Disturbance Scheduling to Unmask Hidden Cells

When researchers measure a tissue or culture in bulk — pooling millions of cells and reading the average signal — the rare cells that matter most, such as stem-like clones that later drive drug resistance, simply disappear from the measurement. Separately, ecologists have long known a counterintuitive fact: communities left alone lose diversity because one dominant competitor monopolizes everything, while moderately disturbed communities stay diverse because each pulse of stress resets the competition before a monopoly forms. This proposal connects those two facts with a single move: treat disturbance not only as a way to keep a biological system healthy, but as a way to keep it legible. If a periodic, calibrated disturbance keeps dominant clones from swallowing the community, then the rare subpopulations that bulk sequencing normally masks rise above the detection floor — and the frequency of that disturbance becomes a control knob on how much information a cheap bulk assay can recover.

The core practical payoff is a scheduler: a procedure that chooses how often to pulse a culture (with low-dose drug, UV, or selective harvesting) so that a downstream deconvolution model — the software that tries to split a pooled bulk measurement back into its constituent cell types — recovers the true composition as accurately as possible, without collapsing productivity. No existing work treats disturbance frequency as a variable tuned for measurement quality rather than yield.

Core Claim

The claim is that deconvolution error from bulk data is a unimodal function of disturbance frequency. At zero disturbance, competitive exclusion drives one clone past roughly 80–90% of the community, entering what we will call the purity trap — the regime where Tran et al. showed, across 38,000 simulated mixtures from a 100,064-cell breast cancer atlas, that every standard deconvolution method systematically misclassifies minority cells as the dominant type. At very high disturbance frequency, the community fragments and, crucially, cell-state expression signatures drift away from the static reference atlases that all current deconvolution methods assume — a failure mode that nobody has benchmarked because perturbation collections like scPerturb catalogue single-timepoint endpoints, not stress–recovery cycles. In between lies an optimal frequency where rare clones persist and their signatures stay recognizable. We further claim that a deconvolution model must be made phase-aware — switching between stress-phase and recovery-phase reference signatures — to realize this optimum, and that the optimum differs from the frequency that maximizes macroscopic yield, creating a real trade-off that a joint scheduler must navigate.

Formally, we seek a disturbance frequency $f$ minimizing expected composition error plus a signature-drift penalty, subject to keeping total biomass $B_T$ above a floor:

$$f^\*=\arg\min_f\; \mathbb{E}\big[\|\hat{\pi}_f-\pi_f\|_2\big]+\lambda\,D_{\mathrm{sig}}(f)\quad\text{s.t.}\;\; B_T(f)\ge B_{\min},$$

where $\pi_f$ is the true clone-proportion vector under frequency $f$, $\hat{\pi}_f$ is the deconvolved estimate from bulk data, and $D_{\mathrm{sig}}(f)$ measures how far current cell-state signatures have moved from the reference atlas. The unimodal form is expected because the first term decreases as disturbance suppresses the dominant clone (the intermediate-disturbance mechanism demonstrated empirically by Yang et al. with periodic low-dose UV breaking up microbial monopolies, and by Wu et al. with biennial mowing maximizing soil nematode diversity) while the second term increases with frequency.

Research Program

Three components. First, a simulation testbed. We extend the kinetic machinery that Yang et al. used — Lotka–Volterra competition dynamics with stochastic growth responses, run by Monte Carlo — from species abundance to a clonal cell community where each clone carries two expression signatures, one for the stress phase and one for the recovery phase, motivated by the cycling physiology documented in hydroponics, where a 72-hour salt-on/salt-off rhythm raised fruit sugar content by 24% without yield loss (Pascual et al.). Disturbance pulses differentially suppress the dominant clone, mimicking how mowing and surfact light gaps act on competitive dominants (Guitet et al.). We then aggregate the simulated community into pseudo-bulk RNA-seq count vectors following the mixture-generation protocol of Tran et al., giving us full ground truth for every time point.

