When a drug kills a bacterium, the decisive moment is when its outer membrane breaks: once the lipid bilayer is breached, the cell has passed a point of no return. Yet the standard ways of measuring this treat it as a snapshot — count how many cells take up a membrane-impermeant fluorescent dye at a fixed endpoint, or measure how much protein and DNA has leaked into the surrounding medium at a fixed hour. Both flatten two things that matter: when each individual cell fails, and how it got there. In parallel, time-course expression measurements show that cells under stress reprogram their gene activity hour by hour, and theory suggests that the speed of that reprogramming, not the drug concentration alone, decides who dies — Kratz and Banerjee's coarse-grained model of proteome allocation concludes that bacterial death under antibiotic stress is governed mainly by how fast gene expression shifts from growth machinery to stress-response machinery Gene Expression Tradeoffs Determine Bacterial Survival and Adaptation to Antibiotic Stress. But nobody has connected these two threads statistically, because the two literatures never meet: expression studies treat time as an index rather than as a countdown to an event, and membrane assays report endpoint fractions rather than event times.
This proposal builds that connection. The idea is to measure membrane rupture as a single-cell time-to-event outcome — each cell in a time-lapse movie has its own rupture time, and cells still intact at the end of the movie are right-censored — and to model the instantaneous rupture hazard with a time-dependent Cox model, a survival regression where the covariates themselves change over time, driven by longitudinal expression measurements from matched cultures. The central scientific claim is that the velocity of stress-response gene-module reallocation predicts rupture hazard better than static expression levels do, and that this can only be seen once membrane failure is treated as a timed event rather than a 24-hour binary outcome.
Core Claim
Formally, for cell $i$ observed from drug exposure at $t=0$, we model the instantaneous hazard of membrane rupture as
$$h_i(t) = h_0(t)\exp\!\left(\beta^\top Z_i(t)\right),$$
where $h_0(t)$ is a baseline hazard left flexible (piecewise-constant, to accommodate the lag-then-burst rupture kinetics that violate the usual constant-hazard-ratio assumption), and $Z_i(t)$ is a vector of time-varying covariates derived from matched longitudinal expression data: for each pre-defined gene module, its smoothed expression level and its velocity (rate of change), estimated by Gaussian Process regression of the kind validated in the Bayesian time-reannotation work, which recovers biological time to within a 30-minute margin and keeps 97% of annotations within 0.5 hours even when 30% of the transcriptome is corrupted A supervised Bayesian method for time (re)annotation of transcriptomics data. The velocity feature deliberately operationalizes the volatility-as-risk idea from the joint longitudinal-survival literature, where a one-unit increase in biomarker trajectory volatility raised the hazard of death 1.2-fold in hemodialysis patients A Flexible Joint Longitudinal-Survival Model for Analyzing Longitudinally Sampled Biomarkers — but transplanted from one clinical protein to genome-scale module dynamics.
The claim has two parts, one methodological and one biological. Methodologically: per-cell rupture times plus time-varying module covariates make the expression-to-death link estimable in a way that neither endpoint assays nor descriptive time-kill curves allow. Biologically: the velocity terms carry predictive information that the level terms alone do not, which would be the first empirical, statistical confirmation of the response-lag hypothesis that Kratz and Banerjee established only by simulation.
Research Program
Three components are built and integrated.
The first is a single-cell rupture assay that fixes the two core defects of current membrane measurements. Present practice is either binary dye exclusion at a static endpoint, as in the flow-cytometric propidium iodide assays of Song et al. Quantifying Cell Membrane Integrity via Permeability and Leakage and the dual-stain live/dead imaging used against vancomycin-resistant enterococci and Candida She et al., or bulk leakage readouts such as optical density at 260/280 nm, which the cinnamaldehyde and emodin studies use to demonstrate membrane failure Antimicrobial and antivirulence function of cinnamaldehyde against <i>Streptococcus suis</i> type 2; The Mechanism of Emodin Against Methicillin-Resistant <i>Staphylococcus aureus</i> Infection and Research on Synergistic Antibiotics. Bulk leakage cannot distinguish uniform partial permeability across all cells from total lysis of a subpopulation, and endpoint counting cannot say when anything happened. We instead run cells in microfluidic chambers under time-lapse imaging with SYTO9 plus propidium iodide, extract a per-cell rupture time for thousands of cells per condition (the moment of dye entry, detected by the 20–30-fold fluorescence enhancement), and treat cells intact at the end of acquisition as right-censored. Multiplexed channels of a membrane-potential dye in the DiSC3(5) style used in the bunamidine study Repurposing bunamidine hydrochloride as a potent antimicrobial agent targeting vancomycin-resistant <i>Enterococcus</i> membranes provide a secondary, graded depolarization trace per cell, giving a time-varying biophysical covariate as well.
