Shotgun-based proteomics of extracellular vesicles in Alzheimer’s disease reveals biomarkers involved in immunological and coagulation pathways

Alzheimer’s disease (AD) is the most common form of dementia and without readily available clinical biomarkers. Blood-derived proteins are routinely used for diagnostics; however, comprehensive plasma profiling is challenging due to the dynamic range in protein concentrations. Extracellular vesicles (EVs) can cross the blood–brain barrier and may provide a source for AD biomarkers. We investigated plasma-derived EV proteins for AD biomarkers from 10 AD patients, 10 Mild Cognitive Impairment (MCI) patients, and 9 healthy controls (Con) using liquid chromatography-tandem mass spectrometry (LC–MS/MS). The ultracentrifuged EVs were washed and confirmed according to the MISEV2018 guidelines. Some AD patients presented with highly elevated FXIIIA1 (log2 FC: 4.6, p-value: 0.005) and FXIIIB (log2 FC: 4.9, p-value: 0.018). A panel of proteins was identified discriminating Con from AD (AUC: 0.91, CI: 0.67–1.00) with ORM2 (AUC: 1.00, CI: 1.00–1.00), RBP4 (AUC: 0.99, CI: 0.95–1.00), and HYDIN (AUC: 0.89, CI: 0.72–1.00) were found especially relevant for AD. This indicates that EVs provide an easily accessible matrix for possible AD biomarkers. Some of the MCI patients, with similar protein profiles as the AD group, progressed to AD within a 2-year timespan.

These necessities have solicited an unbiased approach, where thousands of molecules are simultaneously explored. Proteomics is often used as an untargeted platform for biomarker discovery 6 ; an approach that has already resulted in interesting biomarker candidates for AD, such as neurogranin, neurofilament light chain, TREM2, and YKL-40 3,7,8 . However, failure to replicate results has halted the transition of biomarker candidates from bench to bedside 9 . Blood is a complex source of information, especially for differences in protein concentrations, where the dynamic range can span at least 10 orders of magnitude, from the highly abundant albumin (~ 50 mg/mL) to low abundant cytokines (~ 5 pg/mL) 6 . This dynamic range complicates the extensive profiling using proteomics methods such as mass spectrometry (MS), which is biased towards highly abundant proteins 10 . To partially overcome this obstacle, abundant plasma proteins (up to 22 proteins) are often depleted prior to proteomics analysis. However, several caveats complicate the replicability of these studies, as they introduce variabilities in sample measurements. The caveats include lack of high specificity for antibodies with the removal of non-specific proteins and depletion under denaturing conditions causing co-immunoprecipitation and possible removal of bound proteins 11 . Furthermore, in biomarker studies the presence of confounding factors such as liver and kidney function, and/or other co-morbidities could also affect the reproducibility and interpretation of the results obtained from MS-based proteomics analyses 12 .
In the last decade, extracellular vesicles (EVs) have emerged as important intercellular communicators, regulating both physiological and pathological processes. These vesicles comprise groups of double lipid-layer membranous particles of heterogeneous size and composition. When formed, these entities can be loaded with proteins, lipids, and genetic material such as RNAs and miRNAs. The composition of the EV cargo could therefore reflect the physiological state of the parental cell from where it originated 13 . This important feature has shed light on EVs as novel sources of disease-related biomarkers 14 . EVs can cross the BBB 15 , enabling their release from cells within the CNS, including diseased cells if present and circulate in the blood.
Several studies have investigated biomarkers for AD in EVs; however, their focus was on brain-derived vesicles. These vesicles have indicated that the content of Aβ 42 and various types of tau and, to a minor degree, synaptic and lysosomal proteins may be of use as biomarkers (for references see Refs. 16,17 ). Although an elegant solution to isolate a specific subpopulation of EVs through immunoaffinity-derived methods, extra steps in the isolation procedure are warranted, which concomitantly result in a potential loss of relevant EVs. The specificity of the surface marker used to capture brain-derived EVs, i.e. L1CAM has been questioned due to its implications in cancer metastasis and presence in the tissue of the kidneys 18,19 . In addition, detection of proteins in a minute amount of EVs can be technically demanding. Furthermore, AD is a multifactorial disease, where the peripheral immune system and platelets have been previously implicated in the disease pathology 20,21 . Therefore, we have used a different approach with a simple and standardized procedure using ultracentrifugation of plasma to achieve differential isolation of all types of EVs.
The present study aimed at investigating the potential differences in protein content and relative abundance of blood-derived EVs from patients with AD and Mild Cognitive Impairment (MCI) compared to healthy controls and at elucidating disease-related EV proteome changes and its diagnostic potential.
Using a shotgun-based MS approach and feature selection statistics, we identified a subset of proteins distinguishing healthy from diseased individuals. These proteins were found to primarily be part of immunological and coagulation pathways. Some of the AD patients presented with highly elevated levels of coagulation factor XIII A1 (FXIIIA1) and B (FXIIIB). Additionally, we identified orosomucoid 2 (ORM2), retinol-binding protein 4 (RBP4), and hydrocephalus-inducing protein homolog (HYDIN) to be potential biomarkers for AD.

Materials and methods
Study participants. In total, 30 participants were included; 10 patients with AD, 10 patients with MCI, 10 healthy controls. Technical issues caused one control sample to not be analysed by MS. Patients were clinically verified at the Department of Neurology at Aalborg University Hospital and consecutively enrolled at the time of diagnosis, but before the initiation of treatment. For patients with mild to moderate AD, the diagnosis was based on the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association (NINCDS-ADRDA) 22 and International Classification of Diseases and Related Health Problems 10 th Edition (ICD 10 ) criteria 23 . For MCI patients the diagnosis was based on the Petersen criteria 24 . When found necessary by the physician, patients' cognition was tested using the mini-mental state examination (MMSE), Addenbrooke's Cognitive Examination (ACE), Functional Activities Questionaire (FAQ), as well as paraclinical measurements of CSF markers Aβ, phospho-tau (p-tau), and total-tau (t-tau).
Age-and gender-related healthy donors from the blood bank at Aalborg University Hospital were included for comparison. Donors > 65 years completed a questionnaire related to their mental state, such as memory impairment prior to inclusion. All study participants signed a written informed consent form. The study was approved by the North Denmark Region Committee on Health Research Ethics (N-20150010) and conducted according to the Declaration of Helsinki.
Sample collection and processing. Plasma samples from patients and healthy controls were collected and processed as previously described 25 . Briefly, blood was obtained from the median cubital vein using a 21-gauge needle (Vacuette, Greiner Bio-One, Austria). 9 mL 0.105 M (3.2%) trisodium citrate tubes were used and processed within 2 h after blood collection. Platelet-free plasma was obtained using centrifugation twice at 2500 × g at room temperature for 15 min. After each centrifugation, plasma supernatant was collected until 1 cm above the buffy coat. Samples were subsequently snap-frozen in liquid nitrogen and stored at − 80 °C until analysed. Extracellular vesicle characterisation. EV pellets were characterised using nanoparticle tracking analysis (NTA), western blotting, and transmission electron microscopy (TEM) with immunogold labelling (IEM). The methods describing these analyses and original uncropped images can be found in Supplementary File S1 and Supplementary Fig. S1, respectively.
Preparation by S-Trap micro spin columns. For digestion of proteins in EV pellets, the commercially available S-Trap Micro Spin Columns (Protifi, NY, USA) were used according to the manufacturer's instructions. Briefly, EV pellets were lysed in solubilisation buffer (5% SDS, 50 mM triethylammonium bicarbonate (TEAB), pH 7.55). Proteins were reduced by adding Tris (2-carboxyethyl)phosphine hydrochloride (TCEP) to a final concentration of 10 mM and heated at 95 °C for 10 min, cooled to room temperature, and alkylated using 40 mM iodoacetamide (final concentration) in dark for 30 min. A final concentration of 1.2% phosphoric acid was added to the samples, followed by six times the volume of S-Trap binding buffer (90% MeOH, 100 mM TEAB, pH 7.1). Samples were loaded onto S-Trap spin columns, centrifuged at 4000 × g until all buffer had passed through. The trapped proteins were collected and washed thrice with S-Trap buffer at the same settings. A total of 20 µL digestion buffer (50 mM TEAB) was mixed with 2-5 µg trypsin and incubated overnight at 37 °C. Peptides were eluted using three stepwise buffers; first 50 mM TEAB (40 µL), followed by 0.2 % aqueous formic acid (40 µL), and finally by 50 % acetonitrile with 0.2 % formic acid (35 µL). After the addition of each buffer, samples were centrifuged at 4000 × g. The elutions were pooled, dried by vacuum centrifugation, and resuspended in buffer A (99.9 % water, 0.1 % formic acid). Peptide concentrations were measured by fluorescence using an EnSpire microplate reader (Perkin Elmer, Waltham, MA, USA) and diluted at 1 µg/µL. Label-free quantitative nano liquid chromatography-tandem mass spectrometry analysis. Liquid chromatography-tandem mass spectrometry was performed as previously described 26 using the universal method on an Orbitrap Fusion Tribrid mass spectrometry platform from Thermo Scientific (Waltham, MA, USA). Internal mass calibration was performed by activating the EASY-IC using fluoranthene. Peptides were trapped on a µ-Precolumn (300 µm × 5 mm, C18 PepMap100, 5 µm, 100 Å, Thermo Scientific) and separated on an analytical column (EASY-Spray Column, 50 mm × 75 µm, PepMap RSCL, C18, 2 mm, 100 Å, Thermo Scientific). A 91 min. elution gradient was constructed by mixing buffer A with buffer B (99.9 % acetonitrile, 0.1 % formic acid). Initially, 2 % B was used increasing to 14 % at 3 min, 25 % at 34 min, 40 % at 36 min, 80 % at 37 min, 80 % at 53 min, 2 % at 54 min which was kept at 2 % until 91 min. Samples were injected twice with an amount of 1 µg of sample protein per injection in duplicates, except for one sample which was injected only once due to technical troubleshooting. Replicates were injected with an intermission of several hours to days. The acquisition was performed in the first 60 min of the gradient as previously described 26 with full Orbitrap scans (375 -1500 m/z), a resolution of 120,000, and an automatic gain control (AGC) target of 4 × 10 5 with a maximum injection time of 50 ms. Each cycle time lasted 3 s. Precursor ions with the highest intensity were selected, with an intensity threshold set at 5 × 10 3 , and charge states 2 -7 included. The linear ion trap was used for MS 2 scans at a rapid scan rate with a collision-induced dissociation energy at 35 % and an AGC target of 2 × 10 3 with a maximum injection time of 300 ms. Using the quadrupole, precursor ions were isolated using an isolation window of 1.6 m/z, with dynamic exclusion set to 60 s. The median technical coefficient of variation was calculated for the proteins in each sample. The mean of this was 13.3 %.
Database searches. The 57 raw data files were searched against the human Uniprot database (downloaded 12/03/2019) and using MaxQuant version 1.6.5.0 (Max Planck Institute of Biochemistry, Martinsried, Germany) for label-free quantification (LFQ) analysis 27 . Carbamidomethyl (C) was used as fixed modification. The false discovery rate (FDR) for peptide-spectrum matches (PSMs), proteins, and sites were each set at 1 %. The minimum ratio count for LFQ was set to 1. Tandem MS (MS/MS) was required for LFQ comparisons. For quantification of proteins, unique and razor peptides, unmodified and modified with oxidation (M) or acetyl (protein N-terminal) were used. The function match between runs was used. Reverse sequences were used for decoy search and contaminant sequences were included in the search.

