A method of engineering any polypeptide sequence into an antimicrobial peptide

Abstract

The application discloses a method for transforming any polypeptide sequence into an antibacterial peptide, and belongs to the technical field of bioinformatics. The application develops a hypergraph neural network predictor for predicting the antibacterial activity score of a polypeptide, performs fine-tuning training on a pre-trained protein language model to capture antibacterial peptide sequence features, and trains a strategy network for selecting a to-be-mutated amino acid site in a polypeptide sequence. For any starting polypeptide sequence, the strategy network is first used to select a to-be-mutated amino acid site, then the fine-tuned protein language model is used to give the amino acid type after mutation of the site, and then the hypergraph neural network predictor is used to predict the antibacterial activity score of the sequence after mutation. The process is iteratively executed until an ideal antibacterial peptide sequence output is obtained. The method has the ability to transform any polypeptide sequence into an antibacterial peptide, and the antibacterial activity score of the transformed polypeptide is greatly improved, and can be applied to the actual development of antibacterial drugs, thereby laying a foundation for solving drug-resistant bacterial infections.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for modifying any polypeptide sequence into an antibacterial peptide, in particular to a method for modifying any polypeptide sequence into an antibacterial peptide based on deep learning, and belongs to the field of bioinformatics technology. BACKGROUND

[0002] Antibacterial peptides are a class of short peptides with broad-spectrum inhibitory effect on pathogens such as bacteria, which mainly exert antibacterial effect by destroying bacterial cell membranes. With the development of deep learning technology, developing a computational model provides a more efficient way for designing new antibacterial peptides. Although the number of antibacterial peptides recorded in existing databases has reached thirty thousand, compared with the vast theoretical sequence space, known antibacterial peptides are still very scarce, and not all of them have high antibacterial activity. The existing methods for designing new antibacterial peptides can be divided into two types: modification methods for improving the activity of known antibacterial peptides and generation methods for designing antibacterial peptides from scratch. Due to the limited ability of optimization algorithms, modification methods usually start from known antibacterial peptide sequences, and under the guidance of fitness function, modify the antibacterial peptide sequence through one or more iterations to improve its antibacterial activity. The generation method trains the generator through known antibacterial peptide data, inputs random noise, and obtains a series of antibacterial peptide sequences through the conversion function. Although the HydrAMP model (Nat Commun 2023, 14: 1453) as a generation method can complete the two tasks of unconstrained generation and input known antibacterial peptide sequence to generate analogs, but it needs to be set up independently when used. There is no model that can complete these two tasks in a unified framework with the same settings. SUMMARY

[0003] To solve the above problems, the present application provides a method for modifying any polypeptide sequence into an antibacterial peptide based on deep neural network, which can more efficiently obtain antibacterial peptides with high antibacterial activity.

[0004] As shown in Figure 1 The method for modifying any polypeptide sequence into an antibacterial peptide of the present application comprises the following steps:

[0005] A. Develop a hypergraph neural network predictor to predict the antibacterial activity score of polypeptides;

[0006] B. Fine-tune the pre-trained protein language model with known antibacterial peptide sequence data to make it biased towards capturing the features of antibacterial peptide sequences;

[0007] C. Train a strategy network to select the amino acid sites in the polypeptide sequence to be mutated;

[0008] D. Modify any polypeptide sequence into an antibacterial peptide, comprising:

[0009] D1. Input any starting polypeptide sequence and select the amino acid site to be mutated using the strategy network trained in step C;

[0010] D2. The protein language model fine-tuned in step B is used to determine the amino acid type after the mutation at the site, and the complete mutation sequence is obtained;

[0011] D3. The hypergraph neural network predictor developed in step A predicts the antibacterial activity score of the mutated sequence;

[0012] D4. Iterate steps D1, D2, and D3 until the desired sequence output is obtained.

[0013] In the above method for transforming any polypeptide sequence into an antimicrobial peptide, the specific method for developing a hypergraph neural network predictor in step A is as follows:

[0014] A1. Collect a dataset containing antimicrobial peptide sequences, non-antimicrobial peptide sequences, and the antimicrobial activity scores corresponding to each sequence, and randomly divide the dataset into training, validation, and test sets.

[0015] A2. Build a hypergraph neural network predictor, specifically including:

[0016] A2-1. Each peptide sequence is divided into segments (lengths of 2, 3, and 4 amino acids) using a sliding window, and a hypergraph is constructed for each of the three different segment lengths. The amino acid residues of the entire polypeptide are nodes V, the segmented segments are hyperedges E, and the features of each residue output by the protein language model constitute the node feature matrix X;

[0017] A2-2. Extract feature Y from the peptide hypergraph through hypergraph convolution. The calculation method is:

[0018]

[0019] Where X is the node feature matrix of the hypergraph, H is the incidence matrix of the hypergraph, W is the edge feature matrix (if not specified, W is a unit matrix), D v and D e are the diagonal matrices of node and edge degrees, respectively, and Θ is a parameter learned during training. Features Y2, Y3, and Y4 are obtained from peptide hypergraphs of three different fragment lengths (2, 3, and 4 amino acids, respectively).

[0020] A2-3. The three hierarchical features Y2, Y3, and Y4 output from step A2-2 are fused to obtain the final feature Y of the polypeptide sequence. c , the final feature Y c Input into a fully connected neural network to predict the antibacterial activity score r of the peptide:

[0021] r = dense(Y c )

[0022] where dense(·) denotes a fully connected neural network computation;

[0023] A3. Using the training set data in A1 to train the hypergraph neural network predictor constructed in A2, the loss function (Loss) is mean square error (MSE):

[0024]

[0025] for the i-th polypeptide, r i is the predicted antibacterial activity score output by step A2-3, y i is the known antibacterial activity score contained in the data set. The goal of training is to reduce the difference between the predicted antibacterial activity score and the known antibacterial activity score label, so that the hypergraph neural network predictor can accurately predict the antibacterial activity score of the polypeptide sequence. The validation set is used to select the predictor parameters, and the test set is used to test the performance of the predictor.

[0026] In the above method of modifying any polypeptide sequence into an antibacterial peptide, the strategy network in step C is a fully connected neural network with parameters θ, and the method for training the strategy network is:

[0027] C1. Randomly generate a series of polypeptide sequences with lengths between 10 and 40 amino acids as training data sets, and randomly initialize the strategy network parameters θ;

[0028] C2. Convert the polypeptide sequence into a digital code X aa input into the strategy network, and output the probability prob that each residue is predicted to be a mutation site;

[0029] prob = softmax(X aa θ)

[0030] C3. According to the probability output in step C2, sample to obtain a specific amino acid site a to be mutated;

[0031] a = multinomial(prob)

[0032] where multinomial(·) denotes a sampling operation;

[0033] C4. Replace the mutation amino acid site a output by step C3 with a mask character, input it into the fine-tuned language model in step B, obtain the mutated polypeptide sequence s, and then input it into the hypergraph neural network predictor developed in step A to obtain the antibacterial activity score r;

[0034] C5. Iteratively performing steps C2, C3, C4 for T times, obtaining a trajectory τ: (s1, a1, r1, …, sT, aT, rT), where s is the polypeptide sequence obtained in step C4, a is the mutation site obtained in step C3, and r is the antibacterial activity score obtained in step C4; using the trajectory τ, updating the policy network parameters θ using the REINFORCE algorithm. T , a T , r T ), where s is the polypeptide sequence obtained in step C4, a is the mutation site obtained in step C3, and r is the antibacterial activity score obtained in step C4; using the trajectory τ, updating the policy network parameters θ using the REINFORCE algorithm.

[0035] Preferably, the data set in step A1 contains antibacterial peptide sequences and their logMIC antibacterial activity values from the GRAMPA database, and the conversion method of logMIC and antibacterial activity score y is y = 1 / (1 + 10 logMIC-2 ). The data set also contains non-antibacterial peptide sequences of the same number as the antibacterial peptide sequences, selected from the Uniprot database, and the antibacterial activity score label is filled with 0. In an embodiment of the present application, the proportions of training set, validation set and test set are 8:1:1.

[0036] Preferably, the protein language model used to extract node features in step A2-1 is the Ankh model.

[0037] Specifically, the feature fusion operation in step A2-3 can include concatenation, cross product, Hadamard product, Kronecker product, and addition.

[0038] Preferably, the known antibacterial peptide sequence data in step B is derived from 11 antibacterial peptide databases, APD3, DRAMP, DBAASP, SATPdb, XUAMP, ADAM, AMPfun, LAMP, CAMP, dbAMP, and YADAMP, and the CD-HIT tool is used to remove redundancy with a threshold of 0.6-0.9 to retain representative antibacterial peptide sequences.

[0039] Preferably, the pre-trained protein language model in step B is the Ankh model, which is a T5 architecture model composed of an encoder-decoder and pre-trained on the UniRef50 data set.

[0040] Specifically, the encoding conversion method in step C2 can include one-hot encoding, numerical mapping encoding corresponding each type of amino acid letter to a number, feature encoding including amino acid composition frequency and physicochemical properties, and encoding using a pre-trained language model.

[0041] Preferably, the number of amino acid sites to be mutated in each output of step C3 is 1, and the iteration number T in step C5 is 3-15.

[0042] Specifically, the criteria for obtaining the desired sequence output in step D4 can be that the antibacterial activity score of the sequence predicted by the hypergraph neural network predictor after iteration reaches a target value, or the number of mutations of the sequence after iteration compared with the starting sequence reaches a target value, or a certain physicochemical property of the sequence after iteration meets a target value.

[0043] The present application also provides a system for transforming any polypeptide sequence into an antibacterial peptide, which is used to transform a polypeptide sequence by the above method for transforming any polypeptide sequence into an antibacterial peptide, such as Figure 2 As shown, the system comprises:

[0044] I. A mutation amino acid site designation module for selecting amino acid sites to be mutated from the input polypeptide sequence by using a strategy network;

[0045] II. A mutation amino acid decoding module for performing mutation on the amino acid sites to be mutated designated by the mutation amino acid site designation module by using the fine-tuned pre-trained protein language model, and outputting the mutated amino acid species and the complete polypeptide sequence after mutation;

[0046] III. A sequence evaluation module for scoring the antibacterial activity of the polypeptide sequence after mutation.

[0047] The method provided by the present application has the ability to transform any polypeptide sequence into an antibacterial peptide. When the existing antibacterial peptide, signal peptide and random sequence are transformed by the method of the present application, the antibacterial activity score of the transformed polypeptide can be greatly improved, and the transformed polypeptide has the potential to function as an antibacterial peptide. Therefore, the present application can contribute to the practical development of antibacterial drugs and lay the foundation for solving drug-resistant bacterial infections. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is the flow chart of the method for transforming any polypeptide sequence into an antibacterial peptide of the present application.

[0049] Figure 2 is the composition diagram of the system for transforming any polypeptide sequence into an antibacterial peptide of the present application.

[0050] Figure 3 is the network architecture diagram of the hypergraph neural network predictor developed by the present application.

[0051] Figure 4 is the comparison result of the method for transforming any polypeptide sequence into an antibacterial peptide of the present application and other methods in the same field.

[0052] Figure 5The method of the present invention for transforming any polypeptide sequence into an antimicrobial peptide results in transforming existing antimicrobial peptides, signal peptides and random sequences into antimicrobial peptides. DETAILED DESCRIPTION

[0053] The present invention will be further described below through specific embodiments with reference to the accompanying drawings.

[0054] Example 1 A method for transforming any polypeptide into an antimicrobial peptide

[0055] like Figure 1 As shown, the present invention provides a method for transforming any polypeptide sequence into an antimicrobial peptide, comprising:

[0056] 1) Develop a hypergraph neural network predictor to predict the antimicrobial activity scores of peptides.

[0057] 1.1) The dataset contains 3265 antimicrobial peptide sequences and their logMIC antimicrobial activity values ​​from the GRAMPA database, and 3265 non-antimicrobial peptide sequences from the Uniprot database. logMIC-2 ) The antimicrobial activity score y of the antimicrobial peptide sequence was calculated and used as the label, and the labels of all non-antimicrobial peptide sequences were filled with 0. The data were randomly divided into a training set (5,224 items), a validation set (653 items), and a test set (653 items) in a ratio of 8:1:1.

[0058] 1.2) The architecture of the hypergraph neural network predictor is as follows Figure 3 shown.

[0059] Each peptide sequence was divided into segments with a sliding window (lengths of 2, 3, and 4 amino acids), and hypergraphs were constructed for the peptides after segmentation with three different segment lengths. The amino acid residues of the entire polypeptide are nodes V, the segmented fragments are hyperedges E, and the features of each residue output by the protein language model Ankh constitute the node feature matrix X.

[0060] The three peptide hypergraphs with different fragment lengths are subjected to hypergraph convolution operations to extract features Y2, Y3, and Y4, respectively. The calculation method is:

[0061]

[0062] Among them, X is the hypergraph node feature matrix, H is the hypergraph incidence matrix, W is the identity matrix representing the hyperedge feature, and D v and D e are the diagonal matrices of node and edge degrees, respectively, and Θ is a parameter learned during training.

[0063] The three levels of features Y2, Y3, and Y4 are spliced ​​together to obtain the final feature Y of the polypeptide sequence.c ,

[0064] Y c = concat(Y2, Y3, Y4)

[0065] where concat(·) denotes the concatenation operation.

[0066] The input is fed into a two-layer fully connected neural network to predict the antibacterial activity score r of the polypeptide:

[0067] r = σ(σ(W c1 Y c +b c1 )W c2 +b c2 )

[0068] where σ(·) is the activation function ReLU, W c1 , b c1 are the first layer network parameters, and W c2 , b c2 are the second layer network parameters.

[0069] 1.3) The hypergraph neural network predictor is trained using the training set data, and the loss function (Loss) is the mean square error MSE:

[0070]

[0071] where r i is the predicted antibacterial activity score output for the i-th polypeptide, and y i is the known antibacterial activity score in the data set.

[0072] The optimizer used for training is Adam, the number of training iterations (Epoch) is set to 20, and the learning rate is finally selected as 0.0001 and the batch size is 128 based on the results on the reference validation set. After training is completed, the performance of the model is evaluated on the test set, with a Pearson correlation coefficient of 0.923 and a root mean square error (RMSE) of 0.163, indicating that the hypergraph neural network developed by the present application can accurately predict the antibacterial activity score of the polypeptide sequence, thereby being used for the development of a method for modifying any polypeptide into an antibacterial peptide.

[0073] 2) Fine-tune the pre-trained protein language model Ankh using 7316 known antibacterial peptide sequence data from 11 antibacterial peptide databases, APD3, DRAMP, DBAASP, SATPdb, XUAMP, ADAM, AMPfun, LAMP, CAMP, dbAMP, and YADAMP, to make it biased to capture the features of antibacterial peptide sequences.

[0074] 3) Train a policy network to select the amino acid sites in the polypeptide sequence to be mutated.

[0075] 3.1) Randomly generate 1000 polypeptide sequences with length between 10 to 40 amino acids, remove the redundancy with CD-HIT with threshold 0.7, and the remaining 966 sequences are used as training dataset, and initialize the policy network parameters θ randomly;

[0076] 3.2) Convert the polypeptide sequence into numerical encoding X by numerical mapping encoding method aa , i.e. represent the 20 amino acids letters with corresponding integers 1-20 respectively. Input X aa into the policy network. The policy network is a two-layer fully connected neural network with parameters θ, and the dimension of each layer is 128, which is used to output the probability prob that each residue is predicted to be the mutation site;

[0077] prob = softmax(X aa θ)

[0078] 3.3) Sample the specific amino acid site a to be mutated according to the probability prob;

[0079] a = multinomial(prob)

[0080] Where multinomial(·) represents the sampling operation.

[0081] 3.4) Replace the mutation site a output by the above steps with a mask character, input it into the fine-tuned protein language model in step 2 to obtain the mutated polypeptide sequence s, and then input it into the hypergraph neural network predictor developed in step 1 to obtain the antibacterial activity score r;

[0082] 3.5) Iteratively execute steps 3.2-3.4 for 8 times to obtain a trajectory τ: (s1, a1, r1, …, s8, a8, r8), and use the trajectory τ to update the policy network parameters θ using the REINFORCE algorithm.

[0083] Perform the above training process on all 966 sequences, and train for a total of 40 rounds to obtain the trained policy network.

[0084] 4) Use the above developed modules to realize the modification of any polypeptide sequence into an antibacterial peptide, including:

[0085] 4.1) Input the polypeptide sequence to be modified, and select the amino acid site to be mutated by the policy network;

[0086] 4.2) Obtain the complete mutated sequence by giving the amino acid type after mutation of the site by the fine-tuned protein language model;

[0087] 4.3) predicting the antibacterial activity score of the post-mutation sequence by the hypergraph neural network predictor;

[0088] Steps 4.1, 4.2, 4.3 are iteratively performed until the antibacterial activity score of the post-mutation sequence reaches 0.9, the modification is ended, and the output candidate antibacterial peptide sequence is collected.

[0089] Comparison of Example 2 with other methods

[0090] The method provided by the present application performs better than other known methods in the task of modifying known antibacterial peptides to improve their antibacterial activity. The input known antibacterial peptide data is 200 sequences with a minimum inhibitory concentration greater than 10 μΜ. Since other methods are all sequence optimization methods, the hypergraph neural network predictor provided by the present application is used as an optimization guide scoring. For each method, all sequences generated during the 8 rounds of optimization are collected, and the top 200 sequences with the highest predicted antibacterial activity scores are retained after optimization. The probability of these sequences being antibacterial peptides is predicted using six widely used antibacterial peptide classifiers: AMPScannerV2, CAMPR4-RF, CAMPR4-SVM, CAMPR4-ANN, MACREL, and ampir, the average value of the predicted values of each sequence is calculated, and the distribution is plotted as shown in Figure 4 The predicted probability distribution of the post-evolution sequence of the present application is higher than that of other methods.

[0091] Example 3: Modification of any polypeptide into an antibacterial peptide

[0092] The method described in the present application can be used to modify any polypeptide sequence into an antibacterial peptide. The following examples are provided for modifying existing antibacterial peptides, signal peptides, and random sequences into antibacterial peptides. The input sequences are 200 known antibacterial peptides with a minimum inhibitory concentration greater than 10 μΜ, 708 bacterial signal peptides, and 200 randomly generated polypeptides. Considering the difficulty of synthesis and the potential impact of disulfide bonds on antibacterial peptide activity, the starting polypeptides selected in this example exclude sequences containing cysteine (C). The above sequences are input into the method for 8 iterations, a total of 10 parallel modifications, and the sequence with the best antibacterial activity score predicted by the hypergraph neural network module after modification of each input sequence is collected, and the score distribution is shown in Figure 5 Compared to the antibacterial activity score of the input starting sequence, all three groups of different polypeptides have greatly improved antibacterial activity scores. Further, the top 10 candidate antibacterial peptides with the highest scores after modification are selected for experimental testing of antibacterial activity for the three groups of starting polypeptides. The names and sequences of the 30 antibacterial peptides are shown in Table 1.

[0093] Table 1. Sequences of 30 candidate antibacterial peptides

[0094]

[0095]

[0096] The bacterial strains used in the minimum inhibitory concentration assay include two gram-negative standard strains and two gram-positive standard strains, Escherichia coli ATCC25922, Pseudomonas aeruginosa PAOl, Staphylococcus aureus ATCC6538, Bacillus subtilis ATCC6633. The minimum inhibitory concentration of the candidate antibacterial peptide was determined using the Hancock broth microdilution method. The freeze-dried strains of the four standard strains were recovered, activated by overnight culture, and streaked onto LB (Hibio, #HB0129-2) plates for culture. Four single colonies with similar morphology were picked into 5 mL of LB medium (Hibio, #HB0218) and incubated at 37°C in a shaking incubator at 220 rpm for about 3 hours, then diluted to an OD 625 The OD value was in the range of 0.08-0.13.

[0097] The candidate antibacterial peptide powder was synthesized by Wuhan Dangang Biotechnology Co., Ltd. by solid-phase chemical synthesis method, and the mass spectrum identified the correct molecular weight. The antibacterial peptide sample was purified by high-performance liquid chromatography, and the purity was more than 95%. Each 0.1 mg of candidate antibacterial peptide powder was dissolved in sterile PBS buffer (pH 7.4) to prepare a candidate antibacterial peptide stock solution with a concentration of 512 μΜ.

[0098] The assay was performed using sterilized polypropylene 96-well culture plates (Greiner, #655201). Three parallel controls were set for each candidate antibacterial peptide. Columns 1-10 were for bacterial solution and gradient-diluted candidate antibacterial peptide solution, column 11 was for growth control (without adding antibacterial peptide solution), and column 12 was for sterile control (only adding culture medium). 100 μL of LB culture medium was added to each of columns 1-11, and 200 μL of LB culture medium was added to column 12. The candidate antibacterial peptide stock solution was diluted 8 times with LB culture medium, and 100 μL of the diluted solution was added to column 1. 100 μL was transferred from column 1 to column 2, mixed, and then 100 μL was transferred from column 2 to column 3, mixed, and the process was repeated to obtain a two-fold gradient dilution (concentrations were 128, 64, 32, 16, 8, 4, 2, 1, 0.5, and 0.25 μM, respectively). After mixing in column 10, 100 μL was discarded. 400 μL of the adjusted bacterial solution was diluted in 39.6 mL of LB culture medium to obtain a 100-fold dilution, and 100 μL of the diluted bacterial solution was added to each of columns 1-11. The 96-well culture plate was placed in a 37°C incubator for 16 hours. The growth of bacteria in each well of the culture plate was observed directly. The minimum antibacterial peptide solution concentration observed by naked eye was the minimum inhibitory concentration of the antibacterial peptide under the given conditions. The results are shown in Table 2.

[0099] Table 2. Minimum inhibitory concentrations of 30 candidate antibacterial peptides for four bacterial strains (unit: μM)

[0100]

[0101]

[0102] The minimum inhibitory concentration of at least one bacterial strain was less than or equal to 128 μM, which was the standard for antibacterial peptides. In the above table, Perceptide 01-10 was modified from known antibacterial peptides and still maintained antibacterial activity, and a certain degree of improvement was obtained compared to before modification. Eight of the 10 polypeptides Perceptide 11-20 modified from signal peptides were antibacterial peptides, and six of the 10 polypeptides Perceptide 21-30 modified from random sequences were antibacterial peptides. This demonstrates the effectiveness of the method provided by the present application in modifying sequences from different sources into antibacterial peptides, which can be used in a wider range of antibacterial peptide modification and design to obtain antibacterial peptides with ideal properties.

Claims

1. A method for engineering arbitrary polypeptide sequences into antimicrobial peptides, comprising the following steps: A. Developing a hypergraph neural network predictor for predicting the antimicrobial activity score of polypeptides, comprising steps A1-A3: A1. Collect a dataset containing antimicrobial peptide sequences, non-antimicrobial peptide sequences, and the corresponding antimicrobial activity scores for each sequence. Randomly divide the dataset into training, validation, and test sets. The dataset contains antimicrobial peptide sequences from the GRAMPA database and their logMIC antimicrobial activity values, logMIC and known antimicrobial activity scores. y The conversion method is The dataset also contains the same number of non-antimicrobial peptide sequences from the Uniprot database as the antimicrobial peptide sequences, and the antimicrobial activity score labels are filled with 0; A2. Constructing the hypergraph neural network predictor, comprising: A2-1. Each polypeptide sequence is segmented into fragments with sliding windows of lengths of 2, 3, and 4 amino acids, respectively, and a hypergraph is constructed for each of the three different fragment lengths , wherein each amino acid residue of the entire polypeptide is a node , the segmented fragments are hyperedges , and the features of each residue output by the protein language model form a node feature matrix ; A2-2. Extracting features of polypeptide hypergraph through hypergraph convolution operation The calculation method is: wherein, is a node feature matrix of the hypergraph, is a incidence matrix of the hypergraph, is a hyperedge feature matrix, and are diagonal matrices of the node and edge degree, respectively, are parameters learned during the training process; features are obtained from polypeptide hypergraphs of three different fragment lengths ; A2-3. input the three hierarchical features output from step A2-2 into a fully connected neural network to predict the predicted antibacterial activity score of the polypeptide fuse the final features of the polypeptide sequence , and input the final features into a fully connected neural network to predict the predicted antibacterial activity score of the polypeptide : wherein, denotes a fully connected neural network computation; A3. Using the training set data in A1 to train the hypergraph neural network predictor constructed in A2, the loss function Loss is mean square error MSE: for a polypeptide, a predicted antibacterial activity score output by step A2-3, a known antibacterial activity score contained in the dataset; the goal of the training is to reduce the difference between the predicted antibacterial activity score and the known antibacterial activity score label, so that the hypergraph neural network predictor can accurately predict the antibacterial activity score of a polypeptide sequence; the validation set is used to select the parameters of the hypergraph neural network predictor, and the test set is used to test the performance of the hypergraph neural network predictor;​ B. Fine-tuning a pre-trained protein language model to capture the features of antimicrobial peptide sequences by using known antimicrobial peptide sequence data from 11 antimicrobial peptide databases, APD3, DRAMP, DBAASP, SATPdb, XUAMP, ADAM, AMPfun, LAMP, CAMP, dbAMP, YADAMP, and using the CD-HIT tool to remove redundancy with a threshold of 0.6-0.9 to retain representative antimicrobial peptide sequences; the pre-trained protein language model is the Ankh model; C. training a strategy network for selecting amino acid sites in a polypeptide sequence to be mutated, the strategy network being a fully connected neural network with parameters , the method for training the strategy network comprising: C1. Randomly generate a set of polypeptide sequences of lengths between 10 and 40 amino acids as a training data set, while randomly initializing the policy network parameters ; C2. Converting polypeptide sequences to digital encodings In the input policy network, the output is the probability of each residue being predicted as the site to be mutated ​ C3. Sampling a specific amino acid site to be mutated according to the probability output in step C2 ; wherein, denotes a sampling operation; C4. The mutated amino acid sites output from step C3 Substitute with mask characters, input into the language model fine-tuned in step B to obtain the mutated polypeptide sequence , and input into the hypergraph neural network predictor developed in step A to obtain the predicted antibacterial activity score ; C5. Iteratively performing steps C2, C3, C4 for T times to obtain a trajectory wherein is the polypeptide sequence obtained in step C4, is the mutation site to be mutated obtained in step C3, is the predicted antibacterial activity score obtained in step C4; using the trajectory updating the policy network parameters using the REINFORCE algorithm ;​ D. Engineering arbitrary polypeptide sequences into antimicrobial peptides, comprising: D1. Inputting an arbitrary starting polypeptide sequence, and using the strategy network trained in step C to select the amino acid site to be mutated; D2. Using the fine-tuned protein language model in step B to give the type of amino acid after mutation at the site, and obtaining the complete mutated sequence; D3. Using the hypergraph neural network predictor developed in step A to predict the antimicrobial activity score of the mutated sequence; D4. Iteratively performing steps D1, D2, and D3 until the desired antimicrobial peptide sequence output is obtained.

2. The method of claim 1, wherein, The protein language model used to extract node features in step A2-1 is the Ankh model, and the feature fusion operation in step A2-3 is selected from one or more of splicing, cross multiplication, dot multiplication, Kronecker product, and addition.

3. The method of claim 1, wherein, The encoding conversion method in step C2 is selected from one of the following methods: one-hot encoding, numerical mapping encoding corresponding to each type of amino acid letter, including amino acid composition frequency and physicochemical property feature encoding, and using the encoding of the pre-trained language model.

4. The method of claim 1, wherein, The criteria for the desired antimicrobial peptide sequence in step D4 are that the antimicrobial activity score of the sequence after iteration predicted by the hypergraph neural network predictor reaches the target value, or the number of mutations of the sequence after iteration compared to the starting sequence reaches the target value, or a certain physicochemical property of the sequence after iteration meets the target value.

5. A system for engineering arbitrary polypeptide sequences into antimicrobial peptides, for implementing the method of engineering arbitrary polypeptide sequences into antimicrobial peptides according to any one of claims 1-4, comprising: I. A mutation amino acid site designation module for using a strategy network to select the amino acid site to be mutated for an input polypeptide sequence; II. A mutation amino acid decoding module for using a fine-tuned pre-trained protein language model to perform mutation on the mutation amino acid site designated by the mutation amino acid site designation module, and outputting the type of mutated amino acid and the complete polypeptide sequence after mutation; III. A sequence evaluation module for scoring the antimicrobial activity of the mutated polypeptide sequence.

Images