Circuit Engineering

In our initial circuit, we selected Clostridium butyricum as the chassis organism because it produces butyric acid, which helps restore the normal microbial community structure in the perianal region. For the functional modules, we initially considered using an anti-VEGF single-chain variable fragment (scFv) to inhibit angiogenesis. To facilitate the proper folding of the scFv, we also planned to overexpress the GroEL/GroES chaperones in the chassis bacteria. In the anti-inflammatory module, we employed a ROS-activated promoter to drive the expression of the anti-inflammatory cytokine IL-10. For biosafety, this version of the circuit utilized an HrpR/S AND gate, ensuring that the engineered bacteria would only survive under hypoxic conditions and at sites of inflammation.

However, a limitation of this design was that both IL-10 and the scFv are not ideally expressed in prokaryotic systems. Furthermore, Clostridium butyricumis is notoriously difficult to genetically manipulate.

In the second version, we switched the chassis organism to Escherichia coli Nissle 1917 (EcN). This probiotic strain can naturally migrate towards sites of inflammation, which aids the function of the anti-inflammatory pathway and helps optimize the gut microbiota. To enhance therapeutic efficacy, we focused on the HIF-1α-mediated angiogenesis pathway under hypoxia and adopted the Pepper-tCMA system to target and degrade HIF-1α during hypoxia. We intended to use the basic helix-loop-helix (bHLH) domain to enable our tCMA element to bind HIF-1α. Simultaneously, we replaced the original scFv and IL-10 with a nanobody and an engineered melittin peptide, respectively, as these are more amenable to expression in prokaryotes.

To improve the intracellular delivery of the Pepper-tCMA system, we planned to package it MS2-based virus-like particles (VLPs) and utilize a suicide system to release the assembled VLPs. The VLP surface was designed to display the wound-homing peptide CAR for targeting neo-vascular sites.

At the regulatory level, we introduced an adenosylcobalamin (vitamin B₁₂)-responsive riboswitch, AdoCbl, to create an activation circuit triggered by vascular rupture (a hallmark of advanced hemorrhoids). In this version, we planned to create two separate circuit variants for early-stage and late-stage treatment to broaden the therapeutic applicability.

However, this design had drawbacks: the tCMA system lacked sufficient specificity, its large molecular size prevented efficient packaging VLPs, and the switching mechanism between early and late-stage treatments was somewhat cumbersome. This led to further iterations.

To address the delivery and specificity issues of tCMA, we rationally designed a bioPROTAC (VHH-VHL) based on the nanobody VHH212 (specific for HIF-1α) and the VHL subunit of the E3 ubiquitin ligase complex. This molecule mediates the ubiquitination and degradation of HIF-1α under hypoxic conditions. We replaced VLPs with bacterial outer membrane vesicles (OMVs) for delivery, engineering them for targeted delivery by displaying the OmpA-CAR fusion on their surface.

Inspired by the traditional Chinese medicine concept of "medicinal adjuvants" (药引, a guiding drug that enhances the efficacy of main ingredients), we proposed a "Medicine-Food Collaboration" paradigm using a negative riboswitch and toxin protein. This riboswitch responds to a specific metabolite found in certain foods. In the absence of the ligand, the engineered bacteria express a toxin protein, leading to self-elimination and stopping the treatment. This paradigm aims to help patients develop healthier dietary habits alongside the therapy.

To address the potential risks of engineered bacterial over-proliferation in vivo and associated issues such as therapeutic module expression fluctuations and toxicity, we designed an intelligent global regulatory system based on a LuxR/TetR dual-factor cascade. This system employs a quorum-sensing mechanism to monitor bacterial density and utilizes a negative feedback loop to regulate the expression of downstream genes, thereby achieving precise control.

With the progress of Human Practices, we aimed to expand the application scenarios of HemorrEaser, potentially creating new treatment options for specific groups like pregnant women. Consequently, we designed the anti-angiogenesis module as the core of an oral medication, integrating it with the "Medicine-Food Collaboration" paradigm to combine pharmaceutical treatment with dietary health management. To alleviate the pain and inflammation associated with hemorrhoids, we developed the anti-inflammatory module as a separate topical drug. In the oral administration mode, we plan to process the engineered bacteria into lyophilized powder that retains viability, delivering it via enteric-coated capsules to the small intestine for targeted therapeutic action. ​​In the topical administration mode, the engineered bacteria are formulated into an ointment along with TPP, which serves as a survival signal. This topical formulation incorporates simple suicide elements to ensure biosafety.

Figure 4. Two Administration Routes: oral and topical, targeting angiogenesis and inflammation.

Based on the survey results of stakeholders from our Human Practices work, we incorporated an "expiry-date" circuit the Medicine-Food Collaboration module. This circuit delays the bacterial suicide mechanism, making it more convenient for patients to take the oral formulation of HemorrEaser. You can refer to the Design for details of the final circuit.

Global regulation module

We selected E. coli Nissle 1917 (EcN) as the chassis strain for our engineered bacteria. Our design phase focused on creating a LuxR/TetR-based dual-factor cascade system, autonomously control gene expression in response to bacterial population density, address the risk of uncontrolled proliferation and overtreatment. As shown in Figure 6., the core circuit was designed as follows: constitutive expression of luxI and luxR ensures low-level production of the AHL signal. As the cell density increases, accumulated AHL activates LuxR, which then binds to the P_lux to drive the expression of the TetR repressor. TetR subsequently binds to the P_tet to inhibit the expression of any downstream gene. To macroscopically validate the functionality of this circuit, we chose the AmCyan fluorescent protein as a reporter gene downstream of P_tet in a pET-28b(+) backbone, enabling visual assessment of circuit functionality.

As shown in Figure 7.A, the amCyan gene was placed downstream of the P_tet. After the construct was successfully introduced E. coli BL21(DE3), colony PCR was performed, and as shown in Figure 7.B, the bands matched the expected size. Sequencing analysis further confirmed that the recombinant plasmid was consistent with the design as shown in Figure 7.C.

The finalized plasmid was then transformed the E. coli BL21(DE3) expression host.

