Separating Isolation Artifacts from Biology in Serial Extracellular Vesicle Measurements

Extracellular vesicles (EVs) are tiny membrane packets that cells release into blood. Their protein and RNA cargo is a promising "liquid biopsy," a way to monitor disease from a blood draw instead of a tissue sample. The catch is that purifying EVs from plasma is itself a noisy measurement. Different isolation runs and methods recover very different amounts of vesicles, and they carry over different amounts of non-vesicle plasma protein. When a patient is sampled repeatedly over months, a rise in an EV protein could be disease progression, or it could be a dirtier or higher-yield preparation that week. Current longitudinal models and their explanations cannot tell these two apart.

We propose to learn, from deliberate replicate experiments, how isolation quality distorts EV measurements. That learned distortion is then used to split every between-visit change in a patient's EV marker into a technical part and a biological part, each with an uncertainty interval. The explanation shown to a clinician is this split itself ("of the 2.1-fold rise in ADAM17, about 1.4-fold is explained by lower purity in visit 3; the biological change is 1.5-fold, 90% interval 1.1–2.0"). It is not a post-hoc heatmap or a generated narrative. The core claim is that the split becomes identifiable once technical replicates fix the biology. It cannot be identified from patient data alone, because isolation-quality readouts are themselves correlated with disease.

Why this gap is real

Three bodies of work each stop short of the problem. The EV isolation literature documents the instability but does not model its effect on time series. The microfluidics review Microfluidic Technology for the Isolation and Analysis of Exosomes reports filtration yields 4 to 1000 times higher than ultracentrifugation and acoustic devices at 82% recovery and 98% purity, and it concedes that balancing purity, recovery and throughput remains unsolved. Biomanufacturing reviews (Advancements in extracellular vesicles biomanufacturing; Exosomes as drug delivery systems in glioma immunotherapy) describe persistent batch-to-batch heterogeneity. These papers motivate the proposal by showing that technical variance is large. None of them asks what that variance does to a slope.

Clinical EV biomarker studies are single-snapshot. They motivate the application but leave the time dimension open. Proteomic profiling of plasma exosomes from patients with B-cell acute lymphoblastic leukemia isolates exosomes by ultracentrifugation and reports diagnostic AUCs of 0.989 for ADAM17 and 0.956 for ATG3. Its cohorts are small (10 in discovery, 28 in verification), and it acknowledges that ultracentrifugation co-isolates albumin, immunoglobulins and lipoproteins. The Lewy body versus Alzheimer's abstract 369 Differentiating Dementia with Lewy Bodies and Alzheimer's Disease Using Extracellular Vesicles uses brain-derived plasma EVs from single antemortem draws and reports no numbers yet.

Longitudinal clinical models supply methods but assume stable sensors. SeqRisk combines a Gaussian-process latent trajectory model with a Transformer and a survival head on sparse ICU data. PathwayLLM adds language explanations over learned clinical pathway graphs. Both treat every measured change as patient physiology. PathwayLLM's own failure modes, hallucinated lab values in about 6% of cases and over-attribution of nonspecific signals to the target disease, also constrain our design. Our explanations will be numeric decompositions computed from the model, not generated text, so they cannot invent values.

Harmonization work (ComBat for amyloid PET; ComBat in ENIGMA; Fed-ComBat; Batch-Free Encoding) removes shifts tied to discrete site or batch labels. EV isolation quality varies continuously within a batch and from sample to sample, which these methods do not address.

Core Claim

We write the measured log abundance of marker $k$ for patient $i$ at visit $t$ as

$$ y_{itk} = b_{ik}(t) + g_k(q_{it}) + \varepsilon_{itk}, $$

The terms are defined as follows. $b_{ik}(t)$ is the patient's smooth biological trajectory, given a Gaussian-process prior as in SeqRisk. $q_{it}$ is the vector of quality-control readouts recorded for that preparation: isolation method, particle concentration and size mode from nanoparticle tracking, the tetraspanins CD9, CD63 and CD81, the vesicle marker TSG101, the contamination markers calnexin and albumin, and a lipoprotein marker. $g_k$ is the technical response function, the shift that preparation quality induces in marker $k$. $\varepsilon_{itk}$ is assay noise.

The explanation for a visit-to-visit change is then

$$ \Delta y_{ik} = \Delta b_{ik} + \Delta g_k(q), $$

reported with posterior intervals on both terms.

From patient data alone, $g_k$ and $b_{ik}$ are confounded. Particle concentration and albumin carry real prognostic signal of their own: albumin change predicts mortality in Association between albumin changes and prognosis in older patients with acute myocardial infarction. Regressing them out therefore removes biology along with artifact. Our key ingredient is a replicate arm, in which identical plasma aliquots are processed many times under deliberately varied conditions. Because $b$ is fixed by construction in this arm, $g_k$ is learned there, frozen, and then applied to patients.

Research Program

The first stage is a calibration experiment. Pooled plasma from a small number of donors is split into aliquots. Some aliquots receive cell-line EVs at known ratios (1×, 1.5×, 2×, 4×), which creates known "biological" changes. Aliquots are processed on separate days by ultracentrifugation, size-exclusion chromatography, polymer precipitation and a filtration microfluidic device. Run conditions are varied within each method, including spin time, column loading and input volume, to span a realistic range of purity and yield. Every preparation receives the full quality-control panel and the target-marker readout. The target markers are mass spectrometry proteins, including ADAM17 and ATG3 as a concrete case. We fit $g_k$ as a monotone additive model in $q$, with method-specific offsets, and estimate a held-out-run error for it.

