Pyroptosis is an inflammatory, explosive form of programmed cell death. The dying cell punches pores in its own membrane and then bursts, spilling inflammatory signals into the tissue. Single-cell RNA sequencing (scRNA-seq), which measures gene expression one cell at a time, is now the default way to ask which cells in a diseased tissue are dying this way. The standard workflow then runs a pathway enrichment test, which asks whether a curated set of genes, such as "pyroptosis," is unusually active in some group of cells. This workflow has a structural blind spot. Cells that have burst leak their cytoplasmic RNA and arrive as low-count, mitochondria-heavy droplets. Quality-control filters delete those droplets by design, and the RNA that leaked out ends up in the background "soup" of the experiment, where cleaning tools subtract it as contamination. Enrichment analysis is then run on the survivors. So the cell type that burst the most can look the least pyroptotic.
We propose to turn the two discarded signals into the measurement. The first is the damaged droplets. The second is the ambient RNA, the free-floating RNA captured in empty droplets. We model the ambient RNA as a mixture of leak profiles from each cell type, weighted by how much of that type lysed. From this we estimate a per-cell-type lysed fraction, and we then run pathway enrichment on a corrected population that puts the missing cells back. In plain terms, we infer who died by examining what they left behind, and we test pathways on the whole population rather than on its survivors.
Core Claim
The research file names the "lytic cell death capture paradox" as an open limitation. Standard isolation and filtering systematically discard cells undergoing gasdermin pore formation and NINJ1-driven membrane rupture. The Structural basis paper (Structural basis of NINJ1-mediated plasma membrane rupture in cell death) shows that this rupture is an active, protein-driven, terminal event, which is why the leak is total rather than partial.
Our claim has two parts. First, the ambient RNA of a droplet scRNA-seq run carries a recoverable, cell-type-resolved record of lytic death. Second, correcting enrichment scores for the estimated lost fraction changes which cell types are called pyroptotic, and in the right direction as judged by physical lysis assays.
The closest work applies enrichment to pyroptosis transcriptomes but never questions which cells were sampled. The Rab32 study (Rab32 facilitates Schwann cell pyroptosis in rats following peripheral nerve injury by elevating ROS levels) ran classical threshold-based Gene Ontology and KEGG enrichment on bulk nerve tissue. The COVID-19 PANoptosis work (PANoptosis, an indicator of COVID-19 severity and outcomes) built PANoptosis scores on filtered airway-fluid and blood single-cell data, again with classical enrichment. Both motivate our proposal: they show pyroptosis signatures are clinically interesting. Neither accounts for survivorship.
The obvious follow-up is to rerun such data through newer enrichment tools like DeepGSEA, UNIFAN, pareg or GSPA. That would sharpen the statistics on a sample that is still biased. We fix the sample instead.
Research Program
The core mechanism is a small generative model of the ambient profile. For gene $g$, the ambient expression $a_g$ measured in empty droplets is modelled as
$$a_g = \sum_k w_k \, \mu_{kg} \, \rho_g + b_g + \epsilon_g,$$
where each symbol is defined as follows. The index $k$ ranges over cell types. $\mu_{kg}$ is the mean expression of gene $g$ in intact cells of type $k$. $\rho_g$ is the fraction of that gene's transcripts that sit in the cytoplasm and therefore leak on rupture. $w_k \ge 0$ is the amount of lysed type-$k$ material in the soup. $b_g$ is a non-lytic background term, fitted from a control sample with no death stimulus. $\epsilon_g$ is count noise.
The factor $\rho_g$ is the ingredient that distinguishes lysis from ordinary contamination. A burst cell leaks cytoplasmic mRNA but retains nuclear and intron-rich transcripts. We estimate $\rho_g$ from paired nuclear and whole-cell fractions, or from the ratio of spliced to unspliced reads. Low-count, high-mitochondrial droplets that are currently filtered out are kept. They are assigned to cell types using their retained nuclear reads, which gives a second, independent estimate of each type's damaged fraction.
The lysed fraction $f_k$ combines $w_k$ with the damaged-droplet counts, normalised by the number of intact type-$k$ cells. Enrichment is then run on a reweighted population. Each cell type's pathway score becomes a mixture of the survivors' score, weighted by $1-f_k$, and a score for the lysed compartment, weighted by $f_k$. The lysed compartment's score is computed on the damaged droplets and the deconvolved leak profile.
Functional enrichment enters in two places. The first is the pathway definition. The file notes that NINJ1 filament assembly, non-canonical gasdermin cleavage and PANoptosis are fragmented or absent in GO and KEGG. We therefore assemble an explicit gene set covering the inflammasome sensors, caspase-1, -4, -5 and -11, the gasdermins, NINJ1, and IL-1β and IL-18 processing. We keep the published GO and KEGG terms alongside it so that results can be compared. The second is the test itself, where any single-sample scorer can be plugged in, from AUCell to DeepGSEA.
Baselines
The naive reference is the pipeline that the Rab32 and COVID PANoptosis studies represent. It applies standard quality-control thresholds on counts and mitochondrial fraction, removes ambient RNA as contamination, and scores the pyroptosis gene set per cell with AUCell or ssGSEA. It then compares cell types with a Mann–Whitney test or a classical Fisher's exact test on differentially expressed genes. We run it on the identical count matrices, with thresholds set the way those papers describe or at common defaults. Beating it means ranking the truly lysing cell type first where the naive pipeline does not, and tracking physical lysis better across doses.
The strongest existing enrichment method on this question is DeepGSEA (DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data). It pairs a shared encoder with gene-set-specific heads and learned prototypes, and it reports higher sensitivity than ssGSEA, GSVA, AUCell, Vision, SCPA and scGSEA on four Splatter simulations. The file gives no numeric value, so we record: no reported result found on Lacuna for DeepGSEA sensitivity as a number. We give DeepGSEA the same filtered cells, the same gene set, and a tuning budget matched to our own model's fitting, measured in GPU-hours and seeds. The comparison is fair because DeepGSEA's advantage is sensitivity to weak signal. If survivorship is the real problem, more sensitivity on survivors should not rescue the ranking.
Two further baselines guard against specific objections. The first is pareg (Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression). It regresses differential expression p-values on pathway membership with a network-fusion penalty and reports the highest mean precision-recall area on synthetic benchmarks, though no number is given in the file. We use it to check that our gains are not simply due to coherence across related death pathways.
The second is UNIFAN (Unsupervised Cell Functional Annotation for Single-Cell RNA-Seq). It reports an Adjusted Rand Index of 0.81 on pbmc28k against 0.38 for Leiden clustering. We use it as a guardrail rather than a competitor. Retaining damaged droplets must not reduce clustering agreement on pbmc28k below UNIFAN's reported 0.81 when those droplets are excluded from the clustering step.
Three ablations isolate the key ingredients. The first sets $\rho_g=1$, treating leakage as uniform contamination. The second uses ambient deconvolution alone, without damaged droplets. The third uses damaged droplets alone, without ambient deconvolution.
Why not just…?
Why not just relax the quality-control thresholds and run standard enrichment? Relaxing thresholds admits damaged droplets, but it does not say how many lysed cells left no droplet at all, and fully ruptured cells mostly contribute to the soup. It also mixes true lysis with ordinary handling damage. The damaged-droplets-only ablation is exactly this alternative with cell-type assignment added. If it matches the full model on the lysis-tracking test, we will report that the simpler fix suffices.
Why not just add a proper pyroptosis/NINJ1 gene set to GO and use DeepGSEA or GSPA? This addresses the incompleteness of annotation databases, which is a real gap. Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings reports gains over GSEA on 75.8% of 42 GEO2KEGG datasets by expanding sets along protein-interaction networks. Better gene sets, however, cannot restore cells that were never sampled. Our baselines already receive the curated set, so any remaining gap is attributable to survivorship.
Why not just sequence the culture supernatant? This is a genuine threat in vitro. Supernatant RNA directly measures leaked material, and for cultured cells it may be simpler. It is not available for dissociated tissue or clinical airway samples, which is where the method matters. Supernatant RNA also lacks cell-type resolution unless it is deconvolved, which brings back our model. We therefore use supernatant sequencing as ground truth in the controlled experiments rather than as a competitor. If our ambient estimate fails to track it, the approach fails.
Why not just use existing ambient-RNA removal tools? Those tools estimate the soup in order to subtract it, and they assume it is non-informative. The $\rho_g=1$ ablation approximates their contamination model. If it performs as well as the full model, the cytoplasmic-leak term adds nothing.
Proposed Experiments
The first experiment is simulation. Using Splatter-style synthetic data, the same generator family used in the DeepGSEA benchmarks, we plant a pyroptosis program in one of several cell types. We then remove a controlled fraction of those cells and transfer their cytoplasmic transcripts into the ambient pool with realistic $\rho_g$. The metrics are two. The first is the correlation between the estimated and true lysed fraction. The second is the precision-recall area for calling the pyroptosis gene set enriched in the correct type, as loss rises from 0 to 80%. The decisive pattern is that baseline enrichment should degrade, and eventually invert, as loss rises, while the corrected score should not.
The second experiment is a controlled wet-lab mixture. Human monocytes, which undergo pyroptosis under priming plus a second inflammasome signal, are co-cultured with sample-barcoded lymphocytes, which do not. We sweep the dose of stimulus and add the NLRP3 inhibitor MCC950 or the pore inhibitor disulfiram. Disulfiram is reported to block pyroptotic bubbling and LDH release at 10 μM in Disulfiram alleviates pristane-induced lupus via inhibiting GSDMD-mediated pyroptosis.
Ground truth at each condition comes from three sources: LDH release, counts of dye-positive cells, and supernatant RNA-seq. The decisive check is Spearman correlation of the monocyte lysed fraction $f_k$ with LDH across conditions, with the lymphocyte $f_k$ staying near zero.
A pore-blocked condition gives a sharper test. Disulfiram acts upstream of rupture and should suppress $f_k$. An inhibitor that acts only at the step of membrane rupture, if available, would leave pyroptosis priming visible in the survivors while suppressing $f_k$, which separates priming from execution.
The third experiment applies the method to the public COVID-19 airway-fluid and blood single-cell data used in the PANoptosis study. The question there is whether the corrected scores change which monocyte subsets, including the C1QA+ and S100A8+ populations, appear to die. We also ask whether the lysed fraction of each subset tracks disease severity better than the survivors' scores do. We treat this as exploratory, because it has no lysis ground truth, and interpret it only if the first two experiments succeed.
References
- Structural basis of NINJ1-mediated plasma membrane rupture in cell death
- Rab32 facilitates Schwann cell pyroptosis in rats following peripheral nerve injury by elevating ROS levels
- PANoptosis, an indicator of COVID-19 severity and outcomes
- DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- Unsupervised Cell Functional Annotation for Single-Cell RNA-Seq
- Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings
- Disulfiram alleviates pristane-induced lupus via inhibiting GSDMD-mediated pyroptosis