The first functional test employed our designed macroscopic observation protocol. Transformed E. coli BL21(DE3) cells were cultured in M9 minimal medium supplemented with 50 μg/mL kanamycin. After 48 hours of incubation in a white light environment, we observed distinct phenotypic differences: the experimental group exhibited blue-green fluorescence visible to the naked eye under ambient light, with the fluorescence intensity gradually increasing over time, while the empty-plasmid control group showed no observable fluorescence as shown in Figure 10.

This qualitative result provided crucial initial evidence that our genetic circuit was functioning as intended, with amCyan expression successfully triggered through the density-sensing cascade.

While the macroscopic assay provided preliminary confirmation of functional fluorescence expression and suggested the presence of characteristic kinetic profiles, visual observation alone was insufficient to definitively determine key circuit parameters such as activation kinetics, threshold density, or feedback strength. ​​Furthermore, it remained unclear whether the circuit could maintain downstream promoter activity within a stable range rather than merely providing transient suppression. We learned that a quantitative approach was essential to understand the precise temporal relationship between population growth and circuit output. This insight led to the redesign of our testing strategy, moving from qualitative observation to quantitative, time-resolved measurement.

To address the limitations of the macroscopic assay, we designed a quantitative experiment using a microplate reader. This approach enabled precise quantification of both bacterial growth and gene expression dynamics, allowing us to fit the one-step growth curve and the relative fluorescence unit (fluorescence intensity/OD₆₀₀) profile. Consequently, the P_tet repression kinetics could be characterized through temporal changes in normalized fluorescence intensity. The redefined test protocol involved monitoring the culture in real-time, with simultaneous measurement of optical density (OD₆₀₀) and fluorescence intensity (AU, Ex/Em: 453/486 nm) every 20 minutes over a 5.5-hour growth period. This approach allowed us to generate high-resolution, kinetic data for both growth and circuit output, enabling the calculation of relative fluorescence units (fluorescence intensity/OD₆₀₀) to normalize expression to cell density.

The kinetic data obtained from the microplate reader revealed a characteristic progression of the relative fluorescence. The OD₆₀₀ data exhibited a standard one-step growth curve, while the expression profile showed a rapid increase in fluorescence intensity/OD₆₀₀, which peaked and then declined to a steady state. The kinetic profile obtained was accurately fitted by an ODE model (R² = 0.9078), providing a quantitative representation of the quorum-sensing feedback system, as shown in Figures 11 and 12.

The data clearly demonstrated that the feedback inhibition circuit rapidly reached a peak expression level and subsequently maintained long-term stability in output, thereby achieving effective negative feedback control over the downstream therapeutic circuit.

Experimental data provide clear validation of the circuit's activation characteristics. The quantitative retesting phase was crucial for ODE-based model fitting. RFU intensity peaked at approximately 30 minutes, while the OD₆₀₀ at this timepoint was only 0.15, indicating that basal expression from the constitutive promoter enables preliminary AHL accumulation. This allows partial LuxR-mediated transcriptional activation to occur prior to the population reaching high density. Subsequently, RFU levels decreased rapidly. The high-fitting model further visually demonstrates this "fast-activation/early-repression" kinetic pattern, confirming the success of our circuit design. Notably, the system completed a full activation-repression cycle before the culture entered the stationary phase, demonstrating its ability to preemptively regulate gene expression in response to minor population changes and achieve autonomous control before potential overproliferation risks emerge. The integration of quantitative kinetic data with the molecular mechanism ultimately validates the reliability of this quorum sensing feedback system as the global regulation module in the project.

BioPROTAC

The primary goal of our design was to specifically target and degrade the HIF-1α protein to block its downstream inflammatory signaling cascade, thereby reducing the abnormal expression of target genes and mitigating inflammation in hemorrhoidal lesions. This strategy was initially conceptualized to leverage the chaperone-mediated autophagy (CMA) pathway.

Specifically, we utilized RFdiffusion to design a novel protein that, upon specific binding to HIF-1α, undergoes a conformational switch that exposes the KFERQ degradation tag, consequently inducing its CMA-mediated degradation.

Intriguingly, during the design process, we discovered that HIF-1α carries an autophagic degradation sequence (KFERQ-like motif), suggesting that the CMA pathway is an intrinsic component of its endogenous degradation machinery. Further literature review indicates that under normoxia, HIF-1α is tightly controlled and primarily degraded by the ubiquitin-proteasome system (UPS). Under hypoxia, however, UPS function is impaired, leading cells to primarily depend on the autophagy pathway for HIF-1α homeostasis. Crucially, the suboptimal efficiency of CMA degradation allows HIF-1α to readily accumulate and nuclear-translocate, activating the hypoxia signaling pathway and sustaining its high steady-state levels.

Following discussions with Professor Zhang Xiaoyan and further research, we concluded that merely augmenting CMA activity would be insufficient to robustly decrease HIF-1α's steady-state level. Consequently, we pivoted our strategy to the ubiquitin ligase system. We now propose to construct a bioPROTAC molecule to mediate the selective ubiquitination and subsequent proteasomal degradation of HIF-1α, anticipating a significantly more efficient suppressive outcome. You can see details in Human Practices

During the design of bioPROTAC, the linker's conformation optimization underwent multiple rounds of experimentation. Initially, to prevent structural interference between VHH212 and pVHL, we attempted a fully rigid α-helical linker (EAAAK)n to ensure spatial separation of the two ends. However, excessive rigidity was found to restrict the overall molecular flexibility, hindering the conformational adaptation of the ubiquitinating complex. Subsequently, we tested a fully flexible linker (GGGSS), which offered greater spatial freedom but increased interference between the two protein structures and reduced stability. Balancing the trade-offs between rigidity and flexibility, we ultimately adopted a hybrid linker design: the sequence EAAAKGPGSGGS. The EAAAK segment forms an α-helix, providing initial structural orientation for the amCyan while avoiding interference between structural components. The terminal GPGSGGS segment offers a flexible tail, enabling the two terminal proteins to maintain spatial separation while allowing necessary rotation. This iterative optimization ensures the amCyan preserves the structural integrity of both terminal proteins while providing sufficient flexibility to support bioPROTAC functionality.

Under hyp oxic conditions, the primary functional sites of HIF-1α and its transcriptional regulatory functions are concentrated within the cell nucleus. Therefore, the efficiency of nuclear delivery for bioPROTAC is crucial for achieving effective targeted degradation. Initially, we did not incorporate a nuclear localization signal (NLS) tag the bioPROTAC structure, as the fusion protein's molecular weight of approximately 33 kDa falls below the passive diffusion threshold (around 40 kDa) for the nuclear pore complex (NPC). Theoretically, this allowed passive diffusion the nucleus to bind with nuclear HIF-1α. However, considering that passive diffusion is typically slow and exhibits significant concentration gradient dependence, under cellular hypoxia stress conditions, viscosity differences between the nuclear and cytoplasmic environments and protein interactions may further limit the molecular nuclear transport rate, thereby affecting degradation efficiency.

Based on this, we introduced the classic nuclear localization sequence (NLS) PKKKRKV at the C-terminus of bioPROTAC to enhance its nuclear entry efficiency via importin-mediated active transport. The incorporation of NLS not only significantly enhances the nuclear enrichment and stability of bioPROTAC but also accelerates its spatial interaction with HIF-1α and the subsequent ubiquitination induction process. Through this enhancement, we aim to overcome the rate-limiting constraints of passive diffusion, ensuring bioPROTAC efficiently localizes to the nucleus under hypoxic conditions. This enables precise recognition and degradation of HIF-1α, maximizing its functional efficacy.

Initially, we hypothesized that bioPROTAC could promote modification of HIF-1α at its VHL-mediated classical ubiquitination sites (K532, K538, and K547). However, molecular docking and complex modeling revealed that these lysine residues are partially shielded and significantly less accessible in the spatial conformation formed by the bioPROTAC complex with Elongin B and Elongin C. Further analysis reveals that the long helical backbone of Cullin-2 struggles to bend sufficiently in this conformation to guide ubiquitin molecules to these shielded sites, thereby limiting ubiquitin accessibility and the feasibility of covalent transfer geometry.

Through literature review, we learned that the application of PROTAC technology often leads to alterations in the ubiquitinylation sites of target proteins. As Crowe et al. pointed out, the introduction of PROTAC may cause relocation of ubiquitinylation sites by reshaping the spatial conformation of the POI-PROTAC-E3 ternary complex. This primarily stems from changes in spatial conformation and increased steric hindrance, making it difficult for the E2-E3 complex to effectively recognize the original ubiquitinylation site. Therefore, we urgently need to explore new target residues with ubiquitinylation potential to ensure the degradation process proceeds smoothly.

The essence of the ubiquitination reaction is a nucleophilic addition process: the ε-amino group of lysine (Lys) on the substrate protein undergoes deprotonation to convert into the -NH₂ form, acquiring strong nucleophilicity. This enables it to attack the thioester bond in the ubiquitin-E2 complex, cleaving it and forming a new isopeptide bond, thereby achieving the transfer of the ubiquitin molecule. Based on this, we first employed GPS-Uber to predict potential ubiquitinylation sites on the substrate protein, yielding multiple candidate lysine residues. Subsequently, we employed PROPKA 3 to calculate the pKa values of these sites, evaluating their deprotonation capacity under physiological conditions to infer their potential ubiquitinylation activity.

The results indicate that Lys11 exhibits superior deprotonation capability, with a predicted pKa value of 6.96—significantly lower than the typical value of approximately 10.5 for lysine residues. This indicates that under physiological pH (7.4) conditions, the ε-amino group of Lys11 is predominantly deprotonated (-NH₂), thereby exhibiting enhanced nucleophilicity and favoring attack on the thioester bond to form an isopeptide bond. Structural analysis further supports this conclusion:

1. Lys11 exhibits a surface exposure of up to 98%, indicating that this residue is spatially accessible and readily recognized by the E2-E3 complex;
2. Its desolvation energy of -3.46 indicates that this site resides in a low-dielectric-constant environment, which helps lower the pKa and stabilize the deprotonated state;
3. This residue does not form significant hydrogen bonds, thereby avoiding potential stabilization effects that could restrict the active site.
4. Weak electrostatic interactions with Arg89 and Lys65 (-0.01 and -0.07) provide additional local shielding effects, further promoting the deprotonation process.

Subsequently, we employed HADDOCK and HDOCK for molecular docking and complex modeling of the complete substrate-E3-E2 ubiquitinylation system. The results indicate that the spatial distance between the reaction center of Lys11 and the Gly76 residue of the ubiquitin molecule is approximately 10.6 Å, satisfying the geometric requirements for effective nucleophilic attack. This suggests that this site possesses potential reactivity and can be validated as a novel ubiquitinylation target.

Finally, we performed a 100 ns molecular dynamics simulation of the substrate-E3-E2-ubiquitin system using Amber. From the simulation trajectory, we extracted a plausible attack conformation, i.e., the pre-reactive conformation. Building upon this, we selected the reaction core region for quantum mechanical calculations using Gaussian 16 to determine the energy barrier for nucleophilic attack. The results indicate that the transition state energy barrier for this reaction is approximately 8.2 kcal/mol, falling within a reasonable range and further supporting the feasibility of the Lys11 ubiquitinylation site.

In the design phase of the in vitro ubiquitination assay, we initially planned to use full-length HIF-1α as the substrate. The protein was intended to be recombinantly expressed and purified using a prokaryotic expression system (E. coli BL21(DE3)) to simulate ubiquitination under physiological conditions.

Figure 16. Expression vector with full-length HIF-1α.

We successfully constructed the recombinant plasmid pET-28b(+)-HIF-1α. Positive clones were preliminarily screened by colony PCR, and sequencing confirmed the correct insertion, indicating successful construction of the HIF-1α expression vector (Figure 17).

However, when the expression products were analyzed by 8% SDS-PAGE and Western blot, no clear band corresponding to HIF-1α was observed at the expected molecular weight. This suggested that full-length HIF-1α was either expressed at low levels or unstable in this system, making effective purification difficult.

Literature review revealed that the N-terminal region (amino acids 1-365) of HIF-1α contains both the VHH212 nanobody binding domain and key ubiquitination sites, while the C-terminal portion consists mainly of flexible loop regions with minimal impact on core structure and function. Therefore, we proposed a truncated expression strategy to improve soluble expression levels.

Based on this analysis, we redesigned the experimental approach, replacing the substrate in the ubiquitination assay with a truncated version of HIF-1α(1-365aa) to enhance protein expressibility and experimental feasibility.

Figure 18. Expression vector with truncated HIF-1α.

We successfully constructed the expression vector for truncated HIF-1α. Western blot analysis confirmed its high-level expression in BL21(DE3), and high-purity target protein was obtained through purification.

In the reconstituted ubiquitination reaction system, incubation of VHH-VHL, truncated HIF-1α, E3 ligase complex components (EloB, EloC, Rbx1, Cullin-2), and the ubiquitination system (UBA1, E2, ubiquitin) resulted in a distinct upward shift of the HIF-1α band on Western blot, confirming the occurrence of specific polyubiquitination.

By integrating computational rational design with multidimensional experimental validation, the engineered bioPROTAC molecule VHH-VHL effectively recruits the ubiquitination pathway to achieve targeted protein degradation while maintaining its VHH212 module's high-affinity binding to HIF-1α. These findings demonstrate that VHH-VHL effectively mediates the ubiquitin-proteasome degradation pathway of HIF-1α under hypoxic conditions in vivo, providing a molecular basis and therapeutic potential for inhibiting hypoxia-induced pathological angiogenesis in hemorrhoidal tissues. Next, we will validate the ubiquitination effect in mammalian cells.

Anti-VEGF Nanobody

To enable efficient expression of an anti-VEGF therapeutic in engineered bacteria, we adopted a rational design strategy. Unlike conventional IgG antibodies, nanobodies are intrinsically compatible with prokaryotic expression. We therefore selected a tandem bivalent nanobody construct that exhibits superior pharmacokinetic stability compared with the monovalent form. An N-terminal PelB signal sequence was included to direct the protein to the periplasm, and a C-terminal hexahistidine tag was added to enable purification by Ni-NTA affinity chromatography.

The cassette was cloned pET-28b(+) under the control of the lac operon, and the resulting plasmid was transformed E. coli SHuffle T7 Express.

Agarose gel electrophoresis of the PCR products from the colonies demonstrated that we successfully constructed the expression vector.

In the induction expression experiment, when the culture reached an OD₆₀₀ of 0.5, expression was induced with 0.5 mM IPTG for 16 h at 16 °C. Cells were harvested, periplasmic proteins extracted, and the extract purified by Ni²⁺-affinity chromatography. SDS-PAGE showed no detectable band at the expected ~30 kDa.

Further literature survey and sequence inspection led us to hypothesize that due to the identical repetitive segments in tandem bivalent repeat sequences, erroneous recognition and pairing may occur during homologous recombination. This likely hinders the proper expression of tandem bivalent nanobodies in SHuffle strains.

Professor Xu Shutong pointed out that our tandem sequence repeats were excessively high, which significantly increased the probability of sequence mismatches, making it difficult for the antibody protein to express correctly. Subsequent literature review revealed no significant difference in binding affinity between tandem bivalent nanobodies and monovalent nanobodies against VEGF. To address issues encountered in tandem bispecific nanobody expression, we redesigned the antibody sequence, retaining the N-terminal pelB tag and histidine tag while replacing the tandem bivalent nanobody sequence with a monovalent nanobody sequence.

Figure 23. Anti-VEGF monovalent nanobody.

The tandem bivalent sequence was converted to a monovalent form and constructed into the pET-28b(+) plasmid via homologous recombination, with expression controlled by the lac operon.

Through literature analysis and numerical modeling predictions and analysis, we concluded that the pharmacokinetic stability of monovalent nanobodies is reduced compared to tandem bivalent nanobodies.

To improve the pharmacokinetic stability of anti-VEGF nanobodies, we decided to explore methods to enhance the solubility and stability of monovalent nanobodies.

To optimize the anti-VEGF nanobody and offset the stability loss resulting from the tandem-to-monomeric conversion, we inserted the 33-residue P17 peptide between the PelB signal and the N-terminus of the monovalent nanobody. Studies indicate that the introduction of this P17 short peptide effectively enhances the antibody's stability and solubility.

We constructed the optimized monovalent nanobody sequence containing the P17 fragment into the pET-28b(+) plasmid via homologous recombination, maintaining expression regulation by the lac operon. After transformation into E. coli SHuffle, sequencing verified the correct insert.

E. coli SHuffle carrying the recombinant plasmid was grown at 37 °C, induced with 0.5 mM IPTG at 16°C for 16 h, and periplasmic extracts were purified. SDS-PAGE and Western blot of total and purified fractions showed a single band at the expected 17.7 kDa, confirming expression.

Purified anti-VEGF nanobody was tested in an MTT cell-proliferation assay. Data analysis revealed that the monovalent anti-VEGF nanobody we designed, expressed, and purified—containing the P17 short peptide—effectively binds to VEGF and inhibits its cell proliferation effects.

We also used ProteinSol to simulate the solubility of nanobodies with and without P17; the P17-containing variant showed higher solubility during expression and after secretion.

Although we have essentially achieved validation of antibody expression and function, we still believe there is potential for improvement in its application as a therapeutic drug. We must consider the targeting issue of antibodies. Diffusing antibodies carry a certain degree of systemic toxicity, which may affect normal angiogenesis in other parts of the body. Therefore, we have decided to further optimize the anti-VEGF antibody to enable it to act specifically on the site of hemorrhoids.

Considering potential safety concerns associated with the systemic diffusion of anti-VEGF antibodies, we designed and incorporated a masking peptide at the C-terminus of the antibody to ensure its specific targeting to the hemorrhoidal site. This peptide is linked to the antibody via a linker containing an MMP3 cleavage site. This linker can be cleaved by matrix metalloproteinase 3 (MMP3) expressed at the hemorrhoidal site, causing the masking peptide to separate from the antibody and exposing the antibody binding domain.

