Human-mouse proteomics reveals the shared pathways in Alzheimer's disease and delayed protein turnover in the amyloidome

Abstract

Murine models of Alzheimer's disease (AD) are crucial for elucidating disease mechanisms but have limitations in fully representing AD molecular complexities. We comprehensively profiled age-dependent brain proteome and phosphoproteome (n > 10,000 for both) across multiple mouse models of amyloidosis. We identified shared pathways by integrating with human metadata, and prioritized novel components by multi-omics analysis. Collectively, two commonly used models (5xFAD and APP-KI) replicate 30% of the human protein alterations; additional genetic incorporation of tau and splicing pathologies increases this similarity to 42%. We dissected the proteome-transcriptome inconsistency in AD and 5xFAD mouse brains, revealing that inconsistent proteins are enriched within amyloid plaque microenvironment (amyloidome). Determining the 5xFAD proteome turnover demonstrates that amyloid formation delays the degradation of amyloidome components, including Aβ-binding proteins and autophagy/lysosomal proteins. Our proteomic strategy defines shared AD pathways, identify potential new targets, and underscores that protein turnover contributes to proteome-transcriptome discrepancies during AD progression.

Extended Data Fig. 1 ∣. Workflow and quality control of mouse proteomic samples.

a, TMT-LC/LC-MS/MS and quality control workflow for analysis. Finally, 66 mouse samples were used in our proteome analysis. b, Box plot to show the distribution of normalized protein intensities in the mouse samples. c, Principal component analysis of mouse samples in different TMT batches. No batch bias was observed.

Extended Data Fig. 2 ∣. Proteomic analysis of age-linked alterations in WT mice.

a, Workflow of age-linked proteomics analysis. b, The cluster of upregulated proteins with age. Each line represents one protein. c, The cluster of downregulated proteins with age.

Extended Data Fig. 3 ∣. Phosphoproteomic and quality control analysis of 5xFAD, NLGF and WT mice.

a, Workflow for phosphoproteomics analysis. b, Distribution of measured phosphoprotein intensities in mouse samples. c, Principal component analysis of mouse samples highlighting the normalization of batch. d, Principal component analysis of mouse samples showing no obvious sex effect.

Extended Data Fig. 4 ∣. Whole proteome analysis of 3xTG, BiG and WT mice and the publication record of the AD-mouse shared proteins.

a, Workflow for whole proteome analysis by TMT-LC/LC-MS/MS. b, Distribution of normalized protein intensities in 3xTG and WT samples. c, Distribution of normalized protein intensities in BiG and WT samples. d, Principal component analysis of 3xTG and WT samples highlighting the difference between genotypes. e, Principal component analysis of BiG and WT samples highlighting the difference between genotypes. f, Publication record for the 275 AD-mouse shared proteins is represented by the number of publications and their integrative ranks determined by multi-omics analysis. The number of publications were found by PubMed search using “Alzheimer” and protein/gene names in August 2024. Using 20 publications as the cutoff, 14% of the proteins are well-studied, whereas 86% are under-researched. Notably, Pearson correlation analysis between the rank number and publication count yielded a negative value, suggesting that published research is disproportionately focused on a few well-known AD genes/proteins.

Extended Data Fig. 5 ∣. Plaque proteome analysis of AD mice.

a, Workflow for LCM-TMT-LC/LC-MS/MS. b, Distribution of normalized protein intensities in plaque and non-plaque samples. c, Principal component analysis highlighting the clear difference between the plaque and non-plaque results.

Extended Data Fig. 6 ∣. Comprehensive workflow for mouse pSILAC-TMT-LC/LC-MS/MS analysis and the importance of TMT noise correction.

a, Workflow including mouse pSILAC labeling, TMT-LC/LC-MS/MS experiments, protein identification and quantification via JUMP, batch merging, data curation, and JUMPt analysis for both apparent and corrected T₅₀ turnover rates. b, TMT channels for proteomics experiments designed to include three replicates each of WT and 5xFAD mice, with five-time points (0, 4, 8, 16, and 32 days) distributed across two batches. The fully SILAC-labeled mouse brain tissues from two mouse generations were highlighted in red, serving as a basis for noise correction. c, Schematic diagram illustrating the correction of TMT noise. In the light peptide decay curve, the fully labeled SILAC sample (red) facilitated noise detection since all Lys peptides were in the heavy form. Conversely, for the heavy peptide decay curve, the unlabeled sample (blue) from day 0 was utilized to determine the noise level. d, Comparison of protein half-lives in WT brain samples using the pSILAC and pSILAC-TMT methods, using 2,501 overlapping proteins. In the pSILAC method, each sample was analyzed separately using the traditional approach, effectively avoiding ratio compression in the dataset . Initially, before TMT noise correction, the global half-lives of proteins from the pSILAC-TMT experiment were significantly higher compared to those from traditional pSILAC analysis. After implementing TMT noise correction, the half-lives in the two datasets aligned closely, demonstrating that the TMT noise was effectively eliminated.

Extended Data Fig. 7 ∣. The principle of half-life analysis by JUMPt considering the recycling of Lys in mice.

a, Kinetic model for the JUMPt program: Heavy Lys is absorbed from food and transported into cells, contributing to a mixed pool containing previous light and newly absorbed heavy Lys. A portion of the Lys in cells is degraded, while the remainder is incorporated into proteins. When proteins degrade, the Lys in proteins is recycled back into the free Lys pool. b, The differences between apparent and corrected half-lives. c, Correlation between apparent and corrected half-lives. d, Examples of proteins with very short, intermediate, and very long half-lives: Proteins with very short half-lives have decay curves nearly overlapping with the free Lys curve, complicating accurate half-life calculation. These half-life values were set below 0.5 day. Conversely, proteins that decay very slowly have extremely long half-lives, presenting challenges for accurate computation. These half-life values were set above 100 days.

Extended Data Fig. 8 ∣. Model showing components plaque microenvironment and the changes in protein turnover in amyloid plaque microenvironments.

The core-shell model of plaque microenvironment is adapted from published spatial omics data.

Fig. 1 ∣. Brain proteomics reveals proteomic changes that are shared in AD mouse models.

a, Schematic plan of this study. Mouse cortical tissues from AD models of amyloidosis (5xFAD, NLF, NLGF, and matched WT, total n = 66 for 16 conditions, averaged n = ~4 per condition) were analyzed by TMT-LC/LC-MS/MS and compared with human metadata. b, Proteins quantified at different ages (3-18 months). c, Aβ levels quantified by MS. The values were averaged for each age and model, then normalized to 12-month-old 5xFAD (100%). d, DEPs between AD mice and WT controls, defined by moderated t-test with statistical cutoffs. e, Volcano plot for NLGF-WT comparison (FDR < 0.05, ∣log2FC∣ > 2SD, dashed lines). f, Heatmap of DEPs in AD mice, including the proteins enriched in 5xFAD or NLGF and those shared by both mice. g, Pathway analysis of shared DEPs in 5xFAD and NLGF. FDR was derived from p values (Fisher’s exact test) by the Benjamini-Hochberg procedure. h, Enriched PPI modules from biological processes using the shared DEPs.

Fig. 2 ∣. Tissue phosphoproteomics defines a new layer of regulation beyond whole proteome.

a, Phosphoproteome profiling was performed across 36 5xFAD and NLGF mice and their age-matched controls. We quantified 12,096 phosphopeptides (peptide FDR < 0.01) shared in all mice and performed statistical comparisons by moderated t-test. 122 DE phosphopeptides (80 proteins) were identified (FDR < 0.05, ∣log₂FC∣ > 2SD). b, Distribution of phospho-Ser/Thr/Tyr in identified phosphosites. c, Volcano plot of phosphoproteome data for 12-month-old 5xFAD compared to WT. Dashed lines indicate cutoffs. d, Volcano plot of phosphoproteome data for 12-month-old NLGF compared to WT. e, Heatmap of DE phosphoproteins in 5xFAD and NLGF mice, with protein subcellular location shown. f, PPI modules of DE phosphoproteins. g, The overlap of DEPs in phosphoproteomics and whole proteome analysis. Only consistent DEPs in both AD models were counted. h, Bioinformatics method for identifying altered kinase activities. Kinase-substrate linkages were extracted to infer kinase activities by the KSEA algorithm. An examples of MAPK activation based on KSEA. Phosphopeptide levels are displayed using the accompanying gradients. i, Heatmap of derived kinase activities. The fold change (FC) of 5xFAD and NLGF was calculated by comparison with WT.

Fig. 3 ∣. Comparison of mouse proteomics to human metadata identifies shared proteomic changes.

a, We identified 866 proteins that are consistently altered in more than 30 human AD proteomics studies, 654 of which were quantified in the proteomic analysis of the AD mice. Of these, 196 (30%) are differentially expressed in at least one mouse model (FDR < 0.05, ∣log₂FC∣ > 2SD). b, Number of overlapping DEPs between human AD and different mouse models. c, DEPs shared by human AD, 5xFAD, and NLGF mice. d, Scatter plot comparisons between Z scores of log₂fold change values (log₂FC-Z) of human AD/control cases and mouse models/WT at 12-month ages. Each dot represents one protein, and the color shows the dot density. Pearson correlation (R) values are shown. e, Heatmap showing log₂FC values of human-mouse shared AD proteins, classified by biological pathways. f, Workflow for deriving pathway activities. The FC of proteins in each pathway are integrated to calculate the pathway activity. g, Heatmap of pathway activities in AD and mouse models.

Fig. 4 ∣. Mouse models with additional pathologies beyond amyloidosis increase the similarity to AD.

a, Proteomic profiling of two more mouse models that express additional AD pathologies: WT (n = 8) and 3xTG (Aβ and tau pathologies, n = 19), as well as WT (n = 4) and BiG (Aβ and U1 splicing pathologies, n = 4). All mice were of ~6-month-old. The proteomic data were subjected to DE analysis and comparison with human AD data. b-c, Volcano plots of log2(fold change) and FDR in 3xTG and BiG mice, compared to WT, with DEPs highlighted in colors and cutoffs indicated by dashed lines. d-e, Selected protein-protein interactions of significantly altered DEPs found exclusively in individual mice, such as MAPT interactome in 3xTG, and splicing/synaptic interactome in BiG. f, Numbers of DEPs in AD mouse models that were consistently altered in AD. The percentage was calculated using a denominator of 654 AD DEPs that were detectable by MS in mice. g, Strategy for ranking individual proteins by multi-omics using order statistics. (i) All age-dependent proteomic data from 5xFAD and NLGF were initially consolidated into two datasets for the amyloidosis proteome and phosphoproteome. (ii) These datasets were then integrated with 10 additional datasets, which include the mouse transcriptome (5xFAD), 3xTG/BiG proteomes, human genetic data from GWAS, human transcriptomes, proteomes (MCI and two independent AD studies, n = 3), phosphoproteome, and interactome datasets. h, Protein integrative rankings defined by combining 12 datasets. The entire datasets were ranked based on all identified genes/proteins. Subsequently, we extracted the rankings for the AD-mouse shared proteins (n = 275). The top 20 proteins are displayed, with missing values represented by white boxes.

Fig. 5 ∣. AD and the mouse model show transcriptome-proteome inconsistencies which include RNA-independent upregulated proteins enriched in the amyloidome.

a, Workflow for comparison of protein/RNA data to define protein-RNA consistencies. b-c, Scatterplots of protein-RNA comparisons of log₂FC-Z in human (n = 10,781) and 5xFAD mice (n = 8,840). Density is indicated by color gradients. Pearson correlation (R) values are shown. d, Percentage of protein-RNA consistency in the population of z-score altered proteins. e, Overlap of RNA-independent protein changes between human and mouse. f, Workflow of LCM-MS to compare proteomes in plaque and non-plaque regions, quantifying 5,364 proteins. A Venn diagram illustrated the overlap of 31 shared, RNA-independent, upregulated proteins in both humans and mice with proteins enriched in either plaque or non-plaque regions. g, Volcano plot showing proteins enriched in in plaque or non-plaque regions.

Figure 6 ∣. The analysis of AD mouse proteome turnover confirms distinctly different turnover rates for human APP full-length protein and Aβ peptides.

a, Whole proteome turnover analysis in 5xFAD and WT mice was performed using pulsed SILAC labeling (~9-month-old, 5 data points, 3 replicates, totaling 30 mice), TMT-LC/LC-MS/MS (2 batches), and the JUMPt program. The analysis covered a comprehensive set of 8,492 unique proteins. b, Diagram illustrating the 12 identified peptides in the human and mouse APP or Aβ regions. c, PSM counts for the hAPP-specific peptide (peptide 2) and the hAβ surrogate peptide (peptide 10). d, Apparent T₅₀ values were directly determined from turnover curves for the hAPP- or hAβ-specific peptides. e, The curve of free Lys amino acid. f, Corrected T₅₀ values were calculated based on the distance between the protein curve and free Lys curve, using the JUMPt program, which incorporates a mathematical model to account for delays caused by Lys recycling. g. Summary table of hAPP and hAβ T₅₀ values. The corrected T₅₀ values were much smaller than the apparent T₅₀ values.

Figure 7 ∣. The analysis of AD mouse proteome turnover reveals slow half-lives in amyloidome proteins.

a, Pie chart displaying the average proportions of corrected protein T₅₀ values categorized as very short (<0.5 days), intermediate (0.5-30 days), long (30-100 days), and very long (>100 days) for WT and 5xFAD mice. b, Distribution graphs of T₅₀ values in both genotypes, showing the average values and standard deviations. c, Volcano plots of the log₂ fold change and FDR for T₅₀ in 5xFAD compared to WT, with proteins exhibiting changed T₅₀ highlighted in colors and thresholds marked by dashed lines. d-e, Examples of proteins that have shortened or extended T₅₀ in 5xFAD. f, Heatmap illustrating how some proteins with longer T₅₀ may be explained by their localization in plaques, contributing to RNA-protein discrepancies. The side bar indicates log₂FC-Z values for the first three columns or log₂FC values for the last column.