When a cell detects a threat, a transcription factor called NF-κB rushes into the nucleus and switches on hundreds of inflammatory genes. Turning those genes back off is just as important: the cell's own inhibitor, a protein called IκBα, is thought to enter the nucleus, physically peel NF-κB off promoter DNA, and drag it back to the cytoplasm. The field currently treats this shutdown as a genome-wide, sequence-blind eviction — every NF-κB-bound promoter is assumed to be stripped at roughly the same rate. Yet in bacteria, a short stretch of promoter DNA between the -10 element and the transcription start site (the "discriminator region") is known to determine which repressors win access to which genes, down to single nucleotides. Nobody has asked whether mammalian inflammatory promoters carry analogous core-promoter sequence features that make some genes easy for IκBα to silence and others stubbornly resistant.
This proposal asks that question directly, and it matters for two reasons. First, therapeutic NF-κB blockers like proteasome inhibitors and pan-IKK inhibitors are toxic precisely because they shut everything off; understanding gene-selective termination is a prerequisite for any promoter-selective intervention, a goal motivating existing organ-protection work such as CTRP9-mediated blockade of TLR4/MyD88/NF-κB signaling in cardiomyocytes. Second, three methodological literatures — static co-immunoprecipitation (Co-IP, a technique that pulls one protein out of a cell lysate along with whatever physically sticks to it), end-point transcriptomics, and thermodynamic promoter models built in yeast and flies — have never been combined on a single mammalian signaling system, so the field has no way to even pose the question rigorously.
Core Claim
The kinetics of NF-κB transcriptional termination are gene-specific and dictated by core-promoter sequence architecture — a discriminator-like tract between the core promoter element and the transcription start site — rather than by uniform, sequence-blind stripping by IκBα. Concretely, I predict that (i) different NF-κB target promoters lose NF-κB occupancy at measurably different half-times after the IκBα feedback wave begins, (ii) these half-times are predictable from promoter sequence alone, and (iii) swapping the discriminator-like tracts between a fast-shutoff and slow-shutoff promoter transfers the shutoff kinetics without changing the activation amplitude. If the prediction fails uniformly across a designed promoter library, the negative result itself would settle a live question, since the uniform-stripping assumption in the IκBα post-translational modification literature has never been tested at promoter resolution.
Research Program
The program has three connected workstreams, each contributing a capability the prior literature lacks.
The first workstream builds a time-resolved promoter library for NF-κB targets. The yeast synthetic-promoter study measured 6,500 designed promoters in a single experiment by coupling each to a fluorescent reporter and sorting cells by fluorescence; I will port this design to a mammalian NF-κB context using a massively parallel reporter assay in which the κB motif, its spacing, and the discriminator-like tract (the sequence between the TATA/Inr core element and the start site) are systematically varied, anchored on promoters of the 236 RELA-dependent genes identified in the PDAC live-cell NF-κB atlas plus canonical targets like TNF and IL6. Cells are stimulated with TNFα and the reporter population is quantified at dense intervals (roughly 15 minutes to 6 hours) spanning activation, peak, and the IκBα-mediated decline. This directly addresses the open problem that promoter-binding studies rely on static end-point measurements, and it imports the bacterial discriminator concept of RpoS docking at a single -6G residue into a mammalian inflammatory system where, as the direction survey notes, eukaryotic discriminator-like rules remain unmapped.
The second workstream supplies what transcriptomics cannot: physical evidence of which protein complexes are actually dismantling promoter-bound NF-κB, and when. I will run time-resolved Co-IP on nuclear extracts at the same stimulation intervals, quantifying RELA–IκBα and RELB–RELA complex assembly and disappearance. This targets the documented blind spot that Co-IP proves two proteins bind in a lysate but not that the complex sits on a specific promoter, so Co-IP is paired with RELA ChIP-seq at matched time points to measure per-promoter occupancy decay in the endogenous genome, and with proximity ligation assay as an in-situ orthogonal check of the kind used by Foster and colleagues to avoid lysis artifacts. This fills the gap that no study in the reviewed material has performed time-resolved quantitative Co-IP across an inflammatory transition, and it tests the Butera et al. finding that RELB, induced by RELA, represses RELA nuclear entry — by asking whether RELB also acts at the promoter, as a physical complex measurable by Co-IP.
The third workstream is a mechanistic model that discriminates between the three candidate repression mechanisms the literature currently leaves in tension: direct steric exclusion, short-range local quenching as formalized in GEMSTAT, and one-dimensional sliding competition for a shared local transcription-factor pool as established in the yeast TF-sharing model. I will embed these alternatives in a kinetic ordinary differential equation whose per-promoter parameters are functions of library sequence, fit jointly to the reporter trajectories and the ChIP occupancy decay curves, and compare them on held-out promoters. The scaffold follows FLeCS and PerturbODE, learnable ODE models of single-cell gene dynamics, but with two constraints neither uses: experimentally measured physical interactions from the Co-IP workstream, and nucleotide-level promoter features in the interaction terms. This is deliberate, because PerturbODE validates edges only against database binding records and FLeCS initializes from FANTOM5 priors — neither is anchored to context-specific physical complexes.
Baselines
The naive reference point is the sequence-blind uniform stripping model implied by the IκBα review: every κB-containing promoter decays with the single half-time estimated from bulk nuclear RELA ChIP. It is the strongest null because it is the field's default assumption, it requires no sequence data, and it is what any sequence-dependence claim must beat. It will be fit to the same ChIP time courses with the same number of free parameters per promoter (one shared rate), and beating it means showing that promoter-specific decay half-times vary significantly beyond measurement noise and that sequence features explain a substantial fraction of that variance on held-out promoters.
The strongest existing predictor is the thermodynamic "TF sharing" model of the large-scale yeast promoter study, which that paper reports achieving $R^2 \approx 0.8$ against measured expression across its 6,500-promoter library; on the mechanism side the strongest competing structural model is the short-range repression variant of GEMSTAT, which on its Drosophila segmentation data improved average correlation from 0.547 to 0.587 with self-cooperative binding. I will refit both on our mammalian time-resolved library, giving them the same promoter sequences, the same tuning budget, and the same held-out promoter split as our kinetic model. Because these models were built for static expression in yeast and fly patterning rather than time courses in mammalian inflammation, the fair comparison extends them with the same decay term we use, so any advantage of our model is attributable to its sequence-to-kinetics mapping and physical-interaction constraints, not to having dynamics at all. A sparse linear model mapping promoter sequence features to decay half-times (an ElasticNet, following the comparator in the Virtual Cell Models critique) runs under the same budget; if it matches our model's held-out accuracy, the mechanistic story is unsupported and I will say so.
The key ablation removes the physical anchoring: the same ODE refit to reporter trajectories alone, without Co-IP- and ChIP-derived occupancy constraints. Given that the PerturbCausal benchmark found deep models without such anchoring inflated perturbation effect sizes by approximately 23-fold, I expect this ablation to systematically misestimate repression rates even when it fits the training trajectories well — which would be direct evidence that physical measurements, not just more data, are the binding constraint. As a secondary reference, PerturbODE achieved the highest AUPRC among causal models on its TF Atlas evaluation (no numerical AUPRC value is reported on Lacuna), so it will be run on our data as the strongest published dynamic model; on the question of promoter-level termination kinetics it makes no sequence-level prediction, which is precisely the comparison of interest.
Why not just…?
Why not just take end-point RNA-seq or qPCR snapshots? Because the quantity this proposal measures — how fast each promoter shuts off — is invisible at a single time point. The temporal literature is unambiguous on this: myocardial infarction profiling shows tissue shifting continuously from an inflammatory to a metabolic program, toxic-exposure trajectories show molecular footprints evolving from metabolic to cell-cycle effects, and the PDAC NF-κB study could only identify its delayed negative feedback regulators (NUAK2, ARHGAP31, RELB) across a 10-hour course. A 6-hour snapshot conflates promoters that peaked and shut down with promoters that never fully activated, which is exactly the distinction at stake.
Why not just query STRING or BioGRID for the protein interactions? Aggregate databases cannot tell us whether IκBα, RELA, and RELB form complexes in a TNFα-stimulated nucleus at minute 45, which is the only interaction that matters here. The IκBα synthesis (Post-translational Modifications of IκBα: The State of the Art) shows the IκBα–NF-κB interaction is governed by dynamic post-translational modifications — Ser32/Ser36 phosphorylation, Lys21 SUMOylation, Tyr42 phosphorylation — that no static database encodes, and the IQGAP3–GLI1 study is a direct precedent that Co-IP was required to rule out indirect cascade explanations that correlation-based inference could not. That said, this is a real cost: the Co-IP workstream is the most labor-intensive part of the program, and if the time-resolved ChIP alone already resolves per-promoter decay rates, the Co-IP results become mechanism-confirmation rather than discovery. The experiments are structured so that ChIP and Co-IP arrive together, letting the ChIP data justify or curtail the Co-IP effort.
Why not just assume the repressor sterically blocks the promoter? Because the best quantitative evidence says steric models fail even in simpler systems. Across 6,500 yeast promoters, adding an activator binding site reduced expression in roughly 30% of cases, and steric hindrance could not explain it; only the one-dimensional sliding "TF sharing" model fit, at $R^2 \approx 0.8$ (yeast promoter mapping). GEMSTAT likewise found short-range repression beat competitive binding for some Drosophila repressors but failed for others (Giant, Knirps), meaning the mechanism is repressor-specific and must be tested, not assumed. This is why the model comparison is a designed part of the proposal rather than an afterthought.
Why not just train a deep virtual cell model on perturbation data? Because the PerturbCausal benchmark showed that no deep virtual cell model significantly outperformed a sparse linear comparator on causal recovery, while inflating effect sizes roughly 23-fold. Building on such a model would risk importing inflated repression-rate estimates into the very quantity we are trying to measure. The ODE scaffold here is deliberately constrained by physical measurements for exactly this reason, and the ablation above tests whether that constraint earns its cost.
Proposed Experiments and Decisive Checks
The decisive experiment is a promoter-swap: CRISPR-edit the discriminator-like tract of a fast-shutoff endogenous gene (identified from the library and ChIP decay ranking) into a slow-shutoff gene's promoter, and vice versa. If shutoff kinetics transfer with the tract while activation amplitude is unchanged, sequence-dependence is causal and the uniform-stripping assumption is falsified; if kinetics are unchanged, the negative result constrains all future models of NF-κB termination to be sequence-blind, resolving Gap 1 in the opposite direction. Secondary checks include whether the fitted kinetic model, with its physical-interaction constraints, beats the refit TF-sharing and GEMSTAT baselines and the ElasticNet comparator on held-out promoters under the same budget; whether the no-Co-IP ablation reproduces the effect-size inflation pattern documented by PerturbCausal; and whether RELB, which Butera et al. showed represses RELA nuclear entry, also appears in promoter-proximal complexes measurable by our time-resolved Co-IP. Each check is falsifiable on its own, and any one of them failing tells us something the field currently does not know about how inflammation ends.
References
- C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway
- Post-translational Modifications of IκBα: The State of the Art
- Large-scale mapping of gene regulatory logic reveals context-dependent repression by transcriptional activators
- Actin networks modulate heterogenous NF-κB dynamics in response to TNFα
- Critical role for a promoter discriminator in RpoS control of virulence in Edwardsiella piscicida
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- Thermodynamics-Based Models of Transcriptional Regulation by Enhancers: The Roles of Synergistic Activation, Cooperative Binding and Short-Range Repression
- A scalable gene network model of regulatory dynamics in single cells
- Interpretable Neural ODEs for Gene Regulatory Network Discovery under Perturbations
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis
- New Insights Into the Persistent Effects of Acute Exposure to AFB1 on Rat Liver
- IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer