in Fig. 5, the AVLT was estimated to become abnormal approximately 13 (12, 14) years after amyloid PET in the MCSA and 10 (9, 11) years after amyloid PET in ADNI.

Figure 5.

Timing of biomarker and cognitive progression. By year of age, predicted proportion of MCSA and ADNI participants with abnormal biomarker levels using established cut-off points. These are population mean curves based on the model output. The adjusted age is age in years adjusted for covariates and random effects. ADNI = Alzheimer’s Disease Neuroimaging Initiative; AVLT = auditory verbal learning test sum of trials; C2N = C2N Diagnostics; Lilly = Lilly Research Laboratories; MCSA = Mayo Clinic Study of Aging; p-tau = phosphorylated tau.

Sensitivity analyses

The MCSA model fits with global z-score showed very similar results to the primary data fit with AVLT. The association of individual adjustments, covariate effects and timing of biomarker progression for the model fit with global z-score are shown in Supplementary Tables 5 and 6 and Supplementary Fig. 6.

Discussion

In this work we describe the temporal progression of plasma and PET Alzheimer’s disease biomarkers with cognitive decline. The primary findings are: (i) how early versus late an individual progressed on a plasma p-tau217, amyloid PET, or tau PET compared to the population mean was modestly associated with how early versus late an individual was estimated to undergo cognitive decline in the MCSA; (ii) the association of the timing of Alzheimer’s disease biomarker and cognitive progression was stronger in ADNI; (iii) the strength of the association of plasma p-tau217 progression with cognition was similar to that of amyloid and tau PET biomarkers with cognition; (iv) as seen in prior work, APOE e4 carriership was associated with earlier progression of biomarkers of amyloid pathology and cognitive decline; and (v) beyond age 90 years, AVLT was estimated to become abnormal before tau biomarkers, likely related to increasing prevalence of co-pathology contributing to cognitive decline at older ages.

The correlation of individual adjustments can be thought of as an estimate of the strength of mechanistic associations, in that two processes that have strong mechanistic associations would be reliably linked in time. Note that ‘linked in time’ does not require that two variables change simultaneously—that is, two variables may be highly linked in time but with a temporal offset such that one changes prior to the other. In the MCSA, the association of individual adjustments between biomarkers and cognition was modest, although that may be anticipated in a population-based sample in which individuals may have Alzheimer’s disease pathology as well as other underlying pathologies driving early cognitive decline.⁴⁶ These correlations were higher in the ADNI sample, which was designed to mirror Alzheimer’s disease clinical trial samples and thus included a larger percentage of patients with MCI and dementia secondary to Alzheimer’s disease. Of the biomarkers studied, we anticipated tau PET individual adjustments would be the most strongly associated with cognition adjustments based on the well-established closer temporal proximity of tau accumulation with cognitive decline.^14,47,48 For example, a study by Tosun et al.⁴⁹ showed that up to 56% of variance in longitudinal cognitive decline was explained by amyloid, tau and atrophy, ranging from 16% for amyloid to 46%–47% for tau. In the MCSA, the weaker association of tau with AVLT than with other biomarkers is likely related to the population being mostly CU and with no to minimal measurable tau deposition, which underestimates the effect. The stronger association of the timing of amyloid and cognitive progression corresponds with findings in prior work showing that amyloid is associated with subsequent cognitive decline in participants who are CU or have MCI.^50-52 In ADNI, the associations of each amyloid PET, p-tau217 and tau PET timing with AVLT were strong and quite similar. Notably, the correlation of plasma p-tau217 individual adjustments with AVLT adjustments matched that of PET with AVLT, supporting p-tau217 as an indicator of disease progression—individuals who progress earlier on p-tau217 will on average undergo earlier cognitive decline.

Of the covariates, APOE ɛ4 carriership had the greatest effect. In the MCSA, the relative APOE ɛ4 carriership effects sizes were amyloid and p-tau217 > AVLT > tau PET. This is in keeping with a most proximal effect of APOE ɛ4 carriership on amyloid deposition and agrees with prior work showing that APOE ɛ4 carriership is associated with earlier cognitive decline.^53-56 However, the tau PET effect was likely underestimated by the low prevalence of detectable tau PET signal in the MCSA cohort, which is dominated by unimpaired individuals. The stronger APOE ɛ4 carriership effect on AVLT may be due to the fact that the model fits are bolstered and confidence intervals narrower due to the larger number of participants with AVLT (versus tau PET) and greater extent of longitudinal data. In ADNI, in which APOE ɛ4 carriership and tau accumulation were more common, APOE ɛ4 carriership had a very strong effect on earlier amyloid and tau accumulation, with a lower, yet large, effect size on AVLT.

