Cerebrospinal fluid biomarkers predict frontotemporal dementia trajectory

Participants

Baseline CSF biomarker concentrations, longitudinal clinical data, and longitudinal imaging data were studied in 90 patients with frontotemporal dementia (40 bvFTD, 24 svPPA, 26 nfvPPA) (Table 1). CSF data from 49 healthy age‐matched controls was also included for comparison. Data were collected from three medical centers (University of California, San Francisco [UCSF], Massachusetts General Hospital [MGH], Mayo Clinic, Rochester [MCR]) via the Frontotemporal Lobar Degeneration Neuroimaging Initiative (FTLDNI) and the UCSF program project grant in frontotemporal dementia. The Institutional Review Board (IRB) of each contributing site previously approved study protocols. Patients were diagnosed by a multidisciplinary consensus panel using current (at the time of study enrollment) consensus diagnostic criteria for bvFTD,11, 12 svPPA,13 and nfvPPA.13 Baseline CSF samples were matched with baseline clinical and radiographic data within 90 days of sample collection. The majority (84%) of cases also had one or more additional clinical data and radiographic data points after their baseline visit (Table 1). The average follow‐up interval was between 9 and 10 months. Intervals were not regular within or between patients. There was not a statistically significant difference in follow‐up interval lengths between individual FTD cohorts (Kruskal Wallis P < 0.05). In order to account for the irregularity of follow‐up intervals, time was used as a continuous variable (years from baseline CSF value) in all linear mixed‐effect models.

CSF Assessment

NfL, t‐tau, p‐tau, and Aβ _1‐42 concentrations were measured from frozen CSF samples collected between May 2009 and May 2015. CSF samples were collected via lumbar puncture into sterile polypropylene tubes, using a previously described protocol.8 Within 30 min of collection, CSF samples were centrifuged at 2000g at room temperature (20–25°C) for 5 min, aliquoted into 500 μL cryovials and stored at −80°C, until further analysis. All fluid biomarkers were measured from baseline CSF, previously collected during or prior to earliest clinical and MR imaging assessment. NfL concentrations were also measured in follow‐up CSF samples (within 2 years after baseline) in 27 patients with FTD (14 with bvFTD, 8 with nfvPPA, and 5 with svPPA) to provide longitudinal data on NfL change. Longitudinal data was not collected for t‐tau, p‐tau, or Aβ _1‐42. All biomarkers were measured in duplicate (twice concurrently) to ensure coefficients of variance<25%, and the average concentration was used in our analyses. Aβ _1‐42, total tau (t‐tau), and hyperphosphorylated tau (p‐tau) concentrations were measured with the INNO‐BIA AlzBio3 platform (Fujirebio, Belgium). Neurofilament light‐chain (NfL) levels were quantified using the Uman Diagnostics enzyme‐linked immunosorbent assay kit (Umea, Sweden). All NfL samples were performed using a 1:4 dilution as specified in the Uman Diagnostic kit instructions, and were otherwise performed using previously described kit methods.8 NfL concentrations were first assessed on 7/26/2012 and 4/24/2013 by a clinical research organization, Bristol‐Myers Squibb (BMS), in an initial cohort of 27 patients (14 with bvFTD, 8 with nfvPPA, and 5 with svPPA). Data obtained from BMS (along with contemporaneous clinical and volumetric data) were previously used in separately published cross‐sectional analyses.8 In order to expand our sample size, NfL concentrations in a larger cohort of patients (26 with bvFTD, 22 with nfvPPA, 19 with svPPA, 49 controls) were measured on 7/15/2015 and 1/12/16 by the Fagan lab, Washington University, St. Louis. The Fagan lab also performed CSF Aβ _1‐42, t‐tau and p‐tau assays for all 139 subjects in our study on 7/15/2015 and 1/12/16. The laboratory site of biomarker assessment was included as a covariate in all statistical analyses of NfL levels. Among patients with FTD, there was no difference between NfL concentrations calculated by the Fagan lab and BMS (P‐value = 0.6).

Classification of patients according to causative proteinopathy

To examine whether the predictive value of NfL is different for FTLD‐tau vs. FTLD‐TDP, we assigned cases into groups based on the best available data. Diagnostic information from postmortem brain examination, using previously described methods,14 was used if available. If autopsy data was not available (whether it was not obtained or the patient was still alive), but an FTD‐causing mutation was identified, gene status was used to identify the putative FTLD major molecular class. If neither autopsy nor mutation data were available, we assigned some patients to a proteinopathy group if they developed either a motor neuron disease secondary clinical phenotype (strongly associated with FTLD‐TDP) or a Steele‐Richardson progressive supranuclear palsy (PSP) secondary clinical phenotype15 (strongly associated with FTLD‐tau). Using this enrichment approach, we were able to assign a likely underlying pathology in 35 cases (20 with FTLD‐tau and 15 with FTLD‐TDP) (Table 1). Autopsy data was available in 12 cases with bvFTD, 11 cases with nfvPPA and one case with svPPA. Within the bvFTD group, six patients had FTLD‐tau pathology at autopsy (two with Pick's disease, one with agyrophilc grain disease, one with PSP, and two with corticobasal degeneration), and six patients showed FTLD‐TDP (one with sporadic FTLD‐TDP type A, one with FTLD‐TDP Type A secondary to GRN mutation, three with sporadic FTLD‐TDP Type B and motor neuron disease, and one with unclassifiable FTLD‐TDP secondary to C9ORF72 expansion). An additional two patients with bvFTD were classified as tauopathy based on MAPT mutations and six additional living patients had mutations associated with FTLD‐TDP (five with C9ORF72 expansions and one with a GRN mutation). Two patients with bvFTD due to C9ORF72 expansion also developed motor neuron disease later in their course. One additional patient with bvFTD and motor neuron disease was presumed to have FTLD‐TDP pathology. Within the nfvPPA cohort, 10 cases had FTLD‐tau pathology at autopsy (two with Pick's disease, one with an unclassifiable four repeat tauopathy, six with corticobasal degeneration, and one with PSP pathology), and one had FTLD‐TDP Type B pathology with motor neuron disease. Two living patients with nfvPPA were presumed to have FTLD‐tau pathology because of accompanying clinical features sufficient to also meet clinical research criteria for the Steele‐Richardson PSP syndrome.15 One patient with svPPA had autopsy data that confirmed the presence of FTLD‐TDP type B pathology. A total tau/Aβ _1‐42 ratio of 0.52 was used to define CSF criteria for AD.16 Only five cases met CSF criteria for AD (four cases with bvFTD, one case with svPPA, and no cases with nfvPPA).

Clinical assessment

General functional severity was assessed using the FTLD modified clinical dementia rating scale, sum of boxes score (CDR‐FTLD SB), and established measure of clinical decline in FTD.17 Patients were also assessed using the Functional Assessment Questionnaire (FAQ),18 the Clinician's Global Impression Scale (CGI),19 the Schwab and England Activities of Daily Living scale (SEADL),20 the Neuropsychiatric Inventory (NPI),21 the Mini‐Mental State Examination (MMSE),22 the California Verbal Learning Test, Second Edition‐Short Form (CVLT),23 digit span (backward and forward),24 a modified Trail‐making task,25 the Stroop color naming and inhibition tasks,26 phonemic fluency (D‐words/minute),25 semantic fluency (animals/minute),27 the 15‐item Boston Naming Test (BNT),28 the short version of the Peabody Picture Vocabulary Test‐Revised (PPVT),29 the Pyramids and Palm Trees test of sematic access,30 and a copy of the Modified Rey Benson Figure.25

Structural MRI analyses

We obtained volumetric magnetic resonance images at UCSF and MGH, using a 3T Siemens Tim Trio system equipped with a 12‐channel head coil. Images obtained at MCR were acquired on a 3T GE MRI scanner equipped with an 8‐channel head coil. T1 acquisition and processing were performed under previously described methods.31 Brain volumes were calculated for specific brain regions at each time point by transforming a standard parcellation atlas32 into ICBM space and summing all modulated gray matter within each specified parcellated region. Lobar composites were used to reduce the number of regions of interest (ROI) in our analysis. Right and left frontal lobe composites were created using the sum of volumes within the following regions: anterior cingulate (caudal and rostral), frontal pole, insula, lateral orbitofrontal cortex, medial orbitofrontal cortex, middle frontal gyrus, pars opercularis, pars orbitalis, pars triangularis, superior frontal gyrus, and precentral gyrus. We created right and left temporal lobe composites using the sum of volumes in the following regions: banks of the superior temporal sulcus, entorhinal cortex, fusiform gyrus, middle temporal gyrus, parahippocampal cortex, superior temporal gyrus, temporal pole, and transverse temporal gyrus. We then summed volumes in the resulting right frontal, left frontal, right temporal and left temporal lobe composites to produce total frontotemporal composite ROI. We created a global frontotemporal volumetric ROI by summing the frontal and temporal composite ROIs.

DTI analyses

DTI acquisition and processing were performed using previously described methods using scans from UCSF.33 DTI ROIs were obtained, using the ICVM‐DTI‐81 white matter labels and tract atlas.34 We again created composite ROIs. The right and left frontal composite ROIs represented the average fractional anisotropy (FA) of tracts conveying fibers that entered and exited the frontal lobes, including the superior longitudinal fasciculus and cingulum bundle. The right and left temporal composite ROIs represented the average FA from the fornix, inferior longitudinal fasciculus, sagittal stratum, and hippocampal portion of the cingulum. FA within the genu of the corpus callosum, the right uncinate fasciculus, and left uncinate fasciculus were also analyzed but not included in lobar composites. We created a global frontotemporal DTI ROI by averaging FA in the genu of the corpus callosum, the bilateral uncinated fasciculus, and the right and left constituent ROIs previously used to create our frontal and temporal composite ROI.

Statistics

We performed basic descriptive statistics for demographic data and CSF protein levels. The normality of continuous cross‐sectional data was assessed via the skewness kurtosis test.35 Non‐normally distributed values (CDR‐FTLD SB, NfL, t‐tau, Aβ _1‐42, t‐tau/Aβ _1‐42 ratio) were compared using nonparametric methods, including the Kruskal–Wallis test followed by post hoc Wilcoxon pairwise rank sum testing. Normally distributed measures (p‐tau, p‐tau/t‐tau, age) were compared between cohorts with Analysis of Variance (ANOVA), followed by a post hoc pairwise Tukey test. A Bonferroni correction was used to correct for multiple pairwise comparisons across five differing CSF measures (α = 0.05/5 = 0.01). In light of previously published studies,36, 37 we also used a receiver operating characteristic (ROC) analysis to investigate the utility of p‐tau/t‐tau ration in detecting FTLD‐TDP pathology en vivo.

We hypothesized that CSF protein biomarkers (NfL, t‐tau, p‐tau, Aβ _1‐42) would predict change in three primary selected measures of disease severity: the CDR‐FTLD SB, frontotemporal volume, and frontotemporal fractional anisotropy (FA). In order to compare the effect sizes between CSF markers, the mean and standard deviations (SD) (Table  1) for our entire FTD cohort (n = 90) were used to create Z‐score values for NfL, total tau, p‐tau, and Aβ _1‐42 for each patient (thus allowing effect sizes to be conveyed in terms of standard deviation from the mean). We subsequently used these Z‐scores as continuous variables for cross‐sectional regression models and linear mixed effect models (LME) performed in this study.

In order to leverage data from cases with three or more data points, we used LME models to estimate interactions between biomarkers and time (both continuous variables) in determining disease measures. Biomarker level, age, and sex were included as fixed effects. The lab of NfL measurement and total intracranial volume (TIV) were also included as fixed effects in models containing NfL and volumetric data, respectively. We modeled time as a fixed effect and allowed each patient to have a random intercept to account for differing baseline severity. In order to assess whether the value of biomarkers differed by clinical cohort, we first tested three‐way interactions between biomarkers, diagnosis (a fixed effect), and time. Given the apparent interaction between diagnosis and biomarker predictors, we elected to assess each canonical FTD variant separately in our two‐way LME models. We selected a conservative approach to correct for multiple comparisons to focus on easily reproducible findings for future clinical trial use. A Bonferroni correction (α = 0.05/3 = 0.017) was applied to analyses of our three primary measures. As a sensitivity analysis, all statistically significant findings were reanalyzed after excluding cases with possible contributory non‐FTLD co‐pathology, including three patients with bvFTD and possible AD biomarkers, six cases with bvFTD and evidence of motor neuron disease, one case with nfvPPA and motor neuron disease, and one patient with svPPA and positive AD biomarkers. In order to provide a clinically useful presentation of data, we calculated annual clinical change (using LME models with no interaction) above and below the median CSF concentration for every biomarker that predicted CDR‐FTLD. Follow up post hoc analysis (using previously discussed two‐way interaction LME models) then assessed predictors of change in 30 additional clinical and radiographic measures of disease decline (Tables S1–S3). A Bonferroni correction was applied (α = 0.05/30 = 0.0017) to results for our 30 follow up measures of disease severity. Finally, we modeled interactions between underlying pathology (a fixed effect), biomarkers, and time in determining disease severity in a cohort of cases with bvFTD or nfvPPA phenotypes. The magnitude of the estimated three‐way interaction coefficient was used to assess the relative predictive value of biomarkers in FTLD‐tau relative to FTLD‐TDP. After each LME model analysis a postestimation of standardized residuals was performed to ensure the normality of their distribution. All statistically significant findings were rerun as a sensitivity analysis with patients with motor neuron disease (severe patients with bvFTD, one with nfvPPA) and/or CSF criteria for AD (four cases with bvFTD, one case with svPPA) excluded from analysis.

Linear regression was used to estimate and assess the statistical significance of relationships between CSF protein concentrations and each of our three primary measures of interest within each of the three FTD clinical cohorts of interest. Baseline CDR‐FTLD was not normally distributed in nfvPPA, so a log transformed value was used in our nfvPPA regression models. In cases with longitudinal NfL data, linear regression was also used to estimate the change in CSF NfL over time. We utilized NfL change (NfL value 2 – NfL value 1) as the dependent variable and duration between time points (in years) as an independent variable. Longitudinal t‐tau, p‐tau, and Aβ _1‐42 values were not available for analysis. Age and sex were included as covariates in all regression models. The lab of measurement and total intracranial volume (TIV) were also included as covariates in regression models containing NfL and volumetric data, respectively.

Patient demographics

Basic demographic information is presented in Table  1. Patients with nfvPPA were older than patients with bvFTD, though none of the FTD groups differed from age‐matched controls. Patients with bvFTD had a higher CDR‐FTLD SB score than patients with svPPA or patients with nfvPPA. Additionally, patients with svPPA had a higher FTLD‐CDR SB than patients with nfvPPA. Patients with svPPA tended to have higher CSF NfL levels than other FTD variants, but this difference was not statistically significant. Controls had lower CSF NfL than all three disease variants in all pairwise comparisons (Fig. 1A). CSF p‐tau/t‐tau ratio was lower in patients with bvFTD and nfvPPA compared to healthy controls (Fig. 1F). Additionally, patients with presumed FTLD‐TDP pathology tended to have a lower CSF p‐tau/t‐tau ratio than patients with presumed FTLD‐tau (P = 0.02) (Table  2). A p‐tau/t‐tau ratio less than or equal to 0.29 gave an AUC of 0.7 (SD 0.47, P = 0.003), a specificity of 70% specific and sensitivity of 65.3% in detecting FTLD‐TDP from our cohort with discernable pathology (Table S1). CSF t‐tau/Aβ _1‐42 ratio tended to be lower in controls compared to patients with bvFTD, svPPA, and nfvPPA, but these findings were not significant after adjusting for multiple comparisons. There was no difference in CSF t‐tau, p‐tau, or Aβ _1‐42 concentrations between any of our clinical (controls, bvFTD, nfvPPA or svPPA) or pathologic cohorts.

CSF biomarkers correlate with clinical severity and brain volume

In patients with bvFTD, high baseline CSF NfL was linearly related to high baseline CDR‐FTLD SB (1.36 CDR units/SD NfL, 95% CI [0.41, 2.31], P = 0.006, R ²=0.23) and low baseline frontotemporal volume (−8.4 × 10³ mm³/SD NfL, 95% CI [−15 × 10³, 2.0 × 10³], P = 0.012, R ² = 0.65). In patients with svPPA, high baseline t‐tau (1.44 units/SD NfL, 95% CI [0.29, 2.59], P = 0.016, R ²=0.26) was linearly related to CDR‐FTLD SB.

Baseline NfL and p‐tau predict longitudinal change in bvFTD

Higher baseline levels of CSF NfL, total tau, and p‐tau predicted a faster rate of CDR‐FTLD SB worsening in patients with bvFTD (Table 3, Fig. 2A–C). However, after excluding cases with possible non‐FTLD co‐pathology, only CSF NfL and p‐tau remained statistically significant predictors of CDR‐FTLD progression rate. On average within the bvFTD cohort, patients above the median CSF NfL concentration (4.1 × 10³ pg/mL) increased by 3.4 CDR‐FTLD units annually, patients above the median p‐tau concentration (19.8 pg/mL) increased by 2.9 CDR‐FTLD units annually, and patients below the mean for both biomarker experienced minimal clinical change annually (Table 4). CSF p‐tau/t‐tau ratio was not a useful predictor of clinical change in bvFTD.

Higher CSF NfL levels also predicted more rapid worsening in other clinical measures (Table S2) including FAQ total, MMSE total, CVLT immediate and delayed recall, Stroop interference, and BNT total (even after excluding cases with a CSF tau/Aβ _1‐42 ratio over 0.52). Additionally, higher baseline CSF p‐tau also predicted more rapid decline in the SEADL, after excluding possible comorbid Alzheimer's pathology.

Higher baseline CSF NfL and lower baseline Aβ _1‐42 predicted a faster rate of frontotemporal volume loss (Table 3, Fig. 2E and H). High baseline CSF NfL and low baseline CSF Aβ _1‐42 also predicted a faster rate of decline in frontotemporal FA in patients with bvFTD (Fig. 2I and L). The relationship between NfL and Aβ _1‐42 with frontotemporal imaging measures remained statistically significant even after excluding cases with potential co‐pathology, including patients with motor neuron disease and positive AD biomarkers (Table S2). CSF p‐tau/t‐tau ratio was not a useful predictor of MR imaging changes in bvFTD.

Baseline NfL and total tau predict change in nfvPPA

Higher baseline CSF NfL and higher t‐tau each predicted a faster rate of CDR‐FTLD SB worsening in patients with nfvPPA (Table 3, Fig. 2A and B). On average, within the nfvPPA cohort patients above the median t‐tau NfL concentration (5.17 x 10³pg/mL) increased by 2.4 CDR‐FTLD units annually, patients above the median t‐tau concentration (75.8 pg/mL) increased by 1.9 CDR‐FTLD units annually, and patients below the mean for either biomarker experienced minimal clinical change annually (Table 4).

In post hoc analyses, higher CSF NfL levels predicted a faster rate of worsening in CGI severity, the Stroop inhibition task, and sematic fluency (Table S3). Additionally, high CSF t‐tau predicted a faster rate of change in CGI Severity.

High baseline CSF NFL and total tau were each associated with a faster rate of frontotemporal volume loss in patients with nfvPPA (Table 3, Fig. 2E and F). High baseline CSF NfL and t‐tau also predicted a faster rate of decline in frontotemporal FA in patients with nfvPPA (Fig. 2I and J).

In patients with nfvPPA, the predictive value of CSF NfL remained even after excluding possible co‐pathology (a single case with motor neuron disease). CSF p‐tau and p‐tau/t‐tau ratio were not a useful predictors of clinical change in nfvPPA.

Baseline CSF biomarkers in svPPA

Within the svPPA cohort, the predictive value of CSF biomarkers did not meet our pre‐specified criteria for statistical significance after correction for multiple comparisons (Table 3, Table S4). High baseline NfL did, however, tend to be associated with a faster rate of decline in MMSE score, right temporal volume, right uncinated fasciculus FA, and corpus callosum genu FA. Additionally, high p‐tau tended to be associated with a faster rate of decline in lexical fluency.

CSF biomarkers have less predictive value in svPPA and FTLD‐TDP

In models containing all three clinical FTD variants, clinical diagnosis and CSF biomarkers interacted to determine the rate of disease change over time (Table S5). In general, baseline CSF NfL predicted a greater degree of clinical and volumetric change in bvFTD and nfvPPA compared to svPPA. Baseline CSF t‐tau also predicted a greater degree of clinical change in nfvPPA and bvFTD compared to svPPA. Finally, Aβ _1‐42 predicted a greater degree of volumetric and DTI change in bvFTD compared to svPPA. These three‐way interaction model results were used to validate our approach to analyze each FTD clinical cohort separately in our two‐way interaction models.

High baseline CSF NfL predicted a particularly large increase in the annual rate of CDR‐FTLD SB and frontotemporal volume loss in patients with FTLD‐TDP relative to FTLD tau (Table S5). High baseline CSF p‐tau also predicted a particularly large increase in the annual rate of CDR‐FTLD SB change in FTLD‐tau relative to FTLD‐TDP. High Aβ _1‐42 was also associated with relatively higher rates of frontotemporal volume loss in FTLD‐tau compared to FTLD‐TDP. Given the high burden of motor neuron disease within the FTD‐TDP cohort, we were unable to exclude patients with this co‐pathology in sensitivity analyses.

CSF NfL levels over time

The relationship between time and NfL change did not meet our prespecified criteria for statistical significance in the 27 patients with FTD and longitudinal CSF NfL data (β = 298 pg/mL/year, 95% CI [−4261, 4858], P = 0.893, R ² = 0.02) (Fig. 3). The relationship between time and NfL change was also not statistically significant when bvFTD (n = 14, β = 964 pg/mL/year, 95% CI [−7535, 9465], P = 0.8, R ² = 0.09), svPPA (n = 5, β = 659 pg/mL/year, 95% CI [−4321, 5639], P = 0.34, R ² = 0.96), and nfvPPA (n = 8, β = −193 pg/mL/year, 95% CI [−7058, 6672], P = 0.942, R ² = 0.06) cohorts were analyzed separately. A single case with bvFTD (a 67‐year‐old ambidextrous male with no known pathogenic mutation and a 5‐year prior disease duration) was noted to experience a distinct increase in CSF NfL (+16,605 pg/mL) over 9 months. His level of clinical worsening (+3 CDR‐FTLD SB units), frontotemporal volume loss (−9602 mm³), and frontotemporal FA loss (−0.01 FA) were higher than expected for his baseline CSF biomarker levels (NfL 1149 pg pg/mL, Z score = −1.14; t‐tau 38.3 pg/mL, Z score = −1.2; p‐tau 15.7 pg/mL, Z score = −0.80; Aβ _1‐42 484.96, z score = 0.43) when compared to predicted rates of change (Fig.  2). His baseline CSF t‐tau/Aβ _1‐42 ratio was 0.08 (not suggestive of emerging AD co‐pathology). We did not collect data on longitudinal t‐tau, p‐tau, or Aβ _1‐42 change.