Introduction

To achieve our project goal of developing a comprehensive approach to treating the underlying pathology of Alzheimer’s disease, we had to investigate first what defines Alzheimer’s at the molecular level, as well as to become familiar with the approaches that have been tried so far to address the disease. Apart from initial background research, an important part of our research was studying past iGEM teams’ work on AD or related diseases (such as Parkinoson’s or dementia in general).We used simulations produced in COPASI, NUPAK, MARTINI and other tools to model the behaviour of our biological constructs. The results of the dry lab iterations informed and guided the experiments we carried out in the wet lab.

Stage one: CRISPase

Our first approach was to engineer a single enzyme to directly cleave misfolded proteins linked to Alzheimer’s Disease, amyloid beta and secondarily tau. We opted for:

- a cpCas9 scaffold (from SpCas9) with catalytic cores: MMP-2 catalytic domain (P08253) and TEV protease (P04585), to allow for selectivity and increased activity,
- mutations, so that the DNA-cutting activity of Cas9 was completely eliminated, and
- a lac promoter kill switch.

(1) fusion proteins misfold often, leading to steric occlusion of active sites;
(2) delivery via vesicles or mRNA is challenging due to large size (>4kb); and
(3) there is no previous research in which cpCas9 hosts non-Cas protein domains, thus design is speculative.

First testing cycle demonstrated the single enzyme approach was not feasible. As seen in our AlphaFold attempt, the artificial fusion of Cas9 with a TEV or MMP catalytic core has no natural homologues, and the resulting confidence scores (pLDDT - XX) resulted in AlphaFold not being able to produce a structural prediction. Therefore, this limited our in silico testing capabilities drastically, removing our ability to run virtual docking with amyloid beta.

• CRISPase v2: redesigned with cpCas9-protease hybrids (TEV, MMP) with signal peptides for secretion as well as the addition of a localisation domain for tau/amyloid beta;
• TauTuner: a LOV2-PP2A fusion as an intracellular actuator, with reversible light activation was introduced due to higher correlation between tau phosphorylation and cognitive decline in AD.
• PEGylated Vesicles - lipid vesicles decorated with targeting peptides for BBB transit (such as TfR-binding HAI peptide or NRP-1) as delivery vehicle
• RNA-based

Once tested, this design also failed to produce the predicted structure:

TauTuner. The structure of TauTuner showed minimal steric clashes. The COPASI simulation identified that 30s light pulses led to 80% PP2A active site availability, with half life of activity in the dark of ~5 min, showcasing reversible control.
Vesicles. MARTINI simulations showed a safe shear-induced aggregation, yet ligand density over 2 mol % increases aggregation probability.
Logic gates. NUPACK simulation of initial builds showed very strong secondary structures, with ΔG < −50 kcal·mol⁻¹, resulting in risk of self inhibition. Moreover, large numbers of repeats in design resulted in the constructs being unsynthesizable. Thus, the initial design of RNA-based logic gates was not possible within the limitations of our project.

In conclusion, CRISPase was beyond our in silico-modeling abilities.

Stage 2: Tau-specific approach and LYTAC deployment

Our parallel investigation into current drug developments and clinical trtials convinced the team to drop Amyloid-β focus. As a result, tau pathology became our central target. Whilst tau theraphies, too, failed in clinical trials, recent research showed a new part of tau pathology not explored before as a possible therapeutic target: extracellular tau seeds.

In order to eliminate tau seeds as a therapeutic and profilactic intervention, we decided to use the emergent technology of LYsosome TArgeting Chimeras (LYTAC). LYsosome TArgeting Chimeras are bifunctional chimeric proteins, consisting of two parts: an Internalising Receptor (IR) binder (usually EGF or ASPGR) and a Target binder (usually an antibody specific to the protein of interest). However, we had to address a number of problems before proceeding with it:

• Most LYTACs are obtained through the chemical conjugation of its parts, resulting in higher costs as well as less flexibility in delivery, due to their large size (~110 kDa), hence we had to readjust vesicle size;
• LYTACs were originally designed to target membrane-bound proteins, not extracellular proteins such as p-tau seeds;
• LYTACs are usually used as therapeutic molecules for cancer, therefore, tested IR binders and target binders are specific to that purpose.

As a result, we had to: (1) design a fully genetically encodable LYTAC, to allow for delivery as mRNA as an alternative to vesicles; (2) identify whether or not LYTACs can be used to efficiently target proteins of interest in the extracellular domain ; (3) find relevant Internalising Receptors (IR) specific to immune cells in the brain, to internalise our LYTAC, and sequentially find binding elements (peptides, proteins) specific to those intenalising receptors.

LYTAC docking was simulated in HADDOCK, with a score of -82, and predicted K_d of about 70 nM, showing promising affinity. Moreover, through a coarse grained simulation, the Albumin tagged variant of LYTAC had a longer half life but reduced diffusion across the BBB. Therefore, this variant was disconsidered for wet lab testing.
Vesicle - microglia interaction was also modeled through a boolean network simulation of NF-kB signaling, showing low risks of inflammatory off target effects.
An initial ODE for the split Gal4 logic gate system failed, due to COPASI not being able to perform a stochastic simulation with the current system parameters.

The LYTAC strategy showed strong promise, therefore we rpoceeded with the design of a trifunctional molecule, the TRI-LYTAC. This was inspired by research which showed that the use of TfR binding peptides could be used as both an internalising receptor binder and as a method of transcytosis across the blood-brain barrier (BBB). Moreover, due to TfR’s high rate of internalisation, it could lead to increased efficiency of internalisation. Yet, for our purposes, this approach could not be implemented as is, because (1) the TfR receptor does not internalise and route to the lysosomes reliably, and (2) the TfR receptor has low specificity, and internalises on every cell type, perhaps leading to increased toxicity inside neurones by increasing uptake of tau.

Therefore, our TRI-LYTAC design included the three following elemnts:
(1) a tau binding scFv;
(2) a TfR (HAI) binding peptide to allow for BBB transport and increased rate of internalisation; and
(3) a dimeric IGF-2R binder to route internalised cargo to the lysosomes.

Docking showed a score of -90, and a Kd_pred of about 20-50 nM. HAI-TfR docking received a score of -72, with a Kd_pred of about 200 nM, which is acceptable for our desired shuttle system. Moreover, domain orientation was tested through Alpha fold, revealing that the best fold was obtained through the AT8–HAI–IGF-2(dimer) layout. A RMSD of ~2.8 Å between modeled scFv and antibody template, showed stability. The (G4S)₃ linkers maintained flexibility without steric clashes.

Stage 3: Adding mRNA constructs for blood-brain barrier delivery

Until this stage, our main delivery strategy had been through PEGylated vesicles with TfR/NRP-1 ligand for BBB transcytosis. Yet, feedback from human practices interviews and early modeling revealed that manufacturing vesicles with consistent size and ligand density is costly, slow, and thus an ineffective delivery strategy. Many recommended mRNA/LNP as scalable, already clinically validated (through COVID-19 vaccines) as a potential alternative. Therefore, we added a parallel mRNA/LNP track, because an mRNA strategy would allow for lower cost, faster prototyping, and scalability. Another advantage was the direct secretion of TRI-LYTAC, which removed the need to purify protein, and instead, use cells as “factories” for our therapeutic.

The mRNA constructs did not show any ultra-stable hairpins at the ribosomes entry site. This resulted in an improvement of ~30% in predicted translation efficiency compared to unoptimized UTRs. Moreover, each mRNA construct was tested for compactness, to see whether they fit within mRNA synthesis limits and budget, with both TRI-LYTAC and TauTuner ORF fitting comfortably within limits.

Once the model was assembled, we used it to explore whether the split-GAL4 logic system shows AND behavior. Four input conditions were tested, corresponding to the gate’s truth table: OFF (no input), A-only (inflammation), B-only (oxidative stress), and dual input (both stressors present simultaneously).

In principle, stochastic simulations such as the Gillespie algorithm would capture random promoter switching and cell-to-cell variability. However, due to instability of stochastic solvers when dealing with binary promoter states and very low molecule counts, we were unable to run the model stochastically in COPASI at this stage. Instead, we employed the deterministic LSODA solver, which integrates the same reactions as continuous ordinary differential equations. While this smooths out the binary nature of promoters, it provides a useful approximation of mean system behavior and avoids computational failures.

The deterministic simulations demonstrated clear AND logic. In the OFF condition, reporter protein (REP) levels remained low throughout the 24 hour simulation (2e-20 after 24 hours). In the A-only and B-only conditions, REP levels remained low, slightly above background (6e-19, 7e-19 after 24 hours) but well below the activation threshold. Only in the dual-input condition did the reporter express significantly, showing a x100 fold as compared to the OFF condition.

Dose planning

Modeling of pharmacokinetics suggested that delivery method strongly impacts TRI-LYTAC dynamics in the brain interstitial fluid (ISF). Vesicle formulations were predicted to generate an early concentration peak, while mRNA-based delivery resulted in a slower but more sustained rise. Across both delivery methods, TRI-LYTAC showed a moderate (6-8h) half-life in ISF, indicating that repeated dosing or expression would be necessary to maintain therapeutic levels.

Higher levels of TRI-LYTAC resulted in faster seed degradation, as compared to lower exposures with longer clearance times. TauTuner modeling showed that light-driven activation could substantially decrease phosphorylated tau, with dark recovery phases helping to prevent overshoot or excessive dephosphorylation.

When both modules were combined, the system was predicted to act additively, reducing extracellular tau seeds and intracellular pTau simultaneously without evidence of antagonism. Sensitivity analysis highlighted several key determinants of efficacy, including receptor density, intracellular trafficking efficiency, and the rate of expression from delivered mRNA. Even more so, the integration of logic-gate restriction within the model reduced predicted off-target activity, reinforcing the potential for precise therapeutic control.