Integrative Single-Plaque Analysis Reveals Signature Aβ and Lipid Profiles in the Alzheimer’s Brain

Abstract

Cerebral accumulation of amyloid-β (Aβ) initiates molecular and cellular cascades that lead to Alzheimer’s disease (AD). However, amyloid deposition does not invariably lead to dementia. Amyloid-positive but cognitively unaffected (AP-CU) individuals present widespread amyloid pathology, suggesting that molecular signatures more complex than the total amyloid burden are required to better differentiate AD from AP-CU cases. Motivated by the essential role of Aβ and the key lipid involvement in AD pathogenesis, we applied multimodal mass spectrometry imaging (MSI) and machine learning (ML) to investigate amyloid plaque heterogeneity, regarding Aβ and lipid composition, in AP-CU versus AD brain samples at the single-plaque level. Instead of focusing on a population mean, our analytical approach allowed the investigation of large populations of plaques at the single-plaque level. We found that different (sub)populations of amyloid plaques, differing in Aβ and lipid composition, coexist in the brain samples studied. The integration of MSI data with ML-based feature extraction further revealed that plaque-associated gangliosides GM2 and GM1, as well as Aβ_1–38, but not Aβ_1–42, are relevant differentiators between the investigated pathologies. The pinpointed differences may guide further fundamental research investigating the role of amyloid plaque heterogeneity in AD pathogenesis/progression and may provide molecular clues for further development of emerging immunotherapies to effectively target toxic amyloid assemblies in AD therapy. Our study exemplifies how an integrative analytical strategy facilitates the unraveling of complex biochemical phenomena, advancing our understanding of AD from an analytical perspective and offering potential avenues for the refinement of diagnostic tools.

Introduction

Alzheimer’s disease (AD) is a highly prevalent neurological disorder and a major unmet medical need. Without effective therapies, it is expected that ∼150 million people worldwide will be affected by 2050.¹ Longitudinal studies have shown that AD begins with the accumulation of misfolded Aβ peptides decades before the onset of cognitive symptoms.²⁻⁵ Aβ deposition initiates molecular and cellular cascades that lead to tau hyperphosphorylation and intracellular aggregation, synaptic dysfunction, neuroinflammation, neurodegeneration, and ultimately, cognitive impairment.^6,7 As AD progresses, Aβ peptides continue to aggregate extracellularly to form amyloid plaques, which is a major pathological feature of the AD brain.

A wealth of evidence supports the essential role of Aβ in AD etiology, but the nature of the ultimate neurotoxin remains unclear.⁸ Genetic, clinical, and biochemical evidence indicates that (relative) increases in longer, more aggregation-prone Aβs cause early onset familial AD (FAD), with pathogenic mutations promoting the generation of longer Aβ42/43 peptides at the expense of the shorter Aβ37/38/40 peptides.⁹ In addition, mutation-driven changes in the short-to-long Aβ peptide ratio inform about pathogenicity and clinical onset.¹⁰ In the absence of pathogenic mutations, however, cerebral amyloidosis does not always lead to dementia, as the prevalence of cerebral amyloidosis is ∼60%, while the prevalence of AD (incl. Mild Cognitive Impairment, MCI) is only ∼30% by the age of 85.¹¹ The relatively high prevalence of amyloid pathology even in older cognitively normal individuals^12,13 implies a more complex scenario in which Aβ, causal in early onset FAD, is essential but not sufficient to cause late onset, sporadic AD (SAD).

Amyloid-positive, cognitively unimpaired individuals (AP-CU) present widespread cerebral Aβ plaques and may also present tau pathology as well as other (comorbid) brain pathologies.¹³⁻¹⁷ In a longitudinal study, the cumulative incidence of dementia in cognitively unimpaired amyloid- and tau-positive individuals aged 74 years old was less than 50% after 14 years of follow-up, suggesting that the presence of two major hallmarks of AD (amyloid plaques and tau tangles) does not necessarily imply cognitive decline.¹⁸

Protective or resilience factors prevent, delay, or halt Aβ-induced neurotoxic cascades in AP-CU individuals (for a review, see ref (19)). However, does amyloid pathology differ between AD and AP-CU cases? Pathological studies have shown that the amount of β-pleated sheet (thioflavin S-positive) plaques and oligomeric Aβ, rather than total amyloid burden, and increased levels of activated astro- and microglia better discriminate AP-CU from dementia cases.¹⁶ Furthermore, biochemical analyses have shown similar levels of Aβ_1–42 but lower levels of Aβ_1–40 in the brains of AP-CU individuals, compared to AD cases,²⁰ while mass spectrometry (MS) analyses have revealed that phosphorylated Aβ and pyroglutamylated Aβ_x–42 are significantly higher in AD than in AP-CU brains.^21,22 These data support the notion that Aβ pathology differs among these pathologies. These analyses, however, are based on tissue-wide averages or manually selected regions of interest (ROI) and, therefore, lack spatial and molecular detail on potentially distinct plaque (sub)populations that may better differentiate AD vs AP-CU.

Genome-wide association studies support a probabilistic model of AD,¹⁹ wherein “disease penetrance” of Aβ is modulated by risk and/or protective factors linked to pathways related to Aβ and tau misfolding, inflammation, and cholesterol/lipid metabolism. Additional genetic data²³ and extensive lipid analytical data²⁴⁻²⁹ (for a review, see ref (30)) support a significant role of lipid metabolism in modulating Aβ-driven toxicity. We therefore hypothesized that integrative, spatially resolved single-plaque analysis of (Aβ) peptide and lipid content in brain tissue slices could reveal plaque (sub)populations differentially associated with one or the other condition. We further argued that gaining insight into plaque population statistics in AP-CU and AD brains will highlight potential factors that differentially modulate Aβ-driven toxicity. To test this hypothesis, we assessed Aβ and lipid composition of frontal cortical amyloid plaques from 8 AP-CU and 9 AD cases using MS imaging (MSI), followed by single-plaque analysis and plaque population statistics.³¹ Single-plaque analysis aims at studying large populations of individual plaques, instead of just referring to a population mean, which could hide potential subpopulation. We found that different subpopulations of amyloid plaques coexist to varying degrees in AD and AP-CU cases, and machine learning-based models surprisingly pointed to Aβ_1–38 and gangliosides GM1 and GM2 as the most relevant differentiators between these conditions.

Materials and Methods

Chemicals

Acetonitrile (ACN) and trifluoroacetic acid (TFA) were obtained from Merck (Darmstadt, Germany). Dimethyl sulfoxide (DMSO), ammonium formate, and 1,5-DAN (1,5-diamino-naphthalin) were from Sigma-Aldrich (Munich, Germany). Milli-Q water (ddH₂O; Millipore, Burlington, USA) was prepared in-house. sDHB, the MALDI-MS peptide calibration standard mix II, and the protein calibration standard mix I were obtained from Bruker Daltonics (Bremen, Germany). Synthetic Aβ peptides (rPeptide; Watkinsville, USA) were mixed at 2 μmol/L (equimolar) in ddH₂O (Aβ calibration standard).

Patient Samples and Tissue Sectioning

Brains were donated with the full, informed consent of the Queen Square Brain Bank for neurological disorders (UCL Queen Square Institute of Neurology). Accompanying clinical and demographic data of all cases used in this study were stored electronically in compliance with the 1998 data protection act (for an overview of the full available clinical data, see Table S1).

All cases underwent pathological diagnosis for AD, according to current consensus criteria, regardless of clinical symptoms.⁵⁶⁻⁵⁸ The cohort included pathologically diagnosed cases of AD and AP-CU with varying degrees of underlying pathology as documented by Braak stages, Thal Phase, and CERAD score (Table 1). The ApoE genotype was also determined. Frozen frontal cortices (Brodmann area 9) from all cases were mounted onto a cork block and sectioned on a Bright Cryostat (Bright Instruments Ltd., Huntingdon, UK) at 10 μm and onto ITO slides for MSI analysis.

Table 1. Clinical Data Related to Sporadic Alzheimer’s (SAD), AP-CU, and Amyloid-Negative Control Cases Analyzed in This Study^a.

sample group	AAO [years]	AAD [years]	gender	Braak phase	Thal phase	CERAD	ABC
AP-CU	NA	87	M	2	3	2	A2B1C2
AP-CU	NA	86	F	4	5	2	A3B2C2
AP-CU	NA	88	M	4	3	2	A2B2C2
AP-CU	NA	91	F	4	5	2	A3B2C2
AP-CU	NA	83	M	4	3	2	A2B2C2
AP-CU	NA	89	F	3	3	1	A2B2C1
AP-CU	NA	88	M	3	5	2	A3B2C2
AP-CU	NA	82	F	2	3	1	A2B1C1
CTRL	NA	79	M	2	0	0	A0B1C0
CTRL	NA	86	F	2	0	0	A0B1C0
CTRL	NA	101	M	1	0	0	A0B1C0
CTRL	NA	81	M	2	0	0	A0B1C0
SAD	74	87	F	6	5	3	A3B3C3
SAD	54	65	M	6	5	3	A3B3C3
SAD	72	88	M	6	5	2	A3B3C2
SAD	44	56	F	6	NA	3	A3B3C3
SAD	58	68	M	6	5	3	A3B3C3
SAD	60	68	M	6	5	3	A3B3C3
SAD	48	63	M	6	5	3	A3B3C3
SAD	65	81	F	5	NA	2	A3B3C2
SAD	43	58	F	6	5	2	A3B3C2

^aAAD: age-at-death. Braak phase: Braak stage assessing the progression of tau pathology in the brain (scale from 0 to 6⁵⁶). Thal phase: assessment of Aβ deposition throughout the brain (scale from 0-6,⁵⁷). The mean Thal phase of AP-CU cases was 3.75 and 5 for SAD cases, respectively. The Thal phase is based on a proposed hierarchical progression of Aβ deposition through neuroanatomical areas in the brain⁵⁷ and analogous to Braak staging of neurofibrillary tau tangles⁵⁶ (NA, not available). CERAD and ABC scores are also indicated.

MALDI MS Imaging

Tissue sections on ITO slides (from the −80 °C freezer) were thawed in a desiccator for 20 min. Slides were washed with three dips in 8 °C 50 nM ammonium formate (pH 6.4) and dried in a desiccator (1 h). An optical image was recorded using an Aperio CS2 digital slide scanner (Leica Biosystems GmbH), before 1,5-DAN was deposited on the tissue using an HTX M5 (HTX Technologies, Chapel Hill, USA) sprayer with the following settings: matrix solution: 18 mg/mL 1,5-DAN in 70% ACN, flow rate of 0.1 mL/min, speed of 2250 mm/min, nozzle height of 40 mm, nozzle temperature of 70 °C, plate temperature of 30 °C, track spacing of 3 mm, and layers of 15. MSI of lipids was done on a timsTOF FleX instrument (Bruker Daltonics) using the following parameters: mass range of 300–2100 m/z, negative polarity, laser intensity of 25%, number of shots of 300 at a laser frequency of 2 kHz, laser spot size of 20 × 20 μm, raster size of 20 × 20 μm, and tims-mode default. Instrument calibration was performed in ESI mode with an ESI tuning mix (Agilent Technologies, Santa Clara, USA). The obtained mass resolving power was 40,000 at 400 m/z, and the mass accuracy was better than 5 ppm.

MSI of peptides was performed as previously described.³¹ Briefly, after lipid MSI, the matrix was removed by submerging the slide into 100% ethanol for 60 s at RT. Tissue sections were delipidated and the sDHB matrix applied before being measured on a Bruker Rapiflex MALDI-TOF MS (Bruker Daltonics) in positive linear mode in the range between 2000 and 10,000 m/z with a spot size of 20 × 20 μm, as previously described.³¹

Since the removal of the slide from the instrument was unavoidable to remove the lipids and enhance peptide detection, we chose to use the timsTOF Flex instrument to measure lipids as it offers superior mass resolving power (R = ∼40,000) over the Rapiflex (R = ∼10,000 in reflector mode). Since the timsTOF Flex instrument is not capable of linear-mode measurements, which is a necessity to reliably detect amyloid peptides, the Rapiflex instrument was used.

MALDI MS Imaging Data Processing

Raw data were loaded into SCiLS Lab (Bruker Daltonics) and TIC-normalized, and the top 500 peaks were selected before the centroided data were converted to the open source imzML standard.⁵⁹ Data were loaded into M²aia software for registration.⁶⁰ Lipid and peptide MSI data were acquired using the same slide and the same measurement regions, and the overall alignment was off by only 1–3 pixel positions in the x and y directions. We note that although the same slide and measurement region were used, the registration between the motor coordinates and the optical image had to be redone after the lipid measurement, since the slide was unloaded from the instrument for lipid washes. Using the tissue outline as reference, small deviations were corrected manually (see Figure S1 for an evaluation of the alignment). Data were loaded into R 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria), and the ion images corresponding to Aβ_1–38, Aβ_4–42, Aβ_1–40, and Aβ_1–42 (corresponding to m/z 4129, 4202, 4335, and 4515, respectively, with a tolerance of ±6 Da) were extracted. Using the PLAQUEPICKER workflow,³¹ pixels corresponding to plaques were extracted and unique IDs were assigned to connected pixels now referred to as plaques. Next, plaques coordinates, from peptide MSI data, were used to extract the intensities of lipid species from the corresponding lipid MSI data sets. All ion images colocalizing with plaques were extracted: m/z 1179.72 (GM3(36:1)), 1207.77 (GM3(38:1)), 1382.80 (GM2(36:1)), 1410.83 (GM2(38:1)), 1544.88 (GM1(36:1)), and 1572.89 (GM1(38:1)). At this point, each unique plaque had intensity information for all selected peptide and lipid species. Intensities were square root-transformed. Mean intensities of all m/z features per plaque were calculated (), and Z-scores were derived per m/z feature and measurement run (note that each measurement run comprised one sample each from the tested groups; see Figure S2) to minimize batch effects;⁶¹ see also Figure S3 for a complete overview of the data analysis pipeline.

For modeling, the “tidymodels”-framework in R was used. Data were split into a test set (25% of all observations/plaques) and training set (75%) using stratification on the measurement run, again to control for possible batch effects.⁶¹ The training data were up-sampled by the ROSE algorithm⁶² to control for group imbalances (different numbers of plaques) between the investigated groups. From the training set, a 10-fold cross-validation set was generated, and hyperparameters were tuned for the different models using this set. Once optimal values for hyperparameters of the elastic net model (using the glmnet implementation⁶³), xgBoost (extreme gradient boosted trees, using the implementation from the R-package xgBoost), and MLP (multilayer perceptron, using the nnet implementation⁶⁴) were determined, each of the models was trained using the training set, and its performance was evaluated using the test set and test set. SHAP (SHapley Additive exPlanations) values⁶⁵ to explain the relative importance of variables were computed using the DALEX framework. For a schematic overview of the modeling workflow, see Figure S4.

Results

Differential Accumulation of Aβ Peptides and Gangliosides in Amyloid Plaques from AD and AP-CU Brain Samples

We assessed a cohort of 8 AP-CU, 9 AD, and 4 amyloid-negative and cognitively intact (control) brain samples (Table 1). The post-mortem delay (PM delay) values were 52 ± 29 h for AP-CU and 60 ± 20 h for SAD, resulting in an insignificant difference (two-sided t test, p-value = 0.56) between the groups. Additionally, both groups contained samples with cerebral amyloid angiopathy (CAA) so that the mean CAA level between both groups was equal. To investigate the relationship between peptide and lipid content of amyloid plaques in amyloid-positive cases, we performed two sequential MSI runs (first lipids and then peptides) on the same fresh-frozen post-mortem human brain tissue sections (frontal cortices and Brodmann area 9). Plaques were defined by ion images of Aβ_1–38, Aβ_1–40, and Aβ_1–42,³¹ and lipid data were then integrated to define plaque peptide-lipid composition on a single-plaque level for a total of more than 3000 plaques. Statistical analysis of these revealed two types of amyloid plaques, differing in Aβ peptide composition (Figure 1AD), that were differentially present in AP-CU versus AD samples. Relatively small (type 1) plaques (apparent size, ∼400–2000 μm²) were mostly composed of Aβ_x–42 (arrow heads in Figure 1F,H), presenting roughly 25, 20, and 20% of Aβ_1–42, Aβ_4–42, and pGlu Aβ_3–42, respectively. In addition, larger (type 2) plaques (apparent size, >2000 μm²) were mainly composed of shorter (∼35%) Aβ_1–40 and (20%) Aβ_1–38 species (arrows in Figure 1F). Relative Aβ percentages were estimated based on the respective peptide intensity relative to all detected Aβ species. These findings are consistent with previous observations showing small Aβ_1–42-rich plaques and larger Aβ_1–40-rich plaques in the AD brain.³² No Aβ species were detected in the control cases (Figure S5).

Figure 1.

Ganglioside isoforms are enriched in a subset of plaques from sporadic Alzheimer’s disease (SAD) patients (A, B, E, F) but less so in plaques from AP-CU cases (C, D, G, H). MALDI-MS ion images of m/z 4129 ± 6 (A, C) and m/z 4515 ± 6 (B, D) corresponding to Aβ_1–38 and Aβ_1–42 peptides, respectively. MALDI-MS ion images of m/z 1544.86 ± 0.1 corresponding to GM1(36:1) (E, G). Merged views (F, H). Arrows indicate amyloid plaques containing Aβ_1–38 and other short Aβ species. Arrowheads indicate small Aβ_1–42- and Aβ_4–42-rich plaques. (I) Average spectrum of plaque pixels corresponding to peptides marked by a white rectangle in panels (F) and (H) in red and cyan for SAD and AP-CU cases, respectively. (J) Average spectrum of ganglioside species of plaque pixels marked by the same white rectangle. (K) Mean apparent plaque size per sample; n(SAD) = 8, n(AP-CU) = 9, t test, *p < 0.05. (L) Distribution of total Aβ and total GM plaque intensities (kernel density estimate). Total Aβ (top panel) is comparable in both sample types, whereas total GM (bottom panel) displays high intensities in SAD but low intensities in AP-CU (red and blue, respectively). The dashed line marks the mean per measurement run. Amyloid plaques were considered “low” or “high” in total GM below or above this threshold. (M) Distribution of kernel density estimates for all detected Aβs vs total GM intensity per plaque (high and low total GM contents shown in upper and low panels, respectively).

Note that the here reported (apparent) sizes of plaques are likely an overestimation, but our intention was not to study plaque sizes as absolute but as a relative comparison. Previous publications report the average plaque size to be 400–450 μm² following a log-normal distribution³³ so that many plaques are even smaller than 400 μm² (the used pixel size in our study). As smaller than 400 μm² plaques will either be detected with a size of a single pixel (400 μm²) or will not be detected at all, because of lack of sensitivity, the here reported plaque size distribution overestimates plaque sizes. Both sample types (SAD and AP-CU) were measured with the same method, so a comparison of the apparent size is possible.

Rather than the composition or size, it was the relative abundance of type 1 vs type 2 plaques that differed the most between conditions: plaques in AP-CU samples were mostly type 1 (arrow heads in Figure 1H), while type 2 plaques were sparsely present in AP-CU (Figure 1C) but commonly observed in AD samples (arrows, Figure 1F).

We next used the MS signals corresponding to Aβ peptides to define regions of interest (ROIs) for each plaque and extracted colocalizing features from both lipid and peptide MSI data (see Figure S6 for the MSI scheme). The analysis of the plaque lipid content revealed marked accumulations of some ganglioside isoforms and amyloid. In particular, GM1 isoforms displayed strong colocalization with type 2 plaques that were enriched in shorter Aβs and preferentially populated the AD brain (Figure 1E,F). Remarkably, no ganglioside accumulation was detected in Aβ_1–42-rich type 1 plaques, neither in AD nor in AP-CU cases (Figure 1F,H, respectively). To verify the identity of the detected ganglioside species, we performed on-tissue fragmentation analysis and verified all reported ganglioside species by tandem-MS (Table S2 and Figure S7). No other lipid species other than gangliosides showed strong colocalization with amyloid peptides in plaques (Figure 1I,J).

As ganglioside GM1 isoforms strongly colocalized with heterogeneous type 2 plaques, which mostly populate AD samples, we assessed if Aβ and GM levels (using ion intensities as a surrogate) were correlated at the single-plaque level: We calculated the sum of normalized intensities (Z-scores) for total Aβ and ganglioside species per plaque and used these values to group plaques according to total Aβ or GM levels (Figure 1L). Total Aβ intensities did not differ significantly between the investigated conditions, although a subpopulation of plaques displayed higher Aβ levels in AD samples (seen as a high-intensity shoulder in Figure 1L, upper panel). However, the total GM intensity per plaque substantially differed between the AD (red) and AP-CU (blue) cases, with high GM content plaques being more prevalent in AD than in AP-CU samples. The opposite was true for plaques presenting low GM intensities (Figure 1L, lower panel). Single-plaque analysis also revealed that although there was a trend to lower Aβ_1–38, Aβ_1–40, and Aβ_342pE levels in AP-CU, Aβ peptide levels were largely independent of the GM status (Figure 1M and Table S3). Pairwise correlation analysis between all considered molecules indicated no correlation between the levels of analyzed Aβ species and GM isoforms (Figure S8).

In contrast, associations between Aβs were clear, with the strongest correlation observed between Aβ_1–42 and Aβ_4–42 (values of 0.68 and 0.67 for AD and AP-CU, respectively). Notably, despite their product-precursor link, the correlation between Aβ_1–38 and Aβ_1–42 was weak (−0.26 and −0.16 for AD and AP-CU, respectively) and contrasted with the stronger correlation observed for Aβ_1–38 and Aβ_1–40 (0.66 in AD and 0.39 in AP-CU). These findings suggest similar dynamics in the coaggregation of shorter Aβ_1–38 and Aβ_1–40 peptides in ganglioside-rich amyloid plaques and a disconnection between total (aggregated) Aβ and ganglioside levels in amyloid plaques at least at this end stage of the disease.

Ganglioside/Aβ₃₈/Aβ₄₀-Rich Plaques Differentially Populate the AD and AP-CU Brains

As autocorrelation is a known phenomenon in MSI,³⁴ plaques could not be treated as completely independent samples; instead, the mean tissue-wide composition of all plaques was considered in our analysis. The analysis showed that only Aβ_1–38 and GM2(36:1) levels were marginally significantly different between AD and AP-CU samples (BenjaminiHochberg-adjusted p < 0.05, n(SAD) = 9, n(AP-CU) = 8) (Figure 2A). We therefore investigated whether combinations of multiple attributes could define statistical subpopulations of plaques. To examine this, we split the total pool of plaques (using unsupervised k-means clustering, Figure S9) and performed multivariate statistical data analysis using an extended version of the previously reported plaquepicker software.³¹ This analysis uncovered two clusters of mixed sample origin, implying that plaque types differing in their combined Aβ and GM profiles were present in the pool (Figure 2B). Cluster 1 was enriched 1.6-fold in AD plaques compared to AP-CU plaques and characterized by high ganglioside levels. Cluster 2 was enriched 3.3-fold in plaques from AP-CU cases, relative to AD, and was characterized by lower ganglioside content and lower Aβ_1–38 and higher Aβ_x–42 (Aβ_1–42, Aβ_3–42pE, and Aβ_4–42) levels (Figure 2B). Next, all individual plaques, regardless of sample type (origin), were assigned to either cluster 1 or 2, followed by calculation of mean Z-scores for Aβ and GM isoforms in each cluster (Figure 2C). Aβ peptides showed smaller differences (Aβ_1–38 being the most significant) than GM1–3 isoforms; the latter clearly differentiated clusters. These data support the notion that distinct subpopulations of plaques differing in GM and Aβ levels differentially populate AP-CU vs AD brain samples.

Figure 2.

Single-plaque analysis reveals accumulation of gangliosides in a subgroup of plaques from sporadic Alzheimer’s disease (SAD) but not AP-CU cases. Statistical analysis of Aβ peptide and ganglioside (GM) composition using plaquepicker software.³¹ (A) On a global level mean (tissue-wide) composition of plaques per sample in SAD and AP-CU cases, only Aβ_1–38 and GM2(36:1) were marginally elevated in SAD vs AP-CU. (B) Single-plaque composition: Plaques were grouped into two clusters, independent of sample type, using k-means clustering. (C) Mean composition of plaques from clusters 1 and 2. In cluster 1 plaques, Aβ_1–38 and all ganglioside species were significantly elevated (p < 0.001). In cluster 2, Aβ_1–42 and Aβ_3–42pE (p < 0.01) as well as Aβ₄₄₂ (p < 0.001) were significantly higher. Unpaired t test: *p < 0.05, **p < 0.01, ***p < 0.001. All p-values were adjusted using the BenjaminiHochberg method, n(AP-CU) = 8, n(SAD) = 9.

Aβ_1–38 and GM1/2(36:1) Identified as Most Critical Differentiators of AD, Relative to AP-CU, by Machine Learning

The differential presence of GMs and Aβs in plaque subpopulations prompted us to assess their relative importance in differentiating the conditions studied. To investigate what variables (Aβ and/or GM species) are the most important differentiators between AD and AP-CU cases, we modeled the single-plaque composition for more than 3000 plaques with respect to plaque origin. We reasoned that this modeling would allow more flexibility in dealing with overlapping molecular composition and would allow us to test for multivariate interactions. To make any identified “feature importance” model independent, we trained three models using different machine learning (ML) architectures: an elastic net model (penalized logistic regression), a decision tree-based model (xgBoost), and a multilayer perceptron (MLP) as a simple neural network model (for more details, see Materials and Methods). All three models performed well, as indicated by the area under the receiver operating characteristic curve and accuracy values (Figure S10A,B). We therefore used them all to assess the “importance” of single-plaque molecular features in discriminating plaques from one condition or the other. For each ML model, we first determined the rank of the features and then the average rank across all three models (Figure 3AC and Figure S10C). Notably, all models (i) agreed on Aβ_1–38 as the most important Aβ peptide differentiating plaques, with higher levels of this (aggregated) peptide positively associated with AD, (ii) showed Aβ_1–42 as a feature of low importance, and (iii) consistently ranked the Aβ_3–42pE peptide in the top 5. Furthermore, all models (iv) pointed at (elevated) GM1(36:1) and GM2(36:1) as key features of plaques found in AD brains and intriguingly showed that GM1(38:1) and GM1(36:1) have opposite relationships, with high GM1(36:1) relative to GM1(38:1) levels associated with AD plaques (Figure S11). The influence of other GMs in discriminating between AP-CU and AD cases was marginal (Figure S10C).

Figure 3.

Machine learning (ML) reveals Aβ_1–38 and GM1(36:1) as the best differentiators between SAD and AP-CU conditions.(AC) SHAP values indicating the importance and direction of all variables for elastic net (A), multilayer perceptron (MLP) (B), and extreme gradient boosting (xgBoost) model (C). Each point represents a single observation, with the color indicating the respective feature value (Z-score of measured intensity), whereas the x position indicates the associated SHAP value. Positive SHAP values favor predictions for the SAD class, where negative SHAP values indicate associations with the AP-CU class.

Discussion

Given the essential pathogenic role of Aβ and the key involvement of lipid metabolism in AD pathogenesis, we hypothesized that an integrative, single-plaque analysis of (Aβ) peptide and lipid content in AD vs AP-CU brain samples could reveal amyloid plaque (sub)populations differentially associated with one or the other disease state. We applied spatially resolved MSI and a single-plaque analysis to assess the peptide and lipid profiles of amyloid plaques in AP-CU versus AD-affected cases. Our integrative analysis identified two distinct populations of plaques, differing in ganglioside and peptide compositions. In addition to the type 1 (Aβ_x–42-rich) plaques, we found larger plaques (type 2) enriched in the shorter Aβ_1–38 and Aβ_1–40 peptides and with marked increases in gangliosides GM1 and GM2. Type 1 plaques were found under both disease conditions. In contrast to the reported 20-fold increase in Aβ_1–40 levels in AD vs AP-CU,³⁵ our analysis found that Aβ_1–40 levels were not significantly different (1.5-fold higher but p > 0.05) between the investigated conditions.

In addition, our analysis revealed that although type1 and type 2 plaques coexist in both AD and AP-CU brains, the more complex and larger type 2 plaques preferentially populate the AD brain. Notably, analysis of amyloid pathology in knock-in AD mouse models at different ages also revealed two types of plaques: smaller diffuse plaques (composed mainly of Aβ_x–42 species) and larger, more complex plaques (presenting shorter Aβ peptides that surround a “nucleus” of Aβ_1–42).^31,36 A recent study combining stable isotope labeling kinetic techniques with MSI in the same mouse model demonstrated that Aβ_1–42 aggregates first and shorter Aβ species accumulate later.³⁶ Taken, together, these findings suggest that Aβ_1–42 deposition is an early event in the formation of amyloid pathology in AD and possibly in other brain amyloidosis such as AP-CU. The fact that aggregated Aβ_1–42 is however observed in both conditions suggests that its deposition in amyloid plaques does not per se lead to toxicity. Elevated Aβ_x–42 levels in the brain, relative to shorter Aβ peptides, are however associated with AD,³⁷ suggesting that metabolism of Aβ_1–42 (e.g., into truncated peptides) may be a relevant factor in amyloid-driven toxicity (see below). Whether the more complex plaques, enriched in shorter Aβ species and GMs, elicit cellular toxicity or glial activation and thus contribute to Aβ-driven toxicity early in AD pathogenesis merits further investigation.

We used three different ML architectures (Elastic net, MLP, and xgBoost) to model single-plaque data. The consensus between these models highlighted that aggregated Aβ_1–42 is not a good differentiator of plaques found in AP-CU and AD. In contrast, all ML models pointed to aggregated Aβ_1–38 as the most important discriminating factor between AP-CU and AD amyloid pathology. A recent clinical study showed that high CSF (soluble) Aβ₃₈ levels correlated with slower AD progression in two independent clinical cohorts.³⁸ We speculate that the lower aggregated Aβ_1–38 levels observed in the AP-CU cases may be connected with higher levels of the soluble peptide in the CSF.

In addition, our single-plaque analysis placed the Aβ_3–42pE peptide among the top 5 features differentiating amyloid pathology in AP-CU from AD and associated higher relative Aβ_3–42pE intensities with plaques in AD. This observation is consistent with previous MSI studies showing increased Aβ_3–42pE levels in the amyloid profile of AD, relative to the AP-CU condition,²² and with ex vivo immunostaining analysis showing significantly higher (intraneuronal) deposition of pE-Aβ_3–x in the AD brain compared to nondemented controls.³⁹ The fact that Aβ_3–42pE, but not the Aβ_1–42 peptide, is a distinguishing feature of AD-linked amyloid pathology may suggest that increased metabolism of the Aβ_1–42 peptide (or APP) into Aβ_3–42pE is a relevant factor in AD etiology.⁴⁰ It is worth noting that the Aβ_3–XpE peptide is one of the dominant Aβ species found in the hippocampi and cortices of AD patients⁴¹ and is the target of clinical trials testing donanemab, a plaque-specific humanized anti-Aβ_3–XpE antibody, in early symptomatic AD.⁴²

Notably, our analysis showed a marked increase in gangliosides (GM) in Aβ-heterogeneous plaques, which were found to differentiate AD-associated plaques relative to AP-CU. Levels of GM accumulation and Aβ deposition were not correlated, which might suggest a dissociation between GM accumulation and Aβ deposition. At the global tissue level, there was no significant difference in lipid composition between AP-CU and AD-associated plaques. However, at the single-plaque level, ganglioside differences were more pronounced and significant, with GM1/2(36:1) identified as the most differentiating features of plaques found in AD. In addition, our observations suggest that relative increments in the ratio between GM1(36:1) and GM1(38:1) are a key differentiating feature of AD plaques.

GM2 accumulation has been linked to microglia and astrocyte activation in a mouse model of acute ethanol-induced neurodegeneration;⁴³ thus, these differences might be connected with changes in the local (cellular) environment surrounding plaques. Whether local increases in GM levels modify amyloid behavior and toxicity is unclear, but previous reports have shown that Aβ and GM interact⁴⁴ and postulated that Aβ-GM interactions favor formation of β-sheet-rich structures and plaque seeding.⁴⁵ These effects may contribute to the higher levels of β-sheet-rich amyloid plaques reported in AD vs AP-CU¹⁶ and also observed in our brain samples (Figure S12). In addition, it has been shown that lowering the levels of major brain gangliosides significantly reduces Aβ deposition and improves cognitive performance in AD mouse models,⁴⁶ while inhibiting ganglioside biosynthesis in specific subsets of adult forebrain neurons improves cognition and counteracts dendritic spine loss in the 5XFAD mouse model,²⁸ even though total Aβ burden remains unchanged.^28,47 Indeed, several studies have suggested a pathological link between altered ganglioside levels and Aβ accumulation and/or aggregation.^{44−46,48−51} Changes in ganglioside may also have an effect on Aβ production, as gangliosides are present in lipid rafts, which modulate APP processing.^52,53

Finally, accumulation of GMs in the brain is also a pathological feature of some lysosomal lipid storage disorders, in which lipids accumulate due to loss of function of key enzymes involved in GM catabolism.⁵⁴ Although affected patients often do not survive beyond early childhood, evidence of accumulating Aβ has been found in post-mortem brain samples from Sandhoff’s disease (affecting degradation from GM2 to GM3) and TaySachs disease patients (accumulation of GM2).⁵⁵

In conclusion, our integrative single-plaque analysis provides, for the first time, population-based statistics of amyloid peptide and lipid pathology in post-mortem brain samples from demented versus nondemented amyloid-positive individuals. Although our study is limited by the analysis of amyloid pathology (coaggregated peptide-lipids) in post-mortem, end-point brain samples, it provides important novel insights into amyloid plaque heterogeneity in terms of Aβ peptide and lipid composition in the human brain. Most importantly, it demonstrates the coexistence of distinct plaque subpopulations in AD and AP-CU brain samples and identifies relevant discriminators between these conditions.

Our work opens the door for similar integrative single-plaque analysis to address amyloid heterogeneity at earlier AD stages, while the pinpointed differences may provide molecular clues for improving emerging immunotherapies to effectively target toxic amyloid assemblies in AD therapy.

Glossary

Abbreviations

Aβ

amyloid-β

AP-CU

amyloid-positive but cognitively unaffected

MSI

mass spectrometry imaging

MLP

multilayer perceptron

ROIs

regions of interest

SHAP

SHapley Additive exPlanations

xgBoost

extreme gradient boosted trees

Data Availability Statement

Raw data are available on PRIDE (PXD049325), and the integrated and extracted single-plaque data used for machine learning are available on FigShare (doi: 10.6084/m9.figshare.25186571).

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.analchem.3c05557.

• Supporting methods on immunohistochemistry and MALDI MS/MS and additional figures to demonstrate registration accuracy and to detail slide layout, data analysis, and modeling as well as multimodal workflow; MSI images of control tissue, MS/MS spectra of lipid species, correlation analysis between Aβ and lipid species, and thioflavin S and CR3/43 staining; supporting tables detailing clinical data, observed fragments in MS/MS, and mean Aβ composition of plaques in relation to GM status (PDF)

• ML workflow (PDF)

Author Contributions

L.C.-G. and C.H.: conceptualization of the study, supervision of the experimental research, data analysis, and writing of the manuscript. T.E.: conceptualization of the study, MSI experiments, analysis of data, and writing of the manuscript. T.L.: IHC experiments, analysis of data, and writing of the manuscript. D.A.S.: evaluation of the alignment accuracy between MSI (lipid-peptide) data sets and critical revision of the manuscript.

C.H. is grateful for the support by the Klaus-Tschira Foundation (project MALDISTAR) and acknowledges funding by the Federal Ministry of Education and Research (BMBF; FH-Impuls Partnerschaft M²Aind; Project: M²OGA; Förderkennzeichen 13FH8I02IA). This work was funded by the FWO G0B2519N and G008023N research grants to L.C.-G. and the Stichting Alzheimer Onderzoek (SAO). The Queen Square Brain Bank is supported by the Reta Lila Weston Institute of Neurological Studies, UCL Queen Square Institute of Neurology.

Brains were donated with the full, informed consent to the Queen Square Brain Bank for neurological disorders (UCL Queen Square Institute of Neurology). Accompanying clinical and demographic data of all cases used in this study were stored electronically in compliance with the 1998 data protection act (for an overview of the full available clinical data, see Table S1). Ethical approval for the study was obtained from the NHS research ethics committee and in accordance with the human tissue authority’s (HTA’s) code of practice and standards under license number 12198 and material transfer agreement UCLMTA12-2020.

The authors declare no competing financial interest.