Statistical analysis.
Demographics and clinical characteristics were presented as mean with standard deviations (mean ± SD). Group differences were investigated using either Student's t-test for MMSE, ACE, FAQ, CSF www.nature.com/scientificreports/ Aβ, CSF p-tau, and CSF t-tau or analysis of variance (ANOVA) for age, particle concentration, and particle mean size. Proteins were filtered for potential contaminants, reverse sequences, proteins only identified by site, and at least 2 unique peptides. LFQ values were log 2 transformed and technical replicates were averaged. Distributions were assessed through histograms. Proteins had to have 70 % valid values in at least one group. Venn diagrams were used to investigate proteins common and unique for each group and matched to the top 100 proteins found in EV-related studies using the EV databases; Vesiclepedia 28 and ExoCarta 29 (downloaded 07/04/2020). Unsupervised principal component analysis (PCA) was used to assess data trends. A missing value imputation of width 0.3 and downshift 1.8 was used prior to PCA. Differentially expressed proteins were identified between healthy individuals and AD patients using Student's t-test. A p < 0.05 and log 2 fold change (FC) > 1 or < − 1 was considered statistically significant. A permutationbased FDR < 0.05 with 250 randomizations was adapted to correct for multiple hypothesis testing, reported as q-values. Protein comparisons were depicted using non-log 2 transformed LFQ values. Significant proteins were subjected to enrichment analysis and annotated with the top five significant gene ontology biological process (GOBP) terms using functional annotation clustering analysis by DAVID version 6.8 30,31 . Enrichment scores (ES) and Benjamini-Hochberg corrected p-values, q-values, were shown. STRING analysis was performed using the StringApp in Cytoscape version 3.8.2 32 for protein-protein interactions relating to biological functions for the differentially expressed proteins, requiring a minimum interaction score of 0.4 (medium interaction). When more than one protein ID was listed, the first was used as the allocated protein. Seven of the proteins could not be identified.
The Random Forest algorithm in MetaboAnalyst 4.0 (Xia Lab, Quebec, Canada) 33 was used to select proteins distinguishing AD patients and controls, with missing value imputation as described above. Models selected by Random Forest and biomarker candidates were presented as receiver operating characteristic (ROC) curves.

Results
Clinical characteristics of study populations. For the present study, the protein profiles measured by MS from enriched plasma-derived EVs were compared between the three groups; AD, MCI, and healthy controls. The clinical characteristics and biochemical parameters have been presented in the study by Nielsen et al. 25 . Briefly, measurements of biochemical parameters were within the reference intervals for all participants, except with few anomalies in individuals who had elevated triglyceride and LDL cholesterol levels, though not group-specific. These parameters included markers linked to organ functions, such as liver, kidney, acute-phase reactants, and haematological panels. For clinical measurements, MCI and AD patients presented with lowered MMSE and ACE scores, higher FAQ scores, and slightly lowered CSF Aβ and higher CSF p-tau and t-tau levels ( Table 1). Although the age span of diseased and healthy individuals was overlapping, the difference was statistically significant. Furthermore, AD patients presented with significantly lower MMSE (p: 0.041) and ACE (p: 0.007) scores compared to the MCI patients.
Characterisation of extracellular vesicles. Results from NTA indicated no significant difference in the concentrations of measured particles in the EV pellets between the three groups (Fig. 1A). Similar observations were found for the mean size of the measured particles (Fig. 1B). The size distribution had similar profiles in all three groups (Fig. 1C). Most of the measured particles were within the size range of 100 -200 nm; however, larger particles above 200 nm were also measured in these samples. Western blot confirmed the presence of the EV tetraspanin CD9, as well as the cytosolic marker ALIX (Fig. 1D). Co-isolation of contaminating lipoproteins was confirmed to be present in the samples using Apo-B (Fig. 1D), but was clearly reduced compared to plasma.    (Fig. 1E).

Proteomic analysis of extracellular vesicle related proteins.
