Reading the Wound Edge's Clock Under Low Oxygen

When a sheet of cultured cells is scratched, the cells along the cut begin to crawl into the gap. Low oxygen (hypoxia) changes how fast they close it. The usual way to ask why is to grind up the whole dish at one moment and sequence its RNA, the cell's working copies of active genes. This has two problems. Most cells in a scratched dish sit far from the cut and never change behavior, so the edge program is diluted into noise. And a single moment cannot tell whether hypoxia delays the edge program or weakens it. We propose a cheap way around the first problem and a direct test of the second.

The trick is geometric. Scratch molds with different comb spacings give wells whose fraction of edge cells ranges from small to large. Under the same oxygen condition and at the same time after wounding, bulk RNA from each well is a mixture of an "edge" profile and a "core" profile in proportions that imaging can measure. Regressing expression on edge fraction across geometries recovers the edge-specific program at every time point without dissociating cells. We pair this with frame-by-frame tracking of cell motion in the same wells, which measures the backward-travelling coordination waves that precede directed migration. That lets us ask a sharp question: does hypoxia shift the edge transcriptional clock by the same factor that it shifts the mechanical clock?

Core Claim

Our hypothesis is that hypoxia slows collective wound closure mainly by stretching time, not by shrinking response. Under this hypothesis, the hypoxic edge program is roughly the normoxic one played more slowly. The slowing factor estimated from RNA should then match the slowing factor of the lag between mechanical waves and directional alignment. Propagating Waves of Directionality and Coordination Orchestrate Collective Cell Migration measured that lag as 60 ± 10 min in DA3 tumor cells and about 217 min in MDCK epithelial cells, with no gene expression measured.

The alternative outcomes are just as informative. Matched delays with reduced amplitude would point to an energy limit on the response rather than a timing change. A transcriptional delay that does not track the mechanical one would mean hypoxia decouples the two layers. Either way, the result is a falsifiable, quantitative link between oxygen, gene-expression timing, and tissue mechanics that no existing study supplies.

Research Program

Cells and oxygen conditions. We use MDCK cells, for the epithelial wave lag above, and A549 lung cancer cells. A549 was the line used in the substrate-stiffness scratch study, The effect of stiff substrates on the collective migration of A549 cells, and in FLeCS drug-response trajectories. The primary oxygen arms are normoxia, continuous 1% O₂, and DMOG, a drug that stabilizes the hypoxia-inducible factor (HIF) in normal air. The DMOG arm separates the HIF-dependent component from everything else that low oxygen does.

Intermittent hypoxia cycles form a secondary arm. They are motivated by Chronic intermittent hypoxia reshapes circadian metabolic architecture in a model of sleep apnea, which showed that cyclic hypoxia rewires the timing of gene expression in liver rather than its average level.

Controlling the cell culture. Cells are pre-equilibrated in each oxygen condition for 24 h before scratching, so the measured response reflects wounding under hypoxia and not the onset of hypoxia itself. Mitomycin C blocks division in the main arms, with a parallel untreated arm. This addresses the known confound that gap closure past about 12 h mixes migration with proliferation. Medium height is fixed, and pericellular oxygen is logged with an optical sensor spot. This addresses the documented problem of unphysiological oxygen gradients in 2D culture.

Wounding and sampling. Scratches are guided by laser-cut molds of the kind introduced with An image J plugin for the high throughput image analysis of in vitro scratch wound healing assays, which roughly halves width and angle variability. We use three comb spacings plus a sham mold that touches the cells without cutting. Wells are imaged every 5 min. Closure is measured with that paper's Wound_healing_size_tool, and velocity fields are measured with the wave analysis of Zaritsky et al. RNA is harvested at 0, 0.5, 1, 2, 4, 8, 12, and 24 h with low-cost 3′-tag sequencing, using three replicates per cell. Sampling is dense early because the mechanical lags are on the scale of one to four hours.

Separating edge from core. For gene $g$ at time $t$, the bulk expression in a well with measured edge fraction $f$ is modeled as

$$y_g(t,f) = f\,e_g(t) + (1-f)\,c_g(t) + \varepsilon,$$

where $e_g$ and $c_g$ are the edge and core profiles and $\varepsilon$ is replicate noise. The edge fraction $f$ is the share of nuclei within a fixed distance of the cut, counted from the image taken just before lysis. The model is fit by non-negative least squares with gene-wise shrinkage. A version with two or three distance bands replaces the single edge/core split if the data support it.

Testing for a time-stretch. We compare the hypoxic edge program to the normoxic one using

$$e^{H}_g(t) \approx a_g\, e^{N}_g(t/\tau),$$

where $e^{H}_g$ and $e^{N}_g$ are the hypoxic and normoxic edge profiles. Here $\tau$ is a single shared time-stretch factor and $a_g$ is a per-gene amplitude. This is fit on immediate-early genes, cytoskeletal genes, and genes that respond to the substrate-stiffness pathway. The central comparison is $\tau$ against the ratio of hypoxic to normoxic wave-to-alignment lags measured in the same wells.

Baselines

Naive reference. The do-nothing reference is the sham-molded, unscratched monolayer at each matched time and oxygen condition. It removes handling and hypoxia effects that have nothing to do with wounding.

Standard practice. The standard practice baseline is single-geometry bulk RNA-seq at one endpoint (24 h) analyzed by DESeq2 differential expression. This is the design used in, for example, Palmitic acid activates GPRs/KLF7/HEY1 signal axis to promote the occurrence and development of endometrial carcinoma. It is run on the same libraries by subsetting, so data and cost are matched by construction.

Ablation of the geometric trick. A second baseline isolates our key ingredient. It uses the full time course from one geometry only, with no edge/core separation. If its edge-associated gene signals match ours, varying the geometry adds nothing.

Strongest existing transcriptome-to-phenotype method. The strongest existing method for linking longitudinal RNA to a phenotype is the phase-stratified elastic net of Interpretable transcriptome-to-phenotype modeling of cell-painting nuclear morphology features from RNA-seq under low-dose radiation exposure. We run it on our data with time-blocked cross-validation (whole time points held out), predicting closure rate and wave lag from bulk RNA. No per-feature numerical accuracy for that method was found on Lacuna; the summary reports only that five of nine morphology features had robust predictors. We will therefore report its accuracy under our own protocol and count our method as better if the separated edge profiles predict held-out kinematic readouts with lower error.

Dynamic model of gene regulation. For dynamics we fit FLeCS (A scalable gene network model of regulatory dynamics in single cells, code at Bertinus/FLeCS) to the separated edge trajectories with the same hyperparameter budget given to the time-stretch model. FLeCS matched known biology in 11 of 12 K562 knockdowns, but it assumes linear regulator inputs. We use it to ask whether a learned network model explains the hypoxic edge trajectory better than a single time-stretch does, scored by held-out time-point error.

Kinematic reference values. The reference values for the mechanical side are Zaritsky's lags and the A549 speed of 10.5 µm/h on glass reported by Sasaki et al. We expect our normoxic controls to reproduce these before any hypoxic comparison is trusted.

Why not just…?

Why not just run single-cell RNA-seq on scratched wells? This is the most serious alternative. Dissociating cells destroys the distance-from-edge coordinate, which then has to be inferred. Enzymatic dissociation also adds stress transcripts that overlap with the early wound response. That overlap matters because the burn-wound chromatin atlas (A scATAC-seq atlas of stasis zone in rat skin burn injury wound process) found early-phase activity dominated by Fos/Jun and Stat signals, which are exactly the genes dissociation induces. We therefore run single-cell sequencing at 1 h and 8 h as an independent check. If clusters with edge-like cell shapes reproduce our separated edge profiles at similar cost per time point, the single-cell route is a legitimate competitor and we will say so.

Why not just physically harvest the edge strip? Scraping away the core and lysing only the margin is direct. However, harvesting takes minutes, which blurs the 30-minute resolution the early lags require, and the band width is imprecise. We do this at two time points to validate the separation. If physically harvested strips and separated profiles disagree systematically, the mixture model is wrong.

Why not just use one endpoint, or a HIF drug instead of a hypoxia chamber? An endpoint cannot distinguish a delayed response from a weakened one, and that distinction is the whole question. The endpoint baseline is included precisely to show what it misses. The concern that DMOG behaves differently from real low oxygen is an assumption, not a result, so we include DMOG as an arm. If DMOG reproduces $\tau$, the effect runs through HIF. If it does not, the delay comes from non-HIF effects of oxygen, such as energy supply.

Why not just measure migration speed and skip RNA entirely? Imaging alone can show that hypoxia slows closure. It cannot say whether the slowdown arises in gene programs or in mechanics. The comparison of $\tau$ with the lag ratio needs both measurements in the same wells.

Decisive Checks

The separation is credible only if a model fit on two geometries predicts the bulk expression of the third geometry better than the single-geometry ablation does. It must also agree with the physically harvested strips and single-cell clusters on the leading edge genes.

The biological claim is supported if the 95% confidence interval for $\tau$ overlaps the measured lag ratio in both cell lines, and if amplitudes $a_g$ stay near one for cytoskeletal genes. It is refuted if $\tau$ is near one while the wave lag lengthens, or if amplitudes fall sharply without any delay. Comparing the DMOG and intermittent arms then shows whether the coupling runs through HIF and whether oxygen cycling, rather than its average level, sets the clock.