We designed the gene sequences for the masking peptide and linker, fused them to the C-terminus of the nanobody sequence, and inserted them between the antibody sequence and the histidine tag.

To validate the feasibility of the masking peptide, we analyzed all docking complexes with the Anti-VEGF antibody using Rosetta InterfaceAnalyzer, and identified several sequences with relatively weak interface energies. The representative complex was further examined using PRODIGY, yielding a binding free energy of -9.3 kcal·mol⁻¹ and a dissociation constant of K_d ≈ 2.6 × 10⁻⁷ M (260 nM), indicative of moderate affinity. The binding interface was dominated by hydrophobic (apolar-apolar, 16) and polar-apolar (10) contacts, while charged-charged (5) and polar-polar (1) interactions contributed less. The nonpolar interface area (41.0%) exceeded the charged area (23.1%), suggesting that binding is mainly driven by a hydrophobic core rather than electrostatic interactions.

To assess the dissociation process, steered molecular dynamics (SMD) simulations were performed with GROMACS to generate the unbinding pathway, followed by umbrella sampling and WHAM analysis to compute the potential of mean force (PMF). The resulting binding free energy (plateau-minimum) was -9.8 kcal·mol⁻¹, corresponding to K_d ≈ 1.22 × 10⁻⁷ M at 310 K. These results indicate that the masking peptide binds with moderate affinity and reversible stability, consistent with its expected functional behavior.

We have designed an Anti-VEGF Probody for targeted hemorrhoids therapy, which is activated by haemorrhoid-overexpressed MMP-3 to enable localized VEGF inhibition, potentially reducing systemic toxicity while improving solubility and stability.

However, this study remains preliminary. Functional validation is limited to HUVEC assays, and the full masked construct has not been experimentally built or tested. Its activation mechanism relies solely on computational docking, with no in vivo data on efficacy, stability, or immunogenicity. Another key obstacle is the low oral bioavailability resulting from the inability to penetrate intestinal epithelial cells.

To advance this candidate, future work must focus on constructing and validating the full protein, testing it in animal models, and—critically—engineering it with a cell-penetrating peptide (CPP) to enhance systemic absorption. Overcoming these limitations is essential to advance this targeted nanobody toward preclinical development.

ROS Sensing

In the area of hemorrhoid inflammation flare-up, the local region maintains a significantly higher concentration of reactive oxygen species (ROS) than normal tissue. This characteristic microenvironmental change becomes the key basis for our design of a targeted anti-inflammatory strategy. Accordingly, we selected ROS as the specific signal trigger for the anti-inflammatory functional module. Through genetic engineering modification, the engineered bacteria are endowed with ROS sensing ability—when the bacteria colonize the hemorrhoid lesions and perceive the high ROS signal, they precisely start-up the expression procedures of anti-inflammatory factors. Ultimately, through the targeted release of anti-inflammatory factors, targeted sexual inhibition of the local inflammation is achieved.

We selected SoxR/P_SoxS oxidative stress response promoter as the regulatory element. SoxR is a Transcription factor containing a [2Fe-2S] cluster, while P_SoxS is a promoter sequence. In a high ROS environment, SoxR binds between the -10 and -35 regions, and its Oxidized form distorts the promoter conformation, promoting the binding of RNA polymerase and other transcription factors, thereby activating transcription.

To validate the function of the SoxR/P_SoxS oxidative stress response promoter, we cloned amCyan as a reporter gene into the pUC57 vector, thereby constructing the corresponding reporter plasmid. The recombinant plasmid was then transformed into Escherichia coli BL21(DE3) for subsequent functional analysis.

Agarose gel electrophoresis of colony PCR products showed that we successfully constructed the expression vector.

Before oxidant induction, we used 15% SDS-PAGE electrophoresis to verify that soxR could be normally expressed in E. coli BL21 (DE3).

After verifying the expression of soxR, to verify the functions of the soxR and P_SoxS, we set up a series of tert - butyl hydrogen peroxide (tBHP) concentration gradients for induction (0 μM, 20 μM, 40 μM, 60 μM, 80μM, 100 μM, 120 μM, 140μM, 160 μM, 180μM, 200μM).

We found that when the concentration of tBHP was between 1-100 ppm, the expression of fluorescent protein increased rapidly. We added 200 μL of induction medium inoculated with engineered Bacteria into a 96 wells plate and used a microplate reader to detect the emission light intensity at 486 nm. Three parallel replicates were set for each induction concentration.

We observed that the activity of the P_SoxS is strongly dependent on ROS concentration, and its background expression remains low in the absence of ROS.

The current system relies solely on the positive activation of the P_SoxS by SoxR and does not incorporate Feedback loop such as the negative autoregulation by P_SoxS itself. This may lead to excessive expression of downstream therapeutic factors (such as Di-melittin) in a sustained High ROS environment, increasing the cell metabolic burden or triggering Local toxicity.

Although the system responds well to low concentrations of ROS (20-100 μM tBHP), at higher concentrations (>150 μM), the fluorescence intensity tends to saturate or even decline, possibly due to oxidative damage caused by high ROS levels to the bacterial cells, affecting their survival and expression capacity. This limits the applicability of the system in high-intensity inflammatory environments.

OMV delivery system

Escherichia coli outer membrane vesicles (OMV) can be modified to construct the intended anti-angiogenic hyperplasia module delivery system. To achieve precise targeted delivery of OMV to the Hemorrhoids site, we selected the OmpA-CAR fusion protein as the Targeting element.

To verify the function of OMV, we designed three independent circuits (and combined them into two experimental groups): the three circuits are ompA-CAR, ompA-amCyan, and srp-TagRFP; the two Combination schemes are: Combination 1 (ompA-amCyan and srp-TagRFP), Combination 2 (ompA-CAR and srp-TagRFP).

Among them, group 1 can assess the assembly efficiency of engineered OMV using fluorescence co-localization techniques, while group 2 evaluates whether CAR can exert the wound homing effect through cell-based assays, thereby clarifying its targeted delivery efficacy.

We integrated ompA-CAR and srp-TagRFP into the pET-28b(+) plasmid respectively through one-step In-Fusion cloning, integrated ompA-amCyan into the pET-28b(+) plasmid through two-step In-Fusion cloning, and transformed the constructed plasmids into Escherichia coli Nissle 1917 (EcN) for functional verification.

We submitted the assembled plasmid for sequencing and intended to validate protein expression upon successful sequencing. However, the results revealed that sequence alignment failed for all three vectors, necessitating reconstruction before further experiments could proceed. This indicates that the prior construction attempts were not successful.

Under these circumstances, the recombination process failed to generate the intended plasmid, likely due to reverse insertion of the target fragment, which resulted in unreadable sequencing data. Subsequent steps will involve redesigning the primers and repeating the construction.

Due to sequencing alignment failure, we redesigned the homologous arms flanking each sequence to ensure proper directional insertion.

We proceeded with plasmid construction using new primers, submitted the assembled plasmid for sequencing, and obtained successful results.

We introduced two plasmid combinations (ompA-amCyan with srp-TagRFP and ompA-CAR with srp-TagRFP) into the EcN strain via double transformation. PCR analysis of the transformed colonies revealed double-positive bands in both groups, confirming the successful co-transformation of both plasmid sets into the EcN strain.

To obtain high-purity OMV, we initially selected the density gradient centrifugation method (DGS) for extraction. The key advantage of this approach is its ability to achieve precise separation based on differences in density among various substances, theoretically allowing efficient removal of impurities such as bacterial debris and protein aggregates, thereby maximizing OMV purity. However, upon completion of the experiment, no distinct signs of OMV accumulation were observed in the expected density gradient regions.

Based on the experimental results, the density gradient Centrifugation method did not effectively isolate and extract the target OMV. Therefore, we opted to use differential centrifugation (DS) to repeat the extraction procedure.

We redesigned the extraction protocol for OMV by employing DS for purification, as detailed below:

1. Inoculate glycerol stock into 200 mL LB medium (supplemented with antibiotic) and culture at 37 °C until OD₆₀₀ reaches 1.0.
2. Centrifuge the culture at 12,000 ×g for 20 min at 4°C. Discard the bacterial pellet and collect the supernatant. Filter the supernatant through a 0.45 μm sterile filter to remove residual bacteria and cell debris, obtaining a sterile filtrate.
3. Transfer the sterile filtrate into a 100 kDa ultrafiltration tube and centrifuge at 4,000 rpm. Retain OMVs in the upper chamber of the ultrafiltration device. Repeat centrifugation, wash with PBS, and concentrate the sample to one-tenth of the original culture volume.
4. Ultracentrifuge the concentrated sample using a Beckman Optima XPN ultracentrifuge (SW 41Ti rotor) at 150,000 ×g and 4°C for 4 h. After centrifugation, discard the supernatant, wash the pellet with 2 mL sterile PBS three times, resuspend, and store at -80°C (avoiding repeated freeze-thaw cycles).

We performed the OMV extraction experiment following the protocol and obtained OMV precipitate.

Figure 40. DC Results.

We performed protein concentration determination on the extraction products of DGC and DC using a micro-spectrophotometer, respectively. The results showed that the protein concentration of the DGC product was close to the baseline level, while the protein concentration of the DC product was significantly higher. This finding provides preliminary evidence that the DC method successfully extracted OMV.

Figure 41. Comparison of protein concentrations between DGC and DC.

To further verify the extraction effect of OMVs, we selected the Outer Membrane Vesicle marker protein outer membrane protein A (OmpA, approximately 38.2 kDa) as a key validation marker to conduct a 12% SDS-PAGE experiment. The characteristic band of the outer membrane protein A can be clearly observed in the electrophoresis results, which provides direct evidence for the successful extraction of OMVs; furthermore, through intuitive comparison of the bands from extracts obtained by different centrifugation protocols, the feasibility and success of the Differential centrifugation protocol are once again confirmed.

Figure 42. 12% SDS-PAGE results

M: Protein marker; DC: Differential centrifugation extract; DGC: Density gradient centrifugation extract.

In this round of cycles, we found that although DGC is pure, it almost loses all OMVs, and DC can be used as the final experimental method. In the future, we will measure the Zeta potential of the extract to clarify the stability of the extract dispersion system and preliminarily determine the structure of OMVs. Subsequently, we will conduct fluorescence confocal experiment to confirm the co-localization of AmCyan and TagRFP, verifying the successful loading of Engineered outer membrane vesicle. Finally, we will perform cell scratch assay to determine the role of CAR in promoting cell migration and verify the targeting effect of Engineered outer membrane vesicle.

PelB-Di-melittin

Based on Di-melittin from NKU-China,2024(BBa_K5332002), we fused an N-terminal PelB signal peptide for secretion and a C-terminal 6×His tag for purification.

The sequence was cloned into the pET-28b(+) backbone using homologous recombination and transformed into E. coli BL21 (DE3).

Colony PCR and sequencing confirmed the presence of the correct construct.

The vector was successfully constructed, enabling us to proceed with expression analysis.

Since PelB should guide the peptide into the periplasm, we aimed to extract the protein from the periplasmic fraction.

Cultures were induced with IPTG, and proteins were isolated using osmotic shock methods. SDS-PAGE of the periplasmic extract showed no visible bands at the expected size (9.6 kDa).

The peptide might not have been cleaved/secreting correctly, or degradation may have occurred in the periplasm.

To check whether expression occurred intracellularly, we decided to analyze the whole-cell lysate.

Cells were directly lysed, and proteins were purified using a Ni-NTA column. SDS-PAGE revealed a clear band at 12 kDa, consistent with uncleaved PelB-Di-melittin(12.1 kDa).

Di-melittin was expressed but remained uncleaved, suggesting that the PelB tag was not processed as expected. This guided our future plans to test alternative signal peptides or improve periplasmic extraction conditions.

Suicide and Medicine-Food Collaboration

We selected E. coli Nissle 1917 (EcN) as the chassis strain for our engineered bacteria. EcN is a clinically approved probiotic with excellent safety, strong intestinal colonization ability, and suitability for gut-related therapeutic applications.

To prevent uncontrolled colonization or leakage of engineered bacteria after therapeutic effects, our goal was to construct a suicide module responsive to food-derived small molecules. Based on literature research, we identified hippuric acid—a downstream metabolite of anthocyanins enriched in black rice—as a candidate signaling molecule. We designed a genetic circuit in which a hippuric- acid-responsive riboswitch regulates expression of the toxin gene ccdB. When hippuric acid is present, the riboswitch remains "off" allowing survival during treatment; once hippuric acid is depleted due to diet cessation, the riboswitch is activated, driving ccdB expression and inducing cell death.

Since no hippuric-acid riboswitch could be identified or verified, we used the thiamine pyrophosphate (TPP) riboswitch as a functional analogue for preliminary validation.

We assembled the toxin gene ccdB downstream of the T7 promoter into pET-28b(+). After sequencing confirmation, the recombinant plasmid was transformed into E. coli BL21(DE3) for functional testing.

In parallel, the thiM riboswitch was placed upstream of the reporter amCyan. The TPP riboswitch-AmCyan module was assembled under constitutive P_J23100 in pUC57. After sequence verification, the plasmid was transformed into E. coli BL21(DE3) for testing.

Toxin protein validation. Engineered bacteria carrying the T7-ccdB construct were cultured in LB with IPTG at 0, 0.1, 0.5, and 1 mM; an empty-plasmid control was included. OD₆₀₀ monitoring showed significant growth inhibition at 0.5 mM IPTG. Plating every 30 min under 0.5 mM IPTG revealed time-dependent decreases in colony number and size.

thiM riboswitch validation. Engineered bacteria carrying the thiM riboswitch-ccdB construct were cultured in LB with TPP at 0, 10 μM, 100 μM, 1 mM, and 5 mM. OD₆₀₀ showed inhibition at 1 mM and 5 mM, indicating dose-dependent regulation by the thiM riboswitch.

Figure 52. Relative fluorescence intensity over time under different TPP induction.

Preliminary testing showed no response at low TPP, whereas induced ccdB expression effectively inhibited growth. Further work is needed to quantify the dynamic range and basal leakage of the thiM riboswitch.

We introduced the thiM-ccdB module to achieve orthogonal validation of the suicide system: with TPP present the riboswitch stays "off" and cells survive; without TPP it turns "on", drives ccdB expression, and induces cell death—a complete proof-of-concept.

The widely characterized thiM riboswitch was placed upstream of ccdB. The entire thiM-ccdB module was assembled under P_J23100 in pUC57. After sequence confirmation, the construct was transformed into E. coli BL21(DE3).

Strains carrying the thiM-AmCyan construct were cultured in M9 with a TPP gradient (0-15 mM, 1 mM gradient). OD₆₀₀ and fluorescence were recorded in parallel. Fluorescence remained unchanged at low TPP but dropped markedly at high TPP, confirming dose-dependent repression by the riboswitch.A four-parameter logistic (4PL) fit yielded IC₅₀=7 mM.

To verify the suicide module, strains carrying the thiM-ccdB construct were cultured in M9 with 0 mM vs. 7 mM (IC₅₀) TPP. OD₆₀₀ dynamics showed significant growth inhibition at 7 mM, confirming that the thiM specifically activates ccdB and triggers suicide.

Figure 55. OD₆₀₀ measurement under 7 mM TPP induction.

Induced CcdB expression effectively inhibited growth, and the suicide system exhibited clear dose dependence on TPP. For topical hemorrhoid applications, this module supports safe control; for oral use, the rapid response suggests optimizing concentration-time delivery.

Hippuric acid is a broad-spectrum metabolite widely present in legumes and cereals; its level fluctuates with diet and lacks tissue specificity. We therefore proposed personalized adaptation: patients choose specific foods to activate matching riboswitches, precisely timing bacterial suicide and reducing off-target risks.

Experiments further indicated that once signal supply stops, the suicide system activates within 2-3 h, and CcdB clears most bacteria within ~1 h. To build a clinically relevant time window, we integrated a genetic delay (expiry-date) circuit, extending activation to >40 h, thereby reducing dosing frequency.

To predict the interval from signal withdrawal to toxin-induced death, we modeled the expiry-date circuit and ran simulations with signaling molecules of different half-lives, obtaining corresponding CcdB concentration trajectories. We also simulated different initial doses to derive time-course curves under varied intake levels.

Based on the patient's selected diet, the corresponding molecules are identified. These molecules function as signals to activate the riboswitch, thereby regulating the expression of the CcdB toxin. In the presence of the signaling molecules, the riboswitch remain in the "off" state, allowing the engineered bacteria to survive during treatment. Through the ODE model, we know that after the intake of signal molecules is stopped, the suicide system will be activated after a delay of at least 40 hours.

With the initial concentration fixed at X_init = 20 μM, varying decay constants (per half-life) were used to numerically integrate CcdB expression over time; both shorter and longer half-lives significantly delayed toxin accumulation.

At a fixed half-life, varying initial doses (5, 10, 20, 50 μM) all reduced the accumulation rate of CcdB due to the delay circuit, indicating robustness across intake levels.

Dynamic curves of CcdB concentration based on the constructed circuit, with the dashed line indicating the concentration at which the engineered bacteria are completely eliminated: the top figure, CcdB dynamics induced by TPP as the signaling molecule; the middle figure, simulated CcdB dynamics for signaling molecules with different half-lives under the same dosage; the bottom figure, simulated CcdB dynamics for a specific signaling molecule under varying dosages.

With TPP as the signaling molecule, the engineered bacteria are fully cleared after 41.87 hours. With a signaling molecule half-life of 1000 min^-1, the engineered bacteria are fully cleared after 41.41 hours. With a signaling molecule concentration of 10 µM, the engineered bacteria are fully cleared after 41.54 hours. The suicide circuit can be triggered with a minimum delay of 40 hours across varying types and concentrations of signaling molecules.

Using the time to 10% of maximum CcdB as the "toxicity-threshold trigger time," response surfaces showed that both half-life and dose exert limited influence on CcdB accumulation when the delay circuit is present—supporting reliable timing control across broad conditions.

For hippuric acid—the small molecule candidate we selected—there are currently no publicly reported riboswitch aptamer sequences corresponding to it, nor is there sufficient research on its molecular interaction mechanisms. Consequently, it is difficult to rely on traditional molecular simulations or pure mathematical modeling to make predictions. Given this limitation, we turned to deep learning models for exploration, leveraging the "black-box" capability of neural networks to compensate for the gaps n mechanism research. This design concept is exploratory and innovative: it not only adapts to the current reality of insufficient experimental validation but also opens up a new aptamer prediction pathway based n generative models. Therefore, we propose the Mol2Aptamer model, which generates candidate RNA aptamer sequences by inputting the SMILES representation of small molecules.

First, we constructed a dataset for small molecules and their corresponding riboswitch aptamers. We integrated and filtered three public, experimentally validated nucleic acid aptamer databases: AptaDB, Aptagen, and the Global Nucleic Acid Aptamer Database. Ultimately, we obtained 292 small molecules and their corresponding 794 aptamer sequences. All these data are derived from actual experimental results, thus exhibiting high biological reliability. It should be noted that the number of experimentally validated small molecule-riboswitch aptamer pairs that can be found in the literature is inherently limited, making it difficult to further expand the dataset size at this stage. This objective limitation reflects the initial exploratory state of this research field and is precisely the motivation for our attempt to use deep learning models to expand predictive capabilities.

On the other hand, considering that aptamer sequences in different databases include both DNA and RNA formats, we uniformly converted all sequences into RNA representation. This processing may give rise to certain controversies, but we believe its rationality is primarily reflected in two aspects: First, the molecular recognition function of aptamers mainly depends on their secondary structure and spatial conformation. Since DNA and RNA share high similarity in base-pairing rules and structure formation mechanisms, sequence-level conversion will not alter their intrinsic "structure-function mapping". Second, from a modeling perspective, we regard the small molecule-aptamer prediction task as a sequence generation problem similar to translation. Unifying base types helps reduce the complexity of model learning and avoids noise caused by different coding systems. In summary, this preprocessing is not only biologically self-consistent but also improves the stability and generalization of model training from a methodological perspective.

Subsequently, we trained a dedicated BPE (Byte Pair Encoding) tokenizer based on this dataset to better tokenize and model SMILES strings and RNA sequences. We then constructed a deep learning framework with an encoder-decoder architecture: the encoder is responsible for extracting the structural features of small molecules, while the decoder gradually generates RNA aptamer sequences based on these features. The model architecture incorporates an attention mechanism, enabling the generation process to selectively capture key information in the molecular structure. Through this approach, we established a neural network model capable of end-to-end learning of the small molecule-RNA mapping relationship, laying the foundation for subsequent experiments and inference.

We evaluated the model using an independent validation set, focusing on assessing the diversity and rationality of the generated sequences. Additionally, we conducted preliminary structural and affinity validation of the candidate sequences using secondary structure prediction software and molecular docking tools. Meanwhile, we performed cross-validation on some existing small molecule-RNA data to test the model's generalization ability. Furthermore, we systematically integrated different generation strategies, including greedy decoding, Top-k sampling, Top-p (nucleus) sampling, and temperature adjustment, to balance exploratory generation and sequence rationality. Through these tests, we can preliminarily determine whether the RNA aptamers generated by the model have potential biological value and provide reliable candidate sequences for subsequent wet-lab screening.

During training, the model exhibited stable convergence: both training and validation loss decreased and eventually converged, indicating effective feature learning and good generalization. Token accuracy steadily increased, showing that the model progressively improved its ability to predict individual tokens correctly, while sequence accuracy, though lower, also increased, reflecting the greater difficulty of predicting entire sequences correctly. Meanwhile, perplexity consistently declined, further confirming convergence. Taken together, these results demonstrate that epoch 486 corresponds to the model's optimal performance, where loss is minimized and accuracy is maximized, highlighting a well-trained and balanced model.

We fed back the results of the test phase into model iteration. Through statistical analysis of the generated sequences and extraction of structural features, we can observe the small molecule-RNA interaction rules implicitly captured by the model, thereby providing insights for understanding the aptamer recognition mechanism. Future work can be optimized in two aspects: data and model. On the data side, we will gradually expand the training corpus to improve molecular coverage. On the modeling side, we will explore introducing graph neural networks (GNNs) to enhance molecular structure representation, or adopt contrastive learning strategies to strengthen the association representation between small molecules and RNA sequences. Ultimately, these improvements not only enhance the model's ability to generate aptamers but also lay the foundation for constructing a closed-loop optimization system combined with wet-lab SELEX (Systematic Evolution of Ligands by Exponential Enrichment). As experimental validation results continue to be fed back, the model will undergo continuous iteration, gradually forming an automated Design-Build-Test-Learn (DBTL) cycle that can drive the design of small molecule-RNA aptamers.

Future Work

In our current project, we have achieved significant results, including the successful validation of key biological parts through both wet lab and dry lab approaches. However, HemorrEaser still requires considerable development before it can be considered a fully realized therapeutic. In future work, we plan to characterize the feasibility of delivering bioPROTAC via outer membrane vesicles (OMVs) and investigate whether the OmpA-CAR fusion can maintain the wound-healing properties of CAR while achieving effective targeting. After completing the characterization of all individual components, we will proceed to assemble the complete genetic circuit and introduce it into the chassis organism. It is important to note that since nanobodies require a specific chemical microenvironment for proper expression in prokaryotic systems, we will also constitutively express the disulfide bond isomerase DsbC in the engineered E. coli Nissle 1917 (ΔnlpI) strain to optimize expression efficiency.

Looking ahead, we are committed to systematically addressing these challenges through rigorous characterization and optimization. By integrating these efforts, we aim to advance HemorrEaser from a proof-of-concept study toward a viable therapeutic candidate, ultimately contributing to the development of innovative treatment strategies for hemorrhoidal disease.