A method for producing a dual mechanism antibacterial peptide

Abstract

The application provides a generation method of antibacterial peptides with dual action mechanisms, and belongs to the technical field of bioinformatics. The method comprises the following steps: constructing and training a peptide sequence generation model based on a variational autoencoder; pretraining and fine-tuning the model by using a large-scale peptide sequence dataset and a mechanism-related antibacterial peptide dataset in sequence, so that sequence features related to different antibacterial action mechanisms are mapped to a unified latent space; analyzing the latent space, identifying a latent space region in which the sequence features related to different mechanisms overlap, and sampling candidate antibacterial peptide sequences from the region; then, performing de-redundancy processing, and using multiple antibacterial peptide prediction models to predict antibacterial activity, so as to obtain candidate sequences of antibacterial peptides with dual action mechanisms. Experimental verification shows that the polypeptide sequences generated by the method can act on bacterial cell membranes and inhibit protein synthesis, have dual antibacterial action mechanisms, and provide a new technical means for the discovery of new antibacterial peptide lead molecules and the development of antibacterial drugs.

Description

Technical Field

[0001] This invention relates to a method for designing antimicrobial peptides with dual mechanisms of action, specifically a method for generating antimicrobial peptide sequences based on a deep learning model that simultaneously possess bacterial membrane disruption activity and protein synthesis inhibition activity, belonging to the field of bioinformatics. Background Technology

[0002] Antimicrobial peptides are a class of short peptides widely found in nature. Due to their broad-spectrum antimicrobial activity, rapid bactericidal ability, and relatively low risk of resistance induction, they are considered promising candidates to replace traditional small-molecule antibiotics. Compared to antimicrobial peptides that rely on only a single mechanism of action, dual-mechanism antimicrobial peptides, which simultaneously exhibit membrane disruption and intracellular process inhibition, can interfere with bacterial survival at different levels, thus potentially improving antimicrobial efficiency and reducing the likelihood of bacteria developing resistance through single-target alterations. Traditional dual-mechanism antimicrobial peptide design methods are mainly based on the combination design of known antimicrobial peptide fragments. While these methods can introduce different functional modules to some extent, they are highly dependent on known antimicrobial peptide fragments with well-defined mechanisms, and the splicing and combination methods are limited, resulting in a limited number of obtainable dual-mechanism sequences and a lack of scalability. In recent years, the development of artificial intelligence technology has accelerated the discovery and design of antimicrobial peptides. However, most existing methods mainly focus on antimicrobial activity and rarely explicitly consider the mechanism of action during the design process, especially the design of antimicrobial peptides with dual mechanisms of action, which still requires further exploration. The few existing studies involving mechanisms largely rely on classification or filtering after sequence generation to obtain mechanistic information. This post-processing screening makes it difficult to achieve mechanism-specific targeted guidance during the generation stage, thus affecting the generation efficiency and diversity of dual-action mechanism sequences. Therefore, current technology still lacks a method that can perform mechanism-guided design of different antimicrobial mechanisms during the generation stage and generate antimicrobial peptides with dual-action mechanisms. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a dual-mechanism antimicrobial peptide design method based on a deep generative model, used to obtain antimicrobial peptide candidate sequences with potential bacterial membrane disruption activity and protein synthesis inhibition activity.

[0004] like Figure 1 As shown, the present invention provides a dual-mechanism antimicrobial peptide generation method based on a deep generative model, comprising the following steps:

[0005] A. Construct and train a peptide sequence generation model based on a variational autoencoder, including:

[0006] A1. Construct a sequence dataset for model training, the sequence dataset including a large-scale peptide sequence dataset for model pre-training and a mechanism-related antimicrobial peptide dataset for model fine-tuning; wherein, the mechanism-related antimicrobial peptide dataset includes a membrane-cleavage-type antimicrobial peptide dataset and a proline-rich antimicrobial peptide dataset related to protein synthesis inhibition;

[0007] A2. Construct a peptide sequence generation model, wherein the peptide sequence generation model is a variational autoencoder model based on the Transformer architecture, including an encoder and a decoder;

[0008] A3. The peptide sequence generation model is pre-trained using the large-scale peptide sequence dataset described in step A1, enabling the model to learn the general composition rules and sequence representation features of peptide sequences; subsequently, the pre-trained peptide sequence generation model is fine-tuned using the mechanism-related antimicrobial peptide dataset described in step A1, so that the model obtains sequence feature representations related to different antimicrobial mechanisms, and maps the sequence features related to different antimicrobial mechanisms to a unified latent space;

[0009] B. Analyze the potential space of the peptide sequence generation model after fine-tuning in step A3, identify potential spatial regions where sequence features related to different antibacterial mechanisms overlap, and sample from the potential spatial regions to generate candidate antibacterial peptide sequences through the peptide sequence generation model after fine-tuning in step A3;

[0010] C. Screening the candidate antimicrobial peptide sequences generated in step B, specifically including:

[0011] C1. Redundancy removal is performed on the candidate antimicrobial peptide sequences generated in step B to remove repetitive or highly similar sequences and obtain representative candidate sequences;

[0012] C2. Use multiple antimicrobial peptide prediction models to predict the antimicrobial activity of representative candidate sequences obtained in step C1;

[0013] D. Based on the screening results of step C, select sequences with low sequence similarity and antimicrobial activity potential predicted by multiple antimicrobial peptide prediction models as candidate dual-action mechanism antimicrobial peptide sequences.

