Unmasking Shear-Responsive Cell States in Bulk RNA from Flow Experiments

When researchers simulate blood flow over cells in the lab — using parallel-plate flow chambers or microfluidic channels that impose a known frictional drag, called wall shear stress, on a cell layer — they almost always measure the biological response by grinding up the whole layer and running a bulk assay: one averaged number per gene per condition. This averaging, "heterogeneity masking," hides exactly the thing that matters most in vascular disease: rare transitional cell states. Single-cell sequencing of atherosclerotic tissue has shown, for instance, that smooth muscle cells silently switch into stem-like and macrophage-like states, and that endothelial cells lose their protective flow-response pathway well before plaques form — findings invisible at bulk resolution. Meanwhile, the obvious fix, single-cell sequencing of flow-chamber cells, is compromised because the warm enzymatic digestion needed to free the cells from the chamber itself distorts the transcriptome, destroying the very mechanosensitive states under study.

We propose to close this loop with a computational instrument rather than a new wet-lab assay: a deconvolution model — a method that mathematically unmixed an averaged bulk measurement into underlying single-cell populations and states — whose internal assumptions are conditioned on the measured mechanical environment. Shear stress simulation supplies the physical covariates (shear magnitude, pulsatility, exposure time, gradient); heterogeneity unmasking supplies the inference target. No existing deconvolution framework incorporates mechanical priors of any kind, and no shear-simulation study has ever been paired with modern deconvolution. The payoff is mechanical dose–response curves for rare cell states at bulk-sequencing cost, without dissociation artifacts.

Core Claim

Standard deconvolution models fail on flow-experiment data for a specific, testable reason: when shear stress changes a cell's state without changing its identity, a static model has no vocabulary for that shift and absorbs it into the estimated cell proportions, inventing cells that are not there and missing the state change that is. We hypothesize that conditioning the model's prior on the imposed hemodynamic regime — an explicit function from flow covariates to the expected state distribution of each cell type — separates "which cells are present" from "what state each cell type is in." Formally, in a guided topic model where each bulk sample is a mixture of per-cell-type state distributions, we replace the fixed Dirichlet prior over mixture weights with a shear-conditioned one:

$$\pi_b \sim \mathrm{Dir}\big(\alpha = g_\theta(s_b)\big), \quad \text{where } s_b = (\tau_{\text{mean}}, \tau_{\text{osc}}, t, \nabla\tau)$$

for bulk sample $b$ exposed to mean shear magnitude $\tau_{\text{mean}}$, oscillatory component $\tau_{\text{osc}}$, duration $t$, and shear gradient $\nabla\tau$, with $g_\theta$ a small network predicting prior concentration. The same conditioning also enters the per-cell-type topic distributions, so the model learns, for example, the shear dose at which an endothelial topic expressing KLF2 gives way to a mesenchymal-transition topic. This is a direct extension of GTM-decon's multi-topic design GTM-decon, chosen because its topic-per-cell-type structure already represents within-type state variance; the new ingredient is entirely the mechanical conditioning.

Research Program

The program has three parts. First, we build a calibrated in vitro shear–bulk transcriptome matrix: parallel-plate flow chambers impose steady high laminar shear, low disturbed/oscillatory shear, and static control on endothelial monolayers and on the endothelial–smooth-muscle co-culture geometry of the DKK1 foam-cell system parallel-plate co-culture flow system, with bulk RNA-seq (cold lysis, no dissociation) across a graded matrix of shear magnitudes, waveforms, and timepoints. This follows the established finding that flow regime alone — steady versus pulsatile — remodels endothelial adhesion-molecule expression in a vascular-bed-dependent way Vascular bed origin dictates flow pattern regulation of endothelial adhesion molecule expression, so shear conditioning is demonstrably not redundant with cell-type identity. Second, we construct a small single-cell reference atlas from a few anchor flow regimes, accepting dissociation artifacts only in the reference (and quantifying them by the warm-digestion stress signature documented in the atherosclerosis single-cell literature Single-cell RNA sequencing in atherosclerosis: Mechanism and precision medicine). The economics are the point: one reference amortizes over the full bulk condition matrix, whereas per-condition single-cell runs would multiply cost and artifacts together. Third, we train the shear-conditioned deconvolution model and validate it three ways, including a gradient extension using the stenosis microfluidic geometry in which endothelial cells traverse steep shear gradients microfluidic geometry incorporating a significant stenosis, with spatially segregated lysis so that position-specific shear (from analytic channel formulas or calibrated CFD) can be linked to position-specific state.

A biological anchor keeps the work honest: the DKK1 co-culture study showed shear-induced endothelial secretion of DKK1 driving smooth-muscle foam-cell formation, but measured everything in bulk and could not say whether a rare endothelial subpopulation secretes DKK1 or what fraction of smooth muscle cells actually transitions Endothelial Dickkopf-1 Promotes Smooth Muscle Cell-derived Foam Cell Formation via USP53-mediated Deubiquitination of SR-A During Atherosclerosis. Our model should answer exactly that question, and its answer is checkable by targeted staining. We also test whether the deconvolved endothelial state recovers the hub genes (ICAM1, KLF2, VCAM1) that the observational single-cell study of coronary plaques identified as shear-response hubs Single-cell RNA-seq uncovers distinct pathways and genes in endothelial cells during atherosclerosis progression.

Proposed Experiments

Evaluation uses four tests, in increasing order of biological realism. First, in silico mixtures: we generate synthetic bulk profiles from the anchor single-cell reference at known proportions and known (simulated) flow covariates, following the in-silico-bulk recipe of the Kassandra database (2,286 simulated bulk profiles from sorted populations, used to evaluate Cross-modal transfer learning for mapping bulk transcriptomes at cellular level); metrics are Spearman correlation and RMSE on recovered proportions, the metrics used in the GTM-decon benchmarks GTM-decon. The critical manipulation is a misattribution probe: we construct mixtures where only the state (not the proportion) changes with shear; a correct model reports flat proportions and shifting states, a static model reports shifting proportions. Second, spike-in mixtures of two vascular cell lines at known ratios under flow, giving a ground truth with zero reference ambiguity. Third, held-out single-cell conditions: we hold out one or two flow regimes, run actual scRNA-seq on them, and ask whether the model's bulk-only predictions of state proportions match the held-out single-cell data — restricting the comparison to genes not induced by warm dissociation, to avoid grading against an artifact-confounded key. Fourth, the DKK1 and stenosis-gradient anchors above. For transferability, we also apply the trained model to public bulk vascular datasets if the condition covariates can be reconstructed, though the primary evidence is in vitro where the shear is known exactly. For reference-completeness stress-testing, we deliberately train with the mesenchymal-transition topic removed and measure how conditioning helps the model flag an out-of-atlas state rather than misattribute it, since reference-atlas dependence is a documented failure mode of all three modern deconvolution families Cross-modal transfer learning for mapping bulk transcriptomes at cellular level. No reported result exists on Lacuna for any deconvolution method applied to flow-chamber bulk data, so all numbers reported below from prior work come from other domains and serve as tuning targets, not head-to-head comparators.

Baselines

The do-nothing reference point is bulk differential expression between flow conditions, which is precisely how the existing shear-simulation literature reads out biology — bulk Western blots and qPCR in the flow-chamber studies in vitro flow chamber simulation and the DKK1 co-culture work Endothelial Dickkopf-1 Promotes Smooth Muscle Cell-derived Foam Cell Formation via USP53-mediated Deubiquitination of SR-A During Atherosclerosis. It will be run on the identical bulk matrix with identical sequencing depth; beating it means showing that at least one bulk signature it would read as a uniform shift is in fact a rare-subpopulation phenomenon, verified by staining or held-out single-cell data. The strongest existing competitor is unconditioned GTM-decon, which currently reports the best deconvolution number in this literature — 97% accuracy distinguishing breast cancer subtypes (Basal versus ER+) from phenotypic topic scores on TCGA BRCA GTM-decon — and whose multi-topic structure is the closest existing design to ours. It will be given the same bulk data, the same reference atlas, the same compute, and a matched hyperparameter search budget, including its nested phenotype-guided mode so it can exploit any phenotype structure in the flow labels. Beating it means higher proportion-recovery correlation and lower RMSE on the in silico and spike-in mixtures, and, more decisively, correct flat-proportion behavior on the misattribution probe where the unconditioned model is expected to hallucinate proportion shifts. We will also run CIBERSORTx and MuSiC as classical linear references — no reported result on Lacuna for either on vascular flow data — under the same protocol, expecting them to fail hardest because their fixed marker assumptions break precisely when flow drives marker drift. Finally, the key ablation is our own architecture with the shear conditioning disabled: identical topic model, identical training data, but the covariate vector $s_b$ replaced by a constant. If this ablation matches the conditioned model, the central claim is wrong and the paper becomes a negative result about mechanical priors, which we would report as such.

Why not just…?

Why not just run single-cell RNA-seq on the sheared cells directly? This is the most serious alternative, and we do use single-cell sequencing — but only for the anchor reference and held-out validation, because per-condition single-cell runs multiply two documented costs: the warm enzymatic dissociation at 37°C that induces stress-gene artifacts and destroys fragile states Single-cell RNA sequencing in atherosclerosis: Mechanism and precision medicine, and the cost and dropouts that make a multi-waveform, multi-timepoint shear matrix impractical at single-cell resolution. The design is built so this objection settles itself: held-out single-cell conditions directly grade the deconvolution, restricted to dissociation-insensitive genes. If the deconvolution cannot track held-out single-cell truth, the amortization argument collapses and we will say so.

Why not just apply off-the-shelf GTM-decon or Count Bridges to the new bulk data? That is exactly the unconditioned baseline and ablation, so the question is answered by the main experiment rather than by argument. The concrete reason to expect it to fall short is documented on both sides: flow regime alone rewrites endothelial adhesion-molecule expression in a bed-dependent manner Vascular bed origin dictates flow pattern regulation of endothelial adhesion molecule expression, and marker instability under perturbation is a known failure mode of marker-based unmixing GTM-decon. The misattribution probe is designed so a static model's failure is visible, not debatable: state-only changes must not move its proportion estimates. Count Bridges, which models integer sequencing counts with Poisson birth-death jumps and outperformed CIBERSORTx and MuSiC on proportion recovery (no numerical value reported on Lacuna) Count Bridges enable Modeling and Deconvolving Transcriptomic Data, is an alternative base model; if its integer-count geometry proves more robust than the topic model, we will port the shear conditioning onto it, since the conditioning is architecture-agnostic.

Why not just do conventional CFD plus bulk differential expression, as vascular labs already do? This pairing is the field's default and is adequate when the question is "does flow change average expression." It fails when the question is "which cells changed and into what state": bulk analysis of coronary plaques could not identify that endothelial cells specifically carry the weakest shear-response pathway until single-cell dissection resolved nine clusters Single-cell RNA-seq uncovers distinct pathways and genes in endothelial cells during atherosclerosis progression. On the physics side, CFD is computationally fine for our simple chamber geometries (analytic formulas suffice) but does not scale to deformable or patient-specific geometries, which is why surrogates like GReAT — geometric pre-training on 8,449 vessel meshes, significantly outperforming from-scratch training on the 49-patient IMPACT cohort at $p = 0.016$ — and WSSflow — eightfold resolution refinement and 64× vector density over standard particle image velocimetry, subject to the 1-pixel rule for resolving the near-wall sublayer Uncovering wall-shear stress dynamics from neural-network enhanced fluid flow measurements — exist GReAT: leveraging geometric artery data to improve wall shear stress assessment. Neither surrogate addresses cellular readout at all, which is the gap we fill; our proposal uses neither surrogate as a dependency, so their limitations (domain shift for GReAT, optical access for WSSflow) do not propagate to us.

Decisive Checks

The proposal stands or falls on three outcomes. The misattribution probe must show that shear conditioning converts state changes into state estimates rather than spurious proportion estimates. Held-out single-cell conditions must confirm the model's bulk-only predictions of state proportions within dissociation-insensitive gene sets. And at least one biological quantity previously unresolvable in bulk — the DKK1-secreting endothelial fraction, or the shear dose at which endothelial mesenchymal-transition markers emerge — must be predicted by the model and confirmed by an orthogonal assay. If all three hold, the result is a reusable instrument: any lab with a flow chamber and a bulk sequencer gains single-cell state resolution over its entire mechanical condition matrix.