Protein-Grounded Screening of Dual-Action Antimicrobial Peptides

Antibiotic-resistant infections kill not just through the bacteria themselves but through the host's own runaway response: resistant pathogens trigger a self-amplifying loop in which reactive oxygen species activate the NF-κB transcription factor, which drives inflammatory cytokines, which recruit immune cells that release yet more reactive oxygen species. The most promising countermeasure demonstrated so far is the "dual-action" agent — a molecule or peptide that both kills bacteria directly and simultaneously switches on the host's own antioxidant program (the Nrf2 pathway) while damping NF-κB-driven inflammation. Yet discovering such agents is crippled by a measurement problem. The cheap, scalable readout of host response is RNA sequencing, which measures gene transcripts — but the two master switches of the dual-action response, Nrf2 and NF-κB, are controlled not by how much their mRNA is made but by chemical modification, protein degradation, and nuclear relocation of the proteins themselves. The HUPO blueprint of the human proteome measured that mRNA levels track actual protein abundance only 40% to 60% of the time, and the discordance is worst precisely for these post-translationally regulated regulators.

This proposal builds the missing bridge. I will construct a paired single-cell benchmark in which host immune cells under live drug-resistant infection are measured both at the transcript level and at the protein level for the intracellular redox-inflammation machinery, then train a cross-modal translation model — a generative network that predicts the protein panel from the transcriptome — constrained by Gene Ontology pathway structure, and finally use that model as a cheap transcriptome-only screening readout that scores candidate antimicrobial peptides for dual action at the protein level, where it actually matters.

Core Claim

The claim is that current RNA-to-protein translation models, which are trained on healthy donor blood cells and predict only cell-surface markers, fail exactly where dual-action therapy needs them: under infection perturbation, for intracellular signaling proteins. By building the first paired transcript–intracellular-protein dataset under drug-resistant bacterial challenge and by constraining the translation model with pathway-structure priors, we can (a) predict the functional Nrf2/NF-κB protein state from transcriptomes under a perturbation the model never saw, and (b) use these predictions to rank known dual-action agents above single-node interventions and to prioritize new peptide candidates for wet-lab validation. The methodological bet, supported by Gene Ontology–constrained clustering results, is that pathway-structured connectivity stabilizes the generative model when the data distribution shifts violently away from the healthy atlas it was trained on — the regime where flow-matching and variational methods currently degrade.

Three gaps in the literature motivate this. The generative antimicrobial-peptide models LSSAMP and the ESCAPE benchmark score candidates only for direct microbial killing; neither contains any host-directed label, so a peptide that clears bacteria while igniting a cytokine storm looks identical to one that clears bacteria and calms the host. Conversely, the wet-lab dual-action studies — baicalein-loaded chitosan nanoparticles, SAMP-12aa, and clavanin-MO — demonstrate the biology but evaluate it with a handful of targeted assays, with no genome-wide view and no way to screen candidates computationally. And the cross-modal translation models scChord, scTEL, and totalVI operate only on surface-protein panels from resting donors, leaving the intracellular proteins that carry the dual-action signal unmodeled. My proposal sits precisely in that empty intersection: it does not extend any one of these lines so much as connect them through a new benchmark.

Research Program

The work has three components, in dependency order.

Benchmark construction. I will generate paired single-cell data from human monocyte-derived macrophages across a perturbation grid: unstimulated; stimulated with lipopolysaccharide; infected with live methicillin-resistant S. aureus and with a carbapenem-resistant Gram-negative strain; and each condition crossed with treatment arms — the baicalein–chitosan nanoparticle formulation of the MRSA lung-infection study as a positive dual-action control, clavanin-MO from the synthetic peptide study as a second positive control, a conventional antibiotic plus Vitamin C as a deliberately single-node negative control, and vehicle. Each sample gets full scRNA-seq plus an oligo-tagged antibody panel extended intracellularly to NF-κB p65, phosphorylated p65, IκBα, Nrf2, HO-1, and SOD2, alongside standard surface markers, with targeted mass-spectrometry validation of the signaling state following the signalome-mapping approach of the T-cell NF-κB study (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy). This is the ground truth where transcript–protein discordance should be maximal, because Nrf2 activation reflects Keap1-mediated protein stabilization rather than NFE2L2 transcription, and NF-κB activation reflects IκBα degradation rather than RELA mRNA.

Model. The translation model follows the two-stage design of scChord — a variational autoencoder with a zero-inflated negative-binomial decoder that absorbs technical noise, then a flow-matching map from RNA into the rectified protein manifold — but with two changes. First, the target space is the intracellular signaling panel rather than surface markers, including phosphorylated species. Second, the encoder's hidden units are constrained so that each connects only to genes annotated to a common Gene Ontology process, borrowing the GO-masked connectivity that GOAE showed prevents overfitting to technical artifacts (Adjusted Rand Index 0.917 versus 0.873 for the unconstrained ZIFA). The hypothesis is that this structural prior is what keeps predictions coherent when the perturbation pushes cells far outside the healthy atlas distribution.

Screening loop. For any candidate peptide from a generative pool in the style of LSSAMP, the assay requires only the cheap modality: treat the infected macrophage co-culture, run scRNA-seq, and let the model predict the protein panel. Candidates are ranked by a dual-action score

$$D \;=\; \frac{1}{|A|}\sum_{p \in A} z(\hat{y}_p) \;-\; \frac{1}{|I|}\sum_{p \in I} z(\hat{y}_p),$$

where $\hat{y}_p$ is the predicted abundance of protein $p$, $z$ standardizes within condition, $A$ is the antioxidant set (Nrf2, HO-1, SOD2), and $I$ is the inflammatory set (phospho-p65, nuclear p65). Bactericidal potency is measured separately by standard MIC assays, and only candidates passing both filters go to targeted proteomic validation.

Proposed Experiments

The first experiment tests translation accuracy under distribution shift. The model is trained on the LPS and S. aureus perturbation conditions and evaluated on a held-out perturbation (the Gram-negative challenge) it never saw, reporting Root Mean Square Error, per-protein Pearson correlation, and Correlation Matrix Distance for the cross-protein correlation structure — the metric scChord's own authors flag as degrading on large panels. The decisive comparison is against the same architecture without the GO connectivity mask.

The second experiment tests whether the score separates known agents. The model must rank the baicalein–chitosan nanoparticles (validated dual action: MIC of 15 μg/mL against S. aureus with simultaneous ROS scavenging, SOD restoration, and cytokine suppression in vivo) and clavanin-MO above the antibiotic-plus-Vitamin-C arm, using transcriptome-only input. If a transcript-naive reader — or the model itself — cannot separate these arms, the protein-grounding premise fails.

The third experiment is the screening payoff: rank roughly one to two hundred candidate peptides by predicted $D$, select the top and bottom deciles, and validate with the intracellular antibody panel and MIC assays, asking whether predicted dual action correlates with measured protein state (Spearman correlation) and whether predicted-negative candidates show inflammatory protein signatures despite adequate MIC.

Baselines

The naive reference point is the standard practice in the dual-action literature itself: use transcript abundance of marker genes (NFE2L2 for Nrf2, NFKBIA and RELA for NF-κB, HMOX1 for HO-1) as a proxy for pathway activation, which is exactly what targeted RT-PCR in the baicalein nanoparticle study and ELISA-style readouts in SAMP-12aa effectively do. This baseline costs nothing to run and will be evaluated on the identical paired dataset; beating it means showing that the transcript proxy misclassifies treatment arms that the protein-level predictions get right, with the HUPO 40–60% correlation ceiling as the expected explanation.

The strongest competitor is scChord, which reports the lowest RMSE and highest Pearson correlation across the PBMC benchmark GSE164378 and the CBMC dataset GSE100866, outperforming totalVI, sciPENN, and scTranslator, and which trains on 100,000 cells in approximately 27 minutes with 229 MB of GPU memory — though no exact numerical RMSE value for scChord was reported on Lacuna, so the comparison must be run head-to-head rather than quoted. For fairness I will retrain scChord from scratch on the new intracellular-panel dataset with the same compute and the same hyperparameter search budget as my model, and additionally report scTEL, which achieved 86.6% cell-type classification accuracy on the PBMC benchmark versus sciPENN at 77.0% and Seurat at 76.1%, and totalVI, which processed 33,000 cells in under an hour on a standard GPU. Beating these baselines means lower RMSE and higher per-protein correlation on intracellular targets under the held-out perturbation, and materially better Correlation Matrix Distance — not merely matching them on in-distribution conditions, where I expect them to be strong.

The key ablation isolates the proposal's central ingredient: the identical two-stage translation model with the Gene Ontology connectivity mask removed, leaving a fully unconstrained encoder trained on the same data with the same budget. If the unconstrained version matches the constrained one under the held-out perturbation, the pathway prior is unnecessary and the contribution collapses to the benchmark; I would report that honestly.

Why not just…?

Why not just read out inflammation from RNA markers, since RNA-seq is so cheap? This is the most serious alternative, and the field's default. The evidence against it is direct: the HUPO proteome blueprint measured that mRNA explains only 40–60% of protein variance, and the specific proteins governing dual action are the worst case, because Nrf2 and NF-κB are switched on by protein stabilization and phosphorylation that transcriptomics cannot see in principle. Still, this cannot be settled by argument, so the first experiment is designed to settle it empirically: if the transcript-marker proxy ranks the treatment arms as accurately as the protein-grounded model across all perturbations, the discordance concern is overstated for this setting and the proposal's practical value shrinks. I consider that a real possibility for the Nrf2 arm specifically, since HMOX1 transcription is a decent Nrf2 surrogate; the phospho-p65 arm is where I expect the transcript proxy to break.

Why not just retrain scChord on infection data and skip the pathway prior? That is exactly the strongest baseline described above, and it may well work: scChord's manifold-rectification design exists precisely to handle noisy, zero-inflated single-cell data. The honest position is that the GO constraint is a hypothesis, not a certainty, motivated by GOAE's overfitting protection (ARI 0.917 versus 0.873 unconstrained) and by scChord's own reported degradation in cross-protein correlation as panels scale. The held-out-perturbation experiment with the mask-off ablation is the decisive test; if fine-tuned scChord matches the constrained model, the benchmark itself remains the contribution and the model contribution is negative-result clarity.

Why not just co-administer a potent antibiotic with Vitamin C? The dual-action literature itself explains the failure: dietary antioxidants are consumable, stoichiometric scavengers that are exhausted on oxidation, whereas activation of the Nrf2/HO-1 axis induces self-regenerating enzymatic antioxidants that outperform single-action antioxidants like Vitamin C in sustained tissue protection, and scavenging reactive oxygen species without simultaneously suppressing NF-κB leaves cytokine transcription to replenish the oxidant pool through newly recruited immune cells. This arm is included in the benchmark as an active negative control rather than dismissed, so the experiment — not the argument — will show whether its transcript and protein signatures diverge from the dual-action controls.

Why not just add an "anti-inflammatory" label to the ESCAPE benchmark and train a bigger peptide classifier? The ESCAPE benchmark shows multi-attribute prediction works at the peptide level (its dual sequence-plus-structure transformer gained 2.56% relative mean Average Precision and 1.90% F1 over the next best method, and 35.7% on the rare Antiparasitic class), so this alternative is attractive. It fails for a data reason rather than an architecture reason: there are essentially no labeled examples of peptides annotated with host protein-level Nrf2/NF-κB outcomes, because no such measurements existed at scale — which is what the benchmark component of this proposal creates. A classifier trained on the handful of peptides with measured dual action would have a training set too small to trust; once this benchmark exists, extending ESCAPE with host-response labels becomes the natural follow-up.

Decisive Checks

The proposal stands or falls on three outcomes. The model must beat fine-tuned scChord on intracellular proteins under the unseen perturbation, with the GO ablation showing the gap is attributable to the pathway prior. The dual-action score must rank the two validated dual-action agents above antibiotic-plus-Vitamin-C from transcriptome input alone. And predicted protein state for screened peptides must correlate with directly measured intracellular protein levels in the validation set. If the first fails but the second and third succeed, the benchmark and the screening application still stand, and the modeling contribution reduces to a careful negative result about pathway priors — an outcome worth publishing either way.