The effects of sex and education on earlier versus later biomarker progression was minimal, similar to prior work.^18,21,22 Females appeared to progress on average slightly earlier on amyloid and tau PET biomarkers in the MCSA, although there was no effect of sex on amyloid or tau PET biomarkers in ADNI. Education did not have a significant effect on biomarker timing in either the MCSA or ADNI, as expected given a lack of a clear biological connection. We observed expected effects of sex and education on baseline cognitive metrics, lower scores in males versus females^23,24,57 and those with less years of education.^23-26,58

On average in both the MCSA and ADNI cohorts, markers became abnormal in the expected order of amyloid PET, p-tau217, tau PET and cognition up to approximately age 90 years. This ordering is in keeping with the amyloid cascade hypothesis and supporting data, as well as similar to results of prior biomarker longitudinal studies.^18-20 The estimated years from abnormal amyloid PET to abnormal AVLT was estimated to be 13 (12, 14) in the MCSA and 10 (9, 11) years in ADNI. These estimates are a few years longer than prior work using event-based models to estimate disease timelines, which showed amyloid PET abnormality to precede cognitive changes by approximately 7.5 years.⁵⁹ Rates of tau PET positivity increased from about 5% to 25% in the MCSA and 20% to 45% in ADNI from ages 60 to 80 years. These are similar to rates of 4.2% to 12.8% in CU and 35.4% to 42.8% in MCI in a multicohort study by Ossenkoppele et al.⁶⁰ and to rates reported in another multi cohort study by Moscoso et al.⁶¹ At older ages, the cognition curve crossed and became abnormal before the tau PET and p-tau217 curves, likely secondary to contributions of non-Alzheimer’s disease pathology to cognitive decline. Additionally, Lilly p-tau217 appeared to become abnormal after C2N p-tau217 in the MCSA. This may reflect differences in sensitivity of the assays at low levels of pathology and/or the application of published cut-off points, which were established using different cohorts that may not be entirely comparable. Although both assays have shown good predictive power (e.g. for amyloid positivity⁶²), plots of cross-sectional biomarker concentrations (Supplementary Fig. 2) show that the measures are not the same, and it is not unexpected that the timing curves are different.

As in our prior work and that of others, p-tau217 appears to become abnormal after amyloid PET and before tau PET^18,20,63 and supports rises in plasma p-tau217 to be a response to amyloidosis, indicative of pathologic tau processing that is detected before tau aggregation is detected on PET. This observed timeline of biomarker changes (amyloid PET, plasma p-tau217 and then tau PET) supports that plasma p-tau217 starts to progress relatively early in the disease process and its use as a Core 1 biomarker in the revised Alzheimer’s disease research framework.⁶⁴ Additionally, as seen in prior work, timing of progression on p-tau217 (correlation of individual adjustments) was more strongly correlated with timing of progression on amyloid than tau PET biomarkers,¹⁸ supporting a close link between amyloid deposition and the phosphorylation of tau.

The other primary plasma Alzheimer’s disease biomarker of interest for use in diagnosis and predicting disease progression is Aβ_42/40. Despite good performance of the C2N Aβ_42/40 assay in cross-sectional studies, and higher performance than other assays,⁶⁵ modelling of this biomarker was challenging with our approach. The variability in measurement did not allow for reliable longitudinal modelling. Markers with lower noise (intra-individual variation) to signal (interindividual variation) in longitudinal biomarker measurements (Fig. 1) have been more amenable to longitudinal modelling with our approach and others. Biomarkers with an SD/IQR (a signal to noise measure) of ∼0.1 have been shown to behave well in our and others’ approaches to longitudinal modelling^18,66; for example, modelling amyloid PET consistently works well. For biomarkers with values around 0.25 (e.g. p-tau217), our method works well but requires more care (e.g. starting estimates, priors, convergence criteria). For markers with values around 0.35 (p-tau181 and Aβ_42/40) model results are not reproducible, which can be accounted for by the relatively large variation in values in an individual over time. The cognitive metrics, AVLT and global z-score, showed longitudinal variability similar to p-tau217, and the relatively large number of serial cognitive assessments for participants in both cohorts helped to stabilize the cognitive metric fits.

The model fits and data trajectories (Supplementary Figs 3 and 4) were similar to prior work for the biomarkers. Biomarker trajectories showed relatively flat baselines followed by a tipping point of rapid increase; the more gradual rise seen for tau PET may be related to a more limited dynamic range, weighted to low tau PET values. The shape of the cognitive trajectories mirrored that of the other biomarker trajectories with a relatively flat slope at younger ages followed by rapid decline beyond a tipping point, as has been seen in prior work modelling longitudinal cognitive trajectories.⁶⁷

There are limitations to this study. MCSA and ADNI are cohorts with samples from primarily white individuals with above average education. Future work will include a broader range of covariates to analyse in collaboration with other sites. Cut-off point obtainment is imperfect. Accepted published cut-off points that are widely applied in research were employed and they performed as expected in both the MCSA and ADNI cohorts. As discussed, co-pathology with Alzheimer’s disease is common and may contribute to cognitive decline in individuals along the Alzheimer’s disease spectrum, as studied in this work. Future work will consider temporal progression and contributions of common non-Alzheimer’s disease pathology, such as cerebrovascular disease, to the timing of cognitive decline.

