When Does Blocking NF-κB Stop Helping? Phase-Resolved Single-Cell Pathway Maps of Timed Inhibition in Sepsis

Sepsis begins as a runaway inflammatory reaction and, in patients who survive the first days, often ends in the opposite failure: a paralyzed immune system that cannot fight even opportunistic infections. The master switch driving the first phase is NF-κB, a transcription factor that, once released into the nucleus, commands the production of tumor necrosis factor, interleukin-6, and related alarm molecules. Drug studies consistently show that blocking this switch early — within two hours of inducing sepsis in mice — dampens the cytokine storm and improves survival. What nobody knows is what happens if you block it later, when the immune system has already slid into exhaustion, because every published inhibition study fires the drug at the hyper-inflammatory window and looks no further (Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway; 4-Octyl itaconate alleviates endothelial cell inflammation and barrier dysfunction in LPS-induced sepsis via modulating TLR4/MAPK/NF-κB signaling). The clinical reviews that define late sepsis warn that indiscriminate suppression at that stage could deepen immune paralysis (Sepsis-induced immunosuppression: mechanisms, biomarkers and immunotherapy) — a warning that has never been resolved experimentally.

The reason it has not been resolved is that the answer is not a single number. Different circulating cell types move through the two phases in opposite directions at the same moment: monocytes lose their antigen-presentation machinery and drift toward a wound-heeding, immunosuppressive identity, while T cells die off or become exhausted through inhibitory checkpoints. Bulk assays average these opposing programs into one blurred signal, and the only single-cell machine-learning work on sepsis to date treats the disease as a static yes/no diagnostic label, lumping hyper-inflammatory and paralyzed patients into one class (scCaT: An explainable capsulating architecture for sepsis diagnosis transferring from single-cell RNA sequencing). This proposal builds the missing measurement: a phase-resolved, cell-type-resolved map of NF-κB pathway activity along the sepsis trajectory, used both to read out existing human single-cell data and to predict — then empirically test in mice — at which disease phase and in which cell types NF-κB inhibition flips from protective to harmful.

Core Claim

The claim is that the timing question for NF-κB inhibition can be answered computationally before it is answered pharmacologically, by embedding curated pathway knowledge (Gene Ontology and KEGG gene sets, including the TLR4/MyD88/NF-κB/NLRP3 cascade) directly into two places where it is currently absent: temporal trajectory inference over sepsis single-cell data, and generative perturbation modeling of drug response. Concretely, we will define a per-cell-type, per-phase pathway-activity coordinate for NF-κB signaling and its downstream targets, project longitudinal sepsis blood onto it, and derive a measurable quantity — the phase-conditional inhibition discrepancy — that predicts how strongly a given inhibitor perturbs a given cell type depending on where that cell type sits on the hyper-inflammation-to-exhaustion axis. The decisive test is then a timed-inhibition experiment in a mouse sepsis model with single-cell readout: if the model is right, the same drug at the same dose should protect when given early and measurable specific cell populations when given late, and the model should predict both the direction and the affected lineages in advance.

Background: What Each Thread Supplies

Three bodies of work motivate the gap and supply the parts. First, the sepsis-immunology literature establishes the biphasic phenomenology — monocyte deactivation and HLA-DR loss, T cell apoptosis and PD-1/Tim-3 checkpoint exhaustion, expansion of suppressive myeloid populations (Sepsis-induced immunosuppression: mechanisms, biomarkers and immunotherapy; Tim-3 pathway dysregulation and targeting in sepsis-induced immunosuppression; Increased Percentages of T Helper Cells Producing IL-17 and Monocytes Expressing Markers of Alternative Activation in Patients with Sepsis) — and states explicitly that intervention timing is unresolved and that mouse models diverge from human kinetics. This is the application regime and the cautionary constraint. Second, the enrichment-analysis literature shows that constraining models with pathway structure is not decoration but a large performance lever: on the pbmc28k peripheral-blood benchmark, clustering with a binary GO/KEGG decoder achieved an adjusted Rand index of 0.81 versus 0.38 for standard Leiden clustering (Unsupervised Cell Functional Annotation for Single-Cell RNA-Seq), and removing pathway enrichment from a trajectory method caused a four-fold increase in mismatch scores and a three-fold drop in accuracy (Tempora: Cell trajectory inference using time-series single-cell RNA sequencing data). Pathway priors are the core mechanism we import. Third, the perturbation-modeling literature supplies a strong generative baseline — a diffusion model trained on 61 million cells that predicts post-drug single-cell distributions and wins 96–100% of out-of-distribution drug-prediction comparisons against prior methods (PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling) — but that model, like the trajectory and enrichment tools above, has never been applied to longitudinal in-vivo sepsis data or to NF-κB-specific inhibition across disease phases, and it carries no pathway structure in its loss. The timing paradox of NF-κB inhibition sits precisely in this unoccupied intersection.

Research Program

The program has three linked components. The first is a phase-annotation model for sepsis single-cell data. We start from the observation that the clinical phases are defined not by single genes but by coordinated programs — early NF-κB target transcription, late checkpoint and anti-inflammatory programs — so we build the phase axis out of enrichment scores rather than raw expression, following Tempora's design of converting clusters into pathway-activity profiles and directing edges by collection time (Tempora: Cell trajectory inference using time-series single-cell RNA sequencing data), but extended in two ways that work has not attempted: we apply it to sepsis rather than embryonic development, and we add gene-level alignment to detect where individual genes deviate from the program-level story, using the five-state dynamic-programming alignment of Gene-level alignment of single-cell trajectories, which handles genes that appear in one condition and not the other via insertion/deletion states — a real concern in sepsis, where entire lymphocyte populations drop out through cell death. The phase model outputs, for every cell, a continuous coordinate $s \in [0,1]$ from fully hyper-inflammatory to fully exhausted, plus a confidence, and it identifies the transition point per lineage.

The second component is a phase-conditional perturbation model. We train a generative model that predicts the post-treatment distribution of a cell population given its pre-treatment state and a drug, in the spirit of PerturbDiff, but with one architectural change whose value we will isolate by ablation: the output is factorized through a pathway-activity decoder constrained by the same GO/KEGG matrix, so that the model must express its predictions as shifts in named programs (NF-κB targets, antigen presentation, interferon response, checkpoint ligands) rather than free-floating latent movement. We then define the inhibition discrepancy for NF-κB blockers, cell type $k$, and phase coordinate $s$ as the maximum mean discrepancy between the model's predicted untreated and treated distributions for that cell type at that phase:

$$D_k(s) = \mathrm{MMD}^2\big(\hat{P}_{k,s}^{\,\text{untreated}},\; \hat{P}_{k,s}^{\,\text{inhibited}}\big),$$

where $\hat{P}_{k,s}$ are predicted single-cell distributions restricted to cells of type $k$ within a window around phase coordinate $s$. The signature quantity, however, is not the size of $D_k(s)$ but its direction when decomposed over pathway axes: early, we expect inhibition to shrink NF-κB target and inflammasome programs while sparing antigen-presentation programs (protective); late, we expect the same drug to further suppress antigen-presentation and survival programs in monocytes and T cells (harmful). The crossing phase coordinate $s^*_k$ where the predicted direction flips is the model's headline output.

The third component turns the prediction into an experiment. In a polymicrobial cecal-ligation-and-puncture mouse model, matched to the protocol of the Decitabine study (Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway), we administer an NF-κB inhibitor at two windows: acute (2 hours post-surgery, the only window ever tested) and delayed (chosen by the model's $s^*$, expected in the late immunosuppressive phase), with single-cell RNA sequencing of blood and spleen at 2, 6, 24, and 72 hours, plus vehicle controls and untreated sepsis. The readout is not just survival curves but the cell-type-resolved pathway map: which lineages are rescued, which are further suppressed, and whether the measured direction of change matches the model's phase-conditional prediction.

Proposed Experiments

Experiment 1 (human data, computational): build and validate the phase model on the Broad Single Cell Portal sepsis dataset SCP548 — 126,351 cells from 29 septic patients and 36 controls, the dataset used by scCaT: An explainable capsulating architecture for sepsis diagnosis transferring from single-cell RNA sequencing — augmented with the multi-cohort bulk GEO cohorts from the same paper for transfer testing, and with the Ulinastatin multicenter sepsis datasets (CNCB-NGDC OMIX005600 and PRJCA006118) as an independent immunomodulatory-drug cohort (Correction: Integrated single cell and bulk RNA-Seq analysis revealed immunomodulatory effects of Ulinastatin in sepsis: a multicenter cohort study); no quantitative efficacy values were reported for that cohort, so it serves for qualitative replication only. Because SCP548 lacks time-series sampling by design, phase coordinates are derived from program composition and evaluated against known phase markers (HLA-DR, CD163/CD206, PD-1/Tim-3 ligands) and against the healthy controls. Success is defined as: the phase axis separates cells in an order consistent with the marker progression, and the NF-κB target program is high at one end and low at the other. There is no reported result found on Lacuna for sepsis phase classification accuracy, so we establish this benchmark ourselves.

Experiment 2 (perturbation model): train the phase-conditional perturbation model on held-out single-cell perturbation data — the PBMC cytokine-stimulation benchmark of 9.7 million cells and the large drug-screen atlases used by PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling — then apply it to NF-κB inhibitors (Decitabine, 4-octyl itaconate) conditioned on the phase coordinates from Experiment 1. Evaluation is held-out perturbation prediction scored by MMD-based error and by win rate against PerturbDiff under identical data splits, compute, and tuning budget. The key ablation removes the pathway decoder (same architecture, same budget) to isolate whether pathway structure improves prediction on sparse, subtle perturbations — the regime where deep enrichment methods previously beat classical scorers on sensitivity (DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data; the file reports no numerical sensitivity values for that comparison, so we treat it as directional evidence). A second, cheaper ablation conditions only on cell type, not on phase, to show that phase conditioning carries information beyond lineage identity.

Experiment 3 (decisive wet-lab check): the timed-inhibition CLP study described above, with the delayed-dosing window chosen by the model's predicted flip phase $s^*$, blinded outcome scoring, and the acute arm serving as the replication of the published result — the reference point there is Decitabine's survival improvement at 2 hours post-CLP with a log-rank $p$-value of 0.0081 at 1 mg/kg (Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway). The pre-registered predictions are: which cell types change, in which pathway programs, and in which direction at each window. The model is falsified if the delayed arm's cell-type-resolved effects are indistinguishable from the acute arm's, or if the measured pathway directions contradict the predictions in more than a pre-specified fraction of lineages.

Baselines

The naive reference point is static sepsis classification. The concrete system is scCaT, a capsule network with a Transformer decoder that pre-trains on SCP548 and transfers to bulk cohorts (scCaT: An explainable capsulating architecture for sepsis diagnosis transferring from single-cell RNA sequencing); it reports an AUROC of 0.93 on the single-cell test set and an average AUROC of 0.986 with AUPRC of 0.994 across seven bulk cohorts. It is the strongest competitor on diagnosis, and our phase model must first reproduce its diagnostic performance (sepsis versus control, same SCP548 splits) to prove we lose nothing by reframing — but it cannot, by construction, rank patients or cells by phase, so the informative comparison is on the new phase task, where its representations will be evaluated as features for a phase regressor with the same tuning budget as ours.

The strongest existing method for the trajectory component is Tempora, which already uses pathway enrichment profiles and temporal information for trajectory inference and beat Monocle 3, Slingshot, TSCAN, CSHMM, and PAGA across developmental benchmarks (Tempora: Cell trajectory inference using time-series single-cell RNA sequencing data). Its reported numbers — trajectory accuracy of 0.78 on HSMM, 0.90 on murine cortex, 0.50 on murine cerebellum, and roughly 60 seconds versus 1700 seconds for Monocle 3 — define the bar for trajectory reconstruction speed and fidelity. We will run Tempora on our sepsis time-series exactly as released, with the same gene sets and compute, and beat it only if our gene-level alignment layer materially improves phase-axis recovery on developmental benchmarks first (its home turf), then on sepsis. A fair target is matching its accuracy scores while extending to a regime it has never handled: drug perturbation and a pathology-driven, non-developmental axis.

The strongest competitor for the perturbation component is PerturbDiff itself, a functional diffusion model that reported win rates of 96–100% against the best prior baselines on out-of-distribution drug prediction in PBMC and large drug-screen data, retained accuracy when trained on 1% of the data, and performed best at roughly 114M parameters (PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling). We will run it with its published recipe at its reported best size, same splits, same training data, and same evaluation metrics, and treat parity on generic perturbation prediction as acceptable — our claim is not that pathway decoders beat it everywhere, but that phase conditioning plus pathway decomposition predicts direction-of-effect in sepsis, which PerturbDiff has never been tested on. The decisive ablation is our own model without the pathway decoder and without phase conditioning: if that stripped version matches the full model on the sepsis predictions validated in Experiment 3, the proposal's key ingredient is empty and we will say so.

Finally, the pharmacological baseline for Experiment 3 is the published acute Decitabine result itself — 1 mg/kg at 2 hours post-CLP, survival improvement with log-rank $p = 0.0081$ (Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway) — which our acute arm must replicate before any delayed-arm claim is credible. The file reports no single-cell or cell-type-resolved outcomes for that study, so there is no reported result found on Lacuna for cell-type-level NF-κB inhibition effects in sepsis; that absence is the benchmark gap we fill.

Why not just…?

Why not just use bulk RNA sequencing or cytokine arrays? The published NF-κB inhibition work already used bulk arrays and Western blots and got a clear survival answer. But bulk averaging is precisely what hides the question we are asking: in sepsis, monocytes and T cells move in opposite transcriptionomic directions simultaneously, so a bulk readout cannot distinguish a drug that protects endothelial and myeloid compartments from one that additionally starves residual T cells (Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation makes the averaging problem general; the sepsis-specific opposite-direction programs are documented in Sepsis-induced immunosuppression: mechanisms, biomarkers and immunotherapy). Experiment 3's single-cell readout directly tests this: if the cell-type-resolved effects of delayed inhibition turn out to be uniform across lineages, bulk would have sufficed and we will report that — but the published dose saturation at 1.5 mg/kg Decitabine already hints at non-uniform biology worth resolving.

Why not just cluster with Leiden and run overrepresentation analysis afterwards? This is the standard pipeline, and it is cheaper. The evidence against it is quantitative: on matched peripheral-blood data, Leiden clustering reached an adjusted Rand index of only 0.38 versus 0.81 for a pathway-constrained model (Unsupervised Cell Functional Annotation for Single-Cell RNA-Seq), and thresholded overrepresentation analysis discards coordinated sub-threshold shifts — exactly the regime of a signaling cascade under partial pharmacological inhibition. The pathway-ablation result in Tempora (four-fold mismatch increase, three-fold accuracy drop) is the cleanest proof that pathway structure carries the signal. We will run the Leiden-plus-ORA pipeline as an explicit baseline on our own phase task with the same compute so the comparison is settled on sepsis data rather than borrowed from other tissues.

Why not just use off-the-shelf trajectory tools like Monocle 3 or Slingshot? They ignore real collection times and misroute in noisy data, and they were outperformed by Tempora across three developmental benchmarks (Tempora: Cell trajectory inference using time-series single-cell RNA sequencing data); we adopt the pathway-based Tempora approach as our starting point and baseline rather than re-litigating this. The remaining honest gap in our favor is narrower: Tempora works at cluster-level pathway averages and its own authors note it misses gradual within-type gene changes, while the gene-level aligner we import handles population dropouts through insertion/deletion states (Gene-level alignment of single-cell trajectories) — and lymphocyte dropout is a defining feature of late sepsis. If the Tempora baseline alone recovers the phase axis and predicts the Experiment 3 directions as well as our full model, the gene-alignment layer is unnecessary, and that ablation is in the plan.

Why not just give NF-κB inhibitors early and be done, since the acute result is already proven? This is the most serious objection, because the acute result is real and replicated. The counter-evidence is that late sepsis mortality is driven by immune paralysis, not inflammation, and the sepsis literature explicitly warns that phase-blind immunosuppression risks worsening lymphopenia and secondary infection vulnerability (Sepsis-induced immunosuppression: mechanisms, biomarkers and immunotherapy; Lymphopenia in sepsis: a narrative review) — a warning with no single-cell or temporal evidence behind it in either direction. Clinically, patients rarely present within two hours; the actionable question is what to do for the majority who arrive later, and no reported result found on Lacuna addresses delayed NF-κB inhibition at single-cell resolution. Experiment 3 is designed to settle this directly rather than by argument, and if delayed inhibition turns out benign or even beneficial, that is itself a publishable, clinically important answer.

Decisive Checks

The proposal stands or falls on three outcomes: first, that a pathway-structured phase axis can be recovered from sepsis single-cell data and transfers across cohorts; second, that phase conditioning improves perturbation prediction beyond a cell-type-only model and an unstructured generative baseline at matched budgets; and third, that the timed-inhibition experiment confirms or refutes the model's pre-registered cell-type-level predictions for the delayed window. Any one of the three failing localizes the problem — the axis, the conditioning, or the biology — and the pre-registration makes refutation as informative as confirmation.