Prediction screening method of membrane-rupturing antibacterial peptide

Abstract

The invention discloses a predictive screening method of membrane-rupturing antibacterial peptides, and belongs to the technical field of biological information. According to the method, a sequence classifier based on machine learning is developed, high-throughput screening is carried out on proteome, and a polypeptide sequence which can be in contact with a membrane and has a secondary structure mainly being a spiral is input into the sequence classifier to predict whether the polypeptide sequence is an antibacterial peptide sequence or not; and for the sequences screened by the sequence classifier, sequences with positive net charges and relatively hydrophobic are further screened, the novelty of the sequences is verified through sequence comparison, then the membrane contact probability average value and the oligomerization possibility of forming antiparallel dimers through combination are counted, and the antibacterial peptide sequence with the membrane rupture characteristic is screened out. The accuracy of predicting the antibacterial peptide by the sequence classifier based on machine learning developed by the invention is superior to that of an existing method, a series of membrane-ruptured antibacterial peptides are successfully obtained through the screening method, and the membrane-ruptured antibacterial peptides have good antibacterial activity and low toxicity and are ideal antibacterial drug development objects.

Description

Technical Field

[0001] The present invention relates to a method for screening membrane-perforating antimicrobial peptides, and in particular to a method for predicting and screening membrane-perforating antimicrobial peptides based on machine learning, and belongs to the field of bioinformatics technology. Background Art

[0002] Antimicrobial peptides are a class of short peptides widely present in the natural immune system of organisms. They have diverse mechanisms of action, the most important of which is the destruction of cell membranes, which is also the main feature that distinguishes antimicrobial peptides from traditional small molecule antibiotics. Due to the accelerated growth of antibiotic resistance and the global expansion of pathogenic bacteria, the need to find antimicrobial peptides with broad-spectrum bactericidal ability and not prone to drug resistance is becoming increasingly urgent. Membrane-breaking antimicrobial peptides have the ability to target and destroy cell membranes. This special antibacterial mechanism makes antimicrobial peptides have the potential to become a new generation of antimicrobial drugs, which can treat diseases such as cancer, bacterial and viral infections, and have huge market potential. With the development of artificial intelligence technology, people have developed machine learning models with the ability to identify antimicrobial peptides, and successfully mined a series of new antimicrobial peptides from the genome or proteome. However, the specific mechanism of action of these antimicrobial peptides needs to be further determined, and there is currently no method to directly screen antimicrobial peptides with cell membrane-destroying mechanisms. Summary of the invention

[0003] The purpose of the present invention is to provide a method for screening antimicrobial peptides with a cell membrane-destroying mechanism and predicting their membrane-destroying effects based on machine learning.

[0004] To achieve the above object, the technical solution proposed by the present invention is: a method for predicting and screening membrane-breaking antimicrobial peptides, comprising the following steps:

[0005] A. Develop a sequence classifier based on machine learning, including:

[0006] A1. Organize the public database to form positive and negative data sets, where the positive data set is the antimicrobial peptide sequence data set, and the negative data set is the peptide sequence data set without antibacterial function;

[0007] A2. Use the pre-trained BERT language model to encode the peptide sequence sorted in step A1;

[0008] A3. Input the encoded polypeptide sequence into a support vector machine (SVM) to perform classification prediction training;

[0009] B. Use the SVM sequence classifier developed in step A to perform high-throughput screening on the proteome. The screening process includes:

[0010] B1. predict whether the polypeptide sequence is membrane-contactable, if yes, proceed to step B2;

[0011] B2. Predict whether the secondary structure of the polypeptide sequence is mainly helical, if yes, proceed to step B3;

[0012] B3. Use the SVM sequence classifier developed in step A to predict whether the input polypeptide sequence is an antimicrobial peptide sequence;

[0013] C. Predict the properties of the candidate antimicrobial peptide sequences screened in step B, including:

[0014] C1. Screen sequences based on physicochemical properties, retaining sequences with positive net charge and relatively hydrophobic;

[0015] C2. Determine the novelty of the sequence;

[0016] C3. Predict the average membrane contact probability of the peptide sequence;

[0017] C4. Predict the oligomerization possibility of polypeptide sequences to form antiparallel dimers;

[0018] D. Based on the sequence features obtained in step C, select polypeptide sequences that are novel, have a high average membrane contact probability, and have a high probability of oligomerization to form antiparallel dimers as candidate membrane-permeable antimicrobial peptides.

[0019] In the above step A1, preferably, sequences with a certain length interval, positive charge, and predicted to have membrane contact ability are sorted out from the known major antimicrobial peptide databases APD3, DBAASP, YADAMP, and SATPdb to form a positive data set. The Uniref database is searched to remove all sequences related to "antimicrobial peptides" to obtain polypeptide sequences with a certain length interval to form a negative data set. The ratio of the number of sequences in the positive and negative data sets is 1:5 to 1:10.

[0020] In the above step A2, preferably, a pre-trained ProtBERT-BFD model is used to encode the polypeptide sequence.

[0021] In the above step A3, the polypeptide sequences of the positive and negative data sets are randomly selected to form a training set and a test set in a certain ratio, so as to realize the training of the SVM sequence classifier.

[0022] In the above step B1, the specific criterion for contact with the membrane is: the predicted membrane contact probability (MCP) value of at least one residue in the sequence is above 0.2. The membrane contact probability of amino acid residues in the sequence is predicted using an MCP predictor based on a deep convolutional and recurrent neural network (DCRNN), which can be referenced by Wang et al. (Wang, L., et al., Membrane contact probability: An essential and predictive character for the structural and functional studies of membrane proteins. PLoS Comput Biol. 2022, 18(3): e1009972). The predictor extracts long-range interactions in protein sequences through BGRU, and combines CNN to learn local features to output the MCP prediction value of each residue.

[0023] In some embodiments of the present invention, the criterion for determining that the secondary structure is mainly helical in step B2 is that the number of residues predicted to be helical (H) with a probability value greater than 0.5 accounts for 50% or more of the sequence length.

[0024] In the above step C1, if the proportion of hydrophilic residues in the sequence is below 50%, the sequence can be considered to be relatively hydrophobic.

[0025] In the above step C2, BLAST-p can be used to determine the novelty of the sequence.

[0026] In the above step C3, the average membrane contact probability P of the polypeptide sequence is MCP It can be calculated by the following formula:

[0027]

[0028] Among them, n mem is the number of amino acid residues with a predicted membrane contact probability (MCP) value above 0.2, and L is the length of the sequence.

[0029] In the above step C4, the normalized value R of the antiparallel dimer is standard To predict the oligomerization probability of a polypeptide sequence to form an antiparallel dimer, Figure 2 As shown, the residue contact map (CM) of the sequence formed by connecting two identical polypeptides head to tail, where the upper left quarter is the target area for antiparallel dimer prediction, and the normalized value R standard The calculation formula is as follows:

[0030]

[0031] Among them, R 2 is the coefficient of determination of the linear regression in the prediction of the antiparallel residue contact map, and L is the length of the sequence;

[0032] Normalized value R standard The closer to 1, the better, representing that the sequence is predicted to form an antiparallel dimer.

[0033] In the above step D, preferably, the average membrane contact probability P MCP and the normalized value R of the antiparallel dimer standard Combined to score the peptide sequence:

[0034] F score =P MCP ×P standard

[0035] Final score: F score The interval is (0,1), and the candidate antimicrobial peptides are sorted from high to low according to the score. Select novel ones with a score of F score The sequences with high expression are candidate membrane-permeabilizing antimicrobial peptides.

[0036] Experimental data show that the prediction effect of the antimicrobial peptide sequence classifier based on machine learning developed by the present invention on the antimicrobial peptide classification benchmark data set is better than that of 7 existing methods, and the accuracy rate on the test data set reaches 0.96. Through the membrane-breaking antimicrobial peptide screening method of the present invention, a series of antimicrobial peptides with membrane-breaking mechanisms were successfully obtained, which met the screening expectations, had good antibacterial activity and low toxicity, and were not easy to cause bacterial resistance, and were ideal objects for the development of antimicrobial drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the method for predicting and screening membrane-breaking antimicrobial peptides of the present invention.

[0038] Figure 2 It is a schematic diagram of predicting antiparallel dimers using the residue contact map of connected sequences.

[0039] Figure 3 These are the circular dichroism spectrometer experimental results of 7 antimicrobial peptides screened out in the examples of the present invention.

[0040] Figure 4 The figure shows the leakage of carboxyfluorescein (CF) from liposomes induced by 7 antimicrobial peptides screened in the examples of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below through specific embodiments in conjunction with the accompanying drawings.

[0042] Example 1 Development of an antimicrobial peptide sequence classifier based on machine learning

[0043] From the known major antimicrobial peptide databases APD3, DBAASP, YADAMP, and SATPdb, we sorted out sequences with a length range of 10-40 amino acid residues, positive charge, and predicted to have membrane contact ability (the membrane contact probability of any residue in the sequence is greater than 0.2), totaling 2141 sequences, which constitute the positive set. The negative set sequences came from the search of the Uniref database, removing all sequences related to "antimicrobial peptides". The 9 keywords used included: antibacterial, AMP, antifungal, anticancer, antibiotic, antiviral, toxic, defensive, and secreted. The ratio of the number of positive and negative set sequences was 1:6.

[0044] The peptide sequences input into the classifier were encoded using the pre-trained ProtBERT-BFD model, and the data set was randomly selected in the training process of the support vector machine (SVM) classifier in a ratio of 4:1 to form a training set and a test set. Regarding the SVM classifier, the present invention uses ThunderSVM with a radial basis function (rbf) kernel. The hyperparameters C and gamma were optimized by cross-validated grid search (sklearn.model_selection.GridSearchCV).

[0045] After training, the performance of the classifier model was evaluated using the following metrics: sensitivity, specificity, accuracy, precision, Matthews correlation coefficient (MCC), F1 score, and area under the characteristic curve (AUROC). These metrics provide a comprehensive evaluation of the performance of the classifier model, and each metric highlights different aspects of the classification results.

[0046]

[0047] Among them, TP represents the number of true positive sequences, FP represents the number of false positive sequences, TN represents the number of true negative sequences, and FN represents the number of false negative sequences.

[0048] The comparison between the present invention and other existing methods is shown in Table 1. The test used the same data set as in the literature of Cao et al. (Briefings in Bioinformatics, 24, 2, 2023). The best result of each indicator in the 8 methods is bolded, and the second-ranked result is underlined. The benchmark test results show that when compared with other methods of the same period, the classifier of the present invention performs well in multiple indicators, reflecting its ability to accurately classify antimicrobial peptide sequences.

[0049] Table 1. Benchmark results of SVM classifier

[0050]

[0051] Example 2 Prediction of membrane rupture related characteristics

[0052] First, we counted 30 antimicrobial peptide sequences with sequence lengths between 10 and 40 amino acids, supported by membrane-breaking experimental data or molecular dynamics simulation results, from the database DBAASP to form a benchmark test set of antimicrobial peptides with membrane-related mechanisms. The above sequences were predicted using the membrane contact probability (MCP) predictor based on the DCRNN neural network, and the MCP prediction values ​​of the average amino acid residues were calculated. At the same time, SPIDER3 was used to predict the secondary structure (SS) of the peptide sequence, and the results were divided into three categories: α-helix (H), β-fold (E) and disordered coil (C). The benchmark test results of known antimicrobial peptide sequences are shown in Table 2. The MCP prediction values ​​of 77% of the sequences (23 sequences) were greater than 0.2, and the secondary structure of these sequences was mainly α-helix (H), which is consistent with the structural characteristics of known antimicrobial peptides that are easy to form helical membranes. The MCP and SS prediction standards can be set based on the known membrane-affinity antimicrobial peptide data set.

[0053] Table 2. Statistics of membrane contact probability (MCP) and secondary structure (SS) prediction results of 30 membrane-friendly antimicrobial peptides in the DBAASP database

[0054]

[0055]

[0056] Secondly, the oligomerization of short peptides was investigated, that is, whether they could form antiparallel dimers and then easily form polymers and transmembrane pores. Figure 2 As shown in the figure, by predicting the residue contact map (CM) of the sequence composed of two identical antimicrobial peptides connected head to tail, the contact map between the dimer residues is obtained. The upper left quarter of the entire contact map is the target area for the prediction of antiparallel dimers. The contact information between amino acids from the N-terminus to the C-terminus and in reverse order is predicted. If the peptide chain is predicted to form a regular antiparallel dimer, its contact map will show a standard diagonal pattern, which can be determined by the linear regression coefficient R 2 Quantify and further calculate the standardized result R standard .

[0057]

[0058] Among them, R 2 is the coefficient of determination of the linear regression in the prediction of the antiparallel contact map, and L is the length of the sequence. standard A value greater than 0.60 indicates that the sequence may form an antiparallel dimer, and the closer the value is to 1.0, the better.

[0059] Example 4 Screening of membrane-permeable antimicrobial peptides

[0060] The macroproteome sequences used for screening were collected from the Proteome Identification Database (PRIDE). The sources include human skin: Project PXD020742, containing 101,386 sequences; poison dart frog: Project PXD002326, containing 331,689 sequences; African clawed frog (salamander): Project PXD002326, containing 999,997 sequences. The length of all sequences is between 10 and 40 amino acids. The screening process described in the present invention was used to screen membrane-breaking antimicrobial peptides from these three macroproteomes. For the sequences screened by the SVM sequence classifier, the physicochemical properties of the sequences were statistically analyzed, and sequences with a net positive charge and a hydrophilic residue ratio of less than 50% were screened. This feature is consistent with the common properties of membrane-breaking antimicrobial peptides. Secondly, the novelty of the screened sequences was verified by comparing the BLAST-p sequence alignment with the APD3 database. Thirdly, the average membrane contact probability P was statistically analyzed. MCP Finally, the anti-parallel dimer contact map was used to predict the membrane-breaking antimicrobial peptide sequences, and the final screening score F score The sequences were sorted from large to small, and the top 20 sequences were selected for wet experiment verification, and 7 antimicrobial peptide sequences with effective antibacterial effects were successfully found, confirming the effectiveness of the method of the present invention.

[0061] The names and sequences of the 7 membrane-breaking antimicrobial peptides are as follows:

[0062] L1——KVKVLIIRKLTRHIAKLKSKK

[0063] L2——GFKSKLEQISGICRPLFAAMWK

[0064] L3——GPIHRLINPFFNLLQGIFTN

[0065] L4——AKLVALVNKFIGYLKQNTYC

[0066] L5——KKIASIIYAHVKALRARKILDNLKRLAANRP

[0067] L6——KLAKLLKLLLWAQNELDQKK

[0068] L7——IYKILQLFHKRFKKGFFG

[0069] As shown in Table 3, all 7 antimicrobial peptides were positively charged, which was consistent with the basic characteristics of common antimicrobial peptides. The sequence lengths ranged from 18 to 31.

[0070] The hydrophobicity of the sequence was measured by GRAVY (Grand Average of Hydropathicity). The average hydrophobicity of each amino acid in each sequence was above -0.40, which is consistent with the hydrophobic characteristics of membrane-permeable antimicrobial peptides. In addition, from the perspective of the proportion of hydrophilic residues in the sequence, the proportion of hydrophilic amino acids such as serine (Ser), threonine (Thr), cysteine ​​(Cys), asparagine (Asn), glutamine (Gln) and tyrosine (Tyr) was less than 50%, and the average hydrophilicity of the 7 sequences was 0.38, which is also consistent with the relatively hydrophobic characteristics of membrane-permeable antimicrobial peptides. In terms of sequence similarity, the 7 sequences were compared with the known antimicrobial peptide sequences in the APD3 database by BLAST-p, and the expected values ​​of their sequence alignments were all large (no similar sequences were found for L6 and L7), which can prove the novelty and uniqueness of the 7 sequences as antimicrobial peptides.

[0071] Table 3. Physicochemical properties of 7 antimicrobial peptides

[0072]

[0073] (I) Secondary structures of 7 antimicrobial peptides

[0074] Seven membrane-breaking antimicrobial peptides were synthesized by Qiangyao Biotechnology Co., Ltd. using a solid phase synthesis method, and the purity was verified to be above 95%. The secondary structure of the antimicrobial peptide can be determined by circular dichroism experiments. In this example, a Chirascan V100 circular dichroism spectrometer and a quartz cuvette with a 1 mm optical path length were used to determine the secondary structure of the membrane-breaking antimicrobial peptide. The buffer in the system was 20 mM Tris, 200 mM KCl, pH 7.4, and the final concentration of the antimicrobial peptide was 15 μM. The liposome membrane component used to simulate the bacterial cell membrane was set to DOPE: DOPG = 3:1, with a final concentration of 450 μM. The CD signal was measured at a speed of 1 nm / s in the wavelength range of 200 nm to 260 nm. Figure 3 The distribution of the mean residue ellipicity of 7 antimicrobial peptides with wavelength is shown. Each curve is the average of three spectra, and the data is smoothed in Graphpad Prism10 software. Among them, L2, L5, and L6 have a high proportion of α-helices, and the curves conform to the typical distribution of α-helices, that is, there are negative peaks around 208nm and 222nm. The secondary structures of the remaining sequences are also analyzed to contain at least 28% α-helices (statistics are implemented by the K2D algorithm on Dichroweb).

[0075] (II) Minimum inhibitory concentration of 7 antimicrobial peptides

[0076] The minimum inhibitory concentration (MIC) of antimicrobial peptides was determined by the broth microdilution method. The selected strains included four Gram-negative bacteria (Acinetobacterbaumannii ATCC 17978, Acinetobacter baumannii ATCC 19606, Escherichia coli ATCC 25922, Escherichia coli DH5α) and seven Gram-positive bacteria (Staphylococcus aureus ATCC 25923, Staphylococcus aureus ATCC 6538, Bacillus subtilis ATCC 6633, Enterococcus faecalis ATCC 29212, Micrococcus luteus CMCC(B)28001, Staphylococcus capitis ATCC27842, Staphylococcus hominis ATCC 27844). The bacteria were cultured at 37°C overnight and the bacterial solution was diluted to 5×10 5 cfu / mL, and then add 99μL of bacterial suspension to each well of the 96-well plate. The antimicrobial peptide was dissolved in cation-adjusted Mueller-Hinton medium with an initial concentration of 1mg / mL, and then diluted in a two-fold gradient concentration ranging from 0.125μg / mL to 128μg / mL, and 1μL was added to each well. The 96-well plate was placed in a 37°C incubator and incubated for 20 hours. Observe and record the lowest peptide concentration at which no bacterial growth was observed, which is the minimum inhibitory concentration (MIC) of the peptide.

[0077] The experimental results are shown in Table 4. All 7 peptides can effectively inhibit bacterial growth, and 6 peptides (except L4) have antibacterial properties against multiple strains, with the most effective antibacterial concentration reaching 2 μg / mL. Among them, L1, L5, L6, and L7 showed strong antibacterial ability against both Gram-positive and Gram-negative bacteria, and had excellent minimum inhibitory concentrations for common strains.

[0078] Table 4. Minimum inhibitory concentration experimental results of 7 antimicrobial peptides against 7 Gram-positive bacteria and 4 Gram-negative bacteria

[0079]

[0080] (III) Hemolytic activity of 7 antimicrobial peptides

[0081] Test the hemolytic activity of antimicrobial peptides on rat erythrocytes. Fresh rat blood was centrifuged at 500 g for 5 minutes, and the erythrocyte pellet was retained and resuspended in PBS buffer (pH 7.4). The erythrocyte solution was diluted to 1×10 8 The cells were added to a 96-well plate at a density of 10 cells. The antimicrobial peptides were dissolved in PBS and added to the wells after two-fold gradient dilution, with the final concentration ranging from 0.125 μg / mL to 128 μg / mL. 10 mg / mL Triton X-100 was used as a positive control to induce 100% hemolysis, while PBS was used as a blank control. After incubation for 1 hour, centrifugation was performed at 500g for 5 minutes. Subsequently, the absorbance (OD) of the supernatant was measured at a wavelength of 450 nm.

[0082] The formula for calculating the percentage of hemolysis is: (OD 450,peptide -OD 450,PBS ) / (OD 450,positive -OD 450,PBS ), where OD 450,peptide Represents the absorbance at 450nm of the well treated with the current concentration of antimicrobial peptide, OD 450,PBS Represents the absorbance of the wells treated with PBS solution at 450nm, OD 450,positive Table 5 reports the half maximum hemolysis concentration (HC) of each peptide. 50 ) and the minimum peptide concentration (Minimum Hemolytic Concentration, MHC) value that causes 10% hemolysis. The experimental data were analyzed by Graphpad Prism software 50 Fitting analysis was performed. Each test was repeated three times. The results are shown in Table 5. The half-maximum hemolytic concentrations of the seven peptides were all higher than 128 μg / mL, and the lowest peptide concentration that caused 10% hemolysis was also higher than 128 μg / mL. The above data all prove that the antimicrobial peptides in the present invention have extremely low hemolytic properties and good biosafety.

[0083] Table 5. Hemolytic activity of 7 antimicrobial peptides on rat erythrocytes

[0084]

[0085] (IV) Cytotoxicity of 7 antimicrobial peptides

[0086] The cytotoxicity of seven antimicrobial peptides to HEK293T cells was evaluated using the Alamar Blue method. HEK293T cells were cultured in DMEM medium containing 10% fetal bovine serum and placed at 37°C and 5% CO 2incubated in an environment. Amphotericin B and Staurosporine were used as controls in the test. The antimicrobial peptides were dissolved in phosphate buffered saline (PBS) at pH 7.4, and the final test concentration range was from 1μg / mL to 128μg / mL. After the cells were co-treated with the antimicrobial peptides for 72 hours, 10μL of Alamar Blue reagent was added to each well. Subsequently, the 96-well plate was incubated in dark conditions for 4 hours, and the absorbance (OD) was measured at a wavelength of 570nm. The formula for calculating the cell growth inhibition rate is: (OD 570,peptide -OD 570,PBS ) / (OD 570,staurosporine -OD 570,PBS ), where OD 570,peptide Represents the absorbance of the antimicrobial peptide treated well at 570 nm, OD 570,PBS Represents the absorbance at 570 nm of the wells treated with PBS solution only, OD 570,staurosporine Represents the absorbance of the wells treated with staurosporine at 570 nm. Half Maximal Cytotoxic Concentration (CC 50 ) was fitted and analyzed using Graphpad Prism software. Each experiment was repeated three times.

[0087] As shown in Table 6, the cytotoxicity of the seven antimicrobial peptides was very weak, all higher than the maximum test value of 128 μg / mL, and their safety was in sharp contrast to that of the control group. 50 It performs better than the polyene antifungal drug amphotericin B.

[0088] Table 6. Cytotoxicity of 7 antimicrobial peptides to HEK293T cells

[0089]

[0090] (V) Membrane disruption activity of 7 antimicrobial peptides

[0091] The membrane-disrupting activity of antimicrobial peptides was determined by peptide-induced liposome leakage studies. To prepare liposomes encapsulated with 5(6)-carboxyfluorescein (CF), a lipid membrane composed of 75% POPE and 25% POPG was first prepared, a condition close to the main components of bacterial membranes. Subsequently, the lipid membrane was hydrated in HEPES buffer containing 50 mM CF. The liquid was frozen in liquid nitrogen and then thawed in a 37°C water bath, and this freeze-thaw process was repeated 10 times. The hydrated lipid membrane was extruded through a 100 nm polycarbonate filter (Whatman) at least 21 times using a Mini-Extruder device (Avanti Polar Lipids) to generate liposomes. Subsequently, a PD10 column (GE Healthcare) was used to remove unencapsulated CF.

[0092] In the liposome leakage experiment, CF-encapsulated liposomes (lipid concentration of 5 μg / mL) were incubated with 100 nM peptide protein in HEPES buffer (20 mM HEPES, pH 7.4 and 150 mM NaCl) at room temperature for 1 hour. The emission fluorescence intensity I at 517 nm (excitation wavelength of 494 nm) was recorded using a spectrophotometer (U3900H, Hitachi). Before adding the peptide, the background fluorescence intensity I was monitored. 0 After the incubation, to measure the complete release of CF, 0.2% Triton X-100 was added to the mixture, and the fluorescence intensity I was recorded. m The leakage percentage is calculated according to the following formula: Leakage = [(I – I 0 ) / (I m –I 0 )]×100%.

[0093] The experimental results are shown in Figure 4 , the histogram shows the leakage of CF from liposomes after the addition of peptides, and the leakage is expressed as a percentage of the leakage induced by 0.2% Triton X-100 detergent. The dotted line in the figure represents the background leakage of liposomes incubated only with equal concentrations of buffer to control the effect of buffer. The data are expressed as mean ± standard error (SEM), n = 3 independent measurements. All 7 antimicrobial peptides can cause significant leakage of liposomes, representing a high membrane-breaking activity. The higher liposome leakage ratio verifies the ability of antimicrobial peptides to adsorb to the membrane and break the membrane.

Claims

1. A method for predicting and screening membrane-permeable antimicrobial peptides, comprising the following steps: A. Develop a sequence classifier based on machine learning, including: A1. Organize the public database to form positive and negative data sets, where the positive data set is the antimicrobial peptide sequence data set, and the negative data set is the peptide sequence data set without antibacterial function; A2. Use the pre-trained BERT language model to encode the peptide sequence sorted in step A1; A3. Input the encoded polypeptide sequence into the support vector machine to perform classification prediction training to realize a sequence classifier; B. Use the sequence classifier developed in step A to perform high-throughput screening on the proteome. The screening process includes: B1. predict whether the polypeptide sequence is membrane-contactable, if yes, proceed to step B2; B2. Predict whether the secondary structure of the polypeptide sequence is mainly helical, if yes, proceed to step B3; B3. Use the sequence classifier developed in step A to predict whether the input polypeptide sequence is an antimicrobial peptide sequence; C. Predict the properties of the candidate antimicrobial peptide sequences screened in step B, including: C1. Screen sequences based on physicochemical properties, retaining sequences with positive net charge and relatively hydrophobic; C2. Determine the novelty of the sequence; C3. Predict the average membrane contact probability of the peptide sequence; C4. Predict the oligomerization possibility of polypeptide sequences to form antiparallel dimers; D. Based on the sequence features obtained in step C, select polypeptide sequences that are novel, have a high average membrane contact probability, and have a high probability of oligomerization to form antiparallel dimers as candidate membrane-permeable antimicrobial peptides.

2. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, characterized in that: Step A1: sort out peptide sequences of a certain length range, positively charged, and predicted to have membrane contact ability from the databases APD3, DBAASP, YADAMP, and SATPdb to form a positive data set; use keywords to search sequences in the Uniref database, remove all sequences related to "antimicrobial peptides", and obtain peptide sequences of a certain length range to form a negative data set; the ratio of the number of sequences in the positive and negative data sets is 1:5 to 1:

10.

3. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 2, characterized in that: The length of the polypeptide sequences constituting the positive and negative data sets in step A1 is 10-40 amino acid residues.

4. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, characterized in that: Step A2 uses the pre-trained ProtBERT-BFD model to encode the peptide sequence.

5. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, characterized in that: In step B1, the criterion for determining whether the residue can contact the membrane is that the predicted membrane contact probability value of at least one residue in the sequence is above 0.

2.

6. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, characterized in that: In step B2, the criterion for determining whether the secondary structure is mainly a helix is ​​that the number of residues predicted to be helices with a probability value greater than 0.5 accounts for 50% or more of the sequence length.

7. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, characterized in that: In step C3, the average membrane contact probability P of the polypeptide sequence MCP Calculated by the following formula: Among them, n mem is the number of amino acid residues with a predicted membrane contact probability (MCP) value above 0.2, and L is the length of the sequence.

8. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 1, wherein: In step C4, the normalized value R of the antiparallel dimer is standard To predict the oligomerization probability of a polypeptide sequence to form an antiparallel dimer, the normalized value R standard The calculation formula is as follows: Among them, R 2 is the coefficient of determination of the linear regression in the contact map prediction of antiparallel residues, and L is the length of the sequence.

9. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 8, characterized in that: In step D, the average membrane contact probability and the normalized value for antiparallel dimers are combined to score the peptide sequence: F score =P MCP ×R standard Among them, P MCP is the average membrane contact probability, R standard is the normalized value of the antiparallel dimer, F score That is, the score of membrane-permeable antimicrobial peptides is in the range of (0, 1).

10. The method for predicting and screening membrane-permeable antimicrobial peptides according to claim 9, characterized in that: Step D: Select novel membrane-permeable antimicrobial peptides and score them F. score high sequence.

Images