Cellular transcriptional alterations of peripheral blood in Alzheimer’s disease

Blood expression data was obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database [41, 42], a longitudinal multicenter dataset in the US and Canada containing clinical, imaging, genetic, and biochemical data to identify biomarkers for early detection and progression of AD. Enrollment and diagnosis criteria for ADNI have been described previously in detail [43]. Briefly, enrolled participants were between 55 and 90 years old without other psychiatric and neurologic diseases. Subjects with AD were diagnosed with the National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria for probable AD [44]. Subjects with MCI and normal control (NC) participants were determined through whether meeting the Jak/Bondi Criteria or not with neuropsychological measures [45]. The expression data were sequenced using the Affymetrix Human Genome U 219 array and normalized by the Robust Multichip Analysis (RMA) [46].

To validate the findings in ADNI, we performed the same analysis on other three independent peripheral blood transcriptomic datasets, including the AddNeuroMed datasets ANM1 [47] and ANM2 [47], and the AD cohort in Zhangjiang international Brain Biobank (ZIB-AD [48]). The ANM consortium is a large cross-European, public-private consortium for AD [49]. Blood RNA samples from ANM1 and ANM2 were profiled on Illumina Human HT-12 v3 and v4 Expression BeadChips, respectively. The probe-set level intensities of each set were normalized using the RMA method in the R package affy [50]. The ZIB is an ongoing imaging genetic neuropsychiatric cohort in China. Expression profiling of AD cohort in the ZIB was sequenced on Illumina NovaSeq^TM 6000 with 150-bp paired-end. In addition, we also performed cellular deconvolution on a human AD brain transcriptomic dataset (i.e., Mayo RNAseq [51, 52]) and a spatial transcriptomics dataset in brain of AD mouse model (i.e., Alzmap [53, 54]).

We first applied EPIC [55] to estimate the cellular proportions based on deconvolution algorithm. Briefly, the bulk gene expression (b) is modeled as the sum of the expression from the pure cell types composing this sample, i.e., b = C × p, C is a reference gene expression matrix for cell types; and p is a vector of the proportions from the K cell types. The reference cell gene expression profile was obtained from RNA-Seq data of sorted immune cells [56‐58]. The cellular proportions were then inferred by a least-square optimization. EPIC estimates showed a remarkable agreement with the cell fractions computed with flow cytometry in both blood and tumor samples [55].

To avoid the method bias, we also quantified the immune cells with other computational tools, including quanTIseq [59], CIBERSORT [60], and MCP-counter [61]. QuanTIseq is based on a signature matrix derived from 51 RNA-seq data sets of blood-derived immune cells from ten different immune cell types. Both EPIC and QuanTIseq allow estimating the fraction of uncharacterized cells. CIBERSORT, demonstrating robust deconvolution performance and high accuracy in microarray data via nu support vector regression [60], was also applied to estimate the cell fractions and to infer sample-level gene expression for each cell type. A leukocyte RNA-Seq signature matrix comprised of six peripheral blood immune subsets (GSE60424) was taken as the specialized knowledgebase [57, 60]. The above three methods force the estimates to be non-negative and to sum up to one and therefore could be interpreted directly as cell fractions and compared across different immune cell types. Furthermore, MCP-counter, a transcriptomic marker-based approach [61], was applied to quantify the absolute abundance of immune cells. Here, we applied the R package immunedeconv [62] to implement uniformly the above algorithms with a non-log scale gene expression matrix as the input.

Limma [63], an R package based on linear regression for both RNA-seq and microarray data, was applied to identify bulk differentially expressed genes (DEGs, P < 0.01) of AD and MCI in each dataset. The normalized expression matrixes of microarray and RNA-seq data were used as inputs while adjusting for covariates of age and gender. In addition, the Wilcoxon test that performs robustly with a lower false discovery in population studies [64] was applied to further validate the DEGs.

TOAST [65], a reference-free deconvolution method, was applied to detect differentially expressed genes in each cell type between AD or MCI vs. NC. First, we made a model design using a phenotype matrix (including age, gender, and disease status) and the cellular proportions estimated from EPIC. Then, we fitted the linear model for non-log-transformed expression data and the model design. Finally, we tested the disease effects (i.e., differences) in each cell type while controlling for covariates.

We identified the enriched biological processes in Gene Ontology for DEGs using Metascape [66, 67], which utilizes the hypergeometric test and Benjamini-Hochberg p-value correction algorithm to identify significant terms. To address the issue of term redundancies, Metascape will compute pairwise similarity scores between any two enriched terms based on a Kappa-test score [68] and performs hierarchical clustering using the similarity matrix. Then, the hierarchical tree is trimmed with the 0.3 similarity threshold into separate clusters. Metascape choose the most significant (lowest P-value) term within each cluster to represent the cluster in the bar graph and heatmap representations. We also constructed enrichment networks, in which the nodes represent significant terms and the edges connect these nodes for which the similarity scores are above 0.3. The networks were visualized using Cytoscape v3.8.0 [69].

For DEGs in neutrophils, we re-constructed protein-protein interaction (PPI) network using Metascape under the combined (all) option. Metascape extracts protein interactions from STRING [70], OmniPath [71], InWeb_IM [72], and BioGrid [73] database for input gene/proteins. Then, the Molecular Complex Detection (MCODE), a graph-theoretic clustering algorithm based on vertex weighting by local neighborhood density and outward traversal from a locally dense seed protein to isolate the dense regions, was applied to identify densely-connected modular clusters with default parameters [74]. For each cluster, Metascape further applies function enrichment analysis and uses the top enriched terms to annotate its biological roles. We also identified candidate hub genes with high degree centralites or clustering coefficients using cytoHubba.

All statistical analyses were performed using R (version 4.0.2). ANOVA (analysis of variance) test and chi-squared test were applied to compare continuous demographics variables and categorical variables, respectively. Comparison of cell proportion between groups was performed using the two-tailed Wilcoxon test, whereas the one-tailed Wilcoxon test was applied to determine whether the NLR of AD and MCI is higher than NC or not. Correlations between proportion and phenotype traits were assessed with Spearman’s correlation. Overlaps of DE genes between two sets were assessed using a hypergeometric test. The enrichment of PPI modules in AD risk/associated gene lists was determined with the application of Fisher’s exact test.

We obtained the blood-derived gene expression data from the individuals in the ADNI database [41], including 116 AD, 382 MCI, and 246 elderly NC. We first applied a deconvolution algorithm EPIC [55] to estimate the immune cell proportions for these samples using the gene expression profiles of the reference immune cell types [56‐58] (Fig. 1A). We found that the proportion of B cells was significantly lower in AD compared to NC (P = 3.3e−4, Wilcoxon test) and MCI (P = 9.7e−4, Wilcoxon test), while the proportion of neutrophil was significantly higher in AD compared to NC (P = 4.8e−3, Wilcoxon test). Then, we used the neutrophil-to-lymphocyte ratio (NLR), a simple ratio between the neutrophil and lymphocyte counts measured in peripheral blood, to estimate the neutrophil-mediated systemic inflammation for these samples. We found that the NLR in AD exhibited a significant increase compared to controls (P = 9.9e-3, one-tailed Wilcoxon test) (Fig. 1B). To further validate this observation, we also estimated the immune cell proportions in other three independent bulk blood transcriptomic datasets (i.e., ANM1, ANM2 [47], and ZIB [75]) using three alternative computational tools (i.e., quanTIseq [59], CIBERSORT [60], and MCP-counter [61]). As expected, we found a significantly higher NLR in AD than in NC in most of the datasets when using different computational tools except for ANM2 dataset under CIBERSORT (Fig. 1B). Although the NLR difference between MCI and NC is not statistically significant in most tests (Additional file 1: Fig. S1), we could still observe the same trend as in AD vs. NC. Collectively, these findings revealed an elevation of inflammation response in AD and MCI than in NC.

We then performed principal component analysis (PCA) on the immune cell proportion from these samples. The first two principal components (PCs) can explain 59.9% of the total variance-covariance of the immune cell proportion in the samples. The PC1 (35% of the total variance-covariance) was most strongly contributed by neutrophils (negative) and B cells (positive) (Fig. 1C). The PC2 (24.9% of the total variance-covariance) was most strongly weighted on CD8⁺ and CD4⁺ T cells. The PCA scatterplot revealed that the immune cell profiles in most AD cases could not be distinguished from those in controls (Fig. 1C), but a subset of them showed lower PC1 positivity scores, i.e., a distinctively lower proportion of B cells and/or an increased proportion of neutrophils, indicating a lower adaptive activity and/or a higher innate and pro-inflammatory activity in these samples.

We then estimated the correlations between immune cell proportion, cognition, disease diagnosis, brain structural features for early detection of AD [76] (derived from MRI data), and demographic variables in NC and AD participants (Fig. 1D). Consistent with the above results, NLR and neutrophil proportion showed positive correlation with disease diagnosis, while exhibited negative correlations with hippocampus and left entorhinal cortex volumes, as well as with cognitive functions (Fig. 1D and Additional file 1: Fig. S2). A similar pattern of correlation could be observed when estimating using all AD, MCI and NC individuals (Additional file 1: Fig. S3). Moreover, we noticed that the cellular proportion and NLR could be affected by age and gender (Fig. 1D). Considering the variation of age and gender across different diagnosis groups (Additional file 1: Fig. S4) that could be important confounding factors, we therefore performed logistic regression between disease status and cellular proportion by adjusting for the two confounding factors (i.e. age and gender) (Additional file 2: Table S1). After adjusting for the confounding factors, although with a decreased statistical significance level, the proportions of neutrophils and B cells remained to show significantly positive (t = 2.112, P = 0.035, logistic regression) and negative (t = − 2.553, P = 0.0111, logistic regression) correlation with AD status in ADNI dataset, respectively. We also confirmed these results in ANM1 and ANM2 datasets (Additional file 2: Table S1). We observed the same-direction while insignificant correlation between the proportions of neutrophils and B cells and MCI status (Additional file 2: Table S1). In addition, the difference of the proportion of CD4⁺ T cells between MCI and AD did not remain statistically significant (P = 0.262, logistic regression) after adjusting for the confounding factors of age and gender. Consistent with previous studies [77], these results indicated that in addition to age and gender factors, the occurrence of AD partially drove the increase in neutrophil proportion and NLR.

The differentially expressed genes (DEGs) in AD were identified in each dataset through linear regression by adjusting for the co-founding factors (Fig. 2A), where the significant overlap and strong concordance with those identified using the Wilcoxon test confirmed the reliability of the results (Additional file 1: Fig. S5). Consistent with the changes of immune cell proportion and previous studies [78], we found that the bulk upregulated and downregulated genes in AD were significantly enriched in neutrophil activation/degranulation and B/T cell receptor signaling pathways in three out of four datasets, respectively (Fig. 2B, C). We also observed downregulation of neutrophil activation/degranulation and upregulation of RNA catabolic process in MCI in ANM1 and ANM2 datasets (Additional file 1: Fig. S6).

To further examine the intrinsic expression changes of individual immune cell types, we identified the DEGs in each cell type by adjusting for the cell proportion and co-founding factors in ADNI dataset (Fig. 2D). We compared the bulk DEGs and cell-type-intrinsic DEGs (Additional file 1: Fig. S7) and found that only a small fraction of the DEGs in each cell type overlap with the bulk DEGs. We noticed that four (i.e., neutrophils, monocytes, NK cells and CD4⁺ T cells) out of the six cell types exhibited significant overlap with the bulk DEGs; the neutrophils showed the strongest overlap (P = 9.5e−11, hypergeometric test), which may be due to the high abundance (55-70%) of this cell type in the peripheral blood.

We next explored the shareness of the perturbed biological pathways across the peripheral immune cell types in AD from ADNI dataset. As shown in Fig. 2E, the dysregulated genes in all or most of the cell types are associated with the regulation of immune response, regulation of defense response, cellular response to cytokine stimulus, and regulation of inflammatory response, which is consistent with the common dysregulated signalings among peripheral blood mononuclear cells in AD revealed by single-cell RNA-seq analysis [40]. Moreover, we observed cell-type-specific disruptions in some biological processes, for instance, Ras protein signal transduction in B cells, positive regulation of cell migration in NK cells, regulation of biomineralization in CD4⁺ T cells, antigen processing, and presentation in monocytes. Interestingly, we found that there was a significant enrichment of DEGs in neutrophils for ncRNA metabolic process, peptide and ATP metabolic process, ribosomal subunit biogenesis, and mitochondrion organization, which were also downregulated in the bulk level, indicating that the neutrophils could be the primary factor in the peripheral transcriptomic alterations. The mitochondrion dysfunction and energy hypometabolism have been reported as one of the most consistent and earliest abnormalities in AD and MCI and could be a link to Aβ production [79], highlighting the putative roles of neutrophils in the pathobiology of AD.

The protein-protein interaction (PPI) networks can help elucidate the biochemical complexes or signal transduction components that govern the biological outputs. Given the putative functions of neutrophils in AD pathogenesis, we re-constructed PPI networks for the DEGs in neutrophils using Metascape [66, 67]. We identified densely connected modular clusters (with a minimum size of three proteins) which represent specific molecular complexes and functional units [80, 81] (Fig. 3A and Additional file 2: Table S2). For example, the most significant modular cluster #1 is enriched in ribonucleoprotein complex biogenesis, including ribosome protein families, such as RPL35, RPL7A, and RPL6. The modular clusters #2 and #3 are enriched in cytokine-mediated signaling, including interleukins (ILs) and tumor necrosis factors (TNFs). We also observed other biologically meaningful functional units enriched in clusters #3, #4, and #5, including ATP metabolic process, mitochondrial transport, and mitochondrial outer membrane-associated with mitochondrial organization.

To confirm that the modular clusters are associated with AD pathogenesis, we then estimated the enrichment of the genes in the modular clusters in the AD genetic risk gene set or the AD-associated gene sets from diverse sources (Fig. 3B). We found that modular clusters #2 and #3 were significantly enriched in more than one AD-related gene sets. The modular cluster #2 exhibited mainly enrichment in the regulation of leukocyte cell-cell adhesion and activation. Interestingly, we found that two hub genes in modular cluster #2, IL1B (encoding pro-inflammation members of the interleukin 1 cytokine family) and FAS (encoding TNF-receptor superfamily), were upregulated in neutrophils from AD participants (Fig. 3B, Additional file 1: Fig. S8A-B). We also noticed that modular cluster #3, a mitochondrion organization and cytokine-mediated signaling pathway-related module, was significantly enriched for AD genetic risk genes from GWAS and meta-analysis of large LOAD consortium data sets [82]. For example, the expression level of CD33 (encoding a member of the sialic acid-binding Ig-like lectin family of receptors and expressed on myeloid cells and microglia [83, 84]) shows positive correlation with the hippocampus and entorhinal cortex volumes (Additional file 1: Fig. S9). In previous studies, it has been reported that CD33 might play an important role in Aβ clearance and neuroinflammatory pathways mediated by microglia in the brain, and its structural variants and SNP were associated with a higher risk of AD [15, 85]. In addition, we would like to notice that CD33 was one of the hub genes with high clustering coefficients in modular cluster #3 (Additional file 1: Fig. S8D). These results highlight the putative functions of modular cluster #3 in neutrophils in AD pathogenesis.

We then performed biological process enrichment analysis for the DEGs in neutrophils. We found inhibition of mitochondrion organization and activation of cytokine-mediated signaling pathway and inflammatory response in AD vs. NC (Additional file 1: Fig. S10). To validate the disruptions of signaling pathways in neutrophils, we further performed pathway enrichment analysis using the DEGs identified from different datasets (Fig. 3C-D). Although the DEGs from different datasets exhibited limited overlap (Fig. 3C), we would like to highlight the convergence in the enriched pathway categories across at least three cohorts, including metabolic process, immune system process, and response to stimulus (Fig. 3D, Additional file 2: Table S3). In addition, we observed similar pathway enrichment patterns in MCI across different cohorts (Fig. 3E, Additional file 2: Table S4).

Compelling evidence has demonstrated that the dysfunction and deterioration of blood-brain barrier (BBB) may play key roles in AD pathogenesis through a feedback loop with the accumulation of Aβ and accelerate cognitive impairment and the onset of dementia [86‐88]. Blood-derived leukocyte subpopulations, such as lymphocytes and monocytes, have been also identified in the brains of AD patients and animal models [6, 30, 89].

To check whether the neutrophils infiltrate into AD brains, we estimated the proportion and abundance of neutrophils in AD brains using transcriptomic dataset from 97 AD individuals and 105 controls (i.e., Mayo RNA-seq [51]). Consistent with the extravasation migration of neutrophils into brain [31, 36], we observed that the absolute abundance of neutrophils was significantly increased in AD brains in both cerebellum (CBE) (P = 4.1e−8, Wilcoxon test) and temporal cortex (TCX) (P =1.1e−7, Wilcoxon test) (Fig. 4A). We also found that the abundance of neutrophils exhibited significantly positive correlation with Braak tau neurofibrillary tangle staging score (Fig. 4B and Additional file 1: Fig. S11A) and Thal amyloid phase (Fig. 4C and Additional file 1: Fig. S11B) in the two brain regions. However, we did not observe significant difference of the proportions of neutrophils between ADs and NCs (Additional file 1: Fig. S12), which could be due to the massive elevation of brain immune cells (e.g., microglia and astrocytes). Furthermore, we quantified the abundance of neutrophils in a spatial transcriptomics dataset from AD mouse model (i.e., Alzmap in 3, 6, 12, and 18 months old) [53]. As expected, we found that the abundance of neutrophils show significant increase in both hippocampus (HP) and cortex (CX) at the late stages (i.e., 12 and 18 months) instead of at the early stages (3 and 6 months) (Figs. 4D–E).