CN111477271B - MicroRNA prediction method based on supervised self-organizing mapping neural network - Google Patents

Abstract

The invention discloses a microRNA prediction method based on a supervised self-organizing mapping neural network, which comprises the following steps of: extracting features based on the microRNA sequence, obtaining secondary structure base pairing information of the microRNA sequence to be detected by using an RNAfold program, and converting the secondary structure base pairing information into vectorization expression by combining the features based on the primary sequence and the secondary structure to obtain a feature vector finally used for calculating a prediction means; receiving characteristic data of an input layer, learning a spatial distribution rule of the input data by implicit layer self-organizing mapping, and performing mapping processing between data to form new characteristic representation in the process; and step 3: the output supervision layer calculates output categories and related errors by using the generated new feature representation, and reversely transmits the errors to update the network weight; and setting a threshold value by using a cross validation method, and acquiring a classification result. The invention solves the defects of high cost, long time consumption and the like in the traditional biological prediction means and the existing machine learning prediction method.

Description

MicroRNA prediction method based on supervised self-organizing map neural network

technical field

The invention relates to the field of bioinformatics prediction of microRNA, in particular to a microRNA prediction method based on a supervised self-organizing map neural network.

Background technique

Different from coding RNA (coding RNA) that can be translated into protein, small RNA molecules with a length of about 20 bases that cannot be translated into proteins are collectively referred to as noncoding RNA (ncRNA), which has new and unexplored characteristics. Known gene expression regulation functions. There are many types of non-coding RNAs, including ribosomal RNA (rRNA), transfer RNA (tRNA), and microRNA (microRNA, miRNA), which play important roles in a wide range of biological processes and contribute to many diseases. Such as cancer has a major impact. There are mainly three types of non-coding small RNAs that are known and studied in depth at present, namely siRNA, miRNA and piRNA. These non-coding small RNAs can not only regulate the transcription stage of living species, but also realize post-transcriptional regulation. plays a very important role in development.

MicroRNA (miRNA) is a single-stranded and small non-coding RNA molecule (about 17-25 nt) that is widely present in eukaryotes. The conversion of microRNA molecules from the initial state to mature microRNA molecules requires multi-step processing. Studies have shown that microRNAs regulate genes encoded at the transcriptional level of organisms by binding to mRNAs. Aberrant expression of microRNAs has been observed in many cancers and other disease states, suggesting that aberrant expression of microRNAs is deeply linked to these diseases. Studies have reported that viral microRNAs are associated with human diseases, especially cancer, for example, Epstein-Barr virus, hepatitis B and C viruses, and human papilloma virus are highly associated with gastric, nasopharyngeal, liver, and cervical cancer, respectively. Therefore, it is extremely important to distinguish true miRNAs from pseudo-miRNAs based on hairpin sequences with similar loop structures, and in-depth research on microRNAs is also extremely important for microRNA-based disease treatment.

G,Gong F,Chen R:Integrated sequence-structure motifs suffice toidentify microRNA precursors.PloS One 2012,7.)使用了一组 1300个序列结构模体特征；HuntMi(Gudys A,Szcze′sniak MW,Sikora M,Makalowska I:′ HuntMi:an efficientand taxon-specific approach in pre-miRNA identification.BMC Bioinformatics2013,14:83.)的特征集合并了microPred中使用的特征、triplet-SVM特征集中的四个特征和其他三个特征：序列中检测到的低复杂度区域的百分比、不含终止密码子的氨基酸链最大长度和内环的累积大小。ViralmiR(Kai-Yao Huang,Tzong-Yi Lee,Yu-Chuan Teng,Tzu-Hao Chang.ViralmiR:a support-vector-machine-based method for predictingviral microRNA precursors[C].Asia-Pacific Bioinformatics Conference.CurranAssociates,2015:86-92.)是一种基于支持向量机的病毒microRNA前体预测方法，从先前研究中选择54个特征并使用特征得分评价特征的鉴别能力。Liu等人(Liu B,Fang L,LiuF,et al.iMiRNA-PseDPC:microRNA precursor identification with a pseudodistance-pair composition approach[J].Journal of Biomolecular Structure andDynamics,2015(ahead-of-print):1-13.)以支持向量机(SVM)为分类器，提出了一种新的特征向量，称为伪距离对合成(Pseudo Distance Pair Composition,PseDPC)来保留序列顺序信息。Common methods for identifying microRNAs include molecular biology research methods and bioinformatics research methods. Computational prediction methods mainly include: homologous fragment search methods for finding microRNAs homologous to known microRNAs; based on comparative genomics, conservative Predictive methods for screening and scoring candidate fragments based on sequence and structural features; methods for predicting binding targets of plant miRNAs; based on machine learning (ML) methods, including support vector machines (Support Vector Machines) , SVM), Random Forest (Random Forest, RF), Hidden Markov Model (Hidden Markov Model, HMM), Naive Bayes (Naive Bayes, NB) and Linear Genetic Programming (Linear Genetic Programming, LGP) and so on. Triplet SVM (Xue C, et al: Classification of real and pseudo microRNA precursors using local structure-sequence features and support vector machine. BMC Bioinformatics 2005, 6:310.) used a set of 32 sequence structure features; MiPred (Ng KL, Mishra SK: De novo SVM classification of precursormicroRNAs from genomic pseudo hairpins using global and intrinsic folding measures. Bioinformatics 2007, 23(11): 1321-30) fusion using 32 sequence structural features, Minimum of Free Energy (MFE) ) and stability metric (randfold) feature set; MicroPred (Batuwita R, Palade V: microPred: effective classification of pre-miRNAs for human miRNA gene prediction. Bioinformatics 2009, 25(8): 989-995.) consists of 48 sequences and a subset of the structural feature set, including 21 features; G ² DE (Hsieh CH, Chang DTH, Hsueh CH, Wu CY, Oyang YJ: Predicting microRNA precursors with a generalized Gaussiancomponents based density estimation algorithm.BMC Bioinformatics 2010, 11 (Suppl 1): S52.) is induced by a subset of 7 features from a set of 48 sequence and structural features; Mirident (Liu X, He S,

