2.1 Study participants
ADNI is designed to develop clinical, imaging, genetic, and biochemical biomarkers for the early detection and tracking of AD. The participants are volunteers aged 55–90 years with normal or impaired cognition. Further information can be found at http://www.adni-info.org/ and in previous reports [13–15]. At baseline, each participant underwent an in-person interview of general health and functional ability, followed by a standardized assessment including a battery of neuropsychological tests. Follow-up data were collected during evaluations at sequential intervals of approximately 12 months. ADNI was approved by institutional review boards of all participating institutions, and written informed consent was obtained from all participants or their guardians.
In the present study, a total of 965 participants who had baseline measurements of CSF PGRN and AD core biomarkers, as well as longitudinal measurements of cognitive functions were included. Among these individuals, 228 had measurements of CSF neuroinflammatory markers.
2.2 Classification methods
The classification methods were in line with those proposed by Marc Suárez-Calvet et al. [10, 16] according to 2018 NIA-AA “research framework” for AD diagnosis [17]. In brief, participants were categorized into specific groups based on both biomarker profile (as described by the A/T/N scheme [18]) and clinical stage (as defined by the clinical dementia rating [CDR] score). The A/T/N scheme includes 3 biomarker groups: “A” aggregated amyloid pathology (as indicated by CSF Aβ1−42), “T” aggregated tau (as indicated by CSF p-tau181), and “N” neurodegeneration or neuronal injury (as indicated by CSF t-tau). “A+” participants refer to those with CSF Aβ1−42 < 976.6 pg/ml; “T+” those with CSF p-tau181 > 21.8 pg/ml; and “N+” those with CSF t-tau > 245 pg/ml. The CSF biomarker statuses established by these cutoffs were proven to be highly concordant with PET classification in ADNI [19]. Given that T and N groups were highly correlated, we merged them together to facilitate the analyses, producing a TN group: “TN+” indicates T + or N + and “TN-” indicates T- and N-.
2.3 CSF measurements of PGRN, neuroinflammatory markers, and AD core biomarkers
CSF procedural protocols in ADNI have been described previously [20]. CSF PGRN was measured by a previously reported sandwich immunoassay using the Meso Scale Discovery platform [21]. All CSF samples were distributed randomly across plates and measured in duplicate. All the antibodies and plates were from a single lot in order to exclude variability between batches. The mean intraplate coefficient of variation (CV) was 2.2%; all duplicate measures had a CV < 15%. PGRN levels were corrected by inter-batch variation and corrected values were used for analyses (for the method see Appendix 1).
CSF AD core biomarkers, including Aβ1−42, p-tau181, and t-tau were analyzed by the electrochemiluminescence immunoassays (ECLIA) Elecsys on a fully automated Elecsys cobas e 601 instrument and a single lot of reagents for each of the three measured biomarkers (provided in UPENNBIOMK9.csv file). These measurements are for explorative research use only. A total of 8 types of CSF neuroinflammatory markers, including four anti-inflammatory markers (sTNFR1, sTNFR2, TGF-β1, and IL-10) and four pro-inflammatory markers (ICAM1, VCAM1, IL-6, and IL-7) were measured, using commercially available multiplex immunoassays (Millipore Sigma, Burlington, MA). All samples were run in duplicate along with six standards on each plate. Samples were normalized across plates using CSF standard values. Precision of each analyte was calculated using inter-plate CV < 15%.
2.4 Cognitive measures
Global cognitive function was reflected by the total scores of Alzheimer’s Disease Assessment Scale (ADAS). Composite scores for executive and memory functions were constructed and validated by referring to the neuropsychological batteries [22, 23]. Specifically, the indicators of executive functions include Category Fluency, WAIS-R Digit Symbol, Trails A & B, Digit Span Backwards, and clock drawing. The indicators of memory function include relevant items of the Rey Auditory Verbal Learning Test (RAVLT), ADAS, Logical Memory, and Mini-Mental State Examination (MMSE). The CDR score was used to represent the clinical stage: “0” represents normal cognition, “0.5” represents very mild dementia, and “1” represents mild dementia.
2.5 Statistical analysis
Before the analyses, values of CSF markers as dependent variables were log10-transformed to achieve or approximate normal distributions as assessed by Kolmogorov-Smirnov test. All analyses were adjusted for age, gender, educational level, and CDR, except where specifically noted.
First, one-way analyses of covariance (ANCOVAs) were separately performed to examine the associations of CSF PGRN and neuroinflammatory markers with A/TN status, within the framework combining A/TN classification and clinical stage. Four comparisons were separately conducted for each CDR group, including A-/TN + vs. A-/TN-, A+/TN + vs. A+/TN- (the former two comparisons were to test the associations of tau-related neurodegeneration with biomarkers), A+/TN + vs. A-/TN+, and A+/TN- vs. A-/TN- (the latter two were for the associations with amyloid pathology). Next, multiple linear regressions were conducted to explore the associations of CSF PGRN (an independent variable) with neuroinflammatory markers (dependent variables) within the A/TN framework.
Further, we tested whether PGRN could modulate CSF Aβ1−42 via mediating specific neuroinflammatory markers. To achieve this, causal mediation analyses were conducted using linear regression models fitted based on the methods proposed by Baron and Kenny [24]. The direct effects, indirect effects, and the mediating proportion were estimated by Sobel’s test [25] with the significance determined using 10,000 bootstrapped iterations.
Finally, linear mixed effects (LME) models were used to estimate the longitudinal influences of CSF PGRN on cognitive functions. To facilitate the depiction, CSF PGRN was categorized into three groups (low, moderate, and high) using cutoffs of 1,396 pg/ml and 1,684 pg/ml according to the tertiles of the concentration. The LME models had random intercepts and slopes for time and an unstructured covariance matrix for the random effects, and included the interaction between time (continuous) and the dependent variable (PGRN) as a predictor.
Sensitivity analyses were conducted as follows. a) We repeated the analyses after excluding outlier values of CSF markers, defined as values situated outside the 3 standard deviations from the mean; b) rs5848 genotype of GRN gene, which was associated with PGRN levels, was added as a covariate in analyses with CSF PGRN as the dependent variable. The results barely changed after these analyses. R version 3.5.1 (major packages include “lm”, “ggplot2”, “mediate”, and “nlme”) and GraphPad Prism 7.00 software were used for statistical analyses and figure preparation. All tests were two-tailed, with a significance level of α = 0.05.