[0014] In step A1 above, preferably, peptide sequences of a certain length range (e.g., 10~51) are selected from the UniRef50 database, and redundancy is removed using CD-HIT with a 90% sequence similarity threshold to obtain the large-scale peptide sequence dataset for model pre-training. The membrane-cleavage antimicrobial peptide dataset is an antimicrobial peptide sequence dataset compiled based on reported experimental verification results of membrane cleavage activity; the proline-rich antimicrobial peptide dataset includes sequences explicitly reported as proline-rich antimicrobial peptides in the literature, sequences obtained by mutation based on known representative proline-rich antimicrobial peptide sequences, and potential proline-rich antimicrobial peptide sequences that conform to the characteristics of proline-rich antimicrobial peptides.

[0015] Preferably, the membrane-cleavage antimicrobial peptide dataset is deredundant using CD-HIT with a 90% sequence similarity threshold; the proline-rich antimicrobial peptide dataset is deredundant using CD-HIT with a 70% or 90% sequence similarity threshold depending on the sequence source, and after merging, it is deredundant again using CD-HIT with a 90% sequence similarity threshold to retain representative sequences.

[0016] In step A2 above, the variational autoencoder model based on the Transformer architecture converts amino acid residues in the peptide sequence into numerical indices and obtains a peptide sequence representation through embedding layers and position encoding. In one embodiment of the present invention, both the encoder and decoder consist of four Transformer layers.

[0017] In step A3 above, preferably, the large-scale peptide sequence dataset described in step A1 is randomly divided into a training set and a validation set in a 9:1 ratio for the pre-training stage, respectively used for model parameter training and performance evaluation. The loss function during training is:

[0018]

[0019] Among them, L rec For reconstruction loss, L KL Let λ be the KL divergence loss, and λ be the weight coefficient of the KL divergence loss term, used to balance the contributions of reconstruction loss and KL divergence loss during model training.

[0020] During model training, a linear KL annealing strategy is used to dynamically adjust the weight coefficient λ of the KL divergence loss term, so that the model training process gradually transitions from being dominated by reconstruction loss to a balance between reconstruction learning and latent space regularization.

[0021] In step B above, identifying overlapping regions and generating candidate antimicrobial peptide sequences specifically includes:

[0022] B1. Input the membrane-cleavage antimicrobial peptide sequence and the proline-rich antimicrobial peptide sequence into the fine-tuned peptide sequence generation model to obtain the corresponding latent representation vectors; perform principal component analysis on the latent representation vectors corresponding to the two types of antimicrobial peptides with different mechanisms to reduce their dimension to a two-dimensional latent space, and use kernel density estimation to fit the probability density distributions of the membrane-cleavage antimicrobial peptide and the proline-rich antimicrobial peptide in the two-dimensional latent space, respectively.

[0023] B2. Construct a grid in the two-dimensional latent space, calculate the comprehensive density score of the two types of antimicrobial peptides in the two-dimensional latent space, and determine the overlapping region of the characteristics of the two types of antimicrobial peptides with related mechanisms of action based on the comprehensive density score; the formula for calculating the comprehensive density score is:

[0024]

[0025] Where x represents a grid point in the two-dimensional potential space, P Mem-lytic AMP (x) represents the probability density of this grid point in the membrane cleavage-type antimicrobial peptide nuclear density estimation model, P PrAMP (x) represents the probability density of the grid point in the proline-rich antimicrobial peptide nuclear density estimation model; the higher the overall density score, the more likely the grid point is to be located in the overlapping region of the two types of mechanism-related antimicrobial peptide features.

[0026] In step C1 above, it is preferable to use the CD-HIT tool to perform redundancy removal on the candidate antimicrobial peptide sequences generated in step B with a 70% sequence similarity threshold, so as to retain representative antimicrobial peptide sequences.

[0027] Preferably, the antimicrobial peptide prediction models used in step C2 include AMP-READ, MBC-Attention, SVM-based antimicrobial peptide classification models, AMP-CLIP, and HyperAMP, which are used to comprehensively predict the antimicrobial activity of representative candidate sequences obtained in step C1.

