When a pathologist sees a damaged region on a stained tissue slide — shrunken cells, condensed chromatin, ruptured membranes — they cannot tell from the image alone whether those cells died by programmed suicide with membranes still intact, or whether the membranes had already collapsed into secondary necrosis. That distinction is exactly what the Annexin V/propidium iodide (PI) flow cytometry assay measures: Annexin V detects phosphatidylserine flipped to the outer cell membrane (an early sign of apoptosis), while PI enters only cells whose membranes have broken. The assay therefore sorts cells into four biophysical stages — viable, early apoptotic, late apoptotic/necrotic, and debris — but only after the tissue has been ground into a cell suspension, destroying all spatial context. At the same time, RNA sequencing tells us which gene programs shift during injury, and functional enrichment analysis condenses those into pathway narratives — yet, as commonly practiced, it cannot separate a program driving early membrane asymmetry from one driving lysis. Every cited study that uses these assays treats them as parallel, disconnected readouts (Pulsatilla Decoction and its bioactive component β-peltatin induce G2/M cell cycle arrest and apoptosis in pancreatic cancer; Inhibiting Transglutaminase 2 Mediates Kidney Fibrosis via Anti-Apoptosis).
This proposal claims that the four Annexin V/PI stage fractions can be turned into a supervision signal that no current computational pathology model uses. We will build a split-lobe rodent cohort in which each animal contributes matched whole-slide histology, flow-cytometric quadrant fractions, and bulk RNA-seq from the same organ at the same time point, then train a slide-level model to predict the quadrant composition from H&E morphology, and test whether graph-topological pathway features of cell-death signaling cascades predict quadrant trajectories better than standard gene-list enrichment. If it works, the payoff is a microscope image that reports not just "dead cells here" but "which death stage, driven by which pathway program" — and, on the analysis side, the first enrichment method that resolves stage within a single pathway term.
Core Claim
Two claims, one per direction of the bridge. First, the Annexin V/PI quadrant vector is a learnable whole-slide phenotype: a standard patch-encoder-plus-attention architecture, trained on split-lobe paired data, can predict the four stage fractions from H&E slides well enough to separate early-apoptotic-dominant from late-necrotic-dominant tissue — something TUNEL or cleaved-Caspase-3 staining structurally cannot do, since they mark execution of death without staging membrane state. Second, when RNA-seq from the same tissue is summarized by pathway-graph topology rather than by "bag-of-genes" overrepresentation, the resulting activity scores track the temporal dynamics of the quadrants more tightly. The two claims meet in a cross-modal check: attention regions of the quadrant-supervised model should localize to the necrotic and apoptotic zones a pathologist scores independently.
Existing multimodal models are the foil, not the foundation. TANGLE aligns whole-slide images with bulk transcriptomics through contrastive learning but never validates whether its embeddings correspond to any biophysical death stage (Transcriptomics-Guided Slide Representation Learning in Computational Pathology). STFlow and HaDM-ST generate spatial gene-expression maps from histology — a diffusion-style and a flow-matching-style generator respectively — but evaluate only gene-level spatial accuracy, never membrane-state ground truth, and HaDM-ST restricts itself to 200 pre-selected genes with no pathway-level evaluation (Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching; HaDM-ST: Histology-Assisted Differential Modeling for Spatial Transcriptomics Generation). FAST deconvolves spots into cell lineages but cannot resolve death states (Flexible Analysis of Spatial Transcriptomics Data (FAST): A Deconvolution Approach). On the enrichment side, XPathway brought graph topology to pathway testing but was confined to metabolic pathways in a marine bryozoan (Inferring metabolic pathway activity levels from RNA-Seq data) — we extend its permutation-null idea to KEGG signaling cascades (apoptosis, necroptosis, pyroptosis) and, crucially, give it a physical outcome to predict rather than only a differential label. The β-peltatin and TG2-inhibition studies are the closest in spirit — both pair flow cytometry with tissue histology — but neither builds any computational bridge between them, and neither includes RNA sequencing; they motivate the gap rather than occupy it (Pulsatilla Decoction...; Inhibiting Transglutaminase 2...).
Research Program
The data engine is a split-lobe time-course design in the LPS/D-galactosamine fulminant hepatic failure model, a system where the literature already documents the progression from apoptotic injury to overt necrosis and congestion on H&E (Tetrandrine attenuates lipopolysaccharide-induced fulminant hepatic failure in d-galactosamine-sensitized mice). At each of six time points spanning 0–48 hours after insult, animals are randomized to insult, insult plus a protective compound, or vehicle. One liver lobe is fixed for H&E and Masson's trichrome (the collagen stain used for fibrosis staging in All-trans retinoic acid regulating angiopoietins-1 and alleviating extracellular matrix accumulation in interstitial fibrosis rats); a second lobe is enzymatically dissociated for Annexin V-FITC/PI flow cytometry; a third is snap-frozen for bulk RNA-seq. The time course is the design's key asset: early-apoptotic fractions should peak before PI-permeable fractions rise, giving the quadrants temporal separation that a single end-point could not. A second, smaller arm in the unilateral ureteral obstruction kidney model — where the TG2 study paired Sirius Red/TUNEL histology with Annexin V/PI on tubular epithelial cells (Inhibiting Transglutaminase 2...) — tests organ generalization.
Dissociation is the honest methodological risk: mechanical and enzymatic stress can itself flip phosphatidylserine and inflate the Annexin V-positive quadrants. We treat this as a measurable nuisance rather than a hope: each batch includes a viability-spiked control (undamaged cells carried through the full dissociation protocol) whose quadrant drift is measured and used as a per-batch correction, and a deliberately harsh-dissociation condition quantifies the worst-case artifact. Without this, any morphology-to-quadrant mapping could be an artifact of the grinder, not the biology.
The model reuses the TANGLE architecture rather than inventing one: iBOT-Tox patch encoding, attention-based multiple-instance pooling to a slide embedding, but the prediction head becomes a compositional regressor over the four quadrant fractions $q = (q_\text{via}, q_\text{early}, q_\text{late}, q_\text{debris})$, which sum to one. With $\hat q = \mathrm{softmax}(f_\theta(x))$ for slide $x$, training minimizes
$$\mathcal{L} = -\sum_{c} q_c \log \hat q_c \;+\; \lambda\,\big(\mathrm{logit}\,(\hat q_\text{early}/(\hat q_\text{early}+\hat q_\text{late})) - \mathrm{logit}\,(q_\text{early}/(q_\text{early}+q_\text{late}))\big)^2,$$
where the second term is a stage-ratio penalty (logit is the log-odds transform) that forces the model to get the early-versus-late balance right even when total death fraction is off. Pretraining follows TANGLE on TG-GATEs, whose 6,597 paired rat-liver slides and transcript profiles exist precisely for this (Transcriptomics-Guided Slide Representation Learning...).
On the transcriptomic side, we extend XPathway's permutation logic from metabolic graphs to signaling pathways: build the KEGG apoptosis and necroptosis pathway graphs, score each by observed subgraph density and connectivity of differentially expressed genes against edge-swapping and vertex-label-swapping nulls, and resolve pleiotropic genes shared across death pathways with the same expectation-maximization procedure (Inferring metabolic pathway activity levels from RNA-Seq data). This follows the established critique that standard overrepresentation treats pathways as unstructured gene lists and suffers crosstalk from shared members (Interpreting Gene Expression via Functional Enrichment Analysis; m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease). The testable question is whether topological scores of sub-pathway modules — upstream receptor/signaling nodes versus downstream execution and lytic modules — differentially track the early versus late quadrant trajectories.
Proposed Experiments
Three experiments, in order of decisiveness. The first is the morphology-to-quadrant regression on the hepatic cohort: primary metric is mean absolute error per quadrant fraction, with the early-versus-late log-odds error as the staging-specific metric; evaluation is leave-time-point-out (train on five time points, predict the held-out one), which is the hard test because it requires extrapolating to unseen injury severity. The second is the enrichment head-to-head: regress quadrant trajectories from either topological pathway scores or standard ORA statistics of the same RNA-seq, comparing cross-validated $R^2$; the sharp prediction is that upstream-cascade topology tracks the early-apoptotic curve while downstream lytic-module topology tracks the late/necrotic curve, a split ORA cannot express. The third is the localization check: quadrant-class attention maps from the trained slide model are compared against blinded expert annotations of necrotic zones, in the spirit of the DMSF study's demonstration that its attention clusters recovered expert-marked "Necrosis" and "Hyperplastic Blood Vessels" regions in IvyGAP (Disentangled Multi-modal Learning of Histology and Transcriptomics for Cancer Characterization). HEST-1k serves as an external sanity bench for the patch encoder (its 2.1 million morphology-expression pairs and 76 million segmented nuclei support standard gene-prediction checks, where the best foundation model H-Optimus-0 leads and STFlow reported an 18% average relative improvement over pathology foundation models; HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis; Scalable Generation of Spatial Transcriptomics...), and a kidney transfer test on the UUO arm closes the program. The whole design fits within roughly 120 animals, two organs, one facility.