G, Gong F, Chen R: Integrated sequence-structure motifs suffice to identify microRNA precursors. PloS One 2012, 7.) used a set of 1300 sequence-structure motif features; HuntMi (Gudys A, Szcze'sniak MW, Sikora M, Makalowska I:' HuntMi: an efficient and taxon-specific approach in pre-miRNA identification. BMC Bioinformatics 2013, 14:83.) The feature set incorporates the features used in microPred, four features in the triplet-SVM feature set, and three others Features: Percentage of low-complexity regions detected in the sequence, maximum length of amino acid chain without stop codons, and cumulative size of inner loops. ViralmiR(Kai-Yao Huang,Tzong-Yi Lee,Yu-Chuan Teng,Tzu-Hao Chang.ViralmiR:a support-vector-machine-based method for predictingviral microRNA precursors[C].Asia-Pacific Bioinformatics Conference.CurranAssociates,2015 :86-92.) is a support vector machine-based prediction method for viral microRNA precursors, selecting 54 features from previous studies and using feature scores to evaluate the discriminative power of features. Liu et al. (Liu B, Fang L, LiuF, et al. iMiRNA-PseDPC: microRNA precursor identification with a pseudodistance-pair composition approach [J]. Journal of Biomolecular Structure and Dynamics, 2015 (ahead-of-print): 1- 13.) Using Support Vector Machine (SVM) as the classifier, a new feature vector called Pseudo Distance Pair Composition (PseDPC) is proposed to preserve sequence order information.

Studies based on existing facts have shown that there are multiple associations between microRNAs and various diseases. Traditional biological prediction methods often have long experimental time, high experimental cost or other disadvantages and cannot meet the strong needs in the post-genome era.

SUMMARY OF THE INVENTION

In order to solve the defects of high cost and long time in traditional biological prediction methods and existing machine learning prediction methods, the purpose of the present invention is to propose a microRNA prediction method based on supervised self-organizing mapping network that combines microRNA sequence and microRNA structural features .

In order to achieve the above object, the technical scheme adopted in the present invention is as follows:

A method for microRNA prediction based on supervised self-organizing map neural network, including the following steps:

Step 1: Feature extraction, extract the features based on the microRNA sequence, and use the RNAfold program to obtain the base pairing information of the secondary structure of the microRNA sequence to be tested, and convert the features based on the primary sequence and the features based on the secondary structure into a vectorized representation , obtain the eigenvectors that are finally used to calculate the prediction means;

Step 2: Receive the feature data of the input layer, learn the spatial distribution law of the input data by self-organizing mapping of the hidden layer, and perform mapping processing between the data, and form a new feature representation in the process;

Step 3: The output supervision layer uses the new feature representation generated in step 2 to calculate the output category and related errors, and back-transmits the errors to update the network weights;

Step 4: Use the cross-validation method, set the threshold, and obtain the classification result.

Compared with the prior art, the present invention has significant advantages as follows: (1) based on sequence and structural features, not only sequence features but also secondary structure features of microRNA are extracted, which enriches feature sources and is beneficial to improving the prediction performance of the model ; (2) Based on random down-sampling technology, a balanced sub-sample set is extracted from the original unbalanced sample set, which effectively solves the class imbalance problem, which is beneficial to improve the prediction performance of the model; (3) Use a supervised self-organizing mapping neural network , using the self-organizing mapping neural network to learn the distribution law in the high-dimensional input space while maintaining the characteristics of topological neighborhood information in the low-dimensional output space, calculate the new feature representation and perform subsequent processing, and improve the prediction performance of the model; (4) Use cross-validation to test the performance of the model, not only use the threshold of a single label value to divide the classification data, but also use the adjustment threshold of two output states to divide the samples to improve the prediction performance of the model.

Description of drawings

Figure 1 is a schematic diagram of a prediction method based on a supervised self-organizing map neural network.

FIG. 2 is an encoding position deviation diagram.

Detailed ways

A microRNA prediction method based on a supervised self-organizing map network of the present invention comprises the following steps:

Step 1: Feature extraction, converting the base sequence of the microRNA to be tested into a numerical representation that can be processed by a machine. For a microRNA base sequence, first extract its sequence length, GC content, composition and coding position offset (the codon position bias) and other features that can be calculated from the original base sequence, where the k component represents the length of k The frequency of occurrence of the word string in the sequence (the value of k in the prediction method is 1, 2, 3), thus obtaining the 84 (4 ¹ +4 ² +4 ³ ) dimension k component feature; the coding position deviation indicates that the base is in the Bias at three positions of triplet codons. Secondly, the base pairing information of the microRNA based on the minimum free energy of the secondary structure can be obtained through the RNAfold program. Each base in the microRNA has two states of pairing and unpairing, which are represented by a specific symbol; the secondary structure of the statistical sequence Base pairing information, calculate the frequency of 8 (2 ³ ) possible pairing cases for any three adjacent nucleotides, and combine the 4 possible cases of the intermediate nucleotide to obtain 32 (4 × 8) triple element structures Features; the minimum free energy of the secondary structure obtained by the RNAfold program is also used as a feature and normalized; the ratio of AU pairing to sequence length, GC pairing to sequence length, and GU pairing to sequence length are also calculated. . Finally, all the eigenvectors are serially combined to get the total eigenvector for prediction.

Step 2: Random downsampling of negative microRNA samples using random downsampling techniques. Experiments have verified that the number of labeled positive microRNA samples is much smaller than the number of negative microRNA samples. For the dataset in this case, data preprocessing is required. The samples of negative microRNA are randomly down-sampled, and the obtained subset of negative microRNA samples and positive microRNA sample set are used as experimental data sets to solve the problem of unbalanced learning and maintain the balance of positive and negative samples.

Step 3: Use the extracted microRNA sequence feature vector as input, learn the spatial distribution law of the input in the self-organizing map neural network layer, and fully interconnect with the input nodes. Each neuron in the self-organizing map is represented by a set of weight vectors. The best matching unit is used to continuously adjust the weight vector of the best matching unit and its neighboring neurons, maintain the topology information in the low-dimensional output space, and obtain new feature representations. The self-organizing map output node defaults to a 10×10 grid, and other values can be set in the method according to different data.

Step 4: The supervised output layer uses the new feature representation output by the SOM layer, labels the category information, and transmits the error information in the reverse direction. The gradient descent method is used to update the relevant weights by obtaining the derivative to obtain the trained model, and cross-validation is used. The performance of the model is detected, and the binary classification results are obtained by thresholding.

In order to better understand the technical content of the present invention, specific embodiments are given and described below in conjunction with the accompanying drawings.

Fig. 1 is a schematic diagram of the prediction method of the present invention. With reference to Fig. 1, according to an embodiment of the present invention, the steps of the microRNA prediction method based on the supervised self-organizing mapping neural network are described as follows.

First, for each sequence in the microRNA sequence training set, extract the corresponding sequence information features, use the RNAfold program to obtain the secondary structure base pairing information, and calculate and obtain the corresponding secondary structure features, which are expressed in the form of vectors; The features of the sequence are fused to obtain the final feature vector, and combined with the label of the sequence as the machine learning sample corresponding to the sequence, the microRNA sequence training set can be converted into a microRNA sample training set. Second, based on a supervised self-organizing map three-layer neural network, the SOM computes a new representation, the neurons in the supervised layer compute the result using the new representation, the output and associated errors are computed in the forward pass, and gradient descent is used in the back pass The error is transmitted back to update the network weights. Finally, use cross-validation to test the performance of the model, not only use the threshold of a single label value to divide the classification data, but also use the adjustment threshold of the two output states to divide the sample to obtain the classification result of the data.

