Twist Mismatch as the Origin of One-Way Amyloid Cross-Seeding

When two different amyloid-forming proteins meet, they interact with a strange asymmetry: pre-formed fibrils of amyloid-β 42 potently speed up the aggregation of α-synuclein, but α-synuclein fibrils do nothing to amyloid-β 42 monomers Aβ42 promotes the aggregation of α‐synuclein splice isoforms via heterogeneous nucleation. Nobody can currently say why the catalysis only works one way. The field has exact kinetic numbers — half-times dropping from over 240 hours to 167 hours, seeding thresholds at 0.1 µM — but the physical origin of this directionality is a blank. Meanwhile, in an apparently unrelated corner of physics, condensed-matter researchers solved an almost identical puzzle: at oxide interfaces, whether two layers couple or decouple depends not on how much strain is present but on the symmetry pattern of the structural distortions — antiphase octahedral tilts let interfacial coupling vectors accumulate, while in-phase tilts make them cancel Tilt engineering of exchange coupling at G-type SrMnO3/(La,Sr)MnO3 interfaces.

This proposal imports that idea. An amyloid fibril is, structurally, a small crystal: a repeating cross-β lattice with ~4.7 Å layer spacing and a helical twist. When a monomer of a second species tries to grow on that crystal, it faces an epitaxial mismatch problem — the seed's twist, layer registry, and side-chain packing either accommodate the incoming chain's preferred lattice or geometrically frustrate it. I propose to build the first coarse-grained simulation of amyloid cross-seeding on a deformable lattice, in which local bond angles and spacings can tilt and relax under interfacial strain, exactly as octahedra do at oxide heterointerfaces. The central, testable claim: one-way cross-seeding is a symmetry-selection effect. Aβ42's surface geometry is commensurate with α-synuclein's preferred cross-β registry, so coupling accumulates; the right-handed Lewy fold of α-synuclein fibrils Cryo-EM structures of α-synuclein filaments from Parkinson's disease and dementia with Lewy bodies imposes a phase mismatch that cancels it. A reader can carry the whole idea in two sentences: treat the seed as a strained biological crystal, and ask whether its distortion pattern constructively or destructively interferes with the guest protein's own lattice preference.

Core Claim

The claim has two parts, one mechanistic and one predictive. Mechanistically, heterogeneous amyloid seeding is governed by a geometric commensurability condition between the seed's surface lattice (twist phase per layer, axial layer offset, side-chain contact register) and the recipient monomer's energetically preferred cross-β packing. This is the direct analog of the in-phase versus antiphase tilt selection rule that switches exchange bias on and off in manganite bilayers Tilt engineering of exchange coupling..., and it explains why bulk kinetic fitting, which absorbs cross-seeding into a single effective rate constant, can measure the asymmetry but never explain it.

Predictively, this mechanism can be encoded in a computable quantity — a cross-seeding misfit index computed from any two solved fibril structures — that forecasts the direction and sign of seeding for pairs that have never been mixed in a tube. That is something no existing tool does: the Amyloid Packing Difference metric compares fibril polymorphs within one protein family The amyloid packing difference: a pairwise comparison metric for amyloid structures, and kinetic platforms fit what has already been measured. Neither predicts an untested pair.

Research Program

Three components, in increasing order of empirical risk.

The first is the deformable-lattice model. The starting point is the cubic-lattice Monte Carlo model of A Simple Lattice Model That Captures Protein Folding, Aggregation and Amyloid Formation, which represents residues as lattice vertices with directional side chains and conditional backbone hydrogen bonds, and which already showed that hydrogen bonding alone cannot stabilize fibrils — sequence-specific packing is required. Its lattice, however, is strictly rigid: coordinates sit on integer vertices, bond angles are frozen at 90° and 180°, and nothing can tilt. I will relax exactly this. Each local bond acquires an elastic degree of freedom,

$$H = H_{\text{rigid}} + \sum_{\langle i,j \rangle} \tfrac{1}{2} k_\theta \, (\theta_{ij} - \theta^{0}_{ij})^{2},$$

where $\theta_{ij}$ is the local bond angle between neighboring residues, $\theta^{0}_{ij}$ is its ideal value (90° or 180°), and $k_\theta$ is a single elastic stiffness. Small $k_\theta$ lets the protein "lattice" distort; infinite $k_\theta$ recovers the published rigid model exactly, which gives us a clean ablation. Sampling stays Monte Carlo with parallel tempering, and the contact matrix and hydrogen-bond rules are inherited unchanged, so the only new physics is geometric strain. Heterogeneous seeding is then implemented as it is in the experiments: a pre-equilibrated seed fibril of species A is frozen in place (or held at harmonic restraint) and monomers of species B are introduced. The quantity of interest is how the free-energy barrier for B's primary nucleation changes as a function of the seed's surface geometry — the analog of the tilt-energy-barrier maps computed by DFT for oxide interfaces Tuning Perpendicular Magnetic Anisotropy by Oxygen Octahedral Rotations in (La1-xSrxMnO3)/(SrIrO3) Superlattices.

The second component is the misfit index. From the deformable-lattice free-energy surfaces I will extract which geometric variables actually control the barrier — my working hypothesis is that the phase of the helical twist per ~4.7 Å layer and the axial layer offset between seed and adsorbate dominate, mirroring how the DM-vector argument in Tilt engineering... reduces a complex interface to a phase-accumulation rule. Those variables can then be measured directly on real structures: cryo-EM coordinates of α-synuclein Lewy-fold filaments at 2.2 Å Cryo-EM structures of α-synuclein filaments... and brain Aβ42 Type I (2.5 Å, G9–A42) and Type II (2.8 Å, V12–A42) filaments Cryo-EM Structures of Amyloid-β 42 Filaments from Human Brain. Notably, the APD tool deliberately discards axial layer offsets because of reconstruction artifacts The amyloid packing difference... — and the axial offset is precisely the variable my hypothesis says matters most, so the two metrics are complementary by construction, not competing. The index will also be validated on the prion, tau, TDP-43, and TAF15 structures that the APD study already catalogued, giving a broad structural testbed at no additional data cost.

The third component is a falsifiable set of blind predictions on systems the model was never tuned on: the direction of amylin–α-synuclein co-aggregation, which is documented to occur Lai et al., Theranostics, 2025 but whose seeding directionality is unmeasured; and the mechanistic split between Aβ42, which acts as an external solid surface that triggers α-synuclein condensation, versus Aβ40, which instead gets recruited into pre-formed α-synuclein droplets and accelerates nucleation there by over two orders of magnitude at 20 µM Amyloid-β modulates the phase separation and aggregation of α-synuclein, Recruitment of Aβ into α-Synuclein Condensates.... A geometric mismatch story makes a sharp distinction between these: an externally commensurate seed surface versus a condensate-soluble partner whose partitioning, not its surface lattice, is the rate-limiting step.

Proposed Experiments

The experiments are simulation-first, anchored to reported numbers, with a defined wet-lab confirmation path.

Calibration comes from single-species behavior only. The single elastic constant $k_\theta$ and the species-specific contact parameters are fit once to homogeneous aggregation (α-synuclein alone, Aβ42 alone) and then frozen. Two sanity anchors from prior coarse-grained work must be reproduced before anything else: the four-chain critical nucleus and the free-energy minimum at $n \approx 10$–$15$ peptides found in the flexible-tube simulations Self-Templated Nucleation in Peptide and Protein Aggregation, and the requirement that directional side-chain packing, not hydrogen bonding alone, stabilize the fibril A Simple Lattice Model....

The first real test is the asymmetry itself. The model must reproduce, with frozen parameters, that Aβ42 surfaces catalyze α-synuclein splice isoforms while α-synuclein surfaces leave Aβ42 untouched Aβ42 promotes the aggregation.... Quantitatively, the simulated barrier reduction should track the measured ordering: aggregation half-times of αSyn-112 falling from over 240 hours to about 215 hours with Aβ42 monomers, αSyn-98 falling to about 167 hours, pre-formed Aβ42 fibrillar seeds cutting half-times by roughly 40% relative to monomeric Aβ42, and catalysis persisting down to 0.1 µM Aβ42 FEBS, 2025. I do not expect a lattice model to hit absolute hours; the target is the relative pattern, within roughly a factor of 1.5 on half-time ratios.

The second test is pathway inversion. Inside liquid condensates the aggregation hierarchy of α-synuclein splice isoforms flips: full-length αSyn-140 condenses most easily (critical concentration ~507 µM) and the shortest isoform αSyn-98 needs ~767 µM and aggregates about four times more slowly inside droplets, even though in dilute solution the short isoforms aggregate faster Aggregation of α-synuclein splice isoforms through a phase separation pathway, FEBS, 2025. In the deformable-lattice picture, confinement raises the energy cost of the lattice distortions a chain must tolerate in a dense cluster, and shorter chains with higher charge density have less packing redundancy to absorb that cost — a testable mechanism where currently there is only a documented contradiction. The model should reproduce the flip as a function of crowding volume fraction, at fixed parameters.

The third test is the blind predictions: seeding direction for amylin–α-synuclein, and the external-catalysis versus internal-recruitment split between Aβ42 and Aβ40, checked against microfluidic droplet assays and ThT kinetics fitted in AmyloFit, the standard platform that extracts microscopic rate constants from fluorescence curves. As a stretch target, the model should rationalize the extreme regime — condensate-mediated nucleation accelerated by up to five orders of magnitude Morris et al. and instantaneous co-aggregation at equimolar 100 µM mixtures Amyloid-β modulates the phase separation... — as the misfit penalty being drowned by concentration, which makes a specific prediction: seeding selectivity should weaken as monomer concentration rises.

Baselines

The do-nothing reference is homogeneous bulk kinetics: AmyloFit fitted to each protein aggregating alone, with no cross-seeding term. Its reference numbers are the unperturbed half-times, over 240 hours for αSyn-112 and αSyn-98 in the absence of Aβ42 FEBS, 2025. Any model that predicts the observed cross-seeding reductions beats this trivially, so it serves as the floor and as the null hypothesis that the proteins ignore each other.

The strongest existing competitor is the phenomenological heterogeneous-seeding fit in AmyloFit, as used in the FEBS 2025 study, which reproduces the measured curves essentially exactly — for example, an aggregation half-time of 167 hours for αSyn-98 with Aβ42 present FEBS, 2025. It will be run on identical data with its full per-pair fitting freedom, same kinetic traces, no compute constraint (it is cheap). But it fits each pair after the fact; it has no parameters that transfer to an untested pair and no structural content. Beating it therefore cannot mean fitting the known curves better — it means predicting, with parameters frozen on single-species data, the sign and rough magnitude of seeding for held-out pairs, where the AmyloFit baseline has nothing to predict with. If my frozen-parameter model cannot at least match its fit quality on known pairs and predict direction on held-out pairs, the proposal fails.

The mechanistic baseline is the original rigid lattice model A Simple Lattice Model... — more precisely, my own model with $k_\theta \to \infty$, which recovers it exactly. No reported result found on Lacuna for cross-seeding or multi-species aggregation in that model; it was designed for single-sequence folding-versus-aggregation competition. It will be given the same compute, sampling, and tuning budget, including an equal number of additional free parameters (for instance, extra contact-matrix scaling terms) so the comparison cannot be criticized as unfair on parameter count. Beating it means the deformable version predicts seeding direction above chance while the rigid version does not — this is also the ablation that isolates the proposal's single key ingredient, elastic lattice distortion.

Why not just…?

Why not just fit everything with AmyloFit kinetics? Because the question is not what the rates are but why they are asymmetric and whether the asymmetry transfers. AmyloFit already quantifies that Aβ42 cuts αSyn-98's half-time to 167 hours while α-synuclein seeds do nothing to Aβ42 FEBS, 2025, and that is precisely the phenomenon left unexplained. A fitted rate constant for each pair is a description, not a mechanism; it cannot say in advance whether amylin will seed α-synuclein or the reverse. If, however, AmyloFit-style fitting on one held-out pair turned out to generalize trivially — say, seeding strength correlated with some simple quantity like monomer solubility across all documented pairs — that would undercut the need for a structural mechanism, and I will check exactly that correlation in the calibration phase.

Why not just run all-atom molecular dynamics? Because the timescales are separated by roughly ten orders of magnitude. The measured half-times are hundreds of hours FEBS, 2025, while all-atom simulations reach nanoseconds to microseconds, and the nucleation event involves dozens of chains condensing before a four-chain nucleus forms Self-Templated Nucleation.... Both coarse-grained studies in the file state this limitation explicitly. All-atom MD is the right tool for the last step — verifying, at atomic resolution, the predicted contact geometry of a single commensurate versus frustrated seed–monomer interface — and I will use it there, not for the barrier crossing itself.

Why not just use the existing rigid lattice model with better parameters? The rigid lattice cannot represent the phenomenon by construction. Protofilament twist, axial layer offsets, and interfacial mismatch between two species are all continuous geometric distortions, and the published model restricts residues to integer coordinates with 90°/180° angles A Simple Lattice Model.... The history of the solid-state side is the warning here: macroscopic strain in oxide superlattices predicted that perpendicular magnetic anisotropy should fall as $c/a$ decreased, but measurements showed it rising to a record $K_{\text{eff}} \approx 4 \times 10^{6}$ erg/cm³, tracking internal bond-angle rotations instead Tuning Perpendicular Magnetic Anisotropy.... Getting the right answer required modeling the internal distortion, not just the overall strain — the same lesson applies to protein interfaces. Still, this is a real threat, which is why the rigid-versus-deformable comparison at equal budget is the central ablation rather than an afterthought.

Why not just compare fibril structures with APD or RMSD and call it a day? RMSD fails outright on amyloid filaments because of variable twist and partial disorder, which is why APD was built on contact lists instead The amyloid packing difference.... But APD is a static, within-family comparison tool: it separates disease polymorphs (APD above roughly 20–40% divergence) after the structures are solved, it explicitly discards axial layer offsets, and it has never been used to predict cross-species seeding kinetics. My proposal uses APD's contact machinery where it is strong (measuring seed surface registers) but reinstates the axial-offset information it drops, grounds the index in a simulation of the barrier rather than a structural distance, and validates it against kinetic outcomes, not structural taxonomy.

Decisive Checks

Three outcomes would settle the proposal. First, the ablation: with $k_\theta$ frozen at infinity, seeding-direction prediction should collapse to chance, while the deformable model predicts the documented Aβ42→α-synuclein catalysis and the failure of the reverse FEBS, 2025. Second, transfer: parameters fit only on single-species data must predict the direction of amylin–α-synuclein seeding and the Aβ42-versus-Aβ40 mechanism split before the experiments are run, with the prediction registered in advance and tested by the standard microfluidic and ThT assays. Third, distortion depth: mirroring how substrate-imposed tilt gradients relax with film thickness in oxide heterostructures Tilt engineering..., the model predicts that seed-imposed distortion in the adsorbed layer decays with distance from the interface, so catalytic efficiency should fall with each additional adsorbed layer — a gradient measurable by cryo-EM of seeded products. If the deformable lattice adds nothing over its rigid ablation, or if its frozen-parameter predictions fail on held-out pairs, the symmetry-selection hypothesis for one-way cross-seeding is wrong, and the field will at least know that the answer lies elsewhere — most plausibly in solution chemistry or condensate partitioning rather than geometry.

References

  1. Aβ42 promotes the aggregation of α‐synuclein splice isoforms via heterogeneous nucleation
  2. Tilt engineering of exchange coupling at G-type SrMnO3/(La,Sr)MnO3 interfaces
  3. Cryo-EM structures of α-synuclein filaments from Parkinson’s disease and dementia with Lewy bodies
  4. The amyloid packing difference: a pairwise comparison metric for amyloid structures
  5. A Simple Lattice Model That Captures Protein Folding, Aggregation and Amyloid Formation
  6. Tuning Perpendicular Magnetic Anisotropy by Oxygen Octahedral Rotations in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:msub><mml:mrow><mml:mi>La</mml:mi></mml:mrow><mml:mrow><mml:mn>1</mml:mn><mml:mo>−</mml:mo><mml:mi>x</mml:mi></mml:mrow></mml:msub><mml:mrow><mml:msub><mml:mrow><mml:mi>Sr</mml:mi></mml:mrow><mml:mrow><mml:mi>x</mml:mi></mml:mrow></mml:msub></mml:mrow><mml:mrow><mml:msub><mml:mrow><mml:mi>MnO</mml:mi></mml:mrow><mml:mrow><mml:mn>3</mml:mn></mml:mrow></mml:msub></mml:mrow><mml:mo stretchy="false">)</mml:mo><mml:mo>/</mml:mo><mml:mo stretchy="false">(</mml:mo><mml:msub><mml:mrow><mml:mi>SrIrO</mml:mi></mml:mrow><mml:mrow><mml:mn>3</mml:mn></mml:mrow></mml:msub><mml:mo stretchy="false">)</mml:mo></mml:mrow></mml:math> Superlattices
  7. Cryo-EM Structures of Amyloid-β 42 Filaments from Human Brain
  8. Unraveling α-synuclein and amylin co-aggregation: pathological insights and biomarker development for Parkinson's disease
  9. Amyloid-β modulates the phase separation and aggregation of α-synuclein
  10. Recruitment of Aβ into α-Synuclein Condensates Catalyzes Primary Nucleation of α-Synuclein Aggregation
  11. Self-Templated Nucleation in Peptide and Protein Aggregation
  12. Aggregation of α-synuclein splice isoforms through a phase separation pathway
  13. Aggregation of the amyloid-β peptide (Aβ40) within condensates generated through liquid–liquid phase separation