Machine Learning-Based Discovery of Antimicrobial Peptides and Their Antibacterial Activity Against Staphylococcus aureus

Abstract

The escalating crisis of antibiotic resistance, particularly concerning foodborne pathogens such as Staphylococcus aureus and its biofilm contamination, has emerged as a major global challenge to food safety and public health. Biofilm formation significantly enhances the pathogen’s resistance to environmental stresses and disinfectants, underscoring the urgent need for novel antimicrobial agents. In this study, we isolated Bacillus strain B673 from the saline–alkali environment of Xinjiang, conducted whole-genome sequencing, and applied antiSMASH analysis to identify ribosomally synthesized and post-translationally modified peptide (RiPP) gene clusters. By integrating an LSTM-Attention-BERT deep learning framework, we screened and predicted nine novel antimicrobial peptide sequences. Using a SUMO-tag fusion tandem strategy, we achieved efficient soluble expression in an E. coli system, and the purified products exhibited remarkable inhibitory activity against Staphylococcus aureus (MIC = 3.13 μg/mL), with inhibition zones larger than those of the positive control. Molecular docking and dynamic simulations demonstrated that the peptides can stably bind to MurE, a key enzyme in cell wall synthesis, with negative binding free energy, suggesting an antibacterial mechanism via MurE inhibition. This study provides promising candidate molecules for the development of anti-drug-resistant agents and establishes an integrated research framework for antimicrobial peptides, spanning gene mining, intelligent screening, efficient expression, and mechanistic elucidation.

1. Introduction

The abuse of antibiotics has led to a resistance crisis that is severely undermining their therapeutic efficacy [1]. Among foodborne pathogens, Staphylococcus aureus poses a much greater threat than ordinary spoilage microorganisms [2]. This bacterium is widely distributed and can easily contaminate high-protein foods through personnel, equipment, or raw materials, proliferating rapidly under improper storage conditions. Its greatest hazard lies in its ability to adhere to food or processing equipment surfaces, secrete extracellular matrix, and form biofilms [3]. Such biofilm structures greatly enhance bacterial resistance to conventional disinfectants and thermal treatments, making them difficult to eradicate and thus serving as persistent contamination sources in food processing environments. In addition to providing a protective niche for bacteria, biofilms may also facilitate the dissemination of antimicrobial resistance [4]. Consequently, the biofilm problem of S. aureus has become a serious hygienic challenge in the food industry, posing a significant threat to food safety and potentially causing substantial economic losses [5]. To address this challenge, antimicrobial peptides (AMPs) have attracted considerable attention as novel weapons due to their unique membrane-targeting physical bactericidal mechanisms [6], which make it difficult for bacteria to develop resistance through traditional mutational pathways. AMPs are characterized by abundant positive charges, strong hydrophobicity, and amphipathic structures, whereas microbial membrane components such as phospholipids, teichoic acids, and lipopolysaccharides carry abundant negative charges. This charge complementarity results in strong electrostatic attraction between AMPs and microbial cell membranes, enabling AMPs to attach to the bacterial surface. The hydrophobic face of the peptide interacts weakly with membrane phospholipids to form peptide–lipid supramolecular complexes and insert into the bilayer [7], while the hydrophilic face interacts with lipids to form transmembrane channels. This ultimately leads to leakage of cytoplasmic contents, disruption of homeostasis, and bacterial death. Furthermore, certain AMPs can inhibit the expression of bacterial surface adhesion proteins, thereby fundamentally reducing colonization capacity and effectively preventing or eliminating persistent biofilm-associated infections [8].

However, traditional AMP screening relies heavily on labor-intensive and time-consuming experimental validation, which makes it difficult to address the urgent threat posed by drug-resistant S. aureus [9]. Huang et al. [10] isolated and identified a novel AMP, Dermaseptin-pH, from the skin secretions of captive-bred South American orange-legged leaf frogs. Notably, the frogs required at least three months of acclimation in an artificial environment prior to experimentation, which significantly increased both time and resource costs. Mine et al. [11] isolated, purified, and characterized a class of novel AMPs from enzymatic hydrolysates of egg white lysozyme. The hydrolysates were generated by sequential digestion with pepsin and trypsin, yielding more than 20 low-molecular-weight peptides (<1000 Da), which greatly complicated the separation and identification of bioactive peptides. By contrast, Santos-Júnior et al. [12] proposed a machine learning-based AMP prediction strategy and applied it to global microbiome datasets, achieving high predictive accuracy and offering a new avenue for the efficient discovery of novel AMPs.

Traditional approaches to AMP production also suffer from significant limitations. In plant extraction, AMPs are typically present in host tissues at extremely low abundance and are prone to forming complexes with endogenous components such as polysaccharides, phenolics, and proteins, resulting in complicated extraction procedures, low yields, and poor purification efficiency. For example, Cammue et al. [13] isolated two AMPs from Mirabilis jalapa seeds, with minimum inhibitory concentrations ranging from 6 to 300 μg/mL, but the peptides still suffered from insufficient purity. In chemical synthesis, the efficiency of solid-phase synthesis decreases sharply as peptide chain length increases, while accumulation of side reactions further reduces final yields. Moreover, this process requires large amounts of costly protected amino acids, activating reagents, resins, and organic solvents [14], leading to dramatically increased costs and severely limiting scalability and clinical translation. By contrast, synthetic biology provides a breakthrough pathway for the efficient production of AMPs. Tian et al. [15] successfully achieved heterologous AMP production using an Escherichia coli recombinant expression system. This method not only significantly reduced costs but also offered excellent potential for process scale-up, making it one of the most promising approaches for large-scale AMP production. However, AMPs often display strong hydrophobicity and high cationic charge [16], physicochemical properties that tend to promote misfolding and aggregation during heterologous expression, leading to the formation of inactive inclusion bodies and loss of bioactivity. In addition, AMPs expressed individually frequently suffer from poor stability and low solubility, which remain key technical challenges in recombinant expression strategies [17].

In this study, we analyzed the whole-genome sequence of the dominant strain Bacillus B673 isolated from saline–alkali soils in Xinjiang and applied a deep learning framework integrating LSTM, attention, and BERT neural networks for functional prediction and scoring. By uncovering sequence–activity relationships, this approach enabled the efficient screening of potential antimicrobial peptides (AMPs). Leveraging machine learning to parse large-scale genomic datasets accelerated the translation from gene clusters to novel AMPs, offering a computation-driven strategy to address antimicrobial resistance. The identified AMPs were further expressed in E. coli chassis cells using tandem peptide expression combined with fusion tags [18], which improved the stability and solubility of AMPs during production. This strategy provides a sustainable solution for the large-scale supply of highly active AMPs.

2. Materials and Methods

2.1. Screening of Antimicrobial Peptides Derived from Bacillus

The Bacillus strain, preserved in the laboratory and originally isolated from the saline–alkali soils of Xinjiang, was streaked onto LB solid medium and incubated at 37 °C for 12–16 h. A single colony was then inoculated into 5 mL of LB broth and cultured with shaking at 37 °C for 12 h. The fermentation broth was centrifuged at 5000 rpm for 10 min, and the supernatant was collected and sterilized by filtration through a 0.22 μm membrane for further use. An S. aureus suspension (OD600 = 0.1) was evenly spread on the surface of LB agar plates. Wells were prepared using a sterile hole punch, and 50 μL of the test samples were added to each well. A silver-based antimicrobial agent (Microarmor^® Micro-KS124A, Huizhou, China) and sterile water were used as the positive and negative controls, respectively. Plates were incubated at 37 °C for 18–24 h, after which the inhibition zones were observed and recorded.

2.2. Whole Genome Sequencing of B673

The Bacillus strain preserved at −80 °C was streaked onto LB agar plates for activation and incubated at 37 °C for 12 h. A single colony was then inoculated into LB broth and cultured with shaking at 30 °C for 24–48 h. Genomic DNA was extracted from approximately 2 × 10⁶ cells using a modified CTAB protocol, supplemented with lysozyme treatment (10 mg/mL, 37 °C for 30 min) to enhance the lysis of Gram-positive cell walls. Residual RNA was removed by digestion with RNase A (DNase-free), and DNA quality and concentration were assessed using a Qubit 3.0 fluorometer (Thermo Fisher Scientific, Waltham, MA, USA) and agarose gel electrophoresis [19]. Whole-genome sequencing and analysis were performed by BGI-Shenzhen (Shenzhen, China).

2.3. Using Antismash to Mine Antimicrobial Peptide Sequences

In this study, secondary metabolite biosynthetic gene clusters (BGCs) in Bacillus strain B673 were systematically identified using the antiSMASH platform (v5.0) with default parameter settings [20]. Considering potential fragmentation of contigs in the genome assembly, which may result in incomplete BGC structures, short BGCs containing single genes (1.79%) or two genes (3.73%) were retained in the analysis to avoid potential bias. Subsequently, BiG-SLiCE (v1.1.0) was employed (parameters:—threshold 99,999—complete) to cluster homologous or highly similar BGCs [21], thereby constructing gene cluster families (GCFs). To evaluate the novelty of these clusters, the cosine distance between each BGC and the BIG-FAM database was calculated. For each BGC, the minimum distance value was used to characterize its relatedness to the database, while the mean distance across all members was used to define the threshold of GCF–database association. GCFs with an average distance greater than 2.0 were classified as novel families. Furthermore, based on the presence/absence distribution matrix of GCFs, metabolic diversity of dominant taxa was assessed using the R package iNEXT (v2.0.20). For AMP screening, core peptide sequences predicted from ribosomally synthesized and post-translationally modified peptide (RiPP) BGCs were annotated using a deep learning framework integrating LSTM, attention, and BERT models [22]. The dataset for model training consisted of X sequences, balanced for AMP-positive and AMP-negative examples. The dataset was split into 70% for training and 30% for testing, and 5-fold cross-validation was applied to assess model robustness and generalizability. The LSTM module was responsible for capturing long-range dependencies in amino acid sequences and identifying potential conserved motifs and domains. The attention mechanism assigned weights to different residue positions, enhancing the model’s ability to resolve functionally critical sites and improving interpretability. The BERT model, leveraging large-scale pre-trained language representations, performed deep semantic encoding of core peptide sequences, enabling the identification of atypical AMPs that are difficult to detect through conventional homology-based methods and providing quantitative prediction scores. The model’s performance was evaluated using accuracy, AUC, precision, and recall. We compared our deep learning approach to baseline models based on traditional homology-based methods to assess improvements in AMP prediction. These results were crucial for determining the model’s effectiveness and reliability.

2.4. Heterologous Expression of Antimicrobial Peptides

A tandem expression cassette containing the pelB signal peptide sequence and codon-optimized antimicrobial peptides (AMP2, AMP4, and AMP7), linked via flexible linkers (GGGGS), was constructed in the pET-28a(+) vector. The recombinant plasmid, designated 28a-742, was fully synthesized and sequence-verified by TsingKe Biological Technology Co., Ltd (Nanjing, China). Linearized vector fragments were obtained by PCR amplification of plasmid 28a-742 using primer pair 742-z-sumo1/742-z-sumo2, while the SUMO tag fragment was amplified from plasmid pET-28a-sumo using primer pair 742-p-sumo1/742-p-sumo2. The two purified fragments were ligated using 24317 DNA ligase to generate a SUMO tag fusion expression plasmid with an enterokinase cleavage site. The recombinant plasmid was introduced into E. coli BL21(DE3) competent cells by chemical transformation and plated on LB agar supplemented with kanamycin (50 μg/mL) for positive clone selection. Single colonies were verified by PCR amplification with T7 promoter and T7 terminator primers, followed by agarose gel electrophoresis. Correct clones were inoculated into LB broth containing kanamycin (50 μg/mL) and stored as glycerol stocks. All strains used in this study are provided in Table S1, all plasmids are listed in Table S2, and all PCR primers are listed in Table S3.

For expression, recombinant strains were cultured at 37 °C until reaching an OD600 of 0.6–0.8, followed by induction with 0.5 mM IPTG at 16 °C for 12 h. Cells were harvested, lysed, and centrifuged, and the supernatant was sterilized by filtration through a 0.22 μm membrane. Recombinant peptides were purified using Ni-NTA affinity chromatography. Secretory expression of AMPs was verified by Tricine-SDS-PAGE (BBI Life Sciences Corporation, Shanghai, China). For SDS-PAGE, 20 μL of test sample was mixed with 4× SDS-PAGE loading buffer, loaded, and electrophoresed at 120 V until the dye front reached the bottom of the gel. Gels were stained with Coomassie Brilliant Blue R-250 for 3 h, destained for 3 h in a solution of ultrapure water:glacial acetic acid:95% ethanol (5:0.5:4.5), and further destained overnight with 7% (v/v) acetic acid before gel imaging.

2.5. Mass Spectrometry Method

A Vanquish UHPLC system coupled to a Q-Exactive Plus Orbitrap high resolution tandem mass spectrometer (Thermo Fisher, Waltham, MA, USA) equipped with a Waters ACQUITY UPLC Peptide CSH C18 column maintained at 40 °C (150 × 2.1 mm, 1.7 μm particle size) was used for analysis of synthetic peptide. The mobile phase A (0.1% formic acid in water) and mobile phase B (0.1% formic acid in acetonitrile) were used and the flow rate was 0.3 mL/min. The gradient elution program was as follows: 0.0–1.0 min: 97% A; 1.0–10.0 min: 97–70% A; 10.0–13.0 min: 70–45% A; 13.0–13.1 min: 45–20% A; 13.1–16.0 min: 20% A; 16.0–16.1 min: 97% A; 16.1–20.0 min: 97% A, and the injection volume was 10 µL. MS data acquisition was performed in the full scan MS/ddMS2 mode. The resolutions of full scan MS and ddMS2 were set at 70,000 and 17,500, respectively. The scan range in full scan MS settings was 300 to 1800 m/z, and the automatic gain control (AGC) target and maximum injection time i were 3 × 10⁶ and 100 ms, while their values were 2 × 10⁵ and 50 ms in ddMS2 settings. The TopN was 5, isolation window was 1.6 m/z, and dynamic exclusion was 10 s. Electrospray ionization source was operating in positive mode and the parameters were as follows: auxiliary gas flow 10 arb, sheath gas flow 35 arb, spray voltage 3.5 kV, capillary temperature 320 °C, probe heater temperature 350 °C, and S-lens 60. Raw data was analyzed using Qual Browser module of Thermo Xcalibur software (4.2.47).

2.6. Antimicrobial Activity Assay

The heterologous expression supernatant was freeze-dried and concentrated, and 300 mg of the freeze-dried powder was re-dissolved in sterile water. The solution was then purified using a Ni-NTA affinity chromatography column. The antimicrobial activity of the purified product was measured using the antimicrobial assay method described in Section 2.1.

2.7. Minimum Inhibitory Concentration (MIC) Determination

A S. aureus suspension (approximately 1 × 10⁸ CFU/mL) was mixed with LB medium and dispensed into a 96-well plate (final volume per well: 200 μL, containing 95 μL LB, 5 μL bacterial suspension, and 100 μL antimicrobial peptide solution at gradient concentrations). Positive controls (silver ion antimicrobial agent + LB) and negative controls (sterile water + LB) were also included. After incubating at 37 °C for 18–24 h, the optical density at 600 nm (OD600) was measured using a microplate reader (Thermo Fisher Scientific). The growth inhibition threshold was set at 10% of the OD value of the negative control. The minimum antimicrobial peptide concentration that completely inhibited bacterial growth was determined as the MIC. Each experiment was performed in triplicate to ensure data reliability.

2.8. Molecular Docking Method

To investigate the binding mode between MurE (PDB: 4C12) and the candidate antimicrobial peptide, protein–protein docking was performed using the HDOCK Server. The crystal structure of MurE was obtained from the RCSB PDB database, with non-essential solvent molecules and ligands removed prior to docking. The three-dimensional structure of candidate peptide 1 was generated using AlphaFold3. Docking was carried out with HDOCK employing a template-based and template-free hybrid docking strategy, and scoring was performed using the built-in knowledge-based scoring function. The top-ranked model (Cluster 1, HDOCK score = −276.5, rank #1) was selected as the initial complex for subsequent molecular dynamics simulations. Structural analysis was performed using PyMOL (2.6.0) and Discovery Studio (2019), focusing on hydrogen bonding, salt bridges, hydrophobic interactions, and π–π/π–cation interactions [23].

2.9. Molecular Dynamics (MD) Simulation Method

To evaluate binding stability and conformational adaptation, all-atom molecular dynamics (MD) simulations were performed using GROMACS 2022. The AMBER ff14SB force field was applied for both the protein and peptide, with the TIP3P water model. The complex was solvated in a dodecahedral box with a minimum buffer distance of 1.2 nm and neutralized with 0.15 M NaCl. Energy minimization was conducted using the steepest descent algorithm, followed by equilibration under NVT conditions (500 ps at 303.15 K) and NPT conditions (1 ns at 1 bar). Production simulations were run for 100 ns with three independent trajectories, using a 2 fs time step and saving snapshots every 10 ps. Trajectory analyses included root mean square deviation (RMSD), root mean square fluctuation (RMSF), and radius of gyration (Rg). In addition, an RMSD–Rg free energy landscape (FEL) was constructed to characterize conformational states. Binding free energy was calculated using gmx_MMPBSA v1.6.4 with the MM/GBSA method, based on 400 snapshots extracted from the 60–100 ns interval. Interaction entropy corrections were applied to a subset of frames to refine the estimation.

3. Results

3.1. Screening of Dominant Bacillus Strains

The laboratory currently maintains a collection of Bacillus strains isolated from saline–alkali environments in Xinjiang, with species distribution shown in Figure 1. Bacillus marinus was the predominant group, comprising 33 isolates (21.9% of the total), followed by Bacillus endophyticus (21 isolates, 13.9%) and Bacillus megaterium (16 isolates, 10.6%). In addition, a few strains of Bacillus vallismortis and Bacillus altitudinis were also preserved. To evaluate the antimicrobial potential of these strains, their fermentation broths were tested against S. aureus. Inhibition stronger than that of the positive control was defined as “prominent activity,” weaker inhibition than the positive control as “moderate activity,” and responses comparable to the negative control as “no activity.” The results showed that 14.6% of the fermentation broths exhibited significant inhibitory activity against S. aureus (Figure 2A). Based on inhibition zone diameters (Figure 2B), strain B673 was identified as the most effective isolate, producing an inhibition zone of 1.9 cm. This strain was selected as the focal subject for subsequent genomic analysis and antimicrobial peptide discovery.

3.2. Genome Analysis and Functional Annotation

As shown in Figure 3, based on COG functional classification analysis, Bacillus B673 demonstrates multidimensional potential in antimicrobial peptide biosynthesis and secretion mechanisms. Firstly, the gene abundance in the functional category “Biosynthesis, transport, and catabolism of secondary metabolites” (229 genes) is significant, suggesting that its genome may encode a large number of secondary metabolism gene clusters, including non-ribosomal peptide synthetase (NRPS) or genes related to antimicrobial peptide precursor synthesis. Secondly, the gene enrichment in “Intracellular trafficking, secretion, and vesicular transport” (132 genes) and “Post-translational modification, protein turnover, and chaperones” (201 genes) indicates that this strain has an efficient secretion system and post-translational modification capacity, supporting the transmembrane transport and functional maturation of antimicrobial peptides. Additionally, the high expression of genes related to “Amino acid transport and metabolism” (314 genes) may provide abundant amino acid precursors for antimicrobial peptide synthesis. Although genes related to “Defense mechanisms” (2 genes) are fewer, their synergistic effect with secondary metabolism pathways may still contribute to the stress response regulation of antimicrobial peptides. In summary, the COG functional characteristics suggest that Bacillus B673 has a potential molecular basis for antimicrobial peptide synthesis, modification, and secretion.

As shown in Figure 4, based on KEGG Pathway classification analysis, Bacillus B673 shows significant potential in antimicrobial peptide biosynthesis and regulatory mechanisms. The strain exhibited a high gene abundance in secondary metabolism-related pathways, where “Biosynthesis of other secondary metabolites” (40 genes) and “Terpenoid and polyketide metabolism” (51 genes) may be involved in the synthesis of non-ribosomal peptides (NRPS) or polyketide (PKS) antimicrobial peptides, which often encode complex compounds with antimicrobial activity. Additionally, the high proportion of “Global metabolism” (735 genes) suggested that its metabolic network is highly complex and may contain secondary metabolism gene clusters that have not been fully annotated, which requires further analysis using genomic mining tools such as antiSMASH. Regarding resistance mechanisms, the enrichment of genes in “Antibiotic resistance” (36 genes) and “Membrane transport systems” (186 genes) indicated that the strain may regulate the release of antimicrobial peptides through active secretion or efflux pumps, and its efficient transmembrane transport capacity provides a structural foundation for the functional activity of antimicrobial peptides. The complexity of the signal regulation network (135 genes) suggests that it may dynamically regulate the expression of antimicrobial peptides through quorum sensing or environmental stress responses. In conclusion, the metabolic features and resistance mechanisms of B673 provide a theoretical basis for the discovery of its antimicrobial peptides.

3.3. Antimicrobial Peptide Sequence Mining

In the process of antimicrobial peptide mining, the core peptide sequences predicted from ribosomally synthesized and post-translationally modified peptides (RiPPs) and their biosynthetic gene clusters (BGCs) can be accurately predicted for their functional activity using a deep learning framework that integrates Long Short-Term Memory (LSTM) networks, Attention mechanisms, and BERT neural network structures. This framework provides precise predictions and outputs corresponding antimicrobial probability scores. The specific sequences and scores are shown in

Table 1. Antimicrobial peptide sequences and their corresponding scores.

Name	Sequence	Attention	LSTM	Bert
AMP1	GCATCSIGAACLVDGPIPDFEIAGATGLFGLWG	0.99960017	0.99167371	0.83647186
AMP2	WKSESVCTPGCVTGLLQTCFLQTITCNCKISK	0.99960017	0.99167371	0.985059
AMP3	CTTNTFSLSDYWGNKGGWCTVSKECMAWC	0.99987185	0.99997115	0.98849875
AMP4	TNDAYSKSLANRAGLGNNYGKYCTVSAECFGTISCGS	0.98993313	0.99998486	0.97553307
AMP5	EGSIYTVSHECHMNTWQFVFTCCF	0.99998939	0.9999969	0.98947376
AMP6	GCATCSIGAVCLVDGPIPDFEIAGATGLFGLWG	0.95716596	0.99844491	0.79907644
AMP7	WKSESLCTPGCVTGALQTCFLQTLTCNCKISK	0.70067245	0.99876297	0.9830245
AMP8	TITLSTCAILSKPLGNNGYLCTVTKECMPSCN	0.99987185	0.99997115	0.99422085
AMP9	AFFSCCFSAC	0.99839532	0.99988484	0.98212665

. It is noteworthy that all the antimicrobial peptide sequences mined in this study are novel sequences not found in the antimicrobial peptide database.

3.4. Heterologous Expression and Purification

Through preliminary experiments, we found that expressing individual antimicrobial peptide fragments was challenging. Antimicrobial peptides typically possess strong hydrophobic and cationic properties, which can lead to the formation of inclusion bodies or insoluble aggregates in expression systems such as E. coli, resulting in a loss of activity. In our preliminary experiments, we attempted to express the individual peptides, but the expression results were not satisfactory. Despite our efforts, we were unable to obtain corresponding bands in either the SDS-PAGE gel or Western blot. The SUMO tag, being a highly soluble protein tag, can significantly improve the solubility of antimicrobial peptides when fused to them, promoting their proper folding in a soluble form and reducing inclusion body formation. At the same time, the presence of the SUMO tag can shield the antimicrobial activity of the peptide, thereby reducing its toxicity to the host in the fused state, which in turn enhances cell density and recombinant protein yield. Therefore, this study employed a SUMO tag with an enterokinase cleavage site to connect it to the target antimicrobial peptide sequence. This strategy leverages the SUMO tag to enhance the solubility and expression stability of the fusion protein while using the efficient and specific cleavage ability of thrombin to release antimicrobial peptides with their native sequence in vitro as needed, providing a reliable and controllable material foundation for subsequent activity evaluation and functional studies.

The tandem expression strategy significantly enhances the antimicrobial performance by linking multiple copies of the antimicrobial peptide coding sequence in head-to-tail fashion. This strategy directly improves its efficiency in disrupting biofilms; at the same time, the tandem peptides can bind to multiple pathogen surface molecules simultaneously, generating a synergistic antimicrobial effect that substantially increases the speed and breadth of bacterial killing. In our experiments, we found that different peptide sequence combinations led to varying antimicrobial effects. Among the different combinations, we successfully identified the optimal tandem arrangement for AMP2, AMP4, and AMP7. Among the tested sequences (AMP2, AMP4, and AMP7), the AMP742 sequence demonstrated superior antibacterial efficacy, exhibiting the lowest minimal inhibitory concentration (MIC). Therefore, the AMP742 sequence was chosen as the template for all subsequent investigations. The corresponding MIC values are presented in Table S4. As illustrated in the SDS-PAGE analysis of the SUMO-fused concatenated peptides (Figure S1), all constructs (SUMO-AMP742, SUMO-AMP724, SUMO-AMP472, SUMO-AMP247, SUMO-AMP274, and SUMO-AMP427) showed distinct bands at approximately 24.47 kDa, consistent with their expected molecular weights when fused to the SUMO tag. Following cleavage of the SUMO tag, Tricine-SDS-PAGE analysis (Figure 5) confirmed the successful release and purification of the core peptides. In particular, the AMP742 peptide (theoretical MW: 9.87 kDa) displayed a clear band at the expected position, validating the effective expression and purification of all engineered concatenated variants.

The synthetic peptide AMP742 was successfully identified in fermentation broth by high-resolution LC-MS/MS analysis. Chromatographic data (Figure S2) showed a prominent peak at 13.9 min corresponding to the expected m/z value. MS analysis under positive ionization mode (70,000 resolution) confirmed the presence of AMP742, with ddMS2 fragmentation patterns verifying the peptide identity. Tryptic digest validation demonstrated complete sequence coverage through four key peptides: WTPALSVITGYISSNTCPTTACTR, CVDCTTNTFSLSDYWGNK, GGWCTVSK, and DHMVLHEYVNAAGITEAAAK. These peptides showed high-accuracy precursor ions and characteristic fragment ions (e.g., 1468.65/1355.57, 1431.65/1317.61) whose fragmentation patterns matched theoretical sequences. When mapped to the full-length construct, these peptides formed continuous coverage from N- to C-terminus. The detailed mass spectrometry data are presented in Table S6.

3.5. Antimicrobial Effect

The induced E. coli expressing the recombinant AMP742 protein was collected, lysed, and the supernatant was collected by centrifugation. The supernatant was then sterilized by filtration through a 0.22 μm filter and purified using a Ni-column. The purified liquid was freeze-dried and concentrated for antimicrobial assays. As shown in Figure 6, the antimicrobial effect of the expressed antimicrobial peptide exceeded that of the positive control, silver ion antimicrobial agent.

3.6. Minimum Inhibitory Concentration (MIC)

As shown in Figure 7, the antimicrobial activity of the recombinant antimicrobial peptide AMP742 was evaluated in a 96-well plate. The results indicated that when the antimicrobial peptide concentration was close to 3.13 μg/mL, a slight inhibitory effect on S. aureus was observed, manifested as a significant reduction in bacterial growth within the wells, although complete inhibition was not achieved. However, when the concentration was lower than this value, the antimicrobial peptide showed no significant inhibitory effect, and the bacterial growth was similar to the negative control group. According to the Clinical and Laboratory Standards Institute (CLSI) recommended guidelines, the minimum concentration that completely inhibits visible bacterial growth is determined as the minimum inhibitory concentration (MIC). Therefore, in this study, the MIC of the antimicrobial peptide against S. aureus was determined to be 3.13 μg/mL.

3.7. Molecular Docking

The cytoplasmic enzyme MurE is essential in Staphylococcus aureus, catalyzing a key reaction in the synthesis of peptidoglycan precursors for cell wall biogenesis. There have been numerous studies on the murD enzyme of E. coli, which is similar to the murE we are researching. Its significance lies in the fact that inhibiting MurE will directly lead to bacterial death due to the inability to construct the cell wall, making it a well-recognized and highly efficient antimicrobial target [24]. Of particular importance is that MurE is unique to bacteria and has no homologs in the human host, which means that antimicrobial peptides designed to target it are expected to have very high selectivity [25], thus minimizing potential toxicity to human cells and ensuring drug safety.Initial screening demonstrated potent activity of our antimicrobial peptides against Gram-positive bacteria, but only moderate efficacy against Gram-negative strains. This divergent activity may be caused by the fundamental differences in cell wall structure between these bacteria: the thick, exposed peptidoglycan layer in Gram-positive bacteria renders MurE, a key enzyme in its synthesis, highly accessible and a vulnerable target. In contrast, the enzyme is shielded within the periplasmic space of Gram-negative bacteria, protected by a formidable outer membrane. It is this critical distinction in target accessibility, rather than the mere presence or absence of MurE, that guided our target selection. Therefore, MurE ligase was selected as a primary target for investigation [26], and we employed well-established computational simulation strategies [27] to test the hypothesis that AMP742 acts as a direct inhibitor.

AMP742 was selected as the subject for molecular docking experiments. HDOCK docking results showed that among the top 100 docking conformations, the best model achieved an HDOCK score of −276.5, suggesting a strong binding propensity between MurE and the peptide. As shown in Figure 8, interaction analysis using Discovery Studio revealed multiple stabilizing contacts: hydrogen bonds were formed between AMP742 and residues Glu213, Thr247, and Asn319, while Lys255 and Lys348 established salt bridges with the negatively charged groups of the peptide. Additional hydrophobic and aromatic interactions were identified, including contacts contributed by Phe181, Leu217, and Tyr256, with Tyr256 engaging in π–π stacking with the peptide’s aromatic ring. The MM/GBSA binding free energy decomposition is presented in Table S4 and further supports the favorable binding of AMP742 to MurE. The MM/GBSA binding free energy decomposition is shown in Table S4.

3.8. Molecular Dynamics (MD) Simulation

Using AMP742 as the subject of molecular dynamics simulation. Molecular dynamics simulations characterized the AMP742-MurE complex’s stability and binding properties. As shown in Figure 9, RMSF analysis indicated substantial flexibility in the N-terminal region (residues 1–30) and surface loops, with fluctuations reaching 1.4–1.6 nm. In contrast, the core binding pocket maintained structural rigidity, exhibiting minimal fluctuations between 0.08–0.20 nm. It can be seen in Figure 10, the free energy landscape revealed a deep energy basin centered at RMSD 1.5–2.0 nm and Rg 2.8–3.0 nm, with the global energy minimum reaching 6.6 kJ/mol. This low-energy basin demonstrates the complex adopted a stable, compact conformation. These computational analyses collectively support AMP742 forming a stable complex with MurE, consistent with its proposed mechanism inhibiting this essential cell wall synthesis enzyme.

4. Discussion

As an important natural source of antimicrobial peptides (AMPs), Bacillus species produce a wide variety of ribosomally and non-ribosomally synthesized AMPs that exhibit broad-spectrum activity, favorable stability, and diverse mechanisms of action [28]. In recent years, they have become a hotspot for antimicrobial resource discovery [29]. Compared with previous studies, our work demonstrates distinct advantages in both screening strategies and molecular characteristics. For instance, Jangra et al. [30] identified the lasso peptide Lariocidin (LAR) from Bacillus via genome mining, with their innovation primarily focusing on the in-depth mechanistic dissection of a single lead compound. By contrast, the present study is dedicated to establishing an integrated and efficient AMP discovery platform. This system combines genome mining, intelligent screening based on the LSTM–Attention–BERT model, high-efficiency heterologous expression, and preliminary mechanistic validation. It thus provides a reproducible and broadly applicable framework for the sustainable and large-scale discovery of novel AMPs, rather than being confined to the exploration of single molecules. This systematic approach offers broader applicability and platform value in addressing the challenges of antimicrobial resistance. Similarly, Yang et al. [31] identified the AMP DB16 from Bacillus cereus using a co-cultivation induction and bioinformatics-guided screening strategy. Although DB16 exhibited inhibitory activity against S. aureus, its potency and expression efficiency remained suboptimal, with an MIC of 62.5 μg/mL. In contrast, the synthetic biology-based strategy adopted in our study enabled the heterologous expression and rational engineering of Bacillus-derived AMPs, resulting in engineered peptides with substantially lower MIC values (3.13 μg/mL) against multiple drug-resistant pathogens. Moreover, the approach enhanced the solubility and stability of heterologous expression products, providing a stronger foundation for the future development of AMPs as practical antimicrobial agents.

Machine learning, by analyzing genomic sequences and functional datasets [32], enables efficient prediction and screening of novel peptides with antimicrobial activity, significantly improving the discovery efficiency of candidate peptides and providing new strategies to combat drug-resistant pathogens [33]. In the field of antimicrobial peptide (AMP) discovery, the integration of LSTM, attention mechanisms, and BERT models has shown considerable promise. For example, Huang et al. [34] developed a machine learning workflow capable of screening trillions of sequences from a virtual peptide library composed of 6–9 amino acid residues to identify effective AMPs. This workflow consisted of multiple trainable modules for empirical selection, classification, ranking, and regression, following a coarse-to-fine design to progressively narrow the search space. However, this approach was limited by peptide length, and thus was unable to capture both longer peptides (e.g., AMP1–8) and shorter peptides (e.g., AMP9) as achieved in this study. In another study, Ma et al. [22] integrated multiple natural language processing (NLP) models, including LSTM, attention, and BERT, into a unified pipeline to identify candidate AMPs from the human gut microbiome. Among 2349 candidate sequences identified, 216 were chemically synthesized, of which 181 displayed antimicrobial activity (positive rate > 83%). Despite its remarkable predictive power, this validation strategy based on chemical synthesis was extremely costly, making large-scale experimental verification of candidate molecules impractical and limiting its potential for high-throughput applications. By contrast, the synthetic biology-based heterologous expression strategy employed in this study substantially reduced validation costs, providing a more economical and efficient approach for large-scale functional screening of candidate AMPs. While computational prediction offers a powerful guide for AMP discovery, functional validation through wet-lab experiments remains indispensable to confirm bioactivity and elucidate mechanisms of action, thereby completing the research cycle from “sequence prediction” to “functional confirmation.” The integrated computational–experimental strategy established in this study not only demonstrated better adaptability to peptide length and structural diversity at the sequence mining stage but also provided greater feasibility and scalability during the validation stage, highlighting its potential as a robust framework for large-scale AMP discovery.

In the research and production of antimicrobial peptides (AMPs), synthetic biology has demonstrated significant technological advantages and application potential, providing a systematic solution to the bottlenecks associated with traditional extraction and chemical synthesis methods [35]. In this study, we successfully achieved efficient and soluble heterologous expression of novel AMPs derived from the genome of Bacillus strain B673 using an E. coli expression system, highlighting the core value of synthetic biology strategies for large-scale AMP production. Traditional approaches, such as plant extraction, are often limited by the low abundance of AMPs in host tissues and their coexistence with endogenous compounds such as polysaccharides and phenolics, resulting in poor separation and purification efficiency. For example, Tang et al. [36] reported that the AMP content in plant extracts was considerably lower than that of other metabolites, requiring additional concentration steps that made the process complex and yield-limiting. Chemical synthesis, in turn, is hampered by the low efficiency of solid-phase synthesis for long peptide sequences, accumulation of side reactions, and high reagent costs, restricting large-scale production. Synthetic biology provides an effective means to overcome these limitations. Nevertheless, challenges remain. AMPs often form inclusion bodies or exert cytotoxicity on the host. Xu et al. [37], for instance, found that recombinant cecropins expressed in E. coli were largely deposited as inclusion bodies, which alleviated host toxicity to some extent but increased the complexity of refolding and purification—an issue also observed in our early experiments. In this study, we applied a SUMO-fusion expression strategy, which markedly improved peptide solubility, reduced inclusion body formation, and alleviated host toxicity, thereby enhancing yield and stability of the recombinant peptides. Fusion strategies represent another promising avenue for functional improvement. Zhao et al. [38] linked the protein DAMP4 with the AMP pexiganan via an acid-labile cleavage site, improving peptide stability and antibacterial activity. Building upon this, we rationally designed a tandem expression construct combining AMP2, AMP4, and AMP7, which exhibited significantly enhanced inhibitory activity against S. aureus. Notably, different tandem combinations produced distinct impacts on expression efficiency and antimicrobial potency, suggesting that factors such as linker design, sequence arrangement, and spatial conformation play critical roles in function [39]. Future optimization strategies—such as linker engineering, fine-tuning of induction conditions, and co-expression of molecular chaperones—may further improve soluble expression [40]. Although the recombinant AMPs developed in this study demonstrated excellent antibacterial activity (MIC = 3.13 μg/mL), their large-scale production process, in vivo stability, and biosafety still require systematic evaluation to advance their translational potential as practical antimicrobial agents.

Molecular docking and molecular dynamics (MD) simulations revealed a strong binding affinity between the target antimicrobial peptide and the MurE enzyme (HDOCK score = −276.5), with the interaction stabilized by a combination of hydrogen bonds, salt bridges, and hydrophobic contacts. This finding is consistent with the report by Kumar et al. [41], who proposed MurE as a potential drug target for evaluating peptide–receptor interactions. MD simulations further confirmed the structural stability of the complex over a 100 ns timescale, with RMSD and Rg values indicating equilibrium in the later stages of the simulation and suggesting a rigid binding conformation. MM/GBSA calculations yielded negative binding free energy values, demonstrating the thermodynamic spontaneity of the binding process. Our approach is directly supported by the work of Singh et al., who established MurE as a viable target and provided a computational framework for identifying inhibitors, a paradigm we apply to test AMP742 [26]. Together, these results provide atomic-level insights into the mechanism by which the antimicrobial peptide may exert its activity through inhibition of MurE, a key enzyme in bacterial cell wall synthesis, thereby establishing a structural basis for further mechanistic studies.

5. Conclusions

In recent years, many studies have explored enzymes [42] and bioactive substances [43] from fungi in Xinjiang. In this study, by integrating genomic analysis, machine learning-based prediction, heterologous expression, and molecular simulations, we successfully identified nine novel peptide candidates with potential antimicrobial activity from Bacillus strain B673 isolated from the saline–alkali environment of Xinjiang. Using a SUMO-tag fusion and tandem expression strategy, soluble and efficient expression of the AMPs was achieved in an E. coli system. The purified products exhibited strong inhibitory activity against S. aureus, with a minimum inhibitory concentration (MIC) as low as 3.13 μg/mL. Molecular docking and dynamics simulations suggested that the target AMPs may exert their antimicrobial activity by binding efficiently to MurE, a key enzyme in bacterial cell wall synthesis, thereby disrupting normal bacterial physiology. This work not only provides new candidate molecules for combating drug-resistant bacterial infections but also establishes an integrated research strategy that spans genome mining, machine learning, synthetic biology-based expression, and mechanistic exploration, laying both theoretical and technical foundations for the efficient development and application of AMPs.