Increased amyloidogenic APP processing in APOE ɛ4-negative individuals with cerebral β-amyloidosis.

Increased APP (amyloid precursor protein) processing causes β-amyloid (Aβ) accumulation in autosomal dominant Alzheimer's disease (AD), but it is unclear if it also affects sporadic Aβ accumulation. We tested healthy controls and patients with mild cognitive symptoms (N=331) in the BioFINDER study, using cerebrospinal fluid (CSF) Aβ40 as a surrogate for amyloidogenic APP processing. We find that levels of brain Aβ fibrils (measured by 18F-flutemetamol PET) are independently associated with high CSF Aβ40 (P<0.001) and APOE ɛ4 (P<0.001). The association between CSF Aβ40 and brain Aβ is stronger in APOE ɛ4-negative than in positive people (P=0.0080). The results are similar for CSF Aβ38 and for a combination of CSF Aβ38 and CSF Aβ40. In conclusion, sporadic Aβ accumulation may be partly associated with increased amyloidogenic APP production, especially in APOE ɛ4-negative subjects. The risk for sporadic AD may consequently depend on increased Aβ production, in addition to decreased Aβ clearance.

B rain accumulation of amyloid b (Ab) is a hallmark of Alzheimer's disease (AD) which may precede dementia by up to two decades 1-3 and be quantified by cerebrospinal fluid (CSF) biomarkers or positron emission tomography (PET) imaging 4,5 . Ab accumulation is thought to be caused by an imbalance of Ab production and clearance from the brain 6 . The APOE e4 allele is the main genetic susceptibility factor for late-onset AD and sporadic Ab pathology 7 . This is likely because the APOE e4 gene product apoE4 has reduced capacity to clear Ab peptides from the brain 8 . However, Ab accumulation also occurs in the absence of APOE e4 (ref. 7) and B40-50% of AD patients lack the APOE e4 allele 9 . In autosomal dominant forms of AD, Ab pathology is believed to be caused by increased amyloidogenic processing of APP (amyloid precursor protein), that is, increased Ab production 10 but variations in APP processing have not been thoroughly explored as risk factors in 'sporadic' AD. Using a large cohort of non-demented subjects, the aim of this study was to test if APOE e4 and biomarker surrogates of amyloidogenic APP processing were independently associated with brain Ab accumulation. We used CSF levels of Ab40 to estimate amyloidogenic APP processing. The rationale for this was that Ab40 is a major Ab isoforms produced by neurons by concerted band g-secretase cleavages of APP (the same processing pathway that results in Ab42) 11 but is generally not related to Ab plaque pathology (in contrast to CSF Ab42, which is reduced in the presence of Ab plaques 12 ). Note that previous studies testing the correlation between CSF Ab40 and PET Ab have not co-varied for the presence of APOE e4. We hypothesized that there would be independent correlations between Ab accumulation and the predictors APOE e4 and CSF Ab40, and that increased amyloidogenic APP processing would be related to Ab accumulation mainly in APOE e4-negative subjects. We also hypothesized that CSF Ab40 would not be associated with APOE e4 (that is, CSF Ab40 would not be affected by apoE4-mediated impaired Ab clearance). Finally, we hypothesized to see similar results when using CSF Ab38 instead of CSF Ab40 to estimate amyloidogenic APP processing.
Our results confirmed our hypothesis. We show that 18Fflutemetamol PET levels are independently associated with high CSF Ab40 (Po0.001) and APOE e4 (Po0.001) and that the association between CSF Ab40 and brain Ab is stronger in APOE e4-negative than in positive people (P ¼ 0.0080). The results are similar when using CSF Ab38 or a combination of CSF Ab38 and CSF Ab40 to estimate amyloidogenic APP production. We conclude that sporadic Ab accumulation may be partly associated with increased amyloidogenic APP production, especially in APOE e4-negative subjects. Thus, the risk for sporadic AD may partly depend on increased Ab production, in addition to decreased Ab clearance.

Results
Cohort characteristics. The cohort consisted of 331 participants (cognitively normal controls (CN) 121, subjective cognitive decline (SCD) 102 and mild cognitive impairment (MCI) 108). Demographics and data on cognition and biomarkers are summarized in Table 1 (see Table 2 for demographics stratified by APOE status). In sum, APOE e4 positivity was more common in SCD and MCI than in CN, CSF Ab42 levels were lower in MCI compared with the other groups, and the frequency of PET Ab positivity was lowest in CN and highest in MCI. CSF Ab38 and CSF Ab40 did not differ between the diagnostic groups. APOE e4 was not associated with CSF Ab40 or with CSF Ab38 (Fig. 1). The lack of association between APOE e4 and CSF Ab40 and CSF Ab38 supports our assumption that these CSF Ab peptides are unaffected by apoE4-mediated clearance of Ab.
APOE e4 and high CSF Ab40 independently predict PET Ab. Figure 2 shows the observed PET Ab and CSF Ab40 data, with estimated slopes in the APOE e4-positive and -negative groups. In a linear regression model with PET Ab as the dependent variable, high levels of CSF Ab40 (b ¼ 1.05 Â 10 À 4 , Po0.001), APOE e4-positivity (b ¼ 0.406, Po0.001) and the interaction between APOE e4 and CSF Ab40 (b ¼ À 5.61 Â 10 À 5 , P ¼ 0.0080) were all significant predictors of continuous PET Ab. Note that since APOE e4 and CSF Ab40 were both included as predictors the main effect of CSF Ab40 indicates the effect within APOE e4-negative subjects. The significant interaction between CSF Ab40 and APOE e4 indicates that the correlation between CSF Ab40 and brain Ab was stronger in APOE e4-negative than in positive people (as seen in Fig. 2). The correlation between CSF Ab40 and PET Ab in the APOE e4-positive group was weaker than in the APOE e4-negative group, but remained significant (b ¼ 0.485 Â 10 À 4 , P ¼ 0.010). The results support the hypotheses that high CSF Ab40 and APOE e4 are independent predictors of PET Ab, and that the relationship between CSF Ab40 and PET Ab varies with APOE e4 carrier status. As expected, CSF Ab42 was a significant covariate (low CSF Ab42 was correlated with PET Ab, b ¼ À 0.00120, Po0.001), but CSF Ab40, APOE e4 and the interaction between CSF Ab40 and APOE e4 remained significant also when not adjusting for CSF Ab42 (CSF Ab40: P ¼ 0.0089; APOE e4: Po0.001; interaction: .001) were also significant predictors of PET Ab, but sex was not (P ¼ 0.23). White matter lesions (WML) were evaluated as a covariate, but were not significant (P ¼ 0.68) and were therefore excluded from the final model. We also evaluated plasma levels of Ab40 as a covariate to exclude the possibility that the results depended on peripheral APP processing. Plasma Ab40 was not a significant covariate (P ¼ 0.99) and including it in the model did not change the other estimates.
CSF Ab40 is highest in APOE e4 À PET Ab þ subjects. In a linear regression model with CSF Ab40 as the dependent variable and a four level combination of PET Ab and APOE as the independent variable, the overall highest CSF Ab40 levels were seen in the PET Ab þ & APOE e4 À group (b ¼ 732, P ¼ 0.015, compared with the reference category PET Ab-& APOE e4 À , Fig. 3). PET Ab þ & APOE e4 À subjects had 19% higher mean level of CSF Ab40 (and 26% higher median level) compared with PET Ab-& APOE e4 À subjects. The model was adjusted for When also adjusting for CSF Ab42 as a covariate the effect of PET Ab & APOE e4 was even stronger, with higher CSF Ab40 in the PET Ab

Discussion
We tested the hypothesis that biomarker surrogates of amyloidogenic APP processing (CSF Ab40 and Ab38) and APOE e4 were independent predictors of brain Ab fibril accumulation. In accordance with our hypotheses, CSF Ab40 (and Ab38 in a secondary analysis) and APOE e4 were independent predictors of PET Ab, and the effect of CSF Ab40 was strongest in the APOE e4-negative individuals. To our knowledge, this is the first study showing that increased Ab production are associated with  increased risk for sporadic brain Ab accumulation. These novel results provide indirect evidence that brain Ab pathology in humans may arise from two pathways, where one involves the APOE e4 allele (likely causing reduced apoE4-mediated Ab42 clearance), and the other involves increased amyloidogenic processing of APP. This may correspond to two pathways to sporadic AD, namely reduced clearance and increased production of Ab peptides. The amyloid cascade hypothesis postulates that Ab pathology arises due to an imbalance between Ab production and clearance 6 . It has been suggested that sporadic AD is mainly caused by poor clearance of peptides from the brain, whereas autosomal dominant AD is mainly caused by increased Ab production, especially the Ab42 variant. This is supported by a metabolic labelling study showing reduced Ab clearance in sporadic AD dementia 14 , and studies showing increased amyloidogenic APP processing in early stages of autosomal dominant AD 10 . The main cause of reduced Ab clearance in sporadic AD is likely APOE e4, since the apoE4 protein isoform has reduced capacity to clear Ab peptides compared with other apoE isoforms 8 , although it is possible that APOE e4 may also contribute to increased AD risk by other mechanisms, for example, by affecting inflammation and neuronal repair 15,16 . However, one APP gene polymorphism which reduces Ab production is associated with reduced risk of AD in the general population 17 , which provides genetic evidence that variations in APP processing may also affect the risk for sporadic AD.
The main limitation of this paper was that we used an indirect measure of APP processing, which was estimated by CSF Ab40. The rationale for this approach was that Ab40 is a major Ab isoform produced by neurons 11 , which is not directly influenced by the presence of Ab plaque pathology 12 , and is not influenced by APOE e4-mediated impaired Ab clearance. The later was demonstrated by our finding that there was no overall difference in CSF Ab40 depending on APOE e4 status (Fig. 1). Alterations in CSF Ab40 are therefore more likely to reflect differences in amyloidogenic APP processing rather than differences in Ab clearance. However, we acknowledge that there may be variations in APP processing that are not captured by CSF Ab40. We also performed analyses using CSF Ab38 (another highly expressed Ab isoform) and a combination of CSF Ab38 and CSF Ab40 (based on their molar amounts), with very similar results as when using CSF Ab40 alone, which support our findings. A more direct estimate of Ab production may be done by metabolic labelling 14 , but such methods are liable to bias due to the longitudinal drift of CSF biomarkers during continuous CSF sampling that depends on sampling frequency and volume 18 . Another limitation is that there may be other factors affecting CSF Ab40 besides variations in APP processing. For example, reduced CSF Ab40 is associated with chronic WML 19 , and WML may also be associated with Ab pathology (although this is more common in MCI 20 and AD dementia 20 than in non-demented people 21 ). It is not clear if the association between CSF Ab40 and WML is due to a direct link between Ab production and WML or if lower CSF Ab40 levels reflect reduced neuronal Ab secretion due to decreased brain activity in the presence of WML. Importantly, our results remained significant when adjusting for WML. Theoretically, CSF Ab40 could also be influenced by peripheral APP processing,  Figure 2 | PET Ab as a function of CSF Ab40 and APOE e4. Observed PET and CSF Ab40 data. Slopes are modelled from a linear regression adjusted for CSF Ab42, sex, age and diagnostic group. The shaded areas indicate 95% confidence intervals for the slopes. The dotted line indicates a cutoff for clinically significant PET Ab load (1.42 SUVR). b-coefficients (divided by 10 À 4 ) and P values for the slopes within APOE e4-positive and separately for APOE e4-negative subjects are shown in the legend. The interaction between CSF Ab40 and APOE e4 is significant, indicating that the correlation between CSF Ab40 and PET Ab differs by APOE e4 status (P ¼ 0.0080). The results did not change significantly when removing outliers (CSF Ab40410,000 ng l À 1 ).  but our results were stable when adjusting for plasma Ab40, suggesting that the effects did not depend on peripheral APP processing. We did not measure all other possible factors besides increased amyloidogenic APP that may contribute to Ab deposition in APOE e4-negative subjects. For example, other AD risk genes (including CLU and CR1) may impact Ab clearance in APOE e4-negative subjects 22 . We included several different diagnostic groups, including a SCD group. We noted that the frequency of Ab positivity in our SCD subjects (37%) was higher than in a recent large meta-analysis by Jansen et al. 3 where B22% of SCD subjects were Ab-positive, compared with B25% of CN subjects. The reason for this difference is not clear, but we noted that our SCD subjects were on average 6 years older than the Jansen subjects, which may contribute to higher frequency of Ab pathology. Furthermore, all our SCD subjects were referred to specialized memory clinics because of cognitive symptoms, while some of the Jansen SCD subjects may have been seen at other health care facilities, opening for the possibility that they had less severe complaints than the SCD subjects in our study. Another recent study on PET Ab positivity in memory clinic SCD subjects found that 57% of SCD subjects were Abpositive compared with 31% of CN, which more resembles the findings in our cohort 23 . Finally, we did not include an AD dementia group, since we know from a previous study that patients with severe AD dementia have lower CSF Ab40 than patients with mild dementia (this may reflect reduced capacity to produce Ab peptides as the disease progresses) 24 . Including an AD dementia group in this study would therefore risk confounding the relationship between CSF Ab40 and PET Ab. Until now, there have been few attempts to examine the independent roles of APOE e4 and APP processing in the development of brain Ab pathology in non-demented subjects. Previous studies did not find correlations between CSF Ab40 (or Ab38) and PET Ab (ref. 12). This is likely because they did not covary for APOE e4 (and/or CSF Ab42). Adjusting for APOE e4 is important since the relationship between CSF Ab40 and PET Ab differs between APOE e4-positive and -negative individuals. Furthermore, adjusting for APOE e4 and CSF Ab42 reduces the residual errors of the models, and some of this error may contribute to the variance of CSF Ab40. Once this error is removed the correlation between CSF Ab40 and PET Ab can be better estimated. Our results add novel information and point to different possible pathways to Ab pathology in humans. In sum, our results support the idea that sporadic Ab accumulation may be partly associated with increased amyloidogenic APP production, especially in APOE e4-negative subjects. The risk for sporadic AD may consequently depend on increased Ab production, in addition to decreased Ab clearance. This provides novel insight into disease mechanisms in AD and may be important for development of drugs targeting Ab metabolism in early stages of AD.

Methods
Study population. The study population came from the Swedish BioFINDER study (Biomarkers For Identifying Neurodegenerative Disorders Early and Reliably). All available CN and non-demented patients with mild cognitive symptoms characterized as having SCD or MCI were included.
CN subject were originally enrolled from the population-based EPIC cohort. The inclusion criteria were: age Z60 years old, MMSE 28-30, and fluent in Swedish. Exclusion criteria were: presence of subjective cognitive impairment, significant neurologic disease (for example, stroke, Parkinson's disease, multiple sclerosis), severe psychiatric disease (for example, severe depression or psychotic syndromes), dementia or MCI. All CN subjects underwent a thorough clinical assessment, including neurological, psychiatric and cognitive testing all performed by a medical doctor, in addition to MRI of the brain and relevant blood tests. The cognitive battery included MMSE, ADAS-cog (items 1-3), Trail Making A & B, Symbol Digit modalities, A quick test of cognitive speed, clock drawing, cube coping, letter S fluency and animals fluency. The medical doctor made a global assessment of whether the individual was cognitively healthy based on the test results in relation to education and age. All CN subjects had a Clinical Dementia Rating scale score of 0.
The SCD and MCI cases were recruited consecutively and were thoroughly assessed by physicians with special competence in dementia disorders. The inclusion criteria were: referred to a memory clinic due to possible cognitive impairment, not fulfilling the criteria for dementia, MMSE 24-30, age 60-80 years and, fluent in Swedish. The exclusion criteria were: cognitive impairment that without doubt could be explained by another condition (other than prodromal dementia); severe somatic disease; and refusing lumbar puncture or neuropsychological investigation. The classification in SCD or MCI was based on a neuropsychological battery and the clinical assessment of a senior neuropsychologist. The battery included tests for verbal ability (including A multiple-choice vocabulary test (SRB:1 (ref. 25  The Regional Ethics Committee in Lund, Sweden, approved the study. All subjects gave written informed consent. For more details, see ref. 13 and www.biofinder.se. PET analysis. Brain Ab was measured using 18 F-flutemetamol PET (refs 31,32). PET/CT scanning was conducted at two sites using the same type of scanner, a Philips Gemini TF 16. PET sum images from 90 to 110 min post injection were generated for the average uptake. MRI results were not used since this does not improve the quantification of 18 F-flutemetamol data 33 . The images were analysed using the NeuroMarQ software provided by GE Healthcare. A volume of interest template was applied for nine bilateral regions (prefrontal, parietal, lateral temporal, medial temporal, sensorimotor, occipital, anterior cingulate and posterior cingulate/precuneus), combined in a global neocortical composite   33 . The SUVR was the global composite tracer uptake, normalized for the mean uptake in the cerebellar cortex (note that Thurfjell et al. 35 found that 18 F-flutemetamol PET SUVR had 498% concordance with visual reads independent of which reference region that was used). Most analyses in this study used continuous PET Ab but when indicated a previously defined cutoff for Ab positivity was used (41.42 SUVR, based on mixture modelling analysis 13 ).
Cerebrospinal fluid analysis. All subjects underwent lumbar CSF sampling at baseline, following the Alzheimer's Association Flow Chart 35 . Samples were stored in 1 ml polypropylene tubes at À 80°C until analysis. CSF Ab38, Ab40 and Ab42 were analysed by ELISA assays (EUROIMMUN AG, Lübeck, Germany). All analyses were performed by board-certified laboratory technicians who were blinded for clinical data and diagnoses. The CSF samples were randomized to avoid group bias. The analyses were performed during two different runs in batch, plates 1-20 using lot no. E140224AB for Ab38, E130611AA for Ab40 and E130607AA for Ab42, and plates 21-24 using lot no. E150522BK for Ab38, E150302A1 for Ab40 and E150522AZ for Ab42. Aliquots of two different pools of CSF were used as internal control samples, with CVs of 13.8% for Ab38 for the first control with a mean of 695 pg ml À 1 and 7.9% for Ab38 for the second control with a mean of 1596 pg ml À 1 ; 17.9% for Ab40 for the first control with a mean of 1951 pg ml À 1 and 11.1% for Ab40 for the second control with a mean of 3992 pg ml À 1 ; and 16.3% for Ab42 for the first control with a mean of 227 pg ml À 1 and 15.1% for Ab42 for the second control with a mean of 216 pg ml À 1 . To assure consistency in levels between the two runs, 40 CSF samples from the first run were re-analysed in the second run.
White matter lesions. All patients were examined using a single 3T MR scanner (Trio, Siemens). Automated segmentation of WML was performed using the Lesion Segmentation Tool implemented in SPM8 (http://www.applied-statistics.de/lst.html), generating a total WML volume. Before this, manual segmentation for reference of WML was performed on FLAIR images co-registered to the native MPRAGE in four MCI patients, with the segmented volume ranging from 0.5 to 106.3 ml; the resulting optimal k based on the Dice coefficient was 0.4 (ref. 36) and was used in the subsequent automated segmentation for all participants.
Statistical analysis. We tested correlations between CSF Ab40 and PET Ab in different regression models. The main model was a linear regression model where the dependent variable was PET Ab and the independent variables were CSF Ab40, APOE e4 (dichotomous), and the interaction between CSF Ab40 and APOE e4. Second, we tested the correlation between clinically significant PET Ab accumulation and CSF Ab40 and APOE e4 in a logistic regression model with PET Ab positivity as the dependent variable. Third, we tested a linear regression model with CSF Ab40 as the dependent variable and a four level combination of PET Ab and APOE e4 as the independent variable (PET Ab-& APOE e4 À , PET Ab-& APOE e4 þ , PET Ab þ & APOE e4 À and PET Ab þ & APOE e4 þ ). All models were adjusted for age (years), sex and diagnostic group. We also adjusted for CSF Ab42 to test if CSF Ab40 was associated with PET Ab beyond CSF Ab42, and to reduce the residual error of the model, allowing a better estimate of the correlation between CSF Ab40 and PET Ab. We adjusted for WML (ml) except for when WML was clearly nonsignificant, as detailed in the results section. The primary analyses were done using CSF Ab40, but we also performed analyses using CSF Ab38 and a combination of CSF Ab38 and Ab40 (based on their molar weights, Ab38: 4129.012 g mol À 1 ; and Ab40: 4327.148 g mol À 1 (ref. 37). Statistical significance was determined at Po0.05. All analyses were done using R (v. 3.0.1, The R Foundation for Statistical Computing).