Introduction

Magnetic resonance imaging (MRI)-derived brain-age prediction stands as a significant biomarker of biological brain aging [1]. Brain-age is estimated using machine learning models that predict an individual’s chronological age based on neuroimaging data. These models are trained to learn the association between chronological age and cerebral features from structural MRI in healthy individuals. The difference between the predicted brain-age and the individual’s chronological age, referred as the brain-age delta, provides an estimate of accelerated (or decelerated) brain aging. Brain-age delta is assumed to reflect both past and ongoing neurobiological aging processes [1]. Accelerated brain aging as measured with brain-age delta has been found in Alzheimer’s disease (AD) and other neurodegenerative diseases [2, 3]. It has also been related to mental and cognitive health [2, 4] and to lifestyle factors [4, 5]. Furthermore, we have previously evaluated the associations between brain-age delta and biomarkers of AD and neurodegeneration, finding older brain-aging in more advanced stages of the pathological hallmarks of AD (amyloid-β and tau), as well as with markers of neurodegeneration and cerebrovascular disease [6].

While AD is characterized by the presence of amyloid-β (Aβ) plaques and tau tangles, the disease’s early stages involve a broader spectrum of pathophysiological changes. Among these changes, the aging brain exhibits pathological alterations often associated with chronic inflammation [7]. These changes include a reduction in certain neuronal populations, as well as decreases in dendritic and axonal arborization, post-synaptic densities, presynaptic markers, and cortical volume. Previous work has also evaluated associations between grey matter volume and brain metabolism with cerebrospinal fluid (CSF) biomarkers of several pathophysiological mechanisms altered in preclinical AD stages [8]. Glial reactivity, including that of microglia and astroglia, influence cognitive decline through different mechanisms, such as inflammation regulation, phagocytosis, synaptic pruning, neuroprotection, and blood-brain barrier integrity [9,10,11,12]. Microglia reactivity occurs early in AD and previous studies reported both detrimental and protective effects of microglia on AD progression [13,14,15,16]. Microglia expresses the innate immune receptor triggering receptor expressed on myeloid cell 2 (TREM2) [17], whose activation prompts microglial proliferation, phagocytosis, chemotaxis, and increasing survival and migration [9,10,11]. TREM2 receptor is cleaved in the cell surface and its soluble form (sTREM2) can be detected in cerebrospinal fluid (CSF) [11, 18]. Astroglia, particularly astrocytes, play a key role in the brain’s immune response. When they detect pathological changes, astrocytes become reactive, releasing pro-inflammatory molecules such as cytokines and chemokines [19]. Studies have consistently demonstrated that astrocytes are activated in response to the accumulation of misfolded protein aggregates [20]. Markers of astrocytic reactivity include CSF glial fibrillary acidic protein (GFAP), YKL-40, and S100 calcium-binding protein B (S100b), which increase with higher Aβ load and with age in the Aβ-positive individuals [19, 21,22,23]. We can also measure markers of chronic inflammation in the brain, such as interleukin 6 (IL6), which is increased with age and in AD. In addition to these mentioned mechanisms, synapse loss is closely related to cognitive decline and is an early process in AD [24]. Even though the number of synapses naturally decreases with age, this reduction is significantly accelerated in AD. The elimination of weaker or unnecessary synapses and the maintenance of stronger ones are crucial for balancing brain connectivity. Synapse-specific proteins involved in distinct synaptic pathways can be measured in CSF [24]. These include the postsynaptic protein neurogranin, which is involved in long-term potentiation (LTP) signaling, and presynaptic proteins such as synaptosomal-associated protein-25 (SNAP-25), which is crucial for neuronal survival, cognitive function, vesicular exocytosis, neurite outgrowth, and LTP; growth-associated protein-43 (GAP-43), which plays a key role in memory and information storage; synaptotagmin-1, essential for rapid synchronous neurotransmitter release in hippocampal neurons; and α-synuclein, which is involved in various synaptic processes. Literature has shown that CSF synaptic biomarkers increase early in the AD continuum [25]. Furthermore, several studies suggest that excessive Aβ oligomers can prompt microglia to excessively phagocytose and eliminate synapses, which may contribute to synapse loss in AD progression [26, 27].

Considering that age is the main risk factor for AD, it is essential to explore how distinct disease mechanisms, like glial reactivity and synaptic dysfunction, interact during Alzheimer’s disease development differently from the normal aging brain. Even though the relationship between these CSF glial and synaptic biomarkers with structural and metabolic alterations in the brain have been studied, no previous studies have investigated their relationship with biological brain aging, especially in the preclinical stages of the disease. This is a relevant pursuit, as it may contribute to informing preventive interventions before irreversible neuronal damage. As we previously saw brain-age delta associated with core AD biomarkers and knowing that inflammation and other disease mechanisms play a key role in brain aging and AD, we hypothesized that glial reactivity and synaptic dysfunction might be linked to biological brain aging in AD.

In the present study, we aimed to evaluate how brain-age delta is associated with markers representing pathophysiological processes that are altered in the early stages of AD, beyond amyloid and tau, and their impact on cognition. Additionally, we investigated how these biomarkers are associated with the AD signature, aiming to discern their effects on brain structure concerning both aging and AD. With this aim, we included 380 cognitively unimpaired individuals from ALFA+, for which we had baseline brain-age delta and AD signature, baseline CSF biomarkers of synaptic dysfunction (neurogranin, SNAP-25, GAP-43, synaptotagmin-1, α-synuclein), glial reactivity (sTREM2, GFAP, YKL-40, S100b); and chronic inflammation (IL-6), and cognition measured at baseline and 3.28 (SD 0.27) years later. We used ADNI to reproduce results from ALFA+. For this, 152 cognitively unimpaired and 310 mild cognitive impaired (MCI) participants from ADNI were also included, for which we had available measurements of brain-age delta, CSF sTREM2, and cognition measured 3.23 (SD 0.61) years later. We studied the cross-sectional linear associations between brain-age delta and CSF biomarkers. In addition, we explored the relationship between brain-age delta, CSF sTREM2, and longitudinal changes in cognition measured with the Preclinical Alzheimer’s Cognitive Composite (PACC) in both ALFA+ and ADNI cohorts.

Methods and materials

Participants

We included 380 individuals from the ALFA+ cohort to explore our main analyses. The ALFA+ cohort is a nested longitudinal study of the ALFA (for ALzheimer and Families) parent cohort [28]. The ALFA parent cohort was established as a research platform to understand the early pathophysiological alterations in preclinical AD and is composed of 2743 CU individuals (between 45 and 75 years) and enriched for family history of AD and genetic risk factors for AD [29]. All participants were enrolled in the ALFA (ALzheimer and FAmilies) study (Clinicaltrials.gov Identifier: NCT01835717). The study was approved by the Independent Ethics Committee “Parc de Salut Mar,” Barcelona, and all participants gave written informed consent.

In the present study, we included consecutive participants who met the following criteria: availability of T1-weighted MRI data that had previously passed a quality control QC procedure [30], availability of apolipoprotein E (APOE) categories, and availability of CSF measurements, which were acquired in less than a year from the MRI acquisition.

We also included individuals from the ADNI 1,2, GO, and 3 as a replication cohort [31]. As inclusion criteria, we selected CU and MCI individuals for which we had previously computed brain-age [6] and which had available CSF sTREM2 measurements (Ntotal=463, NCU = 152, NMCI = 310). MCI individuals were specified by a Clinical Dementia Rating = 0.5. ADNI individuals previously selected for brain-age prediction were based on availability of T1-weighted MRI, APOE categories, and CSF biomarkers for amyloid-β and p-tau pathology, which were acquired in less than a year from the MRI acquisition.

Image acquisition and preprocessing

The ALFA+ participants underwent high-resolution 3D T1-weighted MRI scanning using a 3 T Philips Ingenia CX scanner with the following parameters: TE/TR = 4.6/9.9 ms, Flip Angle = 8°, and voxel size = 0.75×0.75×0.75 mm. The acquired images were then subjected to segmentation using Freesurfer 6.0, followed by a rigorous quality control procedure to identify and eliminate any incidental findings [32] and segmentation errors [30]. After the initial FreeSurfer segmentation, the tissue regions were further parcellated into 183 different anatomical regions of interest (ROIs) using FreeSurfer’s widely-used cortical Desikan-Killiany [33] and subcortical aseg [34] labeling pipelines, which also underwent a meticulous quality control procedure [30].

For ADNI participants [31] MRI acquisition methods are described in more detail elsewhere (http://adni.loni.usc.edu/methods/documents/). In brief, most of the T1-weighted MR were MP-RAGE, acquired with 1.5 T or 3 T scanners. Images were segmented with Freesurfer 5.1 and 6.0 and subjected to a quality control procedure.

All volumes were residualized with respect to total intracranial volume (TIV) and scanning site, while all cortical thicknesses were residualized with respect to the scanning site, using linear models.

CSF sampling

In ALFA+, CSF biomarkers (Aβ42, Aβ40, neurogranin, YKL-40, GFAP, sTREM2, S100b, synaptotagmin-1 (SYT-1), IL-6 and α-synuclein) were measured with robust prototype assays as part of the NeuroToolKit, a panel of exploratory robust prototype immunoassays, on cobas® e 411 and cobas e 601 modules (Roche Diagnostics International Ltd, Rotkreuz, Switzerland). CSF p-tau181 was measured using the electrochemiluminescence Elecsys® Phospho-Tau (181 P) CSF immunoassay on a fully automated cobas e 601 module (Roche Diagnostics International Ltd, Rotkreuz, Switzerland). All measurements were performed at the Clinical Neurochemistry Laboratory, Sahlgrenska University Hospital, Mölndal, Sweden, by board-certified laboratory technicians who were blinded to diagnostic and other clinical data. CSF processing and storage in the ALFA+ study have been described previously [21, 35]. CSF biomarkers were grouped into: core AD biomarkers, glial biomarkers and markers of synaptic dysfunction and inflammation.

In ADNI, CSF samples were measured according to the kit manufacturer’s instructions and as described in previous studies [36], using the Elecsys β-amyloid(1–42) CSF and the Elecsys Phospho-Tau (181 P) CSF immunoassays on a cobas e 601 module at the Biomarker Research Laboratory, University of Pennsylvania, USA. CSF sTREM2 data was provided in the ADNI_HAASS_WASHU_LAB.csv file available in the ADNI database (variable “MSD_STREM2CORRECTED”). Measurements of CSF sTREM2 were based in an MSD assay and were previously reported [37]. A detailed description of the methods is found online (https://ida.loni.usc.edu).

Individuals from both ALFA+ and ADNI were classified by Aβ groups using previously stablished cohort-specific cut-offs of CSF Aβ42 (A + : 1098 pg/mL for ALFA+ and 880 pg/mL for ADNI) and p-tau (T + : 24 pg/mL) [21]. Due to the limited sample size in ADNI with available CSF Aβ40 and CSF sTREM2, this study assessed Aβ positivity in this cohort using Aβ42.

Brain-age prediction

Brain-age was estimated using our previously published prediction brain-age model [6]. In brief, we trained two separate XGBoost regressor models from the XGBoost python package (https://xgboost.readthedocs.io/en/) for females and males using 183 FreeSurfer volumes and thickness of the UK BioBank cohort as input (https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf). After training, we performed the brain-age prediction on the ALFA+ and on the ADNI cohort and applied a well-established age-bias correction procedure [38, 39]. This correction consists of deriving values of slope (α) and intercept (β) of a linear regression between age (Ω) and brain-predicted age (Y) on the UK Biobank. These values were used to correct the predicted brain-age in ALFA+ by applying: Corrected Predicted Brain Age = Predicted Brain Age + [Ω − (α × Ω + β)]. By subtracting chronological age from the Corrected Predicted Brain Age, we obtained the brain-age delta. We previously assessed the contribution of each brain variable utilized in training the model for predicting brain age [6].

AD signature

In ALFA+, we computed the Dickerson’s AD signature [40], which is a weighted average of thickness in AD-vulnerable structures, including medial temporal cortex, inferior temporal gyrus, temporal pole, angular gyrus, superior frontal gyrus, superior parietal lobule, supramarginal gyrus, precuneus, inferior frontal sulcus, primary visual cortex. Higher values of the AD signature reflect greater thickness in these regions, while lower AD signature values indicate more reduced thickness in these AD-vulnerable regions.

Cognitive measures

In ALFA+ sample, we computed the Preclinical Alzheimer’s Cognitive composite (PACC) [41,42,43] by standardizing individual raw scores into z-scores using the means and standard deviations from the biomarker-negative sample (A-T-N-) with n = 248 as reference group. Next, z-scores were averaged. The PACC was computed including the Total Immediate Recall from the Free and Cued Selective Reminding Test (FCSRT), the WMS-IV logical memory Total Delayed Recall, WAIS-IV subtest of Coding, and the Semantic Fluency Test [44].

We obtained the already computed modified PACC from ADNI, as explained elsewhere (https://adni.bitbucket.io/reference/pacc.html). Given the absence of FCSRT in ADNI, the Delayed Recall portion of the Alzheimer’s Disease Assessment Scale (ADAS) is used as a proxy. Due to the larger number of individuals with available Trails B test, we used the mPACCtrailsB, which uses (log transformed) Trails B as a proxy for the Digit Symbol Substitution Test (DSST).

We then computed the change of PACC Z-scores representing the longitudinal difference in the cognitive assessment between baseline visit and a follow-up cognitive assessment 3.28 ± 0.27 years later in ALFA+ and 3.23 ± 0.61 in ADNI, which were further annualized.

TREM2 rare genetic variants and genotyping

TREM2 rare variants (L211P, H157Y, T96K, D87N, R62H, R47H) were determined through genome-wide genotyping in ALFA+. DNA was obtained from blood samples through a salting out protocol. Genotyping was performed with the Illumina Infinium Neuro Consortium (NeuroChip) Array (build GRCh37/hg19) [45]. The polymorphisms analyzed herein were imputed using the Michigan Imputation Server with the haplotype Reference Consortium Panel (HRC r1.1 2016) [46] following default parameters and established guidelines. A full description of the genotyping, quality control and imputation procedures is available elsewhere [29]. Among the participants studied, had one or more known TREM2 rare variants (Supplementary Table 1). Results reported in the main text include all participants; sensitive analyses were conducted by excluding carriers of TREM2 rare variants. We applied the same criteria to identify these rare TREM2 variants among ADNI participants who had available CSF sTREM2 measurements.

Statistical analyses

Normality of the distribution for CSF biomarkers was assessed using the Kolmogorov-Smirnov test and visual inspection of histograms. CSF biomarkers which did not follow a normal distribution were log-transformed. Analyses were performed after removing outliers for each CSF biomarker, with outliers defined as those having a Z-score below −4 or above 4.

The first objective of our study was to investigate the associations between brain-age delta and CSF biomarkers of glial reactivity and synaptic dysfunction in all ALFA+ individuals. General linear modeling was used to test the association between each CSF biomarker (predictor) and brain-age delta (outcome) in ALFA+. Given the high cross-correlation between all CSF biomarkers, and especially the high correlation with CSF p-tau181 levels (Supplementary Fig. 1), we computed two models to assess the effects of AD pathology in the associations between brain-age delta and the CSF glial reactivity and synaptic dysfunction biomarkers. Model 1 included chronological age, sex and APOE-ɛ4 status as covariates. Model 2 included, in addition, CSF Aβ42 and p-tau181 measurements as covariates to account for their effects on the associations. We then tested the association between each CSF biomarker and the AD signature, a marker of AD-related neurodegeneration, by computing both Model 1 and Model 2. This was performed to assess if brain-age delta and AD signature are associated with the same or different pathophysiological mechanisms. Brain-age delta should measure age-specific brain changes, while AD signature should measure AD-specific brain changes. Sensitivity analyses of the main results were conducted by excluding individuals with coding TREM2 genetic rare variants and by including CSF Aβ40 as a covariate in both Models 1 and Models 2 (Supplementary Table 2) to account for overall rates of protein production and clearance through the CSF [47].

As per our second main objective, we assessed if the associations previously studied, both with brain-age delta and AD signature, differed in the presence of AD pathology as measured with CSF Aβ42. With this aim, we evaluated the interaction term “biomarker” × Aβ status while adjusting by age, sex, and APOE-ɛ4 status. As post hoc analyses, we stratified the participants into AD pathology status and computed a linear regression model between the significant CSF biomarkers and brain-age delta, including age, sex, and APOE-ɛ4 status as covariates.

For the main analyses, we applied a false discovery rate (FDR) multiple comparison correction following the Benjamini-Hochberg procedure. All tests were two-tailed, with a significance level of α = 0.05.

In exploratory analyses, we selected the previously studied CSF biomarkers that remained significant when adjusting for AD pathology and tested the mediation effect of that biomarker in the association between brain-age delta and cognition changes measured with PACC, for both Model 1 and Model 2. We performed 5000 iterations bootstrapped mediation analyses using the statsmodels package [48]. All mediation analyses were controlled for age, sex, APOE-ɛ4 status, and clinical status. For Model 1, Aβ42 levels and p-tau181 levels were not included as covariates, while for Model 2 they were included.

Lastly, also in confirmatory analyses, we tested whether the significant associations found in ALFA+ could also be seen in ADNI. This was done to validate the effect of the previously selected CSF biomarkers on brain-age delta. We computed linear models for Model 1 and Model 2. As post hoc analyses, we tested the mediation effect of this biomarker in the association between brain-age delta and cognitive changes, following the same procedure performed in ALFA+. Due to the smaller sample size of individuals with available computed brain-age delta, CSF sTREM2, and PACC change, we performed these mediation analyses joining the CU and the MCI individuals and included diagnosis as a covariate in both models.

Results

Participants characteristics

Table 1 summarizes the demographic, CSF biomarkers, and cognitive variables of the study participants, for ALFA+ and ADNI individuals by Aβ status. Cognitive decline measured with PACC was steeper in Aβ+ individuals than in Aβ-, both in ALFA+ and ADNI. As expected, MCI individuals presented a higher cognitive decline than CU individuals.

Table 1 Participants’ characteristics.
Full size table

Brain-age delta and its association with glial reactivity and synaptic dysfunction biomarkers

Table 2 shows the results of the linear regressions between all the glial and synaptic biomarkers for 1) brain-age delta and 2) AD signature for Model 1 (not including CSF Aβ42 and p-tau181 as covariates) and Model 2 (including CSF Aβ42 and p-tau181 as covariates).

Table 2 Effect of CSF Synaptic and glial biomarkers in brain-age delta and AD signature on ALFA+.
Full size table

In model 1, older brain-age was associated with lower CSF GAP43 (PFDR = 0.026), S100b (PFDR = 0.038), synaptotagmin-1 (PFDR =0.015), and sTREM2 (PFDR < 0.001). When correcting for the effect of CSF Aβ42 and p-tau181, only CSF sTREM2 remained negatively associated with brain-age delta (PFDR = 0.015). This indicates that higher concentrations of CSF sTREM2 was associated with younger brain-age independently of AD pathology (Fig. 1). The association between higher CSF sTREM2 and younger brain-age delta remained significant after removing those participants carrying rare TREM2 variants (Supplementary Fig. 2), and correcting the models by CSF Aβ40 (Supplementary Table 2), indicating that our results were not affected by inter-individual differences in CSF production and clearance rate [47]. In contrast, the associations between brain-age delta and CSF GAP43, S100b and SYT1 did not remain significant when removing the effect of Aβ40 (Supplementary Table 2).

Fig. 1: Associations between CSF sTREM2 and brain-age delta in ALFA+ and ADNI participants.
figure 1

Linear regression results for Model 1 (brain-age delta ~ CSF biomarker + age + sex + APOE-ε4 status), in blue (A, B), and for Model 2 (brain-age delta ~ CSF biomarker + age + sex + APOE status + Aβ + p-tau), in purple (C, D). In A and C, scatter plots representing the associations of sTREM2 and brain-age delta in ALFA+ cognitively unimpaired (CU), ADNI CU and ADNI MCI. In B and D, standardized estimates of the regression.

Full size image

None of the associations were significant when using the AD signature as the outcome, neither in Model 1 nor Model 2, indicating that, in our sample, CSF concentrations of these biomarkers are not associated with brain regions that undergo atrophy in clinical stages of AD.

Glial reactivity is reflected in brain-age delta across Aβ groups

After multiple comparisons, we did not find significant interaction effects between Aβ status and any of the CSF biomarkers associated with brain-age delta (Supplementary Table 3). A visual representation of the association between brain-age delta and CSF sTREM2 stratified by Aβ status can be seen in Supplementary Fig. 2.

Conversely, we found significant interaction effects between Aβ status and CSF GAP43 (PFDR = 0.026), GFAP (PFDR = 0.015), and YKL-40 (PFDR = 0.018) with AD signature, showing a different effect of glial reactivity and synaptic dysfunction on brain aging and on AD signature. While we observed that higher cortical thickness in the AD signature was related to lower CSF values of GAP43, GFAP, and YKL-40 in Aβ- individuals, and conversely, lower AD signature values were related with higher values of these CSF biomarkers in Aβ+ individuals (Supplementary Fig. 3), the main effects stratified by Aβ status were not significant.

TREM2-mediated microglial response mediates the link between older brain-age and decline in the PACC

Given that the association between brain age delta and CSF sTREM2 was the only remaining after adjusting for the effect of CSF Aβ42 and p-tau181, we further explored the mechanisms linking brain-age delta, microglial reactivity (as measured with CSF sTREM2) and PACC changes (Fig. 2).

Fig. 2: sTREM2 fully mediates the association between biological brain age and cognitive changes in cognitively unimpaired individuals in ALFA+.
figure 2

Mediation by CSF sTREM2 of the association between brain-age delta and cognitive changes in (A) ALFA+ and (B) ADNI. Results for Model 1 (in blue) and Model 2 (in purple). Blue lines represent significant negative associations. Red lines represent significant positive associations. Significance given by P < 0.05.

Full size image

We found that the association between older brain-age and decline in PACC was fully mediated by CSF sTREM2 in ALFA+, both when not adjusting for CSF Aβ42 and p-tau181 pathology in Model 1 (indirect effect: P = 0.0080), and when adjusting for CSF Aβ42 and p-tau181 pathology in Model 2 (indirect effect: P = 0.034). Older brain-age deltas were associated with lower CSF sTREM2 (Model 1: β = −0.204, P < 0.0001, Model 2: β = −0.159, P < 0.0001), and, in turn, lower CSF sTREM2 was associated with worse PACC evolution (Model 1: β = 0.185, P = 0.004, Model 2: β = 0.194, P = 0.008).

Replication of results in an independent cohort

We conducted confirmatory analyses to reproduce the effect of CSF sTREM2 on brain-age delta seen in ALFA+ in an independent cohort, ADNI.

Table 3 and Fig. 1 show the results of the linear regressions between CSF sTREM2 and brain-age delta for cognitively unimpaired and MCI individuals. In Model 1 (not including CSF Aβ42 and p-tau181 as covariates) brain-age delta was negatively associated with CSF sTREM2 only in MCI individuals (P = 0.006). Even though the association in cognitively unimpaired was not significant, the 95% confidence interval (CI) of the regression coefficients overlapped with the 95% confidence interval (CI) from the ALFA+ (ADNI: 95% CI = [−0.228, 0.104] and ALFA+: 95% CI = [−0.341, −0.125]). In Model 2 (including CSF Aβ42 and p-tau181 as covariates) brain-age delta was still negatively associated with CSF sTREM2 only in MCI individuals (P = 0.023). This indicates that younger-appearing brains were linked with higher CSF sTREM2 independently of AD pathology. The association between higher CSF sTREM2 and younger brain-age delta in MCI individuals was still significant after removing those participants carrying a rare TREM2 variant for Model 1 (Supplementary Fig. 4). We did not find an interaction effect of Aβ status in the associations between brain-age delta and CSF sTREM2 in ADNI, neither in cognitively unimpaired nor MCI (Supplementary Table 5 and supplementary Fig. 2).

Table 3 Effect of CSF sTREM2 on brain-age delta in ADNI.
Full size table

Figure 2 shows the results of the mediation analysis, for which we combined cognitively unimpaired and MCI individuals. We did not find a significant mediation of CSF sTREM2 linking older brain-age with cognitive decline measured with PACC, nor in Model 1 (indirect effect: P = 0.796), nor in Model 2 (indirect effect: P = 0.849). However, the total effect in both models was significant. When not adjusting for CSF Aβ42 and p-tau181, older brain-age deltas were associated with lower CSF sTREM2 (β = ,−0.228 P = 0.003). However, lower CSF sTREM2 was not associated with the decline in the PACC (β = −0.021, P = 0.755). In turn, older brain-age deltas were associated with decline in PACC (β = −0.143, P = 0.031). Conversely, when adjusting for CSF Aβ42 and p-tau181, older brain-age deltas were not significantly associated with lower CSF sTREM2 (β = −0.004, P = 0.629), though CI overlapped with the ones in Model 1. Lower CSF sTREM2 was not associated with decline in the PACC (β = 0.071, P = 0.348). In turn, brain-age deltas were still associated with decline in the PACC (β = −0.170, P = 0.0080).

Discussion

In this work, we assessed, in cognitively unimpaired individuals, the associations between several pathophysiological mechanisms altered in AD and brain-age delta as a measure of biological brain aging. We found that higher CSF sTREM2 was associated with younger-appearing brain-age independently of AD pathology in cognitively unimpaired individuals at risk for developing AD dementia, which was confirmed in MCI individuals from ADNI. Furthermore, CSF sTREM2 fully accounted for the link between older brain-age and cognitive decline. To the best of our knowledge, this study reports for the first time the association of microglial reactivity with biological brain aging and its impact on cognition. Our findings further support that high CSF sTREM2 reflects a protective microglial state and that the modulation of TREM2-related mechanisms in early stages of the disease may delay cognitive decline [49].

TREM2 is an innate immune receptor expressed in the microglia surface and its activation promotes phagocytosis and clearance of neuronal toxic substances and abnormal proteins by microglia. CSF sTREM2 results from shedding the TREM2 ectodomain and is a well-established surrogate measure of TREM2-mediated microglia function [11, 37]. Previous studies support a protective role of the TREM2-mediated microglia activation in the symptomatic phase of AD [50] and preclinical stages. Higher CSF sTREM2 has been associated with attenuated amyloid-β deposition and of tau pathology [51, 52]. A previous study from our group showed that higher CSF sTREM2 was associated with attenuated cognitive decline in cognitively unimpaired individuals [53], in line with other reports across the full AD continuum [14,15,16]. Besides, literature shows that older brain-age delta is associated with cross-sectional measures of cognition and that they can predict longitudinal cognitive decline [54]. Here, we assessed the impact that higher CSF sTREM2 have on brain-age delta and cognitive changes measured with PACC. We found that higher CSF sTREM2 fully mediated the relationship between older brain-age and decline in PACC in ALFA+. This mediation effect persisted even when adjusting for markers of Aβ and tau pathology. Although the mediating effect in ADNI was non-significant, we could essentially reproduce the associations seen in ALFA+. Furthermore, we found a significant total effect in ADNI, in which brain-age delta was negatively associated with PACC changes (both when including and when excluding Aβ and p-tau as covariates). These mediation results might explain biological mechanisms relating brain aging and longitudinal cognitive changes. The possible variability affecting the presence of the mediating effect of TREM2-related response should be further studied.

We have seen that, independently of Aβ or tau pathology, there might be a protective effect of TREM2-related microglia activation, by which microglia is associated with a younger appearing brain morphology. Although CSF sTREM2 increases with age and with AD pathology, previous research has indicated that CSF sTREM2 is associated with a microglia state that may be protective, specifically regarding its impact on Aβ and p-tau accumulation [13, 16, 37, 51, 52]. Our findings extend this protective association to include associations with biological brain aging. In particular, we showed that the protective effect indicated by high CSF sTREM2 mediates the link between younger brain aging and attenuated cognitive decline and this is independent of AD pathology in cognitively unimpaired individuals. Our findings build on a previous research showing a protective role of TREM2-mediated microglial activation, also independent of AD pathology [55]. We replicated the association between CSF sTREM2 and biological brain-age in MCI individuals of the ADNI cohort. In contrast, the results could not be replicated in the cognitively unimpaired individuals of the ADNI cohort. One possible explanation is the age difference, as cognitively unimpaired individuals in ADNI were 10 years older on average than those of ALFA+. TREM2-mediated microglia response may have an effect on middle-aged individuals but not (or to less extent) in older individuals. Despite the nonsignificant results in cognitively unimpaired individuals, the confidence intervals and estimates overlapped with those from ALFA+, suggesting that variability and the low number of cases of ADNI cognitively unimpaired individuals may contribute to these outcomes.

In addition to CSF sTREM2, we detected significant associations between brain-age delta and several synaptic and glial biomarkers. However, these associations became non-significant when accounting for CSF Aβ42 and p-tau181 as covariates, as well as when accounting for CSF Aβ40. Therefore, these associations were driven by AD pathology and by differences in rate of protein production and clearance through the CSF, respectively. The AD signature did not exhibit any associations with the CSF biomarkers examined in this study. This lack of association may be attributed to the AD signature’s focus on regions vulnerable to AD pathology, potentially overlooking changes related to other pathophysiological mechanisms in the earlier preclinical stages of AD, especially when considering cognitively unimpaired individuals who are Aβ-. However, an interaction was observed between CSF GFAP, YKL-40, and GAP43 and Aβ status. This finding is in line with expectations, given that the AD signature characterizes brain structures more prone to AD pathology. Recent work has suggested that alterations in astrocytes occur at an early stage of AD [28], alongside synaptic dysfunction, which is also an early process in AD pathogenesis. We found that a thinner cortical thickness in the AD-signature was seen with higher CSF values of these astrocyte reactivity and synaptic dysfunction biomarkers in Aβ+ individuals, while in Aβ- individuals, a thicker AD-signature was seen with higher CSF values. Therefore, the effects of synaptic dysfunction and astrocyte reactivity measured with these CSF biomarkers on AD-related brain regions in cognitively unimpaired individuals may be modified by AD pathology. These results suggest that AD signature and brain-age delta capture different mechanisms. While brain-age delta captures effects unrelated to AD pathology, mainly the protective effects of reactivity of microglia in aging processes in the brain, AD signature might capture effects related to synaptic and astrocytes changes linked to AD pathology.

Neuroinflammation, particularly microglial response, has long been considered to play a detrimental role in AD. The potential protective or detrimental role of TREM2 in the disease has also been debated [56]. However, most majority of data from genetic studies, animal models and biomarkers studies in human cohorts, support a protective role of TREM2. Our study reinforces this notion and additionally emphasizes the role of TREM2-mediated microglial activation in brain aging, independent of AD pathology. Our findings support the idea of enhancing TREM2 function as a therapeutic strategy not only for AD, but also for other age-related neurodegenerative diseases. Furthermore, our results indicate that this approach may hold potential for addressing brain aging itself, an area that has yet to be extensively explored in relation to TREM2. Currently, there are several TREM2 agonistic antibodies in early clinical trials, which bind at the N-terminal of the ADAM10 and ADAM17 cleavage site, prevent shedding and releasing of sTREM2, cross-link cell surface TREM2 and eventually activate SYK signaling [57, 58].

To the best of our knowledge, this is the first study providing insights into the biological mechanisms associated with a younger-appearing brain morphology by examining different possible factors influencing brain aging. The inclusion of a substantial set of CSF biomarkers enhances our ability to unravel the mechanisms driving age-related changes in the brain. The comparison with the AD signature, an MRI measurement more specific to AD, strengthens the relevance of our findings. Additionally, investigating the biological mechanisms linking brain morphology with longitudinal changes in cognition is important to develop possible clinical trials. Using an independent cohort to validate and corroborate our results adds robustness to the study’s findings. However, certain limitations must be acknowledged. The sample size of the replication cohort, particularly for cognitively unimpaired individuals, was relatively small, leading to increased variability in the results. In addition, ADNI cognitively unimpaired individuals were older than ALFA individuals, which might influence the results. This sample size was further diminished when considering longitudinal cognition, potentially affecting the generalizability of findings. Additionally, it is important to note that the PACC composite computed in ADNI was different from that from ALFA+’s; while ADNI used the tests ADAS and Trails B, ALFA+ used FCSRT and Coding. Regarding the association between CSF sTREM2 and PACC change, the high variability found in the individuals of ADNI, as well as having to combine cognitively unimpaired and MCI together, could be the reason for lack of significance. Furthermore, due to sample size limitations, we could not address differences in sex in this study, which should be further studied in future research. Despite these limitations, the study contributes valuable insights into the biological underpinnings of brain aging and cognition.

In summary, we showed that TREM2-mediated microglial reactivity, as measured with CSF sTREM2, is associated with a younger-appearing brain morphology. Importantly, this microglial reactivity fully accounts for the relationship between younger biological brain age and slower cognitive decline in preclinical AD stages after accounting for the effects of amyloid and tau pathology. Understanding the mechanisms involving biological brain aging, glial activation, and cognitive decline can offer new avenues for potential therapeutic strategies to mitigate cognitive decline.