Baselines
The naive reference point is predicting the cohort-mean quadrant vector for each time point and condition. It is stronger than it sounds, because injury time courses are smooth and most of the variance in quadrant fractions is captured by time alone; any learned model must beat it on leave-time-point-out splits, where the mean predictor degrades gracefully, to justify its existence.
The strongest in-situ competitor is TUNEL and cleaved-Caspase-3 immunohistochemistry quantified as positive-stain density per slide — the standard practice in the closest existing work, which used TUNEL on kidney sections (Inhibiting Transglutaminase 2 Mediates Kidney Fibrosis via Anti-Apoptosis) and Ki-67 plus H&E on xenografts (Pulsatilla Decoction...). It runs on our own slides, same animals, same time points, so the comparison is perfectly paired. No reported result found on Lacuna for a TUNEL-density-to-Annexin-quadrant regression, so we establish that number ourselves. Beating it means specifically lower error on the early-apoptotic quadrant and on the early-versus-late log-odds, since TUNEL marks DNA fragmentation common to both stages and should perform acceptably on total death fraction while failing at staging.
The strongest existing slide-representation method is the TANGLE pipeline: iBOT-Tox patch features with attention-based multiple-instance pooling, pretrained transcriptomics-guided on TG-GATEs, which reported an absolute macro-AUC gain of +5.9% over supervised ABMIL in few-shot liver classification at $k=10$ (Transcriptomics-Guided Slide Representation Learning in Computational Pathology). We run it with its published architecture, our identical pretraining corpus, the same quadrant head, and the same tuning budget in parameter-search trials, so the only difference is whether the model was trained to predict transcriptomic profiles (its own objective) or quadrant fractions (ours). Beating it on quadrant MAE and staging error would show that biophysical supervision adds information that transcriptomic supervision does not supply; losing would itself be informative, showing the transcriptomic signal subsumes the death-stage signal.
Two ablations isolate the proposal's ingredients. Replacing topological pathway scores with standard ORA statistics of the same genes, in the same regression, isolates the topology contribution. And a histology-only variant without any RNA-seq input isolates how much of the quadrant prediction is pure morphology. Finally, for the RNA-to-quadrant direction, an L2-regularized linear model on centered-log-ratio transformed gene abundances is a mandatory baseline given that across roughly 40,000 samples in recount2, it outperformed deep autoencoders on phenotype prediction (Standard machine learning approaches outperform deep representation learning on phenotype prediction from transcriptomics data); we use linear heads throughout precisely so this concern is built in rather than argued away.
Why not just…?
Why not just use TUNEL or Caspase-3 staining on the histology section? Because those assays detect execution of cell death, not the membrane state that distinguishes its stages: a TUNEL-positive cell can be an early apoptotic cell with an intact membrane or a post-apoptotic corpse with a ruptured one, and the distinction is exactly what determines whether, for example, a protective compound is preventing death initiation or merely delaying membrane collapse. Our experimental design settles this empirically rather than rhetorically: TUNEL density is a registered baseline, and the decisive comparison is early-versus-late log-odds error, where we expect TUNEL to fail structurally. We are honest that if TUNEL density turns out to predict the staging ratio well in this model system, the imaging contribution of this proposal weakens substantially — and that outcome would itself be worth reporting.
Why not just run a targeted qPCR panel of canonical apoptosis genes? The recount2 study found that multivariate models over all genes significantly outperformed the best single-gene or targeted predictors across phenotype tasks (Standard machine learning approaches outperform...), and comparative transcriptomics repeatedly surfaces unanticipated drivers outside canonical panels, such as Rab32 in Schwann-cell pyroptosis (Comparative Transcriptomics for Elucidating Cellular Mechanisms). Still, the targeted panel is cheap, so we run it as a baseline in the RNA-to-quadrant regression; if a five-gene panel matches genome-wide topology, the enrichment machinery is not earning its cost and we would say so.
Why not just run standard GO/KEGG overrepresentation on the DEGs? This is the most real threat, because ORA is nearly free and often sufficient. The evidence against sufficiency here is specific: XPathway showed that accounting for graph connectivity and shared-enzyme crosstalk identifies differential pathway activity where unstructured tests add statistical noise, confirmed by a 6.88-fold qPCR-validated change in the fatty-acid-elongation enzyme fadA (Inferring metabolic pathway activity levels from RNA-Seq data) — and our question is harder than a binary differential label, since we need to split one pathway term into stage-tracking sub-modules. The head-to-head regression (Experiment 2) is designed precisely so that if ORA matches topology at predicting quadrant trajectories, the second claim of this proposal dies and we report it.
Why not just use existing multimodal models like TANGLE or STFlow without collecting any flow data? Because they are trained and evaluated against transcriptomic or gene-level targets, and there is no evidence — and no way to obtain evidence post hoc — that their outputs correspond to membrane-state stages; STFlow's 18% relative improvement is measured on gene-expression prediction (Scalable Generation of Spatial Transcriptomics...), not death staging. The flip side of this objection is the deep-versus-linear tension: given the recount2 result, one could argue for skipping deep models entirely, which is why our comparisons run both a linear head and the attention-pooling architecture on identical features and budgets, letting the data decide rather than the ideology.
Why not just bulk RNA-seq without the spatial and imaging machinery? Because bulk expression averages over all cells and dilutes focal events — a small zone of localized apoptosis within an organ is exactly the signal homogenization destroys, which is the documented motivation for spatial resources like HEST-1k (HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis). Our design concedes that the flow assay itself is spatially blind; the whole point of the split-lobe pairing is that the slide image restores the spatial dimension that dissociation destroys, and the localization experiment tests whether it actually does.
Decisive Checks and Risks
The program succeeds if three numbers move: the learned model beats the mean predictor and TUNEL density on early-versus-late log-odds error under leave-time-point-out; topological sub-module scores beat ORA statistics at predicting quadrant trajectories; and model attention aligns with independently annotated necrotic zones above chance. The risks are concrete. Dissociation artifact could swamp the early-apoptotic signal — mitigated, and measured, by the spiked-control correction. The cohort could be too small for the attention model, in which case the linear baseline becomes the finding, echoing the recount2 lesson. And the time-course sampling could miss the early-apoptotic peak — mitigated by a dense pilot at 1–6 hours before committing the full cohort, informed by the dose- and time-dependent apoptosis shifts documented in the flow-cytometry literature (Pulsatilla Decoction...). If the central claim survives, the immediate extension is obvious: the same quadrant-supervision recipe applied to human pathology archives, where billions of stained slides exist and not one carries a flow-cytometry label — this work would be the first proof that such labels are learnable from glass.
References
- Pulsatilla Decoction and its bioactive component β-peltatin induce G2/M cell cycle arrest and apoptosis in pancreatic cancer
- Inhibiting Transglutaminase 2 Mediates Kidney Fibrosis via Anti-Apoptosis
- Transcriptomics-Guided Slide Representation Learning in Computational Pathology
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching
- HaDM-ST: Histology-Assisted Differential Modeling for Spatial Transcriptomics Generation
- Flexible Analysis of Spatial Transcriptomics Data (FAST): A Deconvolution Approach
- Inferring metabolic pathway activity levels from RNA-Seq data
- Tetrandrine attenuates lipopolysaccharide-induced fulminant hepatic failure in d-galactosamine-sensitized mice
- All-trans retinoic acid regulating angiopoietins-1 and alleviating extracellular matrix accumulation in interstitial fibrosis rats
- m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease
- Disentangled Multi-modal Learning of Histology and Transcriptomics for Cancer Characterization
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis
- Standard machine learning approaches outperform deep representation learning on phenotype prediction from transcriptomics data