Rescue-Informed Dose-Response Curves for Iron-Dependent Death in Insecticide-Resistant Aphids

Insecticide resistance in agricultural pests is monitored almost entirely as a single number: the median lethal concentration ($LC_{50}$) from a probit regression — a statistical model that assumes each insect has a random tolerance threshold and that the population's death probability follows an S-shaped curve against log-dose. This practice has a known blind spot. When a resistant aphid strain shows cross-resistance to chemically unrelated insecticides, the scalar $LC_{50}$ ratio says nothing about how the insects die, and decades of resistance research have attributed oxidative-stress lethality to generic apoptosis without ever measuring the alternative. In mammalian cells, a regulated, iron-dependent death pathway — ferroptosis, in which labile iron catalyzes lethal lipid peroxide accumulation in membranes until they rupture — turned out to hide inside what was long assumed to be ordinary oxidative apoptosis. Nothing in the literature has checked whether the same pathway operates in insects, or whether the antioxidant machinery that resistant pests upregulate to survive insecticides also shields them from — or fails against — iron-driven membrane destruction.

This proposal tests ferroptosis directly in insecticide-resistant insects, using dose-response modeling not as a bookkeeping device but as a measurement instrument. The core idea: pharmacological rescues that specifically block one death pathway (iron chelators and lipophilic radical traps for ferroptosis; caspase inhibitors for apoptosis) act as perturbations on the full mortality curve, and the shape of the resulting curve shifts tells us which pathway the toxicant actually engaged, separately for resistant and susceptible strains. Dose-response modeling is the methodological lens, insecticide resistance is the application regime, ferroptosis is the candidate mechanism, and mechanism-of-action (MoA) elucidation — deciding between competing death-pathway hypotheses from evidence rather than assumption — is the point of the whole exercise.

Core Claim

Field-evolved, multi-insecticide-resistant pests differ from susceptible strains in their capacity for iron-dependent lipid peroxidation, and this difference is visible in the full mortality curve, not just in scalar $LC_{50}$ values. The resistant strain's upregulated antioxidant defenses (glutathione-linked systems, glutaredoxins) may either confer incidental cross-protection against ferroptosis or, because resistance carries a metabolic fitness cost, leave a specific vulnerability to iron and lipid-peroxide challenge. We claim that a two-pathway dose-response model, fitted jointly across induction and rescue conditions, can (i) resolve whether death under a given compound is apoptotic, ferroptotic, or mixed, and (ii) predict how resistant and susceptible curves diverge under combination exposures — something no existing probit-based or deep-learning approach can do, because neither represents mechanism at all.

The proposal is deliberately parametric rather than deep. ARCANet, a network that predicts entire dose-response curves from molecular graphs and cell-line transcriptomes, itself reports that below roughly 2,000 data points, simple 3- and 4-parameter logistic models are more robust — and a typical insect bioassay uses 5–7 concentrations with tens to hundreds of insects per dose, exactly the sparse regime where deep curve predictors lose. So we keep the low-dimensional parametric form of classical entomology but make its parameters mechanistic and shared across rescue conditions.

Research Program

Biological system. The primary system is the susceptible (SS) and sulfoxaflor-resistant (SulR) strains of the wheat aphid Sitobion miscanthi characterized by Xinan Li et al., who report resistance ratios of 199.8 to sulfoxaflor, 42.1 to bifenthrin, 14.5 to beta-cypermethrin, and 5.7 to chlorpyrifos, alongside a 27% fitness cost ($R_f = 0.73$) and 4.18 fewer offspring per aphid — a documented, unexplained multi-class cross-resistance whose molecular drivers the authors explicitly left undetermined. The secondary system is Spodoptera exigua with RNAi knockdown of SeGrx1 (glutaredoxin 1), where Fulai Yang et al. showed that camptothecin mortality rises from 32.6% to 72.9% when the antioxidant is silenced, at up to 72.1% knockdown efficiency with 1.5 µg dsRNA per larva. That paper attributed the death to apoptosis purely from morphology; we treat that attribution as the unverified hypothesis it is.

The model. We fit mortality as an independent-hazard mixture of two sigmoidal components, one per death pathway:

$$P(\text{death}\mid d, r) = 1 - \big(1 - \pi_A\, F(d;\alpha_A, \beta_A)\big)\big(1 - (1-r)\,\pi_F\, F(d;\alpha_F, \beta_F)\big)$$

Here $d$ is log-dose, $F$ is the standard normal cumulative distribution function (the probit core), $\pi_A$ and $\pi_F$ are the maximum kill fractions of the apoptotic and ferroptotic components with their own potency $\alpha$ and slope $\beta$, and $r \in [0,1]$ is the rescue penetration — the fraction by which a co-administered pathway-specific inhibitor (ferrostatin-1 or deferoxamine for ferroptosis, a caspase inhibitor for apoptosis) suppresses that component. This structure makes falsifiable predictions: if a compound kills only apoptosis, fitting the induction curve alone must predict the caspase-inhibitor rescue curve and leave the ferrostatin curve untouched, and vice versa. The ferroptosis-specific rescue logic follows the mammalian gold standard noted in the ferroptosis literature, where bulk lipid-peroxidation readouts only count as ferroptosis evidence when rescued specifically by iron chelators or lipophilic radical traps; hydropersulfides and silibinin provide additional mechanistically distinct radical-trapping and cytoprotective controls.

