A hybrid prediction method for virulence factors and antibiotic resistance genes
Through deep learning and classical ensemble learning combined with multiple gene features, a hybrid prediction model was constructed, which solved the problem of low accuracy and recall in the prediction of virulence factor and antibiotic resistance genes by existing tools, and achieved efficient and flexible prediction results.
Patent Information
- Application Number
- CN202210781902.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-30
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2042-06-30
AI Technical Summary
When existing bioinformatics tools identify virulence factors and antibiotic resistance genes, they have low prediction accuracy and recall, high false negative rates, and are sensitive to cutoff thresholds, so they cannot effectively identify non-conservative genes.
Deep learning neural networks and classic integrated learning architecture are adopted, combining a variety of core gene features, including building a hybrid prediction model based on alignment scores, gene evolution information and gene sequence information, and using multiple classification methods to optimize model performance.
High-precision and flexible prediction of virulence factors and antibiotic resistance genes is achieved, improving the accuracy and recall of prediction results, and is superior to traditional methods, especially in the recognition of novel virulence factors and antibiotic resistance genes.
Smart Images
Figure CN115171792B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning and bioinformatics, and in particular to a hybrid prediction method for virulence factors and antibiotic resistance genes. Background Art
[0002] Microbiome science is crucial for the internal ecosystems of hosts (such as humans, animals, and plants) and for maintaining the external environment. In particular, pathogenic microorganisms carry virulence factors (VFs) and antibiotic resistance genes (ARGs), causing disease and even threatening the life of their hosts. Accurate and timely identification of VFs and ARGs can effectively guide medical treatment, reduce host morbidity and mortality, and mitigate economic losses in the livestock and aquaculture industries.
[0003] Furthermore, despite their different evolutionary pathways, VFs and ARGs share common features that are essential for pathogens to adapt to and survive in competitive microbial environments. Specifically, both VFs and ARGs are frequently transferred between bacteria through horizontal gene transfer (HGT) and utilize similar systems (i.e., two-component systems, efflux pumps, cell wall alterations, and porins) to activate or repress the expression of various genes. Pathogens can exploit VFs to cause disease in their hosts, while they can establish themselves in environments with selective antibiotic pressure through the acquisition or presence of ARGs. Therefore, to understand the causal relationships between microbiome composition, function, and disease, it is essential to simultaneously identify VFs and ARGs. Simultaneous prediction of VFs and ARGs can save time in pathogen surveillance, especially for field detection of prevalent pathogens. However, traditional bioinformatics tools for identifying ARGs or VFs usually focus on independently predicting ARGs or VFs. The prediction tools are relatively backward, and the prediction accuracy and recall rate are relatively low. In addition, traditional prediction methods for VFs and ARGs have technical problems such as high false negative rate, high sensitivity to cutoff thresholds, and only able to identify conserved genes, resulting in poor prediction effect. Therefore, it is necessary to design a hybrid prediction method for virulence factors and antibiotic resistance genes. Summary of the Invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a hybrid prediction method for virulence factors and antibiotic resistance genes, so as to solve the technical problems that the prediction tools in the prior art are relatively backward, the prediction accuracy and recall rate are relatively low, and the existing prediction effect is relatively poor. By using computational methods including machine learning and deep learning neural networks, the prediction effect is better.
[0005] In order to achieve the above object, the technical solution of the present invention is as follows:
[0006] The present invention provides a hybrid prediction method for virulence factors and antibiotic resistance genes, comprising the following steps:
[0007] S1. Obtain known antibiotic resistance gene sequence data, virulence factor sequence data, and negative sample gene sequence data from the database respectively;
[0008] S2. Calculate multiple core gene features using gene sequence information and construct deep learning neural network architectures and classical ensemble learning architectures based on these core gene features.
[0009] S3. Take the three types of sequence data in S1 as samples, randomly select them as the total data set, and randomly divide them into five parts. In each division, four parts are used as training data sets, and the remaining part is used as test data sets.
[0010] S4. Use multiple classification methods to obtain a new training data set; build a classification model based on the extreme random tree for the new training data set, and obtain performance evaluation indicators of the classification model.
[0011] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, S1 specifically includes the following steps:
[0012] S11. Obtain known antibiotic resistance gene sequence data from the ARDB, CARD, and Uniprot databases;
[0013] S12. Obtain known virulence factor sequence data from the databases VFDB, PATRIC, Victors, and Uniprot;
[0014] S13. Obtain negative sample gene sequence data from the Uniprot database.
[0015] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, S2 includes the following specific steps:
[0016] S21. Using gene sequence information, calculate similarity features based on alignment scores, simple gene sequence features based on one-hot encoding, features based on gene evolution information, and features based on gene sequence information;
[0017] S22. Build a deep learning network architecture using similarity features based on alignment scores and simple gene sequence features based on one-hot encoding to train a neural network classification model in an end-to-end manner.
[0018] S23. Use features based on gene evolution information and features based on gene sequence information to construct a classic ensemble learning architecture, and use prior feature information to train a classic machine learning classification model.
[0019] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, the calculation of similarity features based on alignment scores in S21 includes the following specific steps:
[0020] The DIAMOND program was selected to align the gene sequences in the training dataset with the remaining known 12,724 ARGs and 30,945 VFs for comparison under sensitive parameters;
[0021] The training dataset has been deduplicated with the dataset used for comparison using the CD-HIT program, and the comparison scores have been normalized to the interval [0,1].
[0022] The bit score-based similarity feature of each gene sequence in the training dataset is converted into a fixed 12724+30945=43669-dimensional feature vector.
[0023] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, the gene evolution information-based features in S21 are composed of three specific features based on a position-specific scoring matrix, including PSSM-component features, RPM-PSSM features, and AADP-PSSM features;
[0024] The PSSM-composition signature is defined as follows:
[0025]
[0026]
[0027] Among them, R i represents the i-th row of the PSSM-composite feature matrix, r k represents the kth row of the normalized PSSM, p k represents the kth amino acid in the protein sequence, a i represents the i-th amino acid among the 20 standard amino acids;
[0028] The RPM-PSSM feature transforms the original PSSM by filtering negative values to 0 while leaving positive values unchanged. The idea of the RPM-PSSM feature comes from the residue probe method, that is, each amino acid corresponding to a specific column in the PSSM is regarded as a probe. The original PSSM is converted into a 400-dimensional feature vector, which is defined as follows:
[0029]
[0030]
[0031] Among them, M i represents the i-th row of the RPM-PSSM feature matrix, m k represents the kth row of PSSM, p k represents the kth amino acid in the protein sequence, a i represents the i-th amino acid among the 20 standard amino acids;
[0032] The AADP-PSSM feature extends the traditional AAC and DPC concepts to PSSM. AAC-PSSM converts the columns of the original PSSM profile into a fixed-length 20-dimensional feature vector, which is defined as follows:
[0033]
[0034] Among them, x j Indicates the replacement of the jth row of the AAC-PSSM feature matrix, representing the average proportion of amino acid mutations during evolution, p i，j represents the entity at row i and column j in the original PSSM;
[0035] The DPC-PSSM is converted into a fixed-length 400-dimensional feature vector to avoid information loss caused by X in the protein, which is defined as follows:
[0036]
[0037] By combining these two components, AADP-PSSM is converted into a fixed-length feature vector of 20+400=420 dimensions.
[0038] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, the features based on gene sequence information in S21 include amino acid composition features, dipeptide composition features, and features of deviations of dipeptides from expected average values;
[0039] The amino acid composition feature represents the frequency of 20 natural amino acids in the protein sequence and is calculated as follows:
[0040]
[0041] Where N(a) represents the number of specific amino acids a, N represents the sequence length of the protein or peptide, and f(a) represents the final generated 20-dimensional feature vector;
[0042] The dipeptide composition feature represents the frequency of dipeptides in a protein or polypeptide sequence, and is calculated as follows:
[0043]
[0044] Among them, Nab represents the number of given dipeptides ab, N represents the sequence length of the protein or peptide, and D(a, b) represents the final generated 400-dimensional feature vector.
[0045] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, the characteristic of the deviation of dipeptides from the expected mean is a combination of three features: theoretical mean TM, dipeptide composition DPC, and theoretical difference TV;
[0046] The calculation formula of TM features is as follows:
[0047]
[0048] Among them, C a and C b Represents the codon numbers encoding amino acids a and b respectively. N is equal to 61, representing the total number of possible codons excluding the three stop codons.
[0049] The calculation formula of TV characteristics is as follows:
[0050]
[0051] Where TM represents TM feature, TV represents TV feature, and N represents the sequence length of the protein or peptide.
[0052] The calculation formula of DDE characteristics is as follows:
[0053]
[0054] Among them, DPC represents DPC characteristics, TM represents TM characteristics, and TV represents TV characteristics.
[0055] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, the prior feature information in S23 is used to train classic machine learning classification models including random forest classification algorithm, extreme random tree classification algorithm, Xgboost classification algorithm, GradientBoosting classification algorithm and Adaboost classification algorithm.
[0056] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, S4 includes the following steps:
[0057] S41. Using multiple classification methods to perform a stacking algorithm, the prediction scores of the training data by different classification methods are used as a new training data set;
[0058] S42. Build a classification model based on the extreme random tree using the new training dataset, score the model using the test dataset, repeat the experiment five times, and take the average result of the five experiments as the performance evaluation indicator of the model.
[0059] As one aspect of a hybrid prediction method for virulence factors and antibiotic resistance genes, S41 specifically includes the following steps:
[0060] S411. Multiple base-level classification models are integrated through a meta-model;
[0061] S412. The base-level classification model is trained using the entire training dataset, and the meta-model uses the output of the base-level classification model as a training feature;
[0062] S413. Use 5-fold cross validation method to train the base-level classification model separately.
[0063] By adopting the above technical solution, the present invention has the following advantages:
[0064] 1. The present invention provides a hybrid prediction method for virulence factors and antibiotic resistance genes, which can make full use of multiple key core gene features and superimpose the power of classic ensemble learning methods and deep learning to efficiently and simultaneously predict potential virulence factors and antibiotic resistance genes. The prediction results have strong scientific performance and high accuracy.
[0065] 2. The present invention can simultaneously and accurately predict virulence factors, drug-resistance genes, and negative sample genes (neither virulence factors nor antibiotic-resistance genes), and can flexibly and accurately predict each individually. This overcomes the shortcomings of traditional best-hit methods, such as high false-negative rates, high sensitivity to cutoff thresholds, and the ability to only identify conserved genes, achieving relatively good prediction results.
[0066] 3. The present invention surpasses previous traditional prediction tools in terms of accuracy and recall of novel virulence factors and resistance genes, virulence factors and resistance genes in real metagenomic data, and pseudo-virulence factors and resistance genes (gene fragments); the present invention uses computational methods including machine learning and deep learning neural networks, and its results are competitive with all the most advanced prediction tools and have higher scientific performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Other features, objects and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.
[0068] Figure 1 It is a flow chart of a hybrid prediction method of virulence factors and antibiotic resistance genes of the present invention;
[0069] Figure 2 The figure is a comparison of the histogram results of the hybrid prediction method of the present invention and other calculation methods in simultaneously predicting virulence factors and antibiotic resistance genes. DETAILED DESCRIPTION
[0070] The technical solution of the present invention is specifically described below in conjunction with embodiments. The detailed features and advantages of the present invention are described in detail in the specific implementation methods. The content is sufficient to enable any person skilled in the art to understand the technical content of the present invention and implement it accordingly. According to the description, claims and drawings disclosed in this specification, those skilled in the art can easily understand the relevant purposes and advantages of the present invention.
[0071] See Figure 1 As shown, a hybrid prediction method for virulence factors and antibiotic resistance genes in microbial data includes the following steps:
[0072] S1. Obtain known antibiotic resistance gene sequence data, virulence factor sequence data, and negative sample gene sequence data (neither antibiotic resistance genes nor virulence factors) from the database;
[0073] S1 includes the following specific steps:
[0074] S11. Obtain known antibiotic resistance gene sequence data from the ARDB, CARD, and Uniprot databases;
[0075] S12. Obtain known virulence factor sequence data from the databases VFDB, PATRIC, Victors, and Uniprot;
[0076] S13. Obtain negative sample gene sequence data from the Uniprot database.
[0077] S2. Calculate multiple core gene features using gene sequence information and construct deep learning neural network architectures and classical ensemble learning architectures based on these core gene features.
[0078] S2 includes the following specific steps:
[0079] S21. Since multiple core gene features include similarity features based on alignment scores, features based on gene evolution information, features based on gene sequence information, and simple gene sequence features based on one-hot encoding; therefore, gene sequence information is used to calculate similarity features based on alignment scores, simple gene sequence features based on one-hot encoding, features based on gene evolution information, and features based on gene sequence information respectively.
[0080] The alignment score-based similarity feature is composed of the alignment scores of virulence factors and antibiotic resistance genes with known virulence factors and antibiotic resistance genes. This feature takes into account the similarity distribution of sequences in the ARGs and VFs databases, rather than just the best hit rate. The alignment score is used as a similarity metric because, unlike the e-value, it takes into account the degree of identity between sequences and is independent of the size of the database.
[0081] Calculating similarity features based on comparison scores in S21 includes the following specific steps:
[0082] The DIAMOND program, which is faster than BLAST, was selected to align the gene sequences in the training dataset with the remaining known 12,724 ARGs and 30,945 VFs for comparison under sensitive parameters;
[0083] The training dataset has been deduplicated with the comparison dataset using the CD-HIT program to avoid the possibility of label leakage. The alignment score is normalized to the interval [0, 1] to represent the similarity of the sequences in terms of distance.
[0084] The bit-score-based similarity feature of each gene sequence in the training dataset is converted into a fixed 12724+30945=43669-dimensional feature vector, where each dimension is the alignment score output by the DIAMOND program between the full-length gene sequence and each available ARG and VF in the comparison dataset.
[0085] The features based on gene evolution information are composed of three specific features based on the position-specific score matrix (PSSM), including PSSM-composition features, RPM-PSSM features, and AADP-PSSM features. The PSSM-composition features are defined as follows:
[0086]
[0087]
[0088] Among them, R i represents the i-th row of the PSSM-composite feature matrix, r k represents the kth row of the normalized PSSM, p k represents the kth amino acid in the protein sequence, a i Indicates the i-th amino acid among the 20 standard amino acids.
[0089] The RPM-PSSM feature transforms the original PSSM by filtering negative values to 0 while leaving positive values unchanged. This method is inspired by the residue probe method, which treats each amino acid corresponding to a specific column in the PSSM as a probe. Ultimately, the original PSSM is converted into a 400-dimensional feature vector, using the following definition:
[0090]
[0091]
[0092] Among them, M i represents the i-th row of the RPM-PSSM feature matrix, m k represents the kth row of PSSM, p k represents the kth amino acid in the protein sequence, a i Indicates the i-th amino acid among the 20 standard amino acids.
[0093] The AADP-PSSM feature extends the traditional AAC and DPC concepts to PSSM. AAC-PSSM converts the columns of the original PSSM profile into a fixed-length 20-dimensional feature vector, which is defined as follows:
[0094]
[0095] Among them, x j Indicates the replacement of the jth row of the AAC-PSSM feature matrix, representing the average proportion of amino acid mutations during evolution, p i，j represents the entity at row i and column j in the original PSSM. Secondly, the DPC-PSSM is converted into a fixed-length 400-dimensional feature vector to avoid information loss caused by X in the protein, which is defined as follows:
[0096]
[0097] By combining these two components, AADP-PSSM is converted into a fixed-length feature vector of 20+400=420 dimensions.
[0098] Features based on gene sequence information include amino acid composition feature (AAC), dipeptide composition feature (DPC), dipeptide deviation from expected average feature (DDE), pseudo amino acid composition feature (PAAC) feature and quasi-sequence order feature (QSO).
[0099] Among them, the amino acid composition feature (AAC) represents the frequency of 20 natural amino acids (i.e., ACDEFGHIKLMNPQRSTVWY) in the protein sequence and can be calculated as:
[0100]
[0101] Among them, N(a) represents the number of specific amino acids a, N represents the sequence length of the protein or peptide, and f(a) represents the final generated 20-dimensional feature vector.
[0102] The DPC feature represents the frequency of dipeptides in a protein or polypeptide sequence and can be calculated as:
[0103]
[0104] where N ab represents the number of given dipeptides ab, N represents the sequence length of the protein or peptide, and D(a,b) represents the final generated 400-dimensional feature vector. The calculation of DPC features refers to the previous description.
[0105] The DDE feature is a combination of three features: theoretical mean (TM), dipeptide composition (DPC), and theoretical difference (TV). Specifically, the TM feature is calculated as follows:
[0106]
[0107] Among them, C a and C b Represents the codon numbers encoding amino acids a and b respectively; C N is equal to 61, representing the total number of possible codons excluding the three stop codons.
[0108] The calculation method of TV features is as follows:
[0109]
[0110] Wherein, TM represents TM feature, the calculation of which refers to the previous description, and N represents the sequence length of the protein or peptide.
[0111] The calculation method of DDE characteristics is as follows:
[0112]
[0113] Wherein, TM represents TM feature, TV represents TV feature, and the calculation refers to the previous description.
[0114] S22. Build a deep learning network architecture using similarity features based on alignment scores and simple gene sequence features based on one-hot encoding to train a neural network classification model in an end-to-end manner.
[0115] S23. Use features based on gene evolution information and features based on gene sequence information to construct a classic ensemble learning architecture, and use prior feature information to train a classic machine learning classification model.
[0116] Among them, the prior feature information in S23 trains classic machine learning classification models including the Random Forest classification algorithm, the Extra Trees classification algorithm, the Xgboost classification algorithm, the GradientBoosting classification algorithm, and the Adaboost classification algorithm.
[0117] S3. Randomly sample the three types of data from S1, using the known antibiotic resistance gene data as the first type of samples, the known virulence factor sequence information data as the second type of samples, and the known negative sample gene sequence information data as the third type of samples. The entire training dataset is randomly divided five times, each time into five parts, with four parts used as training datasets and the remaining part as test datasets.
[0118] S4. Use multiple classification methods to obtain a new training data set, build a classification model for the new training data set based on extreme random trees, and obtain performance evaluation indicators of the classification model.
[0119] S4 specifically includes the following steps:
[0120] S41. Using multiple classification methods to perform a stacking algorithm, the prediction scores of the training data by different classification methods are used as a new training data set;
[0121] To achieve superior prediction performance for virulence factors and antibiotic resistance genes, we combined the power of classical machine learning methods and deep learning in a single stacking algorithm.
[0122] S41 specifically includes the following steps:
[0123] S411. Multiple base-level classification models are integrated through a meta-model;
[0124] S412. The base-level classification model is trained using the entire training dataset, and the meta-model uses the output of the base-level classification model as a training feature;
[0125] S413. Use 5-fold cross validation to train the basic classification model separately to solve the overfitting phenomenon in the final prediction. In a specific embodiment, the stacking algorithm in the present invention is shown in the pseudo code shown in the algorithm as shown in Table 1 below
[0126] Table 1: The stacking algorithm is represented by the pseudo code shown in the algorithm.
[0127]
[0128] S42. Build a classification model based on the extreme random tree using the new training dataset, score the model using the test dataset, repeat the experiment five times, and take the average result of the five experiments as the performance evaluation indicator of the model.
[0129] Example 2
[0130] In order to better illustrate the effectiveness of the prediction method of the present invention, we implemented a rigorous procedure and used a cross-validation step to obtain an unbiased evaluation of the effectiveness of the present invention. Table 2 lists the results of the mixed prediction of virulence factors and drug resistance genes in this example using a five-fold cross-validation method:
[0131] Table 2 Results of simultaneous prediction of virulence factors and drug resistance genes by the present invention under five-fold cross validation
[0132]
[0133] In Table 2: Precision: precision rate Recall: recall rate F1-score: F1 score, VFs: virulence factors, ARGs: drug resistance genes NSs: negative sample genes Micro-average: micro average
[0134] As can be seen from Table 2 above, this embodiment achieved high evaluation scores in multiple cross-experimental results. This result shows that the present invention can not only achieve the simultaneous prediction of virulence factors and antibiotic resistance genes, but also has excellent performance in terms of precision and recall rate.
[0135] Example 3
[0136] In order to test the predictive ability of the present invention for unknown virulence factors (VFs), drug resistance genes (ARGs) and negative sample genes (NSs), the present invention constructed an independent data set including 209 ARGs, 209 VFs and 209 NSs. It is worth noting that these unknown genes are completely independent of the genes in the training data set. By setting the identity threshold of CD-HIT to 100%, all identical or repeated sequences were removed. In addition, we also introduced the currently available VRprofile model (the latest computational model) as a comparison method and the traditional "best hit" method as a baseline (using the Diamond sequence alignment tool) under three different parameters as a comparison method. Table 3 lists the results of the embodiment of the present invention (HyperVR), the VRprofile model and the Diamond sequence alignment tool as a comparison method under three different parameters for the simultaneous prediction of unknown virulence factors and drug resistance genes:
[0137] Table 3: Results of simultaneous prediction of unknown virulence factors and drug resistance genes by the embodiment of the present invention, the VRprofile model, and the baseline comparison method
[0138]
[0139] In the table: Precision: precision rate, Recall: recall rate, F1-score: F1 score, VFs: virulence factors, ARGs: drug resistance genes, NSs: negative sample genes, Micro-average: micro average
[0140] As shown in Table 3, when comparing the present invention (HyperVR) with the VRprofile model and the Diamond sequence alignment tool under three different parameters, the present invention (HyperVR) achieved the highest evaluation score in all experimental results, and outperformed other baseline comparison methods in terms of precision and recall.
[0141] Figure 2 The comparative histogram results of the simultaneous prediction of unknown virulence factors and drug resistance genes by the embodiment of the present invention (HyperVR), VRprofile model (the latest computational model) and baseline comparison method (including three different parameters) are shown. Figure 2 In the figure, column a represents the F1 score, column b represents the recall; column c represents the precision, Diamond-81%, Diamond-64%, and Diamond-21% represent the baselines of the Diamond sequence alignment tool under three different parameters as the comparison method; the height of the bar graph represents the prediction performance of the method, which is shown in Figure 2. Figure 2 From the comparison of the histograms in , it can be seen that the embodiment of the present invention (HyperVR) has higher prediction performance than the latest calculation model (VRprofile) and the baseline comparison method (including three different parameters), and its overall performance is better than other models. Compared with the latest prediction methods, its results are more competitive, have higher scientific performance, and have the best prediction effect.
[0142] Finally, it should be pointed out that although the present invention has been described with reference to the current specific embodiments, ordinary technicians in this technical field should realize that the above embodiments are only used to illustrate the present invention and are not used to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, as long as the changes and modifications to the above embodiments are within the scope of the essential spirit of the present invention, they will fall within the scope of the claims of the present invention.
Claims
1. A hybrid prediction method for virulence factors and antibiotic resistance genes, characterized in that: The following steps are involved: S1. Obtain known antibiotic resistance gene sequence data, virulence factor sequence data, and negative sample gene sequence data from the database respectively; S2. Calculate multiple core gene features using gene sequence information and construct deep learning neural network architectures and classical ensemble learning architectures based on these core gene features. The S2 includes the following specific steps: S21. Using gene sequence information, calculate similarity features based on alignment scores, simple gene sequence features based on one-hot encoding, features based on gene evolution information, and features based on gene sequence information; S22. Build a deep learning network architecture using similarity features based on alignment scores and simple gene sequence features based on one-hot encoding to train a neural network classification model in an end-to-end manner. S23. Build a classic ensemble learning architecture using features based on gene evolution information and features based on gene sequence information, and train a classic machine learning classification model using prior feature information. S3. Take the three types of sequence data in S1 as samples, randomly select them as the total data set, and randomly divide them into five parts. In each division, four parts are used as training data sets, and the remaining part is used as test data sets. S4. Use multiple classification methods to obtain new training data sets; Build a classification model for a new training data set based on extreme random trees and obtain performance evaluation indicators of the classification model; The S4 comprises the following steps: S41. We utilize multiple classification methods in a stacking algorithm, using the prediction scores of different classification methods on the training data as a new training dataset. To achieve superior prediction performance for virulence factors and antibiotic resistance genes, we combine the power of classic machine learning methods and deep learning in a single stacking algorithm. The S41 specifically includes the following steps: S411. Multiple base-level classification models are integrated through a meta-model; S412. The base-level classification model is trained using the entire training dataset, and the meta-model uses the output of the base-level classification model as a training feature; S413. Use 5-fold cross validation to train the base-level classification model separately; S42. Build a classification model based on the extreme random tree using the new training dataset, score the model using the test dataset, repeat the experiment five times, and take the average result of the five experiments as the performance evaluation indicator of the model.
2. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 1, characterized in that: The S1 specifically includes the following steps: S11. Obtain known antibiotic resistance gene sequence data from the ARDB, CARD, and Uniprot databases; S12. Obtain known virulence factor sequence data from the databases VFDB, PATRIC, Victors, and Uniprot; S13. Obtain negative sample gene sequence data from the Uniprot database.
3. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 1, characterized in that: The calculation of the similarity feature based on the comparison score in S21 includes the following specific steps: The DIAMOND program was selected to align the gene sequences in the training dataset with the remaining known 12,724 ARGs and 30,945 VFs for comparison under sensitive parameters; The training dataset has been deduplicated with the dataset used for comparison using the CD-HIT program, and the comparison scores have been normalized to the interval [0,1]. The bit score-based similarity feature of each gene sequence in the training dataset is converted into a fixed 12724+30945=43669-dimensional feature vector.
4. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 1, characterized in that: The gene evolution information-based features in S21 are composed of three specific features based on a specific position score matrix, including PSSM-component features, RPM-PSSM features, and AADP-PSSM features; The PSSM-composition signature is defined as follows: Among them, R i represents the i-th row of the PSSM-composite feature matrix, r k represents the kth row of the normalized PSSM, p k represents the kth amino acid in the protein sequence, a i represents the i-th amino acid among the 20 standard amino acids; The RPM-PSSM feature transforms the original PSSM by filtering negative values to 0 while leaving positive values unchanged. The idea of the RPM-PSSM feature comes from the residue probe method, that is, each amino acid corresponding to a specific column in the PSSM is regarded as a probe. The original PSSM is converted into a 400-dimensional feature vector, which is defined as follows: Among them, M i represents the i-th row of the RPM-PSSM feature matrix, m k represents the kth row of PSSM, p k represents the kth amino acid in the protein sequence, a i represents the i-th amino acid among the 20 standard amino acids; The AADP-PSSM feature extends the traditional AAC and DPC concepts to PSSM. AAC-PSSM converts the columns of the original PSSM profile into a fixed-length 20-dimensional feature vector, which is defined as follows: Among them, x j Indicates the replacement of the jth row of the AAC-PSSM feature matrix, representing the average proportion of amino acid mutations during evolution, p i,j represents the entity at row i and column j in the original PSSM; The DPC-PSSM is converted into a fixed-length 400-dimensional feature vector to avoid information loss caused by X in the protein, which is defined as follows: By combining these two components, AADP-PSSM is converted into a fixed-length feature vector of 20+400=420 dimensions.
5. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 1, characterized in that: The features based on gene sequence information in S21 include amino acid composition features, dipeptide composition features, and features of deviations of dipeptides from expected average values; The amino acid composition feature represents the frequency of 20 natural amino acids in the protein sequence, and the calculation formula is as follows: Where N(a) represents the number of specific amino acids a, N represents the sequence length of the protein or peptide, and f(a) represents the final generated 20-dimensional feature vector; The dipeptide composition feature represents the frequency of dipeptides in a protein or polypeptide sequence, and is calculated as follows: Among them, N ab represents the number of given dipeptides ab, N represents the sequence length of the protein or peptide, and D(a, b) represents the final generated 400-dimensional feature vector.
6. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 5, characterized in that: The characteristic of the deviation of the dipeptide from the expected mean is a combination of three characteristics: the theoretical mean TM, the dipeptide composition DPC and the theoretical difference TV; The calculation formula of TM features is as follows: Among them, C a and C b The codon numbers for amino acids a and b, C N is equal to 61, representing the total number of possible codons excluding the three stop codons; The calculation formula of TV feature is as follows: Where TM represents TM features, TV represents TV features, and N represents the sequence length of the protein or peptide; The calculation formula of DDE characteristics is as follows: Among them, DPC represents DPC characteristics, TM represents TM characteristics, and TV represents TV characteristics.
7. The hybrid prediction method of virulence factors and antibiotic resistance genes according to claim 1, characterized in that: The prior feature information in S23 is used to train classic machine learning classification models including a random forest classification algorithm, an extreme random tree classification algorithm, an Xgboost classification algorithm, a GradientBoosting classification algorithm, and an Adaboost classification algorithm.
Citation Information
Patent Citations
piRNA-disease incidence relation prediction method based on convolution denoising autoencoder
CN113724790A
Bioinformatic processes for determination of peptide binding
US20130330335A1