The process is described in more detail with reference to the drawings.

Step 1: Feature Extraction

For a microRNA base sequence, its sequence length, GC content, composition and coding position bias (the codon position bias) and other features that can be calculated from the original base sequence can be extracted from the original sequence, where the sequence length is The character length of the microRNA sequence consisting of bases (usually A, U, G, C); the GC content is the frequency of occurrences of bases G and C in the microRNA sequence; the k component represents the length of the k word string in the sequence. Occurrence frequency (the value of k in the prediction method is 1, 2, 3), thus obtaining the 84 (4 ¹ +4 ² +4 ³ ) dimension k component features; the coding position deviation indicates that the base is in three types of triplet codons The deviation of the position, for any base x∈{G,C,A,U}, x ₁ is the number of times x appears in 0,3,6,... etc., x ₂ is x in 1,4 ,7,... and other positions, x ₃ is the number of times x appears in 2, 5, 8,... and other positions, and the coding position deviation is shown in Figure 2:

It can be calculated by the following formula:

Secondly, the minimum free energy secondary structure base pairing information of the microRNA can be obtained through the RNAfold program. Each base in the microRNA has two states, paired and unpaired, represented by a specific symbol "(" , ")" and ".", the left bracket "(" indicates that a nucleotide near the 5' end can match another nucleotide near the 3' end, and the corresponding right bracket "(" indicates that a nucleotide near the 3' end A certain nucleotide can be matched with another nucleotide near the 5' end. In the present invention, the two cases can only be represented by "(". Calculate 8 (2 ³ ) kinds of any three adjacent nucleotides. Frequency of possible pairing cases, including "(((", "((.", "(.(", ".((", "(..", ".(.", "..(", and "...", combining 4 possible cases of intermediate nucleotides (A, U, G, C) to obtain 32 (4 × 8) triple element structural features. Minimum free energy of secondary structure obtained by RNAfold program It is also used as a feature and normalized dG (normalized MFE). The ratio of AU pairing to sequence length |AU|/L, the ratio of GC pairing to sequence length |GC|/L and GU pairing to sequence length are also calculated. The ratio |GU|/L. Finally, all the eigenvectors are serially combined to obtain the total eigenvectors for prediction.

Step 2: Self-Organizing Map Network Layer

Self-organizing Map (SOM) is an unsupervised learning model proposed by Kohonen. The training goal is to find a suitable weight vector for each output layer neuron, which can adapt the learning model to the high-dimensional input space. distribution law, and maintain the topological neighborhood information of the input data in the low-dimensional output space. Usually, the output nodes of SOM are arranged in a 2-D regular grid, and each output node is fully interconnected with the input node, corresponding to a prototype vector w _h ∈ R ^m , where m is the dimension of the input space. The training process of SOM is that after receiving a training sample x _i ∈ R ^m , each output layer neuron in the H neuron units of SOM will calculate the distance between the sample and its own weight vector, and the distance between the nearest neuron The unit is called the Best Matching Unit (BMU), and then the weight vector of the best matching unit and its neighboring neurons is adjusted to reduce the distance between the weight vector and the current input sample. This process is iterated until convergence. The best matching unit (BestMatching Unit, BMU) can be calculated by the formula BMU(x _i )=arg min _1≤h≤H ||w _{h -xi} ||, where ||.,.|| are L2 normal form.

Let the learning sample at step t be x _i , update the BMU and its neighbors by the following formula:

为邻域函数，定义了与BMU相邻的输出节点。

其中，r为映射的半径，T是最大迭代次数， d(BMU(x_i),h)是与输入x_i相匹配的最佳神经元BMU(x_i)和神经元h之间的曼哈顿距离(the Manhattan distance)。Among them, α(t) is the learning efficiency, which decreases monotonically with t;

For the neighborhood function, the output nodes adjacent to the BMU are defined.

where r is the radius of the map, T is the maximum number of iterations, and d(BMU( _xi ),h) is the Manhattan distance between the best neuron BMU( _xi ) that matches the input _xi and neuron h (the Manhattan distance).

Step 3: Class Prediction

The activation item of the unit is forwarded through the network to obtain the output value, and the output value is compared with the input category label value to calculate the error value. In order to maintain the organization of the mapping network, the information of the current neuron and its neighbors is obtained.

d(u', u) is the Manhattan distance between neuron u' and neuron u, and α is a constant.

