Enrichment of extracellular vesicles using Mag-Net for the analysis of the plasma proteome

Abstract

Extracellular vesicles (EVs) in plasma are composed of exosomes, microvesicles, and apoptotic bodies. We report a plasma EV enrichment strategy using magnetic beads called Mag-Net. Proteomic interrogation of this plasma EV fraction enables the detection of proteins that are beyond the dynamic range of liquid chromatography-mass spectrometry of unfractionated plasma. Mag-Net is robust, reproducible, inexpensive, and requires <100 μL plasma input. Coupled to data-independent mass spectrometry, we demonstrate the measurement of >37,000 peptides from >4,000 proteins. Using Mag-Net on a pilot cohort of patients with neurodegenerative disease and healthy controls, we find 204 proteins that differentiate (q-value < 0.05) patients with Alzheimer's disease dementia (ADD) from those without ADD. There are also 310 proteins that differ between individuals with Parkinson's disease and without. Using machine learning we distinguish between individuals with ADD and not ADD with an area under the receiver operating characteristic curve (AUROC) = 0.98 ± 0.06.

Fig. 1. Mag-Net affinity capture workflow.

a SAX magnetic beads are incubated with plasma to sieve and bind negatively charged membrane-bound particles. The beads are immobilized with a magnet, and abundant plasma proteins are gently washed away. The membrane particles are lysed on the beads, and proteins are reduced and alkylated, precipitated back onto the beads, and then digested with trypsin. Peptides are collected and analyzed using nLC-MS/MS. b The Kingfisher robot method to automate the Mag-Net workflow.

Fig. 2. DIA-MS/MS analysis of the peptides from enriched membrane particles and unfractionated plasma.

a The number of peptides detected (mean +/− SD) from three preparations of the same plasma sample at a 1% FDR with three different analysis strategies. The different analysis strategies were the use of EncyclopeDIA with a Prosit library (blue), EncyclopeDIA with a chromatogram library generated from gas phase fractionation (yellow), and DIA-NN in library-free mode. EncyclopeDIA reports peptides detected, and DIA-NN reports peptide precursors. The Mag-Net protocol detected a minimum of 451% more peptides than using the same analysis with unfractionated plasma. b Proteins detected (mean +/− SD) at a 1% FDR with the three different analysis strategies. The use of EncyclopeDIA with an on-column chromatogram library returned the most peptides and proteins and, thus, was used for all further analysis. c Differential analysis of proteins between the enriched membrane particles (blue) and the same unfractionated plasma sample (red). d The 2877 proteins that were enriched in the Mag-Net protocol were used to assess the enrichment of Jensen COMPARTMENTS using the Enrichr tool. * DIA-NN reports peptide precursors, whereas EncyclopeDIA reports peptides. This makes the comparison of peptide detections between tools challenging.

Fig. 3. Characterization of Mag-Net enriched membrane particles.

a Fold enrichment of protein groups compared to total plasma (mean +/− SD, N = 3). b Fold depletion of protein groups compared to total plasma (mean +/− SD, N = 3). c Nanosight particle count. d Representative micrographs of captured membrane particles on bead surfaces using transmission electron microscopy.

Fig. 4. Dynamic range of the plasma proteome.

The peptide signal was measured using the average of triplicate measurements performed using PAC total plasma digestion (a) and the Mag-Net extracellular vesicle capture method (b). If a peptide was not detected in a specific LC-MS run, the data from the runs where it was detected was used to provide boundaries to compute a background-subtracted area where the peptide would be located if it were present. The peptide areas were normalized by equalizing the total ion current (TIC) between runs, and the protein abundance calculated by summing the peptides mapping to a parsimonious protein group. The EV marker proteins are shown in blue, and the classic abundant plasma proteins in red. The signal plotted for the protein abundance in the total plasma sample spans ~ 10 logs. Despite the classic plasma proteins being depleted by ~ 95% following Mag-Net, those proteins are still some of the most abundant proteins in the sample.

Fig. 5. Summary of the matrix matched calibration curve data.

a Schematic of the volumetric mixing of human plasma with chicken plasma for 7 of the dilution points. The percentage represents the percent volume of human plasma, where the remainder represents chicken plasma. The mixed plasma samples were enriched and digested using the Mag-Net protocol. b The measured Log₂ area ratios versus the expected % human plasma. The dilution series was measured in triplicate from the lowest to the highest percent human. Following peak detection with EncyclopeDIA, peptides conserved between human and chicken were removed. The peptide background-subtracted peak areas were determined using Skyline and the ratio of the peptide area relative to the area from the 100% undiluted human sample.

Fig. 6. Application of Mag-Net for the assessment of markers of ADD, related diseases, and controls.

a We used a carefully selected cohort of plasma samples from 40 individuals (10 with ADD, 10 with PDD, 10 with PDCN, and 10 who were HCN). The individual samples, in addition to two different controls, were prepared using Mag-Net. The two pooled plasma samples were prepared four separated times and peptides quantified by data-independent acquisition. One sample was used as a quality control for the quantitative measurements, and the second was prepared as a reference control. b The number of peptides detected when the FDR was controlled on the per-run level (red lines) and on the entire experiment level (light gray). The dark gray represents the number of peptides with a non-zero quantity. c The log2 peptide abundances measured between 4 separate preparations and the measurement of the same quality control sample. The lowest Pearson correlation coefficient between pairs was 0.965. d Effect of different data processing steps on the coefficient of variation of different protein quantities. The top row shows the RAW protein quantities, followed by the quantities with between sample median normalization, and the effect of imputation in the last row.

Fig. 7. Measurement of proteins that provide separation between disease cohorts.

a Heatmap of 1093 proteins that had an ROC > 0.7 from any of the six pairwise analyses described in the text. b A receiver operator characteristic (ROC) curve that illustrates our ability to correctly separate people with ADD from other groups (i.e., PDD, PDCN, and HCN) using a hard-margin SVM with linear kernel. The plot shows the mean and standard deviation of 10 iterations of five-fold stratified cross-validation with different random splits. The results shown are the average performance of the model on held-out samples, which were not included in optimizing the machine learning algorithm.

Fig. 8. Selection of proteins with altered abundance in the ADD patient group.

Illustration of selected protein-coding genes that are either increased (INPPL1, VPS13A, DNAJC10, S100B, and HTT) or decreased (VCP, MAP2K2, USP15, JMJD6, and CD36) in ADD plasma when compared with the other three disease groups. The analysis represents a pilot cohort of plasma from 40 individuals: 10 with Alzheimer’s disease dementia (ADD); 10 Parkinson’s disease (PD) cognitively normal (PDCN); 10 PD with dementia (PDD); and 10 age-matched healthy and cognitively normal (HCN) individuals.