The deposition of amyloid-β (Aβ) oligomers in brain parenchyma has been implicated in the pathophysiology of Alzheimer’s disease. Here we present a systems pharmacology model describing the changes in the amyloid precursor protein (APP) pathway after administration of three different doses (10, 30, and 125 mg/kg) of the β-secretase 1 (BACE1) inhibitor MBi-5 in cisterna magna ported rhesus monkeys. The time course of the MBi-5 concentration in plasma and cerebrospinal fluid (CSF) was analyzed in conjunction with the effect on the concentrations of the APP metabolites Aβ42, Aβ40, soluble β-amyloid precursor protein (sAPP) α, and sAPPβ in CSF. The systems pharmacology model contained expressions to describe the production, elimination, and brain-to-CSF transport for the APP metabolites. Upon administration of MBi-5, a dose-dependent increase of the metabolite sAPPα and dose-dependent decreases of sAPPβ and Aβ were observed. Maximal inhibition of BACE1 was close to 100% and the IC50 value was 0.0256 μM (95% confidence interval, 0.0137–0.0375). A differential effect of BACE1 inhibition on Aβ40 and Aβ42 was observed, with the Aβ40 response being larger than the Aβ42 response. This enabled the identification of an Aβ42 oligomer pool in the systems pharmacology model. These findings indicate that decreases in monomeric Aβ responses resulting from BACE1 inhibition are partially compensated by dissociation of Aβ oligomers and suggest that BACE1 inhibition may also reduce the putatively neurotoxic oligomer pool.
The amyloid cascade hypothesis posits that amyloid-β (Aβ) protein peptide levels are increased early in Alzheimer’s disease (AD), leading to the formation of toxic soluble Aβ oligomers (Aβos) and plaques (Di Carlo et al., 2012). According to this hypothesis, a series of causal events initiated by abnormal Aβ levels leads to neuronal cell death and cognitive and functional decline over time (Hardy and Selkoe, 2002). Toxic Aβos are considered to be the drivers of the neurodegeneration (Benilova et al., 2012; Klein, 2013). Such soluble forms of multimeric Aβ peptides are intermediate of soluble Aβ monomers and insoluble Aβ fibrils and likely consist of a mixture of oligomeric species. Aβ dimers, trimers, larger Aβos, and structures such as soluble protofibrils have been isolated from the AD brain (Mc Donald et al., 2010; Esparza et al., 2013; Yang et al., 2013; Savage et al., 2014). Aβos are in a constant equilibrium with Aβ monomers and other Aβ aggregates (Benilova et al., 2012).
Formation of Aβ requires proteolytic cleavage of the transmembrane protein β-amyloid precursor protein (APP). Sequential cleavage of APP by the enzymes β-secretase 1 (BACE1) and γ-secretase leads to the formation of Aβ (Esler and Wolfe, 2001), as schematically depicted in Fig. 1. Here, cleavage by BACE1 leads to the formation of both the N-terminal secreted fragment soluble β-amyloid precursor protein (sAPP) β and the C-terminal membrane-bound 99-amino acid fragment (C99). C99 is subsequently subject to cleavage by γ-secretase, yielding Aβ species of different chain lengths. The most common Aβ isoforms have 38, 40, or 42 amino acids (Aβ38, Aβ40, or Aβ42, respectively) (Wiltfang et al., 2002). In parallel, full-length APP is also cleaved by α-secretase, leading to the formation of sAPPα, which is nonamyloidogenic (Portelius et al., 2011).
Aβ production in the brain is a target for AD therapy, with the potential for a disease-modifying effect by reducing Aβ levels (Husain et al., 2008). Several BACE1 inhibitors are being tested in human clinical trials, but the optimum level of BACE1 inhibition required for the treatment of AD remains to be determined (Yan and Vassar, 2014). A quantitative understanding of the effects of secretase inhibitors on the APP pathway may provide greater insights into dose-response pharmacology relationships.
Generally, measures of Aβ response in humans and primates can only be obtained in the cerebral spinal fluid (CSF) and not in the brain. However, it is believed that changes in Aβ concentrations in CSF reflect changes in brain Aβ (Lu et al., 2012). Thus, CSF Aβ serves as key biomarker for Aβ production–targeted therapies (Jack and Holtzman, 2013). The cisterna magna ported (CMP) rhesus monkey model enables longitudinal sampling in the CSF outflow from the cisterna magna in conscious rhesus monkeys. Because APP is completely homologous between humans and rhesus monkeys, the CMP rhesus monkey model is used to study the effects of secretase inhibitors (Podlisny et al., 1991; Gilberto et al., 2003; Cook et al., 2010).
Several studies on the pharmacokinetics (PK) and the pharmacodynamics (PD) of BACE1 and γ-secretase inhibitors have been reported (Lu et al., 2012, 2013; Parkinson et al., 2013; Janson et al., 2014). Liu et al. (2013) proposed a mechanistic PK-PD model of BACE1 inhibition in monkeys. They identified the β-secretase cleavage step as the rate-limiting step for Aβ formation. However, their model is a simplification of the underlying system because no distinction is made between the β-secretase and γ-secretase cleavage steps and Aβ was modeled as a direct product of APP. Potter et al. (2013) used compartmental modeling to investigate the APP processing pathway based on the results from a metabolic tracer study in humans with rare autosomal dominant AD. A model with 18 compartments accounting for the kinetics of Aβ38, Aβ40, and Aβ42 enrichments, including compartments representing APP and C99, was proposed. However, the reported model is structurally and numerically unidentifiable, considering that not all APP metabolites were measured.
To our knowledge, no systems pharmacology model has been reported that provides an integrated description of the effects of drugs on the APP metabolites. Systems pharmacology modeling is an extension of traditional mechanism-based PK-PD modeling, linking the system that is affected by the drug to its treatment-associated measured biomarkers. This involves computational analysis of the time course of the changes in biomarkers on the basis of a structural mathematical model that describes the underlying biologic processes, while making a strict distinction between drug-specific and systems-specific parameters. It has been demonstrated that such mechanism-based PK-PD models have much improved properties for extrapolation and prediction (Danhof et al., 2005, 2007). Systems pharmacology modeling will provide a quantitative understanding of the effects of drugs on the APP processing pathway to improve the prediction and magnitude of Aβ reducing effects.
The objective of this investigation was to characterize the multistep production of Aβ in the brain and its disposition into CSF in the rhesus monkey and to obtain an indirect impression of Aβο using information from the monomeric Aβ species. To this end, CSF Aβ dynamic data from CMP monkeys treated with the BACE1 inhibitor MBi-5 were analyzed. Inter-relationships of APP metabolites and their responses to MBi-5 were each measured by an enzyme-linked immunosorbent assay and metabolite responses were then integrated by means of a systems pharmacology modeling approach. Comprehensive, model-based information from MBi-5 PK and PD is integrated across time points, doses, and end points, yielding information on dose response and APP metabolite (sAPPβ, sAPPα, and Aβ) responses and inter-relationships. In this manner, invaluable information is obtained on the functioning of the integrated biologic system. The effect of BACE1 inhibition on Aβο is anticipated, which will be measured in future studies.
Materials and Methods
Animal use procedures conformed to the Guide for the Care and Use of Laboratory Animals (1996 National Research Council Institute of Laboratory Animal Resources) and were reviewed and approved by the Institutional Animal Care and Use Committee at Merck Research Laboratories. The CMP rhesus monkey model was described by Gilberto et al. (2003). The rhesus monkeys were chronically implanted with catheters in the cisterna magna, allowing repeated sampling of CSF and plasma in conscious rhesus monkeys. Six male animals, weighing between 5.2 and 11.7 kg (mean 8.7 kg), aged 2–10 years (mean 8 years), were included in the study. These monkeys were captive bred in a closed colony and individually housed.
Drug Administration and Sampling.
The study protocol and pharmacological profile of MBi-5 was described previously by Dobrowolska et al. (2014a). The study protocol is summarized here. In a single-dose, four-way, full crossover study, vehicle (0.4% methylcellulose) or MBi-5 was administered orally at 10, 30, 125 mg/kg (5 ml/kg), with at least 2 weeks washout between each period. Plasma and CSF drug concentrations were collected at 0 (predose) and 3, 5, 7, 9, 13, 16, 19, 22, 25, 28, 31, 49, 55, 58, 73, and 145 hours postdose, resulting in 17 plasma and CSF PK samples for each monkey per treatment group. Two milliliters of blood and 1 ml of CSF were collected at each time point. The concentration of MBi-5 in the plasma and CSF samples was determined using liquid chromatography–tandem mass spectrometry.
The concentrations of Aβ40, Aβ42, sAPPα, and sAPPβ were determined from CSF samples collected at −22, −20, and −1 hours (predose) and 2, 4, 6, 8, 12, 15, 18, 21, 24, 27, 30, 48, 54, 57, 72, and 144 hours postdose, giving 19 measurements of each biomarker for each monkey per treatment group. One milliliter of CSF was collected at each time point. The assays used for the concentration measurements were described previously (Sankaranarayanan et al., 2009; Wu et al., 2011).
The PK-PD model was developed and fitted to the data by means of nonlinear mixed-effects modeling using the NONMEM software package (version VI, level 2; GloboMax ICON Development Solutions, Ellicott City, MD) (Beal, 2008). This approach takes into account structural (fixed) effects and both intra- and interindividual variability. The following parameters were estimated: typical values of structural model parameters (population parameters, which define the mean value for a parameter in a population) (θ), the variance and covariance of the interindividual variability (ω2), and the variance of the residual error (σ2). A stepwise procedure was used to find the model that best fitted the data. A convergence criterion of three significant digits in the parameter estimates was used. The obtained minimum value of the objective function was used for the comparison of nested models. A decrease of 10.8 points in the minimum value of the objective function by adding an additional parameter, corresponding to P < 0.001 in a chi-squared distribution, was considered significant. The first-order conditional estimation approximation with η-ε interaction was used for parameter estimation. Random effects at the individual level were included as exponential (eη), reflecting lognormal distributions of the individual model parameters (eq. 1):(1)in which θi is the value for the ith individual; θ is the typical value for the parameter; and η is an interindividual random effect, which is assumed to follow a normal distribution with mean zero and variance ω2.
The residual variability was explored with additive and proportional error models (eqs. 2 and 3, respectively) or a combination of both (eq. 4).(2)(3)(4)where yijk is the kth observation on the jth occasion for the ith individual; f(θij) is the corresponding model-predicted observation; and ε represents the residual departure of the observed concentration from the predicted concentration, which is assumed to follow a normal distribution with mean zero and variance σ2.
To evaluate the prediction of the central tendency and distribution of the observed data by the model, a visual predictive check was performed in which the median and the 90% interquantile range of the data simulated with the developed model were plotted together with the observations. A validated result would have close agreement of median observed and predicted line, with approximately 90% of the observations falling within the 90% prediction interval.
The NONMEM software package was implemented on an Intel QuadCore (Intel Core i7 CPU860, 2.80 GHz, 3.24 GB RAM; Intel, Santa Clara, CA) and Compaq Visual Fortran (version 6.6; Compaq Computer Corporation, Houston, TX) was used as compiler. Data management and model assessment was done using the S-PLUS for Windows statistical software package (version 8.0 Professional; Insightful Corp., Seattle, WA).
The systems pharmacology model of MBi-5 was developed by sequential analysis of PK and PD data. The PK model of MBi-5 was based on simultaneous analysis of plasma and CSF PK data. The results of the PK data analysis are provided in the Supplemental Material. The PK profiles of MBi-5 observed in plasma and CSF were adequately described by a three-compartmental model. (Supplemental Figs. 1 and 2) and the PK parameters were estimated with good precision (Supplemental Table 1); thus, the model could serve as input for PD model analysis.
The biomarker response profiles of MBi-5 measured in CSF were adequately described by a model containing compartments for five variables: APP, sAPPβ, sAPPα, Aβ40, and Aβ42 (Fig. 2). The production of APP was believed to be zero order (i.e., a constant production of APP). It was assumed that there is no alternative proteolytic enzyme cleaving full-length APP other than α-secretase and BACE1. Because both sAPPβ and C99 are products of APP cleavage by BACE1, sAPPβ and C99 were presumed to follow the same kinetics; therefore, sAPPβ could be used in the model as a surrogate precursor for Aβ. The production of sAPPα, sAPPβ, and Aβ were assumed to be first order (i.e., dependent on the concentration of its precursor). The interaction between APP, sAPPβ, sAPPα, Aβ40, and Aβ42 is described by eqs. 5–9:(5)(6)(7)(8)(9)
The rate of change of APP with respect to time in the presence of the inhibitor is described by eq. 5, in which the BACE1 cleavage inhibition is incorporated by the factor EFF. EFF is the degree of inhibition caused by MBi-5, expressed as shown in eq. 10.(10)where Ctarget is the target site concentration of MBi-5, IC50 is the Ctarget that results in 50% inhibition of BACE1, Imax is the maximum response, and GAM is the Hill coefficient. Ctarget was derived from the PK model as shown in eq. 11:(11)where AUCCSF and AUCplasma are the areas under the CSF and plasma concentration time curves, respectively. Here, Ctarget is assumed to follow the same profile as Cplasma but at a level between CCSF and Cplasma.
It is assumed that the system is in steady state when no treatment is given (EFF = 1). At the treatment free state, the change of the variables with respect to time is shown in eq. 12:(12)
These steady-state conditions were used to derive part of the system parameters. From eqs. 5 and 12, and it follows that the source of APP is shown in eq. 13:(13)where APPbase is the baseline level of APP, which is assumed to be equal to the sum of the baseline levels of sAPPα and sAPPβ, because it was assumed that there is no alternative proteolytic enzyme cleaving full-length APP other than α-secretase and BACE1.
From eqs. 8 and 12, the Aβ40 formation rate (Kin40), equivalent to a γ-secretase cleavage step can be calculated as shown in eq. 16:(16)where Aβ40base is the baseline level of Aβ40. sAPPβbase is the baseline level of sAPPβ, used here as surrogate for the baseline level of C99.
From eqs. 9 and 12, with substitution of Kout from eq. 16, the Aβ42 formation rate (Kin42), equivalent to a γ-secretase cleavage step, is deduced as shown in eq. 17:(17)where Aβ42base is the baseline level of Aβ42.
The model structure includes four transit compartments (Fig. 2), one for each biomarker measured in CSF (sAPPα, sAPPβ, Aβ40, and Aβ42), to account for transport from the target site in the brain to CSF. These transit processes are described, in general, by eq. 18:(18)where Kt is the transit rate for the particular species (KtAP for sAPPα and sAPPβ and KtAB for Aβ40 and Aβ42).
The system defined above can now be extended to incorporate an Aβο pool for Aβ42 oligomerization. The addition of the Aβο pool to the model structure requires adaptation of eq. 9, describing Aβ42 dynamics. The exchange between the Aβο pool and the Aβ42 compartment is described by eqs. 19 and 20:(19)(20)where Kpl and Krev are the Aβ42 oligomerization and dissociation rate, respectively, which are dependent on the baseline values of Aβ42 and the Aβο pool (Aβ42base and Aβοbase, respectively) according to eq. 21:(21)
Separate Empirical Models Described Response of Each APP Metabolite.
Initially, empirical PK-PD models were developed to quantify the exposure-response relationships for each CSF APP metabolite (Aβ40, Aβ42, sAPPα, and sAPPβ) of the BACE1 inhibitor MBi-5 in monkeys. The typical model structure of each APP metabolite-inhibitor combination consisted of a transit model with one or two compartments, with the drug effect modeled relative or subtractive to baseline using an Imax/Emax function. A summary overview of the results of these models is depicted in Table 1. The empirical models provided consistency of drug effects across APP metabolites, with identified potencies [95% confidence intervals (CIs)] of 0.0254 μM (0.0246–0.0262) for Aβ40, 0.0455 μM (0.0351–0.0559) for Aβ42, 0.0490 μM (0.0192–0.0788) for sAPPβ, and 0.0265 μM (0.0135–0.0395) for sAPPα. The mean transit time through the compartments of the models was lower for Aβ40 and Aβ42 than for sAPPβ and sAPPα. This indicates that the response of Aβ40 and Aβ42 will appear earlier in CSF, even though sAPPβ is a sequentially earlier product of the amyloidogenic APP pathway.
A Systems Model to Describe APP Metabolite Responses.
A comprehensive compartmental PK-PD model, incorporating MBi-5 PK and CSF APP metabolite (Aβ40, Aβ42, sAPPα, and sAPPβ) concentrations was developed to quantify APP metabolite responses to BACE1 inhibition in monkeys. The model is schematically presented in Fig. 2. The model describes production, elimination, and brain-to-CSF transport of each APP metabolite, as well as their inter-relationships (Fig. 3). The rate of APP metabolism was assumed to be close to the maximal capacity of the enzymes involved (Nelson and Cox, 2000). Thus, APP production was approximated to follow zero-order kinetics. sAPPβ was used in the model structure as a surrogate substrate for C99 in the γ-secretase cleavage step. Because both sAPPβ and C99 are products of the APP cleavage by BACE1, their formation rates should be the same; thus, use of sAPPβ as a surrogate for C99 was justified. To account for transport from the target site in the brain to CSF, the model included one transit compartment for each APP metabolite. The drug effect was incorporated in the model as the inhibition of loss of the APP precursor pool, equivalent to the BACE1 cleavage step.
MBi-5 Increased sAPPα and Decreased sAPPβ and Aβ in a Dose-Dependent Manner.
APP metabolite CSF concentrations showed a dose-dependent response in the presence of the BACE1 inhibitor. The dose-dependent increase of sAPPα and the corresponding decreases of sAPPβ and Aβ were described by the model with a single drug effect. A potency (IC50) of 0.0256 μM (95% CI, 0.0137–0.0375) was identified. This value is close to the in vitro inhibition constant (Ki) of 10 nM for MBi-5 inhibition of purified BACE1 and is also close to the IC50 for inhibition of Aβ production in intact cells of 24 ± 6 nM (Dobrowolska et al., 2014a). When estimated, the maximal inhibition (Imax) was close to 1. Therefore, Imax could be fixed to 1, indicating 100% inhibition of BACE1 at sufficient high drug concentrations. Figures 4–7 show the model description of each APP metabolite for each dose group.
AβΟ Pool Required to Account for Differential Effect on Aβ40 and Aβ42.
A differential effect of BACE1 inhibition was observed for Aβ40 and Aβ42: a higher response is observed in the data for Aβ40 than for Aβ42 (Fig. 7, E and G). This differential effect could be described by extending the model with an Aβο pool connected to the Aβ42 compartment, resulting in an adequate description of sAPPβ, sAPPα, Aβ40, and Aβ42 CSF concentration time profiles for each dose group (Figs. 4–7, respectively). Incorporating the Aβο pool in the model improved the description of Aβ40 response for the 30- and 125-mg/kg dose group (Figs. 6, E and F, and 7, E and F), as well as the description of the 125-mg/kg dose for Aβ42 response (Fig. 7, G and H). Furthermore, the description of the sAPPβ response for the 125-mg/kg dose (Fig. 7, C and D) was improved. Exchange of an Aβ40 monomer pool with an Aβο pool was evaluated but could not be identified.
The population parameters and intra- and interanimal variability were optimized for the study population and are depicted in Table 2. To select the best random-effects model structure, a sequence of models with interanimal variability on different parameters was tested and the results were compared. The final model included interanimal variability for the baseline of sAPPβ and the IC50 of MBi-5. Both were included as exponential in nature, reflecting lognormal distributions of the individual model parameters. Because the baselines of the other APP metabolites were modeled as function of the baseline of sAPPβ, its interanimal variability reflects also on the other baselines. Residual, or unexplained, variability was described for each APP metabolite (sAPPb, sAPPa, AB40, AB42) separately. Residual variability was assumed to follow a normal distribution. Proportional errors were used to describe the residual variability for each APP metabolite. The identified residual variability was higher for Aβ40 and Aβ42 than for sAPPβ and sAPPα. System-specific parameters could be distinguished from drug-specific parameters (all correlations < 0.95).
Incorporating the Aβο pool into the model improved the description and did not affect the parameter estimate of the IC50 significantly: with the Aβο pool, an IC50 of 0.0269 μM (95% CI, 0.0154–0.0384) was identified; without the Aβο pool, the IC50 was 0.0256 μM (95% CI, 0.0137–0.0375). The incorporation of the Aβο pool affected the Hill coefficient of the sigmoidal Imax concentration response relationship. The Aβο pool resulted in a Hill coefficient slightly deviating from unity: with the Aβο pool, a Hill coefficient of 1.53 (95% CI, 1.14–1.92) was identified; without the Aβο pool, the Hill coefficient was 1 (fixed). This mainly improved the description of the APP metabolite concentration response curves for the higher dose groups (Fig. 7).
Higher Brain-to-CSF Transport of Aβ.
It was not possible to separate the rate of the γ-secretase cleavage from the brain-to-CSF transport. The transit rate for Aβ40 and Aβ42 was assumed to be equal and fast. Therefore, the transit rate from the brain to the CSF for Aβ40 and Aβ42 was fixed to an arbitrary high value (10 h−1). The Aβ half-life of 0.7 hours reflects delays as a result of the γ-secretase cleavage step and brain-to-CSF transfer. For sAPPβ and sAPPα, the transit rate was estimated to be 0.0985 h−1. This value should be interpreted relative to the Aβ transit from the brain to the CSF. For sAPPα, the brain turnover (0.8 hours) could be distinguished from the half-life of the brain-to-CSF transfer (7.0 hours). Aβ is transported from the brain to the CSF approximately 102-fold faster than sAPPα. As a result, the response of Aβ to drug treatment will appear earlier in the CSF than the response of sAPPα, even though sAPPα is a sequentially earlier product of the APP pathway.
It was not possible to identify the brain turnover of sAPPβ as a separate parameter. In the model structure, sAPPβ was used as a surrogate substrate for C99 in the γ-secretase cleavage step, driving the response of Aβ. Therefore, the γ-secretase cleavage step could not be separated from sAPPβ elimination.
APP Metabolite Inter-Relationships.
The proteolytic cleavage rates of APP through the action of BACE1 (Rinβ) and α-secretase (Rinα) were calculated from the model parameters according to eqs. 14 and 15 to be 0.314 h−1 and 0.404 h−1, indicating that 56% of full-length APP is cleaved by α-secretase and 44% by BACE1. The formation rates of Aβ40 and Aβ42 were calculated according to eqs. 16 and 17. The higher Kin40 (0.574 h−1) than Kin42 (0.020 h−1) is in line with previously reported findings of the ratio between Aβ42 and Aβ40 of about 1:10 in the non-AD brain (Iwatsubo et al., 1994). A difference in Aβ40 and Aβ42 degradation rate (Kout) was also evaluated during the model development process, but this could not adequately capture the response profile of Aβ42.
The developed model could be used to predict biomarker inter-relationships in response to BACE1 inhibition and visualize the response of APP and Aβο (Fig. 3A) in brain. APP increases after BACE1 inhibition and appears to be shunted down the α-secretase pathway, resulting in an increase of sAPPα product. The elevation in sAPPα in the data drives the modeling conclusion that there is some increase in APP in the setting of BACE1 inhibition but it is fairly modest. The Aβο level decreases after BACE1 inhibition, indicating that there is reduced formation of Aβο by reduced levels of monomeric Aβ42 and that Aβο dissociates to monomeric Aβ42. The latter influences the shape of the Aβ42 response curve, which is different than the shape of the Aβ40 response curve.
For sAPPβ, sAPPβ, Aβ40, and Aβ42, the time courses of brain versus CSF responses were predicted (Fig. 3, respectively), showing that the earlier appearance of Aβ response in CSF relative to sAPPα and sAPPβ arises from the slower brain-to-CSF transfer.
A systems model of the APP processing pathway was developed describing the inter-relationships of Aβ40, Aβ42, sAPPα, and sAPPβ upon inhibition of BACE1 with MBi-5. All four APP metabolites provided consistent information regarding drug potency. The MBi-5 concentration-dependent decrease of the APP metabolites could be described by incorporating a single drug effect in the model: inhibition of the formation rate of sAPPβ, equivalent to the BACE1 cleavage step. The model supported the notion that MBi-5 can provide 100% inhibition of BACE1, consistent with a complete blockage of Aβ production at high drug concentrations. The fact that MBi-5 can provide 100% inhibition of BACE1 implies that there are no inherent mechanistic limitations of the APP pathway to blocking Aβ production. Therefore, reaching complete inhibition largely depends on drug properties such as having sufficient potency, bioavailability, and tolerability.
The relatively large interanimal variability identified on the baseline level of sAPPβ (also reflecting on the baseline levels of the other APP metabolites) and the IC50 probably denotes the large intrinsic biologic differences in APP processing between subjects. Residual variability was higher for Aβ40 and Aβ42 than for sAPPβ and sAPPα. Residual variability represents the uncertainty in the relationship between the concentrations predicted by the model and the observed concentration and includes any model misspecification error. The higher residual variability for Aβ could be related to the second cleavage step by γ-secretase, yielding Aβ. In this analysis, no direct information was available regarding the γ-secretase cleavage step. This would require data from a γ-secretase inhibitor study. Such data may explain some of the residual variability identified for Aβ.
The systems model identified a higher brain-to-CSF transport of Aβ compared with sAPPα. This result is consistent with the identified lower mean transit time for Aβ40 and Aβ42 compared with sAPPα in the separate empirical models for each APP metabolite. The potencies identified in the empirical models were consistent with the single potency of 0.0256 μM (95% CI, 0.0137–0.0375) identified using the systems model.
In the systems model, the APP production was approximated to follow zero-order kinetics. In reality, APP production is regulated by various factors (above all, the synaptic activity) (Cheng et al., 2014). In this analysis, no quantitative data on the factors involved in APP production were available. It was assumed that the APP production was close to the maximum. Consequently, subtle changes in APP regulation would have little effect on APP metabolite formation.
Our modeling results imply that 56% of full-length APP is cleaved by α-secretase and 44% by BACE1. To our knowledge, there are no quantitative data available from the literature on the ratio of APP moving down the α-secretase pathway and BACE1 pathway. Dobrowolska et al. (2014b) compared sAPPα and sAPPβ levels in human CSF from the lumbar region from cognitively normal and AD participants. They identified an sAPPβ/sAPPα ratio of 0.59 ± 0.4 (n = 15) in cognitively normal healthy controls. Wu et al. (2011) reported sAPPα and sAPPβ levels of 37.1 pmol/g and 50.8 pmol/g, respectively, in human brain cortex samples from elderly subjects without AD (n = 16) (Wu et al., 2011), resulting in a ratio of 0.73. However, both ratios do not directly reflect the ratio of APP cleaved by α-secretase and BACE1, because the steady-state sAPPβ and sAPPα levels in the brain are the result of multiple processes such as production, degradation, and transfer from the brain to the CSF. Levels in the CSF are also affected by transfer through the lumbar region. In our analysis, the developed model facilitated the separation of the different processes involved.
The systems analysis points to a difference in biology of Aβ40 and Aβ42. First, a lower formation rate was identified for Aβ42 than for Aβ40. This is consistent with the composition of Aβ species reported for human CSF, in which Aβ40 is the dominant isoform (Murphy et al., 1999). The relative production of Aβ40 and Aβ42 is probably regulated through changes in the γ-secretase cleavage site (Murphy et al., 1999; Dolev et al., 2013). Second, the model included an Aβο pool in the brain for Aβ42 but not Aβ40. Inclusion of exchange of an Aβ42 monomer pool with an Aβο pool could account for the differential effect of MBi-5 on Aβ40 and Aβ42 response observed in the data, in which the response for Aβ40 was higher than for Aβ42. Without incorporation of the Aβο pool into the model, the Aβ40 and Aβ42 response both could not be described adequately by a single drug effect. The identification of this Aβο constitutes the scientific basis for the identification of BACE1 inhibitor effects on higher-ordered amyloid species. Because of the dissociation of Aβ42 oligomers to Aβ42 monomers during BACE1 inhibition, the response for Aβ42 was lower than for Aβ40. The differential effect of the Aβο pool on Aβ42 is consistent with the biology of oligomer and plaque formation in which Aβ42 plays a more significant role than Aβ40 (Di Carlo et al., 2012). Aβ42 is the major constituent of plaque and other species such as oligomers (Jarrett et al., 1993; Iwatsubo et al., 1994).
The baseline level of the Aβο pool estimated by the model of 278 pM (1255 pg/ml) should be interpreted as the level of Aβ42 monomers that are incorporated in the “oligomer soup” in the brain (i.e., Aβ dimers, trimers, and high molecular weight species) (Benilova et al., 2012). Here, no distinction is made between oligomeric species, because the Aβο pool is modeled as a pool in equilibrium with monomeric Aβ42 without correction for the number of subunits in multimeric species comprising the Aβο pool.
The Aβο dissociation rate (Krev) of 1.308 × 10−5 s−1 identified here is relatively slow compared with the dissociation of Aβ aggregates acquired from in vitro analysis techniques. Grüning et al. (2013) detected the appearance of monomers from Aβ42 and Aβ40 protofibrils: the dissociation rate was 1.4 × 10−4 and 1.2 × 10−4 s−1 for Aβ42 and Aβ40, respectively. Narayan et al. (2012) reported Aβ40 fibrils releasing soluble Aβ40 species at a rate of 9.3 × 10−5 s−1. Sánchez et al. (2011) identified Aβ monomer off rates of 0.6 × 10−2 and 1.0 × 10−2 s−1 for Aβ40 and Aβ42 fibrils, respectively. It is difficult to compare rates obtained in vivo with those determined using in vitro approaches, because the in vitro experimental settings can have a major effect. Moreover, the comparability of the dissociation rates to the value obtained in our analysis is limited, because no particular oligomeric species was characterized in the systems pharmacology approach.
Quantitative measurements of the response of Aβο to drugs targeting the APP pathway, such as BACE1 inhibition, are of interest. Initial results of a new sensitive Aβ oligomer assay (Savage et al., 2014) suggest an Aβο baseline level of approximately 1.5 pg/ml in rhesus CSF from the cisterna magna, which constitutes different oligomeric species (Savage and Kalinina, unpublished observations). This number cannot be compared directly with the model-derived Aβο baseline of 1255 pg/ml representing Aβ subunits in the “oligomer soup” in the brain. In addition, the Aβ-oligomer assay may not pick up all oligomeric species or Aβο may dissociate to monomers during sample preparation. Furthermore, it remains to be seen whether CSF Aβο measurements accurately reflect the brain Aβο concentrations (Rosenblum, 2014). The hydrophobicity of oligomers may make them very low or absent in aqueous fluids such as CSF (Yang et al., 2013). Recent data suggest that human CSF Aβοs range between 0.1 and 10 pg/ml and human brain Aβο levels are 252 pg/ml in the AD brain and 87 pg/ml in the control brain (Savage et al., 2014). Other data suggest 1000-fold higher concentrations of Aβο than monomers in the soluble fraction of the human AD cerebral cortex (Yang et al., 2013). Relative concentrations in the CSF will not necessarily reflect the relative concentrations in the brain because oligomers are likely cleared to the CSF much more slowly than monomeric Aβ. Additional dose-ranging studies of BACE1 inhibition in rhesus monkeys, in which Aβο response is quantified, are ongoing. It is anticipated that including such data in the systems pharmacology model analysis will elucidate the relationship between the Aβο pool in the model and measurements of Aβο.
The identified Aβο pool should be interpreted with caution because an Aβο pool in rhesus monkeys may differ from an Aβο pool in patients with AD with plaque burden. Rhesus monkeys do not develop dementia and neurodegenerative changes that characterize AD (Heuer et al., 2012). It is almost certain that the rhesus monkeys used in this study had far less amyloid deposition than a human patient with AD. Therefore, the most crucial question is the nature of the identified Aβο pool in rhesus monkeys and its pathologic relevance to AD in humans. Aβ oligomerization is a separate aggregation event. Certain oligomers are off-pathway species that do not further aggregate to amyloid fibrils. Coexistence of several oligomeric populations that do or do not propagate into fibrils is possible. If it can be demonstrated that there is a relationship between the soluble Aβο identified in rhesus monkeys and the AD brain-derived soluble Aβο and this could be correlated with neurotoxicity, then the relevance of Aβο in rhesus monkeys would be indisputable.
The lack of success of clinical trials targeting the APP pathway has been ascribed to the failure to reduce the level of toxic Aβο (Rosenblum, 2014). Plaques, toxic Aβο, and Aβ peptides should be targeted to significantly reduce soluble Aβ load because of the relationship between these three. It has been hypothesized that by decreasing Aβ levels, soluble Aβο amounts are also reduced, in turn inducing the release of Aβο from plaques to restore the balance between Aβο in the plaques and the extracellular environment (Rosenblum, 2014).
The ability to identify and estimate the oligomerization effect through modeling suggested that these efforts to model the monomer pathway may also provide information on the higher-ordered amyloid species. The ability to see this effect suggested that Aβ production inhibition by MBi-5 may also have the ability to draw down these forms as well as inhibit Aβ de novo production. In a study with APP transgenic mice, it has been demonstrated that BACE1 inhibition reduces amyloid plaque load (Kennedy and Hide, unpublished observations). This implies that if monomeric Aβ levels decrease as a result of blocked Aβ production, Aβοs dissociate to restore the equilibrium between monomeric Aβ and Aβο. To confirm this, incorporation of Aβο data into the model using rhesus data is ongoing.
A comprehensive model of the APP pathway describing the effects the BACE1 inhibitor MBi-5 has been established, taking into account the kinetics and inter-relationships of sAPPα, sAPPβ, Aβ40, and Aβ42. The effect of BACE1 inhibition was incorporated into the model as inhibition of the formation rate of sAPPβ. Because sAPPβ and C99 are both products of the same BACE1 cleavage step, the response of sAPPβ could be used as driver of Aβ response. However, sAPPβ and C99 could be subjected to different elimination processes because C99 remains membrane bound (Selkoe, 1999). The fact that the Hill coefficient of the concentration response relationship slightly deviates from unity may be a reflection of this simplification of the underlying biologic system. To adequately separate the sequential cleavage steps of BACE1 and γ-secretase from other processes involved, data from a γ-secretase inhibitor study in CMP rhesus monkeys (Cook et al., 2010) will be added to further inform the model (ongoing).
Because BACE1 is the initiating enzyme in Aβ production, its inhibition has been proposed to decrease the amount of cerebral Aβ and to subsequently prevent the development of Aβ-associated pathologies (Cole and Vassar, 2007). With the developed systems pharmacology model, a deeper comprehension of the effects of β-secretase inhibition on the APP processing pathway and the anticipated effect on Aβο was gained. Understanding these effects early in preclinical development could improve the anticipation of the magnitude of Aβ reducing effects in humans. The model forms the first step in developing a translational systems model to predict possible Aβ response of new drug candidates in humans, based on their estimated potency in rhesus monkeys.
The authors thank Randall J. Bateman and Justyna A. Dobrowolska (Department of Neurology, Washington University, St. Louis, MO) for input on study design and data interpretation and Eric M. Parker (Merck Research Laboratories, Whitehouse Station, NJ) for critical comments on the manuscript.
Participated in research design: Michener, Savage, Kennedy, Kleijn, Stone.
Conducted experiments: Michener.
Performed data analysis: van Maanen, van Steeg, Stone.
Wrote or contributed to the writing of the manuscript: van Maanen, van Steeg, Savage, Kennedy, Stone, Danhof.
- amyloid-β oligomer
- Alzheimer’s disease
- β-amyloid precursor protein
- β-secretase 1
- C-terminal membrane-bound 99-amino acid fragment
- confidence interval
- cisterna magna ported
- cerebrospinal fluid
- soluble β-amyloid precursor protein
- Copyright © 2016 by The American Society for Pharmacology and Experimental Therapeutics