Toxicologists summarize a pesticide's potency with a single number: the concentration that kills or inhibits half of a target population (the LC50 or IC50). When researchers want to know why a pest survives that concentration, they sequence its transcriptome — but, almost always, at just one dose, usually the lethal one. This means the field knows what genes look different at the point of death, but not when each biological program switches on as the chemical climbs from harmless to lethal. Yet everything we care about mechanistically is a timing question in disguise: resistance is a detoxification program that turns on far below lethal concentrations, and a synergist is a second compound that pushes that program out of the way. Neither can be seen in a single-dose snapshot.
The proposal here is to sequence pest transcriptomes across the full dose gradient — a titration, not a snapshot — and then fit a dose–response curve not to mortality but to each biological pathway, using the same four-parameter logistic (4PL) curve toxicologists already trust. This yields a "pathway activation midpoint" (a PAC50) for every KEGG and Gene Ontology term, and turns two long-standing descriptive findings into measurable, predictive quantities: the gap between a resistant strain's detoxification PAC50 and its LC50, and the shift in that gap when a synergist is added. The design is deliberately parametric and small-data-friendly, because the agricultural datasets at hand are tiny (16 RNA-seq libraries in the closest prior work) and deep models are known to lose to 4PL fits below roughly 2,000 data points.
Core Claim
Resistance and synergy have dose–response signatures at the pathway level, and these signatures carry information that neither mortality curves nor single-dose transcriptomics contain. Specifically: in a resistant aphid strain, xenobiotic-metabolism pathways should activate at concentrations far below the population's lethal concentration, while in a susceptible strain the same pathways should activate only as death approaches; and adding a synergy partner to a botanical toxicant should shift or suppress the detoxification pathway's activation curve in a way that scales with the empirically measured interaction coefficient. If both hold, the PAC50 — cheap to estimate from a few dozen RNA-seq libraries — becomes a mechanistic intermediate variable connecting potency scalars to pathway biology, usable to diagnose resistance mechanisms and to choose mixture ratios rationally rather than empirically.
Research Program
The core object is the pathway titration curve. For each pathway $p$ (a KEGG or GO gene set), define an activity score at each tested concentration $d$:
$$A_p(d) = \frac{1}{|p|}\sum_{g \in p} z_g(d)$$
where $z_g(d)$ is the log-fold change of gene $g$ relative to a matched untreated control, z-scored across the gradient. A 4PL curve is fit to $A_p(d)$ over log-dose, and the PAC50 is its inflection point: the concentration at which the pathway's response is half-maximal. Because related pathways share genes and standard overrepresentation analysis treats them as independent, the pathway scores will be computed coherently using regularized inter-pathway regression, following pareg, which couples a LASSO sparsity penalty with a network-fusion penalty on pathway overlap and which reported the highest mean area under the precision–recall curve against Fisher's exact test, topGO, MGSA, and blitzGSEA on synthetic and TCGA-BRCA benchmarks (no numeric AUPRC value is reported for pareg in the file) Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression. Execution at scale will use the Spark-backed funkea framework, which unifies five enrichment methodologies funkea: Functional Enrichment Analysis in Python.
The empirical work has three arms. First, a dose-resolved redesign of the closest existing study: Transcriptome-Based Identification and Characterization of Genes Associated with Resistance to Beta-Cypermethrin in Rhopalosiphum padi compared a susceptible strain (SS) and a resistant strain (Beta-R, resistance ratio 4588.48) at a single LC50 challenge with four biological replicates per group. We will run the same insect and insecticide across a strain-specific seven-point gradient spanning roughly LC10 to LC90 plus untreated controls, three replicates per point, mapped to the same reference genome (GCA_020882245.1) — about 48 libraries. Second, a synergy arm built on the one system in the file with quantified botanical synergy: R-carvone plus DL-limonene against the cigarette beetle, where the 1:1 fumigant mixture gave an interaction coefficient R = 1.36, and the enantiomer pair R-carvone plus S-carvone gave R = 3.15 in fumigation but an essentially additive R = 0.98 in contact toxicity Bioactivities and Synergistic Effect of Elsholtzia ciliata Essential Oil and Its Main Components against Lasioderma serricorne. Titration curves for single agents and mixtures, in both fumigant and contact routes, give us a built-in dose–response ladder of synergy strength (0.98, 1.36, 3.15) against which predicted pathway shifts can be ranked. Third, a cross-resistance arm on the grain aphid system, where a sulfoxaflor-resistant field strain (RR = 199.8) also showed elevated resistance to pyrethroids (RR 14.5–42.1) and a fitness cost (relative fitness 0.73) but no transcriptomics at all Fitness Cost of the Field-Evolved Resistance to Sulfoxaflor and Multi-Insecticide Resistance of the Wheat Aphid Sitobion miscanthi (Takahashi); gradient profiles under sulfoxaflor and one pyrethroid test whether shared early-activating detox pathways explain the cross-resistance pattern that paper could only hypothesize.
The interpretive payoff comes from clustering pathways by PAC50 relative to each strain's LC50. Programs activating in the LC10–LC30 range are candidate compensatory resistance mechanisms; programs activating only above LC80 are terminal stress collapse. This distinction is precisely what the single-dose design conflates — a point made from the pharmacology side by Predicting Dose-Response Curves with Deep Neural Networks, which showed that collapsing continuous dose responses into scalar IC50 values hides biphasic and hormetic behavior, and from the biology side by Transcriptomic, Biochemical, and Morphological Study Reveals the Mechanism of Inhibition of Pseudopestalotiopsis camelliae-sinensis by Phenazine-1-Carboxylic Acid, which profiled 3,388 differentially expressed genes after PCA exposure (EC50 = 11.23 µg/mL) but at one arbitrary concentration, so its N-glycan and oxidative-phosphorylation findings cannot be ordered in time or dose.
Baselines
The "do nothing" reference point is the field's default: a four-parameter logistic fit to mortality counts alone, producing an LC50 and nothing else. This is the pure-toxicology answer that every cited bioassay study delivers, and it is what our pathway curves must beat on explanation, not on curve fit — mortality 4PL will, by construction, fit mortality better than anything we build, because that is all it models.
The most direct competitor is the status-quo experimental design: single-dose transcriptomics with standard overrepresentation analysis, exactly as Transcriptome-Based Identification and Characterization of Genes Associated with Resistance to Beta-Cypermethrin in Rhopalosiphum padi did (one LC50 challenge, edgeR, GO and KEGG enrichment, qRT-PCR validation). This baseline is nested inside our design for free: the LC50 point of our titration, analyzed with Fisher's exact test at the same sequencing depth and mapping pipeline, reproduces their study exactly. Beating it means recovering their headline findings at the LC50 point — the P450 xenobiotic-metabolism enrichment, the 37 cuticle protein genes, the ABC transporters — while additionally ordering those same pathways by activation dose; the single-dose design cannot produce a PAC50 at all, so the comparison is whether the extra ~40 libraries buy a mechanistic ordering that changes what a practitioner would conclude (for example, whether cuticle genes are early-compensatory or late-collapse).
The strongest method-level competitor is the best existing enrichment method applied per dose point without any curve fitting: pareg run independently at each concentration, plus a separate 4PL mortality curve Coherent pathway enrichment estimation.... It reported the highest mean precision–recall AUC against Fisher, topGO, MGSA, and blitzGSEA on its own benchmarks (no numeric value reported), though at a runtime of up to ~30 minutes in parallelized mode. We will run it on our gradient data with its default penalties plus one cross-validated sweep, same data and compute as our pipeline. We will also train a CPA-style latent dose model on our pooled gradient libraries, since the Compositional Perturbation Autoencoder is the strongest published model of transcriptomes under continuous doses — it achieved out-of-distribution R² of 0.92 for Nutlin and 0.94 for BMS-345541 in sci-Plex 2, against linear-interpolation baselines of 0.85 and 0.89 Learning interpretable cellular responses to complex perturbations in high-throughput screens — and hold out doses to test whether it can interpolate a pathway titration curve from ~50 libraries. Beating these means narrower bootstrap confidence intervals on PAC50 and correct rank-ordering of the three synergy coefficients, evaluated on held-out mixture conditions.
The key ablations isolate the two ingredients. Dropping inter-pathway regularization (Fisher per dose, then 4PL on raw pathway scores) tests whether pareg-style coherence matters here or only adds a 30-minute runtime. Dropping the gradient (single-dose pareg, no curve) tests whether better enrichment alone, without dose resolution, delivers the same mechanistic ordering. If either ablation matches the full pipeline, the corresponding ingredient is cut and we report that plainly.
Why not just…?
Why not just fit 4PL or probit curves to mortality and skip the transcriptomics? Because that machinery is exactly what we propose to reuse — applied to a different readout. The 4PL fit is the right tool at this sample size; the authors of ARCANet themselves note that below roughly 2,000 data points, 3- and 4-parameter parametric models are more robust than deep networks Predicting Dose-Response Curves with Deep Neural Networks. Our disagreement with the pure-toxicology default is not about the fitting tool but about what is being fit: a mortality LC50 tells you the resistant aphid survives 4,588-fold more beta-cypermethrin, but not how, and therefore offers no lever for breaking that resistance. The pathway titration is the cheapest way to attach a mechanism to that number.
Why not just do single-dose RNA-seq with standard enrichment, as the field already does? This is the real incumbent, and our design contains it as a baseline rather than arguing past it. The concrete deficiency is structural: with one dose you cannot order pathways in dose, so compensatory detoxification and terminal collapse land in the same enriched list. The decisive experiment is whether early-cluster and late-cluster pathways are functionally and predictively different — for instance, whether knocking down one early-cluster gene shifts the LC50 while knocking down a late-cluster gene does not. If the clusters are not separable, the gradient bought nothing and the incumbent wins.
Why not just run targeted enzyme activity assays for P450s, GSTs, and carboxylesterases? This is the most serious alternative because it is an order of magnitude cheaper, and we treat it as a genuine threat rather than a straw man. The evidence against it as a complete substitute is that the transcriptomic studies in this corpus repeatedly found the important players outside the canonical assay panel: the R. padi study found 37 cuticle protein genes and 6 ABC transporters alongside the P450s, which a GST/CarE activity assay would have missed entirely Transcriptome-Based Identification... Beta-Cypermethrin in Rhopalosiphum padi (Transcriptome-Based Identification and Characterization of Genes Associated with Resistance to Beta-Cypermethrin in Rhopalosiphum padi (Hemiptera: Aphididae)), and the PCA study found N-glycan biosynthesis as the target, invisible to oxidative-stress assays Transcriptomic, Biochemical... Phenazine-1-Carboxylic Acid. Our design therefore includes enzyme activity assays at PAC50-adjacent doses as validation, and we state the falsifier honestly: if three-dose enzyme panels recapitulate the pathway ordering and predict the synergy coefficients as well as the full titration, then for the canonical detox families the transcriptome is overkill, and the proposal's scope should shrink to discovering only the non-canonical pathways.
Why not just port CPA or ARCANet to pest data? Because the data does not exist at the scale they need, and their own authors say so: ARCANet's authors concede parametric models win under ~2,000 points Predicting Dose-Response Curves with Deep Neural Networks, and the oncology synergy models were trained on tens of thousands of combinations — DeepSynergy on 23,062 samples, reaching mean squared error 255, Pearson correlation 0.73, and AUC 0.90 DeepSynergy: predicting anti-cancer drug synergy with Deep Learning, and Pisces on over 100,000 combinations Pisces: A multi-modal data augmentation approach for drug combination synergy prediction. Agricultural resistance datasets run to dozens of libraries. Rather than argue, we run a CPA-style model on our pooled data as a baseline; the expectation is underfitting, and if it instead interpolates well, that is a finding worth reporting and a pivot worth making.
Why not just compute interaction coefficients (R or FIC) from bioassays, as toxicologists already do? Because those coefficients are descriptive. The Elsholtzia ciliata study itself notes that essential-oil composition varies across regions, so an empirical R measured in one harvest does not transfer to the next Bioactivities and Synergistic Effect.... A mechanism — here, the specific rightward shift of a detoxification pathway's activation curve — is what lets you optimize ratios, anticipate cross-resistance, and pick a replacement synergist when composition drifts. Our decisive check is whether the predicted pathway shift rank-orders with the three measured coefficients (0.98, 1.36, 3.15); if it does not, the mechanistic story is wrong and the empirical coefficient remains the only game in town.
Proposed Experiments and Decisive Checks
Experiment 1 runs the R. padi titration (SS and Beta-R, beta-cypermethrin, seven strain-specific doses plus control, three replicates, ~48 libraries) and asks three questions. Does the design reproduce the single-dose findings at the LC50 point (P450 xenobiotic-metabolism enrichment, cuticle and transporter gene counts within the reported ranges)? Do pathways cluster into separable early (below LC30) and late (above LC80) groups by PAC50, with bootstrap confidence intervals that do not overlap? And does the ratio LC50/PAC50 for the top early cluster differ between SS and Beta-R by an amount commensurate with the resistance ratio of 4588.48? Validation is qRT-PCR on genes from each cluster, following the original study's protocol.
Experiment 2 runs the synergy titration on L. serricorne with R-carvone, DL-limonene, and their 1:1 mixture under both fumigant and contact routes, plus the enantiomer mixture. The pre-registered prediction is that mixture exposure right-shifts or suppresses the xenobiotic-metabolism PAC50 relative to R-carvone alone, that the magnitude of the shift rank-orders with the empirical coefficients (largest for the enantiomer fumigant pair at R = 3.15, none for contact at R = 0.98), and that damage-associated pathways left-shift correspondingly. The decisive statistic is a Spearman correlation between predicted pathway shift and measured R across the three conditions.
Experiment 3 is the causal check: RNAi knockdown of one early-cluster candidate in R. padi, with the LC50 re-measured — precedent for developmental and physiological RNAi readouts in pest insects comes from the fatty-acid-synthase knockdown study in Spodoptera litura The Key Role of Fatty Acid Synthase in Lipid Metabolism and Metamorphic Development in a Destructive Insect Pest, Spodoptera litura (Lepidoptera: Noctuidae), where silencing impaired larval survival and pupation (no quantitative survival figure is reported in the file). An early-cluster knockdown that shifts the LC50, while a matched late-cluster knockdown does not, is the strongest evidence that the PAC50 ordering carries causal information rather than correlation.
Experiment 4 extends the design to the S. miscanthi sulfoxaflor system and tests whether shared early-activating pathways under sulfoxaflor and a pyrethroid predict the observed cross-resistance pattern (RR 199.8 vs 14.5–42.1) that the original study could only hypothesize. A known risk throughout is annotation quality in non-model genomes — the Aglais urticae assembly yielded 77,624 unigenes with many uncharacterized Analysis of Genes Associated with Feeding Preference and Detoxification in Various Developmental Stages of Aglais urticae — so we will map by orthology to curated aphid and Drosophila references and report the uncharacterized fraction per dose point, since a rising uncharacterized fraction at high doses would itself be a finding about undocumented stress programs. As a cost-reduction path for future scale-up, label-free imaging in the spirit of SIC50 — which reached 99.8% classification accuracy for high-concentration treatments and matched staining-based IC50 values within a 2-fold range SIC50: Determination of IC50 by an optimized Sobel operator and a vision transformer — could replace viability endpoints between sequencing points, though it cannot replace the transcriptomic readout itself.
References
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- funkea: Functional Enrichment Analysis in Python
- Transcriptome-Based Identification and Characterization of Genes Associated with Resistance to Beta-Cypermethrin in Rhopalosiphum padi (Hemiptera: Aphididae)
- Bioactivities and Synergistic Effect of Elsholtzia ciliata Essential Oil and Its Main Components against Lasioderma serricorne
- Fitness Cost of the Field-Evolved Resistance to Sulfoxaflor and Multi-Insecticide Resistance of the Wheat Aphid Sitobion miscanthi (Takahashi)
- Predicting Dose-Response Curves with Deep Neural Networks
- Transcriptomic, Biochemical, and Morphological Study Reveals the Mechanism of Inhibition of Pseudopestalotiopsis camelliae-sinensis by Phenazine-1-Carboxylic Acid
- Learning interpretable cellular responses to complex perturbations in high-throughput screens
- DeepSynergy: predicting anti-cancer drug synergy with Deep Learning
- Pisces: A multi-modal data augmentation approach for drug combination synergy prediction
- The Key Role of Fatty Acid Synthase in Lipid Metabolism and Metamorphic Development in a Destructive Insect Pest, Spodoptera litura (Lepidoptera: Noctuidae)
- Analysis of Genes Associated with Feeding Preference and Detoxification in Various Developmental Stages of Aglais urticae
- SIC50: Determination of IC50 by an optimized Sobel operator and a vision transformer