When cancer cells are exposed to a targeted drug for months, they adapt and survive — a process called acquired resistance induction. Two things are known separately: resistant cells often become more motile, and the molecular profiling of the adaptation process (differential gene expression analysis, which finds genes whose abundance shifts between conditions, followed by functional enrichment analysis, which projects those genes onto curated pathway dictionaries like Gene Ontology) suggests migration programs switch on early. But the motility claim rests on crude, end-point measurements of fully resistant lines, and the transcriptomic claim has never been physically validated. This proposal closes that loop: we will run standardized wound healing assays — a scratch in a cell monolayer whose closing speed measures collective migration — at every stage of a months-long resistance selection, fit each scratch movie with a partial-differential-equation discovery pipeline that separates motility from proliferation and from the mechanical shock of scratching itself, and match those kinetic parameters against stage-resolved transcriptomes. The result is a quantitative, time-resolved map of when and how the migration phenotype is acquired, and which transcriptional programs drive it.
The gap is concrete. Curcumin attenuates lncRNA H19-induced epithelial-mesenchymal transition in tamoxifen-resistant breast cancer cells is the only study in this corpus that physically connects resistance to scratch assays, but it compares just two points (parental versus fully resistant MCF-7/TAMR), monitors the scratch manually, and checks only hand-picked markers (H19, Snail, E-cadherin). Meanwhile the transcriptomic side has found migration signatures without ever measuring migration: Focal adhesion kinase-YAP signaling axis drives drug-tolerant persister cells and residual disease in lung cancer shows that early drug-tolerant persisters survive through FAK — a kinase whose day job is sensing mechanical attachment and driving motility — but runs no motility assay; and Integrating multiscale mathematical modeling and multidimensional data reveals the effects of epigenetic instability on acquired drug resistance in cancer predicts persister enrichment in autophagy and migration pathways purely from models and single-cell RNA data. Nobody has measured wound closure on the cells in between: the drug-tolerant persister (DTP) — a slow-cycling, non-dividing survivor state that appears within days of drug exposure — or the intermediate sublines produced by stepwise dose escalation.
Core Claim
The gain in collective migration during acquired resistance is not a late byproduct of the fully resistant state; it emerges early, in the drug-tolerant persister window, and can be detected as a shift in two fitted wound-closure parameters: the motility coefficient (how fast cells spread into the gap) and the wound-shock delay term (a measured lag in migration and proliferation right after scratching, first demonstrated in Computationally efficient mechanism discovery for cell invasion with uncertainty quantification). We further claim that the timing of these kinetic shifts is predictable from stage-resolved differential expression and pathway enrichment: migration- and mechanotransduction-related gene modules become enriched at the selection stage where the motility coefficient moves, and not before. The sharpest open question we will settle is how two shock dynamics interact: persister cells are already quiescent and stress-adapted, so when we scratch their monolayer, does that adaptation buffer them against the wound-shock delay, or do the two shocks compound into a longer arrest? No study in the corpus has quantified the delay term under any drug pressure.
Research Program
The system is PC9 EGFR-mutant lung cancer cells under stepwise escalation of an EGFR inhibitor, chosen because it directly extends the two motivating studies: the FAK-YAP persister work used PC9 under osimertinib (FAK-YAP...), and the multiscale modeling study analyzed PC9 scRNA-seq under erlotinib and gefitinib (Integrating multiscale mathematical modeling...). A secondary arm uses lapatinib-treated HER2+ breast cancer cells, where Ontogeny and Vulnerabilities of Drug-Tolerant Persisters in HER2+ Breast Cancer showed persisters split into luminal-like and mesenchymal-like branches and explicitly left the mesenchymal branch's functional behavior unresolved; we will measure whether those two branches diverge in wound-closure kinetics.
At each selection stage — parental, DTP window (days 3, 7, 14 under drug), each dose-escalation step, and late stable resistance — we run standardized scratch assays with three deliberate improvements over the literature. Scratches are made with a laser-cut acrylic guidance mold, which An image J plugin for the high throughput image analysis of in vitro scratch wound healing assays showed cuts scratch width and angle variability by roughly 50%; images are quantified with that paper's Wound_healing_size_tool plugin (a 94% analysis speedup over MiToBo, processing 60 images in under 8 minutes); and each stage is run both with and without mitomycin C so proliferation can be separated from pure motility. Critically, every stage is scratched both under continuing drug and after a short drug washout, so we can distinguish the drug's direct effect on motility from the selection process's lasting effect on the cells.
Each 48-hour image stack is then fitted with the equation-learning pipeline of Computationally efficient mechanism discovery..., whose Julia package (EquationLearning.jl) uses Gaussian-process surrogates, over 100 parametric-bootstrap resamples, and AICc model selection, and which runs in under an hour on a desktop. That pipeline, validated on drug-naive PC-3 cells, distinguishes linear diffusion (motility) from logistic growth (proliferation) and quantifies the post-scratch shock as an explicit delay term; we treat that delay term as a phenotype, not a nuisance. In parallel, each stage is profiled by bulk RNA-seq in triplicate, with separate collection of leading-edge and trailing-edge cells from the monolayer to address the spatial blind spot of aggregate scratch measurements. Differential expression follows the negative-binomial count modeling of Detecting Molecular Signatures via Differential Gene Expression Analysis with false-discovery-rate correction; enrichment uses standard GO and KEGG overrepresentation testing; and interpretation of the resulting gene modules is grounded in the evidence-checked agentic style of BioGen: An Evidence-Grounded Framework for Interpreting RNA-seq Gene Clusters in Antimicrobial Resistance Research, which we extend from bacterial antimicrobial resistance to mammalian persister-migration modules. BioGen's critic-agent design reduced hallucination from 0.67 (unconstrained LLM baseline) to 0.00 with a semantic alignment score of 0.59 on its Salmonella benchmark; porting that auditable interpretation to our wound-persister modules is a reusable computational deliverable in itself, since neither BioGen nor GenoMAS has been applied to mammalian phenotypic-assay data.
The integration step is the scientific payoff: we correlate the trajectory of the fitted motility and delay parameters across stages with the trajectory of migration-, FAK-, autophagy-, and EMT-related enrichment scores, and test whether enrichment timing predicts kinetic timing. Two interventions make this causal rather than correlational. First, at whatever stage the motility coefficient first shifts, we add the FAK inhibitor VS-4718 and the YAP/TEAD inhibitor used in FAK-YAP... during the scratch assay; if the migration gain runs through FAK-YAP, the inhibitor should collapse the elevated motility coefficient back toward parental values. Second, we impose a roughly 10-day drug holiday — the schedule the multiscale modeling study identified as optimal for reducing epigenetic instability — and re-measure closure kinetics, asking whether the migration phenotype is reversible plasticity or a fixed selected trait.
Baselines
The naive reference point is the design of Curcumin attenuates lncRNA H19...: compare only parental versus final resistant cells, with manual scratch monitoring and RT-qPCR/Western blots for canonical EMT markers. This is the strongest competitor for simplicity and is the field's default, so we will run it exactly as an arm of our study — same wells, same image stacks, endpoint manual tracing and candidate-marker blots — with identical replicate counts and no extra tuning. It is beaten if stage-resolved fitting finds kinetic differences (for example, a DTP-stage delay-term shift or an intermediate motility gain) that the two-point design structurally cannot see; formally, we require that a model with stage-specific parameters beats a two-state model under the same AICc selection used by the equation-learning pipeline.
The strongest existing method for our measurement is the GP-surrogate equation-learning pipeline itself (Computationally efficient mechanism discovery...); it discovered the logistic delay term in drug-naive PC-3 cells across initial densities of 10,000–20,000 cells per well, using over 100 bootstrap resamples and under one hour of runtime per analysis, though it reports no quantitative wound-closure percentage or motility coefficient value (no reported result found on Lacuna for wound-closure percentages in that study). Because we adopt it unchanged — same software, same bootstrap count, same model library, same compute — it functions as a fair, pre-validated measurement instrument rather than a competitor to out-tune. The genuine scientific competitor is the in-silico claim it must confront: the position of Integrating multiscale mathematical modeling... and FAK-YAP... that migration enrichment in persister transcriptomes implies motility gain. We run that prediction head-to-head: enrichment-only inference versus physical measurement at each stage. Our approach beats it if the transcriptome-only inference mis-calls direction or timing — for instance, if migration pathways are enriched at a stage where the fitted motility coefficient is unchanged, or if the delay-term dynamics (which no enrichment analysis speaks to at all) dominate the phenotype. If enrichment and kinetics agree perfectly, the in-silico shortcut wins and we will say so.
The key ablation isolates longitudinal stage resolution: we re-run the entire pipeline — same images, same transcriptomes, same fitting — collapsing the timeline to parental and late-resistant endpoints only. If the collapsed analysis yields the same scientific conclusions as the stage-resolved one, the longitudinal design (the proposal's central investment) is unjustified; if it misses the DTP-window kinetics that the full design catches, the ablation is the cleanest demonstration of why the timeline matters. A second ablation drops the critic-verification layer from the BioGen-style interpretation, where prior work shows hallucination rates rise to 0.67 without it, testing whether auditable grounding actually matters for mammalian modules.
Why not just…?
Why not just give cells one acute high-dose shock instead of a months-long selection? An acute IC90 exposure is faster, but it selects for intrinsic resistance — rare cells that already survive catastrophe — rather than the progressive adaptation clinicians actually face. Genetic barcoding in FAK-YAP... and Ontogeny and Vulnerabilities... showed persisters arise stochastically across clones through transcriptional plasticity, not from pre-existing mutants, so an acute screen would miss exactly the transitional states whose motility we want to measure. To settle this empirically rather than by argument, we include a 72-hour acute-exposure arm as a control: if acute survivors show the same wound-closure kinetics as stepwise DTPs, the reviewer is right and the long selection adds nothing; we expect them not to, because the plastic persister state needs time under sub-lethal pressure to establish.
Why not just use Transwell chambers instead of a variable, trauma-inducing scratch? Transwell assays cleanly measure single cells crawling through pores toward a chemical gradient (CXCR7 functions in colon cancer cell survival and migration), but our target phenotype is collective migration of a cohesive sheet — the way carcinomas actually invade tissue — which Transwell cannot capture. The two known flaws of scratch assays are handled rather than suffered: mold-guided scratching and the automated plugin cut variability by about 50% and remove observer subjectivity (An image J plugin...), and the scratch trauma itself is measured as the delay term rather than ignored. That said, single-cell chemotaxis could diverge from collective migration in mesenchymal-like persisters, so a Transwell arm on the HER2+ branch is a worthwhile secondary check, not the primary readout.
Why not just run RT-qPCR for E-cadherin, Snail, and Vimentin instead of RNA-seq? Candidate markers confirm what you already suspect; the corpus shows the decisive discoveries came from unbiased profiling. The FAK-YAP bypass in FAK-YAP... — YAP activating FGFR1 and BCL-XL outside the classical Hippo route — and the NPY1R and ER-SGK3 findings in Ontogeny and Vulnerabilities... would have been invisible to an EMT panel, and the multiscale study's autophagy-migration enrichment in DTPs (Integrating multiscale mathematical modeling...) names a program no EMT marker covers. More fundamentally, our central phenotype — the wound-shock delay term — has no candidate marker at all. We will still measure the canonical markers, but as validation of the differential-expression hits, not as a substitute for the unbiased screen.
Proposed Experiments and Decisive Checks
The program delivers four decisive outcomes. First, the delay-interaction result: comparing the fitted delay term across parental, DTP, and resistant stages, with over 100 bootstrap resamples per stage per the equation-learning methodology, settles whether persister quiescence buffers or compounds wound shock — a question flagged by the tension between the delay-term finding of Computationally efficient mechanism discovery... and the quiescent-persister findings of Integrating multiscale mathematical modeling..., and answerable by no existing study. Second, the timing result: whether the motility coefficient shifts first in the DTP window, at an intermediate escalation step, or only at stable resistance, and whether substrate-context effects on migration (as in The effect of stiff substrates on the collective migration of A549 cells) modulate that timing. Third, the prediction result: the correlation between module-enrichment timing and kinetic-shift timing across stages, which either validates or refutes the field's habit of inferring motility from enrichment. Fourth, the causality results: FAK or YAP/TEAD inhibition collapsing the elevated motility coefficient, and the ten-day drug holiday (Integrating multiscale mathematical modeling...) revealing whether the phenotype is reversible plasticity or fixed selection. On the HER2+ branch, divergent closure kinetics between luminal-like and mesenchymal-like persisters would directly resolve the open question Ontogeny and Vulnerabilities... left on the table. Success is defined in advance: a stage-resolved model that beats both the two-point baseline and the enrichment-only inference on the same data, with the inhibitor and holiday experiments confirming the direction of the mechanism the transcriptomes point to.
References
- Curcumin attenuates lncRNA H19‑induced epithelial‑mesenchymal transition in tamoxifen‑resistant breast cancer cells
- Focal adhesion kinase-YAP signaling axis drives drug-tolerant persister cells and residual disease in lung cancer
- Integrating multiscale mathematical modeling and multidimensional data reveals the effects of epigenetic instability on acquired drug resistance in cancer
- Computationally efficient mechanism discovery for cell invasion with uncertainty quantification
- Ontogeny and Vulnerabilities of Drug-Tolerant Persisters in HER2+ Breast Cancer
- An image J plugin for the high throughput image analysis of in vitro scratch wound healing assays
- Transcriptomic Analysis of the Candidate Genes Related to Aroma Formation in Osmanthus fragrans
- BioGen: An Evidence-Grounded Framework for Interpreting RNA-seq Gene Clusters in Antimicrobial Resistance Research
- GenoMAS: A Multi-Agent Framework for Scientific Discovery via Code-Driven Gene Expression Analysis
- CXCR7 functions in colon cancer cell survival and migration
- The effect of stiff substrates on the collective migration of A549 cells