When a mosquito population survives insecticide spraying, biologists measure the response with bulk RNA sequencing: grind whole insects (or dissected midguts) into one tube and read out an average gene-expression profile. That average hides a question that determines how resistance actually works. If detoxifying enzymes like cytochrome P450s are strongly upregulated, is every cell turning them on a little, or is a rare, specialized cell type — such as oenocytes, the insect liver analog — turning them on enormously, or even expanding in number? These three mechanisms have different evolutionary costs, different cross-resistance footprints, and demand different countermeasures, yet current surveillance cannot tell them apart. Meanwhile, the field that solves exactly this problem for human tumors — computational deconvolution, which uses a single-cell reference atlas to unmix a bulk measurement into cell-type proportions and per-cell-type expression — has never been applied to an insect.
This proposal builds the missing bridge. We will construct the first single-nucleus reference atlas of an insect under insecticide selection, use orthology between Drosophila and mosquito vectors to transfer that reference across species, and deploy a hierarchy-aware, integer-count deconvolution pipeline on bulk surveillance transcriptomes from resistant and susceptible mosquito populations. The payoff is a decomposition of resistance into compositional change versus cell-autonomous induction — information that today's single-locus control models and kdr genotyping assays structurally cannot provide.
Core Claim
The central hypothesis is twofold. First, methodological: reference-guided deconvolution can be transferred to non-mammalian, whole-insect bulk transcriptomes — but only if two failure modes are addressed explicitly. Insect whole-body mixtures are dominated (often above 70–90%) by structural tissues such as indirect flight muscle, which triggers the "purity trap" documented in the breast-cancer deconvolution benchmark, where standard algorithms systematically mislabel related minority cell types when one lineage dominates Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. And low-abundance resistance genes, such as voltage-gated sodium channels, are distorted when discrete molecule counts are treated as continuous concentrations, a bias the integer-valued Count Bridges process was designed to fix Count Bridges enable Modeling and Deconvolving Transcriptomic Data. Second, biological: in pyrethroid-selected mosquitoes, the bulk-detoxification signal is not uniform induction but is concentrated in specific cell types — and the partition between per-cell induction and compositional expansion predicts the resistance ratio and cross-resistance breadth measured in the field.
Formally, for a bulk count vector $y$ we fit $y = \sum_{c=1}^{C} p_c\, x_c + \epsilon$, where $p_c$ is the proportion of cell type $c$ and $x_c$ its integer-valued expression profile, with the cell types arranged on a lineage tree so that closely related types are separated only when the data support it (the tree-guided step that SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References introduced to handle dominant lineages). The insect-specific innovation is that the reference profiles $x_c$ are not from the same species: they come from a Drosophila atlas remapped onto mosquito genes through one-to-one orthologs, the mapping that FlyMine: an integrated database for Drosophila and Anopheles genomics already formalizes. Cross-species deconvolution is not fantasy: GTM-decon: guided-topic modeling of single-cell transcriptomes enables sub-cell-type and disease-subtype deconvolution of bulk transcriptomes deconvolved human pancreatic bulk data with a mouse single-cell reference — but only between mammals, never across ~250 million years of dipteran divergence, and never under toxic selection.
Research Program
The program has three components: a reference atlas, a deconvolution pipeline, and a surveillance application that closes the loop with population modeling.
The reference atlas extends the Aging Fly Cell Atlas, which profiled 868,209 nuclei across 163 cell types in adult Drosophila but examined only chronological aging under benign laboratory conditions Aging Fly Cell Atlas Identifies Exhaustive Aging Features at Cellular Resolution. We will generate matched single-nucleus RNA-seq profiles of Drosophila populations evolved under sublethal pyrethroid exposure alongside unexposed controls, across developmental stages, using single-nucleus rather than single-cell protocols precisely because the fat body — a primary detoxification organ — is fragile, polyploid, and ruptures under dissociation, a bias SCDC identifies as the main way droplet platforms lose large cells. This atlas yields reference profiles for xenobiotic-stressed cell states that no existing insect atlas contains, and it directly tests whether insecticide-selected strains show the same cell-type-specific vulnerability that aging showed (the original atlas found adipose and oenocyte populations aged fastest, and that fat body cells divide nuclei without cell division — a mechanism that could plausibly generate compositional resistance).
The pipeline combines three published ingredients in a new regime. From SCDC we take ensemble weighting over multiple references (laboratory-reared and field-derived Drosophila references, plus an eventual mosquito-derived reference) to control cross-species and cross-laboratory batch effects. From Count Bridges we take the integer-valued bridge process with Bessel posterior updates, because per-cell detoxification transcripts in rare cell types sit at counts of tens, where continuous normalization is most distorting. From GTM-decon we take marker-free topic structure, because insecticide exposure is exactly the kind of dynamic stress that shifts cell states away from any predefined marker panel. The output for each field sample is a vector of cell-type proportions and per-cell-type expression, plus an explicit flag when a bulk sample contains a state absent from the reference — the known blind spot of reference-guided methods that Cross-modal transfer learning for mapping bulk transcriptomes at cellular level inherits from its mammalian ontology and that we flag rather than hide.
The application targets bulk RNA-seq from wild-caught resistant and susceptible mosquito populations, of the kind generated by surveillance programs such as the Uganda malaria monitoring cohort, where cases fell 84% below baseline under indoor residual spraying (adjusted IRR = 0.16) and then rebounded to 39% above baseline after the active ingredient was switched Resurgence of malaria in Uganda despite sustained indoor residual spraying and repeated long lasting insecticidal net distributions. Populations like the Jining Culex pipiens pallens strains, with resistance ratios hundreds of times baseline Investigation of mosquito larval habitats and insecticide resistance in an area with a high incidence of mosquito-borne diseases in Jining, Shandong Province, provide the extreme-resistance end of the spectrum. The decisive biological readout is a two-way decomposition: compare resistant versus susceptible populations in (a) per-cell-type detoxification expression within, say, oenocytes and midgut enterocytes, controlling for (b) the proportions of those cell types. Finally, we feed the estimated per-cell-type resistance costs into the compartmental model of Impact of insecticide interventions on the abundance and resistance profile of Aedes aegypti, replacing its single diallelic locus with cell-type-level metabolic costs, and test whether the refined model better predicts the non-monotone outcome it documented — that five insecticide applications per year yielded higher cumulative abundance than four.
Proposed Experiments
Experiment one builds the benchmark. We construct pseudo-bulk mixtures by sampling known numbers of nuclei from the new pyrethroid atlas and the original Aging Fly Cell Atlas, spanning purity levels from 5% to 95% dominant lineage (mirroring the 38,000-mixture, 100,064-cell design of the breast-cancer benchmark Performance of tumour microenvironment deconvolution methods...), and spanning cross-species difficulty by mixing Drosophila references with simulated mosquito-mapped bulk. Metrics are proportion error (RMSE and Pearson correlation against known mixing ratios) and, critically, the false-positive rate on minority detoxification cell types at high purity — the specific failure the purity trap predicts. A small set of experimentally mixed samples with known cell-type ratios, in the spirit of the three-cell-line mixture where SCDC achieved a Pearson correlation of 0.991 with ground truth SCDC, provides an in-vitro anchor; whether any method reaches that level on insect mixtures is exactly what the experiment measures.
Experiment two is the biological decomposition on surveillance data. For paired resistant/susceptible populations (pyrethroid-selected Aedes and carbamate-exposed Anopheles), we test whether the bulk P450 signal — the family repeatedly implicated by whole-organism assays such as the phoxim microarray study, which could not localize its signal Gene expression analysis from phoxim-induced domesticated silkworm (Bombyx mori) by whole-genome oligonucleotide microarray and by stage-specific CYP6B6 profiling Expression of cytochrome P450 CYP6B6 in the different developmental stages of the insect Helicoverpa armigera (Lepidoptera: Noctuidae)) decomposes into induction, composition, or both. The decisive check is orthogonal validation by fluorescence in situ hybridization for the top candidate cell type in a handful of specimens: if the model attributes the signal to oenocyte expansion, oenocyte counts should be visibly higher in resistant insects.
