Long non-coding RNA recognition method based on catboost algorithm

Abstract

The present application relates to the technical field of integrated learning, and more particularly to a long non-coding RNA recognition method based on a CatBoost algorithm, comprising: obtaining codon-related features, guanine-cytosine-related features, transcript-related features, sequence structure-related features, and open reading frame-related features of a plurality of RNA sequences from a database, and establishing an RNA sequence feature vector; performing redundant feature filtering on the RNA sequence feature vector through feature selection; fusing the filtered RNA sequence feature vector; inputting the fused RNA sequence feature vector into a CatBoost model for decision classification, and determining whether it is a long non-coding RNA or a protein-coding RNA according to a probability score. The present application solves the deficiencies of the long non-coding RNA recognition method in the prior art, and realizes a high-accuracy, high-precision, high-Matthews correlation coefficient, and efficient long non-coding RNA prediction method.

Description

Technical Field

[0001] This invention relates to the field of ensemble learning technology, and in particular to a method for identifying long non-coding RNAs based on the CatBoost algorithm. Background Technology

[0002] Multiple studies have shown that over 80% of the human genome has biological function, but less than 2% of this genome can be translated into proteins; the remaining 98% is non-coding RNA. Long non-coding RNAs (LNAs) are an important component of LNAs, accounting for up to 70% of all non-coding sequences transcribed into LNAs. Due to their significant regulatory functions, LNAs have attracted considerable attention from researchers. Studies have shown that LNAs play important biological roles in epigenetic regulation, metabolic processes, chromosome dynamics, and cell differentiation. Furthermore, substantial evidence suggests that LNAs are highly associated with a variety of complex human diseases, such as lung cancer, Alzheimer's disease, and cardiovascular disease.

[0003] Identifying long noncoding RNAs (LNRs) and protein-coding RNAs is a fundamental step in elucidating the functions and mechanisms of LNRs. Currently, numerous methods and tools based on machine learning techniques have been developed to distinguish between LNRs and protein-coding RNAs. Machine learning research methods first process the RNA sequence for coding, then combine machine learning algorithms to predict whether it is a LNR or a protein-coding RNA. CPC (Coding Potential Calculator) compares the sequence with a protein database. CPAT (Coding-Potential Assessment Tool) calculates the Fickett TESTCODE score and hexamer score of the open reading frame (ORF) region to assess differences in nucleotide positions and codon selection between noncoding and protein-coding transcripts. CNCI (Coding-Noncoding Index) uses features based on adjacent nucleotide triplet matrices and codon asymmetry (codon bias). PLEK (predictor of long noncoding RNAs and messenger RNAs based on an improved k-mer scheme) uses an improved k-mer method to classify sequences. Compared to CPC, CPC2 has significantly improved accuracy and efficiency. In addition to classic features such as ORF information and Fickett TESTCOD scores, CPC2 uses isoelectric point (PI) to calculate coding potential, thereby predicting long non-coding RNAs. LncFinder, besides several classic features, incorporates intrinsic sequence features such as logistic distance and Euclidean distance. It also introduces secondary structure features and physicochemical properties based on electron-ion interaction pseudo-potential (EIIP) values. These methods employ different machine learning algorithms to construct classifiers for predicting lncRNAs. For example, CPC, CNCI, and PLEK use support vector machines (SVM), CPAT uses logistic regression (LR), and LncFinder determines which model to use by comparing various machine learning algorithms.

[0004] However, traditional machine learning models have some limitations. For example, before building a regression model, we should fully understand the properties of the data and the capabilities of the model to successfully apply it. Neural network algorithms are difficult to interpret, and the choice of model structure often relies on experience. If the chosen network structure is too large, it can lead to low training efficiency and overfitting. If it is too small, the network may struggle to converge. Ensemble learning, a branch of machine learning, accomplishes learning tasks by building and combining multiple learners, and has already demonstrated its potential in the fields of biology and medicine. Summary of the Invention

[0005] To address the shortcomings of existing algorithms, this invention overcomes the deficiencies of existing long non-coding RNA identification methods, achieving a high-accuracy, high-precision, high Matthews correlation coefficient, and efficient long non-coding RNA prediction method.

[0006] The technical solution adopted in this invention is as follows: The long non-coding RNA identification method based on the CatBoost algorithm includes the following steps:

[0007] S1. Obtain codon-related features, guanine-cytosine-related features, transcript-related features, sequence structure-related features, and open reading frame-related features from the database containing several RNA sequences;

[0008] Datasets for four species—human, wheat, chicken, and zebrafish—were obtained from the GENCODE and Ensembl databases. Each dataset contains several RNA sequences, and each RNA sequence contains an ID in the first row and a string of RNA sequences consisting of A (adenine), T (guanine), C (cytosine), and G (thymine).

[0009] Obtain codon-related features (CF), guanine-cytosine-related features (GC), transcript-related features (TF), sequence structure-related features (SF), and open reading frame-related features (ORF) of the RNA sequence.

[0010] Furthermore, codon-related features include: the number of stop codons, the frequency of stop codons, and the Fickett score;

[0011] In molecular biology, a codon is a triplet of nucleotides in mRNA that together constitutes a unit of genetic code in a DNA or RNA molecule. A stop codon is a codon that indicates the termination of the current protein translation process. There are several stop codon-related features, including the number of stop codons; the frequency of stop codons, which is equal to the number of stop codons divided by the length of the transcript; and the stop codon frequency frame score, which is the variance of the stop codon frequencies in the three open reading frames. A commonly used codon-related feature is called the Fickett testcode score, also known as the Fickett score. For a given transcript, the Fickett score is derived from the weighted nucleotide frequencies of the entire transcript. The Fickett score measures coding potential by assessing the degree of asymmetry in the distribution of the three triplet nucleotides in the sequence, based on compositional bias between codon positions.

[0012] Further, guanine-cytosine related features include: GC frame score and GC content;

[0013] GC content is the percentage of nitrogenous bases in RNA or DNA molecules that are guanine or cytosine. The GC content in coding regions is often higher than that in non-coding sequences. GC, GC1, GC2, and GC3 are calculated from the GC content of the entire transcript and the GC content of the 1st, 2nd, and 3rd positions of the codons, respectively. GC frame score refers to the variance of the number of GCs in the three reading frames. GC1 frame score, GC2 frame score, and GC3 frame score can be obtained in the same way.

[0014] Further transcript-related features include: k-mer, CTD, and Hexamer;

[0015] k-mer refers to a specific subsequence containing k nucleotides, which is one of the most commonly used features in long noncoding RNA recognition tasks; CTD (Composition, Transition, and Distribution) refers to a global transcript sequence descriptor based on nucleotide composition, nucleotide transition, and nucleotide distribution; Hexamer features are a variant of k-mer features, and Hexamer-based features measure the hexamer usage bias between coding and noncoding sequences; Hexamer-based features include Hexamer score, Euclidean and logarithmic distances for long noncoding RNAs, Euclidean and logarithmic distances for protein-coding RNAs, and Euclidean and logarithmic distance ratios.

[0016] Furthermore, sequence structure-related features include: Mw, PI, Gravy, and Instability index. Sequence structure-related features are important indicators for long non-coding RNAs (LNRs) because LNRs and protein-coding RNAs have many different sequence structure properties. Mw refers to the predicted peptide molecular weight, PI refers to the predicted theoretical isoelectric point of the peptide, and Gravy and Instability index represent the average hydrophilicity and stability of the peptide, respectively. EIIP values ​​are used to explore its physicochemical properties; EIIP was originally used to locate exons. Each nucleotide (A, C, G, T) has an EIIP value, which represents the energy of the free electrons in the nucleotide. For any DNA sequence, nucleotides can be replaced by the following EIIP values: (A: 0.1260; C: 0.1340, G: 0.0806, T: 0.1335). EIIP values ​​are directly applied to RNA sequences, avoiding potential biases caused by inferred translation processes. Let X... e [n] is the sequence of EIIP values ​​for the sequence Seq[n], for X e [n] The corresponding power spectrum {S} can be obtained using the Fast Fourier Transform. e [k]}(k=0,1,2,…,N-1): S e [k] = |X e [k]| 2 For protein-coding transcripts, there is usually a distinct peak at the N / 3 position, but such a peak cannot be found in non-coding transcripts.

[0017] Furthermore, features related to open reading frames include: ORF length, ORF coverage, and ORF-attention features;

[0018] ORF is a translatable reading frame, and the ORF of protein-coding sequences is often longer than that of long non-coding RNAs. Commonly used ORF length features include the length of the first ORF and the length of the longest ORF. ORF length is one of the most fundamental features for distinguishing between long non-coding RNAs and protein-coding RNAs because it is unlikely that a long hypothetical ORF will be randomly and accidentally observed in a long non-coding RNA sequence. ORF coverage is the ratio of the longest ORF to the transcript length. This feature also has good classification ability. It is complementary to ORF length and independent of ORF length. To enrich the feature space, ORF-attention features are also introduced. ORF-attention features are derived from the codon-related features, GC-related features, transcript-related features, and structure-related features of the three longest ORFs in the RNA sequence.

[0019] S2. Redundant features are filtered from RNA sequence feature vectors through feature selection;

[0020] Furthermore, the extracted features may include redundant features. Therefore, a feature selection process is used to filter out redundant features for encoding potential prediction. The CatBoost algorithm is used to calculate the feature importance of each feature. Based on the obtained feature importance, the mRMR-IFS method is used to select the best feature subset. mRMR was developed by Peng et al., which selects good features based on the interaction features of the minimum redundancy and maximum association criteria. The features are sorted using the mRMR procedure. Then, incremental feature selection (IFS) is used to add features one by one according to the mRMR sorting results. For each added feature, a new feature subset is generated. Based on the obtained new feature subset, the corresponding feature subset is selected from the entire feature set. The selected feature subset is then subjected to 10-fold cross-validation to select the best feature subset as the final result.

[0021] S3. Fuse the filtered RNA sequence feature vectors;

[0022] Each feature can be obtained from the sequence through scientific computing. There are 5 types of features. Among them, the ORF feature consists of two parts. After encoding, there are a total of 6 feature vectors. The concatenate function in Python is used to concatenate these 6 vectors together to obtain the feature vector of the protein transcript or long non-coding RNA.

[0023] S4. Input the fused RNA sequence feature vector into the CatBoost model for decision classification, and determine whether it is a long non-coding RNA based on the probability score.

[0024] The CatBoost model first selects the tree structure and then assigns values ​​to the leaf nodes of the fixed tree. One of the most important improvements of CatBoost is unbiased gradient estimation to control overfitting. To achieve this, in each iteration of boosting, it excludes the sample from the training set of the current ensemble model to estimate the gradient of each sample in order to truly estimate the gradient. Other improvements include automatically converting categorical features into numerical features without any preprocessing. CatBoost uses a greedy strategy to consider combinations to improve accuracy when creating split points for the current tree, overcome gradient bias, and effectively avoid overfitting.

[0025] Decision trees divide the feature space R into segments based on the values ​​of certain splitting attributes 'a'. mThe Catboost machine learning algorithm is a model built by recursively dividing the model into several disjoint regions (tree nodes). It uses Prediction Values ​​Change (PVC) or Loss Function Change (LFC) to rank the features of the model. PVC is used to calculate the observed change in prediction when the value corresponding to a feature changes. PVC is the default method used in Catboost-based machine learning models and is one of the feature-based ranking methods embedded in the model. LFC is used to rank a specific model among a set of models.

[0026] F = {f1, f2, f3, ..., f n} (1)

[0027] P = β i F j (2)

[0028] F represents a set of input features, f1, f2, f3, ..., f n Let β be a feature of a certain dimension. Equation 1 represents the feature set passed to the machine learning model, and in Equation 2, P represents the predicted value of the substitute numerical factor. i F represents the numerical factor. j Represents a specific feature selected from a given feature set;

[0029] P i+1 =β i+1 F j (3)

[0030] P i＝0 ≠P i ≠P i+1 (4)

[0031] In Equation 3, P i+1 β represents the predicted value as the numerical factor changes. i+1 This represents the modified numerical factor. Equation 4 illustrates how changes in the numerical factor cause changes in the predicted value.

[0032] The CatBoost algorithm was used as the basic classifier for long non-coding RNA identification. It has strong classification and generalization capabilities and can accurately distinguish between long non-coding RNA and protein-coding RNA. It obtains a probability score between 0 and 1. A score greater than or equal to 0.5 is judged as long non-coding RNA, and a score less than 0.5 is judged as protein-coding RNA.

[0033] The beneficial effects of this invention are:

[0034] 1. The method of this invention is compared with LncFinder, CPAT, CPC2, PLEK and CNCT. Through horizontal comparison of sensitivity, specificity, accuracy, F-measure, MCC and accuracy, the overall performance of this invention is significantly better than other algorithms by AUC calculation.

[0035] 2. Compared with the problems of overfitting, large computational load, and long training time of neural networks, the method of this invention has a small computational load, a short training time, and overcomes the overfitting problem. Attached Figure Description

[0036] Figure 1 This is a block diagram of the long non-coding RNA recognition method based on the CatBoost algorithm of the present invention;

[0037] Figure 2 A comparison of the invention before and after using the open reading frame feature;

[0038] Figure 3 This invention provides a comparison of its method with five other commonly used methods on the human dataset.

[0039] Figure 4 This is a comparison of the method of this invention with five other commonly used methods on the wheat dataset;

[0040] Figure 5 This invention provides a comparison of its method with five other commonly used methods on the Chicken dataset.

[0041] Figure 6 This study compares the proposed method with five other commonly used methods on the zebrafish dataset.

[0042] Figure 7 The ROC curves of the method of this invention are compared with those of five other commonly used methods. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0044] like Figure 1 As shown, the long non-coding RNA identification method based on the CatBoost algorithm includes the following steps:

[0045] S1. Obtain codon-related features, guanine-cytosine-related features, transcript-related features, sequence structure-related features, and open reading frame-related features from the database containing several RNA sequences, and establish an RNA sequence feature vector;

[0046] The datasets were obtained from the GENCODE and Ensembl databases. The specific details of the datasets are shown in Table 1.

[0047] Table 1 Dataset

[0048]

[0049] Feature extraction: This step extracts a total of 5 types of features;

[0050] S2. Redundant features are filtered from RNA sequence feature vectors through feature selection;

[0051] S3. Fuse the filtered RNA sequence feature vectors;

[0052] Encoding the RNA sequence involves encoding features extracted from a sequence containing A, T, C, and G, resulting in a 379-dimensional feature vector F = {f1, f2, ..., f...}. 379}

[0053] S4. Input the fused RNA sequence feature vector into the CatBoost model for decision classification, and determine whether it is long non-coding RNA or protein-coding RNA based on the probability score.

[0054] The encoded dataset is fed into the training model built by the CatBoost algorithm with a learning rate of 0.01, a decision tree depth of 10, a loss function of "Logloss", 30,000 training epochs, and 3,000 early-stop epochs.

[0055] To evaluate the performance of the method, we used seven widely used evaluation metrics, including accuracy (ACC), sensitivity (SEN), specificity (SPE), F-measure (F1), Matthew's correlation coefficient (MCC), and area under the ROC curve (AUC). The definitions of ACC, SEN, SPE, PRE, F1, and MCC are as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062] In this method, FP represents a false positive, FN represents a false negative, TP represents a true positive, and TN represents a true negative. Positive examples are long non-coding RNAs, negative examples are protein-coding RNAs, and the area under the ROC curve and the coordinate axes is the AUC.

[0063] The performance of the method of this invention (LncEnsemble) was compared with that of several existing popular long noncoding RNA identification methods (including CPAT, PLEK, CPC2, CNCI, and LncFinder). The comparison results are shown in Tables 2 to 5.

[0064] Table 2 compares the five other methods with our method on the human dataset.

[0065]

[0066] Table 3 compares the five other methods with our method on the wheat dataset.

[0067]

[0068]

[0069] Table 4 compares the five other methods with our method on the Chicken dataset.

[0070]

[0071] Table 5 compares the five other methods with our method on the zebrafish dataset.

[0072]

[0073] As can be seen from Tables 2 to 5, the present invention outperforms several existing methods in multiple metrics across multiple datasets. Furthermore, the performance differences between these methods can be compared more intuitively using ROC curves and bar charts, specifically as follows: Figures 2 to 7 As shown; in summary, the present invention can effectively improve the recognition performance of long non-coding RNAs.

[0074] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for identifying non-diagnostic long non-coding RNAs based on the CatBoost algorithm, characterized in that, Includes the following steps: S1. Obtain codon-related features, guanine-cytosine-related features, transcript-related features, sequence structure-related features, and open reading frame-related features from the database containing several RNA sequences, and establish an RNA sequence feature vector; Codon-related features include: the number of stop codons, the frequency of stop codons, and the Fickett score; guanine-cytosine-related features include: GC frame score and GC content; transcript-related features include: k -mer, CTD, and Hexamer; Sequence structure-related features include: Mw, PI, Gravy, Instability index, and EIIP value; Open reading frame-related features include: ORF length, ORF coverage, and ORF-attention features; Mw represents the predicted peptide molecular weight, while Gravy and Instability index represent the average values ​​of the peptide's hydrophilicity and stability, respectively. ORF-attention features are derived from the codon-related features, guanine-cytosine-related features, transcript-related features, and structure-related features of the first three longest ORFs in the RNA sequence; EIIP values ​​in sequence structure-related features are used to explore the physicochemical properties of sequence structure-related features. EIIP values ​​are directly applied to RNA sequences. For sequence The EIIP value sequence, for The corresponding power spectrum can be obtained using the Fast Fourier Transform. : For protein-coding transcripts, there is a distinct peak at the N / 3 position, but no distinct peak is observed in non-coding transcripts. S2. Redundant features are filtered from RNA sequence feature vectors through feature selection; S3. Fuse the filtered RNA sequence feature vectors; Step S3 includes: using the concatenate function to concatenate the feature vectors of the six filtered RNA sequences together to obtain the feature vectors of the concatenated protein transcript or long non-coding RNA; S4. Input the fused RNA sequence feature vector into the CatBoost model for decision classification, and determine whether it is long non-coding RNA or protein-coding RNA based on the probability score. The feature vectors of the spliced ​​RNA sequence were sorted using the LFC method of Catboost; the decision tree learning rate, decision tree depth, loss function, number of training rounds and number of early-stop rounds were set to obtain the decision probability score, and the decision probability score was used to determine whether it is long non-coding RNA or protein-coding RNA; The CatBoost model includes: in, F Represents a set of input features, For a certain feature, The predicted value representing the alternative numerical factor. Represents numerical factors, Represents a specific feature selected from a given feature set; in, This represents the predicted value as the numerical factor changes. This indicates the modified numerical factor.

2. The method for identifying non-diagnostic long non-coding RNAs based on the CatBoost algorithm according to claim 1, characterized in that, Step S2 includes: S21. Use the mRMR-IFS method to select the best feature subset and use the mRMR program to sort the features. S22. Use incremental feature selection to add features one by one according to the mRMR sorting results. For each additional feature, generate a new feature subset. S23. Based on the obtained new feature subset, select the corresponding feature subset from the entire feature set, and select the best feature subset as the final result by performing 10-fold cross-validation on the selected feature subset.

Images