The second component is the expression layer, designed around the known bottleneck that transcriptomics profiles tens of thousands of genes over very few time points. We aggregate genes into a modest number of modules (ribosomal and growth machinery, oxidative-stress response, envelope repair, and so on, in the spirit of the sector structure Kratz and Banerjee simulate), then fit Gaussian Process trajectories per module over a densely sampled exposure window, using the reannotation machinery above to align replicate time courses whose growth dynamics drift. This mirrors the phase-aware strategy of the low-dose radiation study, which stratified a nine-week RNA-seq course into biological phases and used leave-one-week-out residual Elastic Net regression to isolate dose-independent predictors of imaging phenotypes, successfully predicting 5 of 9 nuclear morphology features Interpretable transcriptome-to-phenotype modeling of cell-painting nuclear morphology features from RNA-seq under low-dose radiation exposure. Our departure from that paper is twofold: the phenotype we predict is membrane rupture, not nuclear morphology, and the statistical engine is a censored time-to-event model, not ordinary least squares.
The third component is the survival engine: a time-dependent Cox model with structured penalties, borrowing the heredity-enforcing group penalties of sox — a method that accommodates time-varying covariates and guarantees that a velocity interaction is only selected when its parent module level is, eliminating the false interaction discoveries that plain Cox-LASSO produced Structured learning in time‐dependent Cox models — combined with FLASH-style association features (rolling means, extrema, signature transforms) to compress module trajectories into a covariate vector small enough for partial-likelihood estimation An efficient joint model for high dimensional longitudinal and survival data via generic association features. Two departures from the existing joint-modeling stack are essential. First, bounded and skewed signals — normalized fluorescence, percentage leakage, normalized counts — violate the Gaussian marker assumption that FLASH and its competitors impose, so the longitudinal submodel uses a logit-transformed or Beta likelihood; this is a named, unaddressed limitation of the whole joint-modeling line. Second, acute membrane failure is exactly the regime where the proportional hazards assumption breaks, so $h_0(t)$ is piecewise-constant with data-driven breakpoints and the module effects are allowed to vary across those pieces, rather than being forced constant as in standard Cox software.
Baselines
The naive reference point is what the field does now: a static endpoint assay — propidium-iodide positivity at a fixed incubation time, scored per cell and fitted with logistic regression on drug concentration, exactly the design implicit in the dual-staining studies Jarosiewicz et al.; She et al.. It will be run on the same microfluidic movies, scored at matched endpoint times, with the same replicate structure. Beating it means, concretely, producing calibrated per-cell rupture-time distributions whose time-dependent concordance index at landmark hours exceeds the endpoint model's discrimination, and predicting an outcome the endpoint model cannot express at all: the post-antibiotic effect, the delay of regrowth after drug removal.
The strongest existing competitor is FLASH, the joint longitudinal-survival framework that handles dozens of time-varying markers with censored outcomes and converged in minutes on the NASA Turbofan run-to-failure benchmark where standard joint software did not finish within 24 hours An efficient joint model for high dimensional longitudinal and survival data via generic association features. Applied to our module-level features, FLASH is a genuine threat: it already does longitudinal-markers-plus-survival. It will be run on exactly the same module trajectories, rupture times, and censoring indicators, with the same Elastic Net tuning grid searched by the same nested cross-validation and the same compute budget; no reported concordance index for FLASH's own benchmarks was found on Lacuna, so all comparison numbers will be generated by our own re-run under identical conditions rather than quoted. Beating FLASH means a higher time-dependent concordance index and lower integrated Brier score on held-out drug concentrations, with differences assessed by bootstrap over cells. A second strong competitor is a two-stage pipeline — Gaussian Process smoothing of module trajectories followed by a static DeepSurv-style Cox network Deep Survival: A Deep Cox Proportional Hazards Network — which represents the "deep survival on omics" state of practice; its reference performance on the METABRIC genomic benchmark is a concordance index of 0.643, versus 0.630 for linear Cox and 0.624 for random survival forests, so on any dataset where our engine does not reach parity with such static-network performance when covariates are frozen to baseline, the temporal machinery has failed a basic sanity check.
Two ablations isolate the proposal's key ingredients. The velocity ablation refits the full engine with all covariates frozen at their $t=0$ values, removing dynamics but keeping module selection and the flexible baseline; if velocity adds nothing, the core biological claim is wrong. The endpoint ablation replaces per-cell rupture times with the bulk leakage readout (optical density at 260/280 nm of matched culture supernatants, the exact assay used in the cinnamaldehyde and emodin studies) as the longitudinal marker, keeping everything else; this isolates whether single-cell resolution is doing the work or whether population-averaged signals suffice.
Why not just…?
Why not just keep the endpoint binary and use logistic regression? Because the binary endpoint destroys exactly the information this proposal is about. The file's own critique applies: population-averaged leakage cannot distinguish uniform partial permeability from catastrophic lysis of a subpopulation, and endpoint counting cannot order events in time. Concretely, cinnamaldehyde produces a post-antibiotic effect exceeding 7 hours Antimicrobial and antivirulence function of cinnamaldehyde against <i>Streptococcus suis</i> type 2 — a quantity that is literally a statement about the timing of events after drug removal and is unrepresentable in a fixed-endpoint logistic model. The decisive experiment is the same-movies comparison above: if the endpoint model matches our dynamic concordance and predicts the post-antibiotic effect equally well, the timing machinery is not earning its complexity and we would report that.
Why not just run DeepSurv on a baseline expression snapshot? This is the static-baseline bias that the survival-omics literature itself flags as a limitation, including in DeepSurv's own METABRIC analysis. More pointedly, Kratz and Banerjee's result says death is governed by the speed of expression reallocation, which a single snapshot cannot encode by construction Gene Expression Tradeoffs Determine Bacterial Survival and Adaptation to Antibiotic Stress. We take this seriously rather than arguing around it: the velocity ablation is precisely this model (same network, same tuning, baseline-frozen covariates), so the experiment itself will settle whether the temporal ingredient matters. If it does not, the honest conclusion is that module levels are sufficient markers of the stress state and the response-lag hypothesis fails empirically at the single-cell level, which would itself be a publishable negative result against the simulation literature.
Why not just use FLASH off the shelf? At module-level aggregation FLASH is a real threat and we treat it as the primary baseline rather than a strawman. It has two documented gaps that our design targets: its longitudinal marker component assumes continuous Gaussian distributions, which bounded fluorescence and percentage-leakage signals violate, and it was demonstrated on tens of sensors, not genome-derived trajectories coupled to a lag-then-burst failure process. Whether those gaps matter in practice is an empirical question, and two experiments answer it: the re-run of FLASH on identical module features settles the head-to-head, and a bounded-marker simulation with known ground truth (calibrated to observed rupture kinetics) settles whether the non-Gaussian likelihood and the non-proportional baseline are load-bearing or cosmetic. If FLASH wins the head-to-head, the contribution reduces to the assay and dataset, and we would say so.
Why not just simulate with Kratz and Banerjee's allocation model instead of collecting new data? Their model is a hypothesis generator, not an estimator: it represents damage thresholds through stochastic differential equations, uses no empirical RNA-seq, and never models membrane permeabilization as the biophysical failure mode. It supplies the prediction we test — velocity matters — but it cannot fit per-cell rupture data or produce hazard ratios. Our work is the empirical instrument for its central claim, and the simulation is retained as a generator of synthetic data for method validation where wet-lab ground truth is unavailable.
Proposed Experiments and Decisive Checks
The program runs five linked experiments. The first is a dry-run on the NASA Turbofan degradation benchmark (200 run-to-failure units, 17 time-varying sensors), an engineering analog of trajectory-driven failure where FLASH converged in minutes and competing joint packages failed within 24 hours An efficient joint model for high dimensional longitudinal and survival data via generic association features; we verify our engine converges at comparable cost and benchmark against our own FLASH re-run on time-dependent concordance, since no concordance number for this benchmark was reported on Lacuna. The second is the bounded-marker, lag-then-burst simulation, which tests the piecewise baseline and non-Gaussian likelihood against standard Cox under controlled proportional-hazards violation. The third is the main wet-lab experiment: Streptococcus suis type 2 exposed to cinnamaldehyde across a concentration ladder from sub-minimum-inhibitory-concentration to 4× MIC — the concentration at which pre-formed biofilm eradication exceeds 73%, with MIC 0.25 µg/mL and MBC 0.5 µg/mL as anchor points Antimicrobial and antivirulence function of cinnamaldehyde against <i>Streptococcus suis</i> type 2 — with microfluidic time-lapse rupture movies in parallel with dense expression sampling, fitted by leave-one-concentration-out cross-validation. The decisive external check is post-antibiotic-effect prediction: the model, fit only on rupture and expression data during exposure, must predict the >7-hour regrowth delay measured after drug removal in held-out conditions. The fourth repeats the design on methicillin-resistant S. aureus with emodin, where leakage, membrane collapse, and a single-interval proteomics snapshot already exist The Mechanism of Emodin Against Methicillin-Resistant Staphylococcus aureus Infection and Research on Synergistic Antibiotics, to test whether fitted hazard structure transfers across organism and compound. The fifth is the static sanity benchmark: on METABRIC (1,980 patients, baseline expression, right-censored survival), our engine with covariates frozen to baseline must match the 0.643 concordance that DeepSurv reports there Deep Survival: A Deep Cox Proportional Hazards Network.
The decisive checks, in order of importance, are these. The velocity ablation must show a measurable concordance gain over baseline-frozen covariates on held-out concentrations — this is the test of the core claim. The head-to-head against FLASH on identical features must be won, or the methodological contribution collapses to the assay. Post-antibiotic-effect prediction must succeed where no endpoint model can even compete. And a sensitivity analysis on censoring — simulating cell loss from the field of view during imaging — must show the hazard estimates are robust to the informative-censoring risk that single-cell movies introduce, since cells swept away mid-movie are not missing at random. If all four hold, the result is a reusable template: any lab already running dye-exclusion assays and expression time courses can, without new instrumentation beyond a time-lapse microscope, convert two standard experiments into a statistical account of when and why individual cells die.
References
- Gene Expression Tradeoffs Determine Bacterial Survival and Adaptation to Antibiotic Stress
- A supervised Bayesian method for time (re)annotation of transcriptomics data
- A Flexible Joint Longitudinal-Survival Model for Analyzing Longitudinally Sampled Biomarkers
- Copper Ions Facilitate the Conjugative Transfer of SXT/R391 Integrative and Conjugative Element Across Bacterial Genera
- Repurposing bunamidine hydrochloride as a potent antimicrobial agent targeting vancomycin-resistant Enterococcus membranes
- Antimicrobial and antivirulence function of cinnamaldehyde against Streptococcus suis type 2
- The Mechanism of Emodin Against Methicillin-Resistant Staphylococcus aureus Infection and Research on Synergistic Antibiotics
- Interpretable transcriptome-to-phenotype modeling of cell-painting nuclear morphology features from RNA-seq under low-dose radiation exposure
- Structured learning in time‐dependent Cox models
- An efficient joint model for high dimensional longitudinal and survival data via generic association features
- The Impact of Glyphosate, Its Metabolites and Impurities on Viability, ATP Level and Morphological changes in Human Peripheral Blood Mononuclear Cells
- Deep Survival: A Deep Cox Proportional Hazards Network