Codon-Tuned Pyroptosis Blockers That Only Work in Reactive Microglia

Every cell's ribosomes read the genetic code through adapter molecules called transfer RNAs (tRNAs), and different codon "spellings" of the same amino acid are read at very different speeds depending on which tRNAs the cell happens to have in stock. This proposal exploits that fact to build a therapy for Alzheimer's disease (AD) that is silent in healthy brain immune cells and switches on only in the diseased ones. The target cells are microglia, the brain's resident waste-clearing immune cells. In chronic AD they tip into a reactive state and die by pyroptosis — a violent, inflammatory self-rupture executed by the protein gasdermin D (GSDMD), which caspase-1 cleaves into pore-forming fragments that perforate the cell membrane and release inflammatory signals like interleukin-1β into tissue. This microglial pyroptosis is now viewed as a driver of AD neuroinflammation, and depleting microglia this way also removes the cells that clear amyloid plaque Microglia Pyroptosis: A Candidate Target for Neurological Diseases Treatment. Existing drugs that block this pathway (NLRP3 inhibitors like MCC950, caspase-1 inhibitors like VX-765) do so indiscriminately, suppressing microglial function everywhere and compromising both infection defense and plaque clearance.

The core idea: instead of chemically inhibiting pyroptosis everywhere, we recode the synonymous codons of a dominant-negative GSDMD construct so that its translation is favored by the tRNA pool of reactive microglia and starved in homeostatic microglia. The translational machinery, not a drug or a delivery particle, becomes the targeting mechanism. Codon deoptimization is an established way to dial expression up or down without touching the protein sequence — L. Zhu et al. used it to attenuate live pathogens for vaccines — but every existing codon design tool conditions on a static, whole-organism host. We propose to condition codon design on cell state, learned from single-cell RNA sequencing (scRNA-seq) atlases of the AD brain.

Core Claim

Three strands of evidence make this plausible but leave the design problem unsolved. First, the tRNA pool of brain tissue is not static: in post-mortem human cortex, one tRNA isoacceptor, tRNA$^{\text{Gly}}_{\text{GCC}}$, is selectively upregulated in disease, and proteins whose transcripts are enriched in the matching GGC codon are selectively over-produced — 34 of 36 upregulated proteins shared this codon bias — with causal validation in a transgenic mouse Altered tRNA expression profile associated with codon-specific proteomic changes in the suicide brain. Second, translation diverges from mRNA abundance in a cell-type-specific way: Translatomer, a model that predicts ribosome footprints from sequence plus bulk RNA-seq, achieved Pearson correlations of 0.72 to 0.80 across 33 paired cohorts, and found that 65% of translation-disrupting disease variants act in a tissue- or cell-type-specific manner, with AD-associated variants hitting macrophages (the circulating cousins of microglia) far harder than neurons Translatomer. Third, codon optimality has direct disease consequences in AD brain: rare synonymous variants that shift codons from non-optimal to optimal in genes like DTL and SERPINB5 are significantly associated with accelerated entorhinal cortical thinning, presumably by abolishing programmed ribosomal pauses needed for co-translational folding Codon bias among synonymous rare variants is associated with Alzheimer's disease imaging biomarker.

What is missing is the bridge: scRNA-seq captures polyadenylated mRNA only, so it cannot see tRNAs or ribosome positions Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation, and every codon optimizer — CodonTranslator, which conditions on species taxonomy embeddings, and CodonMPNN, which conditions on discrete taxon labels — assumes a homogeneous, invariant host. We claim that the gap can be closed computationally: a per-state codon dwell-time profile can be estimated by marrying single-nucleus RNA-seq references (via the EPIC-unmix deconvolution machinery, which beat TCA and bMIND by 4.6% to 109.8% in accuracy and surfaced a microglia-specific eQTL for the AD risk gene AP3B2 that bulk methods missed entirely EPIC-unmix) with bulk Ribo-seq from macrophage/microglial activation time courses, then used as a design dictionary for state-selective constructs. The design objective is simply

$$ S = \frac{\text{TE}_{\text{reactive}}(\mathbf{c})}{\text{TE}_{\text{homeostatic}}(\mathbf{c})}, $$

where $\text{TE}_{s}(\mathbf{c})$ is the predicted translation efficiency of a coding sequence $\mathbf{c}$ in state $s$ — a function of its codon composition against the state's dwell-time profile — and we pick, among synonymous recodings that preserve the protein, the one maximizing $S$ subject to preserving known pause sites at domain boundaries.