The second stage is a semi-synthetic longitudinal benchmark built on PhysioNet Challenge 2012, where SeqRisk was evaluated. We treat a block of lab features as stand-ins for EV markers. We then corrupt each visit with technical shifts drawn from the empirical joint distribution of $(q, g(q))$ measured in stage one, and in some regimes we let $q$ correlate with patient outcome to reproduce the confounding. Because the clean data are retained, we can measure both how much downstream survival prediction is damaged and how much each correction recovers.

The third stage is a pilot on banked serial plasma from a disease with a known EV-cargo hypothesis. B-ALL during induction is the natural choice given the ADAM17 and ATG3 results. In this setting no ground truth exists, so we test internal consistency instead. Technical splits of the same draw should yield the same biological estimate, and biological estimates should track independent clinical response markers better than raw values do.

Baselines

The naive reference is raw EV marker values fed directly into the same longitudinal model. This is what existing EV studies would do if extended to serial draws. Any method that cannot beat it on the calibration data is useless.

The most common practical fix is normalizing each marker to particle count or to CD63 signal. It is the strongest cheap competitor because it already uses part of the quality-control vector. We run it on identical data with the same downstream model. It counts as beaten if, on the spike-in arm, our biological-change estimates have lower absolute error in log fold change. We expect normalization to fail specifically when contamination rather than yield varies.

ComBat with run labels (Harmonization of amyloid PET radiotracers using ComBat) is the established harmonizer. A reference-pool bridge, which normalizes every run to a common pooled sample processed alongside it, is the standard proteomics alternative. We give both the same run labels and the same bridge samples our method sees. For these baselines, beating means lower error on within-run purity variation, since both handle between-run offsets well by design.

The strongest existing longitudinal survival model is SeqRisk. It reports a C-index of 0.789 and an integrated Brier score of 0.135 on PhysioNet 2012 (SeqRisk). We rerun SeqRisk ourselves on clean, corrupted and corrected versions of our semi-synthetic benchmark with an equal hyperparameter search budget, and use our clean-data rerun as the ceiling. Success means closing most of the gap between corrupted and clean C-index while not falling below the clean figure in the regime where $q$ correlates with outcome. Naive regression-out should fall below it there.

For the explanation baseline, we include $q$ as extra features in a Random Survival Forest and read SHAP attributions, following Predicting Progression from Mild Cognitive Impairment to Alzheimer's Disease using Survival Analysis. No reported result exists on Lacuna for technical-versus-biological attribution of EV changes, so the comparison is made only on our spike-in ground truth.

The key ablation fits $g_k$ jointly from patient data without the replicate arm. If it matches the full method, the replicate arm is unnecessary and the core claim fails.

Why not just…?

Why not just track whole-plasma markers and skip EV isolation? For many diseases this is the right answer. Total plasma p-Tau217 alone reaches 4-year AUROC 0.787 in the ADNI survival study, with no vesicle work. Our proposal does not argue that EVs beat plasma. It argues that wherever EV cargo is used, for example brain-derived α-synuclein seeding as in the DLB/AD abstract, serial EV values need correction. In the pilot we report whole-plasma counterparts of the target proteins next to the EV estimates, so readers can see whether the EV channel adds anything at all.

Why not just normalize to particle count or CD63? This is the most serious threat, and the spike-in arm settles it directly. If normalization recovers known fold changes as accurately as our model, we would drop the learned response function and recommend the simpler normalization plus uncertainty propagation. We expect it to fail when albumin and lipoprotein carryover vary independently of yield, because a single scalar normalizer cannot absorb two independent distortions.

Why not just run ComBat or a reference pool per run? Both correct offsets between batches, and we will use reference pools ourselves as covariates. Neither sees differences in preparation quality between two samples in the same run. Their known weakness on longitudinal data is that they can absorb real trajectory change when batch aligns with time (Latent mixed-effect models for high-dimensional longitudinal data). The semi-synthetic regime where run order correlates with disease progression tests this directly.

Why not just add the quality-control readouts as features and explain with SHAP? SHAP will credit albumin when albumin predicts the outcome, whether that albumin is carryover or a real prognostic signal. The attribution therefore cannot answer the clinician's actual question, which is whether the patient changed. Our decomposition answers that question because $g$ is learned where biology is held fixed. The test is whether SHAP's share for the quality-control features tracks the true technical share on spike-in data; we expect it not to.

Proposed Experiments

The decisive check is on the stage-one spike-in arm, using held-out runs. For each pair of preparations we compute the error in estimated biological log fold change, the 90% interval coverage, and the false "changed" call rate when the true fold change is 1×. A false "changed" call means the method reports a change that did not happen. A successful method gets coverage near nominal and halves the false-change rate of particle normalization. A second check is the semi-synthetic PhysioNet recovery of C-index and Brier score under confounded corruption, compared against the SeqRisk reruns. The third check is the pilot consistency test: technical splits of the same B-ALL draw must agree within their intervals.

Following the finding in Algorithmic transparency and interpretability measures improve radiologists' performance in BI-RADS 4 classification that numeric confidence, not heatmaps, drives clinician decisions, the clinician-facing output is the interval-bearing decomposition. A small reader study on spike-in cases would compare how often haematologists correctly call "real change" from raw trends versus the decomposition. If the ablation without the replicate arm performs as well as the full model, or particle normalization matches it on the spike-in data, we report that and recommend the simpler method.