[0028] Experimental data show that a series of candidate antimicrobial peptides obtained by the method of this invention exhibit a dual mechanism of action, capable of disrupting bacterial membranes and inhibiting bacterial protein synthesis, while also demonstrating good antimicrobial activity and low cytotoxicity. Therefore, this invention can be used for the design of antimicrobial peptides with dual mechanisms of action, providing a new technical means for the discovery of novel antimicrobial peptide lead molecules and the development of antimicrobial drugs. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method for generating antimicrobial peptides with a dual mechanism of action according to the present invention.

[0030] Figure 2 This refers to the distribution of membrane-cleaving antimicrobial peptides, proline-rich antimicrobial peptides, and generated sequences in the two-dimensional potential space in Example 2.

[0031] Figure 3 These are the PI uptake results of the candidate antimicrobial peptides in Example 3.

[0032] Figure 4 The results are from the DiSC3(5) uptake experiment of the candidate antimicrobial peptide in Example 3.

[0033] Figure 5 These are the results of the in vitro transcription and translation experiments of the candidate antimicrobial peptides in Example 3. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1: Construction and Training of a Peptide Sequence Generation Model Based on Variational Autoencoders

[0036] 1) Constructing a sequence dataset for model training. This dataset includes a large-scale peptide sequence dataset for model pre-training and a mechanism-related antimicrobial peptide dataset for model fine-tuning. The large-scale peptide sequence dataset for model pre-training was obtained from the UniRef50 database, from which peptide sequences of 10 to 51 amino acids were screened, yielding approximately 312,832 peptide sequences. These peptide sequences were then randomly divided into training and validation sets at a 9:1 ratio for model pre-training. The membrane-lytic antimicrobial peptide dataset was obtained by manually compiling antimicrobial peptide sequences with membrane lysis activity validation results from reported literature. These validation results included one or more of the following: NPN uptake experiments, liposome dye leakage experiments, scanning electron microscopy observation, and membrane integrity detection. During data processing, sequences of 10 to 51 amino acids containing only 20 natural amino acids were retained, while sequences with complex chemical modifications such as cyclization, fatty acid modification, or non-natural amino acids were removed. Redundancy was removed using the CD-HIT tool with a 90% threshold. Finally, 516 membrane-lytic antimicrobial peptide sequences were obtained. The proline-rich antimicrobial peptide dataset includes sequences explicitly reported in the literature as proline-rich antimicrobial peptides, and sequences based on the known representative proline-rich antimicrobial peptide sequence Bac7. 1-23 Sequences obtained through deep mutation scanning, as well as potential sequences matching the characteristics of proline-rich antimicrobial peptides, were analyzed. For sequences obtained through deep mutation scanning, redundancy was removed using the CD-HIT tool with a 70% threshold; for sequences reported in the literature and potential proline-rich antimicrobial peptide sequences, redundancy was removed with a 90% threshold. These three types of sequences were then merged and subjected to a uniform redundancy removal process again using a 90% threshold, ultimately yielding 2,356 proline-rich antimicrobial peptide sequences.

[0037] 2) A peptide sequence generation model is constructed, which is a variational autoencoder model based on the Transformer architecture, including an encoder and a decoder. The input peptide sequence is first converted into an amino acid index sequence by a sequence preprocessing module. The index sequence includes indices corresponding to 20 standard amino acid residues, and includes amidation modification markers, sequence start marker SOS, sequence stop marker EOS, and zero-padding marker PAD. After zero-padding, the input peptide sequence is standardized to a preset maximum length. Subsequently, the index sequence is mapped into a 128-dimensional continuous vector through an embedding layer, and position encoding is added to preserve the order information of amino acid residues in the peptide sequence. The continuous vector is input into the encoder to obtain sequence features. The encoder output is processed by mean pooling, and then passed through two fully connected layers to obtain the mean μ and logarithmic variance logσ of the latent variables. 2 The latent vector z is obtained by sampling using a reparameterization method. In the decoding stage, the sampled latent vector z is first expanded into an input representation matching the sequence length through a fully connected layer, and positional encoding is added. The decoder predicts the residues at the next position position position by position based on the start marker and the generated amino acid residues, and obtains the amino acid probability distribution at each position through the output layer, ultimately generating the target polypeptide sequence.

[0038] 3) The variational autoencoder model is pre-trained using the pre-trained dataset to enable the model to learn the general composition rules and sequence representation features of peptide sequences; subsequently, the pre-trained model is fine-tuned using the membrane cleavage antimicrobial peptide dataset and the proline-rich antimicrobial peptide dataset to enable the model to obtain sequence feature representations related to different antimicrobial mechanisms.

[0039] The model training loss function includes reconstruction loss and KL divergence loss, and the total loss function is:

[0040]

[0041] Among them, L rec For reconstruction loss, L KL Let λ be the KL divergence loss, and λ be the weight coefficient of the KL divergence loss term, used to balance the contributions of reconstruction loss and KL divergence loss during model training.

[0042] During model training, a linear KL annealing strategy is used to dynamically adjust the weight coefficient λ of the KL divergence loss term, achieving a balanced transition from reconstruction-dominated to potential regularization.