A total of 336 proteins were identified.
Among these, 329 proteins were common for all three groups. One protein (HYDIN) was present in both healthy controls and MCI patients, however absent in the AD patient group. Six proteins (F9, TUBB, SLC2A3, CDSN, EHD1, and SACM1L) were present in both AD and MCI patients, but not in the healthy controls ( Fig. 2A). None of the three groups contained uniquely expressed proteins. Furthermore, using the top 100 proteins associated with studies investigating EVs (Supplementary Table S2), based on the databases Vesiclepedia and ExoCarta, we investigated the number of proteins documented in these databases was also present in this study (Fig. 2B). Of the 336 proteins from the current study, a total of 54 proteins overlapped with the top 100 EV proteins, indicating that 16.1 % of the identified proteins were highly associated with EVs, although this observation does not indicate sample purity. The protein profiles were further investigated for intra-and inter-group variations. PCA revealed clustering along the first and second principal components indicating differences in the protein profiles of subjects diagnosed with AD and healthy controls (Fig. 3). Interestingly, samples from the MCI group clustered with both  www.nature.com/scientificreports/ the control (MCI) and AD groups (MCI (AD) ), suggesting some similarities in these patient's protein profiles that were related to either of the groups. When filtering for 100 % valid values, to avoid skewness of data due to imputation, similar observations were made between the groups; however, the distance between the group clusters was less prominent compared to the PCA after 70 % filtration (Supplementary Fig. S2).
Seven of the 10 MCI patients have progressed to AD during the two years from blood sampling until now.
Altered protein expression and pathways related to cognitive impairment. Based on the heterogeneity within the MCI patients as observed in the PCA plot, further comparisons were only focused between AD patients and healthy controls. Analysis of protein expressions between AD patients and healthy controls revealed 63 proteins being differentially expressed, with 57 upregulated in the AD group, and 19 of these being statistically significant after FDR correction (Table 2 and Fig. 4A). Proteins upregulated in the AD group indicated GOBP terms related to the inflammatory response (complement activation, classical pathway and regulation of complement activation) and coagulation processes (fibrinolysis and blood coagulation) (Fig. 4B). STRING analysis of upregulated proteins from AD patients, showed partially biologically related interactions. These proteins had 334 edges, which is significantly more than the 35 expected edges, meaning that the proteins have more interactions among themselves compared to that of a random set of a similar number of proteins. Enrichment analysis revealed that the proteins were mainly involved in biological related to the immune system, leukocyte-mediated immunity, platelet degranulation, and blood coagulation (Fig. 4C,D).
Biomarker candidates for Alzheimer's disease. To identify patterns of regulated proteins enabling classification for diagnosis of cognitive impairment, feature selection using Random Forest was employed.
Based on the ranked important proteins from the Random Forest analysis, three proteins showed a noticeably higher average importance (ORM2, RBP4, and HYDIN). Furthermore, two proteins were of special interest when comparing the protein profiles from AD patients to that of healthy controls. Both coagulation factor XIII subunits, FXIIIA1 and FXIIIB, presented with the highest log 2 FC (FXIIIA1 log 2 FC: 4.6 and FXIIIB log 2 FC: 4.9). Therefore, these proteins were further investigated for their properties as possible single putative biomarker candidates.
ORM2 and RBP4 showed excellent ROC curves with an AUC of 1.00 (95% CI: 1.00-1.00) and AUC of 0.99 (95% CI: 0.95-1.00), respectively (Fig. 6A,B). The protein HYDIN presented with a slightly lower AUC of 0.89 (95% CI: 0.72-1.00) (Fig. 6C). This protein was below the detection limit for the AD group, in contrast to the two former proteins, which were observed to be highly upregulated compared to the control group. Interestingly, as observed in the PCA plot (Fig. 3), six of the MCI patients (MCI (AD) ) clustered with the AD group, and the proteins from the MCI (AD) patients were similarly expressed as the AD patients (Fig. 6A,C). Secondly, the two coagulation factor XIII subunits, FXIIIA1 and FXIIIB, although with the highest log 2 FC, presented with ROC curves with an AUC of 0.88 (95% CI: 0.72-1.00) for FXIIIA1 and an AUC of 0.88 (95% CI: 0.71-1.00) for FXIIIB (Fig. 6D,E). Two of these proteins (RBP4 and HYDIN) were not found to be statistically significantly different in protein comparisons in Table 2, which is due to the missing value imputations used for the Random Forest analysis, since most of their values were below the detection limit in the control group for RBP4 and the AD group for HYDIN.
Due to the observed differences of proteins in EV samples, we investigated if such changes could also be observed in plasma. Therefore, levels of ORM2 were measured in plasma samples, together with levels and activity of FXIII. Interestingly, comparisons of these protein levels and activity in plasma samples did not show any statistically significant differences amongst the groups ( Supplementary Fig. S3), which is in contrast to our findings of relative protein expressions in EV samples.
The results presented in this study further corroborates the importance of investigating EVs and their protein cargo for biological information related to AD pathogenesis and cognitive impairment, as well as relevant biomarker candidates for the diagnosis of AD.

Discussion
The present study investigated the EV proteome derived from cognitively affected AD and MCI patients, and compared it to that of healthy individuals to identify potential biomarker candidates. Distinctive protein profiles were found to efficiently distinguish AD patients from healthy individuals. Ultracentrifugation was selected as the method of choice for our EV enrichment. This selection was based on criteria for easy implementation into a clinical setting, together with the high yield of EVs. Also, for studies of diagnostic character, a high yield is prioritized more than the purity of the sample material 35 , and ultracentrifugation fits these criteria.
Characteristics of EV enrichment revealed no significant differences in concentrations and size of the isolated particles in EV samples. This is in agreement with our previous observations, characterising EVs from a 20,000 × g centrifugation 25 . Another study has investigated plasma-derived EVs from AD patients and found a significant difference in particle concentrations and size compared to healthy individuals, with a lower particle concentration and larger particle size in AD patients 36 . Although our data did not show any significant differences, a similar trend could be observed, with a mean lower amount of particles in AD and larger particles www.nature.com/scientificreports/ measured. Enriched particles in the pellets were confirmed to contain EV specific markers CD9 and ALIX using western blot. CD9 + EVs were further confirmed by IEM. The identified proteins were compared to known EV databases, ExoCarta and Vesiclepedia, and several important EV related proteins were found present in our study.
Based on a study conducted by Kowal et al. 37 , these EV related proteins could be divided into specific subgroups depending on EV properties; proteins in large EVs (ACTN4 and ACTN1), proteins in light small EVs (ADAM10), and proteins in multiple EVs (GAPDH and CD9). Furthermore, the analysed samples also contained proteins associated with lipoproteins APOA1, APOA2, and APOB, which can be co-precipitated contaminants during EV enrichment 38 . Thus, proteins linked to EVs were identified, together with known contaminants, which was an expected outcome considering the nature of the EV enrichment method used in the study 39 .
Neuroinflammation is an integral part of the AD pathology, where both brain-resident immune cells and the peripheral innate immune response play a central role in pathological processes 40 . ORM, an acute-phase protein existing as two subtypes in humans, ORM1 and ORM2, is mainly produced in the liver. ORM1 accounts for approximately 75 % of the subtypes and the protein is sharply increased under pathological conditions, such as inflammatory stimuli 41 . In a mouse model, ORM2 has been shown to be the predominant subtype present in brain tissue 41,42 . During late phase inflammation reactive astrocytes release ORM2 to modulate anti-inflammatory activity on activated microglia with impaired clearance function due to continued Aβ stimulation 42 . The study also observed that only the hippocampal astrocytes appeared to produce ORM2 during systemic inflammation, which is interesting, as the hippocampus is the most vulnerable brain region to an inflammatory response, as well as one of the first areas in the brain to be affected by AD pathology 42 . Moreover, the authors found significantly increased levels of ORM2 in plasma from MCI and AD patients compared to healthy controls 42 . Their findings are somewhat in contrast to ours, as we only detected a difference in EV isolates and not in plasma samples of the AD patients. Similar observations have been made for CRP, where an increased expression was found in EVs; however, no differences were measured in serum CRP 25 . Tight junction proteins are important BBB components to maintain its functional integrity, and a previous study has shown that ORM2 is able to positively affect the BBB integrity through increasing the expression of tight junction proteins zonula occludens 1 and occludin 43 . Our findings of increased ORM2 expression could thus be a positive response to regulate the inflammatory responses and an attempt to re-establish BBB integrity.
In this study, we found FXIIIA1 and FXIIIB to be highly elevated in some AD patients, and the literature shows an interesting connection between levels of FXIII and disease pathology. FXIII exists as a tetramer present in the circulatory system, consisting of two A subunits and two carrier B subunits. The A subunit is the catalytic active part with the transglutaminase function 44 . Aβ proteins accumulate along the brain vasculature in cerebral amyloid angiopathy (CAA), which is present in over 90 % of AD patients 45 . In a mouse model for chronic cerebral hypoperfusion, Shi et al. 46 found significantly expressed FXIIIA in several brain areas such as the hippocampus, thalamus, and neocortex. These findings indicated an association between FXIII, BBB impairment, and cerebrovascular damage, further contributing to AD-related neurodegeneration. Due to the transglutaminase properties, FXIII has been shown to cross-link Aβ into highly stable multimers, which are resistant to proteolytic breakdown 47,48 . These FXIII-Aβ complexes were also shown to co-localize along the cerebral vasculature in CAA 48 . In addition, de Jager et al. 48 showed a protective function of these complex formations, where the binding of FXIII with Aβ protected smooth muscle cells in the cerebral vasculature from the cytotoxicity of Aβ, thus avoiding further damage to the BBB by the formation of a highly stable clot 48 . Lastly, FXIII has been positively associated with neuron-derived EVs released after traumatic brain injury. Furthermore, the FXIII protein was found to bind to neurotoxic forms of Aβ and by EVs delivering these cross-linked proteins into near and distant neurons 49 .
In our study, ORM and FXIII were found significantly elevated in EV enriched samples, but not in plasma. The amount of EVs was comparable among the groups, as indicated by the NTA measurements. These proteins (FXIII and ORM) could either be loaded into the EVs during biogenesis or bound to the surface of EVs, as plasma proteins are known to do 50 . Since the amount of EVs is comparable across all individuals, the increased amount observed could be related to a higher binding affinity of the proteins for the disease-related EVs, if the proteins are bound to these entities. Alternatively, an increased amount of these proteins is possibly loaded into the EVs in pathological conditions.   (334) is identified compared to the expected (35). Enrichment analysis reveals that these proteins are part of the immune system process (blue nodes) and leukocyte mediated immunity (red borders), and (D) platelet degranulation (green borders) and blood coagulation (yellow nodes).   www.nature.com/scientificreports/ RBP4 is secreted by adipocytes as an adipocyte-derived hormone, adipokine 51 . In addition, RBP4 has been shown to bind transthyretin, a carrier protein able to transport Aβ from the brain to the periphery, thus resulting in lower Aβ neurotoxicity 52 . Furthermore, RBP4 has also been shown to transport retinols, such as vitamin A 53 . Vitamin A has been shown to have anti-oxidative and cell-protective effects 54 and be able to inhibit the formation of Aβ oligomers in AD 55 . Thus, the increased intensities observed in this study could potentially be a response to the increasing accumulation of Aβ in relation to the AD pathology. However, in the current literature there are conflicting findings related to RBP4 in AD, probably due to the sample material analysed. In post-mortem brain samples, RBP4 expression has been found increased in AD compared to healthy individuals 56 . Another study found gradually decreased levels of RBP4 in CSF samples from controls to MCI patients and finally to severe cases of AD 57 . Ishii et al. 51 sought to investigate peripheral levels of RBP4 in plasma samples as a potential biomarker for AD; however, they found no differences between AD and cognitively healthy subjects. As we did not use plasma directly, but EVs derived from the circulation, this could be an explanation for the discrepancy. Further similar investigations are needed to determine the role of RBP4 as a potential biomarker for AD. The brain ventricles are covered with motile ciliated epithelium, aiding the distribution and flow of the CSF from the choroid plexus 58 . Cilia consists of an axoneme, a microtubule-based cytoskeletal structure, with HYDIN being a known axonemal protein 59 . Mutations are known to occur in the HYDIN gene, causing dysregulation of ciliary movement and leading to hydrocephalus, an excessive accumulation of CSF within ventricles of the brain 59 . Ventricular enlargement is a characteristic of neuropathological changes associated with cognitive impairments, including MCI and AD 60 . Our EV measurements of HYDIN indicated that it was not measurable in AD, which could be explained by atrophy of the brain tissue in AD, usually identified by magnetic resonance imaging 60 , thus causing dysregulation of the protein expression and ciliary movement, leading to ventricular enlargement. However, the role of HYDIN in relation to AD need to be further investigated before any concise role can be determined for its involvement in disease pathology.
Both STRING and DAVID enrichment analyses showed that proteins upregulated in AD were involved in biological processes related to immunological and coagulation processes. We have presented proteins involved in these processes, such as FXIIIA1, FXIIIB, and ORM2. Furthermore, leukocyte-mediated immunity and platelet degranulation have also been shown to play a role in AD pathology. There is evidence of involvement of the peripheral immune system in AD, with infiltrating leukocytes aiding in phagocytosis of Aβ. However, it is also proposed that these infiltrating cells change to an inflammatory phenotype, due to the neuroinflammation in the brain environment, and thus contributing to this ongoing inflammatory process 20 . Platelets are the main source (~ 90 %) of amyloid precursor protein (APP) and Aβ in blood. Aβ is stored in α-granules and released upon stimulation causing degranulation e.g. by thrombin, thus contributing to the level of circulating Aβ. Soluble APP can exert inhibition of platelet degranulation, while Aβ counteracts this effect 61 .
This study has some limitations. Firstly, small patient and control populations were used for this exploratory study, even though clear differences were observed between healthy individuals and patients with AD. Secondly, not all clinically verified AD and MCI patients had measurements of CSF Aβ and tau biomarkers, as it was not necessary for the diagnosis. These measurements of established biomarkers could have been correlated with our proposed biomarker candidates. Thirdly, the control group had a slightly lower age compared to the patient groups, as recruitment of older blood donors was not possible. This age difference was also significant, however, given that this age span is grouped in proteomics studies, this was probably of minor importance 62 . Fourthly, proteomics data contains missing values, which have been imputed for some of the analyses.
Thus, our findings elaborated on the global effort of identifying blood-based biomarkers for the diagnosis of AD. However, with a discovery-based study, further investigations are warranted to replicate our findings in a larger independent population of cases and controls. Seven of the patients that were initially classified as MCI progressed to AD during the study period. Some of these also revealed increased levels of ORM2 and RBP4, indicating their potential as biomarkers for disease progression.

Conclusions
Based on our initial work, the main finding is that EVs provide an accessible matrix for biomarker discovery, with several proteins involved in inflammation and coagulation processes. We found 10 proteins to distinguish AD patients from healthy individuals. Especially ORM2, RBP4, and HYDIN showed high specificity sensitivity. Furthermore, coagulation factor XIII subunits FXIIIA1 and FXIIIB presented to be significantly upregulated in some of the AD patients. Some of the MCI patients presenting similar protein profiles as the AD patients progressed to clinically verified AD within two years, giving confidence to our findings.