Gel-Locking Peptides: Steering the Liquid-to-Solid Transition in Bacterial Biofilms and Alzheimer's Amyloid

Many proteins spontaneously crowd themselves into dense, liquid-like droplets — membrane-less compartments formed by weak, many-handed interactions between disordered protein segments. Over hours to days, these droplets can harden: their internal bonds reorganize into rigid cross-β sheets, and the once-fluid condensate becomes an irreversible amyloid fibril. This liquid-to-solid transition is the physical event behind both Alzheimer's plaques and the curli amyloid scaffold that holds bacterial biofilms together. Two facts make the shared physics actionable rather than merely poetic. First, the dense interior of a liquid condensate accelerates the birth of the first solid seed by up to five orders of magnitude compared to dilute solution, as shown for the Alzheimer's peptide Aβ40 by Morris et al.. Second, the dangerous state is usually the intermediate one, not the endpoint: in human cells, dynamic liquid-like condensates are the toxic species, while locking them into an inert gel is protective, as Resnick et al. demonstrated for the chaperone DNAJB6, and as Horváth et al. confirmed across more than 20,000 TDP-43 variants. In biofilms the same logic runs in reverse: the solid matrix protects the pathogen by up to 1,000-fold against antibiotics, while fluid, dispersed cells are the vulnerable ones (CRAMP dispersal study).

We propose to treat the liquid-to-solid transition itself as the design target for antimicrobial peptides. The plan is to engineer peptides that "gel-lock" a biofilm matrix — freezing it into a permeable, loose gel that admits antibiotics, without dissolving it into free-swimming cells — while simultaneously proving, as a hard safety constraint, that these peptides do not accelerate the same transition in the Alzheimer's peptide Aβ. A secondary sub-aim tests the converse risk: whether bacterial curli amyloids seed Aβ phase collapse, which would make biofilm control an infection-to-neurodegeneration question rather than three disconnected fields.

Core Claim

The phase state of an amyloid matrix — liquid, gel, or fibrillar solid — can be deliberately redirected by short peptides, and redirecting it is a better therapeutic strategy than either dispersing biofilms or clearing plaques. Concretely, a peptide class exists (or can be found by guided search) that (i) raises the antibiotic permeability and kill rate inside curli-dependent biofilms without triggering the planktonic-cell release that makes dispersal agents clinically hazardous, and (ii) leaves Aβ40/Aβ42 condensation and nucleation kinetics unchanged or slower, so it cannot worsen amyloid pathology in an infected, aging host. No study in the literature has looked for peptides with either property; every existing antimicrobial peptide generator optimizes exclusively for planktonic killing.

One equation captures the design problem. For a candidate peptide $p$, define its phase-modulation index on a target protein as

$$\mathrm{PMI}(p) = \log_{10}\!\left(\frac{k_{\mathrm{nuc}}^{p}}{k_{\mathrm{nuc}}^{0}}\right),$$

where $k_{\mathrm{nuc}}^{p}$ and $k_{\mathrm{nuc}}^{0}$ are the primary nucleation rates fitted by AmyloFit kinetics (Morris et al.) in condensate assays with and without the peptide. The objective is to maximize biofilm antibiotic potentiation subject to $\mathrm{PMI}_{A\beta}(p) \le 0$ and hemolysis below a fixed cytotoxicity threshold. The log-ratio form comes directly from the AmyloFit rate model; the five-order-of-magnitude dynamic range reported for condensate nucleation means PMI is measurable across a wide, informative span.

Mechanistic Background and the Gap

Four lines of work motivate the claim but none completes it. Morris et al. built a microfluidic water-in-oil droplet platform that isolates Aβ40 from air-water interfaces and proved that condensate interiors accelerate primary nucleation by up to $10^5$; they tested only recombinant Aβ40 with a synthetic aminosterol, no bacterial components, and no Aβ42. Resnick et al. showed that DNAJB6 co-partitions into FUS droplets and arrests them in a non-toxic gel with elevated parallel β-sheets and suppressed antiparallel cross-β fibrils, validated by AFM-IR nanoscale spectroscopy and in a FUS-ALS mouse model; this is the mechanistic template for gel-locking, but it is a large human chaperone acting on eukaryotic RNA-binding proteins. The Aβ–α-synuclein co-condensation study (PNAS 2025) established that one amyloid protein can partition into another's liquid condensates or seed them from solid surfaces — the exact mechanism our cross-kingdom sub-aim transposes to CsgA–Aβ pairs, which that paper never touched. On the bacterial side, the CsgA curli subunit has a resolved β-helical monomer fold from coevolutionary contact prediction (Tian et al., JACS 2014), but only as a static structure — nobody has tested whether CsgA assembly passes through a liquid condensate intermediate at all, and the CRAMP dispersal work (Veterinary 2022) never examined the matrix's phase state even while reporting that CRAMP at 62.5 µg/mL cuts one-day biofilm biomass by 56.6% and cell counts by 76.41%. Finally, the generative side is mature: OmegAMP achieves a 96% experimental hit rate (24 of 25 synthesized peptides active) and 92% success at 8 µg/mL against multidrug-resistant clinical isolates, and ApexOracle reaches an ensemble $R^2 = 0.5793$ on strain-wise MIC regression — but both optimize exclusively for killing free-floating cells, with no representation of phase behavior, saturation concentration, or matrix state. The gap is precisely the missing objective.

Research Program

The program has three linked work packages. The first establishes mechanism: we will express and purify CsgA (and the nucleator CsgB) and run them through the microfluidic droplet entrapment platform of Morris et al., measuring whether they phase-separate, whether droplets coalesce and Ostwald-ripen like Aβ40 droplets do, and how the polyanionic matrix components alginate and extracellular DNA shift the saturation concentration. Nanoscale infrared spectroscopy (Ruggeri et al.; Miller et al.) will track the secondary-structure and stiffness trajectory of individual ageing CsgA condensates, exactly as was done for FUS. In parallel, the same platform tests cross-kingdom seeding: CsgA or CsgB added to Aβ40 and Aβ42, measuring aggregation half-times and co-condensation, a direct transposition of the heterologous-seeding logic from the Aβ–α-synuclein study (PNAS 2025). If bacterial amyloid seeds host Aβ the way insoluble Aβ42 seeds α-synuclein, chronic infection gains a concrete biophysical route into Alzheimer's pathology.

The second package builds the first phase-modulation dataset and model. We will screen a few hundred peptides drawn from the DBAASP and ESCAPE collections (the latter holds 21,409 verified antimicrobial peptides among 82,359 total sequences; ESCAPE benchmark) in miniaturized droplet assays against both CsgA and Aβ40, producing PMI labels. These labels then train a generative pipeline in the style of OmegAMP's diffusion model with physicochemical embeddings, but with a dual objective: antimicrobial activity retained as a constraint, biofilm permeabilization maximized, and $\mathrm{PMI}_{A\beta} \le 0$ enforced. The FuzDrop/FuzPred droplet-propensity and binding-mode-multiplicity features that Horváth et al. validated against yeast cytotoxicity (Pearson $r = 0.975$ on single TDP-43 mutants) will be transferred as input features to peptide space. We are explicit that the labeled set will be small (hundreds, not tens of thousands); the design uses active learning, with each synthesis round chosen to shrink uncertainty on the PMI surface, and its success is judged on held-out peptide predictions, not on dataset size.

The third package is empirical validation. Top candidates go into curli-producing E. coli and P. aeruginosa PAO1 biofilms, where we measure matrix phase state by confocal rheology-proxy imaging, antibiotic potentiation (kill rate of a standard antibiotic at sub-minimum inhibitory concentration inside the treated biofilm), planktonic-cell release into the supernatant, and biomass, using the exact assay geometry of the CRAMP study so numbers are comparable. Safety is checked on three axes: hemolysis, mammalian cell viability, and — the proposal's defining filter — the Aβ40/Aβ42 droplet assay must show no nucleation acceleration. A yeast cytotoxicity readout modeled on the multiplex proteotoxicity platform of Resnick et al. serves as a medium-throughput filter for any peptide that also partitions into eukaryotic condensates.

Baselines

The naive reference point is the untreated biofilm, where matrix solidification confers up to 1,000-fold antibiotic tolerance relative to planktonic cells (CRAMP study), and, on the Alzheimer's side, Aβ40 aggregation in bulk solution without condensates, where primary nucleation runs up to five orders of magnitude slower than inside droplets (Morris et al.). These anchor the two ends of every comparison. The strongest existing competitor for biofilm control is CRAMP itself, the mouse cathelicidin that at 62.5 µg/mL reduces one-day-old P. aeruginosa biofilm biomass by 56.6%, biofilm cell count by 76.41%, and total volume by 49.39% by suppressing intracellular c-di-GMP and reactivating motility (CRAMP dispersal study). We will run CRAMP in our assays side by side with our peptides, same strains, same biofilm age, same antibiotic co-treatment, and the same number of peptide candidates screened, so the comparison isolates mechanism rather than effort. Beating CRAMP means matching or exceeding its biomass and cell-count reductions while releasing strictly fewer planktonic cells into suspension and achieving a higher antibiotic kill rate inside the residual matrix — CRAMP's own sepsis hazard is the axis on which it should lose. The strongest competitor for peptide discovery is OmegAMP, whose diffusion-based generator with XGBoost filtering achieved a 96% experimental validation hit rate (24 of 25 peptides) and 92% success at 8 µg/mL against multidrug-resistant isolates (OmegAMP). We will regenerate candidates with OmegAMP's pipeline under the identical synthesis budget and screen them in the same biofilm and Aβ assays; our approach wins only if phase-objective peptides beat MIC-only peptides on biofilm endpoints at equal planktonic potency. Finally, the key ablation isolates our ingredient: the same generator, same data, same tuning budget, with the PMI objective removed. If the phase objective contributes nothing beyond ordinary AMP discovery on biofilm endpoints, the central claim is wrong and we will say so.

Why not just…?

Why not just generate more potent antimicrobial peptides with existing models? Because the existing models are solved for the wrong game. OmegAMP's 96% hit rate and ApexOracle's $R^2 = 0.5793$ strain-wise MIC regression (ApexOracle) measure killing of free-floating cells, while the clinical problem is biofilms whose matrix grants up to 1,000-fold tolerance — a regime where a tenfold more potent peptide on planktonic cells may change little. That said, this alternative is a genuine threat, which is why the head-to-head synthesis-budget-matched comparison against OmegAMP's own output is built into the experimental plan rather than argued around.

Why not just disperse biofilms with agents like CRAMP? The CRAMP study itself documents the hazard: dispersal works by reactivating flagellar motility and rhamnolipid surfactant production, releasing planktonic cells that risk septicemia unless paired with immediate high-dose bactericidal clearance (CRAMP study). Gel-locking inverts the strategy — immobilize the cells, open the matrix — and our decisive metric is precisely the planktonic-release fraction alongside the kill rate. If gel-locked biofilms turn out not to admit antibiotics better than untreated ones, the mechanism fails and the dispersal camp wins; that check is decisive, not rhetorical.

Why not treat Alzheimer's with anti-amyloid antibodies or repurposed drugs instead of this? The trial record assembled in the file argues the single-target amyloid strategy underdelivers: a phase II repurposing study frames AD as multifactorial (Levey et al.), patient-biomarker trajectories show vascular dysregulation precedes detectable amyloid and tau (Iturria-Medina et al.), and cholinergic deficits are downstream of inflammasome activation (Hu). Our proposal is not another plaque-clearing drug: Alzheimer's biology supplies the mechanistic lens (the toxic species is the dynamic condensate, per Horváth et al., and gel-locking is protective, per Resnick et al.) and a hard safety filter — any peptide that accelerates Aβ nucleation, which Morris et al. showed can be sped up 100,000-fold in condensates, is discarded. The cross-kingdom seeding sub-aim additionally tests whether infection itself feeds amyloid pathology, which no antibody strategy addresses.

Why not just characterize CsgA phase behavior and skip the machine learning? The simple biophysics experiment is real and necessary, and it is our first milestone, not a rival. But the peptide search space is combinatorially vast, and unguided screening of even the 17,988 peptides with measured activity in DBAASP (ApexOracle) in multi-day droplet assays does not scale; the active-learning loop is what makes a few hundred measurements sufficient. We are honest about the fallback: if the labeled set proves too small to train a useful model, the CsgA LLPS characterization and cross-kingdom seeding result remain publishable mechanistic findings on their own.

Decisive Checks

Three outcomes decide the proposal. First, CsgA must either pass through a liquid condensate intermediate with measurable AmyloFit kinetics or definitively not — either answer closes an open problem, but gel-locking as a strategy only makes sense if a fluid or soft-gel regime exists to lock. Second, at matched synthesis budget, phase-objective peptides must beat both CRAMP and OmegAMP-derived peptides on the combined endpoint of antibiotic kill rate inside the biofilm and planktonic release, with the PMI-ablation generator in between; if MIC-only peptides match us on biofilms, the added objective is decoration. Third, the safety filter must hold: winning peptides show $\mathrm{PMI}_{A\beta40} \le 0$ and $\mathrm{PMI}_{A\beta42} \le 0$ in the droplet assay and pass hemolysis screening, because a peptide that opens biofilms while seeding Alzheimer's amyloid would be a Pyrrhic victory. The cross-kingdom sub-aim is judged by whether CsgA or CsgB shifts Aβ42 aggregation half-times or droplet formation at physiologically arguable concentrations, using the equimolar and substoichiometric designs of the Aβ–α-synuclein study (PNAS 2025) as the assay template. Resistance risk is acknowledged up front: bacteria can mutate curli production to escape matrix lock, just as they mutate phage receptors (phage resistance study), and we will measure the fitness cost of curli-loss escape mutants as part of the third package — an escape that trades matrix protection for antibiotic vulnerability may be an acceptable failure mode.

References

  1. Aggregation of the amyloid-β peptide (Aβ40) within condensates generated through liquid–liquid phase separation
  2. Multiplex neurodegeneration proteotoxicity platform reveals DNAJB6 promotes non-toxic FUS condensate gelation and inhibits neurotoxicity
  3. Sequence-based Prediction of the Cellular Toxicity Associated with Amyloid Aggregation within Protein Condensates
  4. Pseudomonas aeruginosa biofilm dispersion by the mouse antimicrobial peptide CRAMP
  5. Amyloid-β modulates the phase separation and aggregation of α-synuclein
  6. Structure of a Functional Amyloid Protein Subunit Computed Using Sequence Variation
  7. OmegAMP: Targeted AMP Discovery via Biologically Informed Generation
  8. Predicting and generating antibiotics against future pathogens with ApexOracle
  9. Unraveling the Physicochemical Determinants of Protein Liquid-liquid Phase Separation by Nanoscale Infrared Vibrational Spectroscopy
  10. Nanoscale profiling of evolving intermolecular interactions in ageing FUS condensates
  11. A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification
  12. A phase II study repurposing atomoxetine for neuroprotection in mild cognitive impairment
  13. Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis
  14. Ethanol extract of Evodia lepta Merr. ameliorates cognitive impairment through inhibiting NLRP3 inflammasome in scopolamine-treated mice
  15. Investigation of the therapeutic efficacy and resistance mechanisms of lytic phages targeting ST218 KL57 CR-hvKP