Research Program

The program has three linked components. The first is computational: building a state-resolved codon kinetics atlas. We extend the Translatomer recipe — one-hot sequence (5′ UTR, CDS, 3′ UTR) plus an expression profile fed to a transformer that predicts ribosome occupancy — but the conditioning profile becomes a deconvoluted state profile: single-nucleus RNA-seq microglial states from ROSMAP/MSBB references are matched to bulk paired RNA-seq/Ribo-seq libraries from activated human macrophage and microglial systems, so the model learns how codon-level dwell times co-vary with activation state rather than with species. The second component is an observational test of the disease hypothesis: we ask whether the mRNAs of the pyroptosis executioners (GSDMD, CASP1, NLRP3, IL1B, IL18) are codon-biased in a way that reactive microglia translate preferentially — i.e., whether the disease state happens to be already "primed" at the translational level, which would both explain why mRNA abundance alone correlates with protein output at only about 0.60 and provide the natural codon vocabulary to exploit. The third component is the design and validation of state-selective payloads: a dual-fluorescence reporter to prove state-selective translation, followed by a codon-tuned dominant-negative GSDMD (a non-cleavable full-length mutant whose excess, uncleaved protein saturates caspase-1), packaged as mRNA in the blood-brain-barrier-crossing, microglia-homing nanoparticle architecture already demonstrated for other payloads Enhanced BBB penetration and microglia-targeting nanomodulator.

Baselines

The naive reference point is the untuned construct: the identical dominant-negative GSDMD protein sequence recoded with average human codon usage (or left wild-type), delivered identically. It should express in both resting and reactive microglia; its selectivity ratio $S$ should sit near 1. Any state-selectivity the codon tuning adds is measured against this reference, so the ablation is the experiment. The strongest existing computational competitor for the prediction task is Translatomer itself, which reports Pearson correlations of 0.72 to 0.80 predicting ribosome occupancy on its 33 paired cohorts, and gains of 5.3% accuracy and 20% error reduction when sequence is added to RNA-seq Translatomer. We will run Translatomer unchanged on the same held-out libraries, with the same input budget, and our model must beat it specifically on codon-position-level dwell-time resolution (Translatomer predicts footprint distributions, not codon kinetics; SWAMamba, which does predict elongation rates, reports human Pearson r = 0.641 versus 0.607 for its best baseline, RiboMIMO SWAMamba — that is the number to beat on codon-level elongation prediction in microglia-like data). On the therapeutic side, the strongest existing method in the file is the Res@TcMNP/ASO nanomodulator — roughly 123 nm lipid particles bearing transferrin-receptor and M1-microglia-homing peptides, carrying an antisense oligonucleotide and resveratrol, which reduced IL-6 and TNF-α, cleared plaques, and improved Morris water maze and Y-maze performance in APP/PS1 mice (no reported result found on Lacuna for exact behavioral latencies) Res@TcMNP/ASO. We treat it as the delivery vehicle and the competitor: our head-to-head is the same particles carrying our tuned versus untuned mRNA payload, matched dose and schedule, with the added readout that untuned matched particles already cover. Success means: the tuned payload suppresses cleaved GSDMD-N, IL-1β release, LDH release, and propidium iodide uptake in reactive microglia at least as well as pan-inhibition, while leaving resting microglial phagocytosis intact where small molecules do not, and achieving a reporter selectivity ratio $S$ of at least 5 versus the untuned construct's ~1.

Why not just…?

Why not just use the existing small-molecule inhibitors (MCC950, VX-765, disulfiram)? These are the obvious answer, and they work in animals, but the field's own review of them identifies the failure mode: pan-inhibition of caspase-1 or GSDMD suppresses systemic innate immunity and disables the microglial phagocytosis that clears amyloid and debris, and rodent-to-human discrepancies (mouse caspase-11 versus human caspase-4/5) complicate translation Microglia Pyroptosis review. Disulfiram-based GSDMD inhibition shows the same pan-inhibition concern in lupus models Disulfiram alleviates pristane-induced lupus. Our proposal's decisive check is the paired assay: tuned payload versus VX-765 in the same reactive microglia, plus a phagocytosis assay in resting microglia where the molecule should impair uptake and our construct should be silent. If the tuned construct cannot match small-molecule suppression of IL-1β while sparing phagocytosis, the codon-gating idea loses its main advantage and the project should not proceed to animals.

Why not just deliver an ordinary mRNA or ASO with the existing M1-homing nanoparticles? This is a real threat, because the nanoparticle platform already homes to reactive microglia at the delivery level Res@TcMNP/ASO. But delivery-level targeting is graded, not binary: particles leak into other cells, and activation states shift over time, so a payload that is intrinsically inert except in reactive-state translation adds a second, orthogonal gate. The honest test is the reporter experiment: measure untuned reporter expression in non-reactive cells that nonetheless received particles. If off-state expression is negligible, delivery targeting suffices and codon tuning is redundant; if it is substantial, the translational gate earns its place. We will run exactly that comparison rather than argue it.

Why not just use standard codon optimizers (CAI matching, CodonTranslator, CodonMPNN)? These tools condition on species, not cell state, so they cannot produce a state-selective design by construction — they would give the untuned baseline. Worse, they optimize in exactly the direction the AD genetics warns against: maximizing codon optimality abolishes programmed ribosomal pauses needed for co-translational folding, and in the ADNI cohort, variants shifting codons toward optimality in specific genes were associated with accelerated cortical atrophy Miller et al.. Classical frequency matching also leaves most variance unexplained — CodonMPNN recovers the wild-type codon only 24.8% of the time versus 20.9% for naive most-frequent-codon selection, and prefers the better-expressed synonymous sequence in just 72.4% of yeast fitness pairs CodonMPNN — evidence that there is real, learnable, non-trivial signal beyond frequency tables. Our design differs by conditioning on state profiles and by explicitly preserving pause sites rather than maximizing speed.

Why not just find the pyroptotic microglia with scRNA-seq and target them transcriptionally? Because mRNA is a poor proxy for the thing that matters. Translatomer's analysis puts the mRNA–protein correlation near 0.60, and pyroptosis specifically is governed post-translationally by caspase-1 cleavage of GSDMD, which transcript counts cannot see Translatomer. EPIC-unmix-style deconvolution, however powerful at the mRNA level, is blind to translation EPIC-unmix. scRNA-seq tells us which states exist; only the codon-level model can tell us what those states can build. The proposal uses scRNA-seq as the map and translational modeling as the road — neither alone answers the question.

Proposed Experiments

The computational experiments come first. The state-conditioned dwell-time model is trained on paired bulk RNA-seq/Ribo-seq libraries from human macrophage and microglial activation systems, with single-nucleus reference profiles from ROSMAP/MSBB supplying the state decomposition (following EPIC-unmix, which found 83–100% directional concordance with direct single-cell findings while detecting 57.4% of cell-type-specific AD differentially expressed genes that comparable single-cell studies missed EPIC-unmix). Held-out prediction is scored at two granularities: overall footprint prediction against Translatomer's reported 0.72–0.80 Pearson range, and codon-level elongation against the 0.641 SWAMamba human benchmark SWAMamba. The observational analysis then tests whether reactive-state profiles assign faster codon-level elongation to GSDMD, CASP1, and IL1B transcripts than homeostatic profiles do — the hypothesis is falsified here if pyroptosis-gene translation is state-invariant, which would kill the design premise before any wet-lab cost is incurred.

The wet-lab sequence then runs in human iPSC-derived microglia. First, a dual-reporter assay: the same fluorescent protein recoded twice, once with codons whose cognates are enriched in the reactive-state dwell-time profile and once with homeostatic-enriched codons, activated with oligomeric amyloid-β (the NLRP3 stimulus documented for AD microglia Microglia Pyroptosis review). Readouts are per-cell reporter intensity by flow cytometry across states, plus standard pyroptosis assays — cleaved GSDMD-N and caspase-1 p20 by Western blot, PI uptake kinetics, extracellular LDH, and mature IL-1β. Second, the therapeutic construct: codon-tuned versus untuned non-cleavable GSDMD delivered by the D-T7/MG1 nanoparticle route Res@TcMNP/ASO, assayed head-to-head against VX-765 for pyroptosis suppression and against nothing for preservation of phagocytic uptake of fluorescent amyloid in resting cells. The decisive check for the whole program is a two-by-two: tuned versus untuned payload, reactive versus resting microglia. If the tuned payload suppresses pyroptotic readouts in reactive cells while leaving resting-cell phagocytosis and viability intact, and the untuned payload does neither selectively, then codon-level state gating is doing work that neither chemistry nor delivery can do — and only then would the program move to APP/PS1 mice, using the same behavioral and plaque endpoints (Morris water maze, Y-maze, IL-6/TNF-α reduction) already established for the nanoparticle platform.