A novel antimicrobial peptide design method based on a large pre-trained model

Abstract

This invention discloses a method for designing novel antimicrobial peptides based on a large-scale pre-trained model, which lies at the intersection of bioinformatics and artificial intelligence. The method utilizes a larger-scale pre-trained generative model, ProtGPT2, to generate novel antimicrobial peptides. The resulting sequences better reflect the functional and structural relationships between amino acid residues in proteins. Another large pre-trained protein model is then used to extract protein features. During feature extraction, a character converter is used to convert the peptide sequence into a string of numbers. A feature extractor then converts these strings into an N×L×1024 array. Consequently, only a single discriminator is required, simplifying the screening process and achieving higher accuracy.

Description

Technical Field

[0001] The present invention relates to a novel antimicrobial peptide design method based on a large-scale pre-training model, and belongs to the intersection of bioinformatics and artificial intelligence. Background Art

[0002] Artificial intelligence natural language processing technologies, represented by recurrent neural networks, transformers, pre-trained models, and transfer learning, have achieved breakthroughs in various natural language analysis tasks, such as conversational bots, speech recognition, sentiment analysis, and text mining. In biology, proteins and natural language share a significant similarity: ① The fundamental building blocks of natural language are words (or letters, Chinese characters), while the fundamental building blocks of proteins are amino acid residues; ② The presence of a specific word in a sentence is determined by context; similarly, the presence of a specific amino acid residue is determined by the context of the protein sequence to determine the local structure and functional properties of the protein. Consequently, many mature natural language processing technologies are being applied to protein modeling.

[0003] Millions of natural protein data in public databases provide data support for large-scale pre-trained models. Large-scale pre-trained models with different architectures can achieve different protein modeling purposes, especially protein generation and feature extraction. The most advanced large-scale protein language models can randomly generate artificial sequences that are very similar to the physical and chemical properties of natural proteins. However, if the user wants a protein sequence with a specific function, it needs to be fine-tuned based on the existing data. The sequence generated after fine-tuning will be more likely to be the protein sequence the user needs, but there are still not many excellent methods for screening these candidate sequences. Feature extraction methods based on large-scale pre-trained models can convert each amino acid residue in each protein sequence into a numerical vector of a specific length. The value in this vector depends on the type of residue and the context of the sequence. Many academic evidences have shown that this feature extraction method based on large-scale pre-trained models is significantly better than the traditional one-hot-encoding method.

[0004] Antimicrobial peptides are short-chain polypeptides (also known as short-chain proteins) with antimicrobial effects. Compared to small-molecule antibiotics, antimicrobial peptides offer several advantages, including a more common broad-spectrum activity, a more diverse range of sources, and immunomodulatory effects. Therefore, antimicrobial peptides offer a new avenue for combating "superbugs," addressing the global challenge of a depleted supply of new antibiotics.

[0005] At present, the academic community has proposed a variety of methods for generating antimicrobial peptides through artificial intelligence. For example, the most cutting-edge existing technology is the paper published by Das et al. in 2021 (Das, Payel, et al. "Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations." Nature Biomedical Engineering 5.6 (2021): 613-623.). However, the solution given in this article is a short peptide generation model, so the generated sequence cannot fully reflect the "grammar" of protein language, that is, the functional and structural relationship between amino acid residues; and when training the model, it uses the hidden layer of the generation model as a feature to train multiple models of the physicochemical properties and toxicological properties of polypeptides, and then relies on these multiple prediction models to screen candidate sequences. Therefore, its screening conditions are relatively complex, and multiple discriminators need to be set up. In addition, when extracting features from proteins, the accuracy of feature extraction needs to be further improved. Summary of the Invention

[0006] In order to solve the existing problems, the present invention provides a novel antimicrobial peptide design method based on a large pre-training model, the method comprising:

[0007] Step 1: Obtain training samples; wherein, the sequences of known antimicrobial peptides are used as positive samples, and short-chain polypeptide sequences with no antimicrobial effect are used as negative samples, and the frequencies and lengths of amino acids in the positive and negative samples are ensured to be highly similar;

[0008] Step 2: ProtGPT2 is used as the protein generation model, and a deep convolutional neural network is used as the discriminator to construct a large pre-trained model ProtTrans ProtT5-XL-UniRef50;

[0009] Step 3: Fine-tune the protein generation model ProtGPT2, and use the fine-tuned model to generate a predetermined number of peptides, which are recorded as fine-tuned peptides; use the protein generation model ProtGPT2 before fine-tuning to generate a predetermined number of peptides, which are recorded as unfine-tuned peptides;

[0010] Step 4, feature extraction is performed on the training samples, the peptides generated after fine-tuning, and the peptides generated without fine-tuning using the large pre-trained model ProtTrans ProtT5-XL-UniRef50;

[0011] Step 5: training the discriminator using the feature-extracted positive and negative sample sequences;

[0012] Step 6: Use the trained discriminator to predict the fine-tuned peptides obtained in step 3, and select the candidate sequences that are most likely to be antimicrobial peptides based on the predicted values;

[0013] Step 7: Chemically synthesize the novel antimicrobial peptide based on the candidate sequence.

[0014] Optionally, the deep convolutional neural network comprises: an input layer, a 3×3 convolution layer, a 2×2 maximum value grabbing layer, a 3×3 convolution layer, a 2×2 maximum value grabbing layer, a flattening layer, three fully connected layers, and an output layer;

[0015] The deep neural network is implemented using the Keras toolkit in Python, the optimizer is Adam, the loss function is binary crossentropy, the evaluation method is accuracy, the feature X is the feature extracted in step 4, and the label y is the yes / no antimicrobial peptide in the positive and negative samples.

[0016] Optionally, in step 1, the frequencies and lengths of amino acids in the positive and negative samples are highly similar, which means:

[0017] The number of peptides containing the same number of amino acids in the positive and negative samples is the same.

[0018] Optionally, in step 3, fine-tuning the protein generation model ProtGPT2 includes:

[0019] The protein generation model ProtGPT2 was fine-tuned using the script file run_clm.py in the transformers toolkit in the Python programming language.

[0020] Optionally, in step 3, the process of generating polypeptides using the protein generation model ProtGPT2 utilizes the pipeline function in the transformers toolkit in the Python programming language, with the starting sequence being blank, the maximum generation length being set to 20, the minimum length being set to 6, and other parameters being set to default.

[0021] Optionally, in step 4, feature extraction is performed on the training samples, the peptides generated after fine-tuning, and the peptides generated without fine-tuning using a large pre-trained model ProtTrans ProtT5-XL-UniRef50, including:

[0022] Use the T5Tokenizer.from_pretrained function in the transformers toolkit in the Python programming language to set the character converter and the TFT5EncoderModel.from_pretrained function to set the feature extractor;

[0023] The maximum extraction length of peptides by the feature extractor is set to 50;

[0024] The peptide sequences were converted into digital strings using a character converter, and then these digital strings were converted into N × L × 1024 arrays using a feature extractor, where N is the number of sequences, L is the maximum sequence length of 50, and 1024 is the length of the vector representation of a single amino acid residue;

[0025] The extracted features are used as input to the deep neural network.

[0026] Optionally, when training the discriminator using the positive and negative sample sequences that have been feature extracted in step 5, the training samples are split into a training set and a validation set in a 9:1 ratio; the maximum number of training rounds is set to 75, and the EarlyStopping function in Keras is used to set an early exit mechanism, and the patience value is set to 6, that is, if the loss value does not decrease within six rounds of training, then the training will end early.

[0027] Optionally, in step 6, screening out candidate sequences most likely to be antimicrobial peptides based on the predicted values includes:

[0028] When the probability of the generated polypeptide sequence corresponding to the antimicrobial peptide is >0.95, it is regarded as a candidate sequence.

[0029] The beneficial effects of the present invention are:

[0030] (1) Compared with the prior art method of Das et al., which requires training a short peptide generation model from scratch, the method of the present application directly utilizes a larger-scale pre-trained generative model ProtGPT2 (Ferruz, Noelia, Steffen Schmidt, and Birte "ProtGPT2 is a deep unsupervised language model for protein design." Nature communications 13.1(2022):4348.), the generated sequences can better reflect the "grammar" of protein language (the functional and structural relationships between amino acid residues), and the GPT2 (Generative Pre-trained Transformers-2) architecture used in ProtGPT2 pre-training has been proven to perform well in the field of natural language processing.

[0031] (2) For the screening of candidate sequences generated by the model, the method of Das et al. in the prior art requires the hidden layer of the generated model to be used as a feature, and multiple models of the physicochemical properties and toxicological properties of polypeptides are trained, and then the candidate sequences are screened based on these multiple prediction models; while the method of the present application only requires the training of an antimicrobial peptide discriminator model, which simplifies the screening conditions.

[0032] (3) In addition, since the discriminator model uses another large pre-trained protein model ProtTrans (Elnaggar, Ahmed, et al. "Prottrans: Toward understanding the language of lifethrough self-supervised learning." IEEE transactions on pattern analysis and machine intelligence 44.10(2021):7112-7127.) to extract protein features, it achieves a higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 This is a flow chart of a novel antimicrobial peptide design method based on a large pre-trained model provided by one embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the discriminator architecture in a large pre-training model provided by one embodiment of the present invention.

[0036] Figure 3 This is a simulation of the test results of the growth of test strains using the antimicrobial peptides predicted and randomly generated using the method of the present invention. The tested drug concentration was set at 60 μM, an equal volume of DMSO was used as a negative control, and all conditions were tested in triplicate. A represents the test strain S. aureus GDMCC 1.221, B represents the test strain B. pumilus GDMCC 1.225, C represents the test strain B. subtilis GDMCC 1.222, D represents the test strain K. rhizophila GDMCC 1.226, E represents the test strain E. coli GDMCC 1.335, and F represents the test strain E. coli DH5ɑ. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0038] The peptides used in the following examples were purchased from Nanjing GenScript Biotechnology Co., Ltd. The peptides were produced using solid-phase peptide synthesis. Crude peptides with a purity of 40-70% were used in the antibacterial experiments. For ease of calculation, the purity of all crude peptides was set to 50%. The crude peptides were further purified by high-performance liquid chromatography to obtain high-purity peptides with a purity exceeding 90% for use in MIC experiments. The molecular weights of all synthesized peptides were confirmed by mass spectrometry.

[0039] The test strains used in the following examples, including Staphylococcus aureus GDMCC 1.221, Bacillus pumilus GDMCC 1.225, Bacillus subtilis GDMCC 1.222, Kocuria rhizophila GDMCC 1.226, and Escherichia coli GDMCC 1.335, were purchased from the Guangdong Microbial Culture Collection Center (GDMCC). All can be used for antibiotic potency testing. E. coli DH5α is a commonly used laboratory strain and is maintained in-house.

[0040] Example 1:

[0041] This example provides a novel antimicrobial peptide design method based on a large pre-trained model. Figure 1 , the method comprising:

[0042] 1) Configure the operating environment of Python and its toolkit.

[0043] The hardware requirements are an NVDIA A100 (80GB) graphics card and a Linux Ubuntu 22.04 operating system. First, install the Python environment management software Anaconda and the version control program git. Then, use Anaconda to install Python 3.8 and create and activate a Python environment specific to this method. Next, use the Python toolkit management command pip to install toolkits related to machine learning, including flax, pandas, datasets, evaluate, and scikit-learn. Use Anaconda to install toolkits related to the deep learning framework Pytorch, including pytorch, torchvision, torchaudio, and pytorch-cuda. Use Anaconda to install the deep learning frameworks keras and tensorflow-gpu. Finally, use pip to install the natural language processing toolkit, including SentencePiece and transformers.

[0044] 2) Obtain known antibacterial peptide (hereinafter referred to as antimicrobial peptide) sequences (as positive samples) from the Antimicrobial Peptide Database (SATPdb) and obtain short-chain peptide sequences (as negative samples) and their annotations from the public protein database (UniProt-SwissProt).

[0045] To avoid the mixing of peptides with antibacterial effects into negative samples, peptides containing specific keywords ("antimic*", "defensive", "antibac*", "toxic*") in the peptide annotations were removed.

[0046] 3) To ensure the feasibility of subsequent peptide synthesis experiments and control costs, peptides containing non-natural protein amino acids were eliminated from the positive and negative samples, and the positive and negative samples were limited to within 50 amino acid residues.

[0047] 4) To avoid systematic discrepancies in length or amino acid frequency between positive and negative samples, which could potentially trick the discriminator, negative sample sequences were collected to closely mimic the amino acid frequency and length of the positive samples, ensuring that the final number of positive and negative samples was the same. After data collection was complete, it was randomly split into training and test sets at a ratio of 9:1.

[0048] The highly mimicked amino acid frequency and length of the positive sample ensures that the number of peptides containing the same number of amino acids in the positive and negative samples is the same. For example, the positive sample has 10 6-aa (6 amino acid) peptides and 15 8-aa peptides; the negative sample also contains 10 6-aa peptides and 15 8-aa peptides.

[0049] Regarding the frequency of occurrence of amino acids, this application has counted the frequency of occurrence of each amino acid in antimicrobial peptides in nature, such as arginine 0.06, lysine 0.11, and cysteine 0.09. Therefore, in the process of generating negative samples, this application also generates the probability of occurrence of various amino acids according to the above statistical results.

[0050] 5) A set of negative controls were randomly generated, and for the reasons mentioned above, they also closely mimicked the amino acid frequencies and lengths of the positive samples.

[0051] 6) Use git to get the transformers project directory on GitHub.

[0052] Fine-tune the protein generation model (ProtGPT2, e.g., Figure 1 As shown in the figure), the fine-tuning script file is run_clm.py in the transformers directory, the model source is ProtGPT2, and the hyperparameters are set to the default values.

[0053] 7) Generate a batch of peptides using the fine-tuned generative model (e.g. Figure 1 As shown in the figure, it is called "fine-tuned generated peptides"). The generation process uses the pipeline function in the transformers toolkit in the Python programming language. The model source is ProtGPT2, the starting sequence is blank, the maximum generation length is set to 20, the minimum length is 6, and other parameters are set to default.

[0054] 8) Generate a batch of peptides (referred to as “generated peptides without fine-tuning”) using the unfine-tuned generative model. The model source is the fine-tuned model generated in the above step, and other settings are the same as above.

[0055] 9) Feature extraction (e.g., embedding, etc.) is performed on all sequences (positive samples, negative samples, negative controls, peptides generated after fine-tuning, and peptides generated without fine-tuning). Figure 1 shown).

[0056] The specific method is to use the T5Tokenizer.from_pretrained function in the transformers toolkit in the Python programming language to set up a character converter, and the TFT5EncoderModel.from_pretrained function to set up a feature extractor. The model sources for both the character converter and the feature extractor are "Rostlab / prot_t5_xl_uniref50". The feature extractor sets the maximum extraction length for peptides to 50. Next, the character converter is used to convert the peptide sequence into a numeric string (token). The feature extractor then converts these numeric strings into an N × L × 1024 array (N is the number of sequences, L is the maximum sequence length of 50, and 1024 is the length of the vector representation of a single amino acid residue) to serve as the input for the subsequent neural network.

[0057] 10) Use the feature-extracted positive and negative sample sequences to train a discriminator whose architecture is a deep convolutional neural network (e.g. Figure 2 As shown in the figure, the neural network has the following layers: input layer, 3×3 convolution layer (conv2D), 2×2 maxpooling layer, 3×3 convolution layer (conv2D), 2×2 maxpooling layer, flattening layer, fully connected layer (three layers in total, each with 512 nodes), and output node.

[0058] The neural network was implemented using the Keras toolkit in Python. The optimizer was Adam (learning rate was 0.001), the loss function was binary crossentropy, and the accuracy was evaluated. The features (X) were the features extracted in the embedding step above, and the labels (y) were the positive / negative antimicrobial peptides in the positive and negative samples.

[0059] 11) Use the fit function in the Keras model object to train the neural network on X and y. During training, split the training set into a training set and a validation set in a ratio of 9:1. Set the maximum number of training epochs to 75, and use the EarlyStopping function in Keras to enable early exit. Set the patience value to 6, meaning that if the loss does not decrease within six epochs, terminate the training early.

[0060] 12) Test the trained discriminator on a pre-set test set to evaluate its accuracy (using the evaluate function in the Keras model object). Then, perform predictions on random negative controls and peptides generated without fine-tuning (using the predict function in the Keras model object) to evaluate the changes before and after fine-tuning the generative model for reference.

[0061] 13) The trained discriminator was used to predict the generated candidate peptides. Candidate peptides were screened based on the predicted values (antimicrobial peptide probability > 0.95). Twenty-four candidate sequences were randomly selected, numbered "AMP1" through "AMP24," for experimental validation. Additionally, a set of peptide sequences were randomly generated based on the length and amino acid distribution of "AMP1" through "AMP24" as negative controls, numbered "RP1" through "RP10."

[0062] Example 2

[0063] The bacterial culture process includes:

[0064] Single colonies were picked from fresh plates and inoculated into LB liquid medium. Culture was then shaken overnight at 37°C, 200 rpm. The overnight culture was inoculated into fresh LB medium at a ratio of 1:100 and cultured for 4-5 hours until the logarithmic growth phase (OD600 0.4-0.6) was reached. The logarithmic phase culture was diluted with LB medium to an OD600 of 0.1, then diluted 1000-fold and aliquoted into 96-well plates, 100 μL per well, for use in inhibition and MIC experiments.

[0065] Example 3

[0066] This example conducts an antibacterial experiment based on the bacteria cultured in Example 2, including:

[0067] The crude peptide synthesized according to the polypeptide sequence screened in Example 1 was dissolved in dimethyl sulfoxide (DMSO) to a concentration of 3 mM to prepare a polypeptide stock solution.

[0068] In a 96-well plate containing diluted bacterial suspension, add 4 μL of crude peptide stock solution to each well and add fresh LB medium to a total volume of 200 μL, resulting in a test peptide concentration of 60 μM. An equal volume of DMSO serves as a negative control, and 200 μL of fresh LB medium serves as a blank control.

[0069] After adding the sample, seal the lid of the 96-well plate with a sealing film and place it in a high-speed shaking incubator at 37°C and 500 rpm for 11 to 12 hours.

[0070] Three biological replicates were performed for each condition.

[0071] At the end of the culture, the OD600 of each well was measured with a microplate reader, and the OD600 of the blank control was subtracted.

[0072] It has been verified that under test conditions, the simulation results are as follows Table 1 and Figure 3 As shown:

[0073] AMP3, AMP6, AMP13, AMP14, AMP16, AMP19, AMP23 and RP4 could completely inhibit the growth of strain GDMCC 1.221.

[0074] AMP12 and AMP15 could partially inhibit the growth of the bacterium, and the predicted peptide positive rate was 37.5% (9 / 24), which was higher than the positive rate of randomly generated peptides by 10% (1 / 10);

[0075] AMP3, AMP6, AMP8, AMP13, AMP14, AMP16, AMP19, AMP23 and RP4 could completely inhibit the growth of strain GDMCC1.225.

[0076] AMP4, AMP5, AMP7, AMP12, AMP15 and AMP18 could partially inhibit the growth of the bacterium, and the predicted peptide positive rate was 58.3% (14 / 24), which was higher than the positive rate of randomly generated peptides by 10% (1 / 10);

[0077] AMP3, AMP6, AMP8, AMP13, AMP14, AMP16, AMP19, and AMP23 completely inhibited the growth of strain GDMCC1.222;

[0078] AMP1, AMP4, AMP5, AMP10, AMP12, AMP15, AMP17, AMP18, AMP20, AMP21, RP1, RP2, RP3, RP4 and RP6 could partially inhibit the growth of the bacterium, and the predicted peptide positive rate was 78.3% (18 / 24), which was higher than the positive rate of randomly generated peptides (50% (5 / 10)).

[0079] AMP3, AMP6, AMP8, AMP13, AMP16, AMP19, AMP23, and RP4 completely inhibited the growth of strain GDMCC 1.226;

[0080] AMP24, RP6, RP7 and RP8 could partially inhibit the growth of the bacterium, and the predicted peptide positive rate was 29.2% (7 / 24), which was close to the positive rate of randomly generated peptides (30% (3 / 10)).

[0081] AMP3, AMP6, AMP13, AMP19 and AMP23 could completely inhibit the growth of strain GDMCC 1.335. All random peptides had no inhibitory effect on strain GDMCC 1.335. The predicted peptide positive rate was 20.8% (5 / 24).

[0082] This is much higher than the positive rate of randomly generated peptides, which is 0% (5 / 10);

[0083] AMP3, AMP6, AMP13, AMP19, AMP23, AMP24, and RP4 could completely inhibit the growth of strain DH5ɑ, and the predicted peptide positive rate was 25% (6 / 24), which was higher than the positive rate of randomly generated peptides, which was 10% (1 / 10) ( Figure 2 and Figure 3 ).

[0084] Table 1: Amino acid sequence, length, predicted antibacterial efficiency, relative molecular mass of the antimicrobial peptides predicted and randomly generated using the method of the present invention, and their antibacterial effects

[0085]

[0086] Note: The drug concentration was set at 60 μM, with an equal volume of DMSO as a negative control. All conditions were tested in triplicate. AMP represents a peptide predicted by this method, and RP represents a randomly generated peptide. "+++" indicates complete growth inhibition, "+" indicates some inhibition compared to the control, but not complete inhibition, and "-" indicates no inhibition.

[0087] Example 4

[0088] This example conducts a MIC experiment based on the bacteria cultured in Example 2, including:

[0089] The high-purity peptide sample was dissolved in DMSO to 6 mM to prepare a stock solution, and then diluted with DMSO into 8 gradients using a two-fold gradient dilution method.

[0090] In a 96-well plate containing diluted bacterial solution, 4 μL of sample was added to each well and fresh LB medium was added to a total volume of 200 μL, so that the final concentrations of the tested peptides were 120 μM, 60 μM, 30 μM, 15 μM, 7.5 μM, 3.75 μM, 1.875 μM and 0.9375 μM, respectively.

[0091] After adding the sample, place the 96-well plate in a high-speed shaking incubator at 37°C and 500 rpm for 16 to 18 hours.

[0092] Three biological replicates were performed for each condition.

[0093] After the culture is completed, the bacterial growth in the wells is observed. The lowest antimicrobial drug concentration in the wells without bacterial growth is the minimum inhibitory concentration.

[0094] As shown in Table 2 below, it has been verified that, except for AMP19 whose MIC against some strains is greater than 120 μM, the MICs of AMP3, 6, 13 and 23 are all between 3.75 μM and 120 μM, proving that these antimicrobial peptides have broad spectrum activity, great practical value, and good application in the development and preparation of antimicrobial agents.

[0095] Table 2: MIC test results

[0096]

[0097] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A novel antimicrobial peptide design method based on a large pre-training model, characterized in that: The method comprises: Step 1: Obtain training samples; wherein, the sequences of known antimicrobial peptides are used as positive samples, and short-chain polypeptide sequences without antimicrobial effects are used as negative samples, and the number of polypeptides containing the same number of amino acids in the positive and negative samples is ensured to be the same; Step 2: ProtGPT2 is used as the protein generation model, and a deep convolutional neural network is used as the discriminator to construct a large pre-trained model ProtTrans ProtT5-XL-UniRef50; Step 3: Fine-tune the protein generation model ProtGPT2, and use the fine-tuned model to generate a predetermined number of peptides, which are recorded as fine-tuned peptides; use the protein generation model ProtGPT2 before fine-tuning to generate a predetermined number of peptides, which are recorded as unfine-tuned peptides; Step 4, feature extraction is performed on the training samples, the peptides generated after fine-tuning, and the peptides generated without fine-tuning using the large pre-trained model ProtTrans ProtT5-XL-UniRef50; Step 5: training the discriminator using the feature-extracted positive and negative sample sequences; Step 6: Use the trained discriminator to predict the fine-tuned peptides obtained in step 3, and select the candidate sequences that are most likely to be antimicrobial peptides based on the predicted values; Step 7, chemically synthesizing a novel antimicrobial peptide based on the candidate sequence; The step 4 further comprises: The character converter was set using the T5Tokenizer.from_pretrained function in the transformers toolkit in the Python programming language, and the feature extractor was set using the TFT5EncoderModel.from_pretrained function. The model sources for both the character converter and the feature extractor were Rostlab / prot_t5_xl_uniref50. The maximum extraction length of the feature extractor for peptides was set to 50. First, a character converter is used to convert the peptide sequence into a digital string. Then, a feature extractor is used to convert these digital strings into an N×L×1024 array, where N is the number of sequences, L is the maximum sequence length of 50, and 1024 is the vector representation length of a single amino acid residue. The extracted features are used as input to the deep neural network. The step 5 further comprises: The discriminator is trained using a sequence of feature-extracted positive and negative samples. The discriminator is constructed using a deep convolutional neural network. The deep convolutional neural network consists of the following layers: an input layer, a 3×3 convolutional base layer, a 2×2 maximum-grabbing layer, a 3×3 convolutional base layer, a 2×2 maximum-grabbing layer, a flattening layer, three fully-connected layers, and an output node. The neural network is implemented using the Keras toolkit in Python, the optimizer is Adam, the learning rate is 0.001, the loss function is binary crossentropy, the evaluation method is accuracy, the feature X is the feature extracted in step 4, and the label y is whether the positive or negative sample is an antimicrobial peptide; Use the fit function in the Keras model object to train the neural network on X and y. During training, split the training set into a training set and a validation set in a ratio of 9:

1. Set the maximum number of training epochs to 75, and use the EarlyStopping function in Keras to set an early exit mechanism with a patience value of 6, meaning that if the loss value does not decrease within six epochs, terminate the training early. The trained discriminator is tested using a pre-set validation set, and the accuracy is evaluated using the evaluate function in the Keras model object. The predict function in the Keras model object is then used to predict the random negative control and the peptides generated without fine-tuning to evaluate the changes before and after fine-tuning the generative model. The step 6 further comprises: Candidate polypeptide sequences are screened based on the predicted values, and candidate sequences are randomly selected and numbered for experimental verification; according to the length distribution and amino acid distribution of the candidate sequences selected above, a group of polypeptide sequences are randomly generated as negative controls.

2. The method according to claim 1, characterized in that In step 3, the protein generation model ProtGPT2 is fine-tuned, including: The protein generation model ProtGPT2 was fine-tuned using the script file run_clm.py in the transformers toolkit in the Python programming language.

3. The method according to claim 2, characterized in that In the process of generating peptides using the protein generation model ProtGPT2 in step 3, the pipeline function in the transformers toolkit in the Python programming language is used, the starting sequence is blank, the maximum generation length is set to 20, the minimum length is set to 6, and other parameters are set to default.

4. The method according to claim 3, characterized in that In step 6, the candidate sequences most likely to be antimicrobial peptides are screened out based on the predicted values, including: When the probability of the generated polypeptide sequence corresponding to the antimicrobial peptide is >0.95, it is regarded as a candidate sequence.

Images