A lightGBM-based lncRNA subcellular localization prediction method

Abstract

The application discloses a kind of lncRNA subcellular localization prediction method based on lightGBM, comprising, first, the nucleotide of the front section multiple of known lncRNA sequence is intercepted as sequence sample one;Then by based on single strand multiclass position-specific three nucleotide bias and reverse complementary kmer, sequence sample one is respectively characterized coding, the combination of two kinds of feature coding is vector;Use lightGMB as learning algorithm;Using 5-fold cross validation optimization reverse complementary kmer and LightGBM's hyperparameter;Most intercept unknown lncRNA sequence front end multiple bases as sequence sample two, and its single strand multiclass position-specific three nucleotide bias and the combination of optimized reverse complementary kmer feature coding is input in trained lightGBM, will obtain its localization subcellular type, the patent of the application can be according to long-chain non-coding RNA sequence prediction in cytoplasm, nucleus, ribosome, cytosol, exosome five subcellular positions, the application realizes simply, and prediction precision is high.

Description

Technical Field

[0001] This invention relates to the field of computational biomolecular science, and in particular to a method for predicting the subcellular localization of lncRNAs based on lightGBM. Background Technology

[0002] lncRNAs (long non-coding RNAs) are RNA transcripts longer than 200 nucleotides, transcribed from DNA but never coding proteins. Numerous experimental studies have shown that lncRNAs function by interacting with various types of biomolecules, such as DNA, RNA, or proteins. For example, some lncRNAs control gene expression, some influence DNA damage and repair responses, and others participate in splicing, turnover, translation, and signaling pathways. Xing's paper (Role of lncRNA LUCAT1 in cancer) found that lung cancer-related transcript 1 (LUCAT1) is involved in the regulation of various tumors, including lung cancer, breast cancer, ovarian cancer, thyroid cancer, and renal cell carcinoma, and is therefore considered a potential biomarker for cancer prognosis and a therapeutic target. The role of lncRNAs in cellular processes is closely related to their subcellular localization. Subcellular localization determines which partners they interact with and what post-transcriptional or co-transcriptional regulatory modifications occur, and influences external stimuli that directly affect lncRNA function. In addition, subcellular lncRNAs located in the nucleus are generally more abundant and less stable than those located in the cytoplasm. Therefore, the functions of subcellular lncRNAs located in the nucleus differ from those located in the cytoplasm. The former regulates the transcriptional program through chromatin interactions and remodeling, while the latter controls gene expression post-transcriptionally through signaling pathways, translation processes, and gene expression. Therefore, accurately identifying the subcellular location of lncRNAs is of great significance to the development of bioinformatics.

[0003] With the development of artificial intelligence, in the past five years, an increasing number of scholars have begun to use AI to solve the problem of predicting the subcellular location of lncRNAs. Su's paper (iLoc-lncRNA: predict the subcellular location of lncRNAs by incorporating octamer composition into general PseKNC) uses the feature PseKNC (pseudo k-tuple nucleotide composition, PseKnc) to represent lncRNA sequences and establishes a stochastic vector machine-based model (iLoc-lncRNA) for predicting lncRNA subcellular location. Zeng's paper (DeepLncLoc: a deep learning framework for long non-coding RNA subcellular localization prediction based on subsequence embedding) divides the lncRNA sequence into m consecutive subsequences and uses word2vec to extract semantic features of the sequence for training the deep learning-based model DeepLncLoc. However, the former model, which uses k-mer features, cannot adequately represent lncRNA information. While the latter model extracts semantic features, the large differences in lncRNA lengths inevitably lead to a significant loss of information in the average pooling layer used in the final deep learning model. Therefore, the predictive performance of both classifiers is limited. Summary of the Invention

[0004] To address the problems mentioned in the background section, this invention provides a method for predicting subcellular localization of lncRNA based on lightGBM, thus resolving the issues of existing technologies.

[0005] The technical solution adopted in this invention is:

[0006] A method for predicting subcellular localization of lncRNAs based on lightGBM includes the following steps:

[0007] S1: Collect known lncRNA sequences and extract the first 166 bases of lncRNAs located in the cell nucleus, cell membrane, cytoplasm, ribosome, and exosome as sequence samples for use as the training set;

[0008] S2: The sequence sample is characterized by trinucleotide bias and reverse complementary kmer based on single-strand multi-position specificity. The combination of the two feature codes is then used as a vector.

[0009] S3: Use lightGMB as the learning algorithm;

[0010] S4: Optimize the hyperparameters of inverse complementary kmer and LightGBM on the training set using 5-fold cross-validation;

[0011] S5: Extract the first 166 bases of the unknown lncRNA sequence as sequence sample two, and input the combination of its single-strand multi-position-specific trinucleotide bias and optimized inverse complementary kmer feature encoding into the trained lightGBM to obtain its localized subcellular type.

[0012] Furthermore, the method for feature encoding the sequence samples based on single-strand multi-position-specific trinucleotide bias in steps S2 and S5 is as follows:

[0013] Sequence S is split into three consecutive nucleotides, such as s1s2s3, s2s3s4, ..., s i s i+1 s i+2 ,…,s L-2 s L-1 s L s i s i+1 s i+2 This represents three consecutive nucleotides at the i-th position; a position-specific trinucleotide bias matrix can be represented as...

[0014]

[0015] Where: k represents the subcellular type of lncRNA. This represents the probability that the three nucleotides of the i-th lncRNA sequence appear at position j in the subcellular localization of class k.

[0016] Furthermore, the position-specific trinucleotide bias matrix can be estimated using the corresponding frequencies in the training set. By querying the position-specific trinucleotide bias matrix, a value can be obtained for any three consecutive nucleotides in the lncRNA sequence. make An array is used to record the subcellular type corresponding to three consecutive nucleotides at each position of the lncRNA in the sequence. The subcellular type that appears most frequently in a sequence is counted, and its corresponding position-specific trinucleotide bias matrix is ​​used to encode the lncRNA.

[0017]

[0018] in: j is determined by the types of the three nucleotides at position i in the sequence.

[0019] Furthermore, the method for feature encoding of sequence samples by inverse complementary kmer in steps S2 and S5 is as follows:

[0020]

[0021] Where: N is the number of aggregate types of reverse complementary kmer, N () This represents the number of times the reverse complementary nucleotide i appears.

[0022] Furthermore, the number of polymer types of the reverse complementary kmer is calculated by the following formula:

[0023]

[0024] Furthermore, in step S3, LightGBM is a lightweight gradient boosting tree. A gradient boosting tree is an additive model consisting of a set of decision trees. Each time, an optimal tree is found to fit the difference between the target and the previous decision trees. Constructing the optimal tree is an iterative process that seeks the split that minimizes the following formula:

[0025]

[0026] Among them: I L and I R Let I represent the left and right subtrees to be split, respectively, and g represent the original splitting node. i and h i Let represent the first-order and second-order gradients, respectively, and γ represent the number of leaf nodes.

[0027] Furthermore, in step S4, the 5-fold cross-validation is used to pair k in the inverse complementary kmer, including the following:

[0028] S4.1.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric.

[0029] S4.1.2 Take k as 2, 3, 4, 5, and 6 in the reverse complementary kmer respectively, and calculate the average accuracy in 5-fold cross-validation.

[0030] Furthermore, k in the reverse complementary kmer is set to 5.

[0031] Furthermore, a grid search was performed to optimize four parameters in lightGBM: minimum number of leaf nodes, feature sampling rate, maximum tree depth, and maximum number of leaf nodes. The optimization included the following:

[0032] S4.2.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric.

[0033] S4.2.2 Set the search range of the minimum number of leaf nodes in lightGBM to [10, 30], the range of feature sampling rate to [0.5, 1.0], the range of maximum tree depth to [3.0, 7.0], and the range of maximum leaf nodes to [8, 120], and calculate the average accuracy in 5-fold cross-validation.

[0034] Furthermore, the optimized parameters are: minimum number of leaf nodes is 17, feature sampling ratio is 1, maximum tree depth is 5, and maximum number of leaf nodes is 24.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention primarily utilizes single-stranded, multi-position-specific trinucleotide bias and reverse complementary kmers to encode lncRNA sequences, and constructs a classification model using an optimized LightGBM. This invention can predict five major subcellular localizations: cytoplasm, nucleus, ribosomes, cytosol, and exosomes. Experimental results show that the accuracy of predicting lncRNA subcellular localization using this method is higher than other methods. Attached Figure Description

[0037] Figure 1 : A schematic diagram of the process of this invention;

[0038] Figure 2 The impact of adjusting different parameters of LightGBM on the model's prediction accuracy;

[0039] Figure 3 ROC curves for subcellular locations of each class of lncRNAs after 5-fold cross-validation. Detailed Implementation

[0040] To clearly illustrate the technical features of this solution, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the scope of protection of the present invention.

[0041] Example 1

[0042] One embodiment of the present invention provides a method for predicting subcellular localization of lncRNAs based on lightGBM, comprising the following steps:

[0043] S1: Collect known lncRNA sequences. 857 long non-coding RNA sequences were collected from the paper (DeepLncLoc: a deep learning framework for long non-coding RNA subcellular localization prediction based on subsequenceembedding). Among them, the number of long non-coding RNAs located in the cell nucleus, cell membrane, cytoplasm, ribosome, and exosome were 325, 88, 328, 88, and 28, respectively. The first 166 bases of the lncRNAs located in the cell nucleus, cell membrane, cytoplasm, ribosome, and exosome were extracted as sequence samples 1 for use as the training set.

[0044] S2: The sequence sample is characterized by trinucleotide bias and reverse complementary kmer based on single-strand multi-position specificity. The combination of the two feature codes is then used as a vector.

[0045] A method for feature encoding sequence samples based on single-stranded multi-position-specific trinucleotide bias.

[0046] Sequence S is split into three consecutive nucleotides, such as s1s2s3, s2s3s4, ..., s i s i+1 s i+2 ,…,s L-2 s L-1 s L s i s i+1 s i+2 This represents three consecutive nucleotides at the i-th position; a position-specific trinucleotide bias matrix can be represented as...

[0047]

[0048] Where: k represents the subcellular type of lncRNA. This represents the probability that the three nucleotides of class i appear at position j in the subcellular localization of class k.

[0049] Position-specific trinucleotide bias matrices can be estimated using the corresponding frequencies in the training set. By querying the position-specific trinucleotide bias matrix, a value can be obtained for any three consecutive nucleotides in the lncRNA sequence. make An array is used to record the subcellular type corresponding to three consecutive nucleotides at each position of the lncRNA in the sequence. The subcellular type that appears most frequently in a sequence is counted, and its corresponding position-specific trinucleotide bias matrix is ​​used to encode the lncRNA.

[0050]

[0051] in: j is determined by the types of the three nucleotides at position i in the sequence.

[0052] The method for feature encoding of sequence samples using the inverse complementary kmer algorithm is as follows:

[0053] Reverse complement kmer encoding is a variant of kmer feature encoding. Kmer feature encoding counts the frequency of k nucleotide clusters in a sequence. Reverse complement kmer encoding treats k nucleotide clusters and their corresponding reverse complement k nucleotide clusters as a class and calculates their frequencies. For example, with k=2, the reverse complement of "AA" is "TT", "AC" is "GT", "AG" is "CT", "AT" is "AT", "CA" is "TG", "CC" is "GG", "CG" is "CG", "GA" is "TC", "GC" is "GC", and "TA" is "TA". Therefore, reverse complement kmer encoding only requires counting 10 2-mers, i.e.,

[0054]

[0055] CA(TG), CC(GG), CG, GA(TC), GC, TA}

[0056] Where N is the number of aggregate types of reverse complementary kmer, N () Let N be the number of occurrences of the reverse complementary nucleotide i. For example, to count the reverse complementary 2mers of the sequence AATTGCA, N is 6. AA appears once, TT appears once, so the frequency of AA is 2 / 6; AT appears once, so the frequency of AT is 1 / 6; TG appears once, CA appears once, so the frequency of CA is 2 / 6; GC appears once, so the frequency of GC is 1 / 6. The frequency of reverse complementary kmers of other lengths can be calculated similarly.

[0057] The number of polymer types of reverse complementary kmers is calculated using the following formula:

[0058]

[0059] S3: LightGBM is used as the learning algorithm. LightGBM is a lightweight gradient boosting decision tree (GBDT). Gradient boosting trees are an additive model consisting of a set of decision trees. Each iteration searches for an optimal tree to fit the difference between the target and the previous decision trees. Building the optimal tree is an iterative process that seeks the split that minimizes the following expression:

[0060]

[0061] Where I L and I R Let I represent the left and right subtrees to be split, respectively, and g represent the original splitting node. i and h i Let represent the first and second gradients, respectively, and γ represent the number of leaf nodes. During training, lightGBM employs a histogram algorithm to transform traversing samples into traversing the histogram, reducing time complexity; it uses a one-sided gradient algorithm to filter out samples with small gradients, reducing significant computation; and it uses a leaf-wise growth strategy to construct the tree, further reducing unnecessary computation. Therefore, lightGBM supports highly efficient parallel training and offers advantages such as faster training speed, lower memory consumption, better accuracy, and distributed support for rapidly processing massive amounts of data.

[0062] S4.1 uses 5-fold cross-validation to optimize k in the inverse complementary kmer.

[0063] S4.1.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric.

[0064] S4.1.2 Take k as 2, 3, 4, 5, and 6 in the reverse complementary kmer respectively, and calculate the average accuracy in 5-fold cross-validation.

[0065] Based on the results in Table 1, the inverse complementary kmer encoding with k=5 was finally selected.

[0066] Table 1. Prediction accuracy of combined features of inverse complementary kmer at different k values

[0067] k accuracy 2 0.692 3 0.696 4 0.698 5 0.700 6 0.672

[0068] S4.2 uses 5-fold cross-validation to optimize the hyperparameters of LightGBM on the training set.

[0069] S4.2.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric.

[0070] S4.2.2 Set the search range for the minimum number of leaf nodes in lightGBM to [10, 30], the feature sampling rate to [0.5, 1.0], the maximum tree depth to [3.0, 7.0], and the maximum number of leaf nodes to [8, 120]. Then, calculate the average accuracy in 5-fold cross-validation. Figure 2 As shown.

[0071] The optimized parameters are: minimum number of leaf nodes = 17, feature sampling ratio = 1, maximum tree depth = 5, and maximum number of leaf nodes = 24. Substituting these parameters into lightGBM yields the optimized lightGBM.

[0072] Example 2

[0073] One embodiment of the present invention provides a method for predicting subcellular localization of lncRNAs based on lightGBM, comprising the following steps:

[0074] Step S1 in Example 1: Collect known lncRNA sequences, and extract the front, back, or multiple random bases of lncRNAs located in the cell nucleus, cell membrane, cytoplasm, ribosome, and exosome as sequence samples for use as a training set;

[0075] Step S2 in Example 1: The sequence sample 1 is characterized by trinucleotide bias and reverse complementary kmer (where k=5) based on single-strand multi-position specificity of trinucleotides, and the combination of the two feature codes is connected into a vector;

[0076] Step S3 in Example 1: Use lightGMB as the learning algorithm, where the minimum number of leaf nodes is 17, the feature sampling ratio is 1, the maximum tree depth is 5, and the maximum number of leaf nodes is 24.

[0077] S4. Divide the sample into 5 equal parts, 4 parts for training the light GBM and 1 part for testing the trained light GBM. Repeat this process 5 times, with each sample having a unique test. The average accuracy of the 5 tests is used as the evaluation metric; the results are shown in Table 2. The results from the first part are the best.

[0078] Table 2. Comparison of prediction results for different nucleotide segments

[0079]

[0080] Example 3

[0081] This invention provides an embodiment of a method for predicting subcellular localization of lncRNAs based on lightGBM, including all the contents of Examples 1 and 2. The method proposed in this patent was compared with existing prediction methods (iLoc-lncRNA and DeepLncLoc) on 67 lncRNAs. To quantitatively evaluate the performance of the method, precision, recall, accuracy (Accuracy, Acc), and F1 score were used as evaluation indicators, calculated as follows.

[0082]

[0083] Among them, pred i The predicted category for sample i, label i Let I be the true class, and 1 be the indicator function, where 1 is True and 0 is False. N is the number of samples.

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] Among them: TP () FP represents the number of samples that belong to class i and are correctly predicted as class i. () FN represents the number of samples that do not belong to class i but are predicted to belong to class i. () This represents the number of samples that belong to class i and are predicted to be other classes.

[0090] Table 3. Prediction Results

[0091]

[0092] The results show that the method proposed in this patent is significantly superior to existing methods.

[0093] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting subcellular localization of lncRNAs based on lightGBM, characterized in that, Includes the following steps: S1: Collect known lncRNA sequences and extract multiple bases from the beginning of lncRNAs located in the cell nucleus, cell membrane, cytoplasm, ribosomes, and exosomes as sequence samples for use as the training set; S2: The sequence sample is characterized by trinucleotide bias and reverse complementary kmer based on single-strand multi-position specificity. The combination of the two feature codes is then used as a vector. The method for feature encoding sequence samples based on single-strand multi-position-specific trinucleotide bias is as follows: Sequence S is split into 3 consecutive nucleotides, as follows: , , …, ,…, , This represents three consecutive nucleotides at the i-th position; The position-specific trinucleotide bias matrix is ​​represented as follows: Where: k represents the subcellular type of lncRNA. This represents the probability that the three nucleotides of the i-th lncRNA sequence appear at position j in the subcellular localization of class k; The sequence S is represented as: in: j is determined by the types of the three nucleotides at the i-th position in the sequence; The method for feature encoding sequence samples using the inverse complementary kmer algorithm is as follows: Where: N is the number of polymer types of reverse complementary kmers, calculated using the following formula: ； This represents the number of times the reverse complementary nucleotide i appears; S3: Use lightGMB as the learning algorithm; S4: Optimize the hyperparameters of inverse complementary kmer and LightGBM on the training set using 5-fold cross-validation; S4.

1. The optimization of k in the reverse complementary kmer using 5-fold cross-validation includes: S4.1.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric. S4.1.2 Take the k values ​​of 2, 3, 4, 5, and 6 in the reverse complementary kmer respectively, and calculate the average accuracy in the 5-fold cross-validation. S4.

2. Optimize the following four parameters in LightGBM using a grid search: minimum number of leaf nodes, feature sampling rate, maximum tree depth, and maximum number of leaf nodes. S4.2.1 Divide the training set into 5 equal parts, 4 parts for training and 1 part for testing, repeat 5 times, each sample has a unique test, and take the average accuracy of the 5 tests as the evaluation metric. S4.2.2 Set the search range of the minimum number of leaf nodes in lightGBM to [10, 30], the feature sampling rate range to [0.5, 1.0], the maximum tree depth range to [3.0, 7.0], and the maximum number of leaf nodes in the tree range to [8, 120], and calculate the average accuracy in 5-fold cross-validation respectively; S5: Extract multiple bases from the beginning of an unknown lncRNA sequence as sequence sample two, and input the combination of its single-strand multi-position-specific trinucleotide bias and optimized inverse complementary kmer feature encoding into the trained lightGBM to obtain its localized subcellular type.

2. The method for predicting subcellular localization of lncRNA based on lightGBM according to claim 1, characterized in that, In step S3, LightGBM is a lightweight gradient boosting tree. Gradient boosting trees are additive models consisting of a set of decision trees. Each time, an optimal tree is found to fit the difference between the target and the previous decision trees. Constructing the optimal tree is an iterative process that seeks the split that minimizes the following formula: in: and Let I represent the left and right subtrees to be split, respectively, and let I represent the original splitting node. and These represent the first and second order gradients, respectively. This indicates the number of leaf nodes.

3. The method for predicting subcellular localization of lncRNA based on lightGBM according to claim 1, characterized in that, Set k in the reverse complementary kmer to 5.

4. The method for predicting subcellular localization of lncRNA based on lightGBM according to claim 1, characterized in that, The optimized parameters are: minimum number of leaf nodes is 17, feature sampling ratio is 1, maximum tree depth is 5, and maximum number of leaf nodes is 24.

Images