Calibrating Autophagy Enrichment Against Measured Flux

Autophagy is the cell's internal recycling system: cytoplasmic junk is sealed into membrane vesicles, shipped to acidic lysosomes, and digested. What matters biologically is not how much recycling machinery a cell possesses but how fast cargo actually moves through the pipeline — the "flux." Measuring that speed is genuinely hard, because static snapshots of the pathway's marker proteins are ambiguous: a pile-up of vesicles could mean the conveyor belt sped up, or that it broke downstream. Meanwhile, single-cell RNA sequencing — which reads the full gene-expression profile of individual cells — combined with functional enrichment analysis, which tests whether coordinated gene groups are over-represented among a cell's expressed genes, has become the default way to declare that "autophagy is activated" in a particular cell subpopulation. But that inference reads the parts list, not the conveyor speed.

The literature already contains a proven trap and a proven opportunity. The trap: pantoprazole made p62 protein accumulate in gastric cancer cells, which the textbook would read as blocked autophagy, but the authors proved the accumulation came from Nrf2-driven transcription of the SQSTM1 gene while the proteasome was crippled — a purely transcriptional artifact that any mRNA-based analysis would misdiagnose The proton pump inhibitor pantoprazole disrupts protein degradation systems and sensitizes cancer cells to death under various stresses. Fangjihuangqi Decoction shows the mirror case: autophagy genes can look induced while flux is completely arrested late in the pathway Identification of Fangjihuangqi Decoction as a late-stage autophagy inhibitor with an adjuvant anti-tumor effect against non-small cell lung cancer. And mRNA can be wildly disconnected from protein altogether — the ABC transporter ABCA4 rose ~132-fold at the message level yet its protein was undetectable by Western blot Expression Profiling of ABC Transporters in a Drug-Resistant Breast Cancer Cell Line Using AmpArray. The opportunity: treatment-matched pairing of high-throughput single-cell assays with single-cell RNA sequencing at scale has been demonstrated to work scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling, and a fast flow-cytometric autophagic flux assay using an acid-sensitive tandem fluorescent reporter already exists Concurrent detection of autolysosome formation and lysosomal degradation by flow cytometry in a high-content screen for inducers of autophagy. Nobody has yet connected them.

Core Claim

Autophagy gene-set enrichment computed from single-cell transcriptomes is uncalibrated against functional autophagic flux, and the resulting errors are predictable, not random. I propose to build the first treatment-matched paired benchmark in which the same cell population under the same perturbation is measured both by the tandem mCherry-GFP-LC3 flux reporter — where green-plus-red signal marks early autophagosomes and red-only signal marks acidified autolysosomes, so the yellow-to-red transition directly reads out flux — and by single-cell RNA sequencing. The benchmark will quantify how often transcript-inferred "autophagy activation" is wrong, identify which error modes dominate (transcriptional confounding versus late-stage arrest versus mRNA-protein disconnect), and test whether a calibrated model trained on the paired data can predict true flux state from the transcriptome alone, or whether flux is information-theoretically absent from mRNA and must be measured.

Research Program

The dataset comes first. I will use MCF-7 cells stably expressing mCherry-GFP-LC3, the exact platform of the flow-cytometric screen Concurrent detection of autolysosome formation and lysosomal degradation by flow cytometry in a high-content screen for inducers of autophagy, so the reporter behavior, the GFP/mCherry ratio readout, and the 1-to-6-hour detection window are already validated. Roughly thirty treatments will span the full mechanistic space: rapamycin under both nutrient-rich and starved conditions (the screen showed it induces flux in one context and suppresses it in the other — a built-in test of context dependence); bafilomycin A1 and chloroquine as lysosomal clamps; the cardiac glycosides digoxin and strophanthidin as established inducers; pantoprazole at both neutral and mildly acidic pH as the deliberately planted transcriptional confounder The proton pump inhibitor pantoprazole disrupts protein degradation systems and sensitizes cancer cells to death under various stresses; a late-stage blocker modeled on the Fangjihuangqi phenotype; and DMSO and starvation-only controls. Five biological replicates per condition, following the replication standard of the paired-imaging benchmark, which measured 627,704 single-cell transcriptomes and 716,767 images across 28 treatments scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling. Because single-cell sequencing is destructive — the cell is lysed to read its RNA — pairing is done at the well and treatment level, not on the same physical cell, exactly as in that benchmark.

Second, the computational audit. Each treatment receives a ground-truth flux label (induced, blocked, unchanged) from the reporter kinetics plus clamp controls. Then every standard way of inferring autophagy from the transcriptome is run on the matched single-cell data: classic per-cell scoring (ssGSEA, GSVA, AUCell) and the deep prototype-based enrichment method DeepGSEA, which routes a shared encoder through gene-set-specific heads and has shown higher sensitivity than those six conventional tools on simulated single-cell data DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data. The headline metric is a concordance rate: the fraction of treatments where the direction of autophagy gene-set enrichment agrees with the reporter-measured flux class, and, crucially, the confusion matrix restricted to the confounder and late-blocker arms where I predict systematic false positives.

Third, calibration. A model trained on the paired labels — a fit-to-data single-cell representation in the spirit of scVI, which on the existing paired benchmark generally beat pre-trained foundation models for treatment identification, plus a flux-state head — is evaluated on within-experiment splits, held-out-experiment splits (the split type where imaging foundation models lost 5–7% balanced accuracy on that benchmark), and held-out compounds it never saw. Fourth, the discordance cases get orthogonal protein validation, not just computation: Western blots of LC3-I-to-LC3-II conversion with and without bafilomycin clamp, p62 protein paired with SQSTM1 RT-qPCR to separate transcription from degradation (the exact methodology that exposed the pantoprazole paradox), and cathepsin B, D, and L maturation blots to catch lysosomal-acidity artifacts that fake yellow puncta (the failure mode of the tandem reporter under the Fangjihuangqi-type block Identification of Fangjihuangqi Decoction as a late-stage autophagy inhibitor with an adjuvant anti-tumor effect against non-small cell lung cancer). A later extension moves to a two-cell-type co-culture, since every flux study in the literature uses homogeneous immortalized lines while single-cell work keeps revealing that mixed tissues behave differently per subtype — as with the five pro-inflammatory microglial clusters in gray matter heterotopia Single-cell sequencing revealed the microglia-activated cell distribution in gray matter heterotopia.

Baselines

The naive reference point is the historical default: a static Western blot of LC3-II and p62 protein from bulk lysate. It is the strongest competitor in the sense of being what the field actually does today, and it is precisely the measurement the literature has shown to be misleading — it cannot separate induction from downstream blockage, and it misdiagnosed pantoprazole-treated cells until transcription was checked separately The proton pump inhibitor pantoprazole disrupts protein degradation systems and sensitizes cancer cells to death under various stresses. It will be run on every treatment arm as the floor comparison, with the same cells, the same time points, and the same replicate structure as the paired assay.

The strongest existing method on the computational question is DeepGSEA applied to autophagy gene sets on our own single-cell matrices DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data — it is the most sensitive enrichment method in the file, beating ssGSEA, GSVA, AUCell, Vision, SCPA, and scGSEA on Splatter simulations at comparable specificity, though no numeric sensitivity value is reported for it on Lacuna, only the qualitative superiority claim. It will be given the identical expression matrices, gene sets, and compute budget as our calibrated model, and it will be beaten only if the calibrated model achieves higher balanced accuracy on flux-class labels across all three split types, with the pre-registered tiebreaker being the confounder arms: DeepGSEA is expected to misclassify the pantoprazole arm, and the calibrated model must not. The strongest quantitative anchor for paired-data treatment identification comes from the scGeneScope benchmark, where the best multimodal model achieved 66% TPR for the PD-98059 treatment in the held-out-experiment evaluation and imaging foundation models lost 5–7% balanced accuracy relative to within-experiment splits scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling; a cross-modal diffusion model on that data reached only 0.293 versus 0.270 balanced accuracy for single profiles and recovered about 29% of the gap to real transcriptomics Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, which sets realistic expectations for how far any image-derived proxy can go.

