Enrichment-Scored Design of Cyclic Peptide Inhibitors of the NF-κB Signalosome

NF-κB is the master switch of the mammalian inflammatory response: a transcription factor complex held in the cytoplasm until a chain of intracellular proteins (the TLR4 receptor, the MyD88 adaptor, the IKK kinase complex) relays a danger signal and frees it to enter the nucleus and switch on cytokines like TNF-α and IL-1β. Shutting this cascade down would treat the cytokine storms of sepsis, ischemia-reperfusion injury, and neuroinflammation, but every existing modality fails somewhere: small molecules like the flavonoid morusin hit dozens of unrelated targets, linear peptides are shredded by proteases within minutes, and protein regulators need adenoviral delivery with its own toxicity. Cyclic peptides — peptides whose ends or side chains are covalently linked into a closed, rigid ring — occupy exactly the needed niche: large enough to block the broad, flat protein-protein interfaces that small molecules cannot, and rigid enough to survive proteases, yet fully synthetic.

This proposal argues that the missing piece is not a new generator but a new objective. Generative models for cyclic peptides now sample valid macrocycles in about a second Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints and reach low-energy, ring-closed conformations at high rates MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization, yet all of them are scored on physics alone — potential energy, backbone geometry, or binding free energy — and none has ever been aimed at the NF-κB signaling cascade. Meanwhile, functional enrichment analysis — the statistical method that asks whether a set of co-dysregulated genes is unexpectedly concentrated in one pathway — is used only after the fact, to describe disease tissue. We propose to invert both roles: use cross-dataset enrichment analysis to nominate the target interface, and use pathway-level enrichment signatures as the scoring function that decides which designed cyclic peptide is a good inhibitor. The central claim is that scoring candidates by pathway-level selectivity, rather than by single-gene markers or binding energy, is what will let a cyclic peptide dampen NF-κB without the collateral pathway collapse that makes total NF-κB ablation toxic.

Core Claim

A design loop that (i) selects the target protein-protein interface inside the NF-κB priming cascade by consensus functional enrichment across disease single-cell datasets, (ii) generates cyclic peptide binders to that interface with existing generative scaffolds, and (iii) ranks candidates by a transcriptomic pathway-selectivity score — measured as NF-κB gene-set depletion relative to depletion of unrelated pathways in perturbed immune cells — will produce macrocyclic NF-κB inhibitors that are simultaneously protease-resistant, deliverable, and selective in a way that neither small molecules, linear peptides, nor energy-scored generative design have achieved. The proposal's deliverable is the loop itself plus a small validated candidate set, not a single clinical lead.

Research Program

The work has three connected stages, each grounded in an identified gap in the literature.

Stage 1: Enrichment-based target nomination. Existing NF-κB suppression studies pick their target by intuition or by candidate-gene assays. We instead run overrepresentation analysis (the hypergeometric test that asks whether disease-associated genes are enriched in a given KEGG pathway) on the public single-cell datasets already profiled in the literature: traumatic brain injury microglia (GSE167459), where the NF-κB and chemokine pathways are coordinately upregulated and calcium signaling and stem-cell pluripotency are downregulated Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases, and aortic dissection tissue (GSE213740), where network pharmacology isolated seven hub targets including NFKB1, MAPK8, MAPK14, and CASP3 Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection. The output is a ranked list of intracellular interfaces shared across pathologies — the MyD88–IKK interaction, the IKK substrate-docking surface on IκBα, and the p65/NFKB1 DNA-binding interface are the primary candidates — each backed by a pathway-level rather than single-gene rationale. This directly addresses the topological blindness of overrepresentation analysis itself: enrichment names the pathway, and the protein-interaction wiring of the KEGG NF-κB diagram tells us which node in that pathway is an upstream bottleneck versus a redundant effector.

