AI-guided precise design of antimicrobial polymers through high-throughput screening technology on an automated platform

Abstract

Antimicrobial peptides (AMPs)-mimicking antimicrobial polymers show great potential as therapeutic alternatives to antibiotics in the looming “post-antibiotic era”. However, the discovery of new AMP-mimicking antimicrobial polymers is challenging due to the vast chemical space of side-chain combinations. The advancement of AI-guided high-throughput screening enables more efficient, precise, and intelligent material design. Herein, we integrate combinatorial chemistry, machine learning, and automated high-throughput synthesis and characterization platforms to establish a new paradigm for the design of antimicrobial polymers with excellent biocompatibility. Starting with a library of 13,728 combinations, a seed dataset of 400 structures is generated, followed by four Design-Build-Test-Learn iterations using a new machine learning model. 7 top-performing candidates are screened with a minimum inhibitory concentration (MIC) ≤ 8 μg/mL and an inhibitory concentration causing 20 % cell death (IC₂₀) ≥ 64 μg/mL. The highest-performing polymer (MIC 2 μg/mL, IC₂₀ 256 μg/mL) shows similar in vivo therapeutic efficacy with ceftazidime. Overall, the integration of AI-guided high-throughput screening and combinatorial chemistry accelerates the discovery of new antimicrobial polymers, which provides a scalable strategy for developing novel antimicrobial agents.

Graphical abstract

Highlights

• •A new deep learning model based on Graph Transformer was constructed for antimicrobial polymer screening.
• •The integration of machine learning and high-throughput synthesis accelerated the discovery of new antimicrobial polymers.
• •The top-performing antimicrobial polymer showed excellent in vivo therapeutic efficacy.

1. Introduction

In recent years, multidrug resistance (MDR) has gradually intensified as a global challenge, as traditional antibiotic discovery lags behind the emergence of bacterial resistance[[1], [2], [3], [4], [5]]. Humanity is thus stepping into the “post-antibiotic era” [6]. Antimicrobial peptides (AMPs) are natural antibacterial molecules and form an integral part of the natural immune system in organisms [7]. Due to their multiple site-targeting ability and swift antimicrobial actions, it is too metabolically costly for pathogens to undergo multiple mutations and develop strong resistance[[8], [9], [10], [11], [12], [13]]. As a result, AMPs are promising candidates for combating multidrug-resistant bacteria. Despite their advantages, inherent drawbacks of AMPs such as low stability, difficult scale-up production, and high cost cannot be overlooked [14,15]. Therefore, researchers have started to focus on their chemical mimetics.

Polymers, with their controllable structures, diverse functionalities, and excellent processability, hold significant potential across various applications. Recognizing their potential in integrating antimicrobial properties with biocompatibility by controlling the charge and amphiphilicity of polymers, AMP-mimicking antimicrobial polymers are designed as promising candidates for treating bacterial infections[[16], [17], [18], [19], [20]]. However, since the number of possible structures increases exponentially with the number of parameters in combinatorial chemistry, screening AMP-mimicking antimicrobial polymers through traditional trial-and-error methods is challenging due to the vast chemical space of side-chain combinations of antimicrobial polymers [21,22]. The advent of techniques such as photoinduced electron/energy transfer reversible addition-fragmentation chain transfer polymerization (PET-RAFT) has eliminated the need for stringent anaerobic conditions required for polymerization of well-defined structures, making high-throughput synthesis of polymers feasible [16,[23], [24], [25], [26], [27]]. Moreover, the introduction of automated synthesis and characterization platforms could further improve the throughput, speed, accuracy, reproducibility, and cost-efficiency of experiments[[28], [29], [30], [31], [32]]. Despite advances, current high-throughput technologies and robot-assisted workflows often focus on only one or a few factors, leaving the effective exploration of the full complexity of polymer properties as a remaining challenge.

Machine learning (ML) has achieved significant advances in chemistry and materials science[[33], [34], [35]], owing to its strong capacity to model complex chemical systems and uncover hidden patterns within data. This capability enables accelerated discovery of novel materials and compounds with tailored functionalities. While ML has shown promise in the exploration of polymer chemical space[[36], [37], [38], [39], [40], [41]], applying it to antimicrobial polymers presents substantial challenges. The limitations of existing characterization methods and the absence of standardized testing protocols introduce inconsistencies in the data reported across literature, limiting the availability of large-scale, high-quality datasets. This scarcity of reliable and sufficiently annotated data hinders the development of ML models with strong generalization capabilities. Consequently, there is an urgent need to employ high-throughput experimental approaches and automated techniques to generate large-scale datasets that can support the development of reliable ML models for antimicrobial polymers.

In this study, we combine combinatorial chemistry, machine learning, and automated synthesis and characterization platforms to develop a new paradigm for designing AMP-mimicking antimicrobial polymers with optimal antibacterial activity and biocompatibility. Based on the action mode of most AMPs—which involves initial adhesion to negatively charged bacterial cell membranes via electrostatic interaction, followed by insertion of hydrophobic peptides into the hydrophobic region of the lipid bilayer once a threshold concentration is reached, we selected 11 cationic monomers, 13 hydrophobic monomers, 6 hydrophilic monomers, and 16 different monomer ratios, resulting in a polymer library with a total of 13,728 combinations. A seed dataset containing 400 combinations is first obtained using random sampling. Automated synthesis and characterization are then employed to efficiently produce a high-quality dataset, which is further used to train a machine learning model. Through interactive querying, 150 polymers are recommended in each round of active learning. After four rounds of iterations, the best-performing candidates that met the criteria of a minimum inhibitory concentration (MIC) below 16 μg/mL and a maximum polymer concentration causing 20 % hemolysis (HC₂₀) above 2048 μg/mL are selected for more detailed characterization. Antimicrobial and cytotoxicity characterization was carried out (the maximum inhibitory concentration causing 20 % cell death is denoted as IC₂₀), and the polymer with the lowest MIC value and the highest IC₂₀/MIC ratio was selected as the target polymer. Overall, this work enhances the efficiency of designing antimicrobial polymers with maximized antibacterial potency and minimized cytotoxicity through automated robots and machine learning algorithms, offering valuable insights for novel antimicrobial polymer discovery.

2. Materials and methods

2.1. Materials

Dimethyl sulfoxide (DMSO), diethyl ether, n-hexane, dichloromethane (DCM), tetrahydrofuran(THF), trichlormethane, ethylenediamine, and triethylamine (TEA), hydrochloric acid, sodium bicarbonate, sodium chloride, sodium sulphate, magnesium sulphate, alkaline aluminum oxide, methyl methacrylate, 2-(methacryloyloxy)ethyl (2-(trimethylammonio)ethyl) phosphate were purchased from Sinopharm Chemical Reagent Co., Ltd. 1,5-diaminopentane, dicarbonic acid, 1-pentanamine, neopentylamine, N,N-diethylacrylamide, 2-propenamide, 5,10,15,20-tetraphenylporphyrin (ZnTpp), and 1-N-phenylnaphthylamine (NPN) were purchased from Aladdin. PBS and saline was purchased from Biosharp. acryloyl chloride, 4-(aminomethyl)piperidine, (3-acrylamidopropyl)trimethylammonium chloride, 1-amino-4-methylpiperazine, butyl methacrylate, N-benzylacrylamide, 2-propenoic acid, 2-Propenamide, N-[2-(3,4-dihydroxyphenyl)ethyl]-2-methylprop-2-enamide, morpholine, and 2-[4-(2-hydroxyethyl)piperazin-1-yl]ethanesulfonic acid (HEPES) were purchased from Macklin. Benzyl methacrylate and glucose were purchased from RHAWN. Ceftazidime was purchased from Solarbio. 1,4-butanediamine, ethyl-α-methylacrylate, 1-bsutanamine, n-octyl amine, and Triton X-100 were purchased from Sigma-Aldrich. Tryptamine and 3,3-dipropylthiadicarbocyanine iodide (DiSC3) were purchased from Thermo Fisher. N,N′-Bis(tert-butoxycarbonyl)-S-methylisothiourea was purchased from Shanghai Dibai Experiment Equipment Co.,Ltd. Resazurin was purchased from Shanghai yuanye Bio-Technology Co., Ltd. 2-Propenoic acid and iso-propyl methacrylate were purchased from TCl. 2-Imidazole-1-yl-ethylamine was purchased from Boer. 1-Heptanamine was purchased from Energy Chemical. mPEG-acrylamide (Mn = 350) was purchased from Tanshtech. 2-[[(dodecylthio)thioxomethyl]thio]propanoic acid (DoPAT) was purchased from Leyan. Deuterated DMSO and deuterated chloroform were purchased from J&K Scientific Ltd.

2.2. Synthesis of monomers Terminated with an amino group

A standard procedure was employed for the synthesis of five cationic monomers (N-(2-aminoethyl)acrylamide, N-(3-aminopropyl)acrylamide, N-(4-aminobutyl)acrylamide, N-(5-aminopentyl)acrylamide, and 1-(4-(aminomethyl)piperidin-1-yl)prop-2-en-1-one) with Boc protection from their corresponding diamines [42]. Briefly, 0.3 mol of diamine was dissolved in 300 mL of chloroform, followed by the dropwise addition of di-tert-butyl dicarbonate (0.03 mol in 150 mL of chloroform) over 2 h at 0–5 °C. The reaction mixture was then stirred at 25 °C for an additional 20 h. The resulting white precipitate was filtered, and the organic phase was thoroughly washed with water to remove excess diamine. The organic layer was then dehydrated over Na₂SO₄, filtered, and dried in vacuo. The intermediate obtained was used immediately in the next step without further purification. The intermediate was dissolved in 150 mL of tetrahydrofuran, and triethylamine (36 mmol) and acryloyl chloride (31.5 mmol) were added dropwise to the solution at 0–5 °C under nitrogen bubbling. The reaction mixture was stirred at 25 °C for 1 h. The urea by-product was filtered, and the solvent was removed in vacuo. The crude product was dissolved in chloroform and washed with brine. The organic phase was stirred with MgSO₄ and basic Al₂O₃ for 10 min, filtered, and concentrated in vacuo. The product was further purified by repeating the precipitation step twice in hexane, yielding the tert-butoxycarbonyl (Boc)-protected monomer as a fine white powder, which was then dried in vacuo.

2.3. Synthesis of monomers Terminated with an guanidine group

The Boc-protected cationic monomer N-(2-guanidinoethyl)acrylamide was synthesized by following the protocol that was described by Martin [43]. 1,3-Bis(tert-butoxycarbonyl)-2-methyl-2-thiopseudourea (25.3 mmol in 50 mL DCM) was gradually added to ethylenediamine (70.4 mmol in 60 mL DCM) in a 250 mL round-bottom flask equipped with a magnetic stir bar. Following complete addition, the reaction mixture was stirred at room temperature for 4 h. The resulting solution was washed with water and brine, and the organic layer was dried over MgSO₄. The DCM was then removed under reduced pressure, yielding crude 2-[1,3-Bis(tert-butoxycarbonyl)guanidine]ethylamine as a white solid. The crude product was dissolved in 200 mL of DCM and transferred to a 500 mL round-bottom flask equipped with a magnetic stir bar. Triethylamine (29.4 mmol) was added, and the solution was cooled in an ice bath. Acryloyl chloride (21.4 mmol) was added dropwise, and the reaction mixture was stirred at room temperature overnight. Saturated NaHCO₃ was added, and the aqueous layer was extracted with DCM. The combined organic layers were dried over MgSO₄, and the solvent was removed under reduced pressure. The product was recrystallized from diethyl ether at temperatures below 5 °C and further purified using column chromatography (silica, hexane/ethyl acetate) to yield 2-[1,3-Bis(tert-butoxycarbonyl)guanidine]ethyl acrylamide as a white solid.

2.4. Synthesis of monomers Terminated with an alkyl chain or heterocycle

A standard procedure was employed for the synthesis of three cationic monomers (N-(2-(1H-indol-3-yl)ethyl)acrylamide, N-(4-methylpiperazin-1-yl)acrylamide, and N-(2-(1H-imidazole-1-yl)ethyl)acrylamide) and five hydrophobic monomers (N-pentylacrylamide, N-isopentylacrylamide, N-neopentylacrylamide, N-heptylacrylamide, and N-octylacrylamide) from their corresponding amines [20]. In brief, 0.3 mol of amine was dissolved in 300 mL of THF. TEA (0.36 mol) and acryloyl chloride (0.36 mol) were then added dropwise to this solution at 0 °C under a nitrogen atmosphere. The reaction mixture was stirred overnight at room temperature. By-products were removed by filtration, and the solvent was removed using a rotary evaporator. The resulting crude product was dissolved in chloroform and sequentially washed with 0.1 M hydrochloric acid, saturated sodium bicarbonate solution, brine, and deionized water. The organic layer was dried over MgSO₄ and basic Al₂O₃, then filtered to remove any solids. Finally, the solvent was removed by rotary evaporation to obtain the purified product.

2.5. Synthesis of polymers and deprotection

Polymers were synthesized in 96-well plates using a fully automated PET-RAFT system, with DoPAT as the chain transfer agent, ZnTPP as the photocatalyst, and green light (λ = 520–530 nm, 5 W) as the light source [44]. The monomer, chain transfer agent, and photocatalyst were pre-dissolved in DMSO to achieve concentrations of 0.25 mmol/mL, 0.05 mmol/mL, and 0.5 mg/mL, respectively. The automated synthesis was conducted using a program that directed the Hamilton Microlab STAR liquid-handling robot with inputs such as pipette volume and well plate positions. The volumes handled were 240 μL for the monomer, 30 μL for the chain transfer agent, and 30 μL for the photocatalyst. After the liquids were dispensed, the plate was sealed and irradiated with green light for 16 h, resulting in the synthesis of a series of polymers with a degree of polymerization of 40, according to specific ratios of structural units. For monomers with Boc protection, deprotection was performed post-polymerization [16]. Briefly, 600 μL of TFA was added to the completed solution, and the plate was left for 3 h at room temperature to re-expose the primary amine or guanidine groups.

2.6. Construction of ML model

We constructed graph representations of copolymers by following the framework introduced in our earlier work [45]. Specifically, the repeating units of the three comonomers were individually represented as molecular line graphs {G_i}. A virtual node v_g was introduced and connected to all nodes across the subgraphs to serve as a communication bridge for message passing. Additionally, three separate virtual nodes {v_i} were each connected to one of the subgraphs {G_i}, encoding the relative composition ratios of the corresponding components {r_i}. The integrated graph constructed from all subcomponents was denoted as G. The atom and bond features in the graph are summarized in Tables S1 and S2, while the features of the virtual nodes were initialized as described below:

$$h_v=\{\begin{array}{c}\text{embedding}\left(r_i\right),\text{if}v\in \left\{v_i\right\}\\ h_g,\text{if}v=v_g\end{array}$$ (1)

where embedding(·) denotes the initial feature embedding function that maps ratio into a continuous vector space, and h_g is a learnable vector.

In this study, we employed a graph transformer as the graph encoder f(⋅), which was pretrained via self-supervised learning based on our established strategies [45,46]. The architecture of the graph transformer is illustrated in Fig. S1. Following the graph transformer, a linear projection layer was applied to predict antimicrobial and hemolytic activities. The model was finetuned using a mean squared error (MSE) loss function. Then, 20 models were independently trained and subsequently combined into an ensemble predictor.

The dataset was randomly split into training and test sets in an 8:2 ratio. Ten percent of the training set was further set aside as a validation set. Model training was performed using the Adam optimizer with a learning rate of 3 × 10⁻⁴, dropout of 0.2, and a batch size of 32. Early stopping with a patience of 20 epochs was applied. All hyperparameters were tuned using a grid search strategy, and the corresponding search ranges are shown in Table S3. The loss curves are shown in Fig. S2. The implementation was based on PyTorch and DGL, and all experiments were conducted on NVIDIA RTX 4090 GPU.

2.7. Active learning

In the active learning stage, the trained ensemble predictor was used to estimate the properties of all remaining copolymers in the unlabeled pool. To balance antimicrobial activity and hemolytic toxicity, we defined a composite score for each polymer as:

$$s=w{y}_1+\left(1-w\right)y_2$$ (2)

where y₁ and y₂ denote the predicted antimicrobial and hemolytic activities, respectively, and w is a weighting factor set to 0.8 in this study. The mean (μ) and standard deviation (σ) of the scores predicted by the ensemble model were used to represent the central tendency and uncertainty of each candidate polymer's performance. An upper confidence bound (UCB) acquisition function was employed to prioritize candidate selection, defined as UCB(G) = μ + βσ, where β is a hyperparameter that balances exploitation and exploration, and was set to 2 in this study. To promote diversity among selected samples, the top-ranked (10 %) candidates were clustered using K-means (n = 50), and the top 3 candidates from each cluster were selected, resulting in 150 copolymers for the next round of synthesis and labeling. The newly generated dataset was then incorporated into the existing training set for subsequent iterations.

2.8. Minimum inhibitory concentration (MIC) assay

The MIC of the prepared polymers was determined via the broth microdilution method in accordance with the Clinical and Laboratory Standards Institute (CLSI) guidelines. The bacterial culture was initiated from a single colony and incubated overnight in 10 mL of Mueller-Hinton broth (MHB) at 37 °C with constant shaking at 180 rpm. To prepare a subculture, 100 μL of the overnight culture was diluted into 10 mL of fresh MHB and allowed to grow until reaching the mid-logarithmic phase. This subculture was then further diluted to achieve an approximate cell density of 10⁶ CFU/mL. A two-fold serial dilution of 100 μL polymers (previously dissolved in MHB) was performed in 96-well plates, followed by the addition of 100 μL of the subculture suspension. Positive controls without polymer and negative controls without bacteria or polymer were also included. The plates were then incubated statically at 37 °C for 16 h. After the incubation period, 25 μL of resazurin solution (1 mg/mL in sterile PBS) was added to each well and the plates were incubated for an additional 2 h [44]. The MIC was defined as the lowest concentration of the polymer that prevented a visible color change from blue to pink, indicating no significant bacterial growth. All assays were performed in triplicate and repeated in at least three independent experiments.

2.9. Minimum bactericidal concentration (MBC) assay

The MBC of the prepared polymers was defined as the lowest concentration that kills 99.9 % of the bacteria after incubation. To determine the MBC, 10 μL of the solution from the MIC assay was pipetted onto Mueller-Hinton agar (MHA) plates. The plates were then incubated at 37 °C in a static incubator for 20 h. The lowest concentration of the polymer at which no bacterial colonies were observed was recorded as the MBC [47].

2.10. Hemolysis assay

The hemolysis assay was conducted using rabbit red blood cells (RBCs) to evaluate the hemolytic activity of the polymers. Fresh rabbit blood was collected from Dr.Can Biotechnology (Zhejiang) Co., Ltd., diluted in PBS, and then centrifuged at 1500 rpm for 20 min to separate the RBCs. Specifically, the blood collection followed the guidelines issued by the Lab of Animal Experimental Ethical Inspection of Dr.Can Biotechnology (Zhejiang) Co., Ltd. (License: DRK-20240201112). The RBCs were washed three times with PBS and resuspended in PBS to achieve a 5 % (v/v) RBC solution. Polymer samples were pre-dissolved in PBS at specific concentrations. In a 96-well plate, 100 μL of the polymer solution was added to each well and serially diluted. Subsequently, 100 μL of the RBC suspension was added to each well. The plate was then incubated at 37 °C with shaking at 150 rpm for 1 h. Following incubation, the plate was centrifuged at 1500 rpm for 8 min using a plate centrifuge. 100 μL supernatant was carefully transferred to a new 96-well plate, and the absorbance at 590 nm was measured using a microplate reader. Triton X-100 (1 % v/v in PBS) and PBS were used as positive and negative controls, respectively [48]. All experiments were conducted in triplicate and repeated at least three times independently. The percentage of hemolysis was calculated using the following equation:

$$\text{Hymolysis}\text{ratio}(\%)=\frac{A_{polymer}-A_{negative}}{A_{positive}-A_{negative}}\times 100\%$$ (3)

2.11. Cytotoxicity assay

Cytotoxicity assays were assessed using the 3T3 cell line. 3T3 cells were cultured in Dulbecco's modified Eagle's medium (DMEM) containing 10 % fetal bovine serum (FBS), 100 U/mL penicillin, and 100 μg/mL streptomycin. Cell growth was maintained in an incubator at 37 °C with 5 % CO₂. 3T3 cells were seeded in 96-well plates at a density of 10,000 cells per well and incubated for 24 h to allow for cell attachment. Different concentrations of polymer samples (pre-dissolved in medium and serially diluted) were added to each well, and the incubation was continued for an additional 24 h. Subsequently, the medium was then replaced with 100 μL fresh medium containing CCK-8 (10 %), gently shaken, and incubated for 2 h. The absorbance of each well was measured at 450 nm using a microplate reader. All experiments were performed in triplicate and repeated at least three times independently [49].

2.12. Colony forming unit (CFU) counting assay

E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in PBS to a final concentration of 10⁸ CFU/mL. The polymer was then incubated with the bacteria for 6 h. Subsequently, a CFU counting assay was performed using either the spread plate method or the spot plate method. For the spread plate method, 100 μL of the diluted bacterial suspension was evenly spread on MHA plates. For the spot plate method, 10 μL of different dilutions of the bacterial suspensions were pipetted onto the plates [47].

2.13. Bacterial killing kinetics

E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in PBS to a final concentration of 10⁵ CFU/mL. The suspension was then mixed with the target polymer at a concentration of 2-fold MBC. At the set time point (0 min, 2 min, 4 min, 6 min, 8 min, 10 min, 20 min, 30 min, 40 min, 50 min, 60 min, 70 min, 80 min, 90 min), an aliquot of 10 μL solution was sampled and pipetted on the MHA plate to count the alive bacteria colony [50].

2.14. K⁺ leakage

E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in PBS to a final concentration of 10⁸ CFU/mL. The bacteria were then incubated with the polymer at a concentration of 2 times the MBC for 6 h at 37 °C. After incubation, the supernatant was collected and the concentration of K⁺ in the solution was determined by ICP-MS [51].

2.15. Live/dead cell viability assay

E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in PBS to a final concentration of 10⁸ CFU/mL. The bacteria were then incubated with the polymer at a concentration of 2 times the MBC for 6 h at 37 °C. After incubation, the bacterial cells were stained using the Live/Dead BacLight Bacterial Viability Kit (purchased from Beyotime Biotechnology) according to the manufacturer's instructions. Briefly, an equal volume of the SYTO 9 and propidium iodide staining solutions was added to the bacterial suspension and incubated in the dark at room temperature for 15 min. The stained samples were then analyzed using a fluorescence microscope. Live bacteria exhibited green fluorescence, while dead bacteria showed red fluorescence [6].

2.16. Bacteria resistance study

In the bacteria resistance study, the MIC value in the first round MIC assay was designated as MIC₀. Following the first round, the bacterial suspension was diluted 8000-fold and inoculated into a new 96-well plate. The target polymer and the commercially available antibiotic ceftazidime were then added to the bacterial suspension to achieve a final concentration of 1/2 MIC₀. After 20 h of incubation, the MIC₁ values were determined. For each subsequent round, the working concentration was maintained at 1/2 MIC_n, assuming that bacterial growth proceeded through 10 generations per cycle. MIC values were continuously recorded until 100 generations were reached [47].

2.17. Transmission electron microscope (TEM) characterization

TEM was utilized to observe the ultrastructure of bacteria before and after treated with the polymers. Bacterial cells were cultured overnight in MHB at 37 °C with shaking at 180 rpm, then treated with the polymers at the desired concentration for 6 h. After treatment, the cells were centrifuged at 5000 rpm for 5 min, and the supernatant was discarded. The bacterial pellet was washed three times with PBS and then fixed with 2.5 % glutaraldehyde at 4 °C overnight. Following fixation, the samples were washed with PBS and post-fixed with 1 % osmium tetroxide for 1 h at room temperature. The cells were then dehydrated through a graded series of ethanol concentrations and embedded in epoxy resin. Ultrathin sections were cut using an ultramicrotome, placed on copper grids, and stained with uranyl acetate and lead citrate. The samples were examined using a transmission electron microscope, and images were captured to analyze bacterial structural changes [52].

2.18. Scanning electron microscope (SEM)

SEM was utilized to observe the morphology of bacteria before and after treated with the polymers. Bacterial cells were cultured overnight in MHB at 37 °C with shaking at 180 rpm, then treated with the polymers at the desired concentration for 6 h. After treatment, the cells were centrifuged at 5000 rpm for 5 min, and the supernatant was discarded. The bacterial pellet was washed three times with PBS and then fixed with 2.5 % glutaraldehyde at 4 °C overnight. Following fixation, the samples were washed with PBS and post-fixed with 1 % osmium tetroxide for 1 h at room temperature. The cells were then dehydrated through a graded series of ethanol concentrations and embedded in epoxy resin. Ultrathin sections were cut using an ultramicrotome, placed on copper grids, and stained with uranyl acetate and lead citrate. The samples were examined using a scanning electron microscope, and images were captured to analyze bacterial morphology changes [52].

2.19. Cytoplasmic membrane depolarization

The cytoplasmic membrane depolarization of bacteria was assessed using a membrane potential-sensitive dye DiSC3(5). E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in HEPES to a final concentration of 10⁷ CFU/mL. The bacterial suspension was treated with 0.4 μM DiSC3 for 1 h to assess cytoplasmic membrane depolarization. Potassium chloride was then added to the suspension to achieve a final concentration of 0.1 M. A volume of 195 μL of this suspension was pipetted into each well of a 96-well plate. The fluorescence intensity was measured using a microplate reader with an excitation wavelength of 622 nm and an emission wavelength of 670 nm. Once the fluorescence intensity stabilized, 5 μL of the target polymer was added to the suspension to reach a final concentration of 2-fold MBC. The change in fluorescence intensity was monitored and recorded by the microplate reader. HEPES buffer served as the negative control, and 2.5 % Triton X-100 was used as the positive control [53].

2.20. Outer cell membrane permeability

The outer cell membrane permeability of bacteria was assessed using N-phenyl-1-naphthylamine (NPN) uptake assay. E. coli was cultured overnight in MHB at 37 °C with shaking at 180 rpm, then diluted to a concentration of 10⁸ CFU/mL. The bacterial suspension was centrifuged at 5000 rpm for 5 min, the supernatant was removed, and the cells were washed and reconstituted in HEPES to a final concentration of 10⁷ CFU/mL 185 μL bacterial suspension was pipetted into each well of a 96-well plate, followed by the addition of 10 μL of 10 μM NPN solution. The fluorescence intensity was measured using a microplate reader with an excitation wavelength of 350 nm and an emission wavelength of 420 nm. Once a stable baseline was achieved, 5 μL of the target polymer was added to the suspension to reach a final concentration of 2-fold MBC. The increase in fluorescence intensity, indicating increased membrane permeability, was recorded. HEPES buffer served as the negative control, and ceftazidime was used as the positive control [53].

2.21. Subcutaneous abscess model

All animal experiments were conducted in accordance with the guidelines issued by the Lab of Animal Experimental Ethical Inspection of Dr.Can Biotechnology (Zhejiang) Co., Ltd. (License: DRK-20240201112). At Day 0, 39 female ICR mice (20–22 g, obtained from animal center of Zhejiang Academy of Medical Sciences) were anesthetized with 4 % isoflurane, and the left leg was depilated. Subsequently, 50 μL of 10⁸ CFU/mL of E. coli was injected subcutaneously into the thighs. After 24 h, the modeling was completed, and the mice were randomly and equally divided into three groups, receiving an injection of the target polymer, ceftazidime, or saline at a dose of 0.5 mg/kg, respectively. After seven days, the mice were euthanized, and the injection sites were photographed. Ten homogenates were taken from each group for bacterial counting assays, while the remaining three samples were embedded in paraffin, sectioned, and stained for histological analysis [51].

2.22. Skin wound infection model

All animal experiments were conducted in accordance with the guidelines issued by the Lab of Animal Experimental Ethical Inspection of Dr.Can Biotechnology (Zhejiang) Co., Ltd. (License: DRK-20240201112). At Day 0, 39 female ICR mice (20–22 g, obtained from animal center of Zhejiang Academy of Medical Sciences) were anesthetized with 4 % isoflurane, and the back was depilated. A circular wound approximately 1 cm in diameter was created on the dorsal skin of each mouse. Then, 20 μL of a bacterial suspension containing 10⁸ CFU/mL was added dropwise to the wound to establish the infection model. After 24 h, the mice were randomly divided into three groups and treated with either the target polymers, ceftazidime, or saline at a dose of 0.5 mg/kg. Wound healing was monitored and recorded daily. On day 7, the mice were euthanized for further analysis. Ten homogenates were taken from each group for bacterial counting assays, while the remaining three samples were embedded in paraffin, sectioned, and stained for histological analysis [51].

3. Results and discussion

3.1. Accuracy and efficiency of automated workflows

Considering the structure and mode of action of most AMPs, as well as the unique volumetric and precision requirements posed by machine learning within discrete chemical spaces, we systematically examined a range of structural parameters to inform polymer design. Specifically, to create a diverse antimicrobial polymer library, 11 distinct cationic monomers, 13 hydrophobic monomers, and 6 hydrophilic monomers were chosen, and were then combined in 16 different molecular ratios, resulting in a total of 13,728 unique polymer combinations (Fig. 1a). These monomers were either purchased or synthesized and characterized by ¹H NMR (Fig. S3–S16). This extensive polymer library serves as the foundation for subsequent testing and model training, allowing us to employ machine learning to correlate specific structural features with antimicrobial activity and biocompatibility.

Fig. 1.

Overview of the machine learning-guided high-throughput screen of antimicrobial polymers on an automated platform. a) Different monomers and ratios chosen for antimicrobial library construction. b) Schematic illustration of the seed dataset generation and screen of antimicrobial polymers through the Design-Build-Test-Learn cycle.

We initially employed random sampling to select 400 polymer combinations, thereby creating a diverse seed dataset for subsequent analysis. As illustrated in the dimensionality reduction plot (Fig. 1b), the combinations in this seed dataset are randomly distributed across the entire library, ensuring comprehensive coverage of the design space and minimizing selection bias. The selected 400 polymers were then synthesized by PET-RAFT, which were performed on an automated workflow in iChemFoundry. iChemFoundry is a large-scale platform for molecular manufacturing [53,54], which contains several functional islands for automated synthesis and characterization. In this work, we ingeniously connected the Hamilton Microlab STAR liquid-handling robot and the microplate reader via a tracked robot to enhance the throughput, speed, accuracy, and reproducibility of the experiments. Specifically, the monomers, chain-transfer agent, and photocatalyst were initially dissolved in DMSO at a certain concentration, and information including sequence setting, well position, and reagent volumes was submitted for sample addition and subsequent gradient dilution. After that, the well plates were transferred to the microplate reader by the tracked robot to test the absorbance. The antimicrobial (Fig. S17) and hemolytic (Fig. S18) heatmaps of the seed dataset reveal a low proportion of combinations meeting the required criteria, specifically MIC values below 16 μg/mL and HC₂₀ values above 2048 μg/mL, thereby highlighting the necessity for improved screening efficiency. A Design-Build-Test-Learn cycle was used to accelerate polymer design, beginning with data-driven modeling to analyze the seed dataset. ML models captured structure-property patterns associated with antimicrobial activity and biocompatibility. Based on these insights, a Bayesian optimization-based active learning approach was further employed to iteratively propose new polymer candidates, enabling efficient exploration of the design space and prioritization of high-potential combinations. In the subsequent Build phase, the designed polymers are synthesized through automated PET-RAFT, enabling rapid, high-throughput synthesis of numerous polymer combinations. The synthesis process is tightly controlled to ensure consistency in polymer structure and functionality, essential for reliable downstream testing. Finally, in the Test phase, Escherichia coli (E. coli) and rabbit red blood cells were selected for high-throughput characterization, including antimicrobial assays (using Resazurin staining to assess bacterial viability) and hemolysis assays (measuring optical density at 590 nm to evaluate blood cell compatibility). These assays provide critical data on the antimicrobial activity and biocompatibility, which are fed back into the learning phase for further optimization.

To demonstrate the reliability of the automated synthesis robot, six combinations were randomly selected and synthesized using both the robot and manual methods. No significant difference was found in the composition, molecular weight, and polydispersity of the polymers (Fig. S19). Meanwhile, the monomer conversion rate was approximately 100 %, suggesting that post-treatment to remove unreacted monomers could be omitted, which significantly simplifies continuous automated synthesis and characterization. Additionally, the time required for synthesis and characterization of 400 combinations was compared. While traditional manual methods took approximately 120 h, the automated process completed the task in just 10 h, improving efficiency by 12-fold and significantly reducing labor costs. Thanks to the high efficiency of computer program and significant reduction of time required for synthesis and characterization, we successful screened 7 lead candidates from 13,728 combinations in only 12 days. By comparing the structural characteristics of seven leading candidates, we observed that polymers with primary amine groups as cationic moieties show excellent antibacterial activity, provided moderate hydrophobicity coexists at the same molecular core. In the application scenario focused on in this work, a spacer length of five methylene groups was found to be optimal. It is worth noting that this is not a universal rule, as numerous other parameters influence activity, including the type of functional groups, topological structure, and the number of cationic, hydrophobic, and hydrophilic blocks. More interestingly, in contrast to the common perception that stronger positive charge and hydrophobicity generally correlate with higher antibacterial activity but poorer biocompatibility, our high-throughput experiments revealed that structures with superior antibacterial activity were not necessarily those with the strongest positive charge and hydrophobicity in the library. Some candidates without hydrophilic building blocks also demonstrated excellent biocompatibility. This highlights the difficulty of deriving a comprehensive and universal structure-activity relationship solely through human intuition or experience, underscoring the significance of introducing machine learning in such research endeavors.

3.2. Model construction

Graph-based models have become the mainstream approach for deep learning with small molecules in recent years due to their ability to effectively capture the structural information and relationships within molecules [55]. However [56], polymers are not directly applicable to traditional graph learning methods due to their structural complexity. We employ a graph representation method for copolymers, which we proposed in our previous work [45]. As illustrated in Fig. 2a, each component of the copolymer is represented as a separate graph, with a global virtual node connecting to all nodes in each component, and three virtual nodes are introduced, connected to each of the three components, to specify the feed ratio of each component. Then, a graph transformer pre-trained using our previously established strategy [45,46] is employed to extract polymer features, followed by a readout operator. Finally, a projection head is utilized to map the graph-level representations h_G into property space. Additional details on model construction can be found in the Materials and methods section.

Fig. 2.

Machine learning module. a) Data preprocessing and model training. Copolymers are first transformed into graph representations using the RDKit and DGL packages. A graph transformer is employed as the encoder to extract structural features, followed by a readout operation to generate graph-level representations, which are then passed through a fully connected layer to predict antimicrobial and hemolytic activities. b) Model metrics on initial test set obtained by the trained RF, XGBoost, GPR and ours (RMSE, Top5-RMSE, Top20-RMSE and MAE are normalized for better visualization). c) The measured Log₂MIC and Log₂HC20 versus predicted value. (The zone between the dashed lines represents the predictions with absolute errors in the range of [-2, 2]). d) t-SNE visualization of the samples selected in each round. e) Density plots of measured Log₂MIC and Log₂HC₂₀ in each round.

To evaluate the effectiveness of our approach, we compared its performance with representative baseline models on the test set. For Random Forest (RF) and Extreme Gradient Boosting (XGB), the input features are constructed as the weighted sum of descriptors from each component [57]. For Gaussian process regression (GPR), copolymers are encoded as composition-weighted one-hot vectors, with the active monomer entries scaled by their respective ratios. As shown in Fig. 2b and c, the model constructed in this work exhibits the best performance. We further compared it with mainstream polymer language models, and the superiority of our approach remained consistent (Fig. S20). Therefore, the graph-based deep learning model was selected for subsequent screening.

After predictive model is constructed, we aim to identify polymers with the highest antibacterial activity and the lowest hemolytic properties across the entire chemical space. We formulate the problem as a pool-based multi-objective Bayesian optimization (PMOBO). In PMOBO, a query strategy is employed to select the next samples from the pool to be labeled. To balance the exploration and exploitation processes, both uncertainty and diversity are considered in the query strategy[46,[58], [59], [60]]. We use ensemble approach to estimate the polymer's performance μ and the model's uncertainty σ, and apply upper confidence bound (UCB) as the acquisition function to score the remaining samples in the pool. The samples with higher scores are then selected for K-Means clustering. The top-ranked samples obtained in each cluster are chosen as selected samples. Refer to Materials and methods section for more detailed settings. The selected samples in each round are synthesized and characterized through wet-lab experiments (Fig. S21–S28). These newly labeled data are then added back to the original dataset to refine the model, with the process repeated until convergence is reached. As shown in Fig. 2d, the search scope for antibacterial polymers has been narrowed from the blue region to the red region. As demonstrated in Fig. 2e, as expected, the initial randomly sampled dataset generally exhibits poor antibacterial activity and no significant hemolytic activity. With each iteration, the antibacterial activity of the selected samples significantly increases while maintaining low hemolytic activity. This result indicates that the optimization process has effectively improved the model's predictive accuracy, capturing important features and patterns related to antibacterial and hemolytic activities in polymers. By the fourth round, we concluded that the screening requirements had been met, with 103 polymers selected through the PMOBO process that exhibited high antibacterial activity and low hemolytic activity (MIC ≤16 μg/mL, HC₂₀ ≥ 2048 μg/mL). These 103 polymers were then subjected to further characterization.

3.3. Identification of the target polymer and in vitro antibacterial analysis

To further screen and identify the polymers with the best performance, 103 candidates with a MIC of no more than 16 μg/mL and a maximum polymer concentration causing 20 % hemolysis (HC₂₀) greater than 2048 μg/mL were selected for further MIC testing. We ultimately identified 7 candidates with an MIC no more than 8 μg/mL, which were then subjected to cytotoxicity testing, as demonstrated by the IC₂₀ (Fig. 3a and Fig. S29). Further, we selected two species of Gram-positive bacteria (Staphylococcus aureus, Staphylococcus epidermidis) and two species of Gram-negative bacteria (Pseudomonas aeruginosa, Proteus mirabilis) to verify their broad-spectrum antibacterial activity (Table S4). The chemical and biological characterization results of the 7 candidates are shown in Fig. 3b–f, with their distribution more clearly visualized in Fig. 3g. All the polymers (P1-P7) exhibited low MIC values (Fig. 3c). However, the MBC value of P1 was relatively high (Fig. 3d). P7 exhibited the lowest MIC and MBC. Meanwhile, all the polymers exhibited low hemolysis (Fig. 3e), while the cytotoxicity of P3 was relatively high (Fig. 3f). In short, P7 exhibited the highest IC₂₀/MIC ratio, and was defined as the target polymer, with its structure shown in Fig. 3h. Notably, the target polymer was further proved to be potent for not only E. coli but also Pseudomonas aeruginosa (PA) and methicillin-resistant Staphylococcus aureus (MRSA) (Fig. 3i).

Fig. 3.

Results of active learning-assisted antimicrobial polymer discovery. a) Schematic representation of the process for further characterization. b) Chemical properties of competitive candidates. c) The MIC of competitive candidates. d) The MBC of competitive candidates. e) The HC₂₀ of competitive candidates. f) The IC₂₀ of competitive candidates. g) Visualization of disparities in antimicrobial activity and biocompatibility among competitive candidates. h) Chemical structure of P7. i) Antimicrobial activity of the target polymer against P.A and MRSA (Unit: μg/mL).

Subsequently, we conducted a detailed in vitro characterization of P7. Phosphate-buffered saline (PBS) was used as the negative control, and ceftazidime, a commonly used clinical antibiotic was chosen as the positive control. The polymer at different concentrations was incubated with E. coli at an initial concentration of 10⁸ CFU/mL for 1 h at 37 °C. As shown in Fig. 4a, P7 at the concentration of 4-fold MIC achieved a 3-log reduction of E. coli. The K⁺ efflux under these conditions indicated that P7 exerted its bactericidal activity by disrupting the bacterial cell membrane (Fig. 4b). The bactericidal effect was corroborated by the colony forming unit counting assay on Mueller-Hinton Agar (MHA) (Fig. 4c). Further, scanning electron microscopy (SEM) (Fig. 4d) and transmission electron microscopy (TEM) (Fig. 4e) results demonstrated that the target polymer induced a compromise in the integrity of the cell membrane and the efflux of intracellular material, respectively. To visualize the rapid bactericidal effect, the incubation time was reduced to 10 min, followed by live-dead cell staining. As shown in Fig. 4f, the polymer-treated bacteria exhibited red fluorescence, indicating cell death, while the control group displayed green fluorescence, demonstrating excellent bactericidal activity of the target polymer. Additionally, the bacterial killing kinetics in Fig. 4g revealed that P7 killed 99.9 % of bacteria within 10 min, showing an effect comparable to ceftazidime. Additionally, Fig. 4h showed that the polymer exhibited no significant decrease in activity after 100 generations, unlike ceftazidime with serious resistance after only 40 generations, which demonstrated that compared to traditional antibiotics, P7 is less prone to develop drug resistance. Finally, Fig. 4i and j, showed that the target polymer caused rapid depolarization and significant disruption of the bacterial cytoplasmic membrane, proving its superior efficacy compared to traditional antibiotics in combating MDR bacteria. These results validated the in vitro bactericidal efficacy of the target polymer, and clarified that its primary bactericidal mechanism involves disrupting the permeability or even integrity of bacterial cell membranes. This disruption results in the leakage of intracellular materials, ultimately killing bacteria, which is in accordance with the universal bactericidal mechanism of AMPs[[61], [62], [63]]. The in vivo performance will be further investigated to provide a foundation for clinical translation.

Fig. 4.

Antimicrobial properties and mode of action of the target polymer in vitro. a) Standard plate counting assay of E. coli after 6 h of incubation with polymer at various concentrations. b) K⁺ leakage of E. coli after 6 h of incubation with polymer at various concentrations. c)E. coli on MHA medium after 6 h of incubation with polymer at 4-fold MIC. d) TEM characterization of E. coli cells before and after polymer treatment. (Scale bar: 500 nm) e) SEM characterization of E. coli cells before and after polymer treatment. (Scale bar: 500 nm) f) Live/dead bacteria stain assay of E. coli after 10 min of incubation with polymer at 2-fold MBC. (Scale bar: 50 μm) g) Killing kinetics of target polymer and ceftazidime against E. coli. h) Bacterial resistance development of target polymer and ceftazidime against E. coli. i) Cytoplasmic membrane depolarization induced by the target polymer. j) Outer cell membrane permeability capacity of the target polymer.

3.4. In vivo anti-infectious efficacy via Topical and intravenous administration

To assess the in vivo therapeutic effect of P7, we first constructed a skin wound model with E. coli infection (Fig. 5a). Saline and ceftazidime were used as negative and positive controls, respectively. A 1 cm wound was created on the backs of the mice, followed by the application of an E. coli suspension (10⁸ CFU/mL). The wounds were treated with the target polymer P7, antibiotic, or placebo on day 1 with the dose of 0.5 mg/kg. Photographs were taken on days 1, 3, 5, 7, and 9 (Fig. 5b) to observe wound changes across groups. The results showed that while all groups exhibited some degree of healing due to innate immunity, the polymer-treated and antibiotic-treated groups healed significantly faster than the saline group. On day 9, mice were euthanized and tissues from the wound site were analyzed histologically (Fig. 5c) and for bacterial count (Fig. 5d). In comparison to the saline group, which displayed more pronounced skin notching and inflammation responses, the tissues from the polymer-treated group recovered well, similar to the antibiotic group. No significant differences in body weight were observed among the groups (Fig. 5e). Wound area changes were measured (Fig. 5f), and the results quantitatively demonstrated the marked effect of the target polymers on healing (Fig. 5g).

Fig. 5.

In vivo E. coli -induced skin wound model. a) Schematic diagram of the process for subcutaneous abscess modeling. b) Wound photo after 1, 3,5, 7, 9 days of administration. c) H&E staining of collected skin tissues at day 7. (5X-bar: 320 μm, 10X-bar:160 μm, 20X-bar:80 μm) d) Colony counts after 7 days of administration. e) Weight changes during treatment. f) Digital photo of Wound Area Measurement. g) Wound area changes during treatment.

While local administration is commonly employed in treating wound infections, systemic administration is necessary for broader treatments. However, systemic administration often faces challenges such as drug metabolism, uneven distribution, and side effects. A subcutaneous abscess model is then established to evaluate the therapeutic efficacy of P7 via intravenous injection. E. coli suspension (10⁸ CFU/mL) was injected subcutaneously into the hind legs of mice to establish the subcutaneous abscess model (Fig. 6a). After 24 h of infection, P7, ceftazidime, and saline were injected via the tail vein, respectively. Specifically, the drug dose administered to each group was 0.5 mg/kg. Nine days later, the skin at the infection site was cut open for photographs (Fig. 6b–c). Significant abscesses were observed in saline-treated mice, whereas abscesses were largely eliminated in mice treated with the target polymer and antibiotics. Tissues were collected for histological analysis (Fig. 6d) and bacterial enumeration (Fig. 6e). Body weights of all groups decreased slightly in the first 2 days, but the administered groups regained weight by day 3, while the control group continued to decline (Fig. 6f). H&E staining (Fig. 6d), IL-6 (Fig. 6g) and IL-10 (Fig. 6h) levels showed that target polymers effectively relieved inflammation, comparable to antibiotics. Moreover, major organs (heart, liver, spleen, lungs, kidneys) showed no pathological changes, such as inflammation, necrosis, or fibrosis, indicating the good biological safety of the target polymer.

Fig. 6.

In vivo E. coli -induced subcutaneous abscess model. a) Schematic diagram of the process for subcutaneous abscess modeling. b) Thigh photo after 9 days of administration. c) Skin photo after 9 days of administration. d) H&E staining of collected subcutaneous tissues at day 9. (4X-bar: 400 μm, 10X-bar:160 μm, 20X-bar:80 μm) e) Colony counts after 9 days of administration. f) Weight changes during treatment. g) IL-6 level after 9 days of administration. h) IL-10 level after 9 days of administration. i) Systemic toxicity assessment of the target polymer. (10X-bar:160 μm, 20X-bar:80 μm).

4. Conclusion

In this study, we developed a novel framework that integrates combinatorial chemistry, automated high-throughput synthesis and characterization, and data-driven optimization for the screen of antimicrobial polymers with potent antimicrobial activity and excellent biocompatibility. The antimicrobial polymers were synthesized on an automated high-throughput synthesis platform iChemFoundry via PET-RAFT. By constructing an extensive polymer library and applying advanced machine learning algorithms, we efficiently enabled the targeted design of polymers with enhanced antimicrobial efficacy and minimized toxicity. Briefly, we constructed a seed dataset by extracting 400 candidates through random sampling and utilized the seed dataset to train the machine learning model. Specifically, a copolymer graph representation method was adopted to capture the structural information and relationships within molecules and construct the predictive model, after which a query strategy is employed to select the next samples to be labeled from the entire chemical space. After 4 rounds of Design-Build-Test-Learn cycle, we obtained 103 polymers with MIC <16 μg/mL, and HC₂₀ > 2048 μg/mL, which reached the best level of antimicrobial polymers reported in the literatures. Through further characterization, we acquired the target polymer P7 with a MIC of less than 2 μg/mL and an IC₂₀/MIC exceeding 100.The target polymer P7 exhibited excellent antibacterial activity and biocompatibility both in vitro and in vivo. This integrated approach significantly accelerates the discovery of antimicrobial agents, which enables screening 7 leading candidates from 13,768 combinations in only 12 days, offering a scalable and systematic strategy to tackle the challenge of multidrug resistance. The iterative process of active learning, combined with high-throughput synthesis and characterization, was effective in refining the dataset and selecting the best-performing candidates. The success of this framework highlights its potential to overcome the limitations of traditional trial-and-error methods in polymer research. Moreover, our work demonstrates how automation combined with machine learning can address specific challenges in biomedical research, presenting a promising approach to the rapid discovery of effective antimicrobial agents. We anticipate that this methodology will inspire further innovation and drive progress in combating multidrug-resistant bacterial infections.

CRediT authorship contribution statement

Tianyi Zhang: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Yuhui Wu: Writing – original draft, Software, Investigation, Conceptualization. Ye Tian: Writing – review & editing, Methodology. Youxiang Wang: Supervision, Resources. Peng Zhang: Software. Qiannuan Shi: Software, Resources. Qun Fang: Supervision, Resources. Jianzhang Pan: Visualization, Supervision. Qiao Jin: Writing – review & editing, Visualization, Supervision, Conceptualization. Jian Ji: Visualization, Supervision, Resources, Funding acquisition.

Ethics approval and consent to participate

All animal experiments were conducted in accordance with the guidelines issued by the Lab of Animal Experimental Ethical Inspection of Dr.Can Biotechnology (Zhejiang) Co., Ltd. (License: DRK-20240201112).

Declaration of competing interest

The authors declare no conflict of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: