Longitudinal increase of CSF soluble TREM2 is driven by early aggregation of A β 42 and associates with slower amyloid deposition and clinical decline in autosomal-dominant Alzheimer’s disease

Therapeutic modulation of TREM2-dependent microglial function provides an additional strategy to 2 slow progression of Alzheimer disease (AD). Although studies on animal models suggest that 3 TREM2 is protective, the trigger of increased TREM2 expression during disease progression and its 4 clinical and pathological consequences in AD remain unclear. We measured longitudinally soluble 5 TREM2 (sTREM2) as a surrogate marker for protective TREM2-signalling in cerebrospinal fluid 6 (CSF) from participants in the Dominantly Inherited Alzheimer Network (DIAN) observational study. 7 In mutation carriers (MC), the longitudinal sTREM2 increase followed the earliest aggregation of 8 A β 42 captured by CSF-A β 42 decrease, but not yet by Pittsburg compound-B Positron Emission 9 Tomography (PiB-PET). Higher sTREM2 increase rates provided protection from A β -deposition, 10 whereas lower rates enhanced p-tau increase associated with PiB-PET increase. Moreover, 11 presymptomatic MC with high or low sTREM2 increase rates have opposite associations between 12 CSF A β 42 and PiB-PET longitudinal changes, suggesting that TREM2 modifies A β plaque deposition 13 and compaction. Finally, higher sTREM2 increase rates protected from cortical shrinkage and 14 cognitive decline. Our findings position the TREM2 response within the amyloid cascade right after 15 the first pathological changes in A β 42 aggregation, support ongoing efforts to develop TREM2 16 modulating therapies, and predict a very early window for therapeutic intervention. 17 associated with a higher in CSF sTREM2. This finding suggests that very early Aβ -related pathology, even before are detectable by Aβ -PET imaging, drives sTREM2 generation. In line with that, a very early aggreg ation stage, before are detectable by histology, suggested to play a key role as a target for disease-modifying The Dominantly Inherited Alzheimer Network (DIAN) observational study is a well-described 3 longitudinal and international study, launched in 2008, and recruiting individuals from families 4 carrying mutations in the presenilin 1 ( PS1 ), presenilin 2 ( PS2 ) and amyloid precursor protein ( APP ) 5 genes from 17 sites distributed in USA, Argentina, UK, Germany, Spain, and Australia. This study is supervised by the institutional review board at Washington University (St. Louis, MO, USA), which 7 provided human studies approval. Participants or their caregivers provided written informed consent 8 in accordance with their local institutional review board. Asymptomatic individuals were followed 9 with a 3 year-interval until 3 years after their parental age at onset, when the follow-ups become 10 annual. Symptomatic participants were followed annually. The estimated years from expected 11 symptom onset (EYO) was calculated as the participant’s current age relative to parental age at first 12 progressive cognitive decline for each visit per asymptomatic participant 24 . In the case of symptomatic 13 participants, the EYO was calculated as the participant’s current age relative to the participant´s age at 14 symptom onset.


Introduction 1
Recent advances in understanding the dynamic responses of microglia to pathological challenges such 2 as the deposition of insoluble proteins and cell death during neurodegeneration have increased our 3 knowledge of immune cell function in the brain and even allowed the development of novel 4 therapeutic approaches 1,2 . Historically, microglia were primarily believed to pathologically contribute 5 to disease progression, e.g., by promoting aberrant synaptic pruning and activation of the 6 inflammasome 3,4 . However, single-cell sequencing technologies have identified dynamic microglial 7 populations, which sense their environment and trigger defensive responses to Alzheimer disease 8 (AD) pathology 5 . Moreover, large genome-wide association studies have identified a number of risk 9 variants for sporadic AD in genes expressed exclusively within microglia in the brain. Specifically, 10 analyses of the disease-associated variants in the triggering receptor expressed on myeloid cells 2 11 (TREM2) were extremely intriguing as they strongly suggest that loss-of-function mutations are 12 associated with an increased risk for late onset AD (LOAD) 6-9 . TREM2 loss of function locks 13 microglia in a homeostatic state and prevents the switch to disease-associated microglia (DAM) 5, 10 . 