Inflammasome-driven inflammation sits at a well-documented crossroads of metabolic disease: excess glucose flows into the hexosamine pathway, and the resulting sugar modifications (O-GlcNAcylation) of NF-κB and its downstream machinery both prime and assemble the NLRP3 inflammasome, driving steatohepatitis, cytokine storm, and pyroptotic cell death (O-GlcNAcylation dictates pyroptosis). The obvious interventions — inhibit the sugar-attaching enzyme OGT, or shut down NF-κB transcription — fail for a structural reason: those enzymes are used by every tissue in the body, so systemic blockade poisons glucose homeostasis and cripples baseline immune surveillance. Recent work on other metabolic interfaces has shown a way out: pathological signaling can often be separated from essential housekeeping activity by blocking one specific protein–protein handshake rather than an enzyme's catalytic site (Adipocytic sclerostin loop3-LRP4 interaction required by sclerostin to impair whole-body lipid and glucose metabolism; The metabolic enzyme hexokinase 2 localizes to the nucleus in AML and normal haematopoietic stem and progenitor cells to maintain stemness). Yet no one has applied modern, sample-efficient molecular optimization to design small molecules against such an interface inside the inflammatory cascade, and no one has let the physical-interaction assays that validate these handshakes — co-immunoprecipitation (a pull-down assay that isolates intact protein complexes from cell lysate) and biolayer interferometry (a real-time binding-kinetics measurement) — act as the judge inside the optimization loop itself.
I propose exactly that: a closed-loop lead-optimization program that discovers small-molecule disruptors of the NEK7–NLRP3 physical interaction — the single protein handshake that assembles the inflammasome — where the search algorithm is rewarded only by experimental binding measurements, iterated across rounds, and evaluated for the explicit absence of systemic metabolic toxicity. The core scientific bet is that putting a wet-lab physical-interaction assay inside the loop closes the gap that every current method leaves open: generative lead-optimization models are graded against cheap computational stand-ins that say nothing about real interface binding or whole-body safety (Deep Lead Optimization: Leveraging Generative AI for Structural Modification; Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization).
Core Claim
The NLRP3 inflammasome cannot assemble without the kinase NEK7 physically docking onto NLRP3, and O-GlcNAcylation of NEK7 at Ser-260 actively promotes this docking (O-GlcNAcylation dictates pyroptosis). Because the assembly step is gated by a localized protein–protein interface rather than a catalytic pocket, a small molecule that wedges this interface apart should block mature IL-1β release and pyroptosis while leaving NF-κB transcriptional priming, global glycosylation, and glucose homeostasis untouched — exactly the safety profile that pan-OGT inhibition (which pushes GSDMD, NLRP3, GSDME, and NEK7 in conflicting directions simultaneously) and broad NF-κB shutdown cannot achieve. The claim has two parts: first, that interface-disrupting leads exist in a searchable chemical space at feasible experimental budgets; second, that an optimization loop scored by co-immunoprecipitation and biolayer-interferometry measurements of the NEK7–NLRP3 pair finds them faster than a loop scored by docking scores, because the physical assay — not the rigid-receptor approximation — is what the drug must actually perturb.
This differs from the closest existing work in both halves. Jiang et al. proved that a single interface (sclerostin loop3 binding adipocyte LRP4) mediates whole-body metabolic toxicity and can be blocked by a peptide (LA5) and an aptamer (Apc0010A), but they used biological reagents, did no small-molecule lead optimization, and never touched the inflammatory cascade (Adipocytic sclerostin loop3-LRP4 interaction required by sclerostin to impair whole-body lipid and glucose metabolism). Conversely, the strongest lead-optimization engines — POLO's preference-guided reinforcement learning (POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization), ScaffoldGVAE's scaffold-hopping autoencoder (ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks), and InVirtuoGen's discrete flow refinement (Refine Drugs, Don't Complete Them: Uniform-Source Discrete Flows for Fragment-Based Drug Discovery) — have only ever been graded on abstract chemical-property scores (solubility, QED, docking surrogates) against single enzymes with deep pockets. Neither literature has crossed the other's line: no study in this material applies constrained molecular optimization to a flat intracellular protein–protein interface, and none feeds physical-interaction assay readouts back into the search.
Research Program
The program has three coupled components, and each is grounded in a method already proven in the file.
Target definition from structure. The search space will be narrowed around a defined interface groove rather than left to roam a flat, featureless contact surface — the classic reason PPIs were abandoned. HIGH-PPI, a hierarchical graph network that models protein structure and identifies binding hot-spot residues via its explanation module (Hierarchical graph learning for protein–protein interaction), will be run on the NEK7–NLRP3 structural complex to rank interface residues; the region adjacent to NEK7 Ser-260 is prioritized because O-GlcNAcylation there competitively blocks phosphorylation and promotes the interaction (O-GlcNAcylation dictates pyroptosis). This supplies a 3D pharmacophore constraint, which matters: incorporating protein pocket structure into generative lead optimization raises the predicted binding probability of generated ligands from roughly 0.25 to 0.75 (Deep Lead Optimization: Leveraging Generative AI for Structural Modification).
Sample-efficient search over a synthesis-on-demand space. Because the oracle is a real assay, the budget is a few hundred measurements, not millions of queries. This dictates the algorithm choice honestly. The PMO benchmark showed that under a 10,000-call budget, the classical methods REINVENT and Graph GA beat recent complex deep models across all 23 tasks (Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization), and SynthonBench showed that below roughly 1,000 calls, simple factored bandits outperform genetic algorithms, with SALSA-style Thompson Sampling reaching 39.6x efficiency at a 1,000-sample budget (SynthonBench: Benchmarking Sample-Efficient Optimization in Combinatorial Chemical Spaces). I therefore do not propose a heavy generative model as the search engine. The engine is a typed genetic algorithm over a fragment-combinatorial (synthon) space — the regime where Typed GA closed 76% of the gap between random search and the optimum at a 10,000 budget (SynthonBench) — seeded with fragments that dock into the HIGH-PPI-defined groove. On top of it sits the one deep component that earns its place: a POLO-style preference model, because ranking compounds by measured binding is exactly the pairwise-preference signal that turn-level direct preference optimization consumes (POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization). Preference learning is also the right tool for a coarse oracle: BLI gives a rank order of candidates even when absolute values are noisy.
The experimental oracle loop. Each round works as follows. The algorithm proposes roughly 96 compounds. Recombinant NEK7 and NLRP3 yield biolayer-interferometry association and dissociation rates ($k_{\text{on}}$, $k_{\text{off}}$, and the dissociation constant $K_D$) in the presence of each compound — the same assay format Jiang et al. used to quantify sclerostin loop3 binding to LRP4 (Adipocytic sclerostin loop3-LRP4 interaction...). Compounds that measurably weaken the interaction advance to targeted co-immunoprecipitation of endogenous NEK7–NLRP3 in LPS- and nigericin-stimulated macrophages, and the top performers each round go to a proximity ligation assay in intact cells to rule out lysis artifacts, since detergent extraction can create false associations that the co-immunoprecipitation literature itself flags as a central failure mode. The measured rank order feeds back as preference pairs; the GA propagates winners. Five to eight rounds, a total oracle budget of 500–800 experimental measurements, is the design point.
Proposed Experiments
The in-silico phase first validates the hot-spot definition: HIGH-PPI's explanation module will be checked against interface-contact annotations in the Catalytic Site Atlas and PDBePISA, the ground-truth databases it was originally validated on (Hierarchical graph learning for protein–protein interaction). Its published robustness — up to 19% performance gain over the second-best baseline under 40% interaction-data perturbation — is the reason it can be trusted on a partially characterized interface where structural data is incomplete.
The optimization phase is scored on three metrics adapted from the PMO benchmark: hit rate within budget (fraction of proposed compounds achieving a pre-registered $K_D$ threshold for interface weakening), top-10 area under the accumulation curve as a function of oracle calls, and best-observed dissociation-constant fold-change per round. All search algorithms are run at exactly matched oracle budgets on identical libraries with identical fragment seeds.
The biochemical and cellular phase converts binding into function: reduction of the NEK7–NLRP3 co-immunoprecipitation band, reduced mature IL-1β and caspase-1 activation after LPS plus nigericin stimulation, and — the decisive selectivity signature — preserved TNF-α and IL-6 output, since NF-κB priming should be untouched by an assembly-step inhibitor. The natural-product NF-κB blockers in the file (aucubin, 3-deoxysappanchalcone, 4-octyl itaconate) suppress the whole cascade including priming (Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model; 3-Deoxysappanchalcone attenuates LPS-induced neuroinflammation in microglia cell culture and ameliorates cognitive impairment in traumatic brain injury; 4-Octyl itaconate alleviates endothelial cell inflammation and barrier dysfunction in LPS-induced sepsis via modulating TLR4/MAPK/NF-κB signaling), so they form the "shut everything down" contrast arm.
The safety phase directly tests the metabolic-toxicity premise. Comparator arms are the pan-OGT inhibitor ST045849 and broad AMPK activation, both of which the file documents as systemically toxic (O-GlcNAcylation dictates pyroptosis; Motor protein KIF13B orchestrates hepatic metabolism to prevent metabolic dysfunction-associated fatty liver disease). Readouts are global O-GlcNAc modification levels (which should be unchanged by an interface inhibitor), glucose tolerance, and T-cell glycolytic function, the last because broad metabolic inhibitors paralyze host effector T cells that share the same metabolic machinery as tumors (Metabolic gatekeepers: harnessing tumor-derived metabolites to optimize T cell-based immunotherapy efficacy in the tumor microenvironment). The pattern that counts as success is functional inflammasome blockade with no signal in any of these toxicity readouts.
Baselines
The naive reference point is random selection from the same synthon library at the same measurement budget, with no assay feedback — the reference against which SynthonBench itself calibrates efficiency, where Typed GA "closed 76% of the gap between random search and the global optimum" at a 10,000 budget (SynthonBench: Benchmarking Sample-Efficient Optimization in Combinatorial Chemical Spaces). Random search quantifies how much of the hit rate is simply the library and the pocket constraint rather than the loop.
The strongest existing competitor depends on budget, and both regimes will be run. At the 10,000-call scale, the PMO benchmark's leader is REINVENT, "ranked #1 overall across the metrics, followed closely by Graph GA" (Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization); Graph GA itself will be run because it is the classical search backbone my proposal modifies. At my actual experimental budget of ~500 measurements, the strongest published method is POLO, which achieves an 84% average success rate on single-property tasks within 500 oracle calls, a 2.3× improvement over its baselines, and 50% on simultaneous three-property optimization (POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization). POLO will be retrained on my preference pairs from the BLI oracle — same library, same fragment seeds, same number of wet-lab measurements — so the comparison isolates whether my typed-GA-over-synthons search or POLO's learned policy extracts more from identical experimental spend. Note that POLO's 84% was earned on abstract property oracles; no comparable number exists for a physical PPI assay, so beating POLO here means beating it on my benchmark, not on its own. ScaffoldGVAE is included as a scaffold-diversity reference (it reports a 92.4% success rate on LRRK2 with MM/GBSA energies of −40 to −60 kcal/mol, ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks) but is not the primary competitor, since it was built for monomeric kinase pockets, not interfaces.
The key ablation runs my exact search algorithm with a pure in-silico docking-score oracle instead of the BLI/co-immunoprecipitation oracle, holding everything else fixed. If the assay-in-the-loop version reaches its $K_D$ threshold at fewer experimental measurements than the docking version needs physical measurements to confirm its own top compounds, the central claim — that physical-interaction feedback beats proxy scoring — is supported. For the biological endpoint, the ablation comparator is the OGT inhibitor ST045849: if interface disruption matches ST045849 on IL-1β suppression while leaving global O-GlcNAc levels and glucose tolerance intact where ST045849 does not, the toxicity-decoupling claim survives.
Why not just…?
Why not just use a blocking peptide or an antibody, like LA5 for sclerostin loop3? This is the approach that already worked for the closest interface in the file (Adipocytic sclerostin loop3-LRP4 interaction...), so it is a real threat. It falls short here for two reasons grounded in the same literature: the NEK7–NLRP3 handshake happens in the cytoplasm of living cells, and full-length antibodies cannot cross intact cell membranes; and peptides suffer the hepatic clearance and oral-bioavailability problems documented for natural monomers and peptide-like leads in the file (The traditional chinese medicine monomer Ailanthone improves the therapeutic efficacy of anti-PD-L1 in melanoma cells by targeting c-Jun; Identification of novel leishmanicidal molecules by virtual and biochemical screenings targeting Leishmania eukaryotic translation initiation factor 4A). Still, a cell-penetrant LA5-mimetic peptide against the NEK7 groove is a legitimate fastervalidation tool, and I will run it as a positive-control arm: if the peptide blocks the interface in the co-immunoprecipitation assay but fails in animals, that contrast strengthens the small-molecule rationale; if the peptide succeeds in vivo, the proposal pivots to optimizing the peptide into a macrocycle, and the loop survives intact.
Why not just inhibit the upstream enzymes — IKK for NF-κB, or OGT with ST045849? Because the entire premise of systemic metabolic toxicity is that these enzymes are shared body-wide. AKT blockade causes dose-limiting hyperglycemia (Discovery of Novel AKT Inhibitors with Enhanced Anti-Tumor Effects in Combination with the MEK Inhibitor); broad AMPK activation has extensive off-tissue effects (Motor protein KIF13B orchestrates hepatic metabolism...); and OGT inhibition is worse than merely toxic — O-GlcNAcylation has opposite effects on different proteins in the same pathway, inhibitory on Gasdermin D Ser-338 but pro-pyroptotic on NLRP3, GSDME, and NEK7 (O-GlcNAcylation dictates pyroptosis), so pan-OGT inhibition produces internally conflicting phenotypes. Additionally, kinase-dead mutant experiments show pathological signaling often runs independently of catalytic activity (The metabolic enzyme hexokinase 2 localizes to the nucleus...), so enzyme inhibition can kill housekeeping function while sparing the disease function. The ST045849 arm in my experiments settles this empirically rather than by argument.
Why not just run a simple genetic algorithm or Thompson sampling over a docking-score oracle, with no wet-lab loop at all? This is the most serious objection, because the benchmarks agree with it at scale: REINVENT and Graph GA beat complex deep models under a 10,000-call budget (Sample Efficiency Matters), and simple bandits top SynthonBench below 1,000 calls (SynthonBench). My response is to concede the engine and keep the loop: the proposal already uses a typed GA as its search core, precisely because of this evidence. The disagreement is only about the oracle. Rigid-receptor docking is the documented weak link — the lead-optimization review notes that essentially all current frameworks assume static protein targets and fail on dynamic protein–protein interfaces (Deep Lead Optimization), and the NEK7–NLRP3 assembly is precisely such an induced-fit, post-translationally gated contact (O-GlcNAcylation dictates pyroptosis). The docking-only ablation is designed to answer this objection head-on: if it performs as well as the assay loop at equal spend, the field learns the loop is unnecessary for this target class, and that would itself be a decisive result.
Decisive Checks
The program succeeds if three pre-registered outcomes co-occur. First, the assay-guided loop beats its matched docking-only ablation on experimental measurements-to-hit — fewer wet-lab confirmations wasted on false positives from the proxy scorer. Second, the best lead suppresses mature IL-1β and caspase-1 while leaving TNF-α output and global O-GlcNAc levels intact, distinguishing assembly-step inhibition from the blanket NF-κB suppression produced by agents like 4-octyl itaconate (4-Octyl itaconate alleviates endothelial cell inflammation and barrier dysfunction in LPS-induced sepsis via modulating TLR4/MAPK/NF-κB signaling). Third, glucose tolerance and T-cell glycolytic function are unimpaired at efficacious concentrations, in contrast to the documented systemic failures of AKT, glutaminase, and pan-OGT inhibition. Any one of these failing redirects the program — a loop that cannot beat docking tells us the interface is too rigid-receptor-friendly to need it; a lead that poisons metabolism tells us the toxicity is not gated by the interface alone. The file reports no existing small molecule targeting the NEK7–NLRP3 interaction, so the first round of BLI measurements against naive library compounds is itself a contribution: an initial map of how flat and featureless this interface really is for drug-like chemistry.
References
- O-GlcNAcylation dictates pyroptosis
- Adipocytic sclerostin loop3-LRP4 interaction required by sclerostin to impair whole-body lipid and glucose metabolism
- The metabolic enzyme hexokinase 2 localizes to the nucleus in AML and normal haematopoietic stem and progenitor cells to maintain stemness
- Deep Lead Optimization: Leveraging Generative AI for Structural Modification
- Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization
- POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization
- ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks
- Refine Drugs, Don’t Complete Them: Uniform-Source Discrete Flows for Fragment-Based Drug Discovery
- Hierarchical graph learning for protein–protein interaction
- SynthonBench: Benchmarking Sample-Efficient Optimization in Combinatorial Chemical Spaces
- Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model
- 3-Deoxysappanchalcone attenuates LPS-induced neuroinflammation in microglia cell culture and ameliorates cognitive impairment in traumatic brain injury
- 4-Octyl itaconate alleviates endothelial cell inflammation and barrier dysfunction in LPS-induced sepsis via modulating TLR4/MAPK/NF-κB signaling
- Motor protein KIF13B orchestrates hepatic metabolism to prevent metabolic dysfunction-associated fatty liver disease
- Metabolic gatekeepers: harnessing tumor-derived metabolites to optimize T cell-based immunotherapy efficacy in the tumor microenvironment
- The traditional chinese medicine monomer Ailanthone improves the therapeutic efficacy of anti-PD-L1 in melanoma cells by targeting c-Jun
- Identification of novel leishmanicidal molecules by virtual and biochemical screenings targeting Leishmania eukaryotic translation initiation factor 4A
- Discovery of Novel AKT Inhibitors with Enhanced Anti-Tumor Effects in Combination with the MEK Inhibitor