[0043]

[0044] Where λ(e) is the weight coefficient of the KL divergence loss term in the e-th round, λ min λ is the minimum value of the weight coefficient of the KL divergence loss term. maxis the maximum value of the weight coefficient of the KL divergence loss term, and n is the total number of training rounds.

[0045] The reconstruction loss L rec The accuracy of the decoder's reconstruction of an input polypeptide sequence after it has been represented by a latent vector is measured by the following formula:

[0046]

[0047] Where x=(x1, x2, …, x t ) represents the input amino acid sequence, x <t Let z represent all amino acid residues preceding the t-th position, and z be derived from the posterior distribution. The latent vector obtained by sampling in the middle, This represents the conditional probability predicted by the model for the t-th residue; Represents the posterior distribution The mathematical expectation; T is the length of the input peptide sequence.

[0048] The KL divergence loss L KL The latent variable distribution used to constrain the encoder output approximates a standard normal prior distribution, and its calculation formula is as follows:

[0049]

[0050] Where, μ i and σ i 2 Let be the mean and variance of the i-th dimension of the latent space, respectively, and d be the dimension of the latent space (i.e., the dimension of the latent vector z).

[0051] During the training phase, the Adam optimizer was used to optimize the model, with 200 training iterations (epochs) selected, and the optimal model was chosen based on the validation set loss. In the pre-training phase, the learning rate was set to 0.0005, the weight decay coefficient to 0.000001, and the batch size to 256. During training, a linear KL annealing strategy was used, gradually increasing the weight λ of the KL divergence term from 0 to 0.0001. In the fine-tuning phase, the learning rate was reduced to 0.000001, the batch size was set to 32, and the remaining hyperparameters remained consistent with those in the pre-training phase.

[0052] Example 2: Generation of antimicrobial peptides with a dual mechanism of action

[0053] 1) Antimicrobial peptide sequences from the membrane-cleavage-type antimicrobial peptide dataset and the proline-rich antimicrobial peptide dataset are input into the encoder of the fine-tuned peptide sequence generation model to obtain latent representation vectors corresponding to the two types of antimicrobial peptide sequences. Subsequently, principal component analysis is performed on the latent representation vectors to reduce their dimensionality to a two-dimensional latent space, and kernel density estimation is used to fit the distribution density of membrane-cleavage-type antimicrobial peptides and proline-rich antimicrobial peptides in the two-dimensional latent space, respectively. A grid is constructed in the two-dimensional latent space, and the comprehensive density score corresponding to each grid point is calculated. The formula for calculating the comprehensive density score is:

[0054]

[0055] Where x represents a grid point in the two-dimensional potential space, P Mem-lytic AMP (x) represents the probability density of this grid point in the membrane cleavage-type antimicrobial peptide nuclear density estimation model, P PrAMP (x) represents the probability density of the grid point in the proline-rich antimicrobial peptide nuclear density estimation model.

[0056] 2) The grid points were sorted from highest to lowest based on their comprehensive density scores, and the top 400 grid points with the highest scores were selected as the overlapping region of the antimicrobial peptide features related to the two mechanisms. Subsequently, principal component analysis inverse transformation was performed on these grid points to obtain a high-dimensional latent vector, which was then input into the decoder of the fine-tuned model to generate 236 sequences. The distribution of membrane-cleaving antimicrobial peptides, proline-rich antimicrobial peptides, and the generated sequences in the two-dimensional latent space is shown below. Figure 2 As shown.

[0057] 3) A two-step screening process was used to filter the generated sequences. First, the CD-HIT tool was used to remove redundancy with a 70% threshold, resulting in 75 non-redundant sequences. Then, five antimicrobial peptide prediction models were used: MBC-Attention, AMP-READ, HyperAMP, an SVM-based antimicrobial peptide classification model, and AMP-CLIP. Specifically, sequences ranking in the top 50 of all three prediction models (MBC-Attention, AMP-READ, and HyperAMP) were retained, while sequences predicted as non-antimicrobial peptides by the SVM-based antimicrobial peptide classification model or AMP-CLIP were excluded, ultimately yielding 30 candidate antimicrobial peptide sequences for subsequent experimental validation. The candidate antimicrobial peptide sequences are shown in Table 1.

[0058]

[0059] Example 3: Experimental Validation of 30 Candidate Antimicrobial Peptides

[0060] (a) Minimum inhibitory concentrations of 30 candidate antimicrobial peptides

[0061] Antimicrobial activity was evaluated using the minimum inhibitory concentration (MIC), which was determined by the microbroth dilution method. The tested strains included two Gram-positive strains (Bacillus subtilis ATCC6633 and Staphylococcus aureus ATCC6538) and three Gram-negative strains (Pseudomonas aeruginosa ATCC9027, Escherichia coli ATCC25922, and Acinetobacter baumannii ATCC19606). Single colonies of each strain were picked and inoculated into LB medium, and cultured overnight with shaking at 37 °C and 220 rpm. Subsequently, the bacterial culture in the logarithmic growth phase was collected, and the OD value of the bacterial culture was measured. 625 Adjust the pH to 0.08–0.13, then dilute 100 times with LB medium to prepare a bacterial suspension for MIC determination.

[0062] The candidate antimicrobial peptides were synthesized by Wuhan Dangang Biotechnology Co., Ltd. using solid-phase chemical synthesis. Mass spectrometry analysis showed that the measured molecular weights of each candidate antimicrobial peptide were consistent with the theoretical molecular weights. High-performance liquid chromatography analysis showed that the purity of each candidate antimicrobial peptide was higher than 95%.

[0063] Weigh 1 mg of the candidate antimicrobial peptide and dissolve it in the corresponding volume of sterile PBS buffer (pH 7.4) to prepare a 512 μM stock solution. Add 100 μL of ILB medium to each well of a sterile 96-well polypropylene plate (Greiner, #655201). Add 100 μL of the peptide stock solution to the first column of rows A, B, and C, respectively. After thorough mixing, transfer 100 μL to the next column of wells, performing serial dilutions in the same manner. Discard 100 μL of liquid from the last column of wells to maintain a final volume of 100 μL in each well. Then, add 100 μL of bacterial suspension to each well, bringing the final volume to 200 μL. Each treatment was performed in triplicate. The experiment also included sterile control wells and bacterial growth control wells: sterile control wells contained culture medium but no bacterial suspension, while bacterial growth control wells contained both culture medium and bacterial suspension but no candidate antimicrobial peptide. The 96-well plates were incubated at 37 ℃ for 16–18 h, and bacterial growth in each well was observed visually. The concentration at which no obvious bacterial growth was observed was defined as the minimum inhibitory concentration (MIC) of the candidate antimicrobial peptide. The results are shown in Table 2.

[0064]

[0065] If a candidate sequence exhibits a minimum inhibitory concentration (MIC) of no more than 128 μM against at least one test strain, it is deemed to possess antimicrobial activity. Of the 30 candidate antimicrobial peptides, 20 showed antimicrobial activity, a success rate of 66.7%. Specifically, Seq71 and Seq135 showed antimicrobial activity against all five test strains, and Seq136 inhibited the growth of four of the test strains, indicating that these sequences possess good broad-spectrum antimicrobial activity. These results demonstrate that the method described in this invention can effectively generate candidate polypeptide sequences with antimicrobial activity.

[0066] (II) Verification of the membrane cleavage activity of highly active candidate antimicrobial peptides

[0067] Based on the results of minimum inhibitory concentration (MIC) determination, *Staphylococcus aureus* and *Acinetobacter baumannii* were selected as representative strains, and nine candidate antimicrobial peptides with MICs less than or equal to 32 μM against their respective strains were selected for mechanism of action studies. The effects of the candidate antimicrobial peptides on bacterial cell membrane integrity and permeability were detected using a propidium iodide (PI) fluorescent probe. Single colonies of *Acinetobacter baumannii* and *Staphylococcus aureus* were picked, inoculated into LB broth, and cultured with shaking at 37 ℃ and 220 rpm until the logarithmic growth phase. The bacterial cells were collected, centrifuged to remove the culture medium, washed three times with PBS buffer (pH 7.2-7.4), and then resuspended in PBS buffer. The OD of the bacterial suspension was then measured. 600 Adjust to 1.0. In a sterile black polypropylene 96-well plate (Greiner, #655209), mix 10 μL of bacterial suspension with 10 μL of candidate antimicrobial peptide solution to achieve a final concentration of the candidate antimicrobial peptide in a 20 μL incubation system that is 4 times its minimum inhibitory concentration (MIC). Incubate at 37 ℃ for 1 h. In a sterile black polypropylene 96-well plate (Greiner, #655209), mix 10 μL of bacterial suspension with 10 μL of candidate antimicrobial peptide solution to achieve a final concentration of the candidate antimicrobial peptide in a 20 μL incubation system that is 4 times its MIC against the corresponding strain. Incubate at 37 ℃ for 1 h. After incubation, add 80 μL of PI working solution to each reaction system, mix thoroughly, and incubate at 37 ℃ in the dark for 30 min. Positive and negative control groups were also included in the experiment. In the positive control group, 10 μL of bacterial suspension was mixed with 10 μL of polymyxin B solution; in the negative control group, 10 μL of bacterial suspension was mixed with 10 μL of PBS buffer. Each experimental group and control group had three parallel wells. Fluorescence intensity was measured using a Cytation 5 multi-mode microplate reader (BioTek, USA), with excitation and emission wavelengths set to 535 nm and 615 nm, respectively. Figure 3As shown, compared with the negative control group, the PI fluorescence intensity of Acinetobacter baumannii and Staphylococcus aureus increased to varying degrees after treatment with most candidate antimicrobial peptides, indicating that the candidate antimicrobial peptides could disrupt bacterial cell membrane integrity and increase cell membrane permeability to varying degrees. Among them, candidate antimicrobial peptides such as Seq8, Seq71, Seq129, Seq135, Seq136, and Seq199 showed a more significant effect on enhancing membrane permeability in Acinetobacter baumannii. For Staphylococcus aureus, the PI fluorescence intensity of most candidate antimicrobial peptides was higher than that of the negative control group, but the PI fluorescence intensity of the Seq133 treatment group was close to that of the negative control group, indicating that Seq133 had a weak effect on Staphylococcus aureus cell membrane permeability under the experimental conditions.

[0068] To further evaluate the effect of candidate antimicrobial peptides on bacterial cell membrane potential, membrane depolarization was detected using the membrane potential-sensitive fluorescent dye 3,3'-dipropylthiocyanine iodide [DiSC3(5)]. Single colonies of Acinetobacter baumannii and Staphylococcus aureus were picked and inoculated into LB liquid medium, and cultured with shaking at 37 ℃ and 220 rpm until the logarithmic growth phase. After collecting the bacterial cells and removing the medium by centrifugation, the cells were washed three times with 5 mM HEPES buffer (pH 7.2-7.4) containing 20 mM glucose and 0.1 M KCl. The cells were then resuspended in the same buffer, and the OD of the bacterial suspension was measured. 600 Adjust to 1.0. Add DiSC3(5) to the bacterial suspension to a final concentration of 15 μM, and incubate at 37 °C in the dark for 30 min to allow the dye to reach equilibrium inside and outside the bacterial cell membrane. Subsequently, in a sterile black polypropylene 96-well plate (Greiner, #655209), mix 67 μL of the DiSC3(5)-labeled bacterial suspension with 33 μL of the candidate antimicrobial peptide solution to make the final concentration of the candidate antimicrobial peptide in a 100 μL reaction system 4 times its minimum inhibitory concentration against the corresponding strain, and incubate at 37 °C for 1 h.

[0069] The experiment included both a positive control group and a negative control group. In the positive control group, 67 μL of DiSC3(5)-labeled bacterial suspension was mixed with 33 μL of polymyxin B solution; in the negative control group, 67 μL of DiSC3(5)-labeled bacterial suspension was mixed with 33 μL of HEPES buffer. The total volume and subsequent incubation conditions for each group were the same as those for the candidate antimicrobial peptide treatment group. Three parallel wells were set for each experimental group and control group. Fluorescence intensity was measured using a Cytation 5 multi-mode microplate reader (BioTek, USA), with excitation and emission wavelengths set to 622 nm and 670 nm, respectively. Figure 4As shown, all tested candidate antimicrobial peptides induced varying degrees of cell membrane depolarization in Acinetobacter baumannii and Staphylococcus aureus. Specifically, seven candidate antimicrobial peptides exhibited stronger membrane depolarization against Staphylococcus aureus than the polymyxin B positive control, and eight candidate antimicrobial peptides exhibited stronger membrane depolarization against Acinetobacter baumannii than the polymyxin B positive control.

[0070] Based on the combined results of the PI uptake experiment and the DiSC3(5) membrane depolarization experiment, it can be seen that for Staphylococcus aureus, except for Seq133, the other 8 candidate antimicrobial peptides can all disrupt the bacterial cell membrane; although Seq133 can cause cell membrane depolarization, it does not significantly damage the cell membrane integrity. For Acinetobacter baumannii, all 9 candidate antimicrobial peptides tested can disrupt the cell membrane and cause membrane depolarization.

[0071] (III) Verification of the protein synthesis inhibition effect of highly active candidate antimicrobial peptides

