Evaluation of CSF and plasma tau species as fluid surrogate candidates for tau PET in prodromal to moderate Alzheimer's disease

Abstract

INTRODUCTION

Positron emission tomography (PET) is a valuable tool for assessing tau pathology in Alzheimer's disease (AD), but it is not widely accessible. Clarifying the relationship between fluid tau species and tau PET in AD may allow for the identification of fluid biomarkers that could serve as more accessible surrogates for tau PET.

METHODS

Cerebrospinal fluid (CSF) and plasma tau species levels were assessed across prodromal to moderate AD subjects (CSF: n = 53, plasma: n = 181) via immunoassays and liquid chromatography‐mass spectrometry. Fluid measures were correlated with [¹⁸F]GTP1 tau PET at baseline for head‐to‐head comparisons across analytes.

RESULTS

CSF C2N‐eMTBR‐tau243, tau phosphorylated at threonine 205 (p‐tau205), p‐tau217, and a peptide from the tau‐441 microtubule‐binding region (MTBR) (MTBR/243‐254) exhibited the highest correlations with [¹⁸F]GTP1 standardized uptake value ratio. Plasma p‐tau181 and p‐tau217 demonstrated correlation rankings with tau PET similar to those observed with corresponding CSF analytes.

DISCUSSION

CSF C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217 demonstrated comparable strong correlations with tau PET. These findings may guide future development of plasma biomarker surrogates for tau PET.

Highlights

• Correlations of 20 CSF tau forms with tau PET in AD were assessed.

• CSF C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217 showed comparable correlations.

• CSF C2N‐eMTBR‐tau243 had the numerically highest correlation with cortical tau PET.

• Plasma p‐tau181 and p‐tau217 tau PET correlations matched rankings of their CSF forms.

1. BACKGROUND

Alzheimer's disease (AD) is one of the leading forms of dementia and accounts for an estimated 60% to 70% of over 57 million dementia cases globally. ¹ Based on findings from the Global Burden of Disease Study, the number of dementia cases worldwide is forecasted to reach 152.9 million by 2050, suggesting that AD is continuing to develop into a significant global health concern. ²

Neurofibrillary tangles (NFTs) are one of the pathological hallmarks of AD and other tauopathies and are formed intracellularly in neurons from insoluble aggregates of misfolded tau protein. ³ , ⁴ , ⁵ , ⁶ Under normal physiological conditions, tau functions as a microtubule‐associated protein that facilitates the stabilization and assembly of microtubules, as well as supporting axonal transport. ³ , ⁷ , ⁸ However, post‐translational modifications (PTMs), including hyperphosphorylation, isomerization, acetylation, and truncation, have been shown to interfere with normal interactions between tau with microtubules, resulting in elevated levels of unbound cytosolic tau. ³ , ⁴ , ⁵ , ⁹ , ¹⁰ , ¹¹ The disrupted equilibrium between bound and unbound tau is thought to contribute to the formation of NFT pathology, which is hypothesized to spread through stereotyped patterns across brain regions by the release of pathological tau species behaving in a prion‐like manner. ⁴ , ⁵ , ⁶ , ¹² , ¹³ The spatial progression of NFT pathology with disease severity have been characterized by the six Braak stages where an early presence of tau pathology is observed in the transentorhinal region, followed by eventual spread to the limbic and isocortical regions in later disease stages. ¹³

Imaging studies have shown that NFT pathology can emerge prior to clinical symptoms or neurodegeneration and that the amount of NFT pathology correlates strongly with AD clinical scores across a broad spectrum of disease severity. ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ For such studies and in AD clinical trials, tau positron emission tomography (PET) tracers have been utilized as tools to delineate NFT pathology patterns, stage disease, and assess the impact of interventions targeting tau tangles or amyloid plaques. ¹⁴ , ¹⁵ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ However, utilizing tau PET imaging in large international multicenter clinical trials presents high costs and logistical challenges. These include the limited geographic availability of ligand production networks and difficulty finding access to imaging centers, especially for patients with dementia who may encounter additional challenges related to travel and cognitive function.

As an alternative to tau PET, cerebrospinal fluid (CSF) and plasma tau measures, such as total tau and tau phosphorylated at threonine 181 (p‐tau181), are commonly collected in AD trials to evaluate pharmacodynamic (PD) effects of anti‐tau therapeutic candidates and may be more widely available. ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ However, tau can be highly modified with PTMs and is also known to be highly fragmented, contributing to the heterogeneous nature of tau in CSF. ²⁸ , ²⁹ These variations warrant further investigation to distinguish between tau species and their relationships with NFT pathology in AD.

Despite extensive validation of individual tau biomarkers – including p‐tau217, p‐tau205, and microtubule‐binding region (MTBR)‐related species – the field lacks clarity on whether these leading candidates capture overlapping or biologically distinct aspects of tau pathology. Most prior studies evaluated individual analytes or small panels in separate cohorts, limiting the ability to directly compare performance, biological specificity, and translational potential within the same individuals. ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ Consequently, it remains uncertain whether expanding tau analyte panels meaningfully enhances biological interpretation and staging relevance or simply reiterates established findings from top‐performing markers. Addressing this uncertainty requires systematic, head‐to‐head evaluation of diverse tau proteoforms – including phosphorylation‐specific, truncation‐derived, MTBR‐associated, and domain‐spanning species – anchored to a common tau PET measure within the same cohort.

To fill this gap, this study integrates complementary assay platforms – including data‐independent acquisition mass spectrometry (DIA‐MS), immunoassays, and immunoprecipitation (IP) followed by liquid chromatography‐tandem mass spectrometry (IP‐LC‐MS/MS) – to comprehensively profile CSF tau species and directly compare their relationships with [¹⁸F]GTP1 tau PET across prodromal to moderate AD. Importantly, these analyses were conducted within amyloid‐confirmed, protocolized clinical trial cohorts with harmonized PET acquisition and centralized sample handling, minimizing diagnostic and pre‐analytical variability and enabling high‐confidence cross‐analyte comparisons. ²³ , ²⁴ By anchoring expanded tau profiling to a controlled disease spectrum and a single tau PET tracer, this design clarifies whether broader analyte panels provide added biological and translational value. In addition, by identifying CSF markers with robust and specific associations to tau PET, this work informs the prioritizations for plasma‐based surrogates, which offer greater scalability, accessibility, and feasibility for clinical trials and eventual clinical implementation.