The neurons in the hidden layer and the neurons in the output layer are fully connected, and the two output values of the neurons in the output layer are the sum of the products of the weight vector and the activation term between the H neurons in the hidden layer SOM and the neurons in the output layer. Accumulates the value of the sigmoid function with its output layer bias. The sigmoid function is a common function in biology, and is also often used as the activation function of neural networks. Define sig(x)=1/(1+exp(-x)), which maps a real number to (0,1) interval, used for binary classification.

At the same time, a loss function is introduced to avoid overfitting and improve the generalization ability of the model. In the reverse transmission, the classic gradient descent optimization algorithm is used to adjust the output weight and the weight of the SOM, obtain the relevant weight derivative for the loss function, and update the weight value along the gradient descent direction. In the present invention, two outputs are set, which is different from the single label value setting the threshold value to determine the input category label, and the sample is divided by the adjustment threshold value of the two output states. Instead of taking the maximization of the two output values to divide the samples into 0 and 1, different thresholds are used for different categories of data to maximize the Matthews correlation coefficient in the results.

Claims (6)

1. A microRNA prediction method based on a supervised self-organizing mapping neural network is characterized by comprising the following steps:

step 1: extracting characteristics, namely extracting the characteristics based on the microRNA sequence, obtaining secondary structure base pairing information of the microRNA sequence to be detected by using an RNAfold program, and converting the secondary structure base pairing information into vectorization expression by combining the characteristics based on the primary sequence and the characteristics based on the secondary structure to obtain a characteristic vector finally used for calculating a prediction means;

step 2: receiving characteristic data of an input layer, learning a spatial distribution rule of the input data by self-organizing mapping of a hidden layer, and performing mapping processing between data to form new characteristic representation in the process;

and step 3: the output supervision layer calculates output categories and related errors by using the new feature representation generated in the step 2, and reversely transmits the errors to update the network weight;

and 4, step 4: and setting a threshold value by using a cross validation method, and obtaining a classification result.

2. The microRNA prediction method based on the supervised self-organizing map neural network of claim 1, wherein: in the step 1, for a microRNA sequence with the sequence length of n, the sequence-based characteristics of the microRNA sequence are extracted, and the microRNA sequence length, GC content, composition and coding position deviation characteristics are extracted.

3. The supervised self-organizing map neural network-based microRNA prediction method as recited in claim 1, wherein: in the step 1, for a microRNA sequence, extracting the characteristics related to the structure of the microRNA sequence; extracting the base pairing information of the secondary structure of the sequence by using an RNAfold program, and obtaining the pairing condition of each base in the microRNA sequence based on the minimum free energy principle of the secondary structure at the default temperature; according to the base pairing information, the position composition conditions of any 3 adjacent nucleotides are calculated in a statistical manner, and 32 triplet structure composition characteristics are extracted; extracting the ratio of characteristic AU pairs to n, the ratio of GC pairs to n and the ratio of CU pairs to n according to the ratio of base pairs to sequence length in the secondary structure; and extracting the thermodynamically related characteristics of the microRNA according to the minimum free energy in the secondary structure.

4. The supervised self-organizing map neural network-based microRNA prediction method as recited in claim 1, wherein: in the step 2, the extracted microRNA sequence feature vector is used as an input, an input spatial distribution rule is learned in a self-organizing mapping neural network layer and is fully interconnected with an input node, each neuron in the self-organizing mapping is represented by a group of weight vectors, the weight vectors of an optimal matching unit and adjacent neurons thereof are continuously adjusted by searching for the optimal matching unit, topological structure information is kept in a low-dimensional output space, a new feature representation is obtained, and the self-organizing mapping output node defaults to be a 10 x 10 grid.

5. The microRNA prediction method based on the supervised self-organizing map neural network of claim 1, wherein: in the step 3, two neurons of the output layer are fully connected with the self-organizing mapping layer, new feature representation information output by the self-organizing mapping layer is used, the connection weight of interlayer input and output is utilized to calculate a sample label value, an error value of a predicted sample label and a real label of the predicted sample label is reversely transmitted to the network, the weight of the neural network is updated by using a gradient descent method, and the process is iterated for multiple times.

6. The microRNA prediction method of the supervised self-organizing map neural network of claim 1, wherein: in the step 4, a threshold segmentation method is used to judge whether the sequence belongs to the micro RNA, so that the Mauss correlation coefficient of the prediction result is maximized, and the classification result is obtained.

Images