Cell-of-Origin Deconvolution of Vesicle Proteomes with a Learned Packaging Map

A tube of blood or urine contains trillions of extracellular vesicles. These are tiny membrane bubbles, 30–150 nm across, that cells shed into body fluids, and they carry proteins from the cell that made them. Measuring the proteins in all of these vesicles at once could serve as a repeatable, non-invasive window into tissues that cannot be biopsied. The problem is that the measurement is an average. Vesicles from platelets, red cells, immune cells, kidney epithelium, a tumor and perhaps a few neurons all land in the same mass-spectrometry run. The signal from the rare source, which is usually the clinically interesting one, is diluted and mixed with everything else. This is the same "masking" problem that bulk tissue sequencing has, but it is harder in two ways. A vesicle is not a random sample of its parent cell's proteins, because cells actively load some proteins into vesicles and exclude others. And the proteins that would reveal a rare source sit near the mass spectrometer's detection limit, where they are often simply missing.

We propose to estimate what fraction of a bulk vesicle proteome comes from each cell type. The model has two parts. The first is a learned packaging map that predicts a cell type's vesicle proteome from its whole-cell proteome. The second is a mixture likelihood that treats undetected proteins as "below the detection limit" rather than as unknown values to fill in. The packaging map is what lets the method handle cell types for which nobody has a vesicle reference, which describes most primary and disease-state cells. The censored likelihood is what lets it put an honest lower bound on how rare a source it can see. We test both against ground truth built by spiking known amounts of cell-line vesicles into plasma vesicles.

Core Claim

Two pieces of prior work come closest, and each stops short of this. Shi et al. unmix plasma exosome RNA into tumor and non-tumor compartments using a Bayesian model with a single "packaging coefficient." They reach a response-prediction AUC of 0.805 in discovery and 0.84 in validation. Their coefficient is fixed and linear, and their output is a two-way split. scpDeconv deconvolves protein mixtures using single-cell proteomic references. It fills in missing proteins with an autoencoder and removes platform differences with adversarial training. It was built and tested on solid tissue, where the bulk sample really is a weighted sum of cells.

Our claim has three parts:

  1. Vesicle protein loading is predictable enough from protein-level features to transfer to cell types the model has not seen. Useful features include membrane association, tetraspanin and endosomal-sorting partners, and localization.
  2. A deconvolution that uses this transferred map recovers cell-of-origin fractions in plasma vesicles substantially better than applying whole-cell references directly.
  3. Modeling non-detection as censoring lowers the smallest detectable source fraction compared with imputation.

If the first part fails, so that loading turns out to be idiosyncratic per cell type, that is itself a useful negative result. It would mean vesicle-specific reference atlases are unavoidable.

Model

For protein $j$ in a bulk vesicle sample, the expected intensity is

$$\mu_j = \sum_{k} c_k \,\rho_{jk}\, s_{jk} + c_0\, n_j, \qquad \log \rho_{jk} = f_\theta(z_j) + b_k + \varepsilon_{jk},$$

where the symbols are:

  • $c_k$ is the fraction contributed by cell type $k$.
  • $s_{jk}$ is the whole-cell abundance of protein $j$ in that cell type.
  • $\rho_{jk}$ is its packaging factor.
  • $z_j$ is a vector of protein features.
  • $f_\theta$ is a small shared network.
  • $b_k$ is a cell-type loading offset.
  • $\varepsilon_{jk}$ is a residual with a learned variance.
  • $n_j$ is a measured profile of co-isolated non-vesicular plasma protein, taken from the late, vesicle-depleted size-exclusion fractions. Albumin and apolipoproteins dominate it.
  • $c_0$ is the weight on $n_j$.

Observed proteins enter through a log-normal density around $\mu_j$. A protein not detected in a run enters through the probability that its log intensity falls below that run's detection threshold $\tau$, which we estimate from the run's lowest confidently quantified precursors. The fractions get uncertainty intervals from the posterior. Those intervals are what define a minimum detectable fraction.

Research Program

Paired reference panel. We first build a panel of cultured cell types covering epithelial, fibroblast, monocyte-derived, lymphoid, erythroid/megakaryocytic-like and neuronal-like lines. For each line we collect paired measurements: a whole-cell data-independent acquisition (DIA) proteome and a vesicle proteome. DIA is a mass-spectrometry mode that fragments everything in wide mass windows rather than picking the strongest ions. Vesicles are isolated by size-exclusion chromatography and verified with the standard triad of particle tracking, electron microscopy, and CD9/CD63/CD81 positivity. The paired data train $f_\theta$.

Quantification pipeline. All runs are searched with DIAmeter, which matches theoretical peptides to multiplexed DIA spectra and finds low-abundance peptides that lack a visible precursor signal. Library-free Dear-DIA XMBD serves as a sensitivity check, since it reported 12.3 times more unique peptides than DIA-Umpire on the HYE124 benchmark. These tools supply the method. They resolve peptides, not cells, and we use them as the front end. Precursor-level evidence is kept so that $\tau$ can be estimated per run.

Isolation dependence. Co-isolated contaminants and vesicle subsets differ by isolation chemistry. We therefore repeat a subset of the panel with TiO$_2$ affinity, as in the rectal cancer sEV study, and with EVtrap beads, as in the Parkinson's urinary EV study. We then ask whether one packaging map holds across isolation methods or needs an isolation-specific offset.

Motivating observation. The clinical motivation comes from a pattern in published panels. The LUAD plasma-exosome panel of Feng et al. includes ITGA2B and HBD, which are platelet and erythroid proteins. The urinary study found about 8.9% brain-elevated proteins with no way to say what cells shed them. Both suggest that "biomarker" signals may partly be shifts in which cells contribute vesicles. That is exactly what a cell-of-origin estimate would expose.

Baselines

Naive reference points. There are two, and both use the same DIAmeter quantification and the same reference panel as everything else. The first is a marker score: the summed intensity of proteins enriched in each cell type, normalized across types. This is the logic behind counting "brain-elevated" proteins in the urinary study. The second is non-negative least squares (NNLS) against whole-cell proteomes. NNLS here means fitting the bulk profile as a non-negative weighted sum of reference profiles. Using whole-cell references amounts to assuming vesicles sample their parent cell uniformly. Beating these baselines is the minimum bar.

Strongest existing method. This is scpDeconv, run with whole-cell references from the same panel and with its own pseudo-bulk mixup, autoencoder imputation and adversarial platform correction. It is the only method in the file designed for continuous, missing-value-heavy proteomic mixtures. It reports outperforming Scaden, MuSiC, BayesPrism and DestVI in Lin's concordance correlation and RMSE on the Human Breast Atlas, but no reported result was found on Lacuna for its actual concordance or RMSE values. We will therefore measure it ourselves. It gets the same hyperparameter budget as our model, tuned on the same validation mixtures and never on test mixtures. Beating it means higher concordance and lower RMSE on held-out cell types at spike fractions of 5% and below, with non-overlapping bootstrap intervals across replicate mixtures.

Oracle upper bound. This is NNLS, or our likelihood without the packaging map, given the measured vesicle proteome of every cell type, including the held-out one. It is the real competitor wherever vesicle references can be made, and it sets the ceiling.

Ablations. Two ablations isolate the two ingredients. The first removes the packaging map ($\rho=1$) while keeping censoring. The second replaces censoring with scpDeconv-style imputation while keeping the packaging map. The proposal succeeds if the full model recovers most of the gap between whole-cell scpDeconv and the oracle on held-out cell types.

Why not just…?

Why not just build a vesicle reference for every cell type and run NNLS? This is the most serious threat. Wherever cultured vesicle references exist, it may match our model, and the oracle baseline is there to measure exactly that. The argument for the packaging map rests on cell types that cannot be cultured or whose disease states change their loading, which is where references will not exist. The leave-one-cell-type-out experiment settles this directly. If the oracle wins comfortably and the transferred map closes little of the gap, we will report that vesicle atlases are required. We would then narrow the contribution to the censored likelihood and the non-vesicular plasma component.

Why not just immunocapture vesicles carrying a cell-specific surface marker? The file argues that sequential pulldowns lose material at each wash, which matters most for already trace-level sources. It also notes that non-vesicular particles adsorb to capture matrices (gastric cancer review), and that plastic cells such as stem-like resistant breast cancer cells (tamoxifen-resistance study) can lose lineage markers. None of these points has been quantified here. We therefore treat capture as complementary. On a subset of mixtures we perform marker capture of the spiked line and compare its recovered yield and purity against our computational estimate at each spike level.

Why not just deconvolve vesicle RNA, where the tools are mature? Shi et al. show RNA deconvolution works for a tumor/non-tumor split. However, scpDeconv's authors stress the weak correlation between mRNA and protein. RNA also cannot see the phosphoproteome, and the urinary Parkinson's panels, with a reported AUC of 94.3% for disease, draw on phosphoproteins. RNA methods such as Count Bridges and POPPY are built for counts or transcript embeddings and do not take continuous, censored intensities. We do not run an RNA arm, so this comparison remains an argument rather than a result.

Why not just impute missing proteins and use any good deconvolver? Imputation fills in exactly the low-abundance values that carry the rare-source signal. When a protein is missing because it is rare, imputation tends to pull it toward the values of abundant sources. The imputation-versus-censoring ablation, evaluated as a function of spike fraction, tests this claim. If imputation matches censoring at 0.1–1% spikes, the claim fails.

Proposed Experiments

Ground-truth mixtures. We spike vesicles from each held-out line into pooled healthy-donor plasma vesicles at protein-mass fractions of 30%, 10%, 3%, 1%, 0.3% and 0.1%, in triplicate. We also include a mouse line, borrowing the logic of the HYE124 species-mixture benchmark. Mouse-specific peptides then give a fraction estimate that does not depend on pipetting error or on our model.

Metrics. These are concordance correlation and RMSE of the estimated fractions, the coverage of 90% posterior intervals, and the minimum detectable fraction. The last is defined as the smallest spike at which the interval excludes zero while the estimate stays calibrated.

Decisive checks.

  1. Predicting a held-out line's vesicle proteome from its whole-cell proteome. We measure the across-protein correlation of $\log \rho$ against an identity map and against a cell-type-mean map. If the learned map is no better than the identity, the transfer premise fails.
  2. Leave-one-cell-type-out deconvolution against scpDeconv and the oracle.
  3. Comparing minimum detectable fraction with censoring versus with imputation.
  4. Testing whether a packaging map trained on size-exclusion isolates holds on TiO$_2$ and EVtrap isolates.

Clinical arm. This is secondary and only proceeds if we can obtain raw DIA data from an existing plasma or urine vesicle cohort. We would ask whether estimated platelet, erythroid and brain-like fractions differ between groups, and whether they explain part of the published panel signals. We would treat any classification gain as exploratory, given how small these cohorts are. The lung study has 78 participants.