2. METHODS

2.1. Study design

This study evaluates subsets of participants at baseline from anti‐tau semorinemab phase 2 trials, Tauriel (NCT03289143) and Lauriet (​​NCT03828747). ²³ , ²⁴ The safety and efficacy of semorinemab was evaluated in both randomized, double‐blind, placebo‐controlled trials: in prodromal to mild AD for Tauriel and in mild to moderate AD for Lauriet. Tauriel and Lauriet enrolled 457 and 272 participants, respectively, meeting the following eligibility criteria: mild cognitive impairment or AD dementia diagnosis for Tauriel and probable AD dementia diagnosis for Lauriet, years of age between 50 and 80 for Tauriel and 50 and 85 for Lauriet, Mini‐Mental State Examination (MMSE) score between 20 and 30 for Tauriel and 16 and 21 for Lauriet (inclusive), Clinical Dementia Rating‐Global Score (CDR‐GS) of 0.5 or 1 for Tauriel and 1 or 2 for Lauriet, Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) Delayed Memory scores ≤ 85 for Tauriel, and the presence of amyloid pathology confirmed by amyloid PET imaging or CSF Aβ(1‐42) ≤ 1000 pg/mL by Elecsys B‐Amyloid ¹ , ² , ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² CSF immunoassay (Roche Diagnostics International) for both studies. Apolipoprotein E (APOE)ε status of participants was determined by the presence or absence of at least one APOE ε4 allele. Participants completed the Alzheimer's Disease Assessment Scale‐Cognitive Subscale (ADAS‐Cog13 in Tauriel, ADAS‐Cog11 in Lauriet). To enable comparisons across cohorts, ADAS‐Cog11 scores were calculated for Tauriel participants.

Magnetic resonance imaging (MRI) scans and tau PET imaging was performed as previously described. ²⁴ Briefly, MRI scans were acquired on 1.5T or 3T scanners across different sites. Each scanner was individually qualified by the central imaging vendor (NeuroRx, Montreal, Quebec, Canada) for use in this study via dummy test scans using scanner‐specific sequences provided by the vendor. Tau PET imaging was performed in a subset of participants using Genentech Tau Probe 1 ([¹⁸F]GTP1) ²³ , ²⁴ at several sites, set up by the central PET imaging vendor (Invicro, New Haven, CT, USA). Images were reconstructed in a manner that allowed for harmonization of the data across different PET imaging scanners. Image reconstruction was performed using an iterative/row‐action maximum‐likelihood algorithm reconstruction method with four iterations and 16 subsets (or as close as possible to 16) and a 128 × 128 matrix. A convolution with a three‐dimensional Gaussian kernel was applied to unfiltered reconstructed PET images. The Gaussian full width at half maximum was based on Hoffman phantom data acquired at each site during study set‐up. [¹⁸F]GTP1 PET imaging was done after a dose of 5 to 7 mCi and image acquisition started 60 min after tracer administration for 30 min. Amyloid PET acquisitions were performed as specified by the respective radioligand manufacturer.

RESEARCH IN CONTEXT

1. Systematic review: Information from studies were reviewed from manuscripts, abstracts, and presentations. Many reports examined correlations between tau species with tau PET, but most were limited to a single platform or assay type. No studies assessing 20 CSF tau measures within one cohort were identified. All relevant publications are cited.

2. Interpretation: Using various LC‐MS methods and immunoassays, 20 CSF tau species were ranked by correlation strength with tau PET. CSF C2N‐eMTBR‐tau243 showed the highest correlation out of the leading surrogate candidates, though ρ values did not differ significantly. Plasma p‐tau217 and p‐tau181 exhibited the strongest correlations with tau PET and matched their rankings of their CSF counterparts.

3. Future directions: Future efforts aimed at validating the leading tau PET surrogate candidate biomarkers in more diverse, larger cohorts is needed to strengthen confidence in these findings. Development of new assays may expand the fluid tau comparisons to help inform the rankings of these biomarkers.

CSF was collected by lumbar puncture into low‐binding tubes at baseline or screening from a subset of participants. Sample tubes were inverted to mix followed by centrifugation for 10 min at 2000 × g at 4°C before storage at −70°C in 0.5‐mL aliquots. The discovery CSF cohort consisted of all participants (n = 53) with complete measurements for all CSF tau species described in this study and available tau PET and amyloid PET data. An expanded CSF cohort (n = 107) was subsequently defined to increase the sample size for the top four CSF correlates identified via the discovery cohort analysis. This expanded cohort was inclusive of all 53 participants from the discovery cohort, supplemented by an additional 54 participants with available measurements for the leading analytes. Plasma samples collected by standard venipuncture techniques into vacutainer tubes at baseline were used for cross‐sectional analysis from subjects in Tauriel. Sample tubes were inverted to mix followed by centrifugation for 5 min at 2000 × g at 4°C before storage at −70°C in 1‐mL aliquots. The plasma cohort (n = 181) included all participants with data for the four plasma tau analytes and was limited to Tauriel participants, as plasma p‐tau217 was not measured in Lauriet participants. Baseline demographics of the discovery CSF, expanded CSF, and plasma cohorts are described in Tables 1, 2, and 3.

TABLE 1.

Baseline demographics of discovery CSF cohort.

Characteristic	Discovery CSF, N = 531	Tauriel, N = 261	Lauriet, N = 271
Age	70.0 (63.0, 76.0)	71.0 (62.3, 75.0)	69.0 (64.0, 77.5)
CDR‐SB	4.5 (4.0, 6.0)	4.0 (3.0, 4.5)	5.5 (4.5, 6.0)
ADCS‐ADL	65.0 (59.0, 72.0)	70.5 (64.3, 73.8)	62.0 (59.0, 67.5)
MMSE	20.0 (18.0, 24.0)	24.0 (20.0, 27.0)	18.0 (17.0, 20.5)
Amyloid Centiloid	77.4 (54.0, 95.0)	64.2 (33.5, 79.8)	86.2 (65.6, 96.9)
ApoE 𝛆4+
ApoE ε4−	17 (32%)	5 (19%)	12 (44%)
ApoE ε4+	36 (68%)	21 (81%)	15 (56%)
Sex
F	28 (53%)	11 (42%)	17 (63%)
M	25 (47%)	15 (58%)	10 (37%)
Education level
< High school	7 (13%)	2 (7.7%)	5 (19%)
≥ High school	46 (87%)	24 (92%)	22 (81%)
1Median (IQR); n (%)

TABLE 2.

Baseline demographics of expanded CSF cohort.

Characteristic	Expanded CSF, N = 1071	Tauriel, N = 411	Lauriet, N = 661
Age	72.0 (65.0, 76.5)	72.0 (67.0, 75.0)	71.5 (65.0, 77.8)
CDR‐SB	5.0 (4.0, 6.0)	4.0 (2.5, 4.5)	6.0 (4.5, 7.0)
ADCS‐ADL	64.0 (57.5, 69.5)	69.0 (65.0, 74.0)	61.0 (56.3, 65.0)
MMSE	20.0 (17.0, 23.0)	23.0 (20.0, 25.0)	18.0 (17.0, 21.0)
Amyloid Centiloid	77.1 (53.6, 95.1)	57.2 (37.9, 79.6)	87.7 (69.2, 96.9)
Unknown	45	12	33
ApoE4+
ApoE4−	37 (35%)	9 (22%)	28 (42%)
ApoE4+	70 (65%)	32 (78%)	38 (58%)
Sex
F	56 (52%)	18 (44%)	38 (58%)
M	51 (48%)	23 (56%)	28 (42%)
Education level
< High school	24 (22%)	7 (17%)	17 (26%)
≥ High school	83 (78%)	34 (83%)	49 (74%)
1Median (IQR); n (%)

TABLE 3.

Baseline demographics of plasma cohort.

Characteristic	Plasma N = 181
Age	71.0 (64.0, 75.0)
CDR‐SB	3.8 (2.8, 5.0)
ADCS‐ADL	70.0 (65.0, 73.3)
Unknown	1
MMSE	23.0 (21.5, 25.5)
Amyloid Centiloid	90.4 (70.7, 114.0)
ApoE4+
ApoE4−	45 (25%)
ApoE4+	136 (75%)
Sex
F	103 (57%)
M	78 (43%)
Education level
< High school	30 (17%)
≥ High school	151 (83%)
1Median (IQR); n (%)

2.2. Tau PET image processing and analysis

[¹⁸F]GTP1 uptake was calculated in Montreal Neurological Institute (MNI) space after spatial normalization with SPM12 ³⁷ using standardized uptake value ratios (SUVRs), with the inferior cerebellar gray matter as the reference region. ³⁸ Target regions of interest (ROIs) (Figure 1B,D) included the whole cortical gray matter (WCG), meta‐temporal region (TMP), defined as the entorhinal cortex, amygdala, parahippocampal gyrus, middle and inferior temporal gyrus, and fusiform gyrus from the Hammers atlas, as well as individual regions from the Hammers atlas. ³⁹

FIGURE 1.

(A) CSF tau peptides detected by DIA proteomics mapped across the tau‐441 isoform. Dotted lines indicate the boundaries between protein domains. (B) Brain map highlighting the whole cortical gray matter PET ROI in red. (C) Spearman's rank correlation coefficients and −log10(FDR adjusted p value) between each CSF tau analyte with [¹⁸F]GTP1 SUVR in the whole cortical gray matter. Only analytes with correlations meeting the significance criteria (ρ > 0.4, p(FDR) < 0.05) are annotated. (D) Brain map highlighting meta‐temporal PET ROI in blue. (E) Spearman's rank correlation coefficients and the −log10(FDR adjusted p value) between each CSF tau analyte with [¹⁸F]GTP1 SUVR in meta‐temporal ROI from discovery CSF cohort. (F) Spearman's rank correlation coefficients and −log10(FDR adjusted p value) between each CSF tau analyte with amyloid PET Centiloids from discovery CSF cohort.

2.3. CSF and/or plasma total tau, p‐tau181, p‐tau217, p‐tau205, brain‐derived (BD) tau immunoassays

CSF p‐tau217 and p‐tau205 were measured using Elecsys robust prototype assays (RPAs) (Roche Diagnostics International). CSF total tau and p‐tau181 were measured using the CE‐marked Elecsys in vitro diagnostic immunoassays on cobas e402 or e801 analyzers (Roche Diagnostics International) as previously reported. ²² , ⁴⁰ CSF and plasma levels of BD tau were measured using the BD tau RPA on a cobas e801 or e402. The p‐tau205 (CSF), p‐tau217 (CSF), and BD tau (CSF and plasma) RPAs each employ target‐specific capture antibodies directed toward the respective analyte. The utilized antibodies are proprietary to Roche, and the RPAs were not commercially available at the time of this study. Plasma p‐tau181, p‐tau217, and total tau were measured using Elecsys assays as previously described. ²²

2.4. IP‐MS CSF MTBR‐tau assay (C2N‐eMTBR‐tau243)

All sample analyses for CSF C2N‐eMTBR‐tau243 were performed at C2N Diagnostics (St Louis, MO, USA). On the day of use, frozen calibrators, quality control (QC) samples, test samples, arginine‐C endopeptidase (Worthington Biochemicals, Lakewood, NJ, USA), and isotopically labeled ¹⁵N,¹³C MTBR internal standard (IS, Biosynth, Gardner, MA, USA) were thawed. Once thawed, buffers required for IP were added to each 2‐mL well on a 96‐well plate, and 200 µL of each sample was placed in the respective well of the 96‐well plate (Thermo Fisher Scientific, Waltham, MA, USA). A slurry of conjugated magnetic beads containing a proprietary N‐terminal anti‐tau antibody was then added to each well on the plate for immunocapture of tau proteins. After immunocapture, the magnetic beads bound to tau were removed from their respective matrix and washed with PBS to reduce non‐specifically bound contaminants prior to enzymatic digestion. Once the tau bound magnetic beads were washed, they were placed in a temperature‐controlled buffer where tau species were digested into peptides. After digestion, the reaction was quenched with formic acid. The digested samples were further purified using reverse‐phase solid‐phase extraction to remove contaminants. The 96‐well collection plate was dried under vacuum before samples were reconstituted with the addition of acetonitrile/formic acid.

The 96‐well collection plate containing the resolubilized tau peptides was placed in a temperature‐controlled autosampler within the LC system (Waters Acquity UPLC M‐Class). Reconstituted tau peptides were injected onto the analytical LC column and separated, identified, and quantified by LC‐MS/MS (Waters Acquity UPLC M‐Class LC coupled to a Thermo Fisher Scientific Fusion Lumos Tribrid Mass Spectrometer). The C2N‐eMTBR‐tau243 concentration was calculated by the summation of peak areas from monitored product ions that result after fragmentation of their respective precursor ions. Precursor ions are derived from endogenous and exogenous stable isotope‐labeled IS peptides that correspond to C2N‐eMTBR‐tau243. After summation, the total peak area for the endogenous tau peptides was divided by the total peak area for the corresponding labeled IS peptides to obtain a peak area ratio (PAR). The PAR of the endogenous peptides to their respective labeled IS peptides was determined in unknown samples, and the concentration of each analyte was calculated from exogenous calibration curves. These data were assembled and assessed by TraceFinder 5.1 General Quan software (Thermo Fisher Scientific). 