Second, phase-aware deconvolution. Existing tools assume time-invariant cell-type signatures — an assumption Count Bridges, a generative model that reconstructs single-cell count data from aggregated integer counts via a Poisson birth–death bridge process, has never been tested against, since its benchmarks (MERFISH mouse brain, Onek1k) are static. Our method deconvolves twice, once with baseline references and once with stress-phase references, and uses the E-distance statistic from scPerturb — a permutation-tested distributional distance in cell-state space — to decide which reference the sample currently matches. This is the minimal change that addresses the static-atlas limitation, and it is the key ingredient the ablation will isolate.

Third, the scheduler itself. A small Bayesian-optimization loop treats disturbance frequency as the decision variable, uses the deconvolution model's posterior uncertainty over composition as its acquisition signal, and proposes the next pulse interval; biomass is monitored as the constraint. This mirrors in cell culture what the seagrass dredging study of whole-of-systems dynamic Bayesian networks did for ecosystems — where scheduling disturbance inside identified ecological windows cut recovery time up to fourfold and local extinction risk by 35% — but the objective there was resilience, whereas ours is observability.

Baselines

The do-nothing reference point is an undisturbed culture measured by bulk sequencing and deconvolved with the best current tool. This is the regime the whole proposal attacks: Tran et al. showed that at high purity every one of nine deconvolution algorithms mispredicts normal epithelial cells as cancer cells, though no reported result found on Lacuna for a numeric accuracy value in that purity regime, so we will establish it ourselves on our simulation and co-culture data as the floor to beat.

The strongest existing competitor is the best-performing deconvolution stack from Tran et al.: BayesPrism and DWLS, which achieved the lowest combined false-positive and false-negative counts and the lowest RMSE on the 38,000 pseudo-bulk mixtures generated from the Wu et al. 100,064-cell breast cancer atlas, with BayesPrism, Scaden, and MuSiC the most stable across purities from 5% to 95% (again, no reported result found on Lacuna for the numeric RMSE values themselves). We will run BayesPrism and DWLS, plus the generative Count Bridges — which outperformed CIBERSORTx and MuSiC on bulk proportion recovery — on identical simulated and experimental mixtures, with the same reference atlas, the same read depth, and the same tuning budget per method, so that any difference is attributable to the disturbance schedule and phase-awareness rather than compute. Beating this baseline means a strictly lower proportion RMSE and a higher rare-clone recall (detection of clones below 5% frequency) under the scheduled regime, with confidence intervals over simulation replicates and biological replicates respectively.

Two further comparisons complete the picture. A schedule chosen purely for macroscopic yield — the analogue of the 77 mM / 72-hour cycle that maximized sugar content in Pascual et al. — tests whether yield-optimal and observability-optimal frequencies coincide. And the critical ablation: the same optimized disturbance schedule but with static-reference deconvolution instead of phase-aware deconvolution, which isolates whether cycling signatures are the binding constraint or whether composition change alone carries the gain. If the therapy-response extension is pursued, the number to beat is the median AUROC of 0.74 that SubNetDL reported on the TCGA cohort of 507 patients across ten cancer–drug pairs, and its 0.77 on MSKCC non-small cell lung cancer immunotherapy cohorts.

Why not just…?

Why not just run single-cell sequencing at every time point? If bulk masks heterogeneity, the direct fix is to stop using bulk. The obstacle is scale: scPerturb established that distinguishing a perturbation shift from noise requires at least 300 cells per condition and 1,000 UMI counts per cell, so a grid over frequency, amplitude, and recovery interval multiplies quickly into millions of cells and prohibitive cost, which is exactly why tools like POPPY and Count Bridges exist to transfer single-cell knowledge into cheap bulk assays. Our design uses sparse single-cell anchors — enough to satisfy the scPerturb power thresholds — as ground truth, and dense bulk in between. If single-cell-everywhere were cheap enough, the observability half of the claim would become moot, but the ecological half (which frequency prevents clonal monopoly and state drift) would still need the same sweep, so the scheduler remains useful either way.

Why not just improve the deconvolution algorithms on existing static benchmarks? More data, better priors, or a cleverer generative model might fix the purity trap without touching the biology. The evidence says no: Tran et al. benchmarked nine methods spanning very different algorithmic families, including the deep-learning Scaden and the generative Count Bridges approach, and all of them fail at high purity in the same direction. The reason is physical, not algorithmic — when a clone is 95% of the sample, the information about minority clones is barely present in the reads, and no inference can recover what was never measured. Disturbance changes the sample itself rather than the inference on it. That said, this is partly an empirical wager, and our simulation sweep will quantify the split: how much of the error reduction at intermediate frequencies could have been achieved by BayesPrism alone on an undisturbed sample, versus how much required the disturbance.

Why not just use cheap diversity indices (Simpson, Chao1) to pick the schedule? Yang et al. already showed Simpson diversity rises under periodic UV, so a skeptic could say: measure evenness with a cheap assay, target maximum evenness, and skip all single-cell machinery. This is a genuine threat to the fancier parts of the proposal, and we test it head-on: Experiment 4 regresses deconvolution error against evenness across the frequency sweep. If evenness fully predicts error, the scheduler simplifies to a diversity controller and the phase-aware deconvolution becomes unnecessary — we would report that happily as a useful simplification. But there is a concrete reason to expect a gap: evenness measures who is present, not what state they are in, and the signature-drift penalty is invisible to any abundance index. Pascual et al.'s finding that the same salt cycle that helped in buffered substrate systems caused over 10% fruit deformity in unbuffered ones shows macroscopic readouts miss mechanism; here the analogous miss is state drift, which only a transcriptomic readout can see.

Why not just borrow intermittent-dosing schedules from oncology? Adaptive therapy already pulses drugs. But the closest modeling work, SubNetDL, encodes treatment as a static covariate (which drug, binary exposure) and infers subclones from a single biopsy, whereas we treat frequency and interval as continuously tunable control variables with observability as the explicit objective — a distinction the seagrass study makes vivid, where identical dredging stress caused collapse or recovery depending on timing and the 3-month duration threshold. The two are complementary rather than conflicting: SubNetDL supplies the subclonal architecture we would use as the state model inside our scheduler.

Proposed Experiments

Experiment 1 is the simulation sweep. We vary disturbance frequency from never to near-continuous across the Lotka–Volterra testbed, generate pseudo-bulks at each steady-state sampling point, and measure proportion RMSE and rare-clone recall for BayesPrism, DWLS, and Count Bridges under both static and phase-aware references. The decisive check is the shape of the curve: a unimodal error profile with its minimum at intermediate frequency and a purity-trap plateau at zero disturbance would confirm the core claim; a monotone improvement with frequency would falsify the signature-drift mechanism and the phase-aware component with it.

Experiment 2 is the ablation grid: the same schedules with static versus phase-aware references, crossed with two disturbance modalities (uniform kill versus dominant-preferential kill). This separates the composition effect from the signature effect and tells us which mechanism dominates at which frequency.

Experiment 3 is the wet-lab validation: a co-culture of three to five barcoded cancer cell lines, including a tamoxifen-selected line in which resistance emerged via rare stem-like cells that bulk analysis could not distinguish from uniform population shifts (Lin et al.; the same masking of pluripotency programs is documented across cancers by Palmer et al., and the general averaging problem by Adams et al.). We apply periodic low-dose cytotoxic pulses at the scheduler-chosen frequency and at two controls (never, always), take dense bulk RNA-seq and sparse scRNA-seq anchors meeting the scPerturb power thresholds, and score deconvolution against barcode ground truth. Success is rare-clone recall under the scheduled regime significantly exceeding both controls while biomass stays above the floor.

Experiment 4 closes the loop on the skeptic's alternatives: regressing deconvolution error on Simpson evenness (the cheap-index test) and, in simulation only, asking whether an observability-optimized schedule improves long-horizon prediction of which clone wins — connecting back, if it does, to the therapy-response setting where SubNetDL's median AUROC of 0.74 on TCGA and 0.77 on the MSKCC lung cancer cohorts define the bar a schedule-aware successor would need to clear. The unifying deliverable is a defensible answer to a question nobody has asked: not how often to disturb a biological system for yield, but how often to disturb it so that we can still see what is inside it.