Impact of amyloid-beta changes on cognitive outcomes in Alzheimer's disease: analysis of clinical trials using a quantitative systems pharmacology model

Abstract

Background: Despite a tremendous amount of information on the role of amyloid in Alzheimer's disease (AD), almost all clinical trials testing this hypothesis have failed to generate clinically relevant cognitive effects.

Methods: We present an advanced mechanism-based and biophysically realistic quantitative systems pharmacology computer model of an Alzheimer-type neuronal cortical network that has been calibrated with Alzheimer Disease Assessment Scale, cognitive subscale (ADAS-Cog) readouts from historical clinical trials and simulated the differential impact of amyloid-beta (Aβ40 and Aβ42) oligomers on glutamate and nicotinic neurotransmission.

Results: Preclinical data suggest a beneficial effect of shorter Aβ forms within a limited dose range. Such a beneficial effect of Aβ40 on glutamate neurotransmission in human patients is absolutely necessary to reproduce clinical data on the ADAS-Cog in minimal cognitive impairment (MCI) patients with and without amyloid load, the effect of APOE genotype effect on the slope of the cognitive trajectory over time in placebo AD patients and higher sensitivity to cholinergic manipulation with scopolamine associated with higher Aβ in MCI subjects. We further derive a relationship between units of Aβ load in our model and the standard uptake value ratio from amyloid imaging. When introducing the documented clinical pharmacodynamic effects on Aβ levels for various amyloid-related clinical interventions in patients with low Aβ baseline, the platform predicts an overall significant worsening for passive vaccination with solanezumab, beta-secretase inhibitor verubecestat and gamma-secretase inhibitor semagacestat. In contrast, all three interventions improved cognition in subjects with moderate to high baseline Aβ levels, with verubecestat anticipated to have the greatest effect (around ADAS-Cog value 1.5 points), solanezumab the lowest (0.8 ADAS-Cog value points) and semagacestat in between. This could explain the success of many amyloid interventions in transgene animals with an artificial high level of Aβ, but not in AD patients with a large variability of amyloid loads.

Conclusions: If these predictions are confirmed in post-hoc analyses of failed clinical amyloid-modulating trials, one should question the rationale behind testing these interventions in early and prodromal subjects with low or zero amyloid load.

Keywords: Amyloid load; Amyloid secretase inhibition; Biologics; Modeling; Patient selection; Prevention trials; Trial failure.

Fig. 1

Modeling pipeline for simulating effect of changes in Aβ40 and Aβ42 concentrations on cognitive outcome. For each value of Aβ40 and Aβ42 (monomer, dimer and trimer) levels (left matrix), the effect on changes in excitatory–excitatory (e-e) NMDA conductance and α7 nAChR activation is derived from the coupling equations (top and bottom graphs, respectively) and the outcome on a cognitive neuronal network is calculated for the whole 17 × 17 matrix. Top figure shows the relationship between Aβ40 and Aβ42 load and e-e NMDA conductance with maximum value δ at position x₀ and slopes α and α*. Bottom figure illustrates the relationship between Aβ load and α7 nAChR inactivation with slope β. We calculate the cognitive outcome matrix for a baseline value and for six different disease states (MCI, and mild-to-moderate AD patients at time 0, 12, 26, 52 and 78 weeks). Aβ load (left matrix) of an AD patient can be defined for any region, such as the one in gray which shows a possible low Aβ load case (the high Aβ load case would be everything outside the gray area) and its cognitive status can be determined by averaging the range of results at a particular time (the gray line in the right graph). Similarly, for any change in Aβ40 and Aβ42 low-order aggregates load as a consequence of disease pathology or therapeutic interventions over time span ΔT (going from one box to another box in the Aβ load matrix), corresponding changes in ADAS-Cog can be calculated based on their impact on cognitive readout changes using their glutamate (Glu) and α7 nAChR-mediated effects. Aβ amyloid-beta, ADAS-Cog Alzheimer Disease Assessment Scale, cognitive subscale, NMDA N-methyl-d-aspartate

Fig. 2

Absolute ADAS-Cog predictions in an MCI ‘virtual patient’ for different cutoff values for Aβ + and Aβ– imaging SUVR values. Reported ADAS-Cog values are 8.5 for MCI Aβ– patients and 10.8 for MCI Aβ + patients (blue and green horizontal lines, respectively). Cutoff values define the gray area in the Aβ load matrix of Fig. 1. For the optimal parameter set for δ, α, α* and β (x₀ fixed at 2), the average ADAS-Cog prediction for Aβ– and Aβ + MCI subjects is shown by the green triangles and diamonds, respectively. At a cutoff value of 3, both predicted values (7.8 for Aβ– and 10.7 for Aβ + subjects) are near the clinical values. For the situation of δ = 0 (same values for α, α*, β and x₀), the average ADAS-Cog prediction for Aβ– patients and Aβ + patients is shown by the red triangles and diamonds  (8.4 for Aβ– and 14.2 for Aβ + subjects, respectively). In this case, while the Aβ– MCI subjects cross the blue line in the correct range, the Aβ + MCI subjects result in an ADAS-Cog readout that is much greater than the Aβ + reported clinical values (green line). This suggests that the condition δ = 0 (no beneficial effect of Aβ40) is unable to reproduce this clinical outcome. Aβ amyloid-beta, A–β subjects with x ≤ 2 and y ≤ 2, Aβ + subjects with x > 2 and y > 2, ADAS-Cog Alzheimer Disease Assessment Scale, cognitive subscale, MCI minimal cognitive impairment

Fig. 3

Dose–response of scopolamine-mediated cognitive deficits in a MCI population with low and high Aβ load corresponding to 3 units of Aβ load in our model for δ = 0.025, α = 0.002 and α* = 0.002. Slopes for the Aβ– and Aβ + conditions are −1.08 and −1.36% correct answers/nM scopolamine, respectively, for a 2-back working memory (WM) test. Model outcome suggests that higher Aβ load makes the system more sensitive to scopolamine; that is, a greater deficit for the same scopolamine dose that would correspond to the clinically observed slower recovery after scopolamine in MCI subjects. Pharmacodynamic interaction with Aβ can have important consequences as standard of care for AD patients often includes procholinergic compounds. Aβ amyloid-beta, Aβ– subjects with x ≤ 2 and y ≤ 2, Aβ + subjects with x > 2 and y > 2

Fig. 4

a Simulated outcome of APOE genotype on changes in ADAS-Cog for placebo patients with mild Aβ starting load (4 units) over 78 weeks with APOE4^+/+ genotype implemented as having lower synaptic density (−20%) and lower clearance of Aβ. Platform outcome suggests that under these conditions APOE merely drives the baseline difference but does not affect disease progression. b Slopes of glutamate coupling (expressed as change in ADAS-Cog per % change in glutamate neurotransmission) as a function of disease progression (measured from start of a trial in a mild-to-moderate AD population) and APOE4 genotype. In the pathological state, defined by a range between 12 and 34 weeks into the trial, the system is maximally sensitive to NMDA conductance changes and therefore to Aβ-mediated effects. Note that while the APOE4^+/+ genotype has a greater sensitivity to Aβ load changes in the early pathology stages, the trend is reversed at more extensive pathology. However, with greater Aβ loads seen at later pathology stages, the range available to see an effect is reduced. This would also suggest that changes in amyloid levels from therapeutic interventions tend to have a maximal effect at 10–35 weeks in a trial with mild-to-moderate AD patients. ADAS-Cog Alzheimer Disease Assessment Scale, cognitive subscale, APOE apolipoprotein E, Glu glutamate

Fig. 5

Systematic sensitivity analysis of model outcome for all key parameters of the model with x₀ fixed at 2. Each cell lists which of three conditions are met: Condition 1, outcome of 8.5 ± 1 points on ADAS-Cog in MCI Aβ– patients and 10.7 ± 1 points for Aβ + MCI subjects with cutoff value < 4; Condition 2, greater sensitivity to scopolamine in MCI Aβ + subjects; Condition 3, maximum difference of 10% between cognitive trajectory slopes for APOE4^+/+, APOE4^+/− and APOE4^−/− with the additional requirement that APOE4^+/+ genotype is at least 1.5 points worse than APOE4^−/− at time 0 weeks. Only for some cells with δ > 0.02 and (α + α*) > 0.003 are all three conditions are met simultaneously. Note that for δ = 0 (i.e., no protective effect of the short Aβ form), no case exists where all three conditions are met simultaneously

Fig. 6

a Simulated clinical outcomes in ADAS-Cog in a 78-week trial with mild-to-moderate AD patients with low amyloid baseline (amyloid load negative) after BACE inhibition (verubecestat), gamma-secretase inhibition (semagacestat) and solanezumab (Sola) antibody. Differences with placebo values shown. BACE-I results in a substantial worsening over the whole trial duration, with a somewhat lower negative effect of GSI and almost no effect of solanezumab. b Simulated clinical outcomes in ADAS-Cog in a 78-week trial with mild-to-moderate AD patients with medium to high amyloid baseline (amyloid load positive) after BACE inhibition (verubecestat), gamma-secretase inhibition (semagacestat) and solanezumab antibody. Differences with placebo values are shown. BACE-I improves cognition, with substantial benefit (1–2 points) at longer time points, while GSI has a smaller dose-dependent response (1–1.5 points). Note that higher BACE inhibition is less beneficial. Solanezumab, on the contrary, has a modest dose-dependent clinical benefit (0.5–1 points) ADAS-Cog Alzheimer Disease Assessment Scale, cognitive subscale, BACE-I BACE inhibitor, GSI gamma-secretase inhibitor, wks weeks