Each 96‐well plate included a seven‐point calibrator curve (including the matrix blank), three QC samples, and participant CSF samples. Calibrators were prepared by spiking artificial CSF with known amounts of chemically synthesized, standard peptides corresponding to the region of tau containing C2N‐eMTBR‐tau243. The concentrations of the stock solutions used to prepare the calibrators and IS were value‐assigned by USP‐traceable amino acid analysis. The analytical measurement range of the calibration curves was designed to span the reference ranges of the C2N‐eMTBR‐tau243 measurands in human CSF.

Three QC samples covered the anticipated high, medium, and low CSF concentrations for C2N‐eMTBR‐tau243. The QC samples were randomly analyzed once per run and evenly distributed throughout the unknown samples. QC multi‐rules were applied: The run was rejected when one or more QC concentrations fell outside of three standard deviations (SDs), two or more QC sample concentrations exceeded two SDs, or if the range of two or more QCs exceeded four SDs (1‐3S, 2‐2S, R‐4S rule). The total allowable error for C2N‐eMTBR‐tau243 was defined as 40%.

2.5. IP‐LC‐MS/MS CSF N‐terminal tau and mid‐domain tau assays

CSF concentrations of N‐terminal tau were measured using a targeted LC‐MS multiple reaction monitoring assay as previously described. ²² , ²³ Levels of mid‐domain tau were measured in the same multiplexed assay as N‐terminal tau. Antibodies targeting N‐terminal tau and Mid‐domain tau (N‐terminal: Ab62, Genentech; mid‐domain: 29F12, Genentech) were cross‐linked with Protein A cartridges on the AssayMAP Bravo Platform (Agilent, Santa Clara, CA, USA). Processed CSF samples were spiked with 125 fmol of isotope‐labeled AQUA peptides (Cell Signaling Technology, Danvers, MA, USA) to quantify N‐terminal tau and mid‐domain tau. Diluted samples were loaded onto the cross‐linked cartridges for capture using the affinity purification method on the AssayMAP Bravo Platform. The flow through was collected and loaded onto the same cartridges for a second round of immunocapture. Samples were treated with ammonium hydroxide and iRT standard peptides (Biognosys, Switzerland) before storage at −20°C until LC‐MS analysis. Four transition ions were monitored for the N‐terminal tau precursor ion, A[+42]EPRQEFEVM[+16]EDHAGTYGLGDRK (4+), and three precursor ions were monitored for the mid‐domain tau precursor ion, IATPRGAAPPGQK (3+). Samples were analyzed by an M5 MicroLC (Sciex, Redwood City, CA, USA) coupled with a QTRAP 6500 mass spectrometer (Sciex). Peptides were quantified by calibrating the ratio of the endogenous peptide peak area to the matching isotope‐labeled AQUA peptide using the open source software Skyline (University of Washington, USA).

2.6. CSF proteomics analysis

Protein levels were measured by DIA‐MS on a subset of baseline CSF samples, as previously described. ⁴¹ Peptide‐level data for tau protein were identified by mapping sequences specific to the full‐length protein (UniProtKB accession: P10636). From the 23 detected tau peptides, one was not found in the 2N4R isoform and nine were not detected in all 53 subjects. The remaining 13 peptides were evaluated in the CSF analysis described in the statistical analysis section. Results for PEP.Quantity from Spectronaut were used to report peptide‐level abundances for statistical analyses. The proteomic peptides are referred to by the domain they belong to within the 2N4R isoform, along with the corresponding start and end positions of the peptide, where PRD and MTBR represent the proline‐rich domain and the microtubule‐binding region, respectively.

2.7. Statistical analysis

Spearman's rank coefficients were calculated to test all correlations and the Benjamini‐Hochberg procedure was applied to adjust p values by controlling the false discovery rate (FDR). Meaningfulness and significance criteria for correlation in this study were established as ρ > 0.4 and p(FDR) < 0.05 to filter out analytes with lower correlations and focus on analytes with stronger associations. Given the small cohort size (n = 53) and anticipated noise in the data, the criterion was set to avoid being overly restrictive to minimize risk of excluding analytes that may demonstrate stronger and more robust correlations in larger cohorts. Baseline correlations between [¹⁸F]GTP1 SUVR and amyloid PET with 20 CSF tau species from Tauriel and Lauriet were calculated. Nine additional proteomic tau peptides (N‐term/9‐23, N‐term/12‐24, N‐term/25‐44, PRD/166‐180, PRD/171‐180, PRD/172‐180, PRD/210‐223, MTBR/260‐267, C‐term/386‐395) and BD tau were not included in the correlation matrix analysis due to insufficient overlap in subjects with these measurements. In addition, 95% confidence intervals (CIs) were estimated for Spearman's rank coefficients and rank orders using bootstrap resampling with 2000 replicates. Receiver operating characteristic (ROC) analyses were performed to compare the predictive power of CSF and plasma tau species for tau PET status as high or low burden using previously reported thresholds for Tauriel and Lauriet participants. ²³ , ²⁴ Prediction performance was assessed by calculating the area under the ROC curve (AUC) with 95% CIs, as well as Youden's J index. Spearman's rank coefficients with 95% CIs were calculated to determine associations between tau PET and fluid tau species with clinical scores (MMSE, CDR‐SB, ADCS‐ADL, ADAS‐Cog11) and normalized whole‐brain volume (vMRI) at baseline in the expanded CSF cohort. Longitudinal analysis was performed by estimating subject‐specific rates of change using linear regression models fitted for each participant. For each clinical score or for vMRI, an ordinary least‐squares model was specified, with time as a continuous predictor and the outcome as the dependent variable. The regression coefficient for time was extracted as the individual slope and interpreted as the subject‐specific rate of change over time.

Multiple linear regression models were used to examine the contribution of amyloid PET and tau PET in explaining the variance in CSF and plasma tau analyte levels. Total R ² for each model with the fluid analyte as the dependent variable and amyloid PET, tau PET, age, sex, and APOE status as predictors, was decomposed into a part explained by amyloid PET and tau PET and the remainder (other) using the Lindeman, Merenda, and Gold (LMG) methods. Models with total R ² < 0.1 are not reported to focus on analytes with higher explanatory power. To identify independent predictors of regional tau PET signal in the expanded CSF cohort, Least Absolute Shrinkage and Selection Operator (LASSO) regression and hierarchical modeling were performed for TMP and WCG SUVR. Candidate predictors included CSF tau species (MTBR/243‐254, p‐tau217, p‐tau205, C2N‐eMTBR‐tau243) with age, sex, APOE 4 status, and MMSE as covariates. LASSO regression was conducted with 10‐fold cross‐validation with the penalty parameter chosen by the 1 standard error (λ₁se) rule for feature selection. Selection stability was evaluated using 500 bootstrap resamples, and predictors retained in ≥20% of resamples or selected at λ₁se were carried forward. Hierarchical linear regression models were fitted sequentially using normalized variables: first using covariates only, followed by the addition of C2N‐eMTBR‐tau243, and then incorporating additional tau species. Model performance was evaluated using R ² and change in R ² between nested models to determine incremental explanatory value. The final model for each outcome (WCG and TMP) demonstrated improvement in explanatory power over the preceding model. All analyses were performed using publicly available libraries in R (software version 4.3.1) and Python (version 3.11.4).

3. RESULTS

3.1. Baseline demographics for CSF and plasma cohorts

For the CSF analysis, we first assembled a discovery CSF cohort by incorporating data from 53 prodromal to moderate AD participants enrolled in Tauriel (n = 26) or Lauriet (n = 27) with both tau and amyloid PET data at baseline. The median age of participants was 70 years, 47% were men, 68% were APOE ε4‐positive, and 87% completed the equivalent of high school education or beyond (Table 1). In subsequent CSF analyses using an expanded cohort, we focused on our top candidates (C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau217, and p‐tau205). The expanded CSF cohort consisted of 107 subjects total with 41 from Tauriel and 66 from Lauriet. The median age for this cohort was 72 years, 48% were men, 65% were APOE ε4‐positive, and 78% completed high school education or beyond (Table 2). For the plasma analysis, 181 prodromal to mild AD Tauriel participants with tau and amyloid PET assessments at baseline were evaluated (Table 3). For the plasma cohort, the median age of participants was 71 years, 43% were men, 75% were APOE ε4‐positive, and 83% completed high school education or beyond (Table 3).

3.2. CSF tau and tau PET in prodromal‐to‐moderate AD

Within the discovery CSF cohort, immunoassays, IP‐LC‐MS/MS, and DIA‐MS were used to measure tau species (Table S1). In the DIA‐MS analysis, 13 tau peptides were detected in CSF (Figure 1A), and together with immunoassays or IP‐MS methods directed against total tau, N‐terminal tau, mid‐domain tau, p‐tau181, p‐tau217, p‐tau205, and C2N‐eMTBR‐tau243, we were able to attain a comprehensive assessment of tau domains within the CSF.

We compared the correlations from full brain analyses, represented by the WCG ROI (Figure 1B), to the TMP ROI (Figure 1D) to further investigate areas known to accumulate tau pathology early in AD. Given the wide variability in assay types used to quantify the fluid tau species, the Spearman's rank correlation coefficient (ρ) was employed, as it is based on rank ordering and, thus, robust to large differences in absolute or normalized measurement values. Within this analysis, 12 of the 20 CSF tau measures (five proteomic peptides, C2N‐eMTBR‐tau243, p‐tau217, p‐tau205, N‐terminal tau, p‐tau181, total tau, mid‐domain tau) showed meaningful and significant correlations, ρ > 0.4 and p(FDR) < 0.05, with [¹⁸F]GTP1 WCG and TMP SUVR (Figure 1C,E). Of the well‐correlated peptides, three, referred to as PRD/164‐180, PRD/210‐224, and PRD/181‐194, were from the PRD, while the other two peptides, N‐term/6‐24 and MTBR/243‐254, were from the N‐terminal domain and the MTBR, respectively (Figure 1A). Out of the evaluated CSF species in the discovery CSF cohort, C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217 had the numerically highest correlations with [¹⁸F]GTP1 WCG SUVR (ρ = 0.73, 0.66, 0.65, and 0.63, respectively) (Figure 1C). Strong correlations were observed for the same analytes with TMP SUVR (ρ = 0.72, 0.58, 0.59, and 0.56, respectively) (Figure 1E).

We also investigated the correlation between CSF tau species and amyloid PET. Compared to the correlations observed with tau PET, associations with amyloid PET were weaker, with no CSF tau species meeting meaningfulness and statistical significance (ρ > 0.4, p(FDR) < 0.05) (Figure 1F). This likely reflects the restricted range of amyloid PET values resulting from the exclusion of amyloid‐negative participants from both studies.

Given the relatively small sample size of the discovery CSF cohort, 95% CIs were estimated using bootstrap resampling for Spearman's ρ and rank orders (Figure S1). C2N‐eMTBR‐tau243, p‐tau205, MTBR/243‐254, and p‐tau217 consistently occupied the top ranks for correlations with tau PET signal in the WCG ROI. In the TMP ROI, the overall ordering was largely preserved, but PRD/181‐194 ranked slightly higher than MTBR/243‐254, resulting in a shift in the relative position among the top analytes. Overall, the bootstrap‐derived rankings closely aligned with the corresponding Spearman's ρ estimates and were also consistent with the primary results from this cohort. In contrast, lower‐ranking markers demonstrated greater variability with less concordance between correlation magnitude and rank order.

To prioritize the leading CSF candidates, C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217, we evaluated their correlations with tau PET in the expanded CSF cohort (Figure 2A,B; see also Figure S2A for corresponding scatterplots). This broader group increased the sample size from 53 to 107, facilitating a more robust analysis with stronger statistical power. The top analytes making up the expanded CSF cohort, C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217, remained highly correlated with [¹⁸F]GTP1 WCG and TMP SUVR in the larger cohort (all ρ ≥ 0.58), and the strongest correlates were C2N‐eMTBR‐tau243 and p‐tau205 (ρ = 0.69 in TMP, ρ = 0.65 in WCG). Statistical comparisons of these correlation coefficients confirmed that the associations with tau PET were not significantly different among these top four analytes in the expanded cohort (Table S2). Correlations with amyloid PET were not assessed in the expanded CSF cohort because baseline amyloid PET data were not available for 45 of the 107 participants.

FIGURE 2.

(A) Spearman's rank correlation coefficients and the −log10(FDR adjusted p value) between each CSF tau analyte with [¹⁸F]GTP1 SUVR in the whole cortical gray matter and (B) meta‐temporal ROI from the expanded CSF cohort. Only analytes with correlations meeting the significance criteria (ρ > 0.4, p(FDR) < 0.05) are annotated.

To further assess the ability of CSF tau analytes to discriminate between subjects with low and high tau burden as measured by PET, ROC analyses were performed. Similar to the correlation rankings, C2N‐eMTBR‐tau243, p‐tau217, MTBR/243‐254, and p‐tau205 demonstrated the strongest classification performance, with AUC values ranging from 0.80 to 0.87 in the CSF discovery cohort (Figure S3A) and 0.79 to 0.83 in the expanded cohort (Figure S3B). In plasma, p‐tau217 and p‐tau181 showed similar AUC values (AUC: 0.77 to 0.78), and lower AUC values were observed for total tau and BD tau (AUC: 0.62 to 0.65), consistent with the ranking observed for their correlation strengths with tau PET (Figure S3C).

The clinical relevance of these top correlates was evaluated in the expanded CSF cohort by examining their associations with clinical severity at baseline as well as longitudinal clinical changes (Figure 3). Across most baseline clinical measures, C2N‐eMTBR‐tau243 demonstrated the highest ρ values, followed by [¹⁸F]GTP1, MTBR/243‐254 or p‐tau205, and p‐tau217. Association patterns with whole‐brain volume, represented by vMRI, differed from the other clinical scores, with the strongest correlations observed for tau PET, followed by intermediate comparable correlations, which were calculated for p‐tau217, p‐tau205, and C2N‐eMTBR‐tau243, and the lowest correlation was observed for MTBR/243‐254. Similar comparisons were performed for associations between baseline biomarkers and estimated slopes of clinical progression. [¹⁸F]GTP1 showed the strongest correlations with clinical change for all measures except ADCS‐ADL, where the C2N‐eMTBR‐tau243 demonstrated a higher correlation. For changes in MMSE, CDR‐SB, ADAS‐Cog11, and vMRI, C2N‐eMTBR‐tau243 resulted in ρ values most similar to those observed for tau PET.

FIGURE 3.

Forest plots presenting Spearman's rank correlations with 95% confidence intervals between tau biomarkers and clinical scores, as well as whole‐brain volume at baseline (left panel) and longitudinal change in clinical scores and MRI‐derived atrophy (right panel) in the expanded CSF cohort. Tau biomarkers included CSF C2N‐eMTBR‐tau423, MTBR/243‐254, p‐tau205, p‐tau217, and [¹⁸F]GTP1 meta‐temporal SUVR. Clinical scores included MMSE, CDR‐SB, ADCS‐ADL, and ADAS‐Cog11. Longitudinal change was defined as the subject‐specific slope estimated from linear regression models of cognitive scores over time.

3.3. Plasma tau and tau PET in prodromal to mild AD

Next, we evaluated the correlations between [¹⁸F]GTP1 and amyloid PET with plasma total tau, p‐tau181, p‐tau217, and BD tau from Tauriel prodromal to mild AD plasma. Plasma p‐tau217 and p‐tau181 showed the strongest correlations with [¹⁸F]GTP1 SUVR from the WCG and TMP regions (ρ = 0.46 to 0.53) (Figure 4A,B). BD tau showed higher correlations with [¹⁸F]GTP1 SUVR than total tau, but neither ρ value met the meaningfulness threshold of 0.4. Similar to CSF, none of the plasma tau analytes were significantly and meaningfully correlated with amyloid PET (Figure 4C).

FIGURE 4.

(A) Spearman's rank correlation coefficients and the −log10(FDR adjusted p value) between each plasma tau analyte with [¹⁸F]GTP1 SUVR in the whole cortical gray matter and (B) meta‐temporal ROI from the plasma cohort. (C) Spearman's rank correlation coefficients and the −log10(FDR adjusted p value) between each plasma tau analyte with amyloid PET Centiloids from the plasma cohort.

3.4. Contributions of amyloid and tau PET to tau‐level variation

Next, we investigated the variance of CSF tau levels in the discovery CSF cohort explained by amyloid PET and tau PET by quantifying the relative importance, represented by partial R ², of each variable from multiple linear regression models. In these models, tau species were treated as the dependent variable for each model, while tau PET and amyloid PET were treated as independent variables, and age, sex, and APOE genotype were included as covariates. Tau PET showed the highest explanatory powers for changes in CSF C2N‐eMTBR‐tau243 and p‐tau205 levels (partial R ² [pR ²] = 0.39 and 0.30), while the explanatory power by amyloid PET was relatively low (pR ² = 0.05 and 0.04) (Figure S4A). Variance in p‐tau217, PRD/164‐180, MTBR/243‐254, and N‐terminal tau levels were also explained moderately well by tau PET (pR ² = 0.25), but the explanatory power by amyloid PET or other contributions was higher compared to C2N‐eMTBR‐tau243 and p‐tau205 for MTBR/243‐254 and PRD/164‐180, respectively. The same analysis could not be completed in the expanded CSF cohort because 45 of the 107 participants did not have amyloid PET data at baseline. The variance in BD tau, total tau, p‐tau181, and p‐tau217 levels in the plasma cohort was evaluated. Tau PET accounted for the greatest proportion of variance in p‐tau217 most highly in plasma (pR ² = 0.22), followed by p‐tau217 (pR ² = 0.18) and total tau (pR ² = 0.07) (Figure S4B), while BD tau lacked meaningful explanatory power (R ² < 0.1).

3.5. Regional correlations of top CSF correlates

We evaluated correlations between the top CSF tau PET correlate levels with [¹⁸F]GTP1 SUVR in 58 ROIs from the Hammers atlas (Figure 5). ⁴² Similar correlation patterns were observed across C2N‐eMTBR‐tau243, p‐tau205, and p‐tau217, with the strongest correlations with [¹⁸F]GTP1 SUVR in regions like the left and right fusiform gyrus (ρ = 0.64 to 0.72, all p(FDR) < 0.05), parahippocampal gyrus (ρ = 0.60 to 0.67, all p(FDR) < 0.05), and entorhinal cortex (ρ = 0.61 to 0.65, all p(FDR) < 0.05), ¹³ , ⁴³ which are typically impacted by pathology earlier in the disease course. In comparison, regional correlations for MTBR/243‐254 with [¹⁸F]GTP1 SUVR were strongest in regions typically impacted later in the disease course by tau pathology: posterior temporal lobe (ρ = 0.60 and 0.55 for left and right, p(FDR) < 0.05 for both), the left superior parietal gyrus (ρ = 0.61, p(FDR) < 0.05), the angular gyrus (ρ = 0.60 and 0.59 for left and right, p(FDR) < 0.05 for both), and the left middle frontal gyrus (ρ = 0.61, p(FDR) < 0.05). ¹³ Across all analytes, correlations in the hippocampus were lower than in adjacent medial temporal regions.

FIGURE 5.

Spearman's rank correlation coefficients between the top CSF correlates (C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau205, and p‐tau217) levels and regional [¹⁸F]GTP1 SUVR from the Hammers atlas. Regions displayed have p value < 0.05 after FDR correction and are displayed on a MNI template brain.

In addition, we further evaluated independent and additive explanatory values from CSF tau species for tau PET signal in the expanded CSF cohort given the larger sample size compared to the discovery CSF cohort. C2N‐eMTBR‐243 and p‐tau205 were consistently selected as predictors of tau PET signal in TMP and WCG ROIs using LASSO regression (Figure 6A). High selection stability was observed for p‐tau205 (TMP: 94%, WCG: 96%) and C2N‐eMTBR‐tau243 (TMP: 96%, WCG: 85%), supporting their robustness across resampling. These variables were subsequently incorporated into hierarchical regression models. In the final TMP model, p‐tau205 demonstrated the strongest independent association with [¹⁸F]GTP1 TMP SUVR (β = 0.56), followed by C2N‐eMTBR‐tau243 (β = 0.28) (Figure 6B). In the final WCG model, p‐tau205 and C2N‐eMTBR‐tau243 remained independently associated with tau PET signal (β = 0.39 and 0.15, respectively) though effect sizes were smaller than those observed for TMP.

FIGURE 6.

(A) Bootstrap selection frequencies (500 resamples) for candidate predictors of baseline tau PET signals in the meta‐temporal (TMP; in blue) and whole cortical gray (WCG; in green) regions of interest (ROIs) derived from LASSO regression. Frequencies indicate the proportion of resamples in which each predictor resulted in a non‐zero coefficient. Selection frequency values are annotated to the left of each bar. (B) Bar plot representing standardized regression coefficients from final linear models predicting baseline tau PET signal in TMP (blue) and WCG (green) ROI. Models included age, sex, APOE ε4 status, and MMSE as covariates, with C2N‐eMTBR‐tau243 defined as the primary tau predictor and additional tau species selected by LASSO with bootstrap stability. Estimated coefficient values are annotated to the side of each bar.

4. DISCUSSION

In our comprehensive evaluation of 20 CSF and four plasma tau species in prodromal to moderate AD, CSF C2N‐eMTBR‐tau243, MTBR/243‐254, p‐tau217, and p‐tau205 emerged as the strongest correlates of [18F]GTP1 PET tau burden. The differences between their correlation coefficients were not statistically significant (Table S2), suggesting comparable potential as fluid surrogates for tau pathology, though this may partially reflect limited statistical power.

These findings can be interpreted within established mechanistic frameworks distinguishing phosphorylation sites (e.g., residue 217), which reflect early amyloid‐associated alterations in soluble tau turnover, from MTBR‐derived fragments and truncation‐linked species (e.g., C2N‐eMTBR‐tau243), which more directly track aggregation‐prone processes and tangle burden. ³⁰ , ³¹ , ⁴⁴ , ⁴⁵ , ⁴⁶ Because our symptomatic cohort consisted exclusively of amyloid‐positive individuals, both phosphorylation‐dependent markers (p‐tau217, p‐tau205) and aggregation‐proximal species (C2N‐eMTBR‐tau243, MTBR/243‐254) demonstrated strong global associations with tau PET, suggesting a convergence of soluble tau dysregulation and insoluble tangle accumulation at this disease stage. However, measuring this expanded panel head to head within the same harmonized cohort revealed that these similar global correlations did not imply biological equivalence. Specifically, C2N‐eMTBR‐tau243, p‐tau205, and p‐tau217 showed stronger associations with medial temporal regions classically affected early in AD, whereas the closely related MTBR/243‐254 peptide demonstrated stronger correlations in neocortical regions involved at later stages. By evaluating these analytes systematically, we demonstrate that distinct tau proteoforms capture complementary, biologically distinct aspects of tau pathology rather than fully redundant measures.

Despite these biological distinctions, the comparable performance of p‐tau205 and C2N‐eMTBR‐tau243 in terms of correlation strength, explanatory power, and predictive power in our study is noteworthy, especially given recent literature suggesting that MTBR fragments may more uniquely reflect insoluble tau aggregates than N‐ or mid‐terminal phosphorylated tau. ⁴⁷ Several factors likely contribute to this convergence in our prodromal to moderate AD cohort. First, because our study exclusively enrolled amyloid‐positive individuals, we are observing a disease stage where both soluble tau metabolism and insoluble tangle formation are actively progressing. While p‐tau species typically rise early in response to amyloid, p‐tau205 is a “later‐rising” marker linked specifically to the AT8 epitope found in mature NFTs, aligning its behavior with MTBR species in symptomatic patients. ³¹ , ³⁵ , ⁴⁸

Assay characteristics may also contribute to the similar performance observed. Both C2N‐eMTBR‐tau243 and p‐tau205 were measured using highly optimized, targeted platforms – IP‐LC‐MS/MS and Elecsys RPA, respectively. In a symptomatic population with sufficient tangle burden, these high‐performance assays likely capture the underlying pathology with similar fidelity. This is supported by our regional analysis, where both analytes showed nearly identical correlation patterns with [¹⁸F]GTP1 PET in early‐accumulating regions like the fusiform and entorhinal cortex.

This regional alignment reflects the broader temporal dynamics of AD. Recent work suggests that CSF biomarkers in AD may become abnormal in stages, starting with AB42/40, followed by p‐tau217, p‐tau205, and C2N‐eMTBR‐tau243 in subsequent stages, reflecting, to some extent, the progressive accumulation of tau in the cortex. ⁴⁹ In our study, C2N‐eMTBR‐tau243, p‐tau217, and p‐tau205 were more strongly associated with [¹⁸F]GTP1 in brain regions known to accumulate tau pathology early in AD. Among these markers, C2N‐eMTBR‐tau243 demonstrated stronger correlations (ρ = 0.72) compared to p‐tau217 and p‐tau205 (ρ = 0.67 to 0.69). The highest correlations for MTBR/243‐254 were uniquely observed in regions typically affected in later stages of AD. However, we observed a notable regional exception: Across all markers, [18F]GTP1 SUVR correlated less strongly in the hippocampus. This discrepancy may reflect either off‐target tracer signal in the adjacent choroid plexus ⁵⁰ or altered tracer affinity as tangles advance in pathological maturity. ⁵¹ , ⁵² , ⁵³ Further analysis is warranted to fully understand the long‐term relationship between these key fluid biomarkers and tau PET. ⁴⁹

Our findings align with and extend previous studies examining fluid biomarkers. ³⁰ , ³³ , ⁵⁴ First, the strong performance of p‐tau217 is consistent with previous reports demonstrating its utility as a potential surrogate candidate for tau PET. ³⁰ , ³² , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁵⁵ Second, our findings on CSF p‐tau205 as a top candidate is especially relevant given p‐tau205's known role in NFT formation as part of the AT8 epitope used in neuropathological staging. ⁴⁸ , ⁵⁶ Furthermore, it is encouraging to observe C2N‐eMTBR‐tau243 as a top performer, which adds to prior publications showing that tau fragments closer to or within the MTBR region, which mediate aggregation, may be particularly informative for tau pathology. ⁴⁷ These findings support the growing focus on these analytes as more specific indicators of pathological tau accumulation. Although C2N‐eMTBR‐tau243 and MTBR/243‐254 are structurally similar, the modest differences in their correlation with [¹⁸F]GTP1 SUVR may be attributed to differences in the fragment spans, truncation patterns, and assay methodology. C2N‐eMTBR‐tau243 spans residues 243 to 256, extending two amino acids beyond MTBR/243‐254, which may impact fragment ionization and detectability by MS, as well as proteolytic susceptibility. In addition, as reported in previous studies, ³¹ , ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ tau truncation has been implicated in tau pathogenesis and shown to promote tau aggregation. Different tau fragments, such as C2N‐eMTBR‐tau243 and MTBR/243‐254, may be products of distinct proteolytic processing pathways and may therefore result in fragment‐specific relationships with regional tau PET signals. This is also supported by our feature selection analysis (Figure 5A), where different selection frequencies were observed for the two analytes. C2N‐eMTBR‐tau243 was consistently retained as a candidate predictor of tau PET signal across modeling iterations, while MTBR/243‐254 was selected far less frequently. This difference in selection stability suggests these analytes are not equivalent in relation to tau PET signals, supporting the notion that they represent biologically and analytically distinct tau fragments rather than redundant measures of equivalent species. In addition, C2N‐eMTBR‐tau243 was assayed using IP‐LC‐MS/MS, a highly sensitive and targeted technique that was coupled with IP enrichment and quantification using isotopically labeled internal standards. In contrast, MTBR/243‐254 was detected using an unbiased DIA‐MS approach, which lacks these enrichment techniques and is expected to differ in sensitivity, dynamic range, and quantification accuracy.

For shared analytes measured in both CSF and plasma (p‐tau181, p‐tau217), we observed consistent patterns of correlation with tau PET and higher correlations for p‐tau217 in CSF (plasma: ρ = 0.53 and CSF: ρ = 0.63) and correlations still higher but more similar for p‐tau181 (plasma: ρ = 0.46 and CSF: ρ = 0.51). The preservation of signal across compartments, albeit with different magnitudes, suggests these markers reflect robust biological processes that can be detected throughout the body. This cross‐fluid consistency supports the biological relevance of the markers and the potential to leverage CSF for discovery with the eventual move to a blood‐based biomarker. While we observed a lack of significant and meaningful correlation (ρ > 0.04, p(FDR) > 0.05) between fluid tau measures and amyloid PET, this is likely due to our study design, which only included amyloid‐positive subjects, restricting our ability to assess relationships across the full spectrum of amyloid burden. This methodological difference is critical when reconciling our results with prior work, such as Barthélemy et al., ³³ which reported strong associations between CSF p‐tau species and amyloid PET. Those earlier findings were driven by cohorts that included amyloid‐negative individuals, providing a necessary baseline for correlation that is absent in our trial‐eligible, amyloid‐positive population, where a plateaued amyloid signal limits statistical power to detect a rank‐order relationship. This design choice, while limiting certain analyses, allows for a focused examination of tau dynamics in the presence of established amyloid pathology, reflecting the population most relevant for clinical trials and therapeutic intervention. In addition, the small sample size of the discovery CSF cohort may contribute to differences in correlation strength compared to previous studies where stronger relationships with amyloid PET were reported for p‐tau species in similar cohorts of amyloid‐positive individuals with prodromal to mild AD. ³³ , ⁶¹

Importantly, the intended context of use for these biomarkers should be clearly defined. Based on the cross‐sectional design and amyloid‐confirmed clinical trial populations studied here, our findings support prioritization of select CSF tau species as candidate surrogate markers of tau PET burden within symptomatic, amyloid‐positive AD. The data do not establish utility for primary diagnosis, differential diagnosis, or general population screening, nor do they support replacement of tau PET for all clinical or regulatory purposes. Rather, the most immediate application is in clinical trial settings, where fluid biomarkers may serve as scalable tools for participant stratification, disease staging within established amyloid‐positive cohorts, and prioritization of candidates for plasma translation. Longitudinal and interventional studies will be required to determine whether these markers can reliably monitor disease progression or pharmacodynamic effects over time. ⁶² , ⁶³ , ⁶⁴

Future longitudinal studies across diverse populations and disease stages are needed to determine whether changes in these fluid biomarkers precede or follow tau PET signal alterations. Translating these leading CSF biomarkers into blood‐based assays will be critical for clinical trial integration. This includes further evaluation of plasma eMTBR‐tau243, which has shown stronger correlations with tau PET compared to ratios of p‐tau217 and p‐tau205 to unphosphorylated tau. ⁶⁵

AUTHOR CONTRIBUTIONS

Study design: Julie Lee and Felix Yeh. Sample and data acquisition: Julie Lee, Ellen Casavant, Stephen Schauer, Tobias Bittner, Matthew Meyer, Philip Verghese, Venky Venkatesh, Tim West, Joel Braunstein, Emilia Calderon, Melanie Mandl, Veronica Anania, Sandra Sanabria Bohórquez, and Felix Yeh. Data analysis: Julie Lee, Emily Olafson, Balazs Toth, Sandra Sanabria Bohórquez, and Felix Yeh. Data interpretation: Julie Lee, Emily Olafson, Balazs Toth, Sandra Sanabria Bohórquez, and Felix Yeh. Manuscript writing – original draft: Julie Lee, Emily Olafson, and Felix Yeh. Manuscript writing – review and editing: All authors.

CONFLICT OF INTEREST STATEMENT

COBAS and ELECSYS are trademarks of Roche. All other product names and trademarks are the property of their respective owners. Matthew Meyer, Philip Verghese, Venky Venkatesh, Tim West, and Joel Braunstein are full‐time employees of C2N Diagnostics LLC, St. Louis, Missouri, USA. Melanie Mandl and Anna Bayfield are full‐time employees of Roche Diagnostics GmbH, Penzberg, Germany. All other authors are full‐time employees of Genentech Inc. (member of the Roche group). Edmond Teng is listed as a co‐inventor on the patent for semorinemab. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants in this study and/or their legally authorized representatives provided signed informed consent.

Supporting information

Supporting Information

Supporting Information