Two of the most informative measurements in cancer biology are each blind in exactly the way the other is not. Bulk RNA sequencing of a tissue biopsy tells you what mixture of cells is present — cancer, stromal, immune — but says nothing about the physical state of those cells' genomes. The comet assay, a single-cell gel electrophoresis technique in which a cell's broken DNA fragments literally migrate through agarose under an electric field to form a fluorescent "tail" behind an intact "head," measures that physical state directly in individual cells: strand breaks, repair speed, apoptotic fragmentation. But because the assay dissolves every membrane and strips every protein before imaging, a comet from a $CD8^+$ T cell is indistinguishable from a comet from a cancer stem cell, and the only current fix is to physically sort pure subpopulations before the assay — slow, stressful to the cells, and disruptive to time-course experiments. The same blind spot runs through perturbation genetics: siRNA knockdowns — small RNA duplexes that lure the cell's own silencing machinery into destroying a chosen messenger RNA — are almost always evaluated in a single cell line in isolation, even though we know from pan-cancer cohorts that the same protein can have opposite effects in different immune contexts.
This proposal joins the two blind measurements into one. The idea: run the comet assay directly on a mixed population, treat the resulting distribution of comet phenotypes as a statistical mixture, and unmix it using cell fractions supplied by transcriptomic immune deconvolution — the family of methods that computationally separates a bulk RNA mixture into cell-type proportions — applied to a matched sample of the same culture. The output is per-lineage DNA damage burden and repair kinetics from an unsorted population. We then use this instrument to ask the question monoculture knockdown studies never could: does silencing a candidate gene produce a different DNA-damage phenotype in tumor cells when immune cells are present?
Core Claim
The distribution of comet metrics in a mixed cell population can be unmixed into lineage-specific damage and repair-kinetic profiles, provided (a) each lineage's damage distribution is calibrated in pure-population assays and (b) the mixture fractions are supplied by transcriptomic deconvolution rather than fitted blind. This converts the comet assay from a lineage-blind single-cell measurement and deconvolution from a function-blind fraction estimate into a single joint readout: who is in the dish, and how broken is each population's genome, over time. Applied to siRNA perturbation of genes whose function is suspected to flip with immune context, it can test context-dependent protein function in a regime that no existing study reaches: co-culture, physically measured, per-lineage, kinetic.
The closest existing work sits at pairwise intersections and leaves this joint step open. Hao Huang et al. combined siRNA and the comet assay, showing dose-dependent apoptotic comets after c-myc silencing in Colo 320 cells — but in one monoculture, scored manually, with no immune cells and no transcriptomics (The single-cell gel electrophoresis assay to determine apoptosis induced by siRNA in Colo 320 cells). Xi Zhang et al. combined deconvolution, pan-cancer duality, and siRNA: OAS3 is upregulated in 32 of 34 cancer types, predicts poor survival in glioma and leukemia but improved survival in melanoma, tracks with M1 macrophage and Th2 infiltration, and its knockdown kills THP-1 monocytic leukemia cells — yet the knockdown was done only in monoculture, with no comet assay to test whether the DNA-damage associations are physically real (Pan-cancer multi-omics profiling of OAS3 reveals its immunological and prognostic associations across human cancers). Yuan Meng et al. used siRNA and comet assays to show SIK2 knockdown cripples homologous-recombination repair in colorectal cancer lines, but explicitly without immune deconvolution and without immune co-culture (Salt‐inducible kinase 2 confers radioresistance in colorectal cancer by facilitating homologous recombination repair). Chung-Wah Wu et al. showed the sharpest version of the biological tension: miR-18a suppresses ATM, which both accelerates recurrence and hypersensitizes tumors to genotoxic drugs — a functional switch measured by comet repair kinetics, but correlated against bulk tumors never deconvolved (MicroRNA-18a Attenuates DNA Damage Repair through Suppressing the Expression of Ataxia Telangiectasia Mutated in Colorectal Cancer). No study in the file unmixes comet distributions, and none measures knockdown phenotypes per-lineage inside an immune mixture.