In conclusion, the association of the timing of plasma and PET Alzheimer’s disease biomarker progression with cognitive decline was modest in the MCSA population-based sample and stronger in the Alzheimer’s disease-enriched ADNI cohort. The timing of plasma p-tau217 progression explained a similar degree of variability in AVLT progression as amyloid PET, supporting its utility as a marker of disease progression. The estimated temporal ordering of biomarkers and cognitive abnormality was as anticipated (amyloid, tau, cognition) up to the age of 90, beyond which AVLT was estimated to become abnormal prior to tau biomarkers, likely related to the effects of non-Alzheimer’s disease co-pathologies.

Data availability

MRI, PET and other data from the Mayo Clinic Study of Aging are available to qualified academic and industry researchers by request to the MCSA Executive Committee (https://www.mayo.edu/research/centers-programs/alzheimers-disease-research-center/research-activities/mayo-clinic-study-aging/for-researchers/data-sharing-resources). The code is available at github.com/Therneau/AFTmodel/fourmarkers. The PET and MRI measurement pipeline is available at https://www.nitrc.org/projects/mcalt/.

Funding

This work was supported by the National Institutes of Health [U01 AG006786, P50 AG016574, R37 AG011378, RO1 AG041851, R01 NS097495, R01 AG056366, 5R01AG069052-03].

Data collection and sharing for the Alzheimer's Disease Neuroimaging Initiative (ADNI) is funded by the National Institute on Aging (National Institutes of Health Grant U19AG024904). The grantee organization is the Northern California Institute for Research and Education. In the past, ADNI has also received funding from the National Institute of Biomedical Imaging and Bioengineering, the Canadian Institutes of Health Research, and private sector contributions through the Foundation for the National Institutes of Health (FNIH) including generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; BristolMyers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics.

Competing interests

P.M.C. has received honoraria from Eisai and payment for CME from Kaplan, Medical Learning Institute and PeerView. She serves on a Lilly and Eisai DMC but receives no compensation to self or institution. E.S.L., T.M.T., M.H., M.E.G., H.J.W., N.H.S. and J.L.G. have no disclosures. M.M.M. receives research support from the NIH. J.B.B., T.W. and P.B.V. are paid employees of C2N Diagnostics. J.G.R. serves as an associate editor for JAMA Neurology and receives research support from the NIH. He serves on the DSMB for NINDS StrokeNET. He serves as a site investigator for trials sponsored by Eisai and Cognition Therapeutics. He received honararium for serving as faculty at the IMPACT-AD course and course director for the AAN. A.A.-S. has participated in advisory boards for Roche Diagnostics, Fujirebio Diagnostics and Siemens Healthineers. She also has received honorarium from Roche Diagnostics and Eli Lilly. V.J.L. is a consultant for AVID Radiopharmaceuticals, Eisai Co. Inc., Bayer Schering Pharma, GE Healthcare, Piramal Life Sciences and Merck Research and receives research support from GE Healthcare, Siemens Molecular Imaging, AVID Radiopharmaceuticals and NIH (NIA, NCI). C.G.S. receives research support from the NIH. M.L.S. holds stock in medical-related companies unrelated to the current work: Align Technology, Inc., LHC Group, Inc., Medtronic, Inc., Mesa Laboratories, Inc., Natus Medical Inc. and Varex Imaging Corporation. He has also owned stock in these medical-related companies within the past three years, unrelated to the current work: CRISPR Therapeutics, Gilead Sciences, Inc., Globus Medical Inc., Inovio Biomedical Corp., Ionis Pharmaceuticals, Johnson & Johnson, Medtronic, Inc., Oncothyreon, Inc., Parexel International Corporation. D.S.K. serves on a Data Safety Monitoring Board for the Dominantly Inherited Alzheimer Network Treatment Unit study sponsored by Washington University St Louis and for the SMART-HS clinical trial (Univ of Kentucky). He was an investigator in Alzheimer clinical trials sponsored by Biogen, Lilly Pharmaceuticals and the University of Southern California, both of which have ended, and is currently an investigator in a trial in frontotemporal degeneration with Alector. He has served as a consultant for Roche, AriBio, Linus Health, Biovie and Alzeca Biosciences but receives no personal compensation. He receives funding from the NIH. P.V. receives research support from the NIH. R.C.P. serves as a consultant for Roche Inc., Merck Inc. and Biogen, Inc. He serves on the Data Safety monitoring Board for Genentech, Inc. and receives royalty from Oxford University Press and UpToDate. C.R.J. receives no personal compensation from any commercial entity. He receives research support from NIH and the Alexander Family Alzheimer’s Disease Research Professorship of the Mayo Clinic.

Supplementary material

Supplementary material is available at Brain online.