Experiment three tests predictive value. Using held-out populations with measured resistance ratios, we ask whether the deconvolved features (per-cell-type induction plus composition) predict cross-resistance to an unencountered insecticide class better than bulk expression alone or kdr genotyping alone — the latter being the surveillance standard whose insufficiency the Uganda rebound illustrates, and whose mechanism-level blind spot the metabolic and antioxidant literature documents Effect of RNAi Targeting SeGrx1 on the Cytotoxicity and Insecticide Susceptibility of Camptothecin in Spodoptera exigua Study on the cross-resistance of Aedes albopictus (Skuse) (Diptera: Culicidae) to deltamethrin and pyriproxyfen.
Baselines
The do-nothing reference point is raw bulk differential expression without deconvolution — the current practice in insect resistance functional genomics, exemplified by the phoxim microarray study Gene expression analysis from phoxim-induced domesticated silkworm... and the midgut-dissected Bacillus thuringiensis toxin RNA-seq Transcriptomic Analysis of Aedes aegypti in Response to Mosquitocidal Bacillus thuringiensis LLP29 Toxin. It is the strongest competitor on cost and throughput and sets the biological question: any deconvolution method must beat it at attributing the signal, since on detecting "the population changed" bulk alone already succeeds. We will run it on the identical pseudo-bulk and surveillance data with the same differential-expression tooling, and beating it means recovering per-cell-type structure that bulk provably conflates — demonstrated by constructing pseudo-bulks whose identical total P450 signal arises from different (induction vs. composition) generating mechanisms, which bulk DE cannot distinguish by construction.
The strongest existing deconvolution competitor is Count Bridges, the integer-valued bridge process that outperformed CIBERSORTx and MuSiC on mammalian transcriptomic deconvolution benchmarks Count Bridges enable Modeling and Deconvolving Transcriptomic Data — although on its published benchmarks no numerical error value is reported on Lacuna, only the qualitative ranking, so our insect benchmark will be the first head-to-head with numbers. It is strongest because it shares our integer-count commitment, removing that as a differentiating ingredient. We will run it with the same Drosophila reference, the same ortholog mapping, the same compute and tuning budget, and the same pseudo-bulk splits; our pipeline beats it only if the added hierarchical tree structure and multi-reference ensemble weighting reduce minority-cell-type false positives at high purity — the regime where flat, single-reference methods are documented to fail Performance of tumour microenvironment deconvolution methods.... Off-the-shelf CIBERSORTx and MuSiC run under the same conditions complete the comparison, as does BayesPrism/DWLS, which on the breast-cancer benchmark showed the lowest combined false positives and false negatives among nine algorithms, again without a reported numerical score on Lacuna.
The key ablation isolates cross-species transfer: our full pipeline versus the identical pipeline given a same-species (eventually mosquito-derived) reference. This cleanly measures how much accuracy the ortholog-mapped Drosophila reference costs — the quantity that determines whether the approach is deployable today, before any mosquito single-cell atlas exists. A second ablation removes the tree, collapsing the hierarchy to a flat cell-type list, isolating the purity-trap defense.
Why not just…?
Why not just microdissect the midgut or fat body instead of deconvolving? Microdissection reduces, but does not remove, the heterogeneity problem: the dissected-midgut Bt toxin study still averaged enterocytes, stem cells, and secretory cells into one profile and could not attribute the response Transcriptomic Analysis of Aedes aegypti in Response to Mosquitocidal Bacillus thuringiensis LLP29 Toxin. More practically, the fat body is interwoven with cuticle and tracheoles and resists clean isolation, and dissection of thousands of field-caught specimens — the scale surveillance operates at — is infeasible. Microdissection is a fair validation tool (Experiment two uses it), but as the primary method it fails on scale and on the minority-cell problem: oenocytes are scattered cells, not a dissectible tissue.