Method: Damage-Profile Unmixing
For each comet $i$ we extract a feature vector $y_i$ — percentage of DNA in the tail, tail length, Olive tail moment (tail-DNA fraction times the distance between head and tail centers of gravity), and the morphology class that Huang et al. showed separates apoptotic "teardrop" comets from ordinary damage. The population-level model is a constrained mixture:
$$p(y_i)=\sum_{c=1}^{C}\pi_c\, f_c(y_i;\theta_c),\qquad \pi_c\ge 0,\ \sum_c \pi_c=1,$$
where $f_c$ is lineage $c$'s damage distribution, calibrated by running the comet assay on each pure lineage under the same genotoxic dose, and $\pi_c$ is lineage $c$'s fraction. The fractions come from SCDC-style ensemble deconvolution (SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References) on bulk RNA-seq of a split of the same mixed sample — SCDC is the right source because it pools multiple single-cell references and was the only method to recover a known clinical association in the islet cohort at $p = 0.0019$, and it hit a Pearson correlation of $r = 0.991$ against physically known 6:3:1 mixture fractions. With $\pi$ fixed by deconvolution, the remaining task is not clustering but attribution: posterior assignment of each comet to a lineage, followed by per-lineage summaries at each time point. Repair kinetics per lineage reduce to fitting exponential decay of mean tail moment across staggered recovery intervals (0, 2, 6, 12, 24 hours, following the design of Wu et al.), yielding a repair half-time $t_{1/2,c}$ for each population in the same dish.
Two honest caveats shape the design. First, identifiability: if two lineages have nearly identical calibrated damage distributions, no unmixing method can separate them; Phase 1 exists precisely to map when the method works and when it does not, and we will report the calibration discrepancy needed for reliable attribution rather than assume it. Second, the image front-end must not be the bottleneck: we build on the HiComet pipeline, whose adaptive binarization reached 0.979 recall on its 35 expert-annotated golden sets (702 Jurkat comets after hydrogen peroxide) and whose support-vector-machine classifier distinguished normal, apoptotic, and necrotic cells at 0.904 accuracy (HiComet: a high-throughput comet analysis tool for large-scale DNA damage assessment) — morphology class is a discrete emission that materially helps attribution, since apoptotic fractions differ sharply by lineage.
Research Program
Phase 1 builds and validates the unmixing on controlled mixtures, mirroring the SCDC validation logic. We mix three human lines — a monocytic line (THP-1), a T-cell line (Jurkat, matching the HiComet golden sets), and a carcinoma line — at known ratios bracketing the 6:3:1 design, challenge with hydrogen peroxide, and run the 0–24 hour comet time course on the unsorted mixture. Ground truth comes from parallel fluorescence-activated-cell-sorting of replicate mixtures into pure fractions followed by comets on each. Success is defined before running: per-lineage mean Olive tail moment recovered with Pearson $r \ge 0.9$ against the sorted reference at each time point, per-lineage fraction recovered within the deconvolution's own error, and repair half-times within 20% of sorted values.
Phase 2 applies the instrument to the context-dependence question that motivates it. The primary target is OAS3, chosen because the Zhang et al. pan-cancer study supplies exactly the prerequisite pattern — deconvolution-linked prognostic duality (poor survival in LAML and glioma, improved survival in melanoma, correlated with macrophage and Th2 infiltration), a DNA-damage-response pathway association, and a monoculture-only siRNA result in THP-1 that no one has tested in an immune context. We silence OAS3 in the carcinoma line before mixing (so the siRNA does not hit the immune cells), then co-culture with macrophage-differentiated THP-1 cells and activated T cells across a range of mixture fractions — including the extremes that mimic the melanoma-like immune-rich and leukemia-like immune-poor regimes — and read out per-lineage damage and repair kinetics after genotoxic challenge. The prediction, generated from the pan-cancer inversion, is that the direction of OAS3-silencing's effect on tumor-cell damage burden changes sign or magnitude across immune fraction. The secondary target is the Wu et al. tension made causal: siRNA knockdown of ATM in colorectal cancer cells, with and without immune co-culture, testing whether immune presence shifts the instability-versus-hypersensitivity balance that miR-18a was only correlatively shown to set.
