Prediction method of plant pri-miRNA-encoded peptides based on improved MRMD algorithm and DF model

Abstract

The present invention proposes a method for predicting plant pri-miRNA encoded peptides based on an improved MRMD algorithm and a DF model, and belongs to the fields of bioinformatics and computer science technology. The present invention designs an improved IMMRMD feature selection algorithm, which makes it more suitable for feature data generated by sORF extracted from pri-miRNA, and combines deep forest to realize the prediction of sORF encoded peptides. The present invention can determine whether the input RNA sequence meets the relevant characteristics of the first sORF on the pri-miRNA, thereby providing a basis for determining whether it encodes a peptide, providing a reference for biological experiments, and saving a lot of manpower and material resources.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and computer science technology, and relates to a method for applying an improved MRMD feature selection algorithm to biological sequence data and combining a deep forest (DF) model to construct a plant pri-miRNA encoding peptide prediction model. Background Art

[0002] There are a huge number of RNA sequences in organisms, and researchers divide them into two categories based on whether they can encode proteins, coding RNA and non-coding RNA (ncRNA). As research continues to deepen, it has been found that some ncRNAs also have the ability to encode proteins. MicroRNA (miRNA), as a type of ncRNA, plays an important role in key activities of organisms such as virus defense and organ generation. Mature miRNA is produced by a series of nuclease cleavage processing of a longer primary transcript. The primary transcript is called miRNA primary body (primarymiRNA, pri-miRNA). Since pri-miRNA exists for a short time after being transcribed and is easily cleaved to form miRNA precursor (precursor miRNA, pre-miRNA), there have been few studies on it before. However, studies in recent years have shown that pri-miRNA contains small open reading frames (sORF), and its encoded peptide (miRNA-encoded peptide, miPEP) can promote the accumulation of corresponding miRNA, thereby indirectly regulating the expression of target genes, which has an important regulatory effect on organisms. The discovery of miPEP has provided new ideas for the development of modern agriculture and other fields, making its research a hot topic. How to correctly identify miPEP is crucial for in-depth research on the function of miPEP.

[0003] At present, the use of computational methods to predict plant miPEPs is a relatively new research direction, and relatively few results have been achieved. The main methods used to study it include biological experiments and bioinformatics. The results of biological experiments are highly credible, but they are time-consuming and labor-intensive and cannot be verified on a large scale; bioinformatics methods can quickly identify miPEPs by building machine learning models, providing references for biological experimental identification, thereby saving a lot of time and cost. At present, there are few studies on small open reading frame-encoded peptides (sORF-encoded peptides, SEPs), and existing identification tools are difficult to identify highly credible miPEPs, so there is an urgent need for a fast and accurate bioinformatics tool to assist in the identification of miPEPs. However, the use of bioinformatics to predict plant miPEPs faces great challenges. First, due to the lack of a pri-miRNA database, certain methods need to be used to expand the pre-miRNA sequence to obtain a data set. At the same time, how to use appropriate feature extraction and feature selection methods, and how to select appropriate models to predict sORFs in pri-miRNAs are all problems that must be solved when mining plant miPEPs using bioinformatics methods. Summary of the invention

[0004] Based on the above-mentioned problems to be solved, the present invention adopts a more suitable evaluation coefficient to propose a feature selection algorithm IMMRMD (Improved Max-Relevance-Max-Distance based dimensionality reduction) for the situation that the MRMD feature selection algorithm has poor processing effect on specific biological sequence data, improves the representativeness of feature data, and combines the DF model to achieve the prediction of sORF on plant pri-miRNA. The present invention can determine whether the input RNA sequence meets the relevant characteristics of the first sORF on the pri-miRNA, thereby providing a basis for determining whether the sequence encodes a peptide, providing a reference for biological experiments, and saving a lot of manpower and material resources.

[0005] The technical solution of the present invention is:

[0006] A method for predicting plant pri-miRNA coding peptides based on an improved MRMD algorithm and a DF model comprises the following steps:

[0007] Step 1: Construct the experimental data set. The positive set uses the first sORF obtained from the Arabidopsis pri-miRNA sequence, and the negative set uses the Arabidopsis non-coding sequence (NCDS) data.

[0008] The specific process is as follows:

[0009] 1.1: First, download all the miRNA data of Arabidopsis from the miRBase database (https: / / www.mirbase.org / ) to obtain its precursor miRNA (pre-miRNA). Then, use the Ensemble Plants (http: / / plants.ensembl.org / index.html) tool to expand the data, expand all the obtained pre-miRNAs forward by 1000 bases, and then use the softberry software (http: / / www.softberry.com / berry.phtml?topic=tssplant&group=programs&subgroup=promoter) to find its transcription start site (TSS), so as to obtain the corresponding pri-miRNA. Then use the ORF Finder (https: / / www.ncbi.nlm.nih.gov / orffinder / ) tool to obtain the first sORF sequence of each pri-miRNA, and use the CD-HIT tool to remove sequences with a similarity higher than 80% to obtain the candidate positive set sORF data.

[0010] 1.2: Collect Arabidopsis NCDS data and use it as the negative data source. Use the ORF Finder tool to obtain the sORFs sequence, and remove the sORF sequences with a similarity higher than 80% to obtain candidate negative data.

[0011] 1.3: Select 10% of the total data, that is, randomly select 24 items from the positive set and 25 items from the negative set as the test set to test the prediction performance of the final model.

[0012] Step 2: Extract relevant features of sORF sequence data.

[0013] According to the differences in sequence composition and physicochemical properties between the positive and negative data sets, features such as k-mer features, short sequence motifs (SSM) features, signal-to-noise ratio, GC base content (GC_con), GC base ratio (GC_ra) and sequence length were extracted from both the positive and negative data sets.

[0014] Among them, the k-mer feature describes the composition information of the sORF sequence. It fully considers the dependency between adjacent nucleotides and preserves the information of adjacent nucleotides, thus avoiding the loss of information. k represents the number of adjacent bases in the sequence. There are 4 kIn the present invention, k is 1, 2, 3, 4, with a total of 340 dimensions. A sliding window with a step size of 1 is used for sliding matching, and the formula is as follows:

[0015]

[0016] l k =L-k+1,k=1,2,3,4 (2)

[0017] Among them, l k is the number of slides, s i is the number of occurrences of each k-mer, f i is the final frequency of each k-mer, L is the sequence length, and finally 340-dimensional features are obtained.

[0018] The k-mer feature takes into account the properties of continuous bases, but there are also differences between discontinuous bases, so the SSM feature is used to supplement it. Different base pairs can be recorded as A*B, A**B, A***B, etc., where A and B are any base among A, T, C, and G, and * represents any base. Each group of SSM has a 16-dimensional feature expression. Three groups of SSMs, A*B, A**B, and A***B, are extracted, with a total of 48 dimensions.

[0019]

[0020] u j =Lj-1,j=1,2,3 (4)

[0021] Among them, u j is the number of slides, v i′ is the number of occurrences of each short sequence motif, SSM i′ is the frequency of occurrence of each short sequence motif.

[0022] According to the expression of the sequence itself, the sequence length L, GC base content GC_con, GC base ratio GC_ra and signal-to-noise ratio features are extracted, a total of 4 dimensions. The size of the signal-to-noise ratio can indicate the bias of base usage in the sequence. The power spectrum of sORF at one-third is calculated by the distribution of bases in three phases to obtain the signal-to-noise ratio. The specific formula is as follows:

[0023]

[0024]

[0025]

[0026]

[0027] Where g and c represent the number of bases G and C in the sequence, respectively; L represents the length of the sequence; a x 、b x and c x They are all arrays with one row and three columns, representing the frequency of base x appearing in three phases; W represents the power spectrum; SNR represents the signal-to-noise ratio.

[0028] Finally, the extracted features constitute 392-dimensional feature data.

[0029] Step 3: Design the IMMRMD algorithm and use it to sort and select features of the feature data obtained in step 2.

[0030] Due to the specificity of the data of the present invention, the MRMD algorithm has been improved accordingly. The MRMD algorithm uses the Pearson coefficient as the criterion for judging the importance of features. However, it requires the data to be normally distributed and free of abnormal points, which places relatively high demands on the data. The experimental data of the present invention does not meet this requirement. Therefore, the Spearman coefficient is selected as the judgment criterion. The Spearman coefficient requires a small amount of data, which is in line with the data situation of this experiment. At the same time, it does not require whether the data is normally distributed during evaluation, which is more in line with the actual data situation. The calculation formula of the Spearman correlation coefficient is shown in formula (9), which defines X and Y as two groups of data.

[0031]

[0032] Among them, d y For X y and Y y The level difference between s Between -1 and 1; n is the total number of samples.

[0033] On the other hand, we choose to use adjusted cosine similarity instead of cosine similarity as the criterion for judging the maximum distance. Cosine similarity is a measure of the similarity between two non-zero vectors in the inner product space. It measures the cosine value of the angle between them. It focuses on the difference in direction, but is insensitive to the absolute value. Therefore, it cannot measure the difference in the values ​​of each dimension. Adjusted cosine similarity is an improved form based on vector similarity. While considering the difference in direction, it also considers the difference in value. Its purpose is to solve the problem that cosine similarity only considers the similarity in the direction of the vector dimension and ignores the difference in the dimensions of each dimension. Therefore, when calculating the similarity, a correction operation is performed to subtract the mean of each dimension to make up for the deficiency of cosine similarity.

[0034] In summary, the present invention improves the original evaluation criteria of the original MRMD feature selection algorithm accordingly and designs the IMMRMD algorithm; the specific steps of the IMMRMD algorithm are as follows: the original feature data extracted in step 2 is used as the input of the IMMRMD algorithm, the Spearman correlation coefficient is first used to calculate the correlation between the sub-feature set and the target category, and the sub-feature set with the highest correlation with the target category is selected; secondly, considering that reducing the redundancy between features can better extract relevant features, the maximum distance is used to measure the similarity between two feature vectors. The Euclidean distance is calculated, the cosine similarity and the Tanimoto coefficient are adjusted, and the degree of redundancy between any feature vector and the remaining feature vectors is obtained. Finally, by accumulating and sorting the calculation results of the two parts, the sorted feature data is input into the logistic regression model dimension by dimension for evaluation, and the first m dimensions with the best classification effect are selected, thereby obtaining an m-dimensional sub-feature set with both maximum correlation and minimum redundancy.

[0035] Finally, the original MRMD3.0 algorithm reduces the dimensionality of 392-dimensional feature data to 314 dimensions, while the IMMRMD algorithm of the present invention selects 382 dimensions after changing the evaluation criteria.

[0036] Step 4: Use the DF model to get the prediction results.

[0037] Input the data filtered in step 3 into the DF model to obtain the final prediction result.

[0038] Beneficial effects of the present invention:

[0039] (1) The dependency relationship of nucleotides in gene sequences was comprehensively considered, and relevant features such as k-mer and SSM were extracted.

[0040] (2) The MRMD feature selection algorithm is improved accordingly, and the IMMRMD feature selection algorithm is proposed, which makes the selected feature data more representative and improves the classification effect.

[0041] (3) DF was selected as the classification model in the comparison of multiple models to obtain key feature information, thereby further improving the prediction performance of the present invention for plant pri-miRNA encoded peptides. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall framework flow chart of the present invention.

[0043] Figure 2 This is a comparison chart of the classification effects of the original feature data of the present invention in different models.

[0044] Figure 3This is a diagram showing the classification effect of feature data on SVM after the feature selection algorithm of the present invention.

[0045] Figure 4 This is a diagram showing the classification effect of feature data on LR after the feature selection algorithm of the present invention. DETAILED DESCRIPTION

[0046] The specific implementation of the present invention is further described below in conjunction with the attached drawings, tables and technical solutions.

[0047] The present invention provides a pri-miRNA encoded peptide prediction method combining feature selection algorithm and machine learning, by improving the MRMD algorithm commonly used in bioinformatics to extract more representative features, and then selecting DF with better prediction effect as the prediction model, thereby realizing the prediction of pri-miRNA encoded peptides. The overall framework of the method is as follows Figure 1 As shown, the steps are as follows:

[0048] Step 1: Obtain the sORF sequences in the corresponding pri-miRNA of Arabidopsis as the positive set and the NCDS sequences as the negative set.

[0049] 1.1: Download all miRNA data of Arabidopsis thaliana from the miRBase database to obtain its pre-miRNA data; then, use the Ensemble Plants tool to expand the data, extend all pre-miRNAs forward by 1000 bases, and process the data using the softberry software to obtain the first sORF sequence of the required pri-miRNA. Use the CD-HIT tool to remove sequences with a similarity higher than 80% to obtain a total of 245 candidate positive set sORF data.

[0050] 1.2: Download the NCDS data of Arabidopsis thaliana, use the ORF Finder tool to obtain the sORFs sequence, and use the CD-HIT tool to remove the sequences with similarity higher than 80% to obtain candidate negative set data, a total of 1670. Because the gap between the positive and negative set data is too large, 245 of the 1670 are randomly selected as candidate negative set data.

[0051] 1.3: Randomly select 10% of the positive and negative set data, that is, 24 items from the positive set and 25 items from the negative set, as the model test set.

[0052] Step 2: Extract relevant features of sORF sequence data.

[0053] Studies have shown that the richer the extracted features, the better the performance of the model. First, in order to fully consider the dependent relationship between adjacent nucleotides, the k-mer feature is introduced. This feature describes the composition information of the sORF sequence. It preserves the relationship between adjacent nucleotides and avoids information loss. k represents the number of adjacent bases in the sequence. k adjacent bases will have 4 k In the present invention, k is 1, 2, 3, and 4 respectively, and a total of 340 dimensions of features are obtained. A sliding window with a step size of 1 is used for sliding matching.

[0054] The k-mer feature takes into account the properties of continuous bases, but there are also differences between discontinuous bases. In order to prevent the loss of this part of information, the SSM feature is used to supplement it. Different base pair compositions can obtain different information, which can be recorded as A*B, A**B, A***B, etc., where A and B are recorded as any base among A, T, C, and G, and * represents any base. Each different composition is expressed by 16-dimensional features. The SSM features of three groups of A*B, A**B, and A***B are extracted, with a total of 48 dimensions.

[0055] In addition to considering the different information of base pairs, we also extract features such as sequence length L, GC base content, GC base ratio, and signal-to-noise ratio based on the expression of the sequence itself. The above constitutes a 4-dimensional feature.

[0056] The above features are combined to form a 392-dimensional feature set.

[0057] Step 3: Design the IMMRMD algorithm for feature sorting and selection.

[0058] The MRMD algorithm is often used in biological data dimensionality reduction. The MRMD feature selection method is mainly determined by two parts: the first is the correlation between the feature and the instance label. MRMD uses the Pearson correlation coefficient to calculate the correlation between the feature and the label; the second is the redundancy between features. Three distance functions (Euclidean distance, Cosine distance and Tanimoto coefficient) are used to calculate the redundancy between features. The larger the Pearson correlation coefficient, the closer the relationship between the feature and the label, and the larger the distance, the lower the redundancy between the features. Finally, the MRMD algorithm selects a feature subset that has a strong correlation with the label and low redundancy between features.

[0059] However, the data processed by the Pearson correlation coefficient must conform to the normal distribution and have no abnormal points, which is inconsistent with the experimental data. Therefore, the present invention selects the Spearman correlation coefficient, which is more commonly used in nucleotide sequence research, for calculation. The Spearman coefficient has lower requirements for the amount of data, and the running data is not required to present a normal distribution, which is more in line with the actual data situation. At the same time, the Cosine distance when calculating the distance only considers the difference in direction of the data. Although it can make certain corrections to the bias between individuals, it can only distinguish the difference between individuals in the dimensional direction and cannot measure the difference in the numerical value of each dimension. Therefore, it will lead to deviations in the results due to poor sensitivity to absolute numerical differences. Therefore, in order to correct the irrationality of only considering the similarity in the dimensional direction of the vector without considering the difference in the dimensions of each dimension, the present invention uses the adjusted cosine similarity for calculation, that is, the numerical values ​​in all dimensions are corrected by subtracting the mean, so that the calculation results are more in line with reality.

[0060] Specifically, the original feature data extracted in step 2 is used as the input of the IMMRMD feature selection algorithm. On the one hand, the Spearman correlation coefficient is calculated to obtain the correlation between the sub-features and the classification categories. The correlation between each dimension of feature data and the remaining features is represented by a number between (-1, 1); on the other hand, the redundancy degree of each feature vector and the remaining feature vectors is obtained by calculating the Euclidean distance, adjusting the cosine similarity and the Tanimoto coefficient, which is also represented by numbers; and the calculation results of the two aspects are accumulated and sorted, and input into the logistic regression model dimension by dimension to evaluate the classification effect. The top m dimensional features with the best effect are selected using evaluation criteria such as the F1 value, thereby obtaining an m-dimensional sub-feature set with both maximum correlation and minimum redundancy.

[0061] Step 4: Use SVM, LR, DF and other models for comparative experiments, and select the DF model with the best effect as the prediction model.

[0062] DF is an extended algorithm of random forest, which integrates the idea of ​​deep learning and is a decision tree ensemble method. DF has the advantages of low hyperparameters, high efficiency, good learning performance even with low data volume, and is applicable to data from various fields.

[0063] DF consists of the following two parts: cascade forest and multi-granularity scanning. Multi-granularity scanning is inspired by convolutional neural networks and generates input feature vectors for cascade forests. Sliding windows are used to scan original features, and different sampling windows can enhance the feature expression ability of input data. For example, for 400-dimensional input feature data, if a sliding window of size 100 is used to process the input features, 301 feature vectors will be obtained in the end when the step size is 1. The class vectors generated by window scanning will be input into random forest and completely random forest to generate transformed feature vectors. Assuming it is a binary classification problem, each forest will generate 301 2-dimensional class vectors. Finally, the class vectors generated by the two forests are concatenated to output a 1204-dimensional transformed feature vector corresponding to the original 400-dimensional feature vector. In order to improve the diversity of model features, multi-granularity scanning usually uses multiple windows of different sizes to perform multi-scale sampling on sample data, thereby generating feature vectors of different scales and obtaining more sample information. The DF model specifies the default multi-granularity scanning structure parameters. For raw data with d-dimensional features, feature windows of sizes d / 16, d / 8, and d / 4 will be used for scanning, maximizing the computational efficiency while ensuring the diversity of the generated samples.

[0064] Each layer of the cascade forest structure stacks several random forests in a cascade manner, which makes DF have representation ability and enhances the model's feature representation ability and generalization ability. Each random forest contains several decision trees, each of which outputs a result in the form of a class vector, and then the output class vector results of each decision tree are averaged. The first layer of the cascade forest takes the result of multi-granularity scanning as input, and the input of each layer of the cascade is composed of the enhanced features after the cascade and the results of multi-granularity scanning. Considering that deep structures usually have overfitting problems, the results are verified on the test set after each layer. If the effect is improved, the enhanced vector is passed to the next layer of cascade to generate a new concatenation vector and repeat the above process; if there is no improvement, the training is terminated. Therefore, the number of layers of the cascade forest can be automatically determined, and DF also uses this structure to automatically determine the complexity of the deep structure to achieve adaptability to data sets of different sizes. It is this feature that makes it not only suitable for large-scale training data, but also can obtain good results on small-scale training data sets compared with other deep neural networks.

[0065] The experimental data sets are experimented on SVM, LR and DF models respectively. The present invention uses accuracy (ACC), precision (P), recall (R) and F1 value (F1_score, F1) as evaluation criteria. The calculation formulas of the four evaluation criteria are as follows:

[0066]

[0067]

[0068]

[0069]

[0070] The meanings of TP, FP, TN, and FN are shown in Table 1.

[0071] Table 1 Classification meaning table

[0072]

[0073] The results are attached Figure 2 As shown in Appendix 2, the classification effect of the test set on DF is the best. After the feature data set is processed by the feature selection algorithm MRMD3.0 and IMMRMD, the classification effect after DF is balanced in the prediction ability for positive and negative sets, and the ACC, P value and F1 value are improved, with better prediction effect.

[0074] Table 2 Classification performance of Arabidopsis dataset using different feature processing methods under DF

[0075]

[0076] At the same time, in order to better illustrate the role of IMMRMD in improving the classification effect of this experiment under different conditions, the present invention also conducted the following experiments: Figure 2 It shows that when the original feature data is used for classification prediction through DF, SVM and LR respectively, DF shows a better classification effect, and performs better than other models in ACC, P value, R value and F1 value. Figure 3 , Figure 4 It represents the classification effect on SVM and LR after MRMD3.0, IMMRMD and without feature selection algorithm, respectively. Figure 3 , Figure 4It can be seen that IMMRMD can make the data after feature selection more balanced in the prediction of positive and negative set labels, which means that it is more suitable for the sequence data, and has improved in both ACC and F1 values. Table 2 uses DF as the classification model and compares the classification after processing the original feature data by different methods. In Table 2, MRMD3.0-SC means that only the Pearson coefficient in MRMD3.0 is changed to the Spearman coefficient without changing the cosine similarity, and AutoEncoder means that the data is processed by the autoencoder; As can be seen from Table 2, IMMRMD has significant advantages in ACC, P value and F1 value, and the classification effect of positive and negative sets is more balanced than others, which means that it has better classification effect than other methods.

[0077] The core of the present invention mainly has two aspects. First, the MRMD feature selection algorithm is improved, its evaluation criteria are optimized, and the Spearman coefficient is used instead of the Pearson coefficient to make it more suitable for the data of the present invention itself. At the same time, the adjusted cosine similarity that comprehensively considers the influence of numerical value and direction is used to replace the previous cosine phase velocity, which considers the data more comprehensively. On the other hand, it combines the idea of ​​deep learning, and adopts a more suitable DF as the classification model after comparison to improve the classification effect.

Claims

1. A method for predicting plant pri-miRNA encoded peptides based on an improved MRMD algorithm and a DF model, characterized in that: The method comprises the following steps: Step 1: Build an experimental dataset The positive set used the first sORF obtained from the Arabidopsis pri-miRNA sequence, and the negative set used Arabidopsis non-coding sequence data; Step 2: Extract relevant features of sORF sequence data k-mer features, short sequence motif SSM features, signal-to-noise ratio, GC base content, GC base ratio and sequence length features were extracted from both positive and negative data sets; Step 3: Design IMMRMD algorithm for feature sorting and selection The IMMRMD algorithm uses the Spearman coefficient as the judgment standard and selects the adjusted cosine similarity as the judgment standard of the maximum distance; the IMMRMD algorithm is used to sort and select the feature data obtained in step 2; Step 4: Use the deep forest DF model to get the prediction results Input the data filtered in step 3 into the DF model to obtain the final prediction result.

2. The method according to claim 1, characterized in that The specific process of step 1 is as follows: 1.1: Download all miRNA data of Arabidopsis from the miRBase database to obtain its pre-miRNA data; then, use the Ensemble Plants tool to expand the data, expand all the obtained pre-miRNAs forward by 1000 bases, and use the softberry software to find their transcription start sites to obtain the corresponding pri-miRNAs; then use the ORFFinder tool to obtain the first sORF sequence of each pri-miRNA, and use the CD-HIT tool to remove sequences with a similarity higher than 80% to obtain candidate positive set sORF data; 1.2: Collect Arabidopsis NCDS data, use ORF Finder tool to obtain sORFs sequences, and remove sORF sequences with similarity higher than 80% to obtain candidate negative set data; 1.3: Randomly select 10% of the positive and negative set data as the test set to test the final prediction performance.

3. The method according to claim 1 or 2, characterized in that: The step 2 is specifically as follows: The k-mer feature describes the composition information of the sORF sequence. It fully considers the dependency between adjacent nucleotides and preserves the information of adjacent nucleotides, thus avoiding the loss of information. k represents the number of adjacent bases in the sequence. There are 4 k combinations; k is 1, 2, 3, 4, a total of 340 dimensions; a sliding window with a step size of 1 is used for sliding matching, and the formula is as follows: l k =L-k+1,k=1,2,3,4 (2) Among them, l k is the number of slides, s i is the number of occurrences of each k-mer, f i is the final frequency of each k-mer, and L is the sequence length; The k-mer feature takes into account the properties of continuous bases, but there are also differences between discontinuous bases, so the SSM feature is used to supplement it; different base pairs can be recorded as A*B, A**B, A***B, where A and B are any base among A, T, C, and G, and * represents any base. Each group of SSM has a 16-dimensional feature expression. Three groups of SSM, A*B, A**B, and A***B, are extracted, with a total of 48 dimensions: u j Lj-1, j 1, 2, 3 (4) Among them, u j is the number of slides, v i′ is the number of occurrences of each short sequence motif, SSM i′ is the frequency of occurrence of each short sequence motif; According to the expression of the sequence itself, the sequence length L, GC base content GC_con, GC base ratio GC_ra and signal-to-noise ratio features are extracted, a total of 4 dimensions; the size of the signal-to-noise ratio can indicate the bias of base usage in the sequence, and the power spectrum of sORF at one-third is calculated through the distribution of bases in three phases to obtain the signal-to-noise ratio; the specific formula is as follows: Where g and c represent the number of bases G and C in the sequence, respectively; L represents the length of the sequence; a x , b x and c x They are all arrays with one row and three columns, representing the frequencies of base x appearing in three phases respectively; W represents the power spectrum; SNR represents the signal-to-noise ratio; Finally, the extracted features constitute 392-dimensional feature data.

4. The method according to claim 1 or 2, characterized in that: In step 3, the specific steps of the IMMRMD algorithm are as follows: the original feature data extracted in step 2 is used as the input of the IMMRMD algorithm, the Spearman correlation coefficient is calculated to obtain the correlation between the sub-feature set and the target category, and the sub-feature set with the highest correlation with the target category is selected; the Spearman correlation coefficient calculation formula is shown in formula (9), and X and Y are defined as two groups of data, Among them, d y For X y and Y y The level difference between s It is between -1 and 1, and n is the total number of samples; On the other hand, the Euclidean distance, adjusted cosine similarity and Tanimoto coefficient are calculated to obtain the degree of redundancy between any eigenvector and the rest of the eigenvectors; Finally, by accumulating and sorting the calculation results of the two aspects, the sorted feature data is input into the logistic regression model dimension by dimension for evaluation, and the top m dimensions with the best classification effect are selected, thus obtaining an m-dimensional sub-feature set with maximum relevance and minimum redundancy. Finally, 382 dimensions were selected by the IMMRMD algorithm.

5. The method according to claim 3, characterized in that: In step 3, the specific steps of the IMMRMD algorithm are as follows: the original feature data extracted in step 2 is used as the input of the IMMRMD algorithm, the Spearman correlation coefficient is calculated to obtain the correlation between the sub-feature set and the target category, and the sub-feature set with the highest correlation with the target category is selected; the Spearman correlation coefficient calculation formula is shown in formula (9), and X and Y are defined as two groups of data, Among them, d y For X y and Y y The level difference between s It is between -1 and 1, and n is the total number of samples; On the other hand, the Euclidean distance, adjusted cosine similarity and Tanimoto coefficient are calculated to obtain the degree of redundancy between any eigenvector and the rest of the eigenvectors; Finally, by accumulating and sorting the calculation results of the two aspects, the sorted feature data is input into the logistic regression model dimension by dimension for evaluation, and the top m dimensions with the best classification effect are selected, thus obtaining an m-dimensional sub-feature set with maximum relevance and minimum redundancy. Finally, 382 dimensions were selected by the IMMRMD algorithm.

Images