Ortholog check. Before any bioassay, we search the aphid and S. exigua genomes for orthologs of the ferroptosis gatekeepers — GPX4 (the enzyme that reduces lipid hydroperoxides) and SLC7A11 (the cystine importer that keeps the glutathione pool reduced) — and measure basal labile iron, glutathione, and malondialdehyde in SS versus SulR. The species-conservation gap is a stated open problem in the ferroptosis literature, and this step is the gate for the whole program: if the machinery is absent, the honest outcome is a well-documented negative result.

Morphological profiling arm. In parallel, we adapt high-content imaging to the S. exigua cell line IOZCAS-Spex-II (used by Fulai Yang et al.). TRex showed that a frozen self-supervised vision model (DINO) paired with a lightweight adaptation module doubles mechanism-of-action classification mean average precision from 0.15 to 0.32 on the JUMP-CP Cell Painting benchmark, and generalizes to unseen compounds (0.24 to 0.28). We will fine-tune the same TRex recipe on a small insect-cell image set labeled by our rescue experiments (ferrostatin-rescued vs. caspase-rescued vs. unresolved), testing whether morphological signatures transfer from mammalian profiles to invertebrate cells — a direct attack on the reference-database-coverage problem that UNAGI and connectivity-mapping approaches face, since LINCS L1000 and CMAP are biased toward human cancer lines and contain no invertebrate profiles.

Proposed Experiments

E1 — Full-curve bioassays with rescue panel. For each strain (SS, SulR), each toxicant (sulfoxaflor, beta-cypermethrin, chlorpyrifos, camptothecin, plus a canonical ferroptosis inducer such as a system-$X_c^-$ or GPX4 inhibitor), we run 7–9 concentrations × 5 replicates, crossed with four rescue arms: vehicle, ferrostatin-1, deferoxamine, and a caspase inhibitor. Primary outputs: pathway decomposition ($\pi_A$, $\pi_F$ per compound per strain), rescue-specific curve shifts, and C11-BODIPY lipid-peroxide, iron, and glutathione assays as independent biochemical readouts. The sanity anchor in S. exigua is the replication of the published 32.6% → 72.9% camptothecin-plus-dsSeGrx1 mortality jump, which we then re-interrogate with the rescue panel. A useful mammalian calibration point comes from RRM1 knockdown, where antioxidant-axis collapse roughly doubled radiation-induced lipid peroxidation; a comparable fold-change under insect antioxidant knockdown would be the first invertebrate analogue of that result.

E2 — Resistance-ratio transfer. Does ferroptosis-inducer resistance track insecticide resistance? We compute resistance ratios for the ferroptosis inducers between SS and SulR, exactly as Xinan Li et al. did for insecticides. Two outcomes are both informative: transfer (resistant aphids are also ferroptosis-resistant, implying a shared glutathione-linked shield) or inversion (resistant aphids are more ferroptosis-susceptible, implying the fitness cost is a redox liability — supported by the prolonged development and reduced fecundity of SulR, and consistent with the metabolic-stress-sensitivity vulnerability that remains unmapped in the literature).

E3 — Combination grids. Insecticide × ferroptosis inducer on a 4×4 concentration grid, per strain, testing whether the inducer acts as a resistance-breaking synergist. We compute both classical scalar synergy (Loewe additivity and Bliss independence from $LC_{50}$s) and a joint fit of the mixture model over the grid. The deep-synergy alternative, DeepSynergy, which reached a mean squared error of 255, a Pearson correlation of 0.73, and an AUC of 0.90 on the Merck Oncology Screen, is cited as the reference point for what learned synergy models achieve with ~23,000 samples of transcriptome-paired data; our grids will have orders of magnitude less data and no insect transcriptomic training corpus, which is precisely why we use the 9-parameter mechanistic model instead.

E4 — Profiling transfer. Train TRex-style adaptation on the insect-cell rescue-labeled images; evaluate on held-out compounds. Pre-registered comparison: baseline frozen DINO features with a linear probe (published at mAP 0.15 on JUMP-CP) versus the adaptation module (0.32). No reported result found on Lacuna for TRex performance on any invertebrate dataset, so this is a first-of-kind measurement regardless of outcome.

Baselines

The naive reference point is standard practice itself: leaf-dipping bioassays analyzed with probit regression in PoloPlus, exactly as in the field-evolved sulfoxaflor resistance survey of 58 wheat-aphid populations, producing $LC_{50}$, slopes, and resistance ratios. This is the strongest competitor on the pure monitoring question because it is the field's accepted instrument and it is statistically suited to sparse bioassays. Its published output on our exact biological system is the SulR resistance ratio of 199.8 for sulfoxaflor from Xinan Li et al. — the number our full-curve model must reproduce within confidence intervals on the induction curves. No reported result found on Lacuna for any performance metric of PoloPlus as software; its authority is methodological, not benchmarked. Beating this baseline means matching its $LC_{50}$ estimates while additionally predicting held-out rescue-curve shapes, which PoloPlus cannot represent at all.

The strongest methodological competitor for curve shape is the 4-parameter logistic (4PL) Hill fit applied independently per condition. We choose it as the strongest because ARCANet's own data-efficiency analysis shows parametric models outperform deep networks below ~2,000 points, which is our regime; no reported result found on Lacuna for ARCANet's absolute error values on GDSC or PRISM, so we cannot quote a numeric target from it. We will run the comparison fairly by giving both models the same raw counts, the same concentration grids, and the same Bayesian fitting machinery (only the likelihood differs), with held-out rescue arms as the test set. Beating 4PL means lower held-out deviance on rescue curves and correct prediction of which rescue shifts a given compound's curve — a prediction 4PL cannot make because it shares no parameters across conditions. The key ablation is our own model with the ferroptosis component removed (collapsing to a single-pathway probit): if the two-component version does not beat this ablation on rescue-arm held-out likelihood, the mechanistic claim is dead and we report it.

For synergy, the baseline is scalar Loewe/Bliss co-toxicity computed from per-compound $LC_{50}$s, the standard entomological approach; for the imaging arm, the baseline is the frozen-feature linear probe with its published JUMP-CP mAP of 0.15 versus TRex's 0.32.

Why not just…?

Why not just run PoloPlus probit analysis, as entomologists have for decades? For monitoring, one should — and we do, as our sanity anchor. But the probit scalar compresses the whole response into one number, and the open problem in the resistance literature is precisely mechanistic: the SulR cross-resistance to sulfoximines, pyrethroids, and organophosphates has documented ratios but no determined mechanism (Xinan Li et al.). A scalar cannot distinguish "same death pathway, higher tolerance" from "different death pathway entirely," and the rescue curves that carry that information are invisible to it. Our approach keeps probit's sparse-data strengths and adds only pathway-shared parameters, so the comparison is parametric-versus-parametric, not deep-versus-classical.

Why not just accept that oxidative insect death is apoptosis, as the insect literature reports? This is a real threat and we take it seriously. Fulai Yang et al. observed apoptotic morphology in S. exigua cells, and the zebrafish OBS study (OBS developmental toxicity) attributed oxidative toxicant death to apoptosis. But the mammalian precedent cuts the other way: RRM1 knockdown showed that oxidative chemotherapeutic stress triggers simultaneous apoptotic DNA damage and ferroptotic lipid peroxidation through the same antioxidant-axis collapse — mixed death that morphology alone misassigned for years. Our design settles this empirically rather than rhetorically: if ferrostatin-1 and deferoxamine fail to shift any curve and iron/lipid-peroxide assays stay flat, the apoptosis reading stands and E1 delivers a clean negative result with the first quantitative rescue-panel data on insect cells either way.

Why not just use classical Bliss or Loewe synergy scores instead of a joint model? We compute them as a baseline because they may suffice. Where they fall short is concentration-dependent behavior: scalar synergy at the $LC_{50}$ averages over the grid and cannot represent a synergy that appears only at low doses or a biphasic tail — the exact failure mode that motivated full-curve prediction in ARCANet and that a resistant strain's partial-kill plateau makes likely. The decisive test is pre-registered: wherever scalar and joint-fit synergy disagree, we run the confirmatory grid at those specific concentrations. If the scalar scores match the joint fit everywhere, the simpler method wins and we say so.

Why not just sequence resistance targets and assay P450 activity? Targeted assays only find known mechanisms, which is exactly why the SulR cross-resistance across three distinct chemical classes remains unexplained — no documented VGSC mutation or P450 profile accounts for it (field survey, Lin et al. cover target-site and metabolic mechanisms but not this case). We include enzyme and sequencing assays in E1 as complementary measurements; the unbiased profiling arm exists because systemic, pathway-level adaptation — the kind a fitness-costed resistant strain must make — is what targeted assays miss by construction.

Decisive Checks

Three outcomes would each be decisive. First, if aphid and S. exigua genomes lack functional GPX4/SLC7A11 orthologs and no rescue-by-ferrostatin shift is detectable anywhere, ferroptosis is absent or insensitive in these insects — a publishable negative that closes the Gap-1 question for two major pest species. Second, if the ferroptosis component is real and its resistance ratio in E2 is inverted relative to the insecticide ratios, we have a resistance-breaking entry point with a quantified dose-response surface from E3. Third, if the two-component model fails its own ablation on held-out rescue curves, the mechanistic curve decomposition is over-parameterized and classical per-condition probit remains the correct sparse-data instrument — in which case the surviving contribution is the rescue-panel dataset itself, the first to quantify pathway-specific lethality across a resistant/susceptible pair.