Phase 3 addresses the compensatory-loop problem that monoculture knockdowns miss, exemplified by Yating Liu et al., where E2F1 silencing lowered the cisplatin IC50 by roughly 50% in Y79 and WERI-RB1 cells and raised apoptosis to about 40%, yet boosting the downstream CENPE rescued the phenotype (<i>E2F1</i> Silencing Enhances Cisplatin-induced Apoptosis and DNA Damage in Retinoblastoma: <i>CENPE</i> Overexpression as a Compensatory Mechanism). Because our method measures both lineages in the same dish at every time point, it can detect whether compensation arrives from the tumor compartment or from a paracrine immune response — something neither that study nor any monoculture design can see.
Baselines
The naive reference point is the pooled comet assay as it is practiced today: run comets on the mixed population, report one aggregate tail-moment distribution, and attribute nothing. Its failure mode is quantifiable — in mixtures where one lineage is apoptotic and another is repairing, the pooled distribution confounds the two, and we will report the attribution error of pooled summaries against sorted ground truth in Phase 1 as the "do nothing" number. The strongest existing method is sort-then-comet: fluorescence-activated cell sorting of each subpopulation followed by the comet assay, the standard the field currently accepts as the only route to lineage-resolved comets. We will run it on split samples of every mixture in Phase 1 and every co-culture in Phase 2, with the same genotoxic doses, the same time points, and the same HiComet image pipeline, so the only difference is the sorting step. Beating it does not mean higher accuracy — sorting is close to ground truth by construction — but matching its per-lineage estimates within the thresholds above while removing the sort: the sort costs hours during which repair kinetics are actively progressing (fatal for early time points), imposes mechanical stress that itself damages DNA, and requires pre-selected markers. We will quantify each of these costs directly by comparing sorted-fraction comets to unsorted unmixed comets on identical samples. The key ablation isolates the deconvolution ingredient: run the same unmixing with fractions set to uniform priors, or fitted blind by expectation-maximization, instead of taken from SCDC deconvolution of the matched bulk RNA-seq. If SCDC-informed fractions do not beat uniform and blind priors on per-lineage recovery, the central claim fails, and we will report that. A second ablation removes the calibrated pure-lineage distributions $f_c$ and fits them freely, testing whether calibration is genuinely necessary or a convenience.
Why not just…?
Why not just FACS-sort before the comet assay? This is the real gold standard and we treat it as the validation reference, not a strawman. What the file's evidence shows is its cost, not its absence of merit: sorting adds hours of handling to a time course in which Wu et al. measured tail-moment contraction across minutes-to-hours windows, and the assay's own open-problem list names pre-sorting as the only current workaround for lineage blindness (HiComet). The decisive experiment is in Phase 1: if sorted and unsorted-unmixed per-lineage kinetics agree everywhere, including at the 0- and 2-hour points where sorting artifacts should bite hardest, then the method's value is purely throughput, and we will say so. If they diverge at early points, the sort itself is perturbing the measurement, and the unsorted route is the more faithful one.
Why not just use γ-H2AX phospho-flow with lineage markers? This is the most serious alternative: antibody-based flow cytometry of the phosphorylated histone H2AX, co-stained with lineage markers, gives a single-cell, lineage-resolved readout of double-strand-break signaling without any unmixing. Both Liu et al. and Meng et al. ran γ-H2AX alongside comets, which shows the field treats them as complementary, not redundant. The concrete shortfall is coverage: γ-H2AX is a signaling mark for one lesion class, while the comet assay resolves single-strand breaks, alkali-labile sites, crosslink-associated fragmentation, and the apoptotic teardrop morphology that Huang et al. used as an unambiguous programmed-death readout. We will run phospho-flow in parallel on every co-culture time point and report where the two lineage-resolved readouts agree and where they diverge — if they agree everywhere, our claim narrows to the wider lesion spectrum, and we will state that honestly.