The key ablation isolates the pairing itself: a transcript-only model with identical architecture, trained without flux labels, is compared against the flux-calibrated model on the same held-out compounds. A second ablation swaps the DeepGSEA prototype heads for plain per-cell gene-set scores to test whether the gains come from the learned representation or purely from calibration against measured flux.

Why not just…?

Why not just do bulk RNA sequencing with standard overrepresentation analysis? Bulk averaging destroys exactly the phenomenon under study. In idiopathic pulmonary fibrosis, myofibroblasts expand while another fibroblast subset collapses, and a pro-fibrotic macrophage subset rises while resident alveolar macrophages vanish — opposing shifts that partially cancel in homogenate Single Cell <scp>RNA</scp>‐Seq Identifies Cell Subpopulations Contributing to Idiopathic Pulmonary Fibrosis in Humans. Bulk also cannot ask whether flux varies cell-to-cell within one treatment, which is the heterogeneity question. And classical overrepresentation analysis with hard gene cutoffs is documented to miss subtle coordinated shifts that the deep enrichment methods catch DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data.

Why not just run Western blots for LC3-II and p62? This is a real part of the design, but as validation, not as the primary readout — because the static snapshot cannot tell fast flux from a broken one Autophagy in glaucoma pathogenesis: Therapeutic potential and future perspectives, p62 protein alone was provably wrong for pantoprazole The proton pump inhibitor pantoprazole disrupts protein degradation systems and sensitizes cancer cells to death under various stresses, and bulk blots cannot resolve which cell subpopulation within a culture changed. The paired reporter assay gives a per-cell, kinetic measurement that no endpoint blot can.

Why not just predict flux directly from the single-cell transcriptome? This is honestly the biggest threat, and the proposal is built so that it cannot fail informatively. If the transcript-only ablation generalizes to held-out compounds, the finding is positive: transcripts do carry flux signal, and we have produced the first calibration map plus a usable predictor. If it fails — and the ABCA4 case, where ~132-fold mRNA upregulation produced zero detectable protein, plus the fact that LC3 lipidation and cathepsin cleavage are post-translational events no RNA read can see Expression Profiling of ABC Transporters in a Drug-Resistant Breast Cancer Cell Line Using AmpArray, suggest it might — we have quantified the ceiling of every transcript-only autophagy claim in the literature. Either result is publishable; the experiments settle it.

Why not just use imaging foundation models instead of sequencing? The existing benchmark is unambiguous on this: pre-trained foundation models were consistently outperformed by simple fit-to-data models, and lost 5–7% balanced accuracy under real held-out-experiment variation scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling; a diffusion model translating morphology to transcriptomic embeddings recovered only about 29% of the imaging-to-sequencing performance gap and failed entirely for compounds like Simvastatin that leave morphology quiet Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion. Morphology is a complement, not a substitute — and the tandem flux reporter is itself the imaging channel here, chosen because it reports function rather than shape.

Decisive Checks

Three pre-registered outcomes decide the core claim. First, the confounder arm: enrichment methods should score pantoprazole-treated cells as autophagy-affected while the reporter shows no late-stage block — if the measured discordance rate on this arm is near zero, the central premise is wrong and I would report that. Second, the context arm: the transcript-only model must be tested on rapamycin under fed versus starved conditions, where the same compound flips flux direction Concurrent detection of autolysosome formation and lysosomal degradation by flow cytometry in a high-content screen for inducers of autophagy; failure to track the flip is direct evidence that flux state is not recoverable from the transcriptome. Third, the calibration test: the flux-calibrated model must beat DeepGSEA on balanced accuracy over held-out-experiment and held-out-compound splits, with no misclassifications in the confounder arm. Every discordant call the model or the baselines make gets adjudicated by the orthogonal protein panel — LC3-II clamp blots, p62 protein with SQSTM1 message, and cathepsin maturation — so that every number in the final concordance table rests on biochemistry, not just fluorescence.