Dock-and-Pulldown Inhibitors of the Nuclear Bcl-3–NF-κB Complex in Renal Fibrosis

Whatever the initial insult — diabetes, a blocked ureter, a toxic drug dose — chronic kidney disease tends to end the same way: renal fibrosis, the relentless deposition of collagen scar tissue that replaces working kidney tissue. Two engines drive this scarring and feed each other. One is the TGF-β/Smad circuit, which tells cells to build matrix; the other is NF-κB, the cell's master inflammatory switch, whose activation in injured tubules recruits macrophages and amplifies tissue damage. Both engines converge in the nucleus, where a protein called Bcl-3 — an atypical member of the IκB family that sits inside the nucleus and regulates NF-κB transcription rather than holding it captive in the cytoplasm — has been shown to be an active biomarker and participant in fibrosis in chronic kidney disease patients and animal models (Bcl-3 is a novel biomarker of renal fibrosis in chronic kidney disease). What Bcl-3 physically grips during fibrogenesis, and whether that grip can be pried open with a small molecule, is unknown.

The proposal, in two sentences: we will use co-immunoprecipitation — an assay in which an antibody tethered to a bead fishes one protein out of a cell lysate, dragging along whatever proteins were physically attached to it — both to discover the nuclear protein complex that Bcl-3 assembles in injured kidney cells and to accept or reject candidate drugs designed to break that complex apart. Candidates will come from molecular docking — programs that search for the best-shaped fit of a small molecule into a pocket on a protein surface and score the fit — so the loop runs from pulldown, to structure, to docking, back to pulldown, and finally into mouse models of renal fibrosis.

Core Claim

The claim is that fibrosis progression depends on a physical nuclear scaffold assembled around Bcl-3 that couples NF-κB-dependent inflammation to Smad-driven matrix synthesis, and that this scaffold — not the upstream kinases — is the druggable point where both outputs can be dampened without systemic immunosuppression. Prior work motivates each piece but never connects them. Integrating network pharmacology and experimental validation to decipher the pharmacological mechanism of DXXK in treating diabetic kidney injury docks saponins against TGF-β1, Smad2, and Smad3 with AutoDock Vina (scores below −5 kcal/mol) and shows efficacy in diabetic mice, but performs no physical binding assay and never examines NF-κB. Up-regulated CD38 by daphnetin alleviates lipopolysaccharide-induced lung injury via inhibiting MAPK/NF-κB/NLRP3 pathway docks daphnetin into CD38 to explain NF-κB suppression, but in lung, and again with no pulldown validation. Conversely, ISG15 accelerates acute kidney injury and the subsequent AKI-to-CKD transition by promoting TGFβR1 ISGylation and NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17 use co-immunoprecipitation to validate renal protein complexes but never aim structure-based design at those interfaces. And Cepharanthine Ameliorates Pulmonary Fibrosis by Inhibiting the NF-κB/NLRP3 Pathway, Fibroblast-to-Myofibroblast Transition and Inflammation shows NF-κB blockade can blunt fibrosis, but in lung, with neither docking nor pulldown. No study in the file closes the loop between a computationally designed inhibitor and a pulldown-verified disruption of a specific renal protein complex. That loop, pointed at a nuclear co-factor rather than an upstream kinase, is what is new here.

Research Program

The first stage maps the complex. We will run Bcl-3 bait pulldowns followed by mass spectrometry on human HK-2 proximal tubular epithelial cells stimulated with TGF-β1 plus TNF-α, and on day-7 unilateral ureteral obstruction (UUO) mouse kidney lysates, with mild cross-linking to catch partners that would otherwise dissociate during the wash cycles. Reciprocal pulldowns with p50 and Smad3 as baits will test the specific hypothesis that Bcl-3 physically bridges NF-κB p50/p65 dimers with Smad3 or a shared co-activator such as p300, converting phenotypic crosstalk into a single assembled machine. The precedents are the NAT10–DDX17 discovery by pulldown-mass spectrometry in HK-2 cells (NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17) and the pulldown proof that IQGAP3 activates GLI1 by direct physical association rather than an indirect kinase chain (Chang Li et al.). Because lysis can force proteins that never met in the intact cell to associate, we will confirm every candidate partner in intact cells with a proximity ligation assay, as practiced in Ian T. Foster et al.. Public expression datasets (GSE207587 for UUO, GSE216376 for cisplatin, GSE109564 for human allograft biopsies) will tell us which partners are co-induced in injured tissue.

The second stage builds the structure. We will model the Bcl-3-centered assembly with AlphaFold 2, which ISG15 accelerates acute kidney injury and the subsequent AKI-to-CKD transition by promoting TGFβR1 ISGylation already used to predict an interaction it then confirmed by pulldown. Interface residues will be predicted with the DIPS-Plus-trained interface model, which reports a median AUROC of 0.9473, against 0.902 for the same model trained on the original DB5 and 0.885 for the SASNet baseline (DIPS-Plus: The enhanced database of interacting protein structures for interface prediction). A rigid-body pose will come from DiffDock-PP, which achieves a median complex-RMSD of 4.85 Å on the DIPS benchmark, beating HDOCK at 6.23 Å and EquiDock at 13.30 Å (DiffDock-PP: Rigid Protein-Protein Docking with Diffusion Models), followed by flexible refinement because that method, by its authors' own admission, assumes a rigid backbone.

The third stage docks a curated library of natural products and approved drugs into the modeled Bcl-3 contact groove — deliberately with classical tools under physical-validity checks, not generative models alone. The reason is quantitative: PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences reports that on the 308-complex post-2021 PoseBusters set, AutoDock Vina produced valid-and-accurate poses for 58% of complexes and GOLD for 55%, while DiffDock managed only 12% despite its 38.2% top-1 success rate (RMSD < 2 Å) on the PDBbind 2019 time-split test in its own paper (DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking); on the Astex set, DiffDock's 72% geometric success fell to 47% once basic steric and chemical checks were enforced. Since our targets have no co-crystal structures, we are squarely in the regime where generative docking fails, so Vina and GOLD under PoseBusters checks carry the screen, and DiffDock is run only as a tool ablation we expect to fail the checks.

The fourth stage is the proposal's defining ingredient: docking rank alone never advances a compound; only the pulldown does. Compounds are advanced from the top-scoring decile only if they break the complex in real lysates:

$$d_i = 1 - \frac{I_i}{I_0},$$

where $I_i$ is the Western-blot intensity of p50 (or Smad3) co-precipitated with Bcl-3 from compound-treated HK-2 lysates and $I_0$ the same from vehicle-treated lysates; a compound advances only if $d_i > 0.5$ at a concentration below its cytotoxic threshold. The fifth and sixth stages test function and then the animal: dual-luciferase NF-κB and Smad-responsive reporters, Western blots for collagen I/III, α-SMA, vimentin, phospho-p65 Ser536 and phospho-Smad2/3, followed by UUO mice dosed from day 3 to day 14 and a second cisplatin AKI-to-CKD cohort to probe cross-etiology robustness, with blinded Masson's trichrome and Sirius Red quantification, collagen I and α-SMA immunohistochemistry, and F4/80 macrophage counts. Because pathway blockade is known to shunt signaling into compensatory arms, we will explicitly monitor non-canonical RelB and JAK/STAT readouts.

Baselines

The "do nothing" reference point is the vehicle-treated UUO mouse at day 14, quantified on the same histological metrics as the intervention arms, alongside sham-operated controls. This anchors the natural history of scarring in our hands. For an effect-size scale we use the strongest genetic reference in the file: P311 knockout mice, which showed a roughly 1.43-fold decrease in interstitial collagen deposition, a 3.05-fold decrease in α-SMA protein, and a drop in F4/80-positive macrophages from 256 to 197 cells (P311 promotes renal fibrosis via TGFβ1/Smad signaling); a small molecule that reaches even half of that genetic suppression of collagen would be a serious lead.

The strongest existing method as a comparator arm is cepharanthine, the only compound in the file demonstrated to inhibit the NF-κB/NLRP3 axis and reduce established fibrosis (bleomycin-induced lung fibrosis with reduced myofibroblast transition, Cepharanthine Ameliorates Pulmonary Fibrosis by Inhibiting the NF-κB/NLRP3 Pathway, Fibroblast-to-Myofibroblast Transition and Inflammation). It has never been tested in kidney: no reported result found on Lacuna for cepharanthine in any renal fibrosis model. We will run it through the identical UUO protocol, dosing schedule, and blinded quantification as our leads. Beating it means significantly greater reduction in collagen-positive tissue area, with suppression of both p65 nuclear translocation and Smad2/3 phosphorylation, and without the compensatory pathway rise that a blunt upstream block often triggers.

The strongest existing workflow competitor is the DXXK pipeline of Integrating network pharmacology and experimental validation to decipher the pharmacological mechanism of DXXK in treating diabetic kidney injury: network pharmacology, AutoDock Vina, and in vivo validation, with compounds accepted at Vina scores below −5 kcal/mol and no physical binding assay. We will re-run AutoDock Vina on our modeled Bcl-3 pocket under the same compute and tuning budget, take its top-scoring compounds forward, and compare them head-to-head with our pulldown-confirmed compounds in the disruption assay — the ablation that isolates whether the pulldown-acceptance step adds anything beyond docking rank. A second ablation isolates the interface definition: we will dock the same library against a decoy surface patch of Bcl-3 away from the modeled contact groove; if decoy-pocket compounds disrupt the pulldown as often as interface-pocket compounds, the structure model is not earning its place.

Why not just…?

Why not just screen compounds in fibrotic mice and skip the structural work? Phenotypic screening can find efficacious compounds, but the file shows why it under-determines mechanism and safety. The DXXK network analysis generated 2,216 candidate disease genes — an unmanageable search space that docking had to collapse into three saponins and three targets (Integrating network pharmacology...), and the rutaecarpine case shows a lethal arrhythmia mechanism, binding inside the hERG potassium channel pore, that no phenotypic kidney assay would surface (Rutaecarpine targets hERG channels...). Our functional endpoints are the same histology and blots a screen would use, so we are not slower to efficacy; we are simply able to say which physical interaction each compound breaks, and to triage toxicity by structure. The Vina-only ablation arm settles this empirically: if unguided or rank-only compounds match pulldown-confirmed ones in mice, the reviewer wins and we report it.

Why not just trust AlphaFold and docking predictions and skip the pulldowns? This is the most serious alternative, and the file's evidence cuts against it hard. PoseBusters shows generative docking collapses on unseen targets — 12% valid-and-accurate versus 58% for Vina — and even classical poses are only predictions until a physical assay speaks. Tellingly, the two renal papers that used AlphaFold 2 to predict an interaction (ISG15...) then turned to pulldown to prove it, and NAT10... discovered its key complex by pulldown-mass spectrometry, not computation. Our loop does not discard prediction; it prices it. The dose-response correlation between $d_i$ and downstream reporter suppression across the compound series is the experiment that tells us whether predicted binding tracks real engagement.

Why not just use an existing IKK or TGFβR1 kinase inhibitor? Broad upstream blockade has documented liabilities in this very literature: global TGF-β targeting is clinically unviable, which is why TGF-β/Smad3 Signaling Promotes Renal Fibrosis by Inhibiting miR-29 argues for downstream, fibrosis-specific targets, and systemic NF-κB ablation destroys basal immune defense — the reason C1q/TNF‑related protein 9... and Aucubin... pursue modulated rather than ablated signaling. A nuclear co-factor interface should, in principle, leave IKK-dependent acute host defense intact while disabling the fibrogenic transcriptional module. But an IKK inhibitor might still work in a 14-day mouse model, and we will not argue around that: an IKK inhibitor arm belongs in the UUO comparison, and if it matches our leads on collagen while our compounds show better preservation of basal NF-κB output and less compensatory RelB/JAK shunting, that is the result that justifies the harder route.

Decisive Checks

Three outcomes decide the proposal. First, pulldown-mass spectrometry must identify a Bcl-3-centered nuclear complex containing NF-κB subunits and at least one fibrogenic partner (Smad3 or a co-activator) in UUO kidneys; if Bcl-3's partners are only the known IκB-family proteins, the convergence premise fails and the project honestly pivots to whatever interactome is found. Second, at least one pulldown-confirmed compound must reduce collagen deposition in UUO mice well beyond vehicle and at least match cepharanthine on the same blinded metrics. Third, across the compound series, pulldown disruption must correlate with functional suppression in a dose-dependent manner — if $d_i$ and reporter inhibition are uncorrelated, our compounds are acting off-target and the loop, however elegant, has not delivered a mechanism.

References

  1. Bcl-3 is a novel biomarker of renal fibrosis in chronic kidney disease
  2. Integrating network pharmacology and experimental validation to decipher the pharmacological mechanism of DXXK in treating diabetic kidney injury
  3. Up-regulated CD38 by daphnetin alleviates lipopolysaccharide-induced lung injury via inhibiting MAPK/NF-κB/NLRP3 pathway
  4. ISG15 accelerates acute kidney injury and the subsequent AKI-to-CKD transition by promoting TGFβR1 ISGylation
  5. NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17
  6. Cepharanthine Ameliorates Pulmonary Fibrosis by Inhibiting the NF-κB/NLRP3 Pathway, Fibroblast-to-Myofibroblast Transition and Inflammation
  7. IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer
  8. HPF-MPI: A Programming System Supporting Task and Data Parallelism
  9. DIPS-Plus: The enhanced database of interacting protein structures for interface prediction
  10. DiffDock-PP: Rigid Protein-Protein Docking with Diffusion Models
  11. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences
  12. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking
  13. P311 promotes renal fibrosis via TGFβ1/Smad signaling
  14. Rutaecarpine targets hERG channels and participates in regulating electrophysiological properties leading to ventricular arrhythmia
  15. TGF-β/Smad3 Signaling Promotes Renal Fibrosis by Inhibiting miR-29
  16. C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway
  17. Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model