Method, device and equipment for gene prediction or classification of antibiotic resistance genes or virulence factors

Abstract

The present invention discloses a gene prediction or classification method, device, and equipment for antibiotic resistance genes or virulence factors, and relates to the technical fields of deep learning and bioinformatics. By extracting six gene-related features, including similarity features based on bit scores, sequence information features, physical and chemical property features, evolutionary information features, amino acid composition features, and domain information features, the comprehensive characteristics of genes can be better captured. After feature fusion, the model constructed thereby can improve the accuracy of gene prediction for antibiotic resistance genes or virulence factors. It has broad application prospects for gene screening, classification, and prediction of antibiotic resistance genes or virulence factors.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and bioinformatics, and in particular to a method, device and equipment for predicting or classifying genes of antibiotic resistance genes or virulence factors. Background Art

[0002] Pathogenic microorganisms can cause disease and even threaten the life of the host. Virulence factors in pathogenic microorganisms can induce pathogenicity and can help the microorganisms colonize the host at the cellular level. The success of pathogenic microorganisms depends on their ability to use virulence factors to cause infection, survive in a hostile host environment and cause disease. Various expressions, organizations and combinations of virulence factors can cope with different clinical symptoms of pathogenic infections. VFs can be divided into multiple categories such as adhesion, colonization, exotoxins, endotoxins, iron transport, etc. Different virulence factors work together to enable pathogenic microorganisms (such as bacteria and fungi) to successfully adhere, reproduce and cause damage to host cells.

[0003] However, virulence factor gene identification poses challenges as it requires high precision for actual treatment and robustness to quickly identify problems. Currently, characterization and diagnostic techniques used in laboratories do not provide enough information to effectively conduct surveillance. In addition, test results are inconsistent depending on the environment and laboratory settings. Sequencing technologies, as well as laboratory techniques, can be used to study the genetic variation of widespread antibiotic resistance genes.

[0004] Xie (2021) used a stacked ensemble model to integrate four classic algorithms, including random forest, support vector machine, extreme gradient boosting, and multi-layer perceptron, as well as three deep learning algorithms, including convolutional neural network, long short-term memory network, and deep neural network, to develop DeepVF for identifying and annotating virulence factor genes from metagenomic data. DeepVF uses sequence information features based on bit scores, physicochemical property features, and evolutionary information features. DeepARG uses similarity features based on bit scores when training the model.

[0005] Additionally, antibiotic resistance genes (ARGs) are an obstacle to treating pathogen infections, exacerbating the pathogenicity of microorganisms. Over the years, antibiotics have been shown to be effective in treating a variety of microbial infections, especially bacterial infections. However, as bacterial pathogens become increasingly resistant to antibiotics, treatment of bacterial infections is becoming increasingly limited worldwide. For some patients, such as those caused by multidrug-resistant (MDR) bacteria, effective treatment options did not even exist before, and these bacteria are now unresponsive to traditional first-line treatments.

[0006] However, antibiotic resistance gene identification poses challenges because it requires high precision in actual treatment and robustness to quickly identify the problem. Currently, characterization and diagnostic techniques used in laboratories do not provide enough information to effectively conduct monitoring. In addition, test results are inconsistent depending on the environment and laboratory settings. Sequencing technologies, as well as laboratory techniques, can be used to study the genetic variation of widely spread antibiotic resistance genes.

[0007] Arango-Argoty (2018) developed DeepARG based on a deep learning model to identify and annotate antibiotic resistance genes from metagenomic data. DeepARG uses similarity features based on bit scores when training the model.

[0008] Ji et al. (2023) used a stacked ensemble model, combining random forest, extreme random tree, Xgboost, gradient boosting and Adaboost algorithms, and developed HyperVR software for predicting antibiotic resistance genes and virulence genes. HyperVR uses similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features and amino acid composition features. With the development of deep learning technology, new deep learning network models can be used for drug resistance gene prediction, which will make drug resistance gene prediction more accurate.

[0009] In summary, the accuracy of existing virulence gene prediction methods or drug resistance gene prediction methods needs to be further improved.

[0010] In view of this, the present invention is proposed. Summary of the invention

[0011] The purpose of the present invention is to provide a method, device and equipment for predicting or classifying antibiotic resistance genes or virulence factors so as to improve the accuracy of virulence gene or resistance gene prediction.

[0012] The present invention is achieved in that:

[0013] In a first aspect, the present invention provides a method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning, which comprises the following steps:

[0014] S1. Obtain known antibiotic resistance gene data or virulence factor gene data, as well as negative sample gene data from the database; divide the collected data set into a training set, a test set, and a reference data set;

[0015] S2. Extract the following six features using the gene data collected in S1: similarity feature based on bit score, sequence information feature, physicochemical property feature, evolutionary information feature, amino acid composition feature and domain information feature;

[0016] S3. Adopt the fusion strategy to merge the six features extracted in S2 to form a comprehensive input vector;

[0017] S4. Use deep convolutional neural network to build a classification model; take the six features of the training set and merge them into an input vector to train the classification model;

[0018] S5. Take the six features of the test set and merge them into an input vector to test the classification model trained in S4 and perform model evaluation on the classification model.

[0019] In a second aspect, the present invention also provides a deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors, comprising: an input module, a control module, and an output module;

[0020] The input module is configured to: input known antibiotic resistance gene data or virulence factor gene data, and negative sample gene data obtained from a database;

[0021] The control module includes: a gene data extraction module, a feature fusion module, a classification model construction module and a classification model testing module;

[0022] The gene data extraction module is configured to: extract similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, amino acid composition features and domain information features of the data in the input module;

[0023] The feature fusion module is configured to: merge the six features extracted by the gene data extraction module to form a comprehensive input vector;

[0024] The classification model building module is configured as follows: using ResNet-50 as a deep convolutional neural network to build a classification model; taking the six features of the training set and the merged input vector to train the classification model;

[0025] The classification model testing module is configured to: take the combined input vector of the six features of the test set to test the classification model trained by the classification model building module, and perform model evaluation on the classification model;

[0026] The output module is configured to output at least one of a classification model, a classification result and a prediction result.

[0027] In the third aspect, the present invention also provides a gene prediction or classification device for antibiotic resistance genes or virulence factors based on deep learning, the device includes a processor and a memory, the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement a gene prediction or classification method for antibiotic resistance genes or virulence factors based on deep learning.

[0028] In a fourth aspect, the present invention also provides a computer-readable storage medium, storing at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement a deep learning-based method for predicting or classifying antibiotic resistance genes or virulence factors.

[0029] The present invention has the following beneficial effects:

[0030] The gene prediction or classification method of antibiotic resistance genes or virulence factors based on deep learning provided by the present invention can better capture the comprehensive characteristics of genes by extracting six gene-related features, including similarity features based on bit scores, sequence information features, physical and chemical property features, evolutionary information features, amino acid composition features and domain information features. After feature fusion, the model constructed thereby can improve the accuracy of gene prediction of antibiotic resistance genes or virulence factors. It has broad application prospects for gene screening, classification and prediction of antibiotic resistance genes or virulence factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0032] Figure 1 A detailed flowchart of the gene prediction method for antibiotic resistance genes based on deep learning;

[0033] Figure 2 Detailed flowchart of the deep learning-based virulence gene prediction method. DETAILED DESCRIPTION

[0034] References to embodiments of the present invention will now be provided in detail, one or more examples of which are described below. Each example is provided as an explanation rather than a limitation of the present invention. In fact, it will be apparent to those skilled in the art that various modifications and variations may be made to the present invention without departing from the scope or spirit of the present invention. For example, a feature illustrated or described as part of one embodiment may be used in another embodiment to produce a further embodiment.

[0035] In a first aspect, the present invention provides a method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning, which comprises the following steps:

[0036] S1. Obtain known antibiotic resistance gene data or virulence factor gene data, as well as negative sample gene data from the database; divide the collected data set into a training set, a test set, and a reference data set;

[0037] S2. Extract the following six features using the gene data collected in S1: similarity feature based on bit score, sequence information feature, physicochemical property feature, evolutionary information feature, amino acid composition feature and domain information feature;

[0038] S3. Adopt the fusion strategy to merge the six features extracted in S2 to form a comprehensive input vector;

[0039] S4. Use deep convolutional neural network to build a classification model; take the six features of the training set and merge them into an input vector to train the classification model;

[0040] S5. Take the six features of the test set and merge them into an input vector to test the classification model trained in S4 and perform model evaluation on the classification model.

[0041] In order to improve the accuracy of method prediction or classification, the inventors found through long-term and extensive screening that by adding domain information features to the characteristics of genes, and then fusing the domain information features with similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, and amino acid composition features, the model constructed in this way can improve the gene prediction accuracy and classification accuracy of antibiotic resistance genes or virulence factors.

[0042] The invention has broad application prospects for gene screening, classification and prediction of antibiotic resistance genes or virulence factors.

[0043] The ResNet-50 in step S4 has a deep residual network structure and is suitable for processing complex bioinformatics data. By training on labeled gene data (i.e., genes with labels (whether they are resistant or toxic)), the model can learn complex patterns of genes in the classification of drug resistance or toxicity genes.

[0044] In a preferred embodiment of the present invention, the extraction of domain information features refers to: converting the presence or absence of protein domains into a digital vector.

[0045] In a preferred embodiment of the present invention, genes located in the protein domain are set to 1, and genes located in the non-protein domain are set to 0.

[0046] In a preferred embodiment of the present invention, in step S1: obtaining known antibiotic resistance gene data from at least one database including but not limited to UniProt database, CARD database, ARDB database, SARG database and NCRD database. Any database that can obtain antibiotic resistance gene data can be used as a gene database for feature extraction.

[0047] Gene data of known virulence factors are obtained from at least one database including but not limited to PATRIC, Victors, VFDB and UniProt databases.

[0048] In a preferred embodiment of the present invention, when predicting the genes of virulence factors and / or resistance factors, step S3 includes: first pre-fusing at least two of the six features extracted in S2 to obtain a pre-fused feature vector; and then fusing the pre-fused feature vector with other features of the six features to form a comprehensive input vector.

[0049] This fusion method aims to integrate various features and present them to the deep learning model in a more comprehensive and integrated way. This early fusion step is crucial to preserve the correlation between features and improve model performance.

[0050] In a preferred embodiment of the present invention, when predicting the genes of virulence factors and / or drug resistance factors, step S3 includes:

[0051] (i) Pre-fusion of similarity features and evolutionary information features to form a first fusion feature vector; pre-fusion of amino acid composition features and domain information features to obtain a second fusion feature vector; the design of this step is based on the possible correlation between similarity features and evolutionary information features in gene virulence, so as to more effectively capture the evolution and similarity information of genes;

[0052] (ii) The first fused feature vector, the second fused feature vector, the sequence information feature and the physicochemical property feature are fused to form a comprehensive input vector. This fusion process takes into account the biological structure and function of the gene, which helps to understand the characteristics of the gene more comprehensively.

[0053] The advantages of the above pre-fusion and fusion are: through multi-level feature fusion, the multi-dimensional information of genes is better expressed, providing richer input for the ResNet-50 model, thereby improving the accuracy and performance of the classification model.

[0054] The design of the fusion strategy of the present invention (pre-fusion of similarity features and evolutionary information features to form a first fusion feature vector; pre-fusion of amino acid composition features and domain information features to obtain a second fusion feature vector) takes into account the correlation and biological significance between different features, making the model more interpretable and generalizable, and providing a powerful and comprehensive tool for the study of genetic virulence.

[0055] In a preferred embodiment of the present invention, the deep convolutional neural network is selected from at least one of ResNet-18, ResNet-50, VGGNet, AlexNet, Inception v3, Inception v4, Inception v5, DenseNet, NASNet, GoogleNet, MobileNets and EfficientNets.

[0056] In a preferred embodiment of the present invention, the similarity feature based on the bit score refers to: the local comparison score between the gene sequence in the training data set and the sequence in the reference data set; the higher the bit score, the higher the similarity between the sequences.

[0057] In a preferred embodiment of the present invention, the bit score is calculated by DIAMOND software;

[0058] In a preferred embodiment of the present invention, the reference data set includes o sequences, and the bit score-based similarity feature of the training set obtains an o-dimensional vector composed of bit scores for each sequence.

[0059] In a preferred embodiment of the present invention, the sequence information features include amino acid composition, dipeptide composition, and the deviation of dipeptides from the expected average value;

[0060] Among them, (1) amino acid composition refers to the frequency of 20 natural amino acids in the protein or peptide sequence, and the calculation formula is as follows:

[0061] f(A) = N(A) / T;

[0062] Where A represents a given amino acid, N(A) represents the number of given amino acid A, T represents the total number of amino acids in a protein or peptide, and f(A) represents the number of occurrences of a given amino acid A;

[0063] (2) Dipeptide composition refers to the frequency of dipeptides in a protein or peptide sequence. The calculation formula is as follows:

[0064] f(A1A2)=N(A1A2) / T;

[0065] Where A1A2 represents a given dipeptide, N(A1A2) represents the number of given amino acids A1A2, T represents the total number of amino acids in a protein or peptide, and f(A1A2) represents the number of occurrences of given amino acids A1A2. Finally, a 20*20=400-dimensional feature vector with a sum of 1 is generated.

[0066] (3) the deviation of the dipeptide from the expected mean value is composed of the dipeptide composition, the theoretical mean, and the theoretical variance value;

[0067] (i) The theoretical mean value is calculated as follows:

[0068] TM(A1A2)=(CA1 / CA)*(CA2 / CA);

[0069] Among them, CA is 61, which refers to the number of codons encoding amino acids excluding three stop codons, CA1 is the number of codons encoding amino acid A1, and CA2 is the number of codons encoding amino acid A2;

[0070] (ii) The theoretical variance calculation formula is as follows:

[0071]

[0072] (iii) The deviation of dipeptides from the expected mean is calculated as follows:

[0073]

[0074] In a preferred embodiment of the present invention, the physicochemical property characteristics describe statistical information related to the physicochemical properties of amino acids in a protein sequence. The extraction of the physicochemical property characteristics includes extracting pseudo amino acid composition characteristics and quasi-sequence order.

[0075] Among them, the extraction of pseudo amino acid composition features includes:

[0076] The original side chain mass, hydrophobicity and hydrophilicity of 20 natural amino acids are defined as H(i), HR i and HR j ;

[0077] The physicochemical properties of 20 natural amino acids are normalized to the following formula:

[0078]

[0079] Calculate the correlation between the original side chain mass, hydrophobicity and hydrophilicity physicochemical properties of n amino acids, and set the two physicochemical properties as HR i and HR j , the formula is as follows:

[0080]

[0081] For a sequence of length L, the ordered correlation factor of this sequence is defined as follows:

[0082]

[0083] Where j represents a positive integer, L is an integer used to define the maximum value of j; i represents the position of the amino acid in the protein sequence as i;

[0084] The PAAC features of protein sequences are defined as follows:

[0085]

[0086] where f a represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect;

[0087] The quasi-sequence order is derived by measuring the physicochemical distances between the amino acids of a given protein sequence, resulting in a sequence feature vector; for any given distance matrix:

[0088]

[0089] Where L is the length of the sequence, represents the element in the i-th row and i+d-th column of the distance matrix, and maxlag is the maximum lag value. a represents the normalized frequency of amino acid A in the protein sequence, ω represents the weight factor of the sequence order effect; τ d represents the rank order coupling number, Xr represents the first 20 quasi-sequence order features, and X represents the other quasi-sequence order features;

[0090] In a preferred embodiment of the present invention, the evolutionary information feature in step S2 refers to estimating the evolutionary relationship of different gene sequences based on a set of probability scores for each amino acid at each position.

[0091] The position-specific scoring matrix consists of a set of probability scores for each amino acid (or gap) at each position in the alignment and is used to estimate the evolutionary conservation of genes.

[0092] In a preferred embodiment of the present invention, PSI-BLAST software is used to search for homologous protein sequences to generate a PSSM matrix; the basic idea of ​​PSSM is to match the query sequence in the database with the sequence in the alignment table, thereby providing a higher weight to the conserved position than the variable position.

[0093] The PSSM matrix is ​​defined as follows:

[0094] Based on the PSSM matrix, three specific features are derived, including PSSM component features, RPMPSSM features, and S PSSM features;

[0095] a) PSSM component feature representation The rows of the original PSSM are transformed by summing and averaging the row vectors of amino acids belonging to the same type:

[0096]

[0097] Subject to

[0098]

[0099] Among them, R i refers to the i-th row of the transformation matrix, r k refers to the kth row of the original PSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids, and the result is a 20*20 matrix;

[0100] b) The RPMPSSM feature is obtained by first setting the negative elements in the original PSSM to 0, retaining the positive values, and converting it into a new matrix (PPPSM); then using a formula similar to the PSSM component feature, the RPM-PSSM feature is calculated;

[0101]

[0102] Subject to

[0103]

[0104] Among them, R′ i refers to the i-th row of the filter matrix, r′ k refers to the kth row of PPSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids; the result is a 20*20 matrix;

[0105] c) The S PSSM feature is obtained by first setting the negative elements in the original PSSM to 0 and the values ​​greater than 7 to 7, and converting it into a new matrix S PSSM; the calculation method of S-FPSSM is as follows:

[0106]

[0107] Subject to

[0108]

[0109] Among them, fp k,j refers to the kth row and jth column of FPSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, α i Refers to the ith amino acid of the 20 natural amino acids; S j (i) refers to the S PSSM eigenvalue, resulting in a vector of length 400.

[0110] In a preferred embodiment of the present invention, step S2 of amino acid composition feature extraction refers to: converting the protein sequence into a digital vector, wherein each of the 20 natural amino acids is converted into a 20-dimensional feature vector, and the amino acid position is set to 1 in alphabetical order, and the remaining positions are set to 0.

[0111] In a preferred embodiment of the present invention, negative samples are collected from the PBVF database, genes with 100% similarity are removed using cd-hit software, and the remaining data are used as data for the training set, test set, and reference data set;

[0112] In a preferred embodiment of the present invention, step S5 further includes: classifying or predicting the drug resistance of the gene according to the evaluated optimal classification model, or classifying or predicting the virulence of the gene, and the method is not aimed at diagnosing the disease.

[0113] In a second aspect, the present invention also provides a deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors, comprising: an input module, a control module, and an output module;

[0114] The input module is configured to: input known antibiotic resistance gene data or virulence factor gene data, and negative sample gene data obtained from a database;

[0115] The control module includes: a gene data extraction module, a feature fusion module, a classification model construction module and a classification model testing module;

[0116] The gene data extraction module is configured to: extract similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, amino acid composition features and domain information features of the data in the input module;

[0117] The feature fusion module is configured to: merge the six features extracted by the gene data extraction module to form a comprehensive input vector;

[0118] The classification model building module is configured as follows: using ResNet-50 as a deep convolutional neural network to build a classification model; taking the six features of the training set and the merged input vector to train the classification model;

[0119] The classification model testing module is configured to: take the combined input vector of the six features of the test set to test the classification model trained by the classification model building module, and perform model evaluation on the classification model;

[0120] The output module is configured to: output at least one of a classification model, a classification result, and a prediction result;

[0121] In a preferred embodiment of the present invention, the control module further comprises a data preprocessing module, and the data preprocessing module is configured to: remove genes with 100% similarity from the data of the input module;

[0122] In a preferred embodiment of the present invention, when predicting or classifying genes of virulence factors, the feature fusion module includes a pre-fusion module and a fusion module;

[0123] The pre-fusion module is configured to: pre-fuse the similarity feature and the evolutionary information feature to form a first fusion feature vector; pre-fuse the amino acid composition feature and the domain information feature to obtain a second fusion feature vector;

[0124] The fusion module is configured to fuse the first fused feature vector, the second fused feature vector, the sequence information feature and the physicochemical property feature to form a comprehensive input vector.

[0125] In the third aspect, the present invention also provides a gene prediction or classification device for antibiotic resistance genes or virulence factors based on deep learning, the device includes a processor and a memory, the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement a gene prediction or classification method for antibiotic resistance genes or virulence factors based on deep learning.

[0126] Specifically, the electronic device may include a memory, a processor, a bus, and a communication interface, and the memory, processor, and communication interface are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these elements can be electrically connected to each other via one or more buses or signal lines. The processor can process information and / or data related to target identification to perform one or more functions described in this application.

[0127] The memory can be, but is not limited to, random access memory (RAM), read only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable read-only memory (EEPROM), etc.

[0128] In a fourth aspect, the present invention also provides a computer-readable storage medium, storing at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement a deep learning-based method for predicting or classifying antibiotic resistance genes or virulence factors.

[0129] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0130] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be described clearly and completely below. If the specific conditions are not specified in the embodiments, they are carried out according to conventional conditions or conditions recommended by the manufacturer. If the manufacturer of the reagents or instruments used is not specified, they are all conventional products that can be purchased commercially.

[0131] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0132] Example 1

[0133] This embodiment provides a gene prediction method for antibiotic resistance genes based on deep learning. For a specific flowchart, refer to Figure 1 The gene prediction method of antibiotic resistance gene includes the following steps:

[0134] 1. Preprocessing

[0135] Antibiotic resistance genes recorded in UniProt database, CARD database, ARDB database, SARG database and NCRD database were collected. Genes with 100% similarity were removed using cd-hit software, and the remaining data was used as the database of this embodiment for subsequent analysis.

[0136] The data set of this embodiment is divided into three subsets: 1) a training data set (m sequences), which is used as a training data set for the model of the present invention; 2) a test data set (n sequences), which is used to verify the accuracy of the model of the present invention; 3) a reference true set (o sequences, also known as a reference data set): reference data for comparison with the training data set and the test data set, and is also the largest of the three sub-datasets.

[0137] 2. Six categories of feature extraction

[0138] In order to obtain accurate prediction capabilities for antibiotic resistance genes, the present invention considers a variety of gene-related features, including the following six categories: similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, amino acid composition features, and domain information features.

[0139] 1) Similarity feature based on bit score: The bit score is evaluated by calculating the local alignment score between sequences. The higher the score, the higher the similarity between the sequences. The similarity feature is composed of the bit scores between the gene sequences in the training dataset and the sequences in the reference dataset. The bit score is calculated by DIAMOND software. As a result, each sequence in the training set will get an o-dimensional vector composed of bit scores.

[0140] 2) Sequence information features: Sequence information features include amino acid composition, dipeptide composition and the deviation of dipeptides from the expected average.

[0141] a) Amino acid composition refers to the frequency of 20 natural amino acids in a protein or peptide sequence, calculated as follows:

[0142] f(A) = N(A) / T;

[0143] Where A represents a given amino acid, N(A) represents the number of given amino acid A, T represents the total number of amino acids in the protein or peptide, and f(A) represents the number of occurrences of given amino acid A. Finally, a 20-dimensional feature vector with a sum of 1 is generated.

[0144] b) Dipeptide composition represents the frequency of dipeptides in a protein or peptide sequence. The calculation formula is as follows:

[0145] f(A1A2)=N(A1A2) / T;

[0146] Among them, A1A2 represents a certain dipeptide, N(A1A2) represents the number of given amino acids A1A2, T represents the total number of amino acids in the protein or peptide, and f(A1A2) represents the number of occurrences of given amino acids A1A2. Finally, a 20*20=400-dimensional feature vector with a sum of 1 is generated.

[0147] c) Deviation of dipeptides from the expected mean value composed of dipeptide composition, theoretical mean, and theoretical variance values.

[0148] i. The theoretical mean calculation formula is as follows:

[0149] TM(A1A2)=(CA1 / CA)*(CA2 / CA);

[0150] Among them, CA is 61, which refers to the number of codons encoding amino acids excluding three stop codons, CA1 encodes the number of codons for amino acid A1, and CA2 encodes the number of codons for amino acid A2.

[0151] ii. The theoretical variance calculation formula is as follows:

[0152]

[0153] iii. The deviation of the dipeptide from the expected mean is calculated as follows:

[0154]

[0155] Through three steps of calculation, for each dipeptide, the present invention obtains the deviation of the given dipeptide from the expected average value, and thus the present invention obtains a feature vector with 20*20=400 dimensions.

[0156] 3) Physicochemical property characteristics: describes the statistical information related to the physicochemical properties of amino acids in the protein sequence.

[0157] a) Pseudoamino acid composition characteristics, including the original side chain mass, hydrophobicity and hydrophilicity of 20 natural amino acids, which are defined as H(i), HR i and HR j .

[0158] i. Normalize the physicochemical properties of 20 natural amino acids, the formula is as follows:

[0159]

[0160] ii. Calculate the correlation between the original side chain mass, hydrophobicity and hydrophilicity physicochemical properties of n amino acids, and set the two physicochemical properties of hydrophobicity and hydrophilicity as HR i and HR j , the formula is as follows:

[0161]

[0162] iii. For a sequence of length L, the ordered correlation factor of this sequence is defined as follows:

[0163]

[0164] Where j represents a positive integer, L is an integer used to define the maximum value of j; i represents the position of the amino acid in the protein sequence as i;

[0165] iv. The PAAC features of protein sequences are defined as follows:

[0166]

[0167] where f a represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect;

[0168] b) Quasi-sequence order is derived by measuring the physicochemical distances between the amino acids of a given protein sequence, thereby generating a sequence feature vector. For any given distance matrix:

[0169]

[0170] Where L is the length of the sequence, represents the element in the i-th row and i+d-th column of the distance matrix, and maxlag is the maximum lag value. a represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect. τ d represents the rank order coupling number, Xr represents the first 20 quasi-sequence order characteristics, and X represents the other quasi-sequence order characteristics.

[0171] 4) Features based on evolutionary information: refers to the use of a set of probabilities for each amino acid at each position to estimate the evolutionary relationship of different gene sequences. The position-specific scoring matrix consists of a set of probability scores for each amino acid (or gap) at each position in the alignment table, which is used to estimate the evolutionary conservation of genes. The basic idea of ​​PSSM is to match the query sequence in the database with the sequence in the alignment table, thereby providing a higher weight to the conservative position than the variable position. In this embodiment, the PSI-BLAST software is used to search for homologous protein sequences, thereby generating a PSSM matrix.

[0172] The PSSM matrix is ​​defined as follows:

[0173] Based on the PSSM matrix, three specific features are derived, including PSSM component features, RPMPSSM features, and S PSSM features;

[0174] a) PSSM component feature representation The rows of the original PSSM are transformed by summing and averaging the row vectors of amino acids belonging to the same type:

[0175]

[0176] Subject to

[0177]

[0178] Among them, R i refers to the i-th row of the transformation matrix, r k refers to the kth row of the original PSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids, and the result is a 20*20 matrix;

[0179] b) The RPM PSSM features are obtained by first setting the negative elements in the original PSSM to 0, retaining the positive values, and converting it into a new matrix (PPPSM); then using a formula similar to the PSSM component features, the RPM-PSSM features are calculated;

[0180]

[0181] Subject to

[0182]

[0183] Among them, R′ i refers to the i-th row of the filter matrix, r′ k refers to the kth row of PPSSM, δ krefers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids; the result is a 20*20 matrix;

[0184] c) The S PSSM feature is obtained by first setting the negative elements in the original PSSM to 0 and the values ​​greater than 7 to 7, and converting it into a new matrix S PSSM; the calculation method of S-FPSSM is as follows:

[0185]

[0186] Subject to

[0187]

[0188] Among them, fp k,j refers to the kth row and jth column of FPSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, α i Refers to the ith amino acid of the 20 natural amino acids; S j (i) refers to the SPSSM eigenvalue, and the result is a vector of length 400.

[0189] 5) Amino acid composition features: The protein sequence is converted into a digital vector, where each of the 20 natural amino acids is converted into a 20-dimensional feature vector, with the amino acid position set to 1 in alphabetical order and the remaining positions set to 0.

[0190] 6) Domain information feature: The presence or absence of protein domain is converted into a digital vector, where the domain is set to 1 and the non-domain is set to 0.

[0191] 3. Classification algorithm based on deep learning (CNN)

[0192] In this deep learning-based gene resistance classification algorithm, six key features are extracted from gene sequences, covering similarity, sequence information, physicochemical properties, evolutionary information, amino acid composition, and domain information. The extraction of these features not only helps capture the multifaceted information of genes, but also provides rich input for subsequent prediction models.

[0193] The fusion strategy is used to merge the six extracted features to form a comprehensive input vector. This fusion method aims to integrate the various features and present them to the deep learning model in a more comprehensive and integrated way. The fusion step is crucial to retain the correlation between features and improve model performance.

[0194] ResNet-50 is used as a deep convolutional neural network to build a classification model; the classification model is trained by taking the six features of the training set and merging the input vector. ResNet-50 is selected as a deep convolutional neural network for binary classification of genes, dividing the genes into two categories: drug-resistant and non-resistant. ResNet-50 has a deep residual network structure and is suitable for processing complex bioinformatics data. By training on labeled gene data, the model can learn the complex patterns of genes in drug resistance classification.

[0195] The trained classification model was tested by taking the merged input vector of the six features of the test set, and the classification model was evaluated to verify its accuracy and reliability in identifying drug-resistant genes.

[0196] Example 2

[0197] This embodiment provides a gene prediction method for virulence factors based on deep learning. For a specific flowchart, refer to Figure 2 The gene prediction method of virulence factors includes the following steps:

[0198] 1. Preprocessing

[0199] Virulence factor genes recorded in PATRIC, Victors, VFDB and UniProt databases were collected. In addition, negative samples were collected from the PBVF database. Genes with 100% similarity were removed using cd-hit software, and the remaining data was used as the database of the present invention for subsequent analysis.

[0200] The data set of this embodiment is divided into three subsets: 1) a training data set (m sequences), which is used as a training data set for the model of the present invention; 2) a test data set (n sequences), which is used to verify the accuracy of the model of the present invention; 3) a reference true set (o sequences, also known as a reference data set): reference data for comparison with the training data set and the test data set, and is also the largest of the three sub-datasets.

[0201] 2. Six categories of feature extraction

[0202] In order to obtain accurate prediction capabilities for antibiotic resistance genes, the present invention considers a variety of gene-related features, including the following six categories: similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, amino acid composition features, and domain information features.

[0203] 1) Similarity feature based on bit score: The bit score is evaluated by calculating the local alignment score between sequences. The higher the score, the higher the similarity between the sequences. The similarity feature is composed of the bit scores between the gene sequences in the training dataset and the sequences in the reference dataset. The bit score is calculated by DIAMOND software. As a result, each sequence in the training set will get an o-dimensional vector composed of bit scores.

[0204] 2) Sequence information features: Sequence information features include amino acid composition, dipeptide composition and the deviation of dipeptides from the expected average.

[0205] a) Amino acid composition refers to the frequency of 20 natural amino acids in a protein or peptide sequence, calculated as follows:

[0206] f(A) = N(A) / T;

[0207] Where A represents a given amino acid, N(A) represents the number of given amino acid A, T represents the total number of amino acids in the protein or peptide, and f(A) represents the number of occurrences of given amino acid A. Finally, a 20-dimensional feature vector with a sum of 1 is generated.

[0208] b) Dipeptide composition represents the frequency of dipeptides in a protein or peptide sequence. The calculation formula is as follows:

[0209] f(A1A2)=N(A1A2) / T;

[0210] Among them, A1A2 represents a certain dipeptide, N(A1A2) represents the number of given amino acids A1A2, T represents the total number of amino acids in the protein or peptide, and f(A1A2) represents the number of occurrences of given amino acids A1A2. Finally, a 20*20=400-dimensional feature vector with a sum of 1 is generated.

[0211] c) Deviation of dipeptides from the expected mean value composed of dipeptide composition, theoretical mean, and theoretical variance values.

[0212] i. The theoretical mean calculation formula is as follows:

[0213] TM(A1A2)=(CA1 / CA)*(CA2 / CA);

[0214] Among them, CA is 61, which refers to the number of codons encoding amino acids excluding three stop codons, CA1 encodes the number of codons for amino acid A1, and CA2 encodes the number of codons for amino acid A2.

[0215] ii. The theoretical variance calculation formula is as follows:

[0216]

[0217] iii. The deviation of the dipeptide from the expected mean is calculated as follows:

[0218]

[0219] Through three steps of calculation, for each dipeptide, the present invention obtains the deviation of the given dipeptide from the expected average value, and thus the present invention obtains a feature vector with 20*20=400 dimensions.

[0220] 3) Physicochemical property characteristics: describes the statistical information related to the physicochemical properties of amino acids in the protein sequence.

[0221] a) Pseudoamino acid composition characteristics, including the original side chain mass, hydrophobicity and hydrophilicity of 20 natural amino acids, which are defined as H(i), HR i and HR j .

[0222] i. Normalize the physicochemical properties of 20 natural amino acids, the formula is as follows:

[0223]

[0224] ii. Calculate the correlation between the original side chain mass, hydrophobicity and hydrophilicity physicochemical properties of n amino acids, and set the two physicochemical properties of hydrophobicity and hydrophilicity as HR i and HR j , the formula is as follows:

[0225]

[0226] iii. For a sequence of length L, the ordered correlation factor of this sequence is defined as follows:

[0227]

[0228] Where j represents a positive integer, L is an integer used to define the maximum value of j; i represents the position of the amino acid in the protein sequence as i;

[0229] iv. The PAAC features of protein sequences are defined as follows:

[0230]

[0231] where f a represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect;

[0232] b) Quasi-sequence order is derived by measuring the physicochemical distances between the amino acids of a given protein sequence, thereby generating a sequence feature vector. For any given distance matrix:

[0233]

[0234] Where L is the length of the sequence, represents the element in the i-th row and i+d-th column of the distance matrix, and maxlag is the maximum lag value. a represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect. τ d represents the rank order coupling number, Xr represents the first 20 quasi-sequence order characteristics, and X represents the other quasi-sequence order characteristics.

[0235] 4) Features based on evolutionary information: refers to the use of a set of probabilities for each amino acid at each position to estimate the evolutionary relationship of different gene sequences. The position-specific scoring matrix consists of a set of probability scores for each amino acid (or gap) at each position in the alignment table, which is used to estimate the evolutionary conservation of genes. The basic idea of ​​PSSM is to match the query sequence in the database with the sequence in the alignment table, thereby providing a higher weight to the conservative position than the variable position. In this embodiment, the PSI-BLAST software is used to search for homologous protein sequences, thereby generating a PSSM matrix.

[0236] The PSSM matrix is ​​defined as follows:

[0237] Based on the PSSM matrix, three specific features are derived, including PSSM component features, RPMPSSM features, and S PSSM features;

[0238] a) PSSM component feature representation The rows of the original PSSM are transformed by summing and averaging the row vectors of amino acids belonging to the same type:

[0239]

[0240] Subject to

[0241]

[0242] Among them, R i refers to the i-th row of the transformation matrix, r k refers to the kth row of the original PSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids, and the result is a 20*20 matrix;

[0243] b) The RPM PSSM features are obtained by first setting the negative elements in the original PSSM to 0, retaining the positive values, and converting it into a new matrix (PPPSM); then using a formula similar to the PSSM component features, the RPM-PSSM features are calculated;

[0244]

[0245] Subject to

[0246]

[0247] Among them, R′ i refers to the i-th row of the filter matrix, r′ k refers to the kth row of PPSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, L refers to the length of the sequence, and α i It refers to the ith amino acid of the 20 natural amino acids; the result is a 20*20 matrix;

[0248] c) The S PSSM feature is obtained by first setting the negative elements in the original PSSM to 0 and the values ​​greater than 7 to 7, and converting it into a new matrix S PSSM; the calculation method of S-FPSSM is as follows:

[0249]

[0250] Subject to

[0251]

[0252] Among them, fp k,j refers to the kth row and jth column of FPSSM, δ k refers to the average weight value, p k refers to the kth amino acid in the original sequence, α i Refers to the ith amino acid of the 20 natural amino acids; S j (i) refers to the S PSSM eigenvalue, resulting in a vector of length 400.

[0253] 5) Amino acid composition features: The protein sequence is converted into a digital vector, where each of the 20 natural amino acids is converted into a 20-dimensional feature vector, with the amino acid position set to 1 in alphabetical order and the remaining positions set to 0.

[0254] 6) Domain information feature: The presence or absence of protein domain is converted into a digital vector, where the domain is set to 1 and the non-domain is set to 0.

[0255] 3. Classification algorithm based on deep learning (CNN)

[0256] In this deep learning-based gene resistance classification algorithm, six key features are extracted from gene sequences, covering similarity, sequence information, physicochemical properties, evolutionary information, amino acid composition, and domain information. The extraction of these features not only helps capture the multifaceted information of genes, but also provides rich input for subsequent prediction models.

[0257] The fusion process adopts an ingenious Early Fusion strategy to better capture the comprehensive characteristics of genes.

[0258] Specifically, this embodiment first uses Early Fusion to fuse similarity features and evolutionary information features to form a first fusion feature vector. The design of this step is based on the possible correlation between similarity features and evolutionary information features in gene virulence, so as to more effectively capture the evolution and similarity information of genes.

[0259] Next, the first fused feature vector is obtained by fusing the amino acid composition features and the domain information features. This fusion process takes into account the biological structure and function of the gene, which helps to understand the characteristics of the gene more comprehensively.

[0260] Finally, this embodiment integrates the first fusion feature vector, the second fusion feature vector, the sequence information feature and the physical and chemical property feature to form a comprehensive fusion vector. The advantage of this process is that through multi-level feature fusion, the multi-dimensional information of the gene is better expressed, providing a richer input for the ResNet-50 model, thereby improving the accuracy and performance of the classification model.

[0261] ResNet-50 was used as a deep convolutional neural network to build a classification model; the classification model was trained by taking the combined input vector of the six features of the training set. ResNet-50 was selected as a deep convolutional neural network to predict or classify the virulence of genes, and the genes were divided into virulence and non-virulence factor genes. ResNet-50 has a deep residual network structure and is suitable for processing complex bioinformatics data. By training on labeled gene data, the model can learn the complex patterns of genes in drug resistance classification.

[0262] The trained classification model was tested by taking the merged input vector of the six features of the test set, and the classification model was evaluated to verify its accuracy and reliability in identifying drug-resistant genes.

[0263] Experimental Example 1

[0264] In order to verify the accuracy of drug resistance gene prediction or the feasibility of identification provided in Example 1, a data set unrelated to the training set was constructed to test the method of Example 1 and the three current software. The data set is based on the UniProt database and includes 200 antibiotic resistance genes and 200 non-antibiotic resistance genes.

[0265] Table 1 Results of drug resistance gene prediction accuracy

[0266]

[0267] The results show that Example 1 is superior to the other three software in all three parameters, that is, Example 1 has higher accuracy in the prediction or identification of antibiotic resistance genes.

[0268] Experimental Example 2

[0269] In order to verify the accuracy of gene prediction or feasibility of classification of virulence factors provided in Example 2, a data set unrelated to the training set was constructed to test the method of Example 2 and the three current software. The data set is based on the UniProt database and includes 200 virulence factor genes and 200 non-virulence factor genes.

[0270] Table 2 Gene prediction accuracy results of virulence factors

[0271]

[0272] The results showed that Example 2 was superior to the other three software in all three parameters, that is, Example 2 had higher accuracy in identifying virulence factor genes.

[0273] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning, characterized in that: It includes the following steps: S1. Obtain known antibiotic resistance gene data or virulence factor gene data, as well as negative sample gene data from the database; divide the collected data set into a training set, a test set, and a reference data set; S2. Extract the following six features using the gene data collected in S1: similarity features based on bit scores, sequence information features, physicochemical property features, evolutionary information features, amino acid composition features, and domain information features; the sequence information features include amino acid composition, dipeptide composition, and the deviation of dipeptides from the expected average value; the physicochemical property features describe statistical information related to the physicochemical properties of amino acids in protein sequences; the extraction of domain information features refers to: converting the presence or absence of protein domains into digital vectors; S3. When predicting the genes of virulence factors and / or drug resistance factors, the step S3 comprises: first pre-fusing at least two of the six features extracted in S2 to obtain a pre-fused feature vector; and then fusing the pre-fused feature vector with other features of the six features to form a comprehensive input vector; S4. constructing a classification model using a deep convolutional neural network; taking the six features of the training set and merging them into an input vector to train the classification model; S5. Take the six features of the test set and combine them into an input vector to test the classification model trained in S4, and perform model evaluation on the classification model.

2. According to the deep learning-based gene prediction or classification method for antibiotic resistance genes or virulence factors according to claim 1, genes located in the protein domain are set to 1, and genes located in the non-protein domain are set to 0.

3. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: In the step S1: obtaining known antibiotic resistance gene data from at least one database selected from the group consisting of UniProt database, CARD database, ARDB database, SARG database and NCRD database; The gene data of known virulence factors are obtained from at least one database selected from the group consisting of PATRIC, Victors, VFDB and UniProt databases.

4. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 3, characterized in that: When predicting the genes of virulence factors and / or drug resistance factors, step S3 comprises: (i) Pre-fusing the similarity features and the evolutionary information features to form a first fused feature vector; pre-fusing the amino acid composition features and the domain information features to obtain a second fused feature vector; (ii) fusing the first fused feature vector, the second fused feature vector, the sequence information feature and the physicochemical property feature to form a comprehensive input vector.

5. The method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: The deep convolutional neural network is selected from at least one of ResNet-18, ResNet-50, VGGNet, AlexNet network, Inception v3, Inception v4, Inception v5, DenseNet, NASNet, GoogleNet, MobileNets and EfficientNets.

6. The method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning according to claim 5, characterized in that: The similarity feature based on bit score refers to: the local comparison score between the gene sequence in the training data set and the sequence in the reference data set.

7. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 6, characterized in that: The bit scores were calculated using DIAMOND software.

8. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 6, characterized in that: The reference data set includes o sequences, and the bit score-based similarity feature of the training set obtains an o-dimensional vector composed of bit scores for each sequence.

9. The method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: in, (1) Amino acid composition refers to the frequency of 20 natural amino acids in a protein or peptide sequence. The calculation formula is as follows: ； Where A represents a given amino acid, N(A) represents the number of given amino acid A, T represents the total number of amino acids in a protein or peptide, and f(A) represents the number of occurrences of a given amino acid A; (2) Dipeptide composition refers to the frequency of dipeptides in a protein or peptide sequence; the calculation formula is as follows: ； in Represents a certain dipeptide, N( ) indicates a given amino acid The number of amino acids in a protein or peptide, T represents the total number of amino acids in a protein or peptide, and f( ) indicates a given amino acid The number of occurrences, finally generated Eigenvectors whose dimensions sum to 1; (3) The deviation of the dipeptide from the expected mean value is composed of the dipeptide composition, the theoretical mean, and the theoretical variance value; (i) The theoretical mean calculation formula is as follows: ； CA is 61, which refers to the number of codons encoding amino acids excluding the three stop codons. Encoding amino acids The number of codons, Encoding amino acids The number of codons; (ii) The theoretical variance calculation formula is as follows: ； (iii) The deviation of the dipeptide from the expected mean was calculated as follows: 。 10. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: The extraction of the physicochemical property characteristics includes extracting pseudoamino acid composition characteristics and quasi-sequence order; Wherein, the extraction of the pseudo amino acid composition characteristics comprises: The original side chain mass, hydrophobicity and hydrophilicity of 20 natural amino acids are defined as H(i), and ; The physicochemical properties of 20 natural amino acids are normalized to the following formula: ； Calculate the correlation between the original side chain mass, hydrophobicity and hydrophilicity physicochemical properties of n amino acids, and set the two physicochemical properties as and , the formula is as follows: ； For a sequence of length L, the ordered correlation factor of this sequence is defined as follows: ； Where j represents a positive integer, L is an integer used to define the maximum value of j; i represents the position of the amino acid in the protein sequence as i; The PAAC features of protein sequences are defined as follows: ； in represents the normalized frequency of amino acid A in the protein sequence, and ω represents the weight factor of the sequence order effect; The quasi-sequence order is derived by measuring the physicochemical distances between the amino acids of a given protein sequence, thereby generating a sequence feature vector; for any given distance matrix: ； ； ； Where L is the length of the sequence, represents the element in the i-th row and i+d-th column of the distance matrix, maxlag is the maximum lag value, represents the normalized frequency of amino acid A in the protein sequence, ω represents the weight factor of the sequence order effect; τ d represents the rank order coupling number, Xr represents the first 20 quasi-sequence order characteristics, and X represents the other quasi-sequence order characteristics.

11. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: The evolutionary information feature in step S2 refers to estimating the evolutionary relationship of different gene sequences based on a set of probability scores for each amino acid at each position.

12. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 11, characterized in that: PSI-BLAST software was used to search for homologous protein sequences to generate a PSSM matrix; the PSSM matrix was defined as follows: Three specific features are derived based on the PSSM matrix, including PSSM-component feature, RPM-PSSM feature, and S-PSSM feature; a) The PSSM-component features represent the rows of the original PSSM transformed by summing and averaging the row vectors of amino acids belonging to the same type: ； Subject to ； in, refers to the i-th row of the transformation matrix, refers to the kth row of the original PSSM, δ k is the average weight value, refers to the kth amino acid in the original sequence, L refers to the length of the sequence, It refers to the ith amino acid of the 20 natural amino acids, and the result is a 20*20 matrix; b) The RPM-PSSM feature is obtained by first setting the negative elements in the original PSSM to 0, retaining the positive values, and converting it into a new matrix (PPPSM); then using a formula similar to the PSSM-component feature to calculate the RPM-PSSM feature; ； Subject to ； in, refers to the i-th row of the filter matrix, refers to the kth row of PPSSM, δ k is the average weight value, refers to the kth amino acid in the original sequence, L refers to the length of the sequence, It refers to the ith amino acid of the 20 natural amino acids; the result is a 20*20 matrix; c) The S-PSSM feature is obtained by first setting the negative elements in the original PSSM to 0 and the values ​​greater than 7 to 7, and converting it into a new matrix S-PSSM; the calculation method of S-FPSSM is as follows: ; Subject to ； in, refers to the kth row and jth column of FPSSM, δ k is the average weight value, refers to the kth amino acid in the original sequence, Refers to the ith amino acid of the 20 natural amino acids; S j （i） refers to the S-PSSM eigenvalue, resulting in a vector of length 400.

13. The method for predicting or classifying antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: The amino acid composition feature extraction in step S2 refers to: converting the protein sequence into a digital vector, wherein each of the 20 natural amino acids is converted into a 20-dimensional feature vector, the amino acid position is set to 1 in alphabetical order, and the remaining positions are set to 0.

14. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: Negative samples were collected from the PBVF database, and genes with 100% similarity were removed using cd-hit software. The remaining data were used as the training set, test set, and reference data set.

15. The method for gene prediction or classification of antibiotic resistance genes or virulence factors based on deep learning according to claim 1, characterized in that: The step S5 also includes: classifying or predicting the drug resistance of the gene according to the evaluated optimal classification model, or classifying or predicting the virulence of the gene, and the method is not aimed at diagnosing the disease.

16. A deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors, characterized in that: include: Input module, control module and output module; The input module is configured to: input known antibiotic resistance gene data or virulence factor gene data, and negative sample gene data obtained from a database; The control module includes: a gene data extraction module, a feature fusion module, a classification model construction module and a classification model testing module; The gene data extraction module is configured to: extract similarity features based on bit scores, sequence information features, physical and chemical property features, evolutionary information features, amino acid composition features and domain information features of the data in the input module; The feature fusion module is configured to: merge the six features extracted by the gene data extraction module to form a comprehensive input vector; The classification model building module is configured to: use ResNet-50 as a deep convolutional neural network to build a classification model; take the six features of the training set and the merged input vector to train the classification model; The classification model testing module is configured to: take the six features of the test set and the combined input vector to test the classification model trained by the classification model building module, and perform model evaluation on the classification model; The output module is configured to output at least one of a classification model, a classification result and a prediction result.

17. The deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors according to claim 16, characterized in that: The control module also includes a data preprocessing module, which is configured to remove genes with 100% similarity from the data in the input module.

18. The deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors according to claim 16, characterized in that: When predicting or classifying genes of virulence factors, the feature fusion module includes a pre-fusion module and a fusion module; The pre-fusion module is configured to: pre-fuse similarity features and evolutionary information features to form a first fusion feature vector; The composition features and the structural domain information features are pre-fused to obtain a second fused feature vector; The fusion module is configured to fuse the first fused feature vector, the second fused feature vector, the sequence information feature and the physicochemical property feature to form a comprehensive input vector.

19. A deep learning-based gene prediction or classification device for antibiotic resistance genes or virulence factors, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the deep learning-based gene prediction or classification method for antibiotic resistance genes or virulence factors as described in any one of claims 1-14.

20. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the deep learning-based gene prediction or classification method for antibiotic resistance genes or virulence factors as described in any one of claims 1-14.

Images