Stage 2: Constrained generation with a dual physicochemical filter. We generate cyclic peptide binders at the nominated interface using two complementary scaffolds, both with public implementations: MuCO mianqiu00/MuCO codebase, whose three-stage flow-matching-plus-force-field pipeline reached a mean potential energy of -160.71 kJ/mol on the CPSea PDB gold standard with roughly 95% single-sample success MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization, and CP-Composer, whose composable geometric constraints permit zero-shot generation of disulfide, bicyclic, and head-to-tail topologies against any specified distance and residue-type constraints Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints. Generated candidates then pass through two filters that the literature shows are non-negotiable. First, permeability is screened with MSF-CPMP, which fuses sequence, graph, and 103 tabular physicochemical descriptors and reaches 0.9062 accuracy and 0.9546 AUROC — but which collapses by 80.1% when bulk physicochemical descriptors are removed MSF-CPMP: A Novel Multi-Source Feature Fusion Model for Cyclic Peptide Membrane Permeability Prediction; this ablation result is our license to optimize directly on physicochemical descriptors rather than sequence alone. Second, we explicitly bound the 3D hydrophobic moment rather than maximizing it, because the amphiphilicity-tuning study proved that fully amphipathic cyclic peptides enter cells 8- to 9-fold more efficiently but then remain trapped in endosomes as punctate confocal signal Achieving enhanced cell penetration of short conformationally constrained peptides through amphiphilicity tuning — and the NF-κB machinery sits in the cytosol and nucleus, beyond that trap. This tension between uptake and escape is the least-studied constraint in the file, and we treat the hydrophobic-moment band (enough for entry, not so much that the peptide is endocytosed and sequestered) as a first-class design variable, tuned empirically in the validation stage.

Stage 3: Enrichment-scored phenotypic ranking. Top-ranked candidates are synthesized and applied to LPS-stimulated macrophages, followed by bulk RNA-seq and KEGG/GO overrepresentation analysis. The ranking metric is a pathway selectivity score. Let $z_P$ be the normalized enrichment depletion of pathway $P$ (the negative of the enrichment log-fold change of the upregulated direction), and let $\mathcal{U}$ be the set of pathways upregulated in injury but mechanistically distinct from NF-κB, such as the calcium-signaling and stem-cell-pluripotency programs seen in TBI microglia Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases. Then

$$S = z_{\text{NF-}\kappa\text{B}} - \operatorname{median}_{P \in \mathcal{U}} z_P,$$

so that $S$ is large only when the NF-κB program is selectively suppressed while unrelated programs are spared. This is the formal answer to the field's central open problem — that total NF-κB ablation destroys host immunity and tissue repair — because a candidate that depletes every pathway scores $S \approx 0$. No study in the file has embedded any enrichment statistic in a design or screening loop; enrichment has been strictly post-hoc diagnostic Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases, and existing screening loops score on structural or single-assay labels instead, such as the target-adaptive reinforcement-learning screen that discarded 92.5% of a 36-million-candidate library in 14 days using a cytotoxicity label A scalable reinforcement learning approach for screening large peptide libraries for bioactive peptide discovery.

Proposed Experiments

The computational evaluation uses benchmarks with reported numbers so that progress is measurable. On structure generation, we compare our constrained generation against CP-Composer, which reports constraint-satisfaction success rates of 38% to 84% across cyclization strategies at 1.42 seconds per candidate on the 93-complex LNR benchmark, and against MuCO's -160.71 kJ/mol mean potential energy on the CPSea PDB subset with its ~95% single-sample success rate MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization; both numbers come from the cited papers. Binding-pose quality on the nominated NF-κB interfaces is scored with DockQ, for which PepTri reports 0.63 on PepBench along with a 0.40 binding success rate at $\Delta G < -5$ REU PepTri: Tri-Guided All-Atom Diffusion for Peptide Design via Physics, Evolution, and Mutual Information; since the NF-κB targets are not in PepBench, we build a held-out test set of peptide-protein complexes at kinase-adaptor interfaces and report both in-distribution PepBench numbers (for comparability) and our new interface set (for the actual question). Permeability filtering is validated on the MSF-CPMP benchmark at its reported 0.9062 accuracy. No reported result exists on Lacuna for any NF-κB-targeted cyclic peptide generation task, so our numbers there will be the first.

The wet-lab validation is deliberately staged. First, protease resistance: serum half-life of the top candidates versus their linear precursors, where the benchmarks to beat are the hours-long extension seen with disulfide-constrained angiotensin derivatives Angiotensin II-derived constrained peptides with antiplasmodial activity and suppressed vasoconstriction and the up-to-18-fold enhancement from photo-induced thiol–yne stapling Stapling of unprotected helical peptides via photo-induced intramolecular thiol–yne hydrothiolation. Second, uptake and localization: flow cytometry for entry efficiency plus confocal imaging for endosomal escape, since the amphiphilicity study showed punctate sequestration is the failure mode to overcome Achieving enhanced cell penetration of short conformationally constrained peptides through amphiphilicity tuning. Third, the decisive experiment: LPS-primed macrophages treated with candidate or control, RNA-seq, and the selectivity score $S$ computed from KEGG enrichment, with p65 nuclear translocation and TNF-α/IL-6 levels as orthogonal single-gene confirmations. A candidate succeeds only if it both suppresses the NF-κB enrichment signature and spares the unrelated $\mathcal{U}$ pathways — the picture that single-gene assays demonstrably cannot capture, as the TBI microglia study showed coordinated NF-κB and chemokine upregulation alongside simultaneous downregulation of calcium and pluripotency programs Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases.

Baselines

The naive reference point is the untreated disease state: LPS-stimulated macrophages with no inhibitor, whose RNA-seq defines the ceiling NF-κB enrichment signature, and vehicle-treated cells, which calibrate the noise floor of the selectivity score. Any candidate must move $S$ beyond the vehicle band to count at all.

The strongest computational competitor is CP-Composer run zero-shot with geometric constraints derived from the nominated NF-κB interface Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints. It is the strongest because it is the only existing system that can target an arbitrary new interface without retraining, reporting 38%–84% constraint-satisfaction success across cyclization strategies at 1.42 seconds per candidate on LNR. To make the comparison fair, we give CP-Composer the same interface constraints, the same candidate-generation budget (10,000 structures per target per method), the same molecular-dynamics refinement protocol, and the same permeability and hydrophobic-moment filters; the only difference is our enrichment-guided target nomination and selectivity-based ranking. We beat it if our loop's candidates show higher DockQ at the nominated interfaces and, more importantly, higher $S$ in cells, because CP-Composer optimizes neither. MuCO is run under the same budget as a second generator baseline, scored against its reported -160.71 kJ/mol CPSea PDB energy and ~95% success rate MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization. On the screening side, TARSA's target-adaptive reinforcement learning A scalable reinforcement learning approach for screening large peptide libraries for bioactive peptide discovery is the closest existing loop and is reconfigured with a permeability label in place of cytotoxicity at the same experimental assay budget; it is a real threat because its active-learning machinery is mature, and it would beat us if a generic assay-driven screen finds equally selective candidates without target nomination.

The decisive ablation isolates the proposal's two key ingredients separately. Ablation A removes enrichment-guided target nomination: the same generators and filters are aimed at a single-gene target (the top dysregulated gene from each dataset) rather than the consensus interface. Ablation B removes selectivity scoring: candidates are ranked by binding free energy alone, as all current generators do, then tested in cells. If either ablation matches the full loop's $S$, the corresponding ingredient is doing no work and the claim is weakened accordingly. A biological comparator, morusin, is included in the cell assays at equipotent dosing: it is the strongest existing chemical suppressor of the NF-κB axis in the file, docking to seven hub proteins at -9.0 to -7.0 kcal/mol Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection, and we expect it to suppress NF-κB but with broad collateral depletion across its 84-target polypharmacology — exactly the contrast the selectivity score is built to expose.

Why not just…?

Why not just use small molecules like morusin or aucubin? They are already cell-permeable and have documented NF-κB suppression in aortic dissection and spinal cord injury models Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection, Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model. The honest answer is that they may match our peptides on efficacy in vitro, and our morusin comparator arm is designed to find out. But the specificity argument is grounded in the file: morusin's own network pharmacology shows 84 overlapping targets across three master pathways, and small molecules routinely lack the surface area to selectively block shallow protein-protein interfaces. Our decisive check is the selectivity score — if morusin achieves the same $S$ as our best cyclic peptide, the peptide case weakens to a pharmacokinetic argument only, and we will report that.

Why not just use linear bioactive peptides like YPFPGPIH or AFYRW? Both already suppress NF-κB-driven inflammation in macrophage and endothelial models Bovine β-Casein Peptide YPFPGPIH Regulates Inflammation and Macrophage Activity via TLR/NF-κB/MAPK Signaling, Tartary buckwheat protein-derived peptide AFYRW alleviates H2O2-induced vascular injury via the PI3K/AKT/NF-κB pathway. This alternative fails on a well-quantified axis: unstructured linear backbones are cleaved within minutes in vivo, whereas cyclization extends half-life from minutes to hours Angiotensin II-derived constrained peptides with antiplasmodial activity and suppressed vasoconstriction and stapling yields up to 18-fold extension Stapling of unprotected helical peptides via photo-induced intramolecular thiol–yne hydrothiolation. We will include a cyclized version of the best linear comparator as a control, so the serum half-life and $S$ comparisons between matched linear and cyclic sequences are direct rather than inferred.

Why not just deliver a protein regulator like CTRP9 via adenovirus? Ad-CTRP9 already shuts down TLR4/MyD88/NF-κB in hypoxia/reoxygenation cardiomyocytes, reducing TNF-α and IL-6 and blocking p65 DNA binding C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway. Viral delivery is a genuine threat for ex vivo and preclinical settings, and we do not claim to beat it there. Our claim is about a synthetic, dose-controlled, repeat-dosing modality free of vector immunogenicity and multiplicity-of-infection tuning, which the CTRP9 study itself had to optimize carefully to avoid vector toxicity. A cyclic peptide mimicking the same interface contacts is the like-for-like synthetic counterpart, and the cell-assay arm includes the readouts (LDH, CK-MB analogues, TNF-α, IL-6) needed for a fair mechanistic comparison.

Why not just run RT-qPCR for TNF-α, IL-6, and phospho-p65 instead of RNA-seq plus enrichment? This is the cheapest alternative and the one the field actually uses — the CTRP9 study established efficacy entirely this way. It falls short on the specific question this proposal asks: single-gene assays cannot distinguish selective NF-κB suppression from general pathway collapse or cellular toxicity, which is the central open problem of NF-κB inhibition, whereas the TBI microglia study showed that enrichment analysis reveals simultaneous multi-pathway shifts — NF-κB and chemokine up, calcium and pluripotency down — that isolated markers miss Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases. Our design uses qPCR as a confirmatory tier, not the primary readout, precisely because the selectivity score $S$ is undefinable without transcriptome-wide enrichment.

Decisive Checks

The proposal stands or falls on three outcomes. First, the ablations: if single-gene target nomination or binding-energy-only ranking matches the full enrichment-guided loop's selectivity score, the methodological claim is falsified and we will say so. Second, the morusin comparator: if it matches our best candidate on $S$, the peptide case rests only on pharmacokinetics. Third, the endosomal-escape constraint: if no candidate within the hydrophobic-moment band both enters cells and reaches the cytosolic NF-κB machinery, we will have mapped the delivery boundary quantitatively — a negative result the field currently lacks, since the amphiphilicity study documents the trap but no design method has ever been constrained by it. Each of these outcomes is informative, which is what makes the loop worth building.

References

  1. Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints
  2. MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization
  3. Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases
  4. Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
  5. MSF-CPMP: A Novel Multi-Source Feature Fusion Model for Cyclic Peptide Membrane Permeability Prediction
  6. Achieving enhanced cell penetration of short conformationally constrained peptides through amphiphilicity tuning
  7. A scalable reinforcement learning approach for screening large peptide libraries for bioactive peptide discovery
  8. PepTri: Tri-Guided All-Atom Diffusion for Peptide Design via Physics, Evolution, and Mutual Information
  9. Angiotensin II-derived constrained peptides with antiplasmodial activity and suppressed vasoconstriction
  10. Stapling of unprotected helical peptides via photo-induced intramolecular thiol–yne hydrothiolation
  11. Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model
  12. Bovine β-Casein Peptide YPFPGPIH Regulates Inflammation and Macrophage Activity via TLR/NF-κB/MAPK Signaling
  13. Tartary buckwheat protein-derived peptide AFYRW alleviates H2O2-induced vascular injury via the PI3K/AKT/NF-κB pathway
  14. C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway