A method for intelligent optimization design of cecropin AD by combining VAE sequence amplification, AI evolution and molecular simulation

Abstract

The application discloses an intelligent optimization design method for cecropin AD by fusing VAE sequence amplification, AI evolution and molecular simulation. The method takes cecropin AD as an initial template sequence, uses a variational autoencoder to modify the initial template sequence, and constructs an initial sequence library with the same length as the template and limited mutation sites; then AI evolution is performed to generate candidate mutant sequences, AMPainter-1 and AMP-potency-prediction two prediction models are used for cross scoring to obtain preliminary screening candidate sequences; the preliminary screening candidate sequences are constructed into a gram-negative bacterial membrane system and MD simulation is performed, and the membrane binding force and potential antibacterial activity of the candidate peptide are evaluated based on the peptide-membrane center of mass distance, the number of peptide positive-membrane negative atom pairs and the number of peptide-membrane atomic contacts; the function is verified through an E. coli recombinant expression system, and finally a cecropin AD mutant with development potential is screened. The application effectively improves the efficiency and reliability of the screening of cecropin AD candidate mutants, and provides a molecular mechanism and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of biomedicine and antimicrobial peptide molecular design technology, and relates to a cephalosporin AD intelligent optimization design method that integrates VAE sequence amplification, AI evolution and molecular simulation. In particular, it relates to a cephalosporin AD optimization method that combines variational autoencoder sequence amplification, AI evolution optimization, dual-model antimicrobial potential prediction, molecular dynamics simulation screening and host cell growth verification. Background Technology

[0002] Antimicrobial peptides are a class of naturally occurring immunomodulatory molecules widely found in animals, plants, and microorganisms. They typically consist of short amino acid sequences and possess characteristics such as cationicity, amphiphilicity, and membrane-interacting capabilities. Unlike traditional antibiotics, which primarily rely on specific molecular targets to exert their effects, many antimicrobial peptides can exert their antibacterial effects through electrostatic adsorption, membrane surface enrichment, insertion into lipid bilayers, induction of membrane defects, or disruption of membrane integrity. Therefore, antimicrobial peptides have significant application potential in addressing bacterial resistance, developing novel anti-infective drugs, food preservatives, feed additives, and controlling agricultural pathogens.

[0003] Cecropins are a class of antimicrobial peptides first extracted and isolated from silkworm pupae by Swedish scientists Boman et al. They possess bioactivities including antibacterial, antifungal, antiviral, tumor cell inhibition, and immunomodulatory effects. Currently, cecropin antimicrobial peptides identified include cecropin A, B, C, D, E, P, and some analogues. Cecropin AD is a hybrid antimicrobial peptide formed by combining sequence fragments of cecropin A and cecropin D. Its N-terminus is primarily derived from cecropin A, and its C-terminus is primarily derived from cecropin D. This hybrid design aims to combine the advantageous fragments of different cecropin molecules, giving it strong positive charge, amphiphilicity, and membrane interaction capabilities. Cecropin AD typically interacts electrostatically with anionic phospholipids, lipopolysaccharides, or other negatively charged groups on the bacterial membrane surface through its positively charged amino acid residues, and further interacts with the membrane lipid bilayer through its hydrophobic amino acid regions, thereby causing membrane structural perturbations or even membrane damage. However, existing cephalosporin AD sequences may still have limitations such as limited antibacterial activity, insufficient stability, unsatisfactory selectivity, or unclear potential for further optimization. Therefore, sequence optimization of cephalosporin AD to obtain mutants with stronger membrane interaction capabilities and higher antibacterial potential is of great significance.

[0004] Currently, antimicrobial peptide optimization typically relies on methods such as human experience, site-directed mutagenesis, amino acid substitution rules, or high-throughput experimental screening. Traditional artificial mutagenesis methods are usually designed based on properties such as amino acid charge, hydrophobicity, amphiphilicity, and helical tendency. While these methods offer some interpretability, they limit the sequence space that can be explored and make it difficult to systematically discover performance improvements resulting from multi-site synergistic mutations. Although high-throughput experimental screening can directly evaluate antimicrobial activity, it is costly, time-consuming, and unsuitable for rapidly screening a large number of candidate sequences in the early stages.

[0005] In recent years, AI technology has been increasingly applied to the design of antimicrobial peptides. Sequence optimization methods based on deep learning, reinforcement learning, and protein language models can learn sequence features from existing antimicrobial peptide data and generate candidate sequences with high predictive antimicrobial potential, thus improving the efficiency of antimicrobial peptide design to some extent. However, most existing AI evolutionary models start with a single natural sequence and iteratively optimize through local amino acid substitutions. The generated sequences are usually limited to the local search space near the original sequence, easily getting trapped in local optima and failing to fully explore richer sequence diversity. Furthermore, different AI prediction models use different datasets, network structures, and evaluation criteria, which may lead to inconsistent prediction results between different models, thus affecting the reliability of candidate sequence selection. Summary of the Invention

[0006] To address the problems in existing cephalosporin antimicrobial peptide optimization processes, such as low candidate mutant generation efficiency, insufficient sequence space exploration, difficulty in evaluating multi-site synergistic mutation effects through manual experience alone, lack of membrane action mechanism evaluation in AI predictions alone, and lack of experimental support for computational results, this invention aims to provide a smart optimization design method for cephalosporin AD that integrates VAE sequence amplification, AI evolution, and molecular simulation. This method uses cephalosporin AD as the initial template sequence, constructs a sequence with the same length as the template and a limited number of mutation sites through a variational autoencoder (VAE), obtains an initial sequence library, executes an independent AI evolution process to generate candidate mutant antimicrobial peptides, and combines antimicrobial potential prediction, Gram-negative bacterial membrane system MD simulation, and E. coli expression verification to comprehensively evaluate the candidate sequences, thereby screening out cephalosporin AD mutants with further development potential.

[0007] This invention is achieved using the following technical solution: A smart optimization design method for cephalosporin AD that integrates VAE sequence amplification, AI evolution, and molecular simulation includes the following steps: 1) Using the amino acid sequence of cephalosporin AD as the initial template sequence, the initial template sequence was modified using a variational autoencoder to provide multiple amino acid sequences with no more than five amino acid mutations similar to the initial template, which served as the initial sequence library. 2) The initial sequence library was subjected to multiple rounds of mutation generation using an AI evolutionary model to obtain several candidate cephalosporin AD mutation sequences; 3) The candidate cephalosporin AD mutation sequences were evaluated for activity using two different antimicrobial peptide activity prediction models. Further screening was conducted by combining physicochemical property constraints (such as the overall positive charge of the candidate sequences) to obtain preliminary candidate sequences. 4) Construct peptide-Gram-negative bacterial membrane simulation systems based on the preliminary screening candidate sequences, perform MD simulations, and set up multiple independent replicates for each MD group; 5) Based on the MD simulation trajectory, calculate the peptide-membrane centroid distance, the number of positively charged peptide-negatively charged membrane atom pairs, and the number of peptide-membrane atom contacts for the candidate cephalosporin AD mutant; 6) Based on the indicators described in step 5), the membrane proximity or insertion degree, electrostatic adsorption capacity and membrane contact stability of the candidate sequences are comprehensively evaluated, and mutant antimicrobial peptide sequences with optimization potential compared with the original cephalosporin AD sequence are screened. 7) Construct the candidate sequence encoding gene obtained in step 6) into an *E. coli* expression vector for expression validation in *E. coli*: Using the natural cecropin AD encoding gene as a control, compare the specific growth rates of the host corresponding to different expression constructs, analyze the degree of inhibition of host *E. coli* growth by the candidate cecropin AD compared to the natural cecropin AD, and verify the potential host cell growth inhibition caused by the candidate peptides based on plasmid sequencing results. Based on the validation results, cecropin AD mutants with development potential can be screened.

[0008] Specifically, the AI ​​evolutionary model mentioned in step 2) is the AMPainter-1 model. The AI ​​evolutionary model and antimicrobial peptide activity prediction model mentioned in step 3) include one or both of the AMPainter-1 model and the AMP-potency-prediction model. Both models are existing antimicrobial peptide sequence scoring and evolutionary models based on interpretable deep learning. A higher AMPainter-1 score and a lower AMP-potency-prediction prediction value indicate a stronger predicted antimicrobial potential.

[0009] In some implementations, the multi-round mutation generation in step 2) includes the following steps: introducing substitution mutations at one or more amino acid sites in the initial sequence library sequence; predicting the antibacterial potential of the mutated sequence; retaining sequences that meet a preset threshold (which can be set according to the activity score corresponding to natural cephalosporin AD); using the retained sequences as the input for the next round of evolution; repeating the above process until the preset number of evolution rounds is reached, and using all sequences obtained from all evolution rounds as candidate mutation sequences.

[0010] Preferably, the candidate sequence in step 3) must show a significantly improved prediction score compared to the original cephalosporin AD sequence in at least two antimicrobial peptide prediction models before it can enter the subsequent MD simulation screening.

[0011] In some embodiments, the polypeptide-Gram-negative bacterial membrane simulation system described in step 4) includes one or more of phospholipids, negatively charged lipids, lipopolysaccharides, or other anionic membrane components. The MD simulation includes energy minimization, equilibrium simulation, and production simulation. Simultaneously, three independent replicate MD simulations are performed for each sequence to improve the stability and reliability of the screening results.

[0012] Specifically, in step 5), the peptide-membrane centroid distance is used to characterize the degree to which the candidate peptide approaches or enters the membrane region; the smaller the value, the closer the candidate peptide is to the membrane center. The number of positively charged peptide-negatively charged membrane atomic pairs is used to characterize the electrostatic contact strength between the cation residues and the anion membrane components. The number of peptide-membrane atomic contacts is used to characterize the contact area and binding stability between the candidate peptide and the membrane. These indicators are evaluated by combining the average value and dispersion of multiple independent repeated simulations.

[0013] In some embodiments, step 7) of the *E. coli* expression verification includes: linking the natural cephalosporin AD encoding gene cAD2 (natural peptide) and the candidate mutant encoding genes cAD3 to cAD7 (candidate peptides 1 to 5) into expression vectors containing SUMO fusion tags, and driving expression with strong promoters p00869 or p01902 to obtain multiple expression constructs; introducing the obtained expression constructs into *E. coli* hosts and measuring the host specific growth rate μ; when the μ of the host corresponding to the candidate construct is lower than that of the natural cephalosporin AD control construct, and plasmid sequencing confirms that the coding frame is intact, it can be used as evidence that the candidate peptide may have strong inhibitory effect on host cell growth. For constructs with frameshift, premature termination, or other inactivation mutations, they should be regarded as indirect clues of possible host selection pressure, rather than as definitive evidence of increased antibacterial activity of the candidate peptide alone.

[0014] The present invention also provides a series of optimized cephalosporin AD sequences, which are obtained by the above method and their amino acid sequences are shown in SEQ ID NO.1-5.

[0015] The beneficial effects of this invention are as follows: Compared with the prior art, the present invention has at least the following beneficial effects: 1) This invention uses cephalosporin AD as the initial template sequence and employs a variational autoencoder to modify this initial template sequence, providing multiple amino acid sequences with no more than five amino acid mutations similar to the initial template, serving as an initial sequence library. Utilizing an AI evolutionary model for candidate mutant sequence generation enables the rapid acquisition of candidate mutants with predicted increased affinity or antibacterial potential within a large sequence space, improving the generation efficiency of cephalosporin AD mutants and reducing the dependence on mutation site selection in traditional manual empirical design.

[0016] 2) This invention uses multiple AI models to independently optimize and predict the screening of cephalosporin AD. It can evaluate the optimization potential of candidate mutation sequences from different model perspectives, reduce the impact of single prediction model bias on the candidate sequence screening results, and improve the reliability of the initial screening of candidate sequences.

[0017] 3) This invention introduces MD simulation into the screening process of cephalosporin AD mutants. By constructing a Gram-negative bacterial membrane simulation system, it enables further evaluation of the dynamic interaction process between candidate peptides and bacterial membranes, thereby improving the mechanistic interpretability of the candidate mutant screening process.

[0018] 4) This invention further introduces the measurement of E. coli expression and host cell specific growth rate, which can link the calculated screening results with the preliminary cell phenotype. Sequencing the expression construct can distinguish candidate peptides with potential host toxicity and construction abnormalities such as frameshifting and premature translation termination, enhancing the experimental interpretability of candidate sequence evaluation.

[0019] 5) This invention is not only applicable to the sequence optimization of cecropin AD, but can also be extended to the generation, calculation screening and expression verification of candidate mutants of cecropin A, cecropin D and other cationic amphiphilic antimicrobial peptides. Attached Figure Description

[0020] Figure 1 Peptide-membrane centroid distance diagram. Membrane insertion depth reflects the ability of antimicrobial peptides to penetrate the membrane interior. The smaller the centroid distance, the deeper the peptide is inserted into the membrane, and the stronger the interaction between the peptide and the membrane.

[0021] Figure 2 Peptide positive charge vs. membrane negative charge atom pair diagram. The number of atom pairs formed between positively charged atoms in the peptide and negatively charged atoms on the bacterial membrane surface reflects the strength of the electrostatic interaction between the peptide and the bacterial membrane, and is an important indicator driving the initial adsorption of cationic antimicrobial peptides onto the bacterial membrane.

[0022] Figure 3 Peptide-membrane atomic contact number diagram. The peptide-membrane contact number reflects the contact area and binding stability between the antimicrobial peptide and the bacterial membrane. The higher the number of contacts, the more membrane lipid molecules the peptide can interact with, which is more conducive to membrane structure perturbation.

[0023] Figure 4 This graph shows the specific growth rate of *E. coli* when different strong promoters drive the expression of natural cephalosporin AD and a first-order AI evolution candidate sequence, and is used to compare the effect of candidate peptide expression on host cell growth.

[0024] Figure 5 The sequenced results show the upstream SUMO coding regions of the cAD3 to cAD7 expression constructs. The cAD3 and cAD5 constructs have a single A deletion, which leads to premature frameshift and translation termination. Detailed Implementation

[0025] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, parameter adjustments or process modifications made based on the concept of the present invention should fall within the scope of protection of the present invention.

[0026] In this embodiment, the original sequence of cephalosporin AD was used as the initial template sequence, and its amino acid sequence is shown in SEQ ID NO.6.

[0027] First, a variational autoencoder (VAE) model is used to locally amplify the original sequence of cephalosporin AD to construct an initial sequence library. The core of the VAE model includes a protein sequence feature extraction module, a Transformer variational encoding module, a latent variable sampling layer, and a Transformer decoding module. Specifically, the protein sequence feature extraction module uses the ESM-C protein language model to process the amino acid sequence of cephalosporin AD to obtain the corresponding residue-level high-dimensional embedding matrix. ,in It is the peptide sequence length. This is the output dimension of the ESM-C protein language model. It includes special tokens such as start and end. The Transformer variational coding module consists of a Transformer encoder, a mean prediction layer, and a log-variance prediction layer. The Transformer encoder processes the residue-level high-dimensional embedding matrix to obtain a global representation of the sequence, and then the latent spatial distribution is output through the mean prediction layer and the log-variance prediction layer. The mean prediction layer and the log-variance prediction layer output... , The two together define the potential spatial distribution. : Subsequently, the latent variable sampling layer obtains the latent representation vector from the latent spatial distribution through reparameterized sampling and inputs it into the Transformer decoding module. The Transformer decoder outputs the generation probability distribution of different amino acid types at each amino acid site, thereby generating candidate amino acid sequences similar to the original cephalosporin AD sequence. Further, the candidate amino acid sequences are constrained and screened, retaining those with the same length as the original cephalosporin AD sequence, containing only natural amino acid residues, and having no more than five amino acid substitution sites compared to the original cephalosporin AD sequence, as the initial sequence library.

[0028] In this embodiment, the variational autoencoder model is trained using antimicrobial peptide sequence data and cephalosporin sequence data, enabling the model to learn the statistical distribution characteristics and local mutation patterns of the antimicrobial peptide sequences. During model training, the input sequence reconstruction error is used as the basis for calculation. and latent space regularization term The primary optimization objective is the input sequence reconstruction error. This is used to constrain the consistency between the decoded sequence and the original antimicrobial peptide sequence in terms of amino acid composition and site pattern, specifically represented as follows: in To input the amino acid sequence, The decoder represents the first time. The amino acid corresponding to each output result This represents the number of times the input is z. The probability distribution of amino acids output by the decoder at each position. Latent space regularization term. This is used to maintain the continuity and smoothness of the latent variable distribution, facilitating sampling and sequence generation in the latent space near cephalosporin AD, specifically represented as follows: in Denotes KL divergence, The potential spatial distribution of the encoder output, The prior distribution of the latent variables can be set as a normal distribution. , It is an identity matrix.

[0029] Through the variational autoencoder generation and constraint screening described above, several candidate mutant sequences with no more than five amino acids of mutation sites were obtained that are similar to the initial template sequence of cecropin AD, forming an initial sequence library. This initial sequence library not only retains the basic backbone features and cationic amphiphilic antimicrobial peptide features of the original cecropin AD sequence, but also provides diverse local variants within a limited mutation range, providing richer initial sequences for the multi-round mutation generation of subsequent AI evolutionary models.

[0030] Subsequently, candidate sequences from the initial sequence library were used as input sequences for the AI ​​evolutionary model. The AMPainter-1 model was used to further select mutation sites and perform amino acid substitutions on the input sequences, generating a series of candidate mutant sequences. The antibacterial potential of the candidate sequences was scored using its built-in HyperAMP prediction module. Candidate sequences with higher scores were retained and used as input sequences for the next round of evolution for iterative optimization.

[0031] Simultaneously, the AMP-potency-prediction model was used to predict and score the initial sequence of cephalosporin AD and the mutant sequences obtained in each round of evolution. The two models have different scoring directions: the AMPainter-1 model predicts the probability of antimicrobial peptide activity; a higher value indicates a stronger predicted antimicrobial potential for the candidate sequence. Conversely, the AMP-potency-prediction model predicts the minimum inhibitory concentration (MIC); a lower value indicates a stronger predicted antimicrobial potential for the candidate sequence. By cross-scoring the candidate sequences using the AMPainter-1 and AMP-potency-prediction models, positive candidate sequences with significantly better predicted scores than the original cephalosporin AD sequence, and negative control sequences with lower predicted scores than the original sequence, were selected.

[0032] In this embodiment, five positive candidate sequences (candidate peptides 1 to 5, whose amino acid sequences are shown in SEQ ID NO. 1-5) and two negative candidate sequences (control peptides 1-2, whose amino acid sequences are shown in SEQ ID NO. 7-8) were first obtained through AI evolution and dual-model cross-scoring. Simultaneously, considering that the negative candidate sequences generated by the AI ​​evolution model might not show a significant difference in membrane interaction ability from the original cephalosporin AD sequence, this embodiment further designed a negative control sequence (negative peptide 1, whose amino acid sequence is shown in SEQ ID NO. 9) to enhance the control discrimination in subsequent MD simulation screening. The design principle of negative peptide 1 is: while maintaining the overall sequence length and some backbone features of cephalosporin AD as much as possible, some positively charged lysine or arginine residues in the original sequence are replaced with negatively charged or electrically neutral amino acid residues to reduce the overall positive charge of the candidate peptide and weaken its electrostatic adsorption ability with anionic lipids or negatively charged groups in the Gram-negative bacterial membrane.

[0033] Therefore, this embodiment ultimately obtained nine sequences, including the original cephalosporin AD sequence, the positive evolution sequence, and the negative control sequence, for subsequent MD simulation screening. The sequences and their dual-model scoring results are shown in Table 1, where cephalosporin AD is denoted as natural peptide AD in the table.

[0034] As shown in Table 1, the AMPainter-1 scores of candidate peptides 1 through 5 were all higher than those of cecropin AD, while their AMP-potency-prediction values ​​were all lower than those of cecropin AD, indicating that both models evaluated them as candidate sequences with improved antibacterial potential. Specifically, candidate peptide 1 had an AMPainter-1 score of 1.15 and an AMP-potency-prediction value of 0.26; candidate peptide 2 had scores of 1.09 and 0.22; candidate peptide 3 had scores of 1.14 and 0.34; candidate peptide 4 had scores of 1.06 and 0.41; and candidate peptide 5 had scores of 1.03 and 0.31. The combined prediction results of control peptide 1, control peptide 2, and negative peptide 1 were poor, and they were used as negative or antibacterial controls.

[0035] Table 1. Dual-model scores for cephalosporin AD and primary evolutionary positive and negative candidate sequences. To further evaluate the dynamic interaction between the candidate cephalosporin AD mutant and bacterial membranes, this embodiment constructs a peptide-Gram-negative bacterial membrane simulation system. The membrane system contains negatively charged or polar membrane lipid components to simulate the anionic environment on the surface of Gram-negative bacterial membranes.

[0036] In this embodiment, the membrane system includes membrane components such as EIPA, PPPE, PVPG, and PCL2, and water molecules and Mg are added. 2 + Na + and Cl - Ions were used to construct solvation and electroneutrality simulation systems. Among them, membrane components such as EIPA and PVPG provided negatively charged or polar interaction sites, which are beneficial for evaluating the electrostatic adsorption, salt bridging, and membrane surface binding capacity between cationic cephalosporin AD and its mutants and bacterial membranes.

[0037] Specifically, each candidate antimicrobial peptide sequence was placed in the same Gram-negative bacterial membrane simulation system to ensure consistency in membrane, solvent, and ionic environments among different candidate sequences. Each system underwent energy minimization and equilibrium simulations before entering the production simulation stage. For each candidate antimicrobial peptide sequence, three independent and repeated simulations were constructed to reduce the impact of random fluctuations in a single trajectory on the screening results and improve the stability of the evaluation results.

[0038] Natural cephalosporin AD itself possesses strong antibacterial activity and good bacterial membrane interaction capabilities. Therefore, this embodiment does not require the optimized sequence to be significantly superior to AD in all molecular dynamics indices. For candidate sequences obtained through AI evolution, as long as their key membrane interaction indices are comparable to AD, they can be considered to have essentially maintained the membrane interaction characteristics of natural cephalosporin AD; if some indices are further superior to AD, it suggests that the sequence may have optimization potential in specific aspects.

[0039] like Figure 1As shown, the AD+ series obtained through AI evolution (i.e., candidate peptides 1-5) did not exhibit a consistent trend in terms of peptide-membrane centroid distance. The average centroid distance of candidate peptide 1 was 3.02 nm, lower than the 3.10 nm of natural cephalosporin AD, a decrease of approximately 3%, making it the sequence with the smallest average distance in the AD+ series, suggesting good membrane accessibility or insertion capability. The average centroid distance of candidate peptide 4 was 3.06 nm, roughly equivalent to AD, indicating that it could maintain the membrane accessibility of natural cephalosporin AD well. The average distances of candidate peptides 2, 3, and 5 were slightly higher than AD, but the overall differences were limited, remaining at a level similar to natural AD. In contrast, the average centroid distances of the negative control peptides 1 and 2 were approximately 3.15 nm and 3.22 nm, respectively, while the artificially designed negative control peptide 1 was approximately 3.23 nm. The above results indicate that there are certain differences in membrane proximity or insertion capabilities among candidate sequences in a single AI evolution. Candidate peptide 1 shows relatively outstanding performance, while candidate peptide 4 and the remaining candidate sequences generally maintain membrane interaction levels similar to those of the natural AD. Furthermore, these results also suggest that an improvement in the AI ​​prediction score does not necessarily correspond to a synchronous decrease in the centroid distance of all candidate sequences; therefore, it is necessary to combine other membrane interaction indicators for a comprehensive evaluation.

[0040] The number of positively charged peptide-membrane negatively charged atom pairs formed by different sequences is as follows: Figure 2 As shown in the figure, the positively charged atoms in the peptides mainly originate from the lysine and arginine side chains, while the negatively charged atoms in the membrane mainly originate from anionic phospholipids, lipopolysaccharides, and other negatively charged groups. This indicator can be used to reflect the electrostatic adsorption capacity between cationic antimicrobial peptides and bacterial membranes. The average number of electrostatic atomic pairs for candidate peptides 3 and 4 is approximately 23.24 and 22.67, respectively, which is higher than the average number of electrostatic atomic pairs for natural cephalosporin AD (approximately 20.54), representing increases of approximately 13% and 10%, respectively. This suggests that both may have a stronger electrostatic adsorption capacity for anion exchange membrane components. The mean number of electrostatic atomic pairs for candidate peptide 1 is approximately 22.16, which is generally higher than AD. However, considering the fluctuations between different independent replicates, the degree of improvement still needs to be judged in conjunction with other indicators. The average number of electrostatic atomic pairs for candidate peptides 2 and 5 is close to or slightly lower than AD, indicating that they generally maintain the strong membrane electrostatic interaction characteristics of natural AD. Negative candidate peptides 1 and 2 generally showed lower values ​​than AD. The average value of the artificially designed negative peptide 1 was approximately 9.65, less than 50% of that of AD, indicating that its low positive charge design significantly weakened the electrostatic interaction with the anion exchange membrane components. Therefore, candidate peptides 3 and 4 showed superior electrostatic adsorption, while the results of the remaining AD+ candidate sequences being close to those of natural AD also indicate that these sequences well preserved the cationic membrane binding characteristics of natural cephalosporin AD.

[0041] The number of contacts between different sequences and membrane atoms, such as Figure 3As shown. The peptide-membrane atomic contact number reflects the contact area, binding degree, and interaction stability between the candidate peptide and the membrane. Candidate peptide 4 had the highest average contact number of 2429, approximately 31% higher than the 1851 of natural cephalosporin AD; candidate peptide 3 had 2225, an increase of approximately 20%; and candidate peptide 5 had approximately 1982, an increase of approximately 7%, indicating that these three sequences may have formed a wider direct contact with the membrane. Candidate peptides 1 and 2 had average contact numbers of 1808 and 1682, respectively, slightly lower than natural AD. Given that natural AD itself has good membrane binding capacity, the fact that candidate peptides 1 and 2 achieved slightly lower levels than AD in this metric indicates that they did not significantly lose the original membrane contact capacity of the natural sequences. The negative candidate control peptide 2 had an average contact number of 1771, while the artificially designed negative peptide 1 had approximately 1054, the latter exhibiting the weakest membrane contact capacity. The mean value of control peptide 1 was approximately 2075, but the standard deviation was as high as 678. The fluctuations between different independent replicates were large, and its membrane action ability still needs to be judged by combining centroid distance, electrostatic contact and repeatability stability.

[0042] comprehensive Figures 1 to 3 The results showed that the five AD+ candidate sequences obtained from one AI evolution generally maintained the good membrane interaction characteristics of natural cephalosporin AD, and exhibited differentiated advantages in different indicators. Candidate peptide 1 performed relatively better in terms of peptide-membrane centroid distance; candidate peptides 3 and 4 were more prominent in terms of electrostatic contact; the overall number of peptide-membrane contacts of candidate peptides 3, 4, and 5 was higher than that of natural AD, with candidate peptide 4 showing the greatest improvement. Although candidate peptide 2 did not show a significant advantage in a single indicator, its overall indicators did not deviate significantly from those of natural AD, indicating that it may still maintain good membrane interaction capabilities. These results indicate that dual-model scoring can enrich candidate sequences that maintain the membrane interaction characteristics of natural AD or improve in some indicators from one AI evolution sequence. Since natural cephalosporin AD itself is a benchmark sequence with good antibacterial properties, candidate peptides that reach a level comparable to AD in key MD indicators can be considered worthy of further experimental verification; furthermore, if some indicators are significantly better than AD, they can serve as an important basis for priority screening and ranking. It should be noted that the MD index mainly reflects the potential interaction ability between candidate peptides and bacterial membranes. The final antibacterial performance still needs to be verified by combining experimental results such as E. coli growth inhibition and minimum inhibitory concentration.

[0043] To evaluate the biological effects of an AI-evolved candidate sequence in a real cell system, the genes encoding natural cephalosporin AD and five AD+ candidate sequences were named cAD2, cAD3, cAD4, cAD5, cAD6, and cAD7, respectively. cAD3 corresponds to candidate peptide 1, cAD4 to candidate peptide 2, cAD5 to candidate peptide 3, cAD6 to candidate peptide 4, and cAD7 to candidate peptide 5. These encoding genes were constructed into expression vectors containing SUMO fusion tags and expressed using the strong promoters p00869 or p01902. Using the cAD2 construct expressing natural cephalosporin AD as a control with the same promoter, each construct was introduced into *E. coli* hosts, and the specific growth rate μ during culture was measured.

[0044] like Figure 4 As shown, under the p01902 promoter condition, the specific growth rate of E. coli in the p01902-cAD4 group was approximately 0.19 h. -1 The p01902-cAD2 control group had a response time of approximately 0.38 h. -1 The concentration was reduced by approximately 50%; under the p00869 promoter condition, the p00869-cAD4 group was approximately 0.34 h. -1 The p00869-cAD2 control group had a time interval of approximately 0.38 h. -1 The growth rate was reduced by approximately 10%. Under both promoter backgrounds, cAD4 showed a trend of reducing host growth rate, with p01902-cAD4 exhibiting the most significant effect. Since cAD4 encodes candidate peptide 2, this result provides direct experimental evidence that candidate peptide 2 may have a higher inhibitory effect on E. coli cell growth than natural cephalosporin AD.

[0045] Sequencing of different expression constructs revealed that, for example Figure 5 As shown, both the p00869-cAD3 and p00869-cAD5 constructs exhibit a single A deletion in the SUMO coding region upstream of the candidate peptides. This deletion causes a frameshift and premature termination, preventing candidate peptides 1 and 3 from being expressed completely as expected. This phenomenon may be related to the strong selection pressure exerted on *E. coli* by the complete constructs under constitutive strong promoter conditions, but it could also originate from accidental mutations occurring during cloning or replication. Therefore, frameshift mutations can only serve as indirect clues that candidate peptides 1 and 3 may have strong host cell toxicity, and cannot alone prove that the purified peptides have higher antibacterial activity than natural cephalosporin AD.

[0046] Based on the combined calculation and experimental results, candidate peptide 2 exhibited a relatively clear E. coli growth inhibition phenotype; candidate peptides 3 and 4 showed good performance in terms of MD electrostatic contact or overall membrane contact; candidate peptide 1 was relatively superior in terms of centroid distance, and its expression construct showed signs of inactivation mutation. Therefore, candidate peptides 1 to 5 can be retained as experimental validation candidates, with priority given to inducing expression, purifying proteins, and identifying their antibacterial titers, minimum inhibitory concentrations, hemolytic properties, and other cytotoxicities for candidate peptides 2, 3, and 4.

Claims

1. A method for intelligent optimization design of cephalosporin AD that integrates VAE sequence amplification, AI evolution, and molecular simulation, characterized in that, Includes the following steps: 1) Using the amino acid sequence of cephalosporin AD as the initial template sequence, the initial template sequence is modified using a variational autoencoder to provide multiple amino acid sequences with no more than five amino acid mutations similar to the initial template sequence, which serve as the initial sequence library; 2) The initial sequence library was subjected to multiple rounds of mutation generation using an AI evolutionary model to obtain several candidate cephalosporin AD mutation sequences; 3) The candidate cephalosporin AD mutation sequence was evaluated for activity using two different antimicrobial peptide activity prediction models to obtain preliminary candidate sequences; 4) Using the preliminary screening candidate sequences, construct peptide-Gram-negative bacterial membrane simulation systems, perform MD simulations, and set up multiple independent replicates for each MD simulation group; 5) Based on the MD simulation trajectory, calculate the peptide-membrane centroid distance, the number of positively charged peptide-negatively charged atom pairs, and the number of peptide-membrane atom contacts of the mutant corresponding to the candidate cephalosporin AD mutation sequence; the number of positively charged peptide-negatively charged atom pairs is used to characterize the electrostatic contact strength between the cation residues and the anion membrane components; 6) Based on the peptide-membrane centroid distance, the number of positively charged peptide-negatively charged atom pairs and the number of peptide-membrane atom contacts obtained in step 5), the membrane proximity or insertion degree, electrostatic adsorption capacity and membrane contact stability of the candidate cephalosporin AD mutant sequence are comprehensively evaluated, and mutant antimicrobial peptide sequences with optimization potential compared with the original cephalosporin AD sequence are screened as candidate peptide sequences. 7) The coding genes of the candidate peptide sequences obtained in step 6) were expressed and verified in E. coli. Based on the verification results, cecropin AD mutants with development potential were screened.

2. The intelligent optimization design method for cephalosporin AD integrating VAE sequence amplification, AI evolution, and molecular simulation as described in claim 1, characterized in that, In step 2), the AI ​​evolution model is the AMPainter-1 model.

3. The intelligent optimization design method for cephalosporin AD integrating VAE sequence amplification, AI evolution, and molecular simulation as described in claim 1, characterized in that, Step 2) specifically includes the following steps: An AI evolutionary model is used to introduce substitution mutations at one or more amino acid sites in each sequence in the initial sequence library; the antibacterial potential of the mutated sequences is predicted, and sequences that meet the preset threshold are retained; the retained sequences are used as the initial sequence library for the next round of mutations. Repeat the above process until the preset number of evolutionary rounds is reached, and use the sequences obtained from all evolutionary rounds as candidate mutation sequences.

4. The intelligent optimization design method for cephalosporin AD integrating VAE sequence amplification, AI evolution, and molecular simulation as described in claim 1, characterized in that, In step 3), the two different antimicrobial peptide activity prediction models are specifically the AMPainter-1 model and the AMP-potency-prediction model.

5. The intelligent optimization design method for cephalosporin AD integrating VAE sequence amplification, AI evolution, and molecular simulation as described in claim 1, characterized in that, In step 4), the polypeptide-Gram-negative bacterial membrane simulation system includes one or more of phospholipids, negatively charged lipids, lipopolysaccharides, or other anionic membrane components; the MD simulation includes energy minimization, equilibrium simulation, and production simulation.

6. The intelligent optimization design method for cephalosporin AD integrating VAE sequence amplification, AI evolution, and molecular simulation as described in claim 1, characterized in that, In step 7), the expression verification in *E. coli* specifically includes: linking the natural cecropin AD encoding gene and the candidate mutant encoding gene into an expression vector containing a SUMO fusion tag, and driving expression with a strong promoter p00869 or p01902 to obtain multiple expression constructs; introducing the obtained expression constructs into *E. coli* hosts and measuring the host specific growth rate μ; using the expression construct corresponding to natural cecropin AD as a control, when the μ of the host corresponding to the expression construct corresponding to the candidate mutant is lower than the μ of the host corresponding to the expression construct corresponding to natural cecropin AD, and plasmid sequencing confirms that the coding frame is complete, the candidate mutant is considered a cecropin AD mutant with development potential.

7. A series of optimized cephalosporin AD sequences, characterized in that, The amino acid sequence of the product is as shown in SEQ ID NO.1-5, obtained by the method described in any one of claims 1-6.

Images