[0072] To evaluate the effects of candidate antimicrobial peptides on bacterial protein synthesis, a cell-free in vitro transcription-translation (IVTT) system was used. The IVTT system was the PURExpress system (New England Biolabs), and a plasmid carrying the firefly luciferase gene was used as a reporter template. The inhibitory effect of the candidate antimicrobial peptides on protein synthesis was evaluated by measuring luciferase expression levels. Reaction systems were prepared according to the kit instructions, with each reaction system containing 10 μL of 100 ng of firefly luciferase plasmid and the candidate antimicrobial peptide, resulting in final concentrations of 10 μM or 100 μM for the candidate antimicrobial peptide. A template-free negative control group and a peptide-free positive control group were included. The template-free negative control group did not contain the plasmid template, while the peptide-free positive control group contained the plasmid template but not the candidate antimicrobial peptide. Polymyxin B was used as a mechanism control; polymyxin B can disrupt bacterial cell membranes but does not typically directly inhibit protein synthesis. All reaction systems were incubated at 37 °C for 1 h. After incubation, kanamycin (50 mg / mL) was added to terminate the translation reaction, and the reaction system was incubated on ice for 5 min. Subsequently, 4 μL of the reaction solution from each system was transferred to a white 96-well plate, and 40 μL of luciferase substrate solution (Promega, #E1501) was added. The chemiluminescence intensity of each well was measured using a Cytation 5 multi-plate reader (BioTek, USA). Both the experimental and control groups had three parallel wells. The formula for calculating the relative luminescence intensity is: (L... peptide -L negative ) / (L positive -L negative ) × 100%. Where, Lpeptide L represents the luminescence intensity of the current concentration of candidate antimicrobial peptides treated group. negative L represents the luminescence intensity of the template-free negative control group. positive This represents the luminescence intensity of the peptide-free positive control group. For example... Figure 5 As shown, polymyxin B did not significantly reduce the expression level of firefly luciferase under either 10 μM or 100 μM conditions, indicating that it did not exhibit significant protein synthesis inhibition under the experimental conditions, consistent with its mechanism of exerting antibacterial activity primarily through disruption of the bacterial cell membrane. In contrast, all nine candidate antimicrobial peptides tested reduced the relative luminescence intensity of firefly luciferase to varying degrees under 100 μM conditions, indicating their ability to inhibit in vitro protein synthesis; under 10 μM conditions, five of the candidate antimicrobial peptides showed significant protein synthesis inhibition. In particular, treatment with Seq8, Seq39, Seq71, and Seq161 reduced the relative luminescence intensity of firefly luciferase to below 50%, indicating that these candidate antimicrobial peptides possess strong protein synthesis inhibition capabilities.

[0073] Experimental results show that Seq8, Seq39, Seq71, Seq129, Seq133, Seq135, Seq136, Seq161, and Seq199 can all act on bacterial cell membranes and inhibit protein synthesis, exhibiting a dual antibacterial mechanism and belonging to dual-mechanism antimicrobial peptides.

Claims

1. A method for generating an antimicrobial peptide with a dual mechanism of action, comprising the following steps: A. Construct and train a peptide sequence generation model based on a variational autoencoder, including: A1. Construct a sequence dataset for model training, including a large-scale peptide sequence dataset for model pre-training and a mechanism-related antimicrobial peptide dataset for model fine-tuning; wherein, the mechanism-related antimicrobial peptide dataset includes a membrane-cleavage-related antimicrobial peptide dataset and a proline-rich antimicrobial peptide dataset related to protein synthesis inhibition. A2. Construct a peptide sequence generation model, wherein the peptide sequence generation model is a variational autoencoder model based on the Transformer architecture, including an encoder and a decoder; A3. The peptide sequence generation model is pre-trained using the large-scale peptide sequence dataset described in step A1, enabling the model to learn the general composition rules and sequence representation features of peptide sequences; subsequently, the pre-trained peptide sequence generation model is fine-tuned using the mechanism-related antimicrobial peptide dataset described in step A1, so that the model obtains sequence feature representations related to different antimicrobial mechanisms, and maps the sequence features related to different antimicrobial mechanisms to a unified latent space; B. Input the membrane cleavage-type antimicrobial peptide dataset and the proline-rich antimicrobial peptide dataset into the peptide sequence generation model fine-tuned in step A3, identify potential spatial regions where sequence features related to different antimicrobial mechanisms overlap, and sample from the potential spatial regions to generate candidate antimicrobial peptide sequences through the peptide sequence generation model fine-tuned in step A3; C. Screening the candidate antimicrobial peptide sequences generated in step B, specifically including: C1. Redundancy removal is performed on the candidate antimicrobial peptide sequences generated in step B to remove repetitive or highly similar sequences and obtain representative candidate sequences; C2. Use multiple antimicrobial peptide prediction models to predict the antimicrobial activity of representative candidate sequences obtained in step C1; D. Based on the screening results of step C, select sequences with low sequence similarity and antimicrobial activity potential predicted by multiple antimicrobial peptide prediction models as candidate dual-action mechanism antimicrobial peptide sequences.

2. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 1, characterized in that, In step A1, peptide sequences of a certain length range are screened from the UniRef50 database and deredundant is processed using CD-HIT with a 90% sequence similarity threshold to obtain the large-scale peptide sequence dataset. The membrane cleavage-type antimicrobial peptide dataset is an antimicrobial peptide sequence dataset compiled based on reported membrane cleavage activity experimental verification results, and deredundant is processed using CD-HIT with a 90% sequence similarity threshold. The proline-rich antimicrobial peptide dataset includes sequences explicitly reported as proline-rich antimicrobial peptides in the literature, sequences obtained by mutation based on known representative proline-rich antimicrobial peptide sequences, and potential proline-rich antimicrobial peptide sequences that meet the characteristics of proline-rich antimicrobial peptides. These sequences are deredundant using CD-HIT with a 70% or 90% sequence similarity threshold, and after merging, are deredundant again using CD-HIT with a 90% sequence similarity threshold.

3. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 1, characterized in that, In step A2, the variational autoencoder model based on the Transformer architecture converts the amino acid residues in the peptide sequence into corresponding digital indices, and obtains the peptide sequence representation through embedding layers and position encoding.

4. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 3, characterized in that, In step A2, the input polypeptide sequence is first converted into an amino acid index sequence by a sequence preprocessing module. The index sequence includes indices corresponding to 20 standard amino acid residues, and includes amidation modification markers, sequence start marker SOS, sequence stop marker EOS, and zero-padding marker PAD. After zero-padding, the input polypeptide sequence is standardized to a preset maximum length. Subsequently, the index sequence is mapped into a 128-dimensional continuous vector through an embedding layer, and position encoding is added to preserve the sequence information of amino acid residues in the polypeptide sequence. This continuous vector is then input into an encoder to obtain sequence features. The encoder output is processed by mean pooling and then passed through two fully connected layers to obtain the mean μ and log-variance logσ of the latent variables. 2 The latent vector z is obtained by sampling through a reparameterization method. In the decoding stage, the sampled latent vector z is first expanded into an input representation that matches the sequence length through a fully connected layer, and position encoding is added. The decoder predicts the next position of residues based on the start marker and the generated amino acid residues, obtains the amino acid probability distribution at each position through the output layer, and finally generates the target polypeptide sequence.

5. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 1, characterized in that, The loss function in step A3 of the training process is: Among them, L rec For reconstruction loss, L KL λ is the KL divergence loss, and λ is the weight coefficient of the KL divergence loss term, which is used to balance the contributions of reconstruction loss and KL divergence loss in the model training process. A linear KL annealing strategy is adopted to dynamically adjust the weight coefficient λ of the KL divergence loss term, so that the model training process gradually transitions from reconstruction loss to a balance between reconstruction learning and latent space regularization.

6. The method for generating the dual-mechanism antimicrobial peptide as described in claim 5, characterized in that, In step A3, during model training, the weight coefficients of the KL divergence loss term are: Where λ(e) is the weight coefficient of the KL divergence loss term in the e-th round, λ min λ is the minimum value of the weight coefficient of the KL divergence loss term. max is the maximum value of the weight coefficients of the KL divergence loss term, and n is the total number of training rounds; Reconstruction loss L rec The calculation formula is: Where x=(x1, x2, …, x t ) represents the input amino acid sequence, x <t Let z represent all amino acid residues preceding the t-th position, and z be derived from the posterior distribution. The latent vector obtained by sampling in the middle, This represents the conditional probability predicted by the model for the t-th residue; Represents the posterior distribution The mathematical expectation; T is the length of the input peptide sequence; KL divergence loss L KL The calculation formula is: Where, μ i and σ i 2 Let be the mean and variance of the i-th dimension of the latent space, respectively, and d be the dimension of the latent space.

7. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 1, characterized in that, Step B includes: B1. Input the membrane-cleavage antimicrobial peptide sequence and the proline-rich antimicrobial peptide sequence into the fine-tuned peptide sequence generation model to obtain the corresponding latent representation vectors; use principal component analysis to reduce the dimensionality of the latent representation vectors to a two-dimensional latent space, and obtain the probability density distributions of the two types of antimicrobial peptides by kernel density estimation. B2. Construct a grid in the two-dimensional potential space, calculate the comprehensive density score of each grid point according to the following formula, and determine the overlapping region of the two types of antimicrobial peptides based on the comprehensive density score: Where x represents a grid point in the two-dimensional potential space, P Mem-lytic AMP (x) represents the probability density of this grid point in the membrane cleavage-type antimicrobial peptide nuclear density estimation model, P PrAMP (x) represents the probability density of the grid point in the proline-rich antimicrobial peptide nuclear density estimation model.

8. The method for generating the dual-mechanism antimicrobial peptide as described in claim 7, characterized in that, In step B, the top 400 grid points in terms of comprehensive density score are selected, and principal component analysis inverse transformation is performed on the grid points to obtain a high-dimensional latent representation vector. The high-dimensional latent representation vector is then input into the decoder of the fine-tuned peptide sequence generation model to generate candidate antimicrobial peptide sequences.

9. The method for generating antimicrobial peptides with a dual mechanism of action as described in claim 1, characterized in that, Step C1 uses CD-HIT with a 70% threshold to remove duplicate or highly similar sequences from the candidate antimicrobial peptide sequences generated in step B, obtaining representative candidate sequences.

10. The method for generating the dual-mechanism antimicrobial peptide as described in claim 1, characterized in that, The antimicrobial peptide prediction models used in step C2 include AMP-READ, MBC-Attention, SVM-based antimicrobial peptide classification models, AMP-CLIP, and HyperAMP.

Images