Normalization of CSF pTau measurement by Aβ40 improves its performance as a biomarker of Alzheimer’s disease

Background Alzheimer’s disease (AD)-related tauopathy can be measured with CSF phosphorylated tau (pTau) and tau PET. We aim to investigate the associations between these measurements and their relative ability to predict subsequent disease progression. Methods In 219 cognitively unimpaired and 122 impaired Alzheimer’s Disease Neuroimaging Initiative participants with concurrent amyloid-β (Aβ) PET (18F-florbetapir or 18F-florbetaben), 18F-flortaucipir (FTP) PET, CSF measurements, structural MRI, and cognition, we examined inter-relationships between these biomarkers and their predictions of subsequent FTP and cognition changes. Results The use of a CSF pTau/Aβ40 ratio eliminated positive associations we observed between CSF pTau alone and CSF Aβ42 in the normal Aβ range likely reflecting individual differences in CSF production rather than pathology. Use of the CSF pTau/Aβ40 ratio also increased expected associations with Aβ PET, FTP PET, hippocampal volume, and cognitive decline compared to pTau alone. In Aβ+ individuals, abnormal CSF pTau/Aβ40 only individuals (26.7%) were 4 times more prevalent (p <  0.001) than abnormal FTP only individuals (6.8%). Furthermore, among individuals on the AD pathway, CSF pTau/Aβ40 mediates the association between Aβ PET and FTP PET accumulation, but FTP PET is more closely linked to subsequent cognitive decline than CSF pTau/Aβ40. Conclusions Together, these findings suggest that CSF pTau/Aβ40 may be a superior measure of tauopathy compared to CSF pTau alone, and CSF pTau/Aβ40 enables detection of tau accumulation at an earlier stage than FTP among Aβ+ individuals.


Background
Extracellular amyloid-β (Aβ) peptides in cortical Aβ plaques and intracellular phosphorylated tau protein as neurofibrillary tangles are key hallmarks of Alzheimer's disease (AD) that can be measured in vivo with positron emission tomography (PET) imaging and biofluid markers including plasma and cerebrospinal fluid (CSF) assays. The relationship between CSF Aβ and Aβ PET in AD has been widely reported [1][2][3][4][5][6][7][8], but relationships between CSF tau and tau PET are uncertain [9][10][11][12][13]. Recent studies reported that individuals with abnormal CSF phosphorylated tau (pTau) were more prevalent than individuals with abnormal tau PET only [14], and that abnormal tau PET but not CSF pTau was related to cognitive decline [15], suggesting that CSF and PET may not be interchangeable indices of tau pathology.
There are also remaining technical questions involved in measurement of CSF biomarkers. Elevated (abnormal) CSF pTau has been observed in cases with exceptionally elevated CSF Aβ 42 in the Aβ− range [7,16]. Positive correlations between these measurements in the Aβ− range are likely not AD-related but are instead due to individual variability in CSF production. This would suggest that abnormal CSF pTau in individuals with elevated CSF Aβ 42 lack a pathological basis and instead reflect disease-invariant CSF increases that would be observed across all CSF markers. To address this phenomenon, use of the CSF Aβ 42 /Aβ 40 ratio has been proposed over CSF Aβ 42 alone [6,7,[17][18][19], since Aβ 40 is most abundant Aβ species in CSF [19,20], and expected to increase due to higher overall Aβ production but not sensitive to AD [21][22][23][24][25][26][27][28][29]. We hypothesize that a similar adjustment of CSF pTau using CSF Aβ 40 may reduce noise and improve associations with other biomarkers.
In this study, we used Alzheimer's Disease Neuroimaging Initiative (ADNI) participants to explore the utility of a CSF pTau/Aβ 40 ratio to reduce noise in pTau measurements and improve associations with downstream markers of AD progression. We then examined the biological plausibility of this biomarker in relation to regional 18− Flortaucipir (FTP) PET as well as subsequent tau PET and cognitive changes.

Participants
Data used in this study were obtained from the ADNI database (ida.loni.usc.edu; specific datasets used in this study are named below). The ADNI study was approved by institutional review boards of all participating centers, and written informed consent was obtained from all participants or their authorized representatives. In total, 219 cognitively unimpaired (CU) elderly adults, 91 mild cognitive impairment (MCI), and 31 AD patients with concurrent (acquisition interval within 1 year) Aβ PET ( 18 F-florbetapir (FBP) or 18 F-florbetaben (FBB)), CSF Aβ 40 , Aβ 42 and pTau 181 , FTP tau PET, structural MRI, and cognitive test were included in this study.

PET and MRI imaging
PET data was acquired in 5-min frames from 50 to 70 min (FBP), 90-110 min (FBB), and 75-105 min (FTP) post-injection (http://adni-info.org). PET and structural MRI scans were downloaded from the Laboratory of NeuroImaging (LONI) (ida.loni.usc.edu) and processed with Freesurfer V5.3.0. All fully pre-processed PET scans were co-registered to the structural MRI scan that was closest in time to the baseline PET. Regions of interest (ROIs) were defined on each structural MRI scan using Freesurfer (V5.3.0) and used to extract regional FBP, FBB, and FTP measurements from the co-registered PET images as described previously [30,31].
For FTP (BERKELEYAV1451_05_12_20.csv), composite Temporal-metaROI (including entorhinal, parahippocampal, fusiform, amygdala, inferior temporal, and middle temporal) [32] and entorhinal cortex SUVRs were calculated using inferior cerebellar cortex intensity normalization [31]. To define FTP SUVR thresholds, we carried out ROC analyses with Temporal-metaROI and entorhinal SUVR values using the Youden index classifying 280 Aβ PET− ADNI CU participants and 183 Aβ PET+ ADNI MCI and AD patients as the endpoint (Supplemental Figs. 1-4). This resulted in a threshold of 1.25 for the Temporal-metaROI and 1.21 for entorhinal cortex. Among these 463 ADNI participants for the definition of tau PET cutoffs, 217 (47%) participants were included in the following analyses of this study. We also examined alternative thresholds for these regions defined by the mean + 2SD of 280 Aβ PET-ADNI CU participants. These resulted in more conservative thresholds of 1.34 for the Temporal-metaROI and 1.31 for entorhinal cortex. In total, 34% of 341 participants had longitudinal FTP data. FTP slope (ΔFTP, SUVR units per year) was calculated based on longitudinal FTP data for each individual using linear mixed effects (LME) model, including the following independent variables: time, APOE-ε4 status, age and gender, and a random slope and intercept. Since white matter intensity normalization has shown less variability for longitudinal tau PET changes [33][34][35], we calculated FTP slopes using a white matter reference region.
Hippocampal volume (HCV) (mm 3 ) was calculated across hemispheres from the structural MRI scan that was closest in time to the baseline PET scan and for subsequent MRI scans using Freesurfer, and adjusted by estimated intracranial volume (ICV) using the regression approach [36]: adjusted HCV (aHCV) = HCV − 0.0017 × (ICV -1 498858), where 0.0017 and 1498858 represent the correlation coefficient between HCV and ICV, and the mean of ICV in Aβ− 323 ADNI CU participants. In total, 41% of 341 participants had longitudinal aHCV data. aHCV slope (ΔaHCV, mm 3 units per year) was calculated based on longitudinal aHCV data for each individual using LME model, including the following independent variables: time, APOE-ε4 status, age, gender and education, and a random slope and intercept.
CSF Aβ 40 , Aβ 42 , and pTau CSF Aβ 40 , Aβ 42 , and pTau were analyzed by the University of Pennsylvania ADNI Biomarker core laboratory using the fully automated Roche Elecsys and cobas e 601 immunoassay analyzer system [16,37]. CSF data (UPENN-BIOMK10_07_29_19.csv) were downloaded from ADNI website. A threshold for abnormal CSF pTau was defined as ≥22 pg/mL based on an ROC analysis using the Youden index classifying 320 Aβ PET− ADNI CU participants and 429 Aβ PET+ ADNI MCI and AD patients as the endpoint . We also defined an alternative threshold of ≥31 for CSF pTau which was based on the mean + 2SD of CSF pTau in 320 Aβ PET− ADNI CU participants. We calculated the CSF pTau/Aβ 40 ratio threshold as ≥0.0012 according to the same ROC approach classifying 169 Aβ PET− CU participants and 161 Aβ PET+ MCI and AD patients as the endpoint (Supplemental Figs. 7-8), and the alternative threshold was ≥0.0014 based on the mean + 2SD of the CSF pTau/Aβ 40 ratio in 169 Aβ PET− ADNI CU participants. Among these 749 ADNI participants for the definition of CSF pTau, 212 (28%) participants were included in the following analyses of this study. Among these 329 ADNI participants for the definition of CSF pTau/ Aβ 40 , 201 (61%) participants were included in the following analyses of this study.

Cognition
The Delayed Recall portion of the Alzheimer's Disease Assessment Scale (ADASSCORES.csv and ADAS_ ADNIGO23.csv downloaded at April 28, 2020), the delayed recall score on the logical memory IIa subtest from the Wechsler Memory Scale, the digit symbol substitution test score from the Wechsler Adult Intelligence Scale-Revised (NEUROBAT.csv downloaded at April 28, 2020), and the MMSE total score (MMSE.csv downloaded at April 28, 2020) were transferred to standard z scores (using the mean values of ADNI CU participants). Preclinical Alzheimer Cognitive Composite (PACC) scores [38] were calculated by combining these 4 cognitive z scores to one composite score. In total, 59% of 341 participants had longitudinal PACC data. PACC slope (ΔPACC) was calculated for each participant based on longitudinal PACC scores using LME model, including the following independent variables: time, APOE-ε4 status, age, gender and education, and a random slope and intercept.

Statistical analysis
Normality of distributions was tested using the Shapiro-Wilk test and visual inspection of data. Data are presented as median (interquartile range (IQR)) or number (%). Baseline characteristics were compared between Aβ− and Aβ+ groups by using a two-tailed Mann-Whitney test or Fisher's exact test.
In order to evaluate the feasibility of using CSF pTau/ Aβ 40 as an alternative to CSF pTau, we first used generalized linear models (GLM) to examine the relationships of CSF Aβ 40 with Aβ PET and tau PET to confirm that CSF Aβ 40 is not related to AD biomarkers, and subsequently investigated the cross-sectional associations between CSF Aβ 42 , pTau and pTau/Aβ 40 , and controlling for APOE-ε4 status, diagnosis, sex, and age. A false discovery rate of 0.05 using the Benjamini-Hochberg approach was employed for 35 regions.
The slopes of FTP SUVR, aHCV, and PACC post baseline CSF collection were calculated using LME models over time from the first measurement point post baseline CSF collection (time = 0) to the last measurement point for each participant. The time variable is anchored to the baseline CSF measurement. In order to study whether elevated CSF pTau/ Aβ 40 is more related to the progression of AD than high CSF pTau, we also used GLM models to investigate the associations of CSF pTau and pTau/Aβ 40 with Aβ PET, tau PET, aHCV, ΔaHCV, PACC, and ΔPACC, controlling for APOE-ε4 status, diagnosis, sex, age, and education. Since there was a time difference between baseline CSF collection point and the first measurements of FTP SUVR, aHCV, and PACC post baseline CSF collection, we included these time differences in the GLM models. Because we found use of the CSF pTau/Aβ 40 ratio abolished the positive correlation between CSF pTau and Aβ 42 among Aβ PET− range (see Fig. 1c, d in "Results") and improved the associations with Aβ PET, tau PET, aHCV, ΔaHCV, PACC, and ΔPACC (see Fig. 2 in "Results"), we used this ratio in subsequent analyses.
In order to investigate the predictive effect of baseline Aβ PET, CSF pTau/Aβ 40 , and FTP on subsequent ΔFTP and ΔPACC, we used these variables at baseline to predict subsequent ΔFTP and ΔPACC in participants with longitudinal tau PET and PACC data respectively. In order to explore temporal relationships between Aβ and tau, we also examined the sequential associations between baseline Aβ PET, CSF pTau/Aβ 40 ratio, FTP, and ΔFTP in Aβ+ participants using latent variable modeling (R; Lavaan package) [39].
For GLM models with non-Gaussian distribution outcomes (Aβ and tau PET), we used a "log" link function in the Gaussian family to study the associations between predictor and outcome. Spearman's rank correlation coefficient (rho) was calculated between predictor and outcome. We selected p < 0.05 as the significance level. All statistical analyses were performed in the statistical program R (v3.6.2, The R Foundation for Statistical Computing).

Demographics
Measurements were acquired between September 21, 2015 and April 9, 2020. Demographics can be found in Table 1. In total, 341 participants had contemporaneous CSF Aβ 40 , Aβ 42 and pTau, Aβ PET, tau PET, structural MRI, and PACC cognitive score. At baseline, Aβ+ participants were significantly older and had greater CSF pTau, CSF pTau/Aβ 40 and Temporal-metaROI FTP SUVR, lower aHCV, lower cognitive test scores, and a higher percentage of APOE-ε4 carriers than Aβ− participants. Longitudinally, 116, 139, and 202 participants had Use of CSF Aβ 40 to adjust CSF pTau CSF Aβ 40 was not associated with Aβ PET or tau PET regardless of Aβ PET status (Fig. 1a, b). Before normalizing to CSF Aβ 40 , CSF pTau was positively (standardized β (β std ) = 0.59[95% confidence interval (CI), 0.48, 0.71]) associated with CSF Aβ 42 in Aβ PET− participants, whereas no association was found in Aβ+ participants (Fig. 1c). We also verified that there was a similar positive association between CSF pTau and CSF Aβ 42 analyzed with mass spectrometry rather than the Roche Elecsys immunoassay in a partially overlapping (9.8%) sample of 384 Aβ− participants (Supplemental Fig. 9). After normalizing CSF pTau using CSF Aβ 40   Notably, the association with Aβ PET increased from rho value 0.51 when using CSF pTau alone to 0.67 using the CSF pTau/Aβ 40 (Fig. 2a, b). Likewise, the association with tau PET increased from rho value 0.43 when using CSF pTau alone to 0.46 using the CSF pTau/Aβ 40 (Fig. 2c,  d). We also compared CSF pTau and CSF pTau/Aβ 40 in terms of their associations with other measures of neurodegeneration biomarkers and cognition in order to further investigate the validity of CSF pTau/Aβ 40 . CSF pTau/Aβ 40 but not CSF pTau was negatively associated with baseline aHCV (Fig. 2e, f), and the association with aHCV slope increased from rho value − 0.18 when using CSF pTau alone to − 0.38 using the CSF pTau/Aβ 40 (Fig. 2g, h). The association with baseline PACC and PACC slope increased from rho values − 0.33 and − 0.24 when using CSF pTau alone to − 0.45 and − 0.39 using the CSF pTau/Aβ 40 respectively (Fig. 2i, l).
Based on these findings, CSF pTau/Aβ 40 was used to represent tauopathy in CSF instead of CSF pTau for all subsequent analyses.
We also found that CSF pTau and CSF pTau/Aβ 40 were both more strongly associated with Aβ PET than they were with tau PET (Fig. 2a-d).
Regions with significant associations between CSF pTau/ Aβ 40 and tau PET CSF pTau/Aβ 40 was significantly associated with tau PET SUVRs in all the 35 ROIs, and the strongest  association regions were within the Temporal-metaROI region (Fig. 3). We repeated these analyses in Aβ−, Aβ+, CU, and non-demented (CU and MCI) participants. The results were similar for Aβ+ participants (supplemental Fig. 10A), whereas no association was found for Aβ− participants. Similar features were observed for CU and non-demented (CU and MCI) participants (supplemental Fig. 10B-C). Because the strongest associations between CSF pTau/Aβ 40 and tau PET were within the Temporal-metaROI (Fig. 3), which has been commonly used to detect tau deposition in brain [40][41][42][43][44][45][46], temporal tau PET (Temporal-metaROI FTP SUVR) was selected to represent tau deposition for further analyses unless otherwise noted.
The conservative cutoffs of CSF pTau, CSF pTau/Aβ 40 , entorhinal tau PET, and temporal tau PET were higher and defined fewer "T+" individuals, while the results of concordance of different biomarkers were substantially the same as the initial cutoffs (Supplemental Figs. [12][13]. with subsequent tau PET increase (ΔFTP) in Aβ+ participants (Fig. 5a-c). In contrast, no predictive effect was found in Aβ− participants.

Discussion
This study had several primary findings: (1) use of a CSF pTau/Aβ 40 ratio reduced noise in pTau likely introduced by individual variability in CSF production rates, and increased associations with Aβ PET, tau PET, hippocampal volume, and cognition compared with CSF pTau alone.
Our motivation to adjust CSF pTau measurements was based on our observation that Aβ PET-negative individuals had abnormal ("positive") CSF pTau that correlated positively with high ("normal") CSF Aβ 42 (Fig. 1c), suggesting that these elevated measurements reflect high CSF total production rate but not abnormal tau. Similar patterns of elevated pTau and CSF Aβ 42 in the negative range that are presumably artifactual have been observed in other recent studies from ADNI, BIOFINDER, and Washington University [7,16], and with CSF data analyzed with mass spectrometry (Supplementary Fig. 9) and immunoassays. CSF pTau/Aβ 40 appears to be a compelling strategy for improving sensitivity to CSF tau pathology, since this approach reversed the biologically implausible association between CSF pTau and Aβ 42 and improved associations with downstream markers of AD progression compared with CSF pTau alone. Because CSF Aβ 40 was not associated with PET measures of either Aβ or tau (Fig. 1a, b) and is not elevated in AD [21][22][23][24][25][26][27][28][29], its use as a normalization variable is unlikely to bias estimates of CSF pTau. This strategy is in line with recent work supporting use of CSF Aβ 42 /Aβ 40 instead of CSF Aβ 42 alone [6,7,[17][18][19], and use of CSF pTau/tTau instead of CSF pTau [47]. However, our results did not exclude other possibilities for the enhanced associations between CSF pTau/Aβ 40 and downstream markers of AD progression. For example, a few studies [48][49][50][51] have reported that CSF Aβ 40 may decrease in cognitively impaired individuals, which may thereby increase the CSF pTau/Aβ 40 ratios of cognitively impaired individuals. In addition, one animal study [52] observed that CSF Aβ 40 may increase in the earliest phase of Aβ accumulation in mouse models, which may delay the increase of CSF pTau/Aβ 40 in the preclinical stage of AD. We found only trend-level decreases in CSF Aβ 40 in Aβ− unimpaired and Aβ+ impaired groups relative to Aβ+ unimpaired individuals (data not shown), but it is possible that early and late changes in CSF Aβ 40 may contribute to the taurelated effects we observed.
Associations between CSF pTau/Aβ 40 and tau PET were stronger in ROIs in the temporal lobe than other areas such as frontal and occipital lobes that accumulate tau in later stages of disease [53,54], consistent with our observation and recent studies [14,15,55] that CSF tauopathy is an early marker of tau pathology. The strongest associations were within the medial and lateral temporal regions that overlapped with a tau composite region (Temporal-metaROI) reported previously as well as a "Braak III/IV" like ROI [40,41,45,56]. Notably, the relationship between CSF pTau/Aβ 40 and tau PET was primarily driven by Aβ PET positivity and less influenced by clinical diagnosis (Supplementary Fig. 10), which could also reflect a greater range of tau pathology in Aβ+ individuals and a stronger relationship between Aβ and tau than between tau and clinical symptoms [57,58]. Consistent with the present study, Chhatwal et al. [10] reported a significant association between CSF pTau and tau PET in limbic regions of the temporal lobe in CU elderly adults. However, two studies [9,12] did not find significant association between CSF pTau and tau PET in CU individuals, perhaps due to methodological factors such as sample size and the use of CSF pTau alone rather than the CSF pTau/Aβ 40 ratio.
In line with our findings, one recent study [15] also reported that CSF pTau mediated the association between Aβ PET and tau PET, and higher CSF pTau was associated with faster tau PET increase rates in cognitively impaired individuals. Unlike this study, we found baseline tau PET was also related to the tau PET rate. The discrepancy may be explained by the larger sample size and the use of white matter reference for longitudinal tau PET in the present study. In the mediation analyses, two significant CSF pTau/ Aβ 40 -linked pathways were identified, which explained 70% of the association between Aβ PET and longitudinal brain tau accumulation among Aβ+ individuals.
Finally, consistent with three recent reports [14,15,66], we found that tau PET was more predictive of subsequent cognitive decline than CSF tau among Aβ+ individuals, suggesting brain tau may reflect a later tau stage closer to cognitive decline than CSF tau on the Alzheimer's continuum. Interestingly, previous comparisons of CSF and PET measurements of Aβ were analogous in showing that cognitive decline is more related to Aβ PET than CSF Aβ [1,3,67,68]. We also noticed that higher CSF pTau/Aβ 40 was significantly related to faster longitudinal cognitive decline in amyloid-negative individuals. No previous studies reported the association between CSF pTau and cognitive decline in amyloidnegative individuals, which should be cautious to interpret this result and may need to be validated in other samples.
This study has several limitations. The CSF pTau/Aβ 40 threshold was derived from the existing sample of ADNI participants and only pTau 181 was available in the ADNI sample at this time, so it would be helpful to validate the findings in other samples and with other phosphorylation sites (i.e., pTau 217 [47,69]) and tau PET ligands. Furthermore, only 9% (31/341) of the participants in this study were AD patients and the longitudinal observation was of relatively short duration, so it would be helpful to confirm those findings using additional participants and extended longitudinal data. Finally, one possible explanation for the differences we observed between tau PET and CSF pTau measurements is that CSF pTau may reflect Aβ in addition to tau pathology. Our observation that both CSF pTau and CSF pTau/Aβ 40 had stronger associations with Aβ PET than they did with tau PET (Fig. 2a-d) is consistent with this possibility, but further pathology studies are needed to verify this interpretation.

Conclusions
In summary, we found that the use of a CSF pTau/Aβ 40 ratio improves the sensitivity to detect CSF tau by adjusting for individual differences in CSF production. Furthermore, although PET and CSF measures of tau are broadly concordant in the majority (76%) of individuals when measured dichotomously, our findings support recent work [14] indicating that CSF and PET measures of tau may not be interchangeable in the A/T/N research framework [70]. Among amyloid-positive individuals, higher tauopathy measured with CSF and PET is related to faster tau accumulation, while tau PET was more predictive of subsequent cognitive decline than CSF tau. Taken together, these findings suggest that the interchangeability of PET and CSF measures of tau likely depends on the goals of the study, the phase of AD being studied, and the clinical characteristics of the population.
Authors' contributions T.G contributed to the study design, drafting and editing of the manuscript, data and statistical analysis, and interpretation of results; D.K contributed to acquiring data and editing the manuscript; R.L contributed to interpretation of results and editing the manuscript. L.M.S. and J.Q.T contributed to acquiring data, interpretation of results, obtaining funding, and editing the manuscript; W.J.J and S.M.L contributed to acquiring data, interpretation of results, obtaining funding, editing the manuscript, and study supervision. The author(s) read and approved the final manuscript.

Funding
Not applicable.

Availability of data and materials
The dataset supporting the conclusions of this article is available in the ADNI repository (ida.loni.usc.edu). Derived data is available from the corresponding author on request by any qualified investigator subject to a data use agreement.
Ethics approval and consent to participate All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the principles of the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Formed written consent was obtained from all participants at each site of ADNI.

Consent for publication
Not applicable.