Gradient recognition method for antibacterial peptides based on protein sequence multi-feature fusion

Abstract

The application discloses an antibacterial peptide gradient recognition method based on protein sequence multi-feature fusion, belongs to the technical field of bioinformatics and protein sequence analysis, and comprises the following steps: reading positive and negative antibacterial peptide data sets, extracting 17-dimensional features and constructing a benchmark positive sample feature matrix after sequence recognition, verification and length screening; calculating a mean vector, a standard deviation vector and a regularized inverse covariance matrix, defining a general scoring function to calculate a fusion score, determining a positive threshold and a final threshold, and forming a persistent model file; extracting features from a to-be-tested sequence and calculating a fusion score, comparing the fusion score with the threshold, and outputting a three-gradient classification label. The application adaptively determines a grading boundary in a data-driven manner by constructing a benchmark positive sample feature matrix and defining a general scoring function, achieves the antibacterial peptide recognition effect of establishing an objective quantitative standard and realizing grading, and solves the problems of scattered feature representation, single scoring mechanism and grading relying on artificial experience in the prior art.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and protein sequence analysis, and in particular to a gradient identification method for antimicrobial peptides based on the fusion of multiple features of protein sequences. Background Technology

[0002] In the research and development of antimicrobial peptides, gradient identification based on protein sequence information and multi-feature fusion strategies can provide technical support for the accurate screening of candidate antimicrobial peptides.

[0003] Existing antimicrobial peptide screening systems perform grading assessments by calculating physicochemical parameters or use dual-channel neural networks to fuse sequence and structural features for prediction. These systems tend to focus on dimensional superposition during feature fusion, lacking in-depth analysis of the relationships between features. The scoring process relies on linear weighting or a single fusion method, allowing some dimensional biases to be masked by compensation from other dimensions. Classification decisions depend on fixed thresholds or manually preset parameters. The grading criteria lack adaptability to the distribution patterns of sample data, making it difficult to balance screening sensitivity and specificity.

[0004] For example, Chinese invention patent CN111415706B discloses a method and application for high-throughput mining and screening of antimicrobial peptides. This method includes: calculating the physicochemical and structural parameters of amino acid sequences based on naturally occurring or artificially designed base sequences or amino acid sequences; simultaneously performing graded screening using multiple limiting conditions; and then evaluating and analyzing the obtained polypeptide sequences to screen out antimicrobial peptide sequences with antimicrobial activity. This invention screens antimicrobial peptides using some or all of 59 parameters and obtains antimicrobial peptides with better activity after using the MD Index scoring system.

[0005] For example, the Chinese invention patent with announcement number CN120432006B discloses an antimicrobial peptide identification method and system based on a sequence-structure dual-channel neural network, which includes: splicing amino acid features extracted by ProtT5 and amino acid-level handcrafted features to obtain peptide embeddings and inputting them into a sequence channel composed of multiple Transformer blocks to extract peptide sequence features; using ESM-Fold to predict the three-dimensional structure of the peptide to construct an adjacency graph, using the amino acid features obtained by ESM-2 as node features of the adjacency graph, and performing layer-by-layer extraction and enhancement by fusing multi-head graph attention, residual network, layer normalization and feedforward neural network structure channels; performing max pooling and splicing on the sequence features and structure features, and then using them for antimicrobial peptide prediction.

[0006] While existing antimicrobial peptide identification technologies have broken through the single-feature recognition mode and can achieve basic multi-feature combination analysis and simple hierarchical judgment, the overall technical solution still has many inherent defects: First, existing antimicrobial peptide identification multi-feature combinations mostly adopt a scattered and disordered simple superposition form, lacking a standardized and systematic dedicated feature architecture. The feature dimension coverage is limited, and it cannot simultaneously take into account the protein sequence distribution pattern and multi-dimensional physicochemical structural properties, resulting in weak overall feature characterization specificity and completeness. Second, existing antimicrobial peptide identification multi-feature fusion methods are all shallow feature splicing or single-model voting fusion, with relatively simple computational logic. They lack the deep computational capabilities of independent scoring and complementary collaboration among multiple models, and the effect of distinguishing positive and negative samples is limited. First, existing antimicrobial peptide identification technologies have significant limitations, making it difficult to further improve the accuracy of antimicrobial peptide identification. Second, current antimicrobial peptide classification identification only uses a two-level activity division, resulting in a single identification level that cannot adapt to diverse practical application scenarios. The classification thresholds are generally set manually based on experience, without considering objective data patterns such as sequence length distribution and score distribution. Furthermore, gradient determination and confidence rating are independent of each other, making it impossible to achieve refined and hierarchical accurate identification of antimicrobial peptides. Third, existing antimicrobial peptide identification technologies can only achieve local optimization of a single module. The various technical links are isolated and disconnected, failing to form a coherent integrated workflow. This makes it impossible to simultaneously meet the requirements of screening large batches of protein sequences and high-precision identification, resulting in poor engineering adaptability of the overall technical solution. Summary of the Invention

[0007] To address the technical problems of fragmented feature representation, simplistic scoring mechanisms, and reliance on manual experience in existing technologies, this invention provides a gradient identification method for antimicrobial peptides based on multi-feature fusion of protein sequences. This method includes: S1: Read the positive antimicrobial peptide dataset and the negative non-antimicrobial peptide dataset, and perform independent sequence entry identification, sequence content legality verification and length filtering on these two types of benchmark sequence files to obtain the benchmark positive antimicrobial peptide qualified sequence set and the benchmark negative non-antimicrobial peptide qualified sequence set.

[0008] The baseline positive antimicrobial peptide qualified sequence set and the baseline negative non-antimicrobial peptide qualified sequence set were subjected to 17-dimensional full-dimensional antimicrobial peptide feature quantification extraction, and the baseline positive sample feature matrix and the baseline negative sample feature matrix were constructed.

[0009] S2: Calculate the mean vector, corrected standard deviation vector, weight coefficient vector, and regularized inverse covariance matrix based on the feature matrix of the benchmark positive samples, and define a general scoring function for calculating the fusion score.

[0010] A general scoring function is used to calculate the positive and negative nonlinear fusion scores respectively. A positive threshold is determined based on the positive nonlinear fusion score, and a negative threshold is determined based on the negative nonlinear fusion score. The positive nonlinear fusion score sequence is arranged in ascending order of value, and the positive low threshold is obtained. The final threshold is determined based on the negative threshold and the positive low threshold.

[0011] S3: Save the mean vector, the corrected standard deviation vector, the weight coefficient vector, the regularized inverse covariance matrix, the positive threshold, and the final threshold to form a persistent model file.

[0012] S4: For a single protein sequence to be tested, the same feature extraction process is called to obtain a 17-dimensional original feature vector. The fusion score to be tested is calculated using a persistent model file. The fusion score to be tested is compared with the positive threshold and the final threshold, and the first classification label, the second classification label, or the third classification label is output.

[0013] S5: Batch import the protein sequence files to be tested, call S4 one by one to get the prediction results, record and output the first output file, the second output file and the third output file.

[0014] S6: Declare the versions of the programming language interpreter and libraries that the runtime environment depends on, and specify the consistency between input and output files.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. The antimicrobial peptide gradient identification method based on multi-feature fusion of protein sequences provided by this invention extracts seventeen-dimensional features covering four major categories: physicochemical properties, structure, sequence, and distribution from benchmark positive samples and constructs a feature matrix. Based on this feature matrix, the mean vector, the corrected standard deviation vector, and the regularized inverse covariance matrix are calculated as a unified scoring benchmark. A general scoring function is used to nonlinearly fuse the local single-dimensional feature similarity of the test sequence with the global multi-dimensional distribution consistency, thereby obtaining a fusion score that takes into account both individual feature differences and overall distribution position. This achieves an objective quantitative assessment of the antimicrobial activity potential of the test protein sequence, effectively solving the problem that the scoring methods of single dimensions or simple linear combinations in the prior art cannot fully reflect the synergistic effect of multiple features of antimicrobial peptides, resulting in limited identification accuracy.

[0016] 2. This invention designs a general scoring function as a nonlinear fusion of weighted similarity scoring and Mahalanobis distance density scoring. The weighted similarity scoring uses the reciprocal of the corrected standard deviation as the weight coefficient to perform weighted summation and normalization of the Gaussian similarity of each feature, highlighting key features that are highly conserved in positive samples. The Mahalanobis distance density scoring eliminates the redundant influence of correlation between features by regularizing the inverse covariance matrix. This ensures that abnormal deviations in any dimension cannot be fully compensated by normal performance in other dimensions, thereby achieving refined scoring of candidate antimicrobial peptide sequences. This effectively avoids the problem of the shortcomings of linear scoring methods in the prior art being masked, leading to a high false positive rate.

[0017] 3. This invention sets the positive threshold as the high quantile of the fusion score for positive samples and the negative threshold as the high quantile of the fusion score for negative samples, while introducing a low quantile for positive samples as a reference. The larger of the negative threshold and the low quantile for positive samples is taken as the final threshold. Simultaneously, when there is an abnormal intersection between the candidate final threshold and the positive threshold, a preset interval value is automatically used for correction. This ensures that the three-gradient classification boundary is based on the fusion score distribution of positive and negative samples, combined with a preset false positive tolerance rate and preset percentiles. Furthermore, it possesses fault tolerance capabilities for abnormal operating conditions, thus achieving a tiered response to screening requirements of varying stringency. This effectively solves the problem in existing technologies where the threshold determination relies on manual experience or external parameter coverage, making it difficult to balance sensitivity and specificity. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the antimicrobial peptide grading and identification method provided in the embodiments of this application; Figure 2 This is a schematic diagram of 17-dimensional protein sequence feature extraction provided in an embodiment of this application; Figure 3 This is a schematic diagram of dual-score fusion and gradient threshold recognition provided in an embodiment of this application; Figure 4 The sequence overlap map of high-confidence AMP provided in the embodiments of this application; Figure 5 A high-confidence AMP quantity map provided for embodiments of this application; Figure 6 A comprehensive score distribution chart provided for embodiments of this application; Figure 7Weighted score comparison chart provided for embodiments of this application; Figure 8 Sequence length distribution diagram provided for embodiments of this application; Figure 9 The average comprehensive score chart provided for the embodiments of this application. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Embodiment 1 of the present invention provides a gradient identification method for antimicrobial peptides based on multi-feature fusion of protein sequences, such as... Figure 1 The flowchart of the antimicrobial peptide grading and identification method is shown. The preset parameters involved in this embodiment and the basis for their setting are as follows: The fixed angle value of the projection rotation is preset to the standard rotation value for detecting α-helical amphiphilism, so as to accurately extract the amphiphilic features of the sequence; at the same time, another fixed angle value of the projection rotation is preset to the standard rotation value for detecting β-sheet amphiphilism, which is used for amphiphilicity calculation, so as to accurately extract the β-sheet amphiphilic features of the sequence and ensure that amphiphilism and hydrophobic moment use different projection rotation angles to eliminate feature redundancy. The mapping scaling factor is preset to a scalar value that ensures the scaled coefficient of variation falls within the middle of a preset numerical range, guaranteeing that the cation homogeneity characteristic value is reasonably mapped to a bounded interval. The first preset coefficient for alanine, the second preset coefficient for valine, and the third preset coefficient shared by isoleucine and leucine in the fat index are set according to the coefficient values ​​in the standard formula for the fat index field. The percentage scaling factor is specifically set based on engineering practice to conform to the industry conventions for fat index calculation. The preset minimum positive number is set based on the precision of computer floating-point operations to prevent abnormal subsequent division operations caused by a zero standard deviation. The preset minimum length threshold is set according to the industry's general screening standards to remove excessively short sequences that cannot stably extract spatial structure features, sequence motif features, and distribution pattern features. The preset minimum sample size threshold is set according to statistical robustness requirements to determine whether the number of rows in the feature matrix of the benchmark positive sample is sufficient to support... Reliable calculation of mean, standard deviation, and covariance matrix; preset false positive tolerance rate is manually specified by the operator based on the expected stringency of the results, used to control the quantile level corresponding to the negative threshold; the first and second fixed indices are determined based on the classification performance evaluation results on the benchmark positive antimicrobial peptide qualified sequence set and the benchmark negative non-antimicrobial peptide qualified sequence set, to balance local feature similarity and global distribution consistency; the preset regularization coefficient λ is set as a small positive constant based on engineering practice experience, used to ensure the invertibility of the covariance matrix and improve numerical stability; the first, second, and third preset percentiles are set according to the need to distinguish between high-scoring boundaries of positive samples, high-scoring boundaries of negative samples, and low-scoring boundaries of positive samples; preset maximum constant, preset minimum constant, and preset constant default values ​​are set according to the range of computer numerical representation, used to handle the default threshold assignment in the case of an empty set.

[0023] The processing flow of this method may include the following steps: S1: Benchmark sequence selection and feature matrix construction.

[0024] Read the local pre-set positive antimicrobial peptide dataset and negative non-antimicrobial peptide dataset, which are composed of protein sequences. The two types of sequences are collectively referred to as the baseline sequences. The baseline sequence file name in FASTA (FAST-All, a text-based bioinformatics storage format for representing nucleotide or protein sequences) format is used as the source file name and is only used for model parameter statistics and recognition rule construction.

[0025] During the reading process, the baseline sequence file is automatically split to identify individual sequence entries, and the following operations are performed on each sequence entry: The sequence identifier is extracted from the description line starting with a greater than sign. During extraction, if a space immediately following the greater than sign exists, the string after the greater than sign and before the first space is extracted. If no space exists, the string is truncated from the greater than sign to the end of the description line, or to before the first tab, bar, or comma separator, and used as the sequence identifier. The default truncation rule is to truncate from the greater than sign to the end of the description line, or to before the first space, tab, bar, or comma separator. The sequence line string immediately following the description line is extracted. This string consists of amino acid residue abbreviations arranged linearly from the amino terminus to the carboxyl terminus, and is used as the sequence content. The filename of the baseline sequence file is directly obtained as the common source filename for all sequences in that file. Simultaneously, the sequence content is validated for legality, and abnormal sequences containing missing characters, garbled characters, or invalid characters are removed. The resulting sets are the baseline positive antimicrobial peptide screening sequences and the baseline negative non-antimicrobial peptide screening sequences.

[0026] Then, for the baseline positive antimicrobial peptide screening sequence set and the baseline negative non-antimicrobial peptide screening sequence set, the sequence content of each sequence is a string. The length of each amino acid contained in the sequence is counted, and the number of abbreviation letters of the amino acid residues contained in the string is counted. This count value is the sequence length.

[0027] Based on the common screening criteria in the field, short sequences with amino acid lengths less than the preset minimum length threshold are removed. In this embodiment, the preset minimum length threshold is set to 10, which makes it difficult to stably extract spatial structure features, sequence motif features, and distribution pattern features of such short sequences under the 17-dimensional feature system adopted by this model. This can easily lead to increased feature noise in subsequent identification operations. After screening, the qualified sequence set of benchmark positive antimicrobial peptides and the qualified sequence set of benchmark negative non-antimicrobial peptides are retained.

[0028] Using a set of qualified positive antimicrobial peptide sequences as the processing object, 17-dimensional full-dimensional antimicrobial peptide features were extracted from each sequence. The 17 features cover 17 features in 4 major categories: physicochemical, structural, sequence, and distribution. This forms a complete antimicrobial peptide characterization space, including: basic physicochemical features, spatial structural features, sequence motif features, and distribution pattern features.

[0029] Among them, the basic physicochemical characteristics include net charge, charge density, total average hydrophilicity, isoelectric point and aliphatic index; the spatial structure characteristics include hydrophobic moment, amphiphilicity, proportion of hydrophobic residues, proportion of cationic residues and proportion of aromatic residues; the sequence motif characteristics include hydrophobicity of amino-terminal pentapeptides, hydrophobicity of carboxyl-terminal pentapeptides, number of hydrogen bond donors, number of hydrogen bond acceptors and dipeptide preference; and the distribution pattern characteristics include disorder tendency and uniformity of cation distribution.

[0030] like Figure 2 The diagram shows the overall process for extracting 17-dimensional protein sequence features. For the input protein sequence, it is first determined whether the sequence is empty. If it is empty, all 17 feature values ​​are set to zero. If the sequence is not empty, the 17 features in four categories—basic physicochemical features, spatial structure features, sequence motif features, and distribution pattern features—are calculated in sequence. After arranging all feature values ​​in a fixed order, the original 17-dimensional feature vector of the sequence is output.

[0031] The method for quantitative extraction of 17-dimensional antimicrobial peptide features is as follows: Based on engineering practice, the fixed angle value of the projection rotation is preset to the standard rotation value for detecting α-helical amphiphilicity, the mapping scaling factor is preset to a scalar value that makes the coefficient of variation fall into the middle of the preset value range after scaling, the first preset coefficient corresponding to alanine, the second preset coefficient corresponding to valine, and the third preset coefficient shared by isoleucine and leucine in the fat index are set according to the coefficient values ​​in the standard formula of the fat index field, and the percentage scaling factor is preset to one.

[0032] For protein sequences from which features are to be extracted, first determine whether the sequence is empty. If it is empty, set all seventeen feature values ​​of net charge, charge density, total average hydrophilicity, isoelectric point, lipid index, hydrophobic moment, amphiphilicity, proportion of hydrophobic residues, proportion of cationic residues, proportion of aromatic residues, hydrophobicity of amino-terminal pentapeptide, hydrophobicity of carboxyl-terminal pentapeptide, number of hydrogen bond donors, number of hydrogen bond acceptors, dipeptide preference, disorder tendency, and cation uniformity to zero and end feature extraction for that sequence. If the sequence is not empty, continue with subsequent operations.

[0033] When calculating the net charge, each amino acid residue in the protein sequence is traversed. Based on the pre-constructed amino acid charge parameter table, the category of the residue is determined as a positively charged residue, a negatively charged residue, or a neutral residue. The negative logarithm of the acid dissociation constant of the dissociable group of the residue's side chain and the hydrogen ion activity index under the preset physiological conditions are combined to calculate the charge contribution value of the residue under the preset physiological conditions. The charge contribution values ​​of each residue are accumulated, and the accumulated result is rounded to the preset decimal place to obtain the net charge.

[0034] The amino acid charge parameter table includes the standard charge state of each amino acid residue under preset physiological conditions and the acid dissociation constant of its side chain dissociable groups.

[0035] When calculating charge density, the absolute value of the net charge is divided by the total number of amino acid residues in the sequence to obtain the characteristic value of charge density.

[0036] To calculate the hydrophobic moment, each amino acid residue in the sequence is traversed, and its hydrophobic value is retrieved from a pre-constructed amino acid hydrophobicity scaling table. From the beginning to the end of the sequence, each residue is sequentially assigned an increasing projection angle with a fixed projection angle step size. The product of the hydrophobic value and the corresponding sine of the projection angle, as well as the product of the hydrophobic value and the corresponding cosine of the projection angle, are accumulated. The square of the accumulated sine and cosine values ​​is then summed and taken as a square root. The square root result is divided by the total number of amino acid residues in the sequence to obtain the hydrophobic moment. For amphiphilicity, the projection angle is set to a preset angle for detecting β-sheet amphiphilicity. The accumulated sine and cosine values ​​are recalculated, and then normalized to the same square root and length as the hydrophobic moment.

[0037] When calculating cation uniformity, the position indices of amino acid residues belonging to a preset cation residue set are extracted from the sequence to form a position index set. This preset cation residue set consists of lysine, arginine, and histidine. If the number of elements in this set is less than two, the cation uniformity is assigned a value of zero. If the number of elements is greater than or equal to two, the interval values ​​of adjacent position indices are calculated to obtain an interval value sequence. The mean and standard deviation of the interval value sequence are calculated. If the standard deviation is zero, the cation uniformity is assigned a value of 1. If the standard deviation is not zero, the coefficient of variation is calculated (i.e., the standard deviation divided by the mean). The coefficient of variation is then divided by a mapping scaling factor to obtain an intermediate mapping value. The smaller value between the intermediate mapping value and 1 is taken, and then 1 is subtracted from this smaller value to obtain the cation uniformity.

[0038] When calculating the proportion of cationic residues, the sequence is traversed to count the number of amino acid residues belonging to the preset cationic residue set. This number is then divided by the total number of amino acid residues in the sequence to obtain the characteristic value of the proportion of cationic residues.

[0039] When calculating dipeptide preference, a target dipeptide combination set is pre-constructed, consisting of eight dipeptide combinations: lysine-arginine, arginine-lysine, alanine-lysine, lysine-alanine, lysine-lysine, arginine-arginine, phenylalanine-lysine, and lysine-phenylalanine. For each pair of adjacent amino acid residues in the sequence, if the residue pair belongs to the target dipeptide combination set, the total dipeptide hit count is incremented by one. After the iteration is complete, the total dipeptide hit count is divided by the total number of amino acid residues in the sequence minus one to obtain the dipeptide preference feature value. If the total number of amino acid residues in the sequence is less than two, the dipeptide preference feature value is assigned to zero.

[0040] When calculating the disorder tendency, the sequence is traversed to count the number of amino acid residues falling into a preset disorder tendency set. This number is then divided by the total number of amino acid residues in the sequence to obtain the disorder tendency characteristic value. When calculating the adipose index, the sequence is traversed to count the occurrences of alanine, valine, isoleucine, and leucine. The occurrences of alanine are multiplied by their corresponding preset coefficients, the occurrences of valine are multiplied by their corresponding preset coefficients, and the sum of the occurrences of isoleucine and leucine is multiplied by their corresponding preset coefficients. These three results are then added together, divided by the total number of amino acid residues in the sequence, and multiplied by a percentage scaling factor to obtain the adipose index.

[0041] The module calls upon a pre-packaged protein analysis module from a third-party biological computing library. The current sequence is input, and the module internally calculates the overall average hydrophobicity based on the Kyte-Doolittle Hydrophobicity Scale (a scale that assigns a specific hydrophobicity index to each amino acid based on its side chain hydrophobicity). The negative value of the overall average hydrophobicity is used as the overall average hydrophilicity characteristic value. The isoelectric point is calculated based on the amino acid acid dissociation constant. The proportion of hydrophobic residues is statistically analyzed based on a preset hydrophobicity threshold and a preset aromatic residue set. The overall average hydrophilicity, isoelectric point, proportion of hydrophobic residues, and proportion of aromatic residues are obtained. If a null value is returned, the corresponding characteristic value is set to zero. The same library is then called upon again, inputting the current sequence. This library performs theoretical statistics at the sequence level based on the preset number of donor atoms and acceptor atoms corresponding to the amino acid residue type, obtaining the number of hydrogen bond donors and acceptors. The number of hydrogen bond donors and acceptors is obtained. If a null value is returned, the corresponding characteristic value is set to zero. The hydrophobicity of the amino-terminal and carboxyl-terminal pentapeptide was obtained by extracting five residues from the sequence and calculating the average hydrophobicity of each terminal fragment.

[0042] All the calculated feature values ​​are arranged in a fixed order: net charge, charge density, total average hydrophilicity, isoelectric point, aliphatic index, hydrophobic moment, amphiphilicity, proportion of hydrophobic residues, proportion of cationic residues, proportion of aromatic residues, hydrophobicity of amino-terminal pentapeptide, hydrophobicity of carboxyl-terminal pentapeptide, number of hydrogen bond donors, number of hydrogen bond acceptors, dipeptide preference, disorder tendency, and cation uniformity, forming the original 17-dimensional feature vector of the sequence.

[0043] The seventeen feature values ​​corresponding to a single benchmark positive antimicrobial peptide qualified sequence are arranged in a predetermined dimensional order to form an independent 17-dimensional feature vector; then the feature vectors of all sequences in the set are arranged row by row and integrated to construct the benchmark positive sample feature matrix.

[0044] Using the same 17-dimensional full-dimensional antimicrobial peptide feature quantification extraction method, a 17-dimensional original feature vector was extracted from each sequence in the set of qualified negative non-antimicrobial peptide sequences. The feature vectors of all negative sequences in the set were then arranged and integrated row by row to construct the feature matrix of the benchmark negative samples.

[0045] In the baseline positive sample feature matrix and the baseline negative sample feature matrix, each row corresponds to all the feature information of a protein sequence, and each column corresponds to the same type of fixed feature dimension.

[0046] Before entering S2, the integrity of the baseline positive sample feature matrix and the baseline negative sample feature matrix are checked respectively. If the number of rows in the baseline positive sample feature matrix or the baseline negative sample feature matrix is ​​less than the preset minimum sample size threshold, an insufficient baseline sample size prompt message is output, the model parameter calculation and threshold calibration are terminated, and a persistent model file is not generated. Only when the number of rows in both types of baseline sample feature matrices reaches the preset minimum sample size threshold, does it enter S2.

[0047] S2: Statistical calculation, scoring function definition and threshold calibration.

[0048] During the threshold calibration stage, the 17-dimensional original feature vector X of known positive protein sequences is extracted row by row from the feature matrix of the benchmark positive samples. p And extract the 17-dimensional original feature vector X of known negative protein sequences row by row from the feature matrix of the benchmark negative samples. n The scoring function is used to calculate the fusion score CS for any 17-dimensional original feature vector X, and is called in the subsequent prediction execution stage to calculate the test fusion score CS of the protein sequence to be tested. t .

[0049] Based on the established baseline positive sample feature matrix, and following the fixed feature order of the 17-dimensional original feature vector, the mean μ corresponding to the i-th feature is calculated column by column. i Compared with the original standard deviation, where i = 1, 2, ..., 17; then the standard deviation σ of the i-th feature is... i The lower limit is defined as a preset minimum positive number, and in this embodiment, the preset minimum positive number is 10. -6 This avoids the standard deviation being zero due to consistent feature values, thus preventing the abnormal situation of a zero denominator in subsequent normalization calculations. The mean μ corresponding to the seventeen features is... i The mean vector μ is formed according to a fixed feature order. vec =(μ1,μ2,...,μ 17 The corrected standard deviation σ corresponding to the seventeen features is calculated. i The corrected standard deviation vector σ is formed according to a fixed feature order. vec =(σ1,σ2,...,σ 17), and according to w i =1 / σ i Calculate the weight coefficient corresponding to the i-th feature, and arrange the seventeen weight coefficients into a weight coefficient vector w according to a fixed feature order. vec =(w1,w2,...,w 17 Thus, the standardized parameters of the baseline features, which remain fixed within the same model parameter set, are obtained.

[0050] Define a general scoring function to calculate the fusion score CS for any 17-dimensional original feature vector X. For any protein sequence, this includes a baseline positive antimicrobial peptide qualified sequence, a baseline negative non-antimicrobial peptide qualified sequence, and the protein sequence to be tested. The calculation process is as follows: First, X is standardized using the Z-value by applying the fixed mean and corrected standard deviation previously obtained from the baseline positive samples. The standardized feature values ​​are then obtained using the following formula: Where, x i Let μ be the i-th original feature value in the original feature vector X of any protein sequence. i The mean vector μ vec The mean of the i-th term, σ i Let i be the i-th term of the corrected standard deviation vector, where i = 1, 2, ..., 17.

[0051] Therefore, by standardizing and transforming each original feature value in the original feature vector X of any protein sequence, a standardized feature vector Z is obtained. vec Standardized feature vector Z vec It contains 17 standardized eigenvalues ​​Z i Each original feature value x i Each corresponds to a calculated Z. i .

[0052] Furthermore, based on the corrected standard deviations σ in the original feature matrix of the benchmark positive sample... i The weight coefficients for each feature are calculated using the inverse of the standard deviation weighting method. The calculation formula is as follows: , The weight coefficients comprise seventeen feature dimensions, which, after being calculated one by one according to the formula, correspond to seventeen independent feature weight values.

[0053] Furthermore, for each standardized feature value Z, the Gaussian similarity sim is calculated one by one: , Each protein sequence has 17 Z values, which correspond to 17 single-dimensional similarity components (sim) that can be calculated.

[0054] The seventeen single-dimensional similarity components of each protein sequence are denoted as sim i From the weight coefficient vector w vec Extract the i-th weight coefficient from the sim of the same protein sequence. i with w i Weighted summation and normalization yield a weighted similarity score (WS) for this protein sequence: , For any protein sequence, calculate its original feature vector X and its 17-dimensional mean vector μ. vec Mahal distance between them: , Each sequence corresponds to a Mahalanobis distance. Where Σ -1 The method for obtaining the covariance matrix is ​​as follows: Calculate the covariance matrix based on the feature matrix of the benchmark positive samples. Add the covariance matrix to the matrix obtained by multiplying the regularization coefficient λ by the identity matrix I to obtain the regularized covariance matrix. Then, invert the regularized covariance matrix to obtain the regularized inverse covariance matrix Σ. -1 .

[0055] The purpose of regularization is to address the issue that when the feature matrix of the baseline positive samples contains highly correlated feature dimensions, or when the sample size is insufficient relative to the feature dimensions, the original covariance matrix may be singular or nearly singular, and direct inversion would lead to numerical instability. By introducing a regularization coefficient λ to diagonally load the covariance matrix, the existence and computational stability of the inverse matrix can be ensured.

[0056] Furthermore, for each protein, the calculated Mahalanobis distance value is substituted one by one, and the density score DS corresponding to that protein sequence is mapped using a chi-square distribution: , Where, χ 2 .cdf represents the cumulative distribution function of the chi-square distribution; df represents the degrees of freedom. Since the original feature vector X is a 17-dimensional feature vector, the value of df is 17.

[0057] Finally, when WS>0 and DS>0, a weighted similarity score WS and a density score DS corresponding to the same protein sequence are non-linearly fused to obtain a non-linear fusion score CS for that protein sequence. , When WS≤0 or DS≤0, CS is set to 0 directly; 0.4 and 0.6 are fixed index weights, which are determined based on the classification performance evaluation results on the benchmark positive antimicrobial peptide qualified sequence set and the benchmark negative non-antimicrobial peptide qualified sequence set. The evaluation indicators include accuracy, recall, F1 value or false positive rate, which can take into account both local feature similarity and global distribution consistency.

[0058] During the threshold calibration phase, a general scoring function was used to calculate the positive nonlinear fusion score (CS) for each known positive sequence in the baseline set of qualified positive antimicrobial peptide sequences. p And the negative nonlinear fusion score CS of each known negative sequence in the baseline set of qualified negative nonmicrobial peptide sequences. n The prediction execution phase calculates the test fusion score (CS) for each protein sequence based on the original feature matrix of the test sample. t .

[0059] like Figure 3 The diagram illustrates the dual-score fusion and gradient threshold identification. Positive sample fusion scores are sorted in ascending order, and the highest quantile is used as the positive threshold. Negative sample fusion scores are sorted in ascending order, and a negative threshold is determined based on a preset false positive tolerance rate. Simultaneously, the lowest quantile of the positive sample fusion score is used as the low positive threshold. The negative threshold and the low positive threshold are compared, and the larger one is taken as the final threshold. The positive threshold and the final threshold form the dual-threshold classification boundary, dividing the test sequence into the first gradient, second gradient, and third gradient.

[0060] First, calculate all positive nonlinear fusion scores CS. p Sort by numerical value in ascending order, and take the CS value that is in the first preset high percentile after sorting. p The value is used as a positive threshold. This positive threshold represents the strong positive boundary corresponding to high-scoring samples in the distribution of positive samples. If the set of qualified positive antimicrobial peptide sequences is empty, the positive threshold is set to a preset maximum constant, and a prompt message is output indicating that the positive threshold is invalid.

[0061] Next, a preset false positive tolerance rate is set, the specific setting of which is manually specified based on the expected stringency of the results. The second preset percentile of the preset false positive tolerance rate is subtracted from 1, and the CS score of all negative results is non-linearly fused. n Sort by numerical value in ascending order, and take the CS value that is in the second preset percentile after sorting. n The value is used as a negative threshold. This negative threshold represents the high quantile boundary of the baseline negative non-antimicrobial peptide qualified sequence set under the scoring function. Samples below this value fall into the main score range of negative samples, while samples above this value have a higher risk of being falsely judged as positive. If the baseline negative non-antimicrobial peptide qualified sequence set is empty, the negative threshold is set to a preset minimum constant, and a prompt message is output indicating that the negative threshold is invalid.

[0062] Furthermore, the fusion score (CS) of the benchmark positive antimicrobial peptide qualified sequence set was calculated. p The sequence is then sorted again in ascending order of value, and the CS value at the third preset percentile is taken. p The value is used as the positive low threshold, where the third preset percentile is lower than the first preset high percentile. The negative threshold is compared with this positive low threshold, and the larger of the two is taken as the candidate final threshold. If the candidate final threshold is less than the positive threshold, it is determined as the final threshold; if the candidate final threshold is greater than or equal to the positive threshold, the positive threshold is subtracted from the preset threshold interval to obtain the final threshold. The preset threshold interval is greater than zero and is preset based on engineering experience. An abnormal threshold interval warning message is output. When both the negative threshold and the positive low threshold are corresponding extreme constants, the final threshold is set to the preset constant default value.

[0063] Using the positive threshold and the final threshold, construct hierarchical classification rules: For any protein sequence to be tested, if the fusion score is greater than a positive threshold, the first classification label is output; if the fusion score is less than or equal to the positive threshold but greater than or equal to the final threshold, the second classification label is output; if the fusion score is less than the final threshold, the third classification label is output. The first, second, and third classification labels are all antibacterial activity potential classification labels, indicating different levels of antibacterial activity potential.

[0064] S3: Persistently save model parameters.

[0065] The various model parameters calculated by S2 are serialized and written to the same file to form a persistent model file, which constitutes a complete set of model parameters for use in subsequent prediction stages.

[0066] The model parameters include: a mean vector consisting of the mean of each feature dimension of the baseline positive sample feature matrix; a corrected standard deviation vector consisting of the corrected standard deviation of each feature dimension; a weight coefficient vector calculated from the reciprocal of the standard deviation; a regularized inverse covariance matrix calculated and inverted based on the baseline positive sample feature matrix; and the positive threshold and final threshold determined in the S2 threshold calibration stage.

[0067] S4: Tri-gradient prediction of the sequence to be tested is executed.

[0068] Read the mean vector, corrected standard deviation vector, weight coefficient vector, regularized inverse covariance matrix, positive threshold, and final threshold from the persistent model file saved in S3.

[0069] Import the raw protein sequence to be tested. When importing the raw protein sequence file, the file is automatically split and individual sequence entries are identified. The sequence identifier of the sequence to be tested is extracted from the description line of each sequence entry. The description line begins with a greater-than sign, followed by the string after the greater-than sign and before the first space, or a string truncated from the description line according to a preset truncation rule; this is the sequence identifier of the sequence to be tested. Extract the sequence line string immediately following the description line from each sequence entry. This string consists of amino acid residue abbreviations arranged linearly from the amino terminus to the carboxyl terminus; this is the protein sequence of the sequence to be tested, i.e., the sequence content. Directly obtain the filename of this file, which serves as the common source filename for all sequences to be tested within that file.

[0070] For each sequence in the qualified sequence set obtained after screening, during the screening process using the same screening rules as the benchmark sequence, the length of each sequence string is counted by counting the number of amino acids contained in the sequence and the number of abbreviations of amino acid residues in the string. This count value is the sequence length. After comparison with a preset minimum length threshold, short sequences that are too short are removed, and the remaining sequences and their length values ​​are retained along with the qualified sequence set.

[0071] For the protein sequence to be tested, the same feature extraction process as in S1 is used to obtain a 17-dimensional raw feature vector. This raw feature vector is then input into the general scoring function defined in S2. Using the mean vector, corrected standard deviation vector, weight coefficient vector, and regularized inverse covariance matrix read from the persistent model file saved in S3, the weighted similarity score WS for the sequence to be tested is calculated. t Density score to be tested (DS) t and the fusion score CS to be tested t .

[0072] For the input 17-dimensional original feature vector, the weighted similarity score channel and the Mahalanobis distance density score channel are calculated independently. The obtained weighted similarity score and density score are nonlinearly fused to obtain a fusion score. Then, the fusion score is compared with a positive threshold and a final threshold. Based on the comparison result, a first classification label, a second classification label, or a third classification label is output, corresponding to three gradients: high activity potential, medium activity potential, and low activity potential, respectively. The specific steps are as follows: The fusion score of the original protein to be tested (CS) t Compare with the positive threshold and the final threshold: If CS t If the value is greater than the positive threshold, output the first category label; if CS t If the value is less than or equal to the positive threshold and greater than or equal to the final threshold, then output the second category label; if CS tIf the value is less than the final threshold, the third category label is output. The first, second, and third category labels are all antimicrobial activity potential category labels, indicating different levels of antimicrobial activity potential.

[0073] Simultaneously, the fusion score CS to be tested is calculated. t The difference between the positive threshold and the final threshold is output as a threshold distance index.

[0074] S5: Batch prediction and result file output.

[0075] Import the protein sequence file to be tested. The protein sequence file to be tested is in FASTA (FAST-All, a biological sequence storage format) format, which contains a preset number of proteome sequences of the species to be tested.

[0076] For each imported protein sequence to be tested, the S4 triple gradient prediction process is invoked to obtain the fusion score CS of that sequence. t And the corresponding classification labels, including the sequence identifier, sequence content, sequence length, and fusion score CS. t Weighted similarity score WS t Density score DS t The prediction result record consists of the category label, threshold distance index, and source file name. The threshold distance index is based on the fusion score CS. t The difference between the positive threshold and the final threshold is determined. If the protein sequence file to be tested is empty or all sequences fail the screening in S4, only a header line containing the field names is generated as the prediction result record, and an empty statistical result and prompt message are generated for the statistical report.

[0077] Write all predicted results to the first output file. Select records from the first output file that match the first category label and write these records to the second output file. Count the number of records for each category label in the first output file and assign all records to the fusion score CS. t Sort in descending order. A preset number of sequences is set based on the user's requirement for the number of sequences displayed in the statistical report, controlling the number of detailed information entries recorded in the third output file. The detailed information of the first preset number of sequences is written to the third output file, which is the prediction result statistical report.

[0078] The first output file records the prediction results for all test sequences, facilitating comprehensive data review. The second output file quickly identifies the sequence candidates with the highest antibacterial activity potential, which can be directly used for priority ranking in downstream experimental validation. The third output file provides a quick overview of the overall statistical distribution of the batch predictions.

[0079] like Figure 4The image shows a sequence overlap diagram of a high-confidence AMP sequence provided in an embodiment of this application. It illustrates the number of overlaps in the high-confidence AMP sequences under different stringency conditions, with 367 overlaps under the lenient condition, 145 overlaps under the moderate condition, and 47 overlaps under the stringent condition. Figure 5 The figure shown is a graph illustrating the number of high-confidence AMPs provided in the embodiments of this application. It shows the number of high-confidence AMPs identified under different stringency conditions, with 675 identified under the lenient condition, 232 under the moderate condition, and 62 under the stringent condition; as shown... Figure 6 The figure shown is a comprehensive score distribution diagram provided in an embodiment of this application. It illustrates the comprehensive score distribution under different stringency conditions, with each stringency level exhibiting a specific concentrated distribution trend. This diagram verifies that the general scoring function in the embodiment can create clearly separated concentrated distributions of comprehensive scores under different gradient thresholds, proving the effectiveness of the hierarchical boundaries set by the positive threshold and the final threshold. Figure 7 The figure shown is a weighted score comparison chart provided in an embodiment of this application. It illustrates a comparison of weighted scores under three conditions: lenient, medium, and strict. The median weighted score under the lenient condition is approximately 0.882; under the medium condition, it is approximately 0.903; and under the strict condition, it is approximately 0.918. This chart reflects the central tendency of the weighted scores under each condition. It shows that the median weighted similarity score increases with increasing strictness, indicating that the scoring function can sensitively reflect the severity of the grading standard, corresponding to the gradient recognition logic of the embodiment. Figure 8 The figure shown is a sequence length distribution diagram provided in an embodiment of this application. It illustrates the distribution of sequence lengths under different stringency conditions. The density values ​​under relaxed, moderate, and strict conditions all show a distribution trend of first increasing and then decreasing. This figure verifies that the sequence length density distribution in the embodiment conforms to the following: short sequences have low density due to incomplete features, medium-length sequences have high density due to complete features, and excessively long sequences have decreased density due to feature dilution. This indicates that the 17-dimensional feature extraction and length screening rules of the embodiment are reasonable, and the grading results have biological interpretability. Figure 9 The figure shown is an average comprehensive score graph provided by the embodiments of this application. It shows the average comprehensive score under different stringency conditions. The average comprehensive score is 0.949 under the lenient condition, 0.956 under the medium condition, and 0.964 under the strict condition. The graph verifies that the average comprehensive score is highest under the strict condition, followed by the medium condition, and lowest under the lenient condition in the embodiments. This indicates that the setting of the positive threshold and the final threshold effectively progressively improves the stringency of recognition, which is consistent with the gradient recognition target of the embodiments.

[0080] The above six figures together constitute an overview of the performance comparison of the method of the present invention under different gradient thresholds, fully demonstrating the technical effects from the calculation of benchmark sample statistics, the construction of general scoring functions to the output of three-gradient classification labels.

[0081] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0082] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0083] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0084] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0085] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A gradient identification method for antimicrobial peptides based on multi-feature fusion of protein sequences, characterized in that, Includes the following steps: S1: Read the positive antimicrobial peptide dataset and the negative non-antimicrobial peptide dataset, and perform independent sequence entry identification, sequence content legality verification and length filtering on these two types of benchmark sequence files to obtain the benchmark positive antimicrobial peptide qualified sequence set and the benchmark negative non-antimicrobial peptide qualified sequence set. The baseline positive antimicrobial peptide qualified sequence set and the baseline negative non-antimicrobial peptide qualified sequence set were subjected to 17-dimensional full-dimensional antimicrobial peptide feature quantification extraction, and the baseline positive sample feature matrix and the baseline negative sample feature matrix were constructed. S2: Calculate the mean vector, corrected standard deviation vector, weight coefficient vector, and regularized inverse covariance matrix based on the feature matrix of the benchmark positive samples, and define a general scoring function for calculating the fusion score; A general scoring function is used to calculate the positive nonlinear fusion score and the negative nonlinear fusion score respectively. A positive threshold is determined based on the positive nonlinear fusion score, and a negative threshold is determined based on the negative nonlinear fusion score. The positive nonlinear fusion score sequence is arranged in ascending order of numerical value, and the positive low threshold is obtained. The final threshold is determined based on the negative threshold and the positive low threshold. S3: Save the mean vector, corrected standard deviation vector, weight coefficient vector, regularized inverse covariance matrix, positive threshold, and final threshold to form a persistent model file; S4: For a single protein sequence to be tested, the same feature extraction process is called to obtain a 17-dimensional original feature vector. The fusion score to be tested is calculated using a persistent model file. The fusion score to be tested is compared with the positive threshold and the final threshold, and the first classification label, the second classification label or the third classification label is output. S5: Batch import the protein sequence files to be tested, call S4 one by one to get the prediction results, record and output the first output file, the second output file and the third output file.

2. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The process of identifying independent sequence entries, validating the validity of sequence content, and filtering length includes: Extract the string immediately following the greater than sign and before the first space from the description line starting with the greater than sign in the baseline sequence file. If there is no space immediately following the greater than sign in the description line, extract the string from the description line as the sequence identifier according to the preset truncation rules. The preset truncation rule is to truncate from after the greater than sign to the end of the description line, or to the first space, tab, vertical bar, or comma; Extract the string of the sequence line immediately following the description line as the sequence content; The filename of the baseline sequence file is directly obtained and used as the source filename common to all sequences in that file; The sequence content is validated for legality, and abnormal sequences containing incomplete characters, garbled characters or invalid characters are removed. The results are then aggregated to obtain the baseline positive antimicrobial peptide screening sequence set and the baseline negative non-antimicrobial peptide screening sequence set. The number of abbreviations of amino acid residues in each sequence content string is counted, and short sequences with amino acid lengths less than a preset minimum length threshold are removed to obtain a set of qualified positive antimicrobial peptide sequences and a set of qualified negative non-antimicrobial peptide sequences.

3. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The 17-dimensional full-dimensional antimicrobial peptide feature quantification extraction includes: The net charge, charge density, total average hydrophilicity, isoelectric point, aliphatic index, hydrophobic moment, amphiphilicity, proportion of hydrophobic residues, proportion of cationic residues, proportion of aromatic residues, hydrophobicity of amino-terminal pentapeptide, hydrophobicity of carboxyl-terminal pentapeptide, number of hydrogen bond donors, number of hydrogen bond acceptors, dipeptide preference, disorder tendency, and cation uniformity are extracted in a fixed order to form a 17-dimensional original feature vector. The 17-dimensional original feature vectors of all sequences in the benchmark positive antimicrobial peptide qualified sequence set are arranged and integrated row by row to construct the benchmark positive sample feature matrix. The 17-dimensional original feature vectors of all sequences in the baseline negative non-antimicrobial peptide qualified sequence set are arranged and integrated row by row to construct the baseline negative sample feature matrix.

4. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The calculation of the mean vector, corrected standard deviation vector, weight coefficient vector, and regularized inverse covariance matrix based on the feature matrix of the benchmark positive samples includes: Based on the fixed feature order of the 17-dimensional original feature vector, the mean and original standard deviation corresponding to the i-th feature are calculated column by column; The original standard deviation of the i-th feature is compared with a preset minimum positive number, and the larger of the two values ​​is taken as the corrected standard deviation of the i-th feature. The means corresponding to the seventeen features are arranged into a mean vector according to a fixed feature order, and the corrected standard deviations corresponding to the seventeen features are arranged into a corrected standard deviation vector according to a fixed feature order. The weight coefficients corresponding to the i-th feature are calculated using the inverse standard deviation weighting method, and the seventeen weight coefficients are arranged into a weight coefficient vector according to a fixed feature order. The covariance matrix is ​​calculated based on the feature matrix of the benchmark positive sample. The covariance matrix is ​​then added to the matrix obtained by multiplying the regularization coefficient by the identity matrix to obtain the regularized covariance matrix. Finally, the regularized inverse covariance matrix is ​​obtained by inverting the regularized covariance matrix.

5. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The general scoring function includes: For any 17-dimensional original feature vector, the mean vector and the corrected standard deviation vector obtained from the benchmark positive sample are called, and Z-score standardization is performed to obtain the standardized feature vector. Calculate the Gaussian similarity components one by one based on the standardized eigenvalues ​​in the standardized eigenvector; The 17 Gaussian similarity components of the same protein sequence and their corresponding feature weight coefficients are weighted, summed, and normalized to obtain a weighted similarity score for the protein sequence. For any protein sequence, calculate the Mahalanobis distance between its original feature vector and the mean vector, and obtain the density score corresponding to the protein sequence based on the chi-square distribution cumulative distribution function mapping. When the weighted similarity score is greater than 0 and the density score is greater than 0, the weighted similarity score and the density score are non-linearly fused according to a fixed exponential weight to obtain the non-linear fusion score value corresponding to the protein sequence. When the weighted similarity score is less than or equal to 0 or the density score is less than or equal to 0, CS is set directly to 0.

6. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 5, characterized in that: The Z-value standardization includes: For each dimension of the original feature vector, calculate the difference between the original feature value of that dimension and the mean of that dimension in the mean vector, and then divide it by the standard deviation of that dimension in the corrected standard deviation vector to obtain the standardized feature value of that dimension.

7. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The process of determining a positive threshold based on a positive nonlinear fusion score, a negative threshold based on a negative nonlinear fusion score, and a final threshold based on a negative threshold and a low positive cutoff value includes: All positive nonlinear fusion scores are sorted in ascending order of numerical value, and the positive nonlinear fusion score that is in the first preset high percentile after sorting is taken as the positive threshold. The second preset percentile is determined based on the preset false positive tolerance rate. All negative nonlinear fusion scores are arranged in ascending order of value, and the negative nonlinear fusion score at the second preset percentile after arrangement is taken as the negative threshold. The positive nonlinear fusion scores are sorted in ascending order, and the positive nonlinear fusion scores that are in the third preset percentile after sorting are taken as the positive low threshold value. The third preset percentile is lower than the first preset high percentile. The negative threshold is compared with the positive low threshold, and the larger of the two is taken as the candidate final threshold. If the candidate final threshold is less than the positive threshold, the candidate final threshold is determined as the final threshold. If the candidate final threshold is greater than or equal to the positive threshold, the positive threshold is subtracted from the preset threshold interval to obtain the final threshold, and an abnormal threshold interval prompt message is output.

8. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: The saved mean vector, corrected standard deviation vector, weight coefficient vector, regularized inverse covariance matrix, positive threshold, and final threshold form a persistent model file, including: Save the mean vector composed of the mean of each feature dimension of the baseline positive sample feature matrix, the standard deviation vector composed of the corrected standard deviation of each feature dimension, the weight coefficient vector calculated from the reciprocal of the standard deviation, the regularized inverse covariance matrix of the baseline positive sample feature matrix, the positive threshold, and the final threshold. The above data is serialized and written to the same file to form a complete set of model parameters, thereby saving all the statistics and matrices required to construct the hierarchical classification rules as a single persistent model file.

9. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: For a single protein sequence to be tested, the same feature extraction process is called to obtain a 17-dimensional original feature vector. The fusion score to be tested is calculated using a persistent model file, including: Read the mean vector, corrected standard deviation vector, weight coefficient vector, regularized inverse covariance matrix, positive threshold, and final threshold from the persistent model file; For a single protein sequence to be tested, import the original protein sequence file to be tested, automatically split the file to identify independent sequence entries, and extract the sequence identifier, sequence content and source file name of the sequence to be tested; The protein sequence to be tested was screened according to the same screening rules as the baseline sequence to obtain a set of qualified sequences to be tested; The same feature extraction process as the baseline sequence was used to extract the original 17-dimensional feature vector from the protein sequence to be tested. The original feature vector is input into a general scoring function, and the weighted similarity score and density score are calculated using the mean vector, the corrected standard deviation vector, the weight coefficient vector, and the regularized inverse covariance matrix. The fusion score is determined based on the weighted similarity score and the density score.

10. The method for gradient identification of antimicrobial peptides based on multi-feature fusion of protein sequences as described in claim 1, characterized in that: After outputting the first category label, the second category label, or the third category label, the method further includes: Calculate the difference between the fusion score to be tested, the positive threshold, and the final threshold, and output it as the threshold distance index. For each imported protein sequence to be tested, the three-gradient prediction process is invoked to obtain the fusion score and corresponding classification label of the sequence; The prediction result record is composed of the sequence identifier, sequence content, sequence length, fusion score, weighted similarity score, density score, classification label, threshold distance index, and source file name; Write all the predicted results to the first output file. Filter out the records with the first category label from the first output file and write these records to the second output file. Count the number of records for each category label in the first output file. Sort all records in descending order of fusion score and write the detailed information of the first preset number of records to the third output file.

Images