Why not just run single-nucleus RNA-seq on every surveillance sample? This is the real threat to the proposal's long-term relevance, and we should say so: as single-nucleus costs fall, deconvolution's value shrinks. Today, though, per-sample cost and logistics rule it out for routine monitoring across hundreds of sites, and droplet protocols systematically under-sample exactly the large, fragile cells (fat body) where detoxification matters, a bias documented by SCDC. The honest settlement: our benchmark reports, per method, both accuracy and cost per sample, so the community can see the accuracy-per-dollar frontier directly rather than take our word for it.
Why not just run CIBERSORTx or MuSiC off the shelf on insect count matrices? The documented failure modes align precisely with the insect regime: whole-body mixtures with a >70% dominant lineage trigger the purity trap that the breast-cancer benchmark exposed in standard tools Performance of tumour microenvironment deconvolution methods..., single-reference methods destabilize under batch effects, and stress-shifted cell states break marker-based assumptions GTM-decon. But this objection might be partly right — if off-the-shelf tools with a same-species reference perform adequately on our benchmark, that is a useful finding, and we will report it. The experiments settle it rather than argue it, which is why the baselines run under identical conditions.
Why not just genotype known kdr mutations? Target-site genotyping is cheap and entrenched, and for pyrethroid-specific resistance it partially works. But the Uganda rebound after an active-ingredient switch — from 84% suppression to 39% above baseline — shows known-marker surveillance failing to predict sudden operational loss Resurgence of malaria in Uganda..., and metabolic plus antioxidant mechanisms underlie much high-level and cross-class resistance Expression of cytochrome P450 CYP6B6... Effect of RNAi Targeting SeGrx1.... Experiment three makes this empirical: if kdr genotyping predicts cross-resistance as well as deconvolved features, the proposal's surveillance claim dies, and we will say so.
Decisive Checks
The proposal lives or dies on three outcomes. The benchmark must show that some method recovers minority detoxification cell types at insect-realistic purity on cross-species references; if none does, the negative result still redirects the field toward building mosquito-native atlases first. The in-situ validation must confirm at least one predicted cell-type attribution; without it, the decomposition is unmoored from biology. And the deconvolved features must predict cross-resistance and the population-level rebound dynamics Impact of insecticide interventions... better than bulk or genotyping baselines; that predictive margin, not methodological elegance, is the reason for the whole program.
References
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- FlyMine: an integrated database for Drosophila and Anopheles genomics
- GTM-decon: guided-topic modeling of single-cell transcriptomes enables sub-cell-type and disease-subtype deconvolution of bulk transcriptomes
- Aging Fly Cell Atlas identifies exhaustive aging features at cellular resolution
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Resurgence of malaria in Uganda despite sustained indoor residual spraying and repeated long lasting insecticidal net distributions
- Investigation of mosquito larval habitats and insecticide resistance in an area with a high incidence of mosquito-borne diseases in Jining, Shandong Province
- Impact of insecticide interventions on the abundance and resistance profile of Aedes aegypti
- Gene expression analysis from phoxim-induced domesticated silkworm (Bombyx mori) by whole-genome oligonucleotide microarray
- Expression of cytochrome P450 CYP6B6 in the different developmental stages of the insect Helicoverpa armigera (Lepidoptera: Noctuidae)
- Effect of RNAi Targeting SeGrx1 on the Cytotoxicity and Insecticide Susceptibility of Camptothecin in Spodoptera exigua
- Study on the cross-resistance of Aedes albopictus (Skuse) (Diptera: Culicidae) to deltamethrin and pyriproxyfen
- Transcriptomic Analysis of Aedes aegypti in Response to Mosquitocidal Bacillus thuringiensis LLP29 Toxin