RNA secondary structure prediction method, device and system, and storage medium

Abstract

The invention discloses an RNA secondary structure prediction method, device and system, and a storage medium. The RNA secondary structure is predicted based on the combination of a convolutional neural network model and a heuristic algorithm. Wherein the convolutional network part adopts AlexNet as a preposed basic network, the subsequent network uses a forward feature matrix corresponding to deconvolution to carry out superposition output, 10 comparatively typical physicochemical properties in an RNA structure are selected as research objects in the training process, input data is an L * L * 10 matrix, output is an L * L * 1 matrix, the numerical value is [0, 1], and the numerical value is [0, 1]. 1]; and calculating a stem region pool based on the pairing probability matrix, then searching the stem region pool by using a heuristic algorithm, and predicting the RNA secondary structure according to the type of RNA or a minimum free energy principle.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and particularly relates to a method, apparatus, system, and storage medium for predicting RNA secondary structure. Background Technology

[0002] RNA plays many crucial roles in basic cellular regulation and gene expression. It is a single-chain biopolymer composed of basic nucleotide units, with bases of adenine (A), cytosine (C), guanine (G), and uracil (U). Due to local hydrogen bonding interactions between bases, it can have more complex structures. Related studies have shown that non-coding RNA (ncRNA) plays important roles in chromosome replication, transcriptional regulation, and RNA processing in cells. ncRNA also plays a vital role in clinical applications in diagnosis, prognosis, vaccines, and treatment. The secondary structure of RNA refers to the folding of the primary structure and the interaction of hydrogen bonds between base pairs. The classic Watson-Crick rule includes the interaction between AU and GC base pairs; however, wobble pairs (GU base pairs) also frequently appear in RNA secondary structure pairing.

[0003] Predicting RNA secondary structure essentially involves predicting the stem and loop regions formed by the folding of RNA's primary structure. In 1981, Zuker et al. proposed a thermodynamic energy model to optimize the folding of large RNA sequences, aiding in the prediction of RNA secondary structure. In 1984, they proposed a dynamic programming-based method to find the optimal secondary structure by minimizing free energy. Later, Turner's nearest neighbor model calculated the free energy of each nearest neighbor loop by summing the free energy parameters characterizing the loops. This model, with adjustments to certain energy parameters, led to numerous methods such as UNAFold, RNAstructure, and Vienna RNAfold.

[0004] Recently, researchers have also developed machine learning-based methods for predicting RNA secondary structure. These methods train alternative parameters for thermodynamic parameters by using a large number of RNA sequence pairs and their reference secondary structures as training data. The MXFold2 method combines RNA folding fractions learned by a deep neural network with Turner's nearest neighbor free energy parameter, and trains the model with thermodynamic regularization to ensure that the folding fraction and the calculated free energy are as close as possible, thus improving the robustness of the prediction. Zhang et al. proposed an RNA secondary structure prediction algorithm, ContextFold, based on a convolutional neural network model and dynamic programming. By analyzing current experimental RNA sequence and structural data, a deep convolutional network model is constructed, and then implicit features for effective classification are extracted from large-scale data to predict the pairing probability of each base in the RNA sequence. Based on the obtained RNA sequence base pairing probabilities, an enhanced dynamic programming method is used to obtain the optimal RNA secondary structure. SPOT-RNA [ CONTRAfold and Ufold have been developed for predicting RNA secondary structure using deep neural networks (DNNs). SPOT-RNA uses multiple deep context learning models combined with Long Short-Term Memory (LSTM) to predict the base pairing probability of RNA structure, achieving a significant improvement in prediction performance. UFold uses the U-Net model to capture contextual information in the sequence, thereby improving the accuracy of RNA secondary structure prediction. CDPfold uses a convolutional neural network (CNN) to predict the probability of each base pairing in a sequence and uses dynamic programming to find the optimal secondary structure after a successful prediction, achieving some success. These algorithms all formulate RNA secondary structure prediction as multiple binary classification problems, using deep neural networks trained with a large amount of training data to predict whether each nucleotide pair forms a base pair. Although machine learning-based models have achieved extremely high performance in terms of prediction accuracy, they are subject to the risk of overfitting. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, apparatus, system, and storage medium for predicting RNA secondary structure.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting RNA secondary structure, comprising:

[0008] Step S1: Obtain the RNA dataset;

[0009] Step S2: Based on the RNA dataset, predict the secondary structure of RNA using a convolutional neural network model and heuristic algorithm.

[0010] Preferably, in step S2, the convolutional network part of the convolutional neural network model uses AlexNet as the front-end base network. In terms of usage, deconvolution is followed by the corresponding positive feature matrix for superposition and output. During training, 10 physicochemical properties in the RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix with base pairing probability matrix values ​​between [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the RNA secondary structure according to the RNA type or the principle of minimum free energy.

[0011] As a preferred approach, 10 physicochemical properties were selected as the research objects: PC1: rise; PC2: roll; PC3: shift; PC4: slide; PC5: tilt; PC6: twist; PC7: enthalpy; PC8: entropy; PC9: stack energy; and PC10: free energy.

[0012] The present invention also provides an RNA secondary structure prediction device, comprising:

[0013] The first processing module is used to acquire the RNA dataset;

[0014] The second processing module is used to predict the secondary structure of RNA based on the RNA dataset using a convolutional neural network model and heuristic algorithms.

[0015] As a preferred approach, the convolutional network part of the convolutional neural network model uses AlexNet as the front-end base network. In terms of operation, deconvolution is followed by the corresponding positive feature matrix for superposition and output. During training, 10 physicochemical properties of RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix with base pairing probability matrix values ​​between [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the RNA secondary structure according to the RNA type or the principle of minimum free energy.

[0016] As a preferred approach, 10 physicochemical properties were selected as the research objects: PC1: rise; PC2: roll; PC3: shift; PC4: slide; PC5: tilt; PC6: twist; PC7: enthalpy; PC8: entropy; PC9: stack energy; and PC10: free energy.

[0017] The present invention also provides an RNA secondary structure prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs an RNA secondary structure prediction method when executed by the processor.

[0018] The present invention also provides a storage medium storing a computer program that executes an RNA secondary structure prediction method when running.

[0019] To further improve the accuracy of RNA structure prediction, this invention proposes a method combining a convolutional neural network model and a heuristic algorithm for predicting RNA secondary structure. The convolutional network uses AlexNet as the initial base network, followed by deconvolution operations and superposition with the corresponding positive feature matrix for output. During training, ten typical physicochemical properties of RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix containing base pairing probability matrices with values ​​between [0, 1]. Based on the pairing probability matrix, a stem pool is calculated, and then a heuristic algorithm is used to search the stem pool. The RNA secondary structure is predicted based on the RNA type or the principle of minimum free energy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This is a flowchart of the RNA secondary structure prediction method according to an embodiment of the present invention;

[0022] Figure 2 It is a convolutional neural network model;

[0023] Figure 3 A graph of the loss function;

[0024] Figure 4 For stem region pool;

[0025] Figure 5 This is an incompatible situation;

[0026] Figure 6 For compatible cases;

[0027] Figure 7 This is the first compatible case;

[0028] Figure 8 This is the third compatible scenario;

[0029] Figure 9 This is the fourth compatible scenario;

[0030] Figure 10 This is part of the No. 1 stem area pond;

[0031] Figure 11 This refers to part of the No. 2 stem area pond;

[0032] Figure 12 This is a partial rhizome area pond;

[0033] Figure 13 This is part of the left-side frame stem area pool;

[0034] Figure 14 This is part of the right-side frame stem area pool;

[0035] Figure 15 This refers to a portion of the left stem region pool;

[0036] Figure 16 This refers to a portion of the right stem region pool;

[0037] Figure 17 This is a partial stem region pool;

[0038] Figure 18 This is a partial prediction result;

[0039] Figure 19 For replacement operation;

[0040] Figure 20 This is a mutation operation;

[0041] Figure 21 This is the overall prediction result;

[0042] Figure 22 For comparison of the average number of stem regions. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Example 1:

[0046] like Figure 1 As shown, this embodiment of the invention provides a method for predicting RNA secondary structure, including:

[0047] Step S1: Obtain the RNA dataset;

[0048] Step S2: Based on the RNA dataset, predict the secondary structure of RNA using a convolutional neural network model and heuristic algorithm.

[0049] In one embodiment of the present invention, in step S1, the dataset consists of three datasets: TrainSetA, TestSetA, and TestSetB, created by Rivas and 15 others. TrainSetA is used as the training set, containing 3166 structured data entries, while TestSetA and TestSetB are used as the test sets.

[0050] Training set data selection:

[0051] (1) Delete data in the training set where the number of unpaired bases in the hairpin loop of the RNA secondary structure is less than 3.

[0052] (2) Since RNA structures with less than 32 bases have low structural complexity, sequences with less than 32 bases suffer severe feature loss during training. Therefore, data with RNA sequences less than 32 bases in length are deleted.

[0053] 2177 records remain.

[0054] Test set data filtering:

[0055] (1) Delete data from the test set in which the number of unpaired bases in the hairpin loop of the RNA secondary structure is less than 3.

[0056] (2) Delete base pairings that do not conform to the Watson-Crick rule and open base pairs with a stem region of 1.

[0057] (3) To fill in RNA sequence lengths less than 32, fill the rightmost column of the matrix and the bottom row of the matrix, padding all the filled parts with 0 until the length is 32.

[0058] There are 574 test data entries remaining.

[0059] The input data was based on a method published by Zi Liu et al. in 2016: pRNAm-PC, which predicts N6-methyladenosine sites in RNA sequences based on physicochemical properties. The pRNAm-PC method selected 10 typical physicochemical properties as research objects: PC1: rise; PC2: roll; PC3: shift; PC4: slide; PC5: tilt; PC6: twist; PC7: enthalpy; PC8: entropy; PC9: stack energy; PC10: free energy. Table 1 shows the raw values ​​of the 10 physicochemical properties for 16 different dimers.

[0060] Table 1

[0061] Code Diner PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 1 AA 3.18 7 0.08 -1.27 -0.8 31 -6.82 -18.4 -13.7 -0.9 2 AC 3.24 4.8 0.23 -1.43 0.8 32 -11.4 -26.2 -13.8 -2.1 3 AG 3.3 8.5 -0.04 -1.5 0.5 30 -10.48 -19.2 -14 -1.7 4 AU 3.24 7.1 -0.06 -1.36 1.1 33 -9.38 -15.5 -15.4 -0.9 5 CA 3.09 9.9 0.11 -1.46 1 31 -10.44 -27.8 -14.4 -1.8 6 CC 3.32 8.7 -0.01 -1.78 0.3 32 -13.39 -29.7 -11.1 -2.9 7 CG 3.3 12.1 0.3 -1.89 0.1 27 -10.64 -19.4 -15.6 -2 8 CU 3.3 8.5 -0.04 -1.5 0.5 30 -10.48 -19.2 -14 -1.7 9 GA 3.38 9.4 0.07 -1.7 1.3 32 -12.44 -35.5 -14.2 -2.3 10 GC 3.22 6.1 0.07 -1.39 0 35 -14.88 -34.9 -16.9 -3.4 11 GG 3.32 12.1 -0.01 -1.78 0.3 32 -13.39 -29.7 -11.1 -2.9 12 GU 3.24 4.8 0.23 -1.43 0.8 32 -11.4 -26.2 -13.8 -2.1 13 UA 3.26 10.7 -0.02 -1.45 -0.2 32 -7.69 -22.6 -16 -1.1 14 UC 3.38 9.4 0.07 -1.7 1.3 32 -12.44 -26.2 -14.2 -2.1 15 UG 3.09 9.9 0.11 -1.46 1 31 -10.44 -19.2 -14.4 -1.7 16 UU 3.18 7 -0.08 -1.27 -0.8 31 -6.82 -18.4 -13.7 -0.9

[0062] The original values ​​of various physicochemical properties vary considerably. If used directly, the smaller values ​​may be underweighted, resulting in the loss of certain characteristics. Therefore, standardization was performed. Table 2 shows the results after standardization.

[0063] Table 2

[0064]

[0065]

[0066] After standardization, the data is incremented by 10 so that every number in the matrix is ​​greater than 0. The physicochemical properties of each column can be used as a matrix, with the rows of the matrix being RNA sequences and the columns being RNA sequences. The matrix is ​​filled according to the base matching, and a total of 10 matrices are constructed as input data.

[0067] The expected output of the data is a pairing probability matrix. The rows and columns of the matrix are the dotted bracket structure of the actual RNA secondary structure. Since all numbers in the input matrix are not 0, and the number of paired bases in the expected output matrix is ​​relatively small, the expected output matrix has a pairing probability of 0.01 and an unpaired probability of 0.99.

[0068] For example, the existing sequence: UUACAUGGCAACAGCUGGUGA, has the following actual structure: ((((.((((....)))))))). Taking the physicochemical properties of the first column as an example, Table 3 shows the first matrix of the input data for constructing the data.

[0069] Table 3

[0070] U U A C A U G G C A A C A G C U G G U G A U 9.17 9.17 10.09 11.47 10.1 9.17 8.13 8.13 11.47 10.09 10.1 11.47 10.09 8.13 11.47 9.17 8.13 8.13 9.17 8.13 10.09 U 9.17 9.17 10.09 11.47 10.1 9.17 8.13 8.13 11.47 10.09 10.1 11.47 10.09 8.13 11.47 9.17 8.13 8.13 9.17 8.13 10.09 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17 C 10.55 10.55 8.13 10.78 8.13 10.6 10.55 10.6 10.78 8.13 8.13 10.78 8.13 10.55 10.78 10.6 10.55 10.55 10.55 10.55 8.13 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17 U 9.17 9.17 10.09 11.47 10.1 9.17 8.13 8.13 11.47 10.09 10.1 11.47 10.09 8.13 11.47 9.17 8.13 8.13 9.17 8.13 10.09 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 C 10.55 10.55 8.13 10.78 8.13 10.6 10.55 10.6 10.78 8.13 8.13 10.78 8.13 10.55 10.78 10.6 10.55 10.55 10.55 10.55 8.13 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17 C 10.55 10.55 8.13 10.78 8.13 10.6 10.55 10.6 10.78 8.13 8.13 10.78 8.13 10.55 10.78 10.6 10.55 10.55 10.55 10.55 8.13 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 C 10.55 10.55 8.13 10.78 8.13 10.6 10.55 10.6 10.78 8.13 8.13 10.78 8.13 10.55 10.78 10.6 10.55 10.55 10.55 10.55 8.13 U 9.17 9.17 10.09 11.47 10.1 9.17 8.13 8.13 11.47 10.09 10.1 11.47 10.09 8.13 11.47 9.17 8.13 8.13 9.17 8.13 10.09 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 U 9.17 9.17 10.09 11.47 10.1 9.17 8.13 8.13 11.47 10.09 10.1 11.47 10.09 8.13 11.47 9.17 8.13 8.13 9.17 8.13 10.09 G 9.86 9.86 11.47 10.37 11.5 9.86 10.78 10.8 10.37 11.47 11.5 10.37 11.47 10.78 10.37 9.86 10.78 10.78 9.86 10.78 11.47 A 9.86 9.86 9.17 9.86 9.17 9.86 10.55 10.6 9.86 9.17 9.17 9.86 9.17 10.55 9.86 9.86 10.55 10.55 9.86 10.55 9.17

[0071] The remaining nine matrices are processed in a similar way to this matrix.

[0072] Table 4 shows the expected output matrix representation of the data. The actual expected output data is obtained by removing the first row and first column of the table below. This table is for reference only.

[0073] Table 4

[0074]

[0075]

[0076] As one embodiment of the present invention, in step S2... Figure 2 The constructed convolutional neural network model consists of four convolutