Anti‐amyloid antibody equilibrium binding to Aβ aggregates from human Alzheimer's disease brain

Abstract

INTRODUCTION

Lecanemab binds “protofibrils,” which are poorly characterized in human brain. It is unknown why lecanemab caused fewer amyloid‐related imaging abnormalities (ARIAs) than other antibodies in trials. The apolipoprotein E (APOE) ε4 allele increases ARIA risk through unknown mechanisms.

METHODS

Equilibrium binding constants (K_D ) and total amyloid beta (Aβ) binding (B _max) of aducanumab, lecanemab, and donanemab equivalents to soluble and insoluble amyloid plaque‐enriched and cerebral amyloid angiopathy (CAA)‐enriched Aβ were compared across 17 Alzheimer's disease (AD) cases by mixed models. Titrated immunofluorescence (IF) staining compared antibody binding.

RESULTS

Lecanemab and aducanumab had indistinguishable preference for “protofibrils.” Antibody preference for plaque‐enriched versus CAA‐enriched Aβ did not differ in soluble extracts or by IF staining but differed slightly in insoluble extracts. The APOE ε4 allele was associated with more soluble antibody‐accessible Aβ.

DISCUSSION

Lecanemab's binding target is similar to other antibodies’. Differences in antibody preference for plaque versus CAA Aβ may not explain differences in ARIA with edema rates.

Highlights

• We tested binding of clinical anti‐amyloid antibodies to human brain amyloid beta (Aβ).

• Lecanemab, donanemab, and aducanumab bound similar populations of Aβ.

• The apolipoprotein E ε4 allele increased solubility of antibody‐bound Aβ.

1. BACKGROUND

The introduction of disease‐modifying therapy is changing the standard of care for Alzheimer's disease (AD). Three anti‐amyloid antibodies, aducanumab, lecanemab, and donanemab, have received US Food and Drug Administration approval to treat early AD. Lecanemab and donanemab have entered clinical use. All three antibodies target misfolded, aggregated amyloid beta (Aβ), the principal component of amyloid plaques. All three clear amyloid plaques as measured by amyloid positron emission tomography (PET) and avoid binding to monomeric Aβ. However, all three are thought to derive their preference for misfolded, aggregated Aβ over monomeric Aβ through different mechanisms. Donanemab binds to pyroglutamate‐3 Aβ, a posttranslational modification found in plaques but not newly synthesized Aβ monomers. ¹ Lecanemab and aducanumab both bind conformational epitopes of misfolded Aβ. Between lecanemab and aducanumab, lecanemab is purported to have greater specificity for toxic “protofibrils,” rather than “mature” amyloid fibrils, and this distinction is postulated to account for its better observed efficacy. ² The more soluble, diffusible pool of Aβ aggregates in human AD brains have correlated better with dementia than the total or insoluble Aβ, ³ , ⁴ , ⁵ , ⁶ making it plausible that a more soluble Aβ (protofibril)‐preferring antibody could be more effective. Binding preferences of all three antibodies have been published almost entirely using synthetic in vitro Aβ assemblies, rather than Aβ aggregates derived from human brain tissue. ² , ⁷ , ⁸ Structures of in vitro amyloids may differ from in vivo ones, ⁹ , ¹⁰ and therefore binding properties to in vitro assemblies do not infer binding properties to those present in the human brain. No structural differences between soluble and insoluble Aβ from the human brain have been demonstrated; rather, we have found identical fibrillar structures in both. ¹¹ , ¹² Whether and how lecanemab has a soluble Aβ preference in the human brain compared to other anti‐Aβ antibodies is unclear.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using PubMed and conference proceedings. There are conflicting data regarding the binding preferences of anti‐amyloid antibodies to different amyloid beta (Aβ) forms in the human brain and insufficient understanding of antibody binding characteristics as explanators of trial results.

2. Interpretation: Our findings are inconsistent with the literature that lecanemab selectively binds “protofibrils” in the human brain, and inconsistent with differences in amyloid‐related imaging abnormality with edema (ARIA‐E) rates being due to differences in cerebral amyloid angiopathy versus amyloid plaque binding. The apolipoprotein E (APOE) ε4 allele may increase ARIA‐E risk by increasing the solubility of antibody‐accessible Aβ.

3. Future directions: Antibodies with greater “protofibril” preference may not be more efficacious. Future work may focus on modulating Aβ solubility independent of its structure. Contributors to ARIA‐E rates beyond antibody affinity, such as effector function and titration schemes, should continue to be evaluated. Future hypotheses for how the APOE ε4 allele increases ARIA include increased Aβ solubility.

All plaque‐clearing anti‐amyloid antibodies cause amyloid‐related imaging abnormalities with edema/effusions (ARIA‐E), a side effect that can rarely be fatal, and the cause of ARIA‐E is not currently known. One hypothesis for the cause of ARIA‐E is inflammation and blood–brain barrier breakdown due to binding of anti‐amyloid antibodies to cerebral amyloid angiopathy (CAA), Aβ aggregates in the tunica media of brain arterioles and small arteries instead of (or in addition to) in the neuropil. ¹³ Support for this hypothesis comes from the observations that magnetic resonance imaging (MRI)‐detectable lobar microhemorrhages, usually caused by CAA in the AD population, increase risk for ARIA‐E; that ARIA‐E exhibits clinical and radiographic similarity to CAA‐related inflammation (CAAri), a sporadic disease possibly mediated by endogenous anti‐Aβ autoantibodies; and autopsies of rare fatal ARIA‐E cases have demonstrated vasculitis. ¹⁴ , ¹⁵

These observations can be compiled into a theory in which (1) lecanemab, aducanumab, and donanemab bind structurally different classes of Aβ aggregates in the human AD brain, and (2) observed differences in ARIA‐E rates in clinical trials (aducanumab > donanemab > lecanemab) are attributable to differences in binding preference to CAA Aβ versus plaque Aβ. In the current study, we sought to test predictions of this theory: (1) certain populations of Aβ aggregates from the human AD brain (i.e., “protofibrils”) will be accessible to binding to lecanemab but not aducanumab or donanemab, (2) lecanemab will exhibit greater binding affinity to aqueously extracted (“soluble,” “protofibrillar”) Aβ aggregates than aducanumab, and (3) binding preference of all three antibodies for CAA‐enriched versus plaque‐enriched Aβ extracts from the human brain will follow the inverse rank order of observed ARIA‐E rates in phase 3 clinical trials.

2. METHODS

2.1. Case selection

For antibody equilibrium affinity determination, we selected 18 cases from the Brigham and Women's Hospital Neuropathology Tissue Repository and Clinical Data Bank with the presence on the autopsy report of both amyloid plaques and CAA, to represent a population of patients who might undergo therapy with anti‐amyloid immunotherapy. We studied the occipital lobe parenchyma and meninges because ARIA‐E most typically occurs in the occipital lobe. As expected for a population enriched for CAA, the apolipoprotein E (APOE) genotypes were enriched for the ε4 allele. All cases subsequently underwent Consortium to Establish a Registry for Alzheimer's Disease (CERAD) plaque staging ¹⁶ and Olichney CAA staging ¹⁷ in the occipital lobe. One case was found to have sparse plaques only in the temporal lobes and none in the occipital lobes not rising to the diagnostic criteria for AD; this case was excluded for lacking the neuropathological findings associated with amyloid PET positivity. ¹⁸ Three cases had definite pathologic diagnoses of CAA reflected in focal or multifocal areas throughout the cortex, but had occipital Olichney scores of 0. These cases were included due to the patchy nature of CAA and the likelihood that at least some unsampled CAA was present in the occipital lobe; these cases would likely have had positive amyloid PETs and been considered for anti‐amyloid therapy. The final sample size was thus 17.

We extracted Aβ in two ways as described in detail in the supporting information: (1) an aqueous extraction composed of material in the supernatant after homogenization in Tris‐buffered saline (TBS) followed by centrifugation, and (2) an insoluble extraction from the pellet of the first extraction. Separate extracts were prepared from the gray matter of the occipital lobe and from the overlying leptomeninges. Validation of centrifugation speeds and Aβ42:40 ratios are discussed in the Supplementary Methods and in Figures S1 and S2 in supporting information.

For tissue immunofluorescence (IF), we selected five separate pathologic AD cases from which we had paraffin‐embedded blocks of occipital cortex that had been fixed for only 2 hours in 10% neutral buffered formalin to preserve antibody epitopes. These cases were banked prior to the widespread adoption of Braak staging. A Braak stage was inferred from the pathologic description. APOE genotypes, CAA, and CERAD staging were not available for these cases.

All case information is described in Table 1.

TABLE 1.

Human case characteristics.

Equilibrium binding study
Case	Age (y)	Sex	APOE genotype	Braak stage	Overall pathologic CAA diagnosis	Occipital CERAD neuritic plaque score	Occipital Olichney CAA score
1	73	F	ε3/ε4	6	Mild to moderate, multifocal	1	1
2	68	F	ε4/ε4	4	Moderate	2	1
3	66	F	ε3/ε4	5	Moderate, diffuse	2	4
4	75	M	ε3/ε4	4	Mild to moderate, multifocal	1	0
5	73	F	ε3/ε4	5	Moderate, widespread	1	2
6	70	F	ε4/ε4	6	Mild, focal	2	0
7	79	M	ε3/ε3	3	Mild to moderate, widespread	1	3
8	73	M	ε3/ε4	5.5	Mild to moderate, multifocal	2	0
9	73	M	ε3/ε4	5.5	Mild to moderate, multifocal	1	0
10	82	M	ε4/ε4	3	Moderate to severe, multifocal	1	3
11	68	F	ε4/ε4	6	Mild, multifocal	2	1
12	65	F	ε3/ε4	5.5	Mild, focal	2	1
13	88	M	ε3/ε3	5.5	Mild, multifocal	2	2
14	74	M	ε3/ε3	5	Severe, multifocal	1	4
15	79	M	ε4/ε4	5	Moderate, widespread	2	4
16	80	F	ε3/ε4	3	Mild to moderate, multifocal	1	2
17	77	F	ε4/ε4	5	Mild to moderate	1	4

Immunofluorescence pathology study
Case	Age (y)	Sex	APOE genotype	Braak stage	Overall pathologic CAA diagnosis	Occipital CAA pathologic description
A	55	M	NA	6	NA	Moderate
B	84	F	NA	5	Focal severe	Moderate to severe
C	88	F	NA	4	Moderate to severe	Moderate to severe
D	96	F	NA	5	NA	Moderate
E	69	M	NA	6	Severe	Severe

Abbreviations: APOE, apolipoprotein E; CAA, cerebral amyloid angiopathy; CERAD, Consortium to Establish a Registry for Alzheimer's Disease.

2.2. Measurement of equilibrium binding affinity

A full description of the biochemical methods, rationale, and validation is available in the supporting information. Briefly, liquid extracts (aqueously soluble or insoluble, from parenchyma or from meninges) were diluted and immunoprecipitated with serially diluted anti‐amyloid antibody on magnetic beads. The beads were washed and then simultaneously eluted, monomerized, and denatured in 5 M guanidine hydrochloride (GuHCl) to expose C‐terminal epitopes. Monomer‐preferring Meso Scale Discovery enzyme‐linked immunosorbent assays (ELISAs) quantified Aβ₄₂ in parenchymal samples and Aβ₄₀ in meningeal samples. Selective detection of Aβ isoforms permits enrichment for amyloid plaque‐derived Aβ over microvascular CAA in the parenchyma, and CAA over adhered amyloid plaque in the meninges. The resulting titration curves were fitted to a one‐site binding model resulting in a K_D (approximation of binding affinity) and B _max (total Aβ available to the antibody to bind) in GraphPad Prism software (Figure 1). We validated that binding in our assay occurs at equilibrium (Figure S3 in supporting information); avoids unwanted antibody titration (Figure S4 in supporting information); and reproduces binding to synthetic, pure, Aβ (Figure S5 in supporting information). We confirmed that the recombinant aducanumab and donanemab produced at contract research organization Evitria used for our study demonstrated equivalent binding characteristics to brand‐name Aduhelm and Kisunla purchased from their drug manufacturers (Figure S6 in supporting information). Example binding curves from our full dataset are shown in Figure S7 in supporting information.

FIGURE 1.

Experimental overview. Occipital lobe gray matter and overlying meninges from 18 cases with AD and CAA were processed to extract aqueously soluble and insoluble fractions. These were immunoprecipitated with serially diluted anti‐amyloid antibodies followed by wash, elution, denaturation into monomers, and quantitation of total Aβ₄₂ (for parenchyma) or Aβ₄₀ (for meninges). Binding curves were fit to a one‐site–specific model and generated a K_D , expressed in nM antibody and approximating equilibrium binding affinity, and a B _max, expressed in ng/ml Aβ and reflecting the total amount of Aβ accessible to the antibody. Aβ, amyloid beta; AD, Alzheimer's disease; CAA, cerebral amyloid angiopathy; IP, immunoprecipitation.

2.3. Size exclusion chromatography

Parenchymal homogenates (1 mL) were separately chromatographed on a Sephacryl S500 HR 16/60 column (Cytiva) run on an Akta FPLC (Amersham), in TBS running buffer at 0.5 mL/minute at 4°C, collecting 2‐ml fractions. Detection of Aβ proceeded using sandwich immunoassays on the SMCxPRO platform. Streptavidin‐coated high‐capacity microplates (Pierce) were coated with 50 µL of 6 µg/mL solution of biotinylated capture antibodies (lecanemab, donanemab, or aducanumab, recombinantly produced at Evitria) diluted in oAβ Assay Buffer (EMD Millipore). The coated plates were incubated undisturbed overnight at 4°C and washed 3x in TBS‐T (Pierce) immediately prior to sample loading. Fifty µL of each size exclusion chromatography (SEC) fraction were added to wells in duplicate and incubated for 30 minutes shaken at 600 rpm. The plates were washed 3x with TBS‐T, and 50 µL of Alexa‐647 conjugated 3D6 (Evitria) at 100 ng/mL in oAβ Assay Buffer was added to each well and shaken for 1 hour in the dark. Wells were then washed 3x with TBS‐T and eluted with 25 µL of 0.1 M glycine, pH 2.7, supplemented with 0.01% Triton X‐100 at 600 rpm on a plate shaker for 15 minutes. The eluent solution was then transferred to a new 96‐well plate (Axygen V‐bottom), in which each well contained 25 µL of neutralization buffer (EMD Millipore). The 50 µL of neutralized mixture was transferred a 384‐well plate (Aurora ABB2‐00160A) and the fluorescent readout was measured using an SMCxPRO plate reader.

2.4. Statistics

For our primary statistical analyses, we used the ratios of meningeal:parenchymal or insoluble:aqueous K_D s and B _maxs as dependent variables. Because different antibody preparations can have different degrees of degradation or inactivity, normalizing by within‐antibody ratios avoids this artifact. Expressing the K_D as a ratio can be interpreted as an antibody's preference for one type of Aβ over another (e.g., soluble vs. insoluble). We treated the unadjusted K_D and B _max without ratios as secondary outcomes. We refer to the meningeal:parenchymal ratios as “MP ratio” and the insoluble:soluble ratio as “IS ratio.”

All ratios were log‐transformed for statistical analyses. We used a linear mixed model of the log(ratio) as the dependent variable; the random effect of patient (case); and the fixed effects of age, sex, antibody, and number of APOE ε4 alleles. In all cases, the mixed models were fit by restricted maximum likelihood. T tests used the Satterthwaite method setting alpha to 0.05. Pairwise contrasts and 95% confidence intervals (CIs) were determined from the estimated marginal means. All mixed models were analyzed using R package lme4. Pearson correlations were analyzed using GraphPad Prism.

2.5. Tissue sectioning and IF

Brain blocks were sectioned at 12 µm (Leica RM2125 RTS Rotary Microtome). Tissue was deparaffinized with Histo‐Clear (National Diagnostics) and rehydrated in graded ethanol solutions. After washing with phosphate‐buffered saline (PBS), slides were treated with 88% formic acid for 10 minutes, washed again, and treated with sodium citrate solution at 60°C for 45 minutes. Slides were washed then blocked in 5% normal goat serum (NGS) in PBS for 1 hour at room temperature (RT). Slides were then incubated with serial dilutions of lecanemab (Evitria), aducanumab (Aduhelm, Biogen), or donanemab (Kisunla, Eli Lilly) in 3% NGS in PBS overnight at 4°C. After washing, fluorescently labeled secondary antibody (goat anti‐human immunoglobulin G [IgG] AlexaFluor‐647, Southern Biotech) was applied at 500x dilution for 2 hours at RT. Slides were washed and mounted with ProLong Gold Antifade Mountant with DAPI (Invitrogen). Negative controls omitted the primary antibody. Images were captured using a Zeiss Axioscan 7 Microscope Slide Scanner at 20x magnification.

Image analysis was done using ImageJ v1.54p. A rectangular region of interest (ROI; 1620 µm x 1247 µm) was analyzed in six serial sections stained with dilutions of lecanemab, aducanumab, or donanemab. A seventh serial section for each antibody was the no‐primary control, which was used to estimate the overall background intensity. The ROI position was approximated to include the same tissue in all sections. Within the rectangular ROI, a smaller CAA sub‐ROI was drawn manually around the vasculature, and the staining intensity was calculated using the measure tool in ImageJ (Figure S8 in supporting information). Subsequently, the mean background intensity for the CAA ROI was measured. Next, the intensity was measured for the tissue not containing CAA (i.e., the rest of the parenchyma and amyloid plaques). The reported intensities, here denoted as ΔΔ Intensity, were calculated in two steps: The Δ intensity was the total intensity within the sub‐ROI (CAA or plaque/parenchyma) minus the background intensity (in an unstained region) in the exact same ROI. The ΔΔ intensity was then calculated by subtracting the Δ Intensity of the no‐primary‐control in the matching ROI (representing tissue autofluorescence or non‐specific background) from the Δ intensity of the antibody of interest. A summary of the calculations is given below:

$$\begin{array}{ccc}\hfill \mathrm{\Delta }{\text{Intensity}}_{\text{ROI lecanemab}}& =& {\text{Intensity}}_{\text{ROI lecanemab}}\hfill \\ & & -\left(\text{Mean Background}{\text{Intensity}}_{\text{ROI}}\times {\text{Area}}_{\text{ROI}}\right)\hfill \\ \hfill \mathrm{\Delta }\mathrm{\Delta }{\text{Intensity}}_{\text{ROI lecanemab}}& =& \mathrm{\Delta }{\text{Intensity}}_{\text{ROI lecanemab}}\hfill \\ & & -\mathrm{\Delta }{\text{Intensity}}_{\text{ROI no--primary}}\hfill \end{array}$$

3. RESULTS

3.1. Lecanemab does not exhibit greater preference for aqueously soluble (protofibrillar) Aβ compared to aducanumab

Aβ protofibrils extracted from the AD brain have not been defined structurally, but rather arbitrarily as those which remain in the supernatant after centrifugation. ¹⁹ , ²⁰ , ²¹ , ²² According to this definition, sufficient centrifugal force renders all detectable protofibrils insoluble. ¹¹ That said, if lecanemab had greater preference for more‐soluble Aβ aggregates (protofibrils) than less‐soluble aggregates compared to other antibodies, then one would expect a higher ratio of its K_D for pellet‐to‐supernatant (insoluble‐to‐soluble) aggregates (IS K_D ratio) compared to other antibodies, in particular for parenchymal samples, in which all the intended antibody targets lie. However, we detected no statistically significant difference between the lecanemab and aducanumab parenchymal IS K_D ratios (Figure 2A). The 95% CI for the difference between lecanemab and aducanumab log(IS K_D ratios) was 0.245 to 0.163, meaning our results are 95% confident that lecanemab has between a 1.75‐fold lesser and 1.46‐fold greater preference for aqueously extracted Aβ (i.e., protofibrils) than aducanumab (Figure 2B). Thus, we conclude that lecanemab is unlikely to have greater preference for aqueously soluble Aβ aggregates than does aducanumab. However, our results did suggest that lecanemab and aducanumab have greater preference for aqueously extracted Aβ compared to donanemab (lecanemab 95% CI 0.029–0.437, 1.07‐fold to 2.74‐fold; aducanumab 95% CI 0.070–0.477, 1.17–3.00‐fold; Figure 2B). Full model results are presented in Table S1 in supporting information. We found that B _max did not correlate with K_D for any of the antibodies, meaning that total antigen concentration did not influence apparent affinity, as would be expected for a valid measurement of equilibrium binding (Figure S9 in supporting information).

FIGURE 2.

Binding profiles to insoluble versus soluble parenchymal Aβ₄₂ aggregates. A, The log(IS K_D ratio) reflects the binding preference of antibodies to insoluble versus soluble aggregates. A higher ratio implies greater preference for soluble aggregates. Donanemab exhibited a log ratio < 0, reflecting a slight preference for insoluble aggregates. Error bars = mean ± standard deviation. B, Model estimates for the pairwise mean differences ± 95% confidence interval in log(IS K_D ratio). There was no statistically significant difference between lecanemab and aducanumab, but donanemab had a statistically different ratio compared to lecanemab and aducanumab. C–F, Correlations between B _max, the total Aβ accessible to the antibody, across soluble and insoluble extracts reveal near‐perfect correlations. Aβ, amyloid beta.

3.2. Lecanemab and aducanumab access the same pool of Aβ aggregates

The B _max, the amount of Aβ immunoprecipitated at saturating antibody concentrations, is a measure of the total Aβ in a brain extract accessible to an antibody. If lecanemab bound a distinct pool of Aβ aggregates to which aducanumab or donanemab could not bind, then one might expect the B _max of lecanemab to exceed that of the others. Alternatively, if different AD patients’ brains contained different amounts of protofibrillar versus fibrillar Aβ, then there would be an imperfect correlation between B _maxs of different antibodies across different cases. However, we found a perfect 1:1 correlation of lecanemab and aducanumab B _max in both the aqueously soluble (r = 0.97, P = 3.3E‐10) and insoluble (r = 0.98, P = 3.9E‐11) fractions (Figure 2C,D). The donanemab B _max also correlated with the B _max of lecanemab for soluble and insoluble fractions albeit with slightly weaker correlations (soluble r = 0.94, P = 1.7E‐8; insoluble r = 0.86, P = 7E‐6, Figure 2E,F). Overall, we conclude that all three antibodies access essentially identical populations of aggregates. Donanemab may bind only a subset of pyroglutamate‐3 epitopes, but we reason that these are distributed evenly enough among the Aβ aggregates that donanemab can bind at least one site on all aggregates to which aducanumab and lecanemab can also bind.

Because there is no gold‐standard centrifugation protocol for protofibril isolation from human brain, protofibrils have sometimes been subdivided as those which elute close to the void volume of SEC columns. ⁸ Most SEC columns used in the literature, such as the Superdex 200 column or even Superose 6 column, ⁸ have relatively low size cutoffs compared to the molecular weight of Aβ aggregates, resulting in poor or no resolution for separating material in the putatively protofibrillar size range; or by using even lower resolution techniques like rate‐zonal ultracentrifugation. ²³ Thus, we used a Sephacryl S500 HR column (Cytiva), the commercially available column with the highest molecular weight cutoff (≈ 20,000 kDa), followed by high‐sensitivity ELISA using lecanemab, aducanumab, or donanemab as capture antibody each paired with detector antibody 3D6, which detects the Aβ N‐terminus. Although the raw ELISA signal differed between antibodies, all three antibodies recognized similar size distributions of Aβ particles even if the size of the particles differed among cases (Figure 3). This is consistent the conclusion that all three antibodies recognize the same population of Aβ aggregates in the human brain. If lecanemab recognized a smaller sized aggregate compared to aducanumab or donanemab, then one would expect a right‐shifted distribution of lecanemab ELISA signal compared to the other antibodies, which we did not observe.

FIGURE 3.

Size distribution of soluble Aβ aggregates recognized by anti‐amyloid antibodies. Parenchymal soluble extracts of three cases were fractionated on a Sephacryl S500 HR 16/60 column, which possesses a very high molecular weight cutoff (≈ 20,000 kDa) for separating large particles. Particles elute from the column in reverse order of size, such that the largest particles are present in the lowest‐numbered fractions. Fractions were then tested by single‐molecule counting sandwich enzyme‐linked immunosorbent assay on the SMCxPRO platform, using the clinical anti‐amyloid antibody as capture, with N‐terminally directed antibody 3D6 as detector. Raw response units on SMCxPRO presented in (A, C, E), then normalized to the maximum within‐antibody signal in (B, D, F). The size distribution of Aβ aggregates recognized by all three antibodies was similar. Aβ, amyloid beta.

3.3. Antibody binding preference does not explain ARIA‐E rates in clinical trials

In phase 3 trials, aducanumab caused ARIA‐E in 35.2% of subjects receiving the drug, ²⁴ followed by 24.0% by donanemab ²⁵ and 12.6% by lecanemab. ²⁶ If the differences in ARIA‐E rates were due to differences in binding preferences for CAA versus plaque Aβ, then one would expect the ratio of K_D for meningeal Aβ₄₀‐rich to parenchymal Aβ₄₂‐rich aggregates (MP K_D ratio) to differ among the three antibodies and follow the order lecanemab > donanemab > aducanumab. We detected no statistically significant differences among the MP K_D ratios of the three antibodies in the soluble pool (Figure 4A). The 95% CI for the difference in MP K_D ratio of lecanemab versus aducanumab in the soluble pool was 0.121 to 0.305 (Figure 4B). Thus, we conclude that in aqueous extracts, lecanemab has between a 1.32‐fold lesser and a 2.02‐fold greater preference for plaque‐enriched (parenchymal Aβ₄₂) versus CAA‐enriched (meningeal Aβ₄₀) aggregates compared to aducanumab, unlikely to explain fully the ≈ 2.8‐fold difference in ARIA‐E rates in phase 3 clinical trials. Full model results are presented in Table S2 in supporting information.

FIGURE 4.

Plaque versus CAA antibody binding preferences. A, C, The log(MP K_D ratio) reflects binding preferences of antibodies to meningeal Aβ₄₀‐rich aggregates (CAA‐enriched) versus parenchymal Aβ₄₂‐rich aggregates (plaque‐enriched). A higher ratio implies greater preference for plaque versus CAA aggregates. Error bars = mean ± standard deviation. B, D, Model estimates for the pairwise mean differences ± 95% confidence interval in log(MP K_D ratio) reveals no significant differences in the soluble fraction but significant differences in the insoluble fraction. The difference in insoluble log(MP K_D ratio) between lecanemab and aducanumab reflects a 1.01‐ to 2.07‐fold greater lecanemab preference for plaque compared to aducanumab, less than the ≈ 2.8‐fold difference in phase 3 clinical trials. Aβ, amyloid beta; CAA, cerebral amyloid angiopathy.

In insoluble extracts, there were statistically significant differences in MP K_D ratio, including all three pairwise comparisons (Figure 4C). However, the rank order was donanemab > lecanemab > aducanumab, as opposed to the expected order lecanemab > donanemab > aducanumab based on phase 3 trial ARIA results. Examining only lecanemab and aducanumab, which do exhibit the hypothesized order based on trial results (lecanemab > aducanumab), the 95% CI for the difference in MP K_D ratio was 0.0044 to 0.317 (Figure 4D), meaning lecanemab exhibited between a 1.01‐ and 2.07‐fold greater preference for plaque‐enriched (parenchymal Aβ₄₂) versus CAA‐enriched (meningeal Aβ₄₀) compared to aducanumab. Although this difference is statistically significant, it may not fully explain the observed nearly 3‐fold difference in ARIA rates, nor account for donanemab having still greater preference for plaque compared to lecanemab or aducanumab. Full model results are presented in Table S3 in supporting information.

3.4. Lecanemab and aducanumab do not differ in immunostaining CAA and amyloid plaques

Although highly quantitative, the immunoprecipitation ELISA method lacks spatial resolution and cannot distinguish morphologically between CAA versus amyloid plaque binding. In previous work, lecanemab has been found to stain CAA, but only at single concentrations and/or not directly compared in serial sections to aducanumab and donanemab. ²⁰ , ²⁷ , ²⁸ Thus, we pursued a modified serial dilution IF method. In a cohort of five cases for which occipital blocks were briefly fixed for 2 hours in formalin at autopsy (to preserve antibody epitopes) and embedded in paraffin, we quantified CAA and amyloid plaque labeling intensity by the three antibodies in serial sections (Figure 5). We found that all three antibodies labeled both amyloid plaques and CAA, consistent with prior reports (Figure 5A). With serial dilution, the signal intensities for lecanemab and aducanumab disappeared to the no‐antibody background levels at approximately the same dilution for both CAA and plaque (Figure 5B,C). Donanemab staining was apparent only at the highest concentration, but this staining was equivalent for CAA and plaque. These qualitative results are consistent with our quantitative affinity measurements, suggesting that CAA versus plaque preference of the three clinical antibodies does not explain their differences in ARIA‐E rates.

FIGURE 5.

Titration immunostaining of plaque and CAA. A, Representative images from patient D across all concentrations in the three antibodies. Lowest concentration on the left and highest concentration on the right. Scale bar 200 µm. B, Intensity of plaque staining in five cases at increasing concentration of lecanemab, aducanumab, and donanemab. Data presented as mean and standard deviation. C, Intensity of CAA staining at increasing concentrations of the three antibodies. Data presented as in (B). CAA, cerebral amyloid angiopathy.

3.5. Effect of APOE genotype on antibody binding

The APOE ε4 allele increases the risk of AD through multiple mechanisms, including an increase in amyloid plaques, CAA, and in plaque‐mediated tangle accumulation. ²⁹ , ³⁰ , ³¹ , ³² The APOE ε4 allele also increases ARIA risk in a dose‐dependent manner, and carriers may also benefit less from lecanemab and donanemab compared to non‐carriers. We explored the effect of APOE ε4 gene dosage in our antibody binding data to help explain these phenomena. Treating the number of APOE ε4 copies as a continuous variable (i.e., the gene dose), we found the APOE ε4 allele lowered the parenchymal IS B _max ratio; in other words, the APOE ε4 allele rendered the pool of Aβ accessible to each anti‐amyloid antibody more soluble (P = 0.01; Figure 6A, Table S4 in supporting information). Treating the APOE genotype as a categorical variable rendered a statistically significant effect of APOE ε4 homozygosity compared to other genotypes (P = 0.02). In meninges, the dose of APOE ε4 also had a significant effect on increasing the solubility of antibody‐accessible Aβ (P = 0.01). In categorical analyses, homozygotes possessed a lower IS B _max ratio than heterozygotes and non‐carriers (P = 0.04; Figure 6B). There was no statistically significant effect of APOE ε4 dosage on the IS K_D ratio in either parenchyma or in meninges; in other words, APOE ε4 did not affect the affinity with which antibodies bound to soluble versus insoluble Aβ, only their total availability to bind. We also detected no effect of APOE ε4 dosage on either the absolute K_D to meningeal extracts or those normalized to parenchymal extracts (the MP K_D ratio); in other words, APOE ε4 did not change Aβ conformation in such a way as to significantly alter antibody affinity. Full model results are presented in the supplementary tables in supporting information.

FIGURE 6.

APOE genotype effects on Aβ accessible to antibody binding. The log(IS B _max ratio) reflects the solubility of the Aβ pool accessible to the antibody. A higher log(IS B _max ratio) reflects a less soluble pool of Aβ. We found that the APOE ε4 dosage increased the solubility (decreased the log(IS B _max ratio)) of both parenchymal Aβ₄₂ (A) and meningeal Aβ₄₀ antibody targets. Error bars = mean ± standard deviation. Aβ, amyloid beta; APOE, apolipoprotein E; SD, standard deviation.

4. DISCUSSION

We find that (1) aducanumab, lecanemab, and donanemab bind substantially the same population of Aβ aggregates in the human brain; (2) differences in equilibrium binding affinity preferences for CAA versus amyloid plaque between the antibodies are either undetectable or insufficient to explain differences in ARIA‐E rates; and (3) the APOE ε4 allele renders the Aβ aggregates that these antibodies bind more soluble. However, this study has important limitations. First, we only measured equilibrium binding constants (K_D ) as an approximation of affinity. We were unable to measure association (k _on) and dissociation (k _off) constants. Antibodies with identical equilibrium binding affinities can have different kinetics, and thus we cannot exclude that association and dissociation rates could differ among antibodies for parenchymal, vascular, soluble, or insoluble Aβ aggregates. We are developing techniques to measure these kinetics in AD brain extracts.

Second, the extraction method from the parenchyma could not separate plaque Aβ aggregates from microvascular CAA aggregates. Thus, some contamination in the parenchymal preparation with CAA aggregates was likely. However, we could enrich for plaque aggregates through the selective detection of Aβ₄₂ over Aβ₄₀ in the parenchymal fraction. Further, the meningeal preparations were unlikely to contain any plaque aggregates because amyloid plaques do not occur in the meninges. Thus, at best, the calculated MP ratios are of binding affinity to CAA versus plaque aggregates, and at worst, a ratio of binding affinity to CAA versus total (CAA + plaque) aggregates. This also assumes that CAA aggregates in meningeal blood vessels are biochemically equivalent to those in the parenchymal microvasculature, which they may not be. A last disadvantage of our approach is the exclusion of Aβ₄₀‐rich aggregates from the parenchyma; certain plaques are Aβ₄₀ rich, ³³ and antibody binding to these would not have contributed to our measures. We can further not exclude that, especially in the parenchymal portion, competition for antibody binding by non‐Aβ₄₂–containing Aβ aggregates could alter the measured binding parameters. Principally analyzing within‐case ratios across insoluble/soluble or meningeal/parenchymal likely helped to mitigate this effect.

Third, we used antibodies produced recombinantly from patent sequences for affinity determinations. While at least aducanumab and donanemab displayed similar binding in a subset of extracts (Figure S4), we cannot exclude subtle differences in binding compared to the brand‐name product given to patients. Brand‐name aducanumab (Aduhelm) and donanemab (Kisunla), but not lecanemab (Leqembi), were used for IF.

Fourth, we cannot exclude the extraction method itself causing structural alterations in the antibody epitopes. We take particular caution in interpreting donanemab results, because this is a non‐conformational (linear) epitope. Homogenization may have sheared free pyroglutamate‐3 Aβ monomers off aggregates, leading to an excess of monomeric pyroglutamate‐3 Aβ as opposed to that which naturally occurs in the aggregates. This artificially sheared pyroglutamate‐3 Aβ could result in falsely low K_D and B _max values. Because aducanumab and lecanemab recognize conformational epitopes, shearing of monomers should not have caused this artifact, though we cannot exclude alterations in their epitope abundance during extraction.

Within these limitations, we conclude that (1) lecanemab does not bind a distinct population of Aβ aggregates inaccessible to aducanumab or donanemab, (2) that lecanemab exhibits equivalent binding affinity to soluble protofibrils (aqueously extracted Aβ aggregates) compared to aducanumab or donanemab, and (3) that the observation from phase 3 trials that lecanemab caused less ARIA‐E than donanemab, in turn less than aducanumab, cannot be explained by differences in binding preference to CAA. In previous work, we found that anti‐amyloid antibodies including lecanemab could bind to aqueously extracted short Aβ fibrils from the human brain, ¹¹ suggesting that much of what has previously been termed protofibrils and presumed to possess a different structure (and thus different antibody‐binding characteristics) compared to fibrils, are in fact structurally the same amyloid fibrils as those found in amyloid plaques. Our data from this study are consistent with this model: a putatively protofibril‐preferring antibody (lecanemab) exhibited the same binding preference to the aqueous (protofibrillar) extraction as did the putatively non‐protofibril–preferring antibody (aducanumab). Previous comparisons of lecanemab to other antibodies used in vitro synthetic preparations of Aβ, ² the atomic structures of which are unknown and may not exist in the AD brain. One previous measurement comparing lecanemab to other antibodies’ binding to human meningeal extracts did not include a ratio to parenchymal extracts, tested only N = 4 brains and excluded one of these as an outlier. ³⁴

Our results do not exclude antibody binding to CAA as a cause of ARIA‐E. Rather, we conclude only that differences between antibodies’ ARIA‐E rates and efficacy cannot be explained by differences in the equilibrium binding constants we measured. A small dose titration adjustment in the Trailblazer‐Alz 6 trial of donanemab appeared to reduce ARIA‐E rates, ³⁵ implying that at least some differences in ARIA‐E rates from other antibodies could be explained by dosing or monitoring factors rather than differences in antibody affinity. The phase 3 trials for aducanumab, donanemab, and lecanemab differed in titration regimens and in the number of monitoring MRIs obtained; the latter could lead to sampling bias. While all three antibodies are IgG1 subclass, differences in Fc receptor engagement, half‐life, off‐rates, or other antibody‐dependent but binding equilibrium affinity‐independent factors could additionally affect ARIA‐E and efficacy.

We found that the APOE ε4 allele was associated with an overall more soluble pool of Aβ that the antibodies could bind to. Future studies may relate how this alteration may affect therapeutic parameters such as clearance from the brain and neutralization of toxicity, and whether a more soluble pool of CAA or plaque Aβ results in an elevated risk of ARIA‐E.

Future efforts to improve anti‐amyloid therapy for AD will require a deeper mechanistic understanding of how antibody‐antigen interactions shape plaque clearance, clinical efficacy, and ARIA.

CONFLICT OF INTEREST STATEMENT

Dennis J. Selkoe is a director of Prothena Biosciences and ad hoc consultant to Roche and Eisai. Cynthia A. Lemere is a consultant or scientific advisory board member to Acumen Pharmaceuticals, ADvantage Therapeutics, Apellis Pharmaceuticals, Cyclotherapeutics, Eli Lilly, Merck, MindImmune, Novo Nordisk, Receptive bio, Switch Therapeutics, and Therini Bio. The other authors declare no conflicts of interest.

CONSENT STATEMENT

All subjects or their health‐care proxies whose brain tissue was used for this study provided informed consent for autopsy and research tissue use at Brigham and Women's Hospital.

Supporting information

Supplementary Methods

Figure S1. Identical crude parenchymal homogenates were subjected to centrifugation by the two protocols used in this study, followed by 5 M GuHCl denaturation and Aβ₄₂ ELISA, resulting in substantially similar concentrations in a pilot set of two cases during protocol optimization.

Figure S2. After 5 M GuHCl denaturation of soluble extracts the Aβ₄₂:Aβ₄₀ was found to be consistently higher for parenchyma than meninges, as expected in a pilot set of four cases during protocol optimization. Note the log scale on the Y‐axis.

Figure S3. Past 6 h, the binding curve for test antibody h1C22 equilibrated. Thus, we used a 24‐h immunoprecipitation reaction to measure equilibrium binding of anti‐amyloid antibodies. Each dot is one antibody at one concentration at one timepoint, with N = 3 technical replicates. Error bars represent standard deviation.

Figure S4. As a human AD brain extracted is diluted 5‐, 10‐, or 20‐fold, the B_max diminishes proportionately to the dilution factor but the K_D does not, implying a measurement of antibody affinity rather than antibody titration. Each dot is one concentration, with N = 3 technical replicates. Error bars represent standard deviation (left graph) or 95% CI of model fit (right graph).

Figure S5. The binding of test antibody m266, which binds the Aβ mid‐region and selectively binds monomers, exhibits the same binding profile with synthetic Aβ₄₀ monomers as those found naturally in a human AD extract. This antibody was chosen to test the validity of the method because the m266 epitope, being linear and present on monomers, is expected to have the same structure in synthetic as human brain‐derived extracts. Each dot is one antibody at once concentration, with N = 3 technical replicates. Error bars represent standard deviation.

Figure S6. Binding profile of recombinant aducanumab and donanemab equivalents used for this study were identical to those of their brand‐name products. Each dot is one antibody at once concentration, with N = 3 technical replicates. Error bars represent standard deviation.

Figure S7. Example binding curves used to fit the K_D and B_max values from the first four cases of our cohort. Each dot is one antibody at once concentration, with N = 3 technical replicates. Error bars represent standard deviation.

Figure S8. Identification and definition of ROICAA and ROIPlaque A: Example of how the ROIs were defined in case C for ROI_CAA and ROI_Plaque, respectively including how background intensity was measured for plaque (pq’) and CAA (caa’), respectively. Subsequently, each ROI was cleared of the corresponding compartment and analyzed separately as exemplified in B and C. All scale bars 200 µm.

Figure S9. Lack of correlation between K_D and B_max for any antibody or extract fraction as expected for a valid measurement of antibody affinity, independent of antigen concentration. Each dot is one case from the full N = 17 cohort.