Why not just use single-cell RNA sequencing of damage-response genes? Sequencing DDR gene expression per cell would give identity cheaply and richly. But this is exactly the limitation the deconvolution literature already documents — fractions and expression scores without physical function (SCDC). The file's clearest evidence that transcription is a poor proxy for physical damage is Kell et al., where low-dose rapamycin in ageing human T cells produced a 3-fold survival gain 24 hours after genotoxic exposure, measured physically by comet assay, while the transcriptional readout was a change in p21 (Rapamycin Exerts Its Geroprotective Effects in the Ageing Human Immune System by Enhancing Resilience Against <scp>DNA</scp> Damage); and in Wu et al., ATM mRNA correlated with miR-18a at only $r = -0.4562$, far too weak to substitute for the functional repair measurement. Still, scRNA-seq is the right tool for discovering which lineage expresses a dual-function gene, and we use it only in that supporting role.
Why not just model context computationally instead of measuring it? CDBridge conditions biological predictions on tissue embeddings but only as static GTEx tissue categories, and ProtQueSt conditions on text queries bounded by existing annotations; both explicitly exclude dynamic mixture proportions and functional readouts (CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling, ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation). Our context variable — a continuous immune fraction measured on the same sample as the phenotype — is precisely what they lack, and we measure rather than predict it. CDBridge's reported $R^2$ of 0.387 on the GTEx expression benchmark also suggests context-aware prediction is far from settled, which argues for experiments before models.
Decisive Checks
Three results would settle the proposal's fate. First, on the controlled 6:3:1-style mixtures, per-lineage mean tail moment recovery of $r \ge 0.9$ against sorted ground truth, with the SCDC-informed ablation beating uniform-prior and blind-fitted variants — this is the instrument's license to exist. Second, on the OAS3 co-cultures, either a detectable shift in the sign or magnitude of the knockdown's per-lineage damage effect across immune fractions (supporting the context-flip hypothesis, consistent with the pan-cancer survival inversion) or a clean null, which would itself constrain the melanoma-versus-leukemia duality to non-DNA-damage mechanisms — both are publishable, but only the instrument built here can distinguish them. Third, the early-time-point comparison between sorted and unsorted kinetics: if sorting artifacts are real, the field's current gold standard is biased precisely where repair measurement matters most, and the pooled-comet baseline everyone currently tolerates in mixed samples is worse still. The one outcome that would kill the idea is Phase 1 failing identifiability — calibrated lineage distributions too similar to separate across realistic damage regimes — and we run that check first, cheaply, before building anything downstream.
References
- The single-cell gel electrophoresis assay to determine apoptosis induced by siRNA in Colo 320 cells
- Pan-cancer multi-omics profiling of OAS3 reveals its immunological and prognostic associations across human cancers
- Salt‐inducible kinase 2 confers radioresistance in colorectal cancer by facilitating homologous recombination repair
- MicroRNA-18a Attenuates DNA Damage Repair through Suppressing the Expression of Ataxia Telangiectasia Mutated in Colorectal Cancer
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- HiComet: a high-throughput comet analysis tool for large-scale DNA damage assessment
- E2F1 Silencing Enhances Cisplatin-induced Apoptosis and DNA Damage in Retinoblastoma: CENPE Overexpression as a Compensatory Mechanism
- Rapamycin Exerts Its Geroprotective Effects in the Ageing Human Immune System by Enhancing Resilience Against <scp>DNA</scp> Damage
- CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling
- ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation