A method for predicting cross-species polypeptide encoding sORFs

Abstract

This invention, belonging to the field of bioinformatics, discloses a cross-species prediction method for peptide-encoding sequence-dependent open-type recurrent (sORF) proteins. The method integrates protein-encoding DNA sequences from human and mouse sORF databases, Arabidopsis thaliana from the TAIR database, and some prokaryotes from the NCBI database. Positive and negative samples corresponding to each species are constructed using data filtering and negative sample generation strategies. Sequence features are extracted and the training set is screened using maximum relevance minimum redundancy and incremental selection methods to obtain the optimal feature sets corresponding to different methods and construct a peptide-encoding sORF prediction model. Parameters are optimized using a greedy grid search method to construct the optimal prediction model based on a support vector machine. The prediction model is evaluated using sensitivity, specificity, accuracy, and the Matthews correlation coefficient. This method facilitates the classification and identification of peptide-encoding sORFs and is of great significance for the study of peptide-encoding sORFs and gene annotation.

Description

Technical Field

[0001] The present invention specifically relates to a method for predicting cross-species polypeptide encoding sORF, and belongs to the field of bioinformatics. Background Art

[0002] Small open reading frames (sORFs) are DNA sequences less than 100 amino acids in length and are found in the genomes of many organisms. For over a decade, researchers have often overlooked sORFs in genome annotations due to their short sequence length, low expression levels, and limited data. With the advancement of sequencing technology, it has been discovered that many sORFs can also encode proteins and are ubiquitous across the genome. In 2016, Professor Eric from the University of Texas and his colleagues discovered that DWORF, a small protein encoded by sORFs, plays a significant role in myocardial contractility, prompting researchers to rethink and rethink sORFs.

[0003] In recent years, small proteins encoded by small open-type recurrent (sORF) proteins (sORFs) have been detected in bacteria, yeast, and humans. These proteins play important roles in life processes such as embryonic development, muscle function, and apoptosis. Therefore, peptide-encoding sORFs have gradually become a research hotspot in biology. However, traditional sequence analysis methods, such as genome sequencing, transcriptome sequencing, and proteome sequencing (mass spectrometry), have shown limited success in identifying sORFs. In recent years, ribosome profiling has emerged as a new technology, succeeding traditional sequencing methods such as mass spectrometry, and has been widely used to classify and identify sequences as protein-encoding. This approach primarily involves analyzing protein synthesis through ribosome imprinting. However, evidence suggests that many noncoding RNAs (ncRNAs) also bind to ribosomes, making traditional sequencing methods alone insufficient for classifying and identifying peptide-encoding sORFs. Currently, numerous tools exist for distinguishing between coding and noncoding RNAs, including CPAT, CNCI, PLEK, CPC2, CPPred, LGC, MePiped, DeepCPP, and CPPred-sORF. These methods are all based on certain data sets and can effectively distinguish between "normal" length coding RNA and ncRNA, but their accuracy in classifying and identifying peptide-encoding sORFs is not high. Therefore, the development of effective sORF classification and identification technology is of great significance. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the existing background technology and provide a method for predicting cross-species polypeptide encoding sORFs, solve the problem of sample size imbalance, and facilitate the extraction of robust and efficient DNA sequence features.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] A method for predicting cross-species polypeptide encoding sORF (small open reading frame), comprising the following steps:

[0007] The non-coding sequence generation strategy was applied to the peptide-coding sORFs datasets of multiple species to obtain the corresponding non-coding sORFs datasets.

[0008] The peptide-coding sORFs and non-coding sORFs of each species were removed for redundancy (Max-Relevance and Min-Redundancy, MRMR) to obtain the corresponding positive and negative samples of each species and construct the training set and test set;

[0009] Extract the corresponding feature parameters from each data set; combine the MRMR strategy and incremental selection method to select the features with better performance and construct the corresponding feature set;

[0010] A prediction model for peptide-encoding sORFs based on a support vector machine (SVM) was constructed, and the feature set of the training set was used for model training. The training model was further optimized using a greedy grid search method to obtain the best prediction model within the specified parameter range.

[0011] The prediction model is used to predict the test set, the prediction results of each data set are analyzed, the prediction efficiency of different feature selection strategies is compared and evaluated, and the best performing feature set and prediction model are obtained as the optimal feature set and prediction model.

[0012] Furthermore, the coding sequences (CDS) of humans and mice were downloaded from the sORF database, the CDS of Arabidopsis thaliana was downloaded from the TAIR database, and the CDS of some prokaryotic genomes were downloaded from the NCBI database; "erroneous" sequences were filtered out according to the data filtering strategy to obtain a data set of peptide-encoding sORFs from multiple species.

[0013] Further, "error" sequence filtering strategy:

[0014] sORFs with a length ≥100 aa were filtered out;

[0015] sORFs whose sequence length is not divisible by 3 were filtered out;

[0016] sORFs starting with a stop codon were filtered out;

[0017] sORFs that do not end with a stop codon were filtered out;

[0018] sORFs with stop codons in the sequence were filtered out;

[0019] Furthermore, the non-coding sequence generation strategy is:

[0020] The start and stop codons were fixed, and each positive sORF sequence was randomly shuffled;

[0021] Make sure there is no stop codon before the stop codon at the end of the sequence;

[0022] Furthermore, the redundancy removal method is:

[0023] The CDHit program was used to remove redundancy from peptide-coding sORFs and non-coding sORFs of each species, and the corresponding positive and negative samples of each species were obtained to construct training and test sets.

[0024] The de-redundancy threshold was set to 0.80, and DNA sequences with similarity greater than 80% were filtered out;

[0025] Furthermore, the corresponding feature parameters of each data set were extracted according to 9 different feature selection strategies; the 9 different feature selection strategies were: CPPred, 2mer, 3mer, TN, ITN, Cylindrical, Spherical, Codon, and Amino.

[0026] Furthermore, PyCharm software was used to extract the characteristic parameters of the DNA sequences in each data set; the pymrmr package of PyCharm software was used to implement the MRMR strategy and incremental selection method, complete feature sorting, and construct the corresponding feature set.

[0027] Furthermore, the parameters of the grid search method are set as follows:

[0028] cmin: The minimum value of the penalty parameter c; the default value is -5;

[0029] cmax: The maximum value of the variation range of the penalty parameter c; the default value is 5;

[0030] gmin: The minimum value of the range of parameter g; the default value is -5;

[0031] gmax: The maximum value of the range of parameter g; the default value is 5;

[0032] v: cross validation parameter; the default is 3;

[0033] cstep: parameter c step size; the default is 1;

[0034] gstep: the size of the parameter g step; the default is 1;

[0035] accstep: The step size when displaying the final accuracy graph; the default is 1.5.

[0036] Furthermore, the prediction model was used to perform cross-species prediction on the test set using the libsvm package of Matlab. The prediction results were analyzed based on four indicators, namely Sn, Sp, ACC, and MCC. The prediction models were compared and evaluated, and the best performing feature set and prediction model were used as the feature set and prediction model for the prediction method of cross-species coding polypeptide sORFs.

[0037] Furthermore, the evaluation index calculation formula is as follows:

[0038]

[0039]

[0040]

[0041]

[0042] Among them, Sn is sensitivity, Sp is specificity, ACC is accuracy, MCC is Matthews correlation coefficient, TP is true positive, FN is false negative, TN is true negative, and FP is false positive.

[0043] Beneficial effects

[0044] The present invention proposes a cross-species prediction method for peptide-coding sORFs. Based on high-throughput sequencing data, it integrates and screens peptide-coding sORFs from human, mouse, Arabidopsis, and some prokaryotic genomes. Combined with a strict non-coding sORF generation strategy, it addresses the problem of sample size imbalance and facilitates the extraction of robust and efficient DNA sequence features.

[0045] The present invention is based on support vector machines (SVMs), utilizes the maximum relevance minimum redundancy (MRMR) strategy and the incremental selection method to screen DNA sequence feature parameters and construct an optimal feature set for prediction, which helps to classify and identify peptide-encoding sORFs and develop a cross-species prediction method for peptide-encoding sORFs. This is of great significance for the research on peptide-encoding sORFs and gene annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A data screening strategy for a method for predicting cross-species polypeptide encoding sORFs according to an embodiment of the present invention;

[0047] Figure 2 This is a flow chart of a method for predicting cross-species polypeptide encoding sORFs according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] like Figure 1 As shown in FIG, a data screening strategy for a method for predicting cross-species polypeptide encoding sORFs according to an embodiment of the present invention includes the following steps:

[0050] Step 1) Download the CDS of human and mouse from the sORF database, download the CDS of Arabidopsis thaliana from the TAIR database, and download the CDS of some prokaryotic genomes from the NCBI database; filter out "erroneous" sequences according to the data filtering strategy to obtain a dataset of peptide-encoding sORFs of each species;

[0051] This paper constructed two training sets: a prokaryotic mixed training set (TR1) and a eukaryotic mixed training set (TR2); and eight test sets: Arabidopsis thaliana test set 1 (Ara1), Arabidopsis thaliana test set 2 (Ara2), human test set 1 (Hum1), human test set 2 (Hum2), mouse test set 1 (Mou1), mouse test set 2 (Mou2), prokaryotic test set (Pro), and Escherichia coli test set (Bac). TR2 was composed of 10% of sequences randomly sampled from each of the three datasets: Ara1, Hum1, and Mou1. The remaining 90% of sequences from Ara1, Hum1, and Mou1 constituted Ara2, Hum2, and Mou2, respectively.

[0052] To construct TR1, CDS sequences of NC_009089, NC_003103, NC_012962, NC_000913, and NC_008380 were downloaded from NCBI, and sORFs with a length of ≤100 aa and clear functions were retained. To construct Ara1, Hum1, and Mou1, 2888 Arabidopsis thaliana sORFs were downloaded from the TAIR database, and 10,000 human sORFs and 10,000 mouse sORFs were downloaded from the sORF database, respectively. To construct the Pro_6318-6318 dataset, sORFs with clear functions were screened from 56 prokaryotic genomes (Table 1). The genomic GC content of the 56 selected prokaryotic genomes ranged from 20% to 70%. The Bac_150-53 dataset is an experimentally validated dataset published by Hemm et al.

[0053] Table 1 56 prokaryotic genomes

[0054]

[0055]

[0056] According to the data filtering strategy, the “erroneous” sequences in the downloaded data were filtered out to obtain a dataset of peptide-encoding sORFs of each species;

[0057] "Error" sequence filtering strategy:

[0058] (i) sORFs ≥100 aa were filtered out;

[0059] (ii) filtering out sORFs whose sequence length is not divisible by 3;

[0060] (iii) filtering out sORFs starting with a stop codon;

[0061] (iv) filtering out sORFs that do not end with a stop codon;

[0062] (v) filtering out sORFs with stop codons in the sequence;

[0063] Step 2) applying the non-coding sequence generation strategy to the datasets of peptide-coding sORFs of each species in step 1 to obtain the corresponding non-coding sORFs datasets;

[0064] Non-coding sequence generation strategy:

[0065] (i) The start and stop codons were fixed and each positive sORF sequence was randomly shuffled;

[0066] (ii) ensuring that there is no stop codon preceding the stop codon at the end of the sequence;

[0067] Step 3) Using the CDHit program, the peptide-coding sORFs and non-coding sORFs of each species were de-redundant, and the corresponding positive and negative samples of each species were obtained (Table 2) to construct the training set and test set;

[0068] The de-redundancy threshold was set to 0.80, and DNA sequences with similarity greater than 80% were filtered out;

[0069] Table 2 Dataset statistics

[0070]

[0071] Figure 2 Flowchart of a method for predicting cross-species polypeptide encoding sORF according to an embodiment of the present invention, comprising the following steps:

[0072] Step 4) Using PyCharm software, we extracted characteristic parameters of the DNA sequences in each dataset based on nine feature selection strategies (CPPred, 2mer, 3mer, TN, ITN, Cylindrical, Spherical, Codon, and Amino). We also used the pymrmr package in PyCharm to implement the MRMR strategy and incremental selection method, complete feature sorting, and construct the corresponding feature set.

[0073] (i) 9 feature selection strategies:

[0074] CPPred: CPPred-sORF is an improvement to CPPred that can be used to predict the coding potential of sORFs. In this work, we extracted all the features involved in the CPPred-sORF prediction tool, including sequence length (Length), coverage (Coverage), integrity (Intergrity), Fickett score, Hexamer score, isoelectric point (pI) of predicted peptides, average hydrophilicity of predicted peptides (Gravy), instability of predicted peptides (Instability), global descriptor (CTD), GC content, and the content of 11 attention codons (UAC, AAC, UAU, AUC, UUC, GAG, AAG, GAU, GAC, AAU, and GUG), a total of 40 features. Among them, CTD contains a total of 30 features: C represents nucleotide composition, with 4 features, describing the proportion of each nucleotide in the transcript sequence, which can be represented by A, T, C, and G; T represents nucleotide transition, which contains 6 features, describing the percentage frequency of the transition of four nucleotides between adjacent positions, which can be represented by AT, AG, AC, TG, TC, and GC; D represents nucleotide distribution, which contains 20 features, describing the proportion of each nucleotide before 5 relative positions, which are 0 (first), 25%, 50%, 75%, and 100% (last).

[0075] 2mer, 3mer: K-mer is a subsequence of length k, where k is an integer. For any DNA sequence, specify a reading frame length of K, a step size of 1, and move from the first base to the end of the sequence. The subsequence of length K intercepted within the reading frame is the K-mer feature of the DNA sequence. Generally, any DNA sequence contains four bases: A, T, C, and G. Therefore, when different K values ​​are taken, the number of K-mers in the DNA sequence is different, which is 4. K Taking the sequence length into consideration, we selected K = 2 and K = 3, and analyzed the proportion of each DNA subsequence under different K values, and finally obtained 16 and 64 features respectively.

[0076] TN: The TN curve is a three-dimensional graphical representation of DNA sequences based on trinucleotides. In this work, we extract six parameters of the TN curve for each DNA sequence as sequence features.

[0077] ITN: The I-TN curve is an improved graphical representation based on the TN curve. In this work, we extract 18 parameters of the TN curve of each DNA sequence as sequence features.

[0078] Cylindrical: Cylindrical is a cylindrical representation of protein sequences. In this work, we convert DNA sequences into protein sequences and use the parameters of the cylindrical representation of protein sequences as sequence features.

[0079] Spherical: Spherical is a spherical coordinate representation of protein sequences. In this work, we converted DNA sequences into protein sequences and used the physicochemical properties of each amino acid in the cylindrical spherical representation of protein sequences as sequence features.

[0080] Codon: The percentage of 64 codons.

[0081] Amino: The percentage of 20 amino acids.

[0082] (ii) MRMR and incremental selection:

[0083] The MRMR (maximum relevance minimum redundancy) method is used to rank the correlation between the features and coding potential in each feature extraction strategy. That is, the higher the ranking, the greater the correlation between the features and the sequence coding potential, and the lower the redundancy between the features.

[0084] The sorted feature set is subjected to an incremental selection method. That is, according to the sorting results, the first N features are selected in sequence as a new feature subset (N = 1, 2, 3...). Each feature subset is subjected to 10-fold cross-validation, and the average prediction accuracy is analyzed. The feature subset with the highest accuracy is selected as the best feature set under this feature extraction strategy for model training.

[0085] Step 5) Using the libsvm package in Matlab, a prediction model for peptide-encoding sORFs based on Support Vector Machines (SVM) was constructed; a prediction model based on prokaryotic training was constructed using the nine feature sets of training set TR1, and a prediction model based on eukaryotic training was constructed using the nine feature sets of training set TR2; the prediction model was further optimized using a greedy grid search method to obtain the best prediction model within the 18 specified parameter ranges;

[0086] Parameter settings for the grid search method:

[0087] cmin: Minimum value of the range of variation of the penalty parameter c (after taking the base 2 logarithm). Default is -5.

[0088] cmax: The maximum value of the range of variation of the penalty parameter c (after taking the base 2 logarithm). The default value is 5.

[0089] gmin: The minimum value of the range of variation of parameter g (after taking the base 2 logarithm). The default value is -5.

[0090] gmax: The maximum value of the range of g (after taking the base 2 logarithm). The default value is 5.

[0091] v: Cross-validation parameter. Defaults to 3.

[0092] cstep: The size of the parameter c step. The default is 1.

[0093] gstep: The size of the g step. The default value is 1.

[0094] accstep: The step size when displaying the final accuracy graph. The default value is 1.5.

[0095] Step 6) Use the prediction model to predict the test set, analyze the Sn, Sp, ACC, and MCC of the prediction results of each data set, compare and evaluate the prediction efficiency of different feature selection strategies, and obtain the best performing feature set and prediction model as the optimal feature set and prediction model.

[0096] According to the 18 prediction models obtained in step 5), the libsvm package of Matlab was used to perform cross-species prediction on the test set, where the test sets Ara1, Hum1, Mou1, Pro and Bac were predicted on the prediction model trained with the prokaryotic mixed training set, and the test sets Ara2, Hum2, Mou2, Pro and Bac were predicted on the prediction model trained with the true mixed training set.

[0097] According to the four indicators of Sn, Sp, ACC, and MCC, the prediction results are evaluated (Table 3, Table 4). The calculation formula of the evaluation indicators is as follows:

[0098]

[0099]

[0100]

[0101]

[0102] Among them, Sn is sensitivity, Sp is specificity, ACC is accuracy, MCC is Matthews correlation coefficient, TP is true positive, FN is false negative, TN is true negative, FP is false positive. Table 3 Prediction results of the prediction model constructed based on the original kernel mixed training set

[0103]

[0104]

[0105] Table 4 Prediction results of the prediction model constructed based on the prokaryotic mixed training set

[0106]

[0107] Based on the prediction results, prediction models for prokaryotic peptide-encoding sORFs and eukaryotic peptide-encoding sORFs were constructed based on separate prokaryotic and eukaryotic training sets, respectively, to enable cross-species prediction of sORFs within the same kingdom. For prokaryotic sORFs, the model trained with the optimal feature set extracted using the Codon strategy performed best, with a prediction accuracy of up to 0.91. For eukaryotic sORFs, the model trained with the optimal feature set extracted using the 3mer strategy performed best, with prediction accuracies ranging from 0.83 to 0.87.

[0108] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting cross-species polypeptide encoding sORF, characterized in that: The following steps are involved: The non-coding sequence generation strategy was applied to the datasets of peptide-encoding sORFs from multiple species to obtain the corresponding non-coding sORFs datasets. The non-coding sequence generation strategy was as follows: the start and stop codons were fixed, and each positive sORF sequence was randomly shuffled; ensuring that there was no stop codon before the stop codon at the end of the sequence. The peptide-coding sORFs and non-coding sORFs of each species were de-redundant, and the corresponding positive and negative samples of each species were obtained to construct the training and test sets. The de-redundancy method was as follows: the peptide-coding sORFs and non-coding sORFs of each species were de-redundant, and the corresponding positive and negative samples of each species were obtained to construct the training and test sets. The de-redundancy threshold was set at 0.80, and DNA sequences with a similarity greater than 80% were filtered out. According to 9 different feature selection strategies, the corresponding feature parameters in each data set are extracted; the 9 different feature selection strategies are: CPPred, 2mer, 3mer, TN, ITN, Cylindrical, Spherical, Codon, and Amino; the maximum correlation minimum redundancy strategy and the incremental selection method are combined to select features with better performance and construct the corresponding feature set; a prediction model based on support vector machine peptide-encoded sORFs is constructed, and the feature set of the training set is used for model training; the training model is further optimized using the greedy grid search method to obtain the best prediction model within the specified parameter range; the prediction model is used to predict the test set, the prediction results of each data set are analyzed, the prediction efficiency of different feature selection strategies is compared and evaluated, and the best-performing feature set and prediction model are obtained as the optimal feature set and prediction model.

2. The method for predicting cross-species polypeptide encoding sORF according to claim 1, characterized in that: Human and mouse coding sequences were downloaded from the sORF database, Arabidopsis CDS was downloaded from the TAIR database, and CDS of some prokaryotic genomes were downloaded from the NCBI database. "Error" sequences were filtered out according to the data filtering strategy to obtain a dataset of peptide-encoding sORFs from multiple species.

3. The method for predicting cross-species polypeptide encoding sORF according to claim 2, characterized in that: The "error" sequence filtering strategy is: sORFs with a length of ≥100 aa were filtered out; sORFs whose sequence length is not divisible by 3 were filtered out; sORFs starting with a stop codon were filtered out; sORFs that do not end with a stop codon were filtered out; sORFs with stop codons in the sequences were filtered out.

4. The method for predicting cross-species polypeptide encoding sORF according to claim 1, characterized in that: PyCharm software was used to extract the characteristic parameters of DNA sequences in each data set; the pymrmr package of PyCharm software was used to implement the MRMR strategy and incremental selection method, complete feature sorting, and construct the corresponding feature set.

5. The method for predicting cross-species polypeptide encoding sORF according to claim 1, characterized in that: Parameter settings for the grid search method: cmin: The minimum value of the penalty parameter c; the default value is -5; cmax: The maximum value of the variation range of the penalty parameter c; the default value is 5; gmin: The minimum value of the range of parameter g; the default value is -5; gmax: The maximum value of the range of parameter g; the default value is 5; v: cross validation parameter; the default is 3; cstep: parameter c step size; the default is 1; gstep: the size of the parameter g step; the default is 1; accstep: The step size when displaying the final accuracy graph; the default is 1.

5.

6. The method for predicting cross-species polypeptide encoding sORF according to claim 1, characterized in that: Using the libsvm package of Matlab, the prediction model was used to perform cross-species prediction on the test set. The prediction results were analyzed according to four indicators: Sn, Sp, ACC, and MCC. The prediction models were compared and evaluated, and the best performing feature set and prediction model were used as the feature set and prediction model for the prediction method of cross-species coding polypeptide sORF.

7. The method for predicting cross-species polypeptide encoding sORF according to claim 6, wherein the evaluation index calculation formula is as follows: in, Sn is sensitivity, Sp is specificity, ACC is accuracy, MCC is Matthews correlation coefficient, TP is true positive, FN is false negative, TN is true negative, and FP is false positive.

Images