14 Since DAM facilitate lipid metabolism, efficiently remove amyloid β-peptide (Aβ) seeds, and form a 15 barrier around amyloid plaques, their protective activities are now explored for disease modifying 16 therapeutic strategies 1,2 . In that regard, a variety of different TREM2-agonistic antibodies have been 17 generated, all triggering several protective microglial functions, by stabilizing and cross-linking 18 signalling-competent TREM2 1 . Moreover, one of these antibodies already passed a phase-1 clinical 19 trial and multiple therapeutic antibodies will soon be tested for their efficacy in AD patients 11 . It is, 20 therefore, of greatest importance to translate our knowledge on protective TREM2 functions from 21 animal models to AD patients. Quantitative analysis of soluble TREM2 (sTREM2) in cerebrospinal 22 fluid (CSF) as a surrogate marker for TREM2-dependent microglial-activation allows such 23 translational efforts. We have previously shown that cell-surface full-length TREM2 is shed by 24 proteases of the ADAM family, releasing soluble TREM2 (sTREM2) into biological fluids including 25 CSF 9,12 . Since only cell-surface full-length TREM2 is capable to efficiently initiate downstream 26 signalling, and ADAM proteases cleave TREM2 preferentially on the plasma membrane 9,12 , CSF 1 sTREM2 can be considered as a surrogate maker for TREM2 expression and signalling. 2 Cross-sectional studies have shown that CSF sTREM2 levels are elevated in late presymptomatic and 3 in early symptomatic stages both in sporadic and in autosomal-dominant AD (ADAD), and correlate 4 with tau pathology-related markers in CSF (total tau -t-tau-and phosphorylated tau on threonine 181 -5 p-tau-), reinforcing the idea of TREM2-dependent microglial activation as a defensive response to AD 6 pathology 13-16 . Higher CSF sTREM2 levels at baseline are also associated with a slower hippocampal 7 shrinkage and slower memory decline in symptomatic phases of sporadic LOAD supporting the 8 potential beneficial role of TREM2 17 . However, cross-sectional CSF sTREM2 levels only give an 9 estimate of the microglial activation state at a single time-point and are influenced by a high 10 interindividual variability. Therefore, cross-sectional CSF sTREM2 levels do not represent the 11 dynamic TREM2-dependent microglial response during AD progression. In fact, longitudinal studies 12 are more accurate than cross-sectional studies to investigate pathological processes occurring in AD 13 as they have a higher power to discriminate the temporal changes and the dynamic relationships 14 between them 18-20 . Moreover, based on longitudinal amyloid PET imaging studies and seeding 15 experiments in mice 21 , the beneficial effect of TREM2-dependent microglial functions is expected to 16 be most effective in the earliest stages of Aβ-deposition. Thus, any effects of TREM2 on disease 17 initiation and progression must be studied during an asymptomatic phase of patients known to 18 develop AD symptoms at later time points. In contrast to sporadic AD, ADAD has a predictable 19 clinical onset in each family and penetrance of the involved mutations is mostly complete 22 . That 20 allows to stage individuals relative to their expected year of symptom onset and enabled us to study 21 biomarker dynamics from a very early presymptomatic phase in a way that is not possible with 22 sporadic AD patients. The Dominantly Inherited Alzheimer Network (DIAN) observational study has 23 recruited a large number of participants from families suffering from ADAD, many of them with 24 longitudinal markers of Aβ-deposition, tau-related pathology, neuronal death and dysfunction, and 25 longitudinal cognitive evaluations 23,24 . We studied the longitudinal change of CSF sTREM2 26 throughout the course of AD in the observational DIAN cohort, the factors triggering this change, and 27 the relationship between the dynamics of CSF sTREM2 and the dynamics of other biomarkers 1 representing Aβ-accumulation and deposition, tau-pathology and brain structure, as well as the 2 influence of the dynamics of CSF sTREM2 on cognitive decline. Our goal was to explore potential 3 protective activities of TREM2 during the presymptomatic phase of the disease, to search for an 4 interaction of sTREM2 rate of change with amyloid and tau pathology and to identify a window for 5 therapeutic modulation of TREM2. 6

7
Early Aβ changes drive CSF sTREM2. We measured sTREM2 in the longitudinal CSF samples 8 from participants in the DIAN observational study. Table 1 summarizes baseline characteristics of the 9 participants. First, we assessed cross-sectionally the time point at which CSF sTREM2 levels 10 significantly differ in mutation carriers (MC) from non-carriers (NC) according to the methods 11 already described elsewhere 19 . We found that CSF sTREM2 levels started to be significantly higher in 12 MC than in NC 21 years prior to the expected symptom onset (Fig. 1a). Interestingly, this time point 13 is close to the point at which the cortical uptake of PiB-PET starts to be significantly higher in MC 14 than NC (approximately at -22 estimated years from expected symptom onset (EYO)) 19,22 . Although, 15 CSF sTREM2 levels were significantly higher in MC than in NC 21 years before the expected 16 symptom onset, we could not find a time point at which the rate of CSF sTREM2 change was 17 significantly different in MC from NC (Fig. 1b). We did not find any influence of sex, educational 18 level, age, or EYO at baseline on the subsequent rate of CSF sTREM2 change studied by univariate 19 Linear Mixed Effects (LME) models in MC and NC (supplementary table 1). The mutation status 20 (NC vs MC) and the mutant gene involved did not significantly affect the rate of sTREM2 change 21 (supplementary table 1 and 2). 22 Next, we searched for the pathological factors influencing the longitudinal change of CSF sTREM2. 23 To do so, we investigated whether baseline levels of Aβ-related, tau-related or structural MRI 24 biomarkers available in DIAN participants were related to the subsequent rate of CSF sTREM2 25 change, both in MC and NC, using univariate LME models. For Aβ, we found that lower levels of 26 both, CSF Aβ42 and CSF Aβ42/Aβ40 at baseline independently predicted a subsequent higher rate of 1 CSF sTREM2 increase in MC but not in NC (Fig. 2a and supplementary table 3). In contrast, we 2 did not find any association between total cortical uptake in PiB-PET at baseline and the subsequent 3 rate of CSF sTREM2 change ( Fig. 2b and supplementary table 3). CSF p-tau, t-tau and structural 4 MRI biomarkers at baseline showed no relationship to the subsequent longitudinal change of CSF 5 sTREM2 in MC nor NC (Fig. 2c and 2d and supplementary table 3). Altogether, these results 6 suggest that very early Aβ-pathological changes, even before fibrillary deposits are detectable by 7 amyloid-PET imaging 25-27 , trigger the subsequent longitudinal increase of CSF sTREM2 in ADAD. 8 Higher sTREM2 increase rate slows Aβ-deposition. To further understand the relationship between 9 the longitudinal dynamics of CSF sTREM2 and the evolution of AD pathology represented by the 10 correspondent biomarkers, we studied whether the rate of CSF sTREM2 change was associated with 11 the rate of change of Aβ-accumulation and tau-related pathology biomarkers in MC by bivariate LME 12 models 28,29 . We found that a higher rate of CSF sTREM2 increase was related to a slower decrease in 13 CSF Aβ42 (lower rate of CSF Aβ42 decrease) in presymptomatic MC (r = 0.56, p = 0.01, Fig. 3a) and 14 with a slower increase in the total cortical PIB-PET uptake in symptomatic MC (r = -0.67, p= 0.006, 15 Fig. 3b). For tau-pathology related markers, we found a trend for an association between a higher rate 16 of CSF sTREM2 increase and a higher rate of t-tau increase when considering all MC together, but 17 not when stratifying by the clinical state (r = 0.34, p = 0.08, Fig. 3c). No association with the rate of 18 p-tau change in MC nor in any MC subgroup was found (presymptomatic/ symptomatic MC, Fig. 3d). 19 Higher sTREM2 increase rate attenuates tau pathology. Although we did not find a clear 20 connection between the rates of change of CSF sTREM2 and tau-related markers, we further explored 21 the influence of the longitudinal sTREM2 change on the association between Aβ-related markers and 22 CSF p-tau longitudinal changes. In presymptomatic MC, but not in symptomatic MC, the association 23 between the rate of CSF p-tau change and the rate of change in the PiB-PET cortical mean differed by 24 the rate of CSF sTREM2 change above or below the median (β = -0.394 (high sTREM2 change), p = 25 0.005 for the linear interaction) ( Fig. 4a and 4b; appendix 1, supplementary tables 6 and 7). This 26 suggests a slower p-tau increase rate relative to PiB-PET in presymptomatic MC with higher rates of CSF sTREM2 increase. Regarding the rate of CSF Aβ42 change, its association with the rate of CSF p-1 tau change also differed by the rate of CSF sTREM2 change in symptomatic MC with CSF sTREM2 2 increases above or below the median (β = -4.065 (high sTREM2 change), p = 0.027 for the quadratic 3 interaction) ( Fig. 4c and 4d, appendix 1, supplementary tables 6 and 7). That indicates a slower p-4 tau increase related to the CSF Aβ42 change in symptomatic MC with a higher rate of CSF sTREM2 5 change. Taken together these results suggest that TREM2 activation attenuates the evolution of tau-6 related pathology in ADAD in an Aβ-dependent manner. 7 The relationship between CSF Aβ42 and PiB-PET longitudinal changes is modified by the 8 sTREM2 increase rate. Our results indicate a seemingly discrepant relationships between the CSF 9 sTREM2 dynamics and the longitudinal changes in CSF Aβ42 and the cortical uptake in PiB-PET. 10 That could reflect putative effects of TREM2-activation on both, clearance of early Aβ42-seeds and 11 compaction of Aβ-plaques 21,30-34 , which may affect the equilibrium of CSF Aβ42 and PiB-PET cortical 12 uptake. To explore this possibility, we assessed whether the rate of CSF sTREM2 change influenced 13 the relationship between the rates of change of CSF Aβ42 and PiB cortical uptake. Interestingly, the 14 quadratic model predicting the rate of CSF Aβ42 change by the rate of PiB-PET uptake change 15 significantly improved when including the binary rate of CSF sTREM2 change and its interaction 16 with PiB-PET cortical mean, especially in presymptomatic MC (multiple r-squared = 0.51 (vs. 0.15), 17 adjusted r-squared = 0.45 (vs. 0.09), appendix 1, supplementary tables 8 and 9). The association 18 between CSF Aβ42 and the PiB PET cortical mean rates of change differed significantly in the groups 19 with high (above median) versus low (below median) rate of CSF sTREM2 change (β = 0.932 (high 20 sTREM2 change), p = 0.004, for the linear interaction; β = -6.082 (high sTREM2 change), p < 0.001, 21 for the quadratic interaction) (Fig. 4e and 4f). In presymptomatic MC with a high rate of CSF 22 sTREM2 change, CSF Aβ42 decreased with the increase of PIB-PET cortical uptake, whereas those 23 with a low rate of CSF sTREM2 change showed an opposite relation. In symptomatic MC, this 24 interaction effect was not significant, maybe due to the low number of participants in this group. 25 These results suggest an influence of CSF sTREM2 dynamics on the relationship between CSF Aβ42 26 and PIB-PET longitudinal changes in early stages of the disease, which may be in line with the well-1 described role of TREM2 in clearance and compaction of Aβ-aggregates 21,30,32,34,35 . 2 Higher sTREM2 increase rate protects from cortical shrinkage in the precuneus. A higher rate of 3 sTREM2 increase was associated with slower cortical shrinkage in the precuneus of the 4 presymptomatic MC (r = 0.46, p = 0.04; Fig. 5a). We also found a trend for an association between 5 higher sTREM2 increase rate and slower cortical shrinkage in the precuneus in the case of 6 symptomatic MC ( Fig. 5a and 5b). Although we did not find any association between the rate of 7 sTREM2 change and the rate of hippocampal shrinkage in presymptomatic nor symptomatic MC 8 ( Fig. 5c and 5d), we observed that a higher CSF sTREM2 level at baseline was associated with 9 slower hippocampal shrinkage in line with our previous results 17 (supplementary table 4). 10 Higher sTREM2 increase rate is associated with slower cognitive decline. Finally, we studied the 11 influence of the longitudinal dynamics of CSF sTREM2 on cognitive decline. To do so, we used again 12 a bivariate LME model assessing the association between the rate of CSF sTREM2 change and the 13 rate of cognitive decline, measured by a cognitive composite 36 . We found a strong association 14 between a higher rate of CSF sTREM2 increase and a slower cognitive decline in presymptomatic 15 MC (r = 0.62, p = 0.003, Fig. 5e and 5f). This finding is in line with the effects of a higher rate of 16 sTREM2 increase on CSF Aβ42 and slower cortical shrinkage in the precuneus in the presymptomatic 17 MC. On the other hand, we did not find any significant association between the rate of CSF sTREM2 18 change and the rate of cognitive decline in symptomatic MC (r = -0.07, p = 0.79, Fig. 5e). Thus, 19 these findings indicate an early beneficial effect on cognition of the TREM2-dependent microglial 20 response. 21

22
Although major advances have been made in the past to develop amyloid-lowering therapeutic 23 strategies, their clinical outcome remained disappointing. Clinically, secretase inhibitors caused 24 major side effects probably due to the inhibition of proteolytic processing of physiologically required 25 substrates 37,38 . Anti-Aβ immunotherapeutic approaches efficiently removed Aβ-plaques 39 , but had so 26 far mostly minute and inconsistent effects on cognitive and functional outcome. Furthermore, the 1 recent approval for Aducanumab has not been exempt of controversy 40,41 . Thus, intense efforts are on 2 the way to identify new therapeutic targets. It is now widely accepted that TREM2-dependent 3 microglial functions are required to limit initiation and progression of AD pathology in animal models 4 of amyloidosis. Agonistic antibodies were developed with the goal to increase TREM2-dependent 5 protective functions 1,2 . They promote TREM2 expression, increase clustering of microglia around 6 amyloid plaques, reduce amyloid plaque load and neuritic dystrophies, and even improve cognition in 7 mouse models 1,2 . It is therefore of greatest importance, to prove that TREM2-dependent protective 8 functions observed in mouse models and cultured human microglia also occur in AD patients and to 9 define a therapeutic window, when elevation of TREM2 activity is most efficacious. The only way to 10 study that is using a biomarker approach 42 . Considering sTREM2 as a surrogate marker for cell-11 surface signaling-competent TREM2 9,13-15,17 , we studied its longitudinal changes in CSF during 12 disease development in the longitudinal DIAN cohort, the largest and best characterized ADAD 13 cohort 19,22,24 . The predictable nature of ADAD allowed us to determine the relationship between the 14 longitudinal dynamics of CSF sTREM2 and other longitudinal biomarkers associated with 15 progression of AD pathology, along with the longitudinal cognitive evolution from the very early 16 preclinical phases, decades before the first symptom, until a late phase with fully developed AD 17 symptomatology. Correlating the concentrations of CSF sTREM2 with pathological measures, such as 18 Aβ-and tau-related changes, as well as hippocampal and precuneus shrinking, and cognition allowed 19 us to draw conclusions on a potentially protective function of TREM2 in human brains and may also 20 predict an optimal time window for pharmacological intervention. 21 Interestingly, we found that lower baseline levels of CSF Aβ42, but not higher baseline cortical uptake 22 of PIB-tracer nor markers for tau-related pathology or neuronal death, were associated with a higher 23 increase in CSF sTREM2. This finding suggests that very early Aβ-related pathology, even before 24 Aβ-deposits are detectable by Aβ-PET imaging, drives sTREM2 generation. In line with that, a very 25 early Aβ aggregation stage, before seeds are detectable by histology, has been recently described in 26 animal models, and suggested to play a key role as a target for disease-modifying drugs 43 . Further 27 supporting the relationship between CSF sTREM2 increase and the early Aβ-pathological changes, 1 CSF sTREM2 levels begin to be significantly higher in MC 21 years before the expected symptom 2 onset, close to the time-point where the longitudinal change of CSF Aβ42 and Aβ-PET imaging starts 3 to diverge in MC compared to NC 19,22 . Thus, microglia may be able to extremely sensitively monitor 4 and respond to the slightest Aβ related pathological challenges in the human brain. Using a different 5 analytical approach, we previously reported that CSF sTREM2 levels at baseline were higher in MC 6 than in NC in a time window of five years before and after the expected symptom onset 13 . The 7 difference between these findings may be due to a higher discriminative power of our current 8 method 19,22 detecting differences in the extremes of the EYO time line, where the number of included 9 participants is low. 10 Our current work also shows that a higher sTREM2 increase rate is associated with a slower decrease 11 in CSF Aβ42 in presymptomatic MC, and a slower increase in total cortical PIB-PET uptake in 12 symptomatic MC. This association agrees with our previous results in symptomatic LOAD 44 . 13 Moreover, our current results show differential associations between the rate of CSF sTREM2 change 14 and CSF Aβ42 or PiB-PET cortical uptake in different clinical phases. That supports a potential dual 15 protective effect in line with previous findings in animal models, aimed to reduce plaque-associated 16 toxicity 21,35 . In the early presymptomatic stages, the association between higher sTREM2 increase 17 rates and a slower CSF Aβ42 decrease may be related to microglia clustering around the smallest 18 amyloid seeds, limiting their growth and spreading 21 . At later stages during the symptomatic phase of 19 the disease, when plaques are fully developed, a protective function of microglia may be carried out 20 by their barrier function and their ability to compact amyloid plaques 21,30,32,34,35 . Microglia may 21 contribute to amyloid plaque compaction by secreting ApoE, which drives Aβ aggregation directly 22 within the plaque 21 . Additionally, we found that in MC with a high sTREM2 increase rate, a higher 23 increase rate in the PIB-PET uptake is related to a higher decrease rate in the CSF Aβ42 levels, 24 probably reflecting the sequestration of Aβ42 in dense core amyloid plaques in an attempt to reduce 25 the most toxic amyloid aggregates as recently suggested by Huang et al 35 . The opposite is found in 26 MC with lower sTREM2 increase rates, possibly reflecting a failure of microglia in compacting Aβ-27 plaques. Additionally, CSF Aβ42 and Aβ-PET imaging are suggested to measure different aspects of 1 Aβ-pathology as, so far, no relationship was found between their longitudinal changes 25 . We now 2 solved this problem by the introduction of CSF sTREM2 in their relationship. Thus, we obtained a 3 relatively accurate model to predict the CSF Aβ42 changes by the PiB-PET changes. This highlights 4 the important role of TREM2 in amyloid-plaque metabolism, and points to CSF sTREM2 as a 5 relevant marker for a better interpretation of the Aβ-pathology related markers and their dynamics in a 6 clinical setting. 7 Regarding tau-related AD pathology, we could not detect any significant association between the 8 baseline levels of tau-related markers and the subsequent longitudinal sTREM2 change in CSF and 9 neither between the rate of CSF sTREM2 change and the dynamics of CSF t-tau and p-tau. Previous 10 cross-sectional studies in both sporadic and genetic AD clinical cohorts have shown a strong 11 correlation between CSF sTREM2 and CSF t-tau and p-tau 13-16 . We interpret the strong cross-12 sectional correlation between CSF sTREM2 and tau-related markers as the static view of AD 13 evolution where both markers are sequentially higher, reflecting the disease progression, as a result of 14 Aβ deposition. Albeit the lack of evidence of a direct relationship between the dynamics of CSF 15 sTREM2 and tau-related markers, we found that the rate of CSF sTREM2 increase influenced the 16 association between Aβ-pathology markers and CSF p-tau dynamics. A higher rate of CSF sTREM2 17 increase attenuated the rate of p-tau increase related with the PiB-PET increase rate in the 18 presymptomatic MC and with the rate of CSF Aβ42 change in symptomatic MC. Nevertheless, 19 difficulties with the interpretation of the rates of change of tau-related CSF markers have been already 20 described. An unexpected longitudinal decrease in CSF p-tau was found in the DIAN cohort before 21 the symptom onset, being associated with its sequestration in neurofibrillary tangles, while the rate of 22 t-tau change remained stable along the entire disease evolution 19 . On the other hand, studies in 23 sporadic AD have shown an increasing rate of change in tau-related CSF markers 45 . Although we 24 should interpret our current results with caution, our data support a protective role of TREM2 25 functions on tau-related pathology dependent on Aβ-pathology, which is in line with recent results in 26 mouse models 46 . 27 Protective effects of TREM2 on Aβ and tau related pathology might lead as a consequence to reduced 1 cortical thinning. Accordingly, we found a clear relationship between a higher sTREM2 increase rate 2 and slower cortical shrinkage in the precuneus in presymptomatic MC and a trend for an association 3 in symptomatic MC. However, we could not detect an association between the longitudinal sTREM2 4 change and the longitudinal evolution of hippocampal volume. The striking association between a 5 higher rate of increase in CSF sTREM2 and slower cortical shrinkage in the precuneus with a lack of 6 association in the hippocampus may not occur by chance. The precuneus is the first region affected by 7 Aβ accumulation in ADAD followed by a decrease in cortical FDG-PET signal and subsequent 8 cortical shrinkage 22 . This canonical sequence is not followed by the hippocampal region, where the 9 atrophy is the main event, not following a significant increase in amyloid deposition 22 . That suggests 10 that the beneficial effect of TREM2 follows a regional pattern, being more evident in early phases in 11 those brain areas with a higher amyloid accumulation rate. That is also in line with the triggering of 12 TREM2 protective functions by the early Aβ-pathological changes. However, we have already shown 13 a protective role of higher CSF sTREM2 levels at baseline on the hippocampal shrinkage in 14 symptomatic phases of sporadic LOAD 17 , and we replicate that here in ADAD. 15 We found a very striking association between a higher rate of CSF sTREM2 increase and a slower 16 cognitive decline in the presymptomatic AD stage. This result is consistent with the association 17 observed between a higher rate of CSF sTREM2 increase and a slower pathological progression in the 18 presymptomatic phase of the disease, represented by a slower decrease in CSF Aβ42 and slower 19 cortical shrinkage in the precuneus, highlighting the association between the Aβ-pathological process, 20 the neurodegeneration process and the consequent cognitive decline. We reported earlier that higher 21 CSF sTREM2 levels at baseline exert a beneficial effect on cognition, specifically on memory 22 domains and not on general cognitive composites, in symptomatic sporadic LOAD 17 . The current 23 results did not show a significant effect of the CSF sTREM2 levels at baseline on cognition, probably 24 because the cognitive composite we used in our current work is not specific for the memory domain 25 and also because our study is focused on the presymptomatic AD phase, including a limited number 26 of symptomatic participants. Taken together our results suggest that a higher rate of CSF sTREM2 increase slows Aβ-deposition and precuneus shrinkage, having a clear clinical readout via its strong 1 association with a slower cognitive decline in a presymptomatic AD stage. The beneficial effect of 2 TREM2-functions may continue in symptomatic stages by slowing hippocampal shrinkage in patients 3 with highest CSF sTREM2 levels. 4 The main limitation of our study is that this is an observational cohort, thus any causative 5 relationships must be interpreted with caution. Moreover, we used a variety of biomarkers, which are 6 indirect measures for studying pathological processes. Furthermore, the study of the relationship 7 between the longitudinal changes of tau-related markers and CSF sTREM2 was restricted to CSF 8 markers, with no available tau-imaging. Finally, CSF sTREM2 is a surrogate of TREM2 signaling and 9 expression of the entire brain, allowing no regional conclusions. Our recently developed TREM2 10 reporter mouse will allow us to address this pivotal question. Our study has important strengths. This 11 is the first study assessing longitudinally CSF sTREM2 changes through an extensive period from 12 very early presymptomatic AD stages until a late symptomatic stage. We report a comprehensive and 13 complete set of highly consistent findings including the biological triggers of the increase of CSF 14 sTREM2 and the effects on Aβ deposition, tau-related pathology, brain structure and cognition. 15 provided human studies approval. Participants or their caregivers provided written informed consent 8 in accordance with their local institutional review board. Asymptomatic individuals were followed 9 with a 3 year-interval until 3 years after their parental age at onset, when the follow-ups become 10 annual. Symptomatic participants were followed annually. The estimated years from expected 11 symptom onset (EYO) was calculated as the participant's current age relative to parental age at first 12 progressive cognitive decline for each visit per asymptomatic participant 24 . In the case of symptomatic 13 participants, the EYO was calculated as the participant's current age relative to the participant´s age at 14 symptom onset. Alzheimer disease mutations and the Apolipoprotein E (APOE) genotyping was performed according 20 to the methods already described 24 . Clinical evaluators were blinded to mutation status of participants. 21 Regarding biomarker measurements, CSF was obtained in the morning by lumbar puncture and 22 followed the pre-analytical processing described elsewhere 48 . Amyloid β-peptide1-42 (Aβ42), Amyloid 23 β-peptide1-40 (Aβ40), total tau (t-tau), and tau phosphorylated at threonine 181 (p-tau) were measured 24 by immuno-assay using the LUMIPULSE platform. Samples were run in duplicates and those 25 measurements with a coefficient of variation (CV) ≥25% were excluded. 26 MRI was performed using the Alzheimer disease Neuroimaging Initiative (ADNI) protocol 49 , by a 3T 1 scanner with regular quality control assessments. T1-weighted images were acquired for all 2 participants. Volumetric segmentation and cortical surface reconstruction were done as described 3 elsewhere 22 . In our study we analysed longitudinally cortical thickness in the precuneus, and 4 hippocampal volume. Hippocampal volume was corrected for intracranial volume as already 5 described 22 . Cortical thickness and hippocampal volume measurements were averaged across 6 hemispheres. Amyloid imaging was done using the 11C-Pittsburgh Compound B (11C-PiB) as 7 already described 22 . For the longitudinal analysis we used the total cortical Aβ uptake in the PIB-PET 8 corrected by a Regional Spread Function (RSF) as it demonstrated a better sensitivity to longitudinal 9 changes 50 . 10 All participants with genetic, clinical, CSF and neuroimaging longitudinal data that passed the quality 11 control from the 14 th data freeze were included for sTREM2 quantification in CSF. As usually done in 12 DIAN observational studies 22,51 , families carrying the APP E693G (Dutch) mutation were excluded 13 from the statistical analysis. These mutations often present with predominant cerebral amyloid 14 angiopathy and diffuse Aβ plaques with little neurofibrillary tangle pathology 52 . 15

Quantitative sTREM2 determination in CSF 16
To avoid detection of sTREM2 variants, which are created by alternative splicing and are not 17 surrogate markers for cell surface signalling competent TREM2 (supplementary Fig. 2a), we 18 developed a novel MSD-based immunoassay using a novel neo-epitope specific antibody (1H3), 19 which allows the selective detection of ADAM10/17 cleaved sTREM2 (supplementary Fig. 2 and 3  The mean intraplate CV% was 4.8% and the interplate CV% was 8.4%, calculated as the mean 26 between interplate CV% from the three included IS. To account for the interplate variability of the measurements, the sTREM2 concentrations were adjusted according to the method already described 1 for the ADNI cohort 15 . 2 3

Statistical analysis 4
The main goals were (1) to determine which factors were related with the longitudinal change of CSF 5 sTREM2 and (2) to study the effect of the longitudinal change of CSF sTREM2 on AD evolution. For 6 the first, we analysed the relationship between longitudinal CSF sTREM2 as outcome and baseline 7 markers for Aβ accumulation (CSF Aβ42, ratio CSF Aβ42/Aβ40 and total cortical PIB-PET uptake), 8 tau-related pathology (CSF t-tau and p-tau) and brain structure (cortical thickness in the precuneus, 9 hippocampal volume) as predictor variables. The second outcome was assessed by analysing the 10 correlation between the longitudinal change of CSF sTREM2 and the longitudinal change of CSF 11 Aβ42, total cortical PIB-PET uptake, CSF t-tau and p-tau, cortical thickness in the precuneus, 12 hippocampal volume, cortical FDG-PET uptake in the precuneus and cognition as measured by a 13 cognitive composite already describe elsewhere 36 . In brief, this cognitive composite comprises the 14 following tests: DIAN Word List Test, Logical Memory delayed recall, Digit Symbol Coding test 15 (total score), and Minimental Status Examination (MMSE). As secondary outcomes, we analysed the 16 relationship between baseline CSF sTREM2, as predictor, and the longitudinal change of CSF Aβ42, 17 total cortical PIB-PET uptake, CSF t-tau and p-tau, cortical thickness in the precuneus, hippocampal 18 volume, and cortical FDG-PET uptake in the precuneus as the outcomes. 19 CSF biomarker variables were log-transformed to follow a normal distribution. We calculated the 20 time point at which the CSF sTREM2 levels were significantly different in MC than in NC cross-21 sectionally as already described elsewhere 19 . Cross-sectional analysis focused on the descriptive 22 characteristics at baseline of the different clinical groups, including demographic variables and 23 biomarker values at baseline were done by chi-square tests for categorical variables, and ANOVA or 24 analysis of covariance (ANCOVA) for continuous variables. Age and sex were included as covariates 25 in the ANCOVA studying the differences between biomarkers at baseline across the different groups. 26 Presymptomatic MC were defined as MC with a CDR score at baseline equal to zero. Symptomatic 1 MC were those MC with a CDR score at baseline higher than zero. 2 We calculated the raw rate of biomarker change as the individual slope using linear regression for 3 each participant. We used this raw rate of change for illustrational purposes in figures 1, 4 and 5. In 4 order to reassure the validity of our results we defined participants with extreme rate of CSF sTREM2 5 change as those with a raw rate of CSF sTREM2 change higher or lower than the mean +/-three SD, 6 respectively (7 symptomatic MC and 2 NC). The participants with extreme rate of CSF sTREM2 7 change are described in the appendix 2 in the supplementary material. We performed the entire 8 analysis by both, excluding and including them (see appendix 2), and both sets of results were highly 9 consistent. We show in the main text the results excluding these participants and in the appendix, 10 including them. 11 We based our longitudinal analysis in Linear Mixed Effects (LME) models. Univariate LME models 12 were used to assess the influence of baseline biomarkers (predictor) on the longitudinal change of the 13 outcome biomarker. The fixed effects in the models included baseline EYO, baseline predictor-14 biomarker, time from baseline (time), interactions EYO*time and predictor-biomarker*time effects. 15 The random effects included random intercept for each family cluster, individual intercept and slope. 16 The interaction term predictor-biomarker*time was interpreted as the effect of the baseline predictor-17 biomarker on the subsequent rate of outcome-biomarker change. The models were also evaluated by 18 adjusting for baseline CSF Aβ42 and its interaction with time (when analysing the relationship between 19 CSF sTREM2 and tau-related markers), baseline CSF p-tau and its interaction with time (when 20 analysing the relationship between CSF sTREM2 and Aβ accumulation related markers) or both and 21 their interaction with time (in the case of neuroimaging). 22 The correlations between the annual rate of change of CSF sTREM2 and that of other outcomes were 23 evaluated using bivariate LME models that simultaneously model the longitudinal courses of both 24 CSF sTREM2 and another biomarker/cognitive outcome 28,29 . The bivariate LME models included the 25 covariates of baseline parental EYO, baseline CSF p-tau (this is not included if p-tau or t-tau is the 26 outcome), baseline CSF Aβ42 (this is not included if Aβ42 is the outcome) and their interaction with time. The random effects included the random intercept for family cluster and random slope for each 1 participant. Unstructured covariance matrix was used for the random effects. Changes from baseline 2 were used as the outcomes instead of the values at each visit that are usually used in LME models. 3 This was done to reduce the dimension of the covariance matrix from 4D (two random intercepts and 4 two random slopes) to 2D (only two random slopes) so that it is easier to converge. 5 The modification effect of the rate of CSF sTREM2 on the association between the rates of change of 6 Aβ-deposition and tau-related pathology markers were explored using linear or quadratic regressions. 7 The raw rate of change calculated for each participant based on linear regression were used for this 8 analysis. For each pair of biomarkers, for example CSF Aβ42 and PiB PET cortical mean, the rate of 9 change of CSF Aβ42 was used as the outcome. Baseline CSF Aβ42, baseline PiB PET cortical mean, 10 rate of change of PiB PET cortical mean (and its quadratic term if the model fits better based on AIC), 11 rate of change of CSF sTREM2 group and its interaction with the rate of change of PiB PET cortical 12 mean (and its interaction with the quadratic rate of change of PiB PET cortical mean if the model fits 13 better based on AIC) were used as predictors. The correlations between each pair of the three 14 biomarkers stratified by the CSF sTREM2 rate of change group were also explored using similar 15 bivariate LME models as described above. 16 Statistical analyses were conducted using SAS 9.4 (SAS Institute Inc., Cary, NC) and R version 3.6.1 17 (2019-07-05). All p values were based on two-sided tests and values < 0.05 were considered statistical 18 ly significant. 19

Data availability 20
All the data used in this study is available upon request from DIAN at https://dian.wustl.edu/our-21 research/observational-study/dianobservationalstudy-investigator-resources/. 22

Code availability 23
The codes used for data analysing in our study can be requested from the corresponding author.           1 cognitive decline in presymptomatic MC. a, significant association between a higher sTREM2 increase rate 2 and slower cortical shrinkage in the precuneus in the presymptomatic MC (pale violet, n =100) and a trend for a 3 similar association in the symptomatic MC group (dark red, n =48). b, shows the raw rate of cortical shrinkage 4 in the precuneus over the EYO in MC. MC were divided in two groups according to their raw rate of sTREM2 5 (above the median, in green, or below the median, in light blue). c, no significant relationship between the rate 6 of sTREM2 change and the hippocampal shrinkage rate in the presymptomatic or symptomatic MC was 7 observed. d, raw rate of hippocampal shrinkage over the EYO in MC divided in two groups according to their 8 raw rate of sTREM2 (above the median, in green, or below the median, in light blue). e, a strong correlation 9 between higher sTREM2 increase rates and slower cognitive decline in presymptomatic MC (pale violet, n 10 =100), but not in the symptomatic MC (dark red, n =48) was observed. f, raw rate of cognitive decline over the 11 EYO in MC divided in two groups according to their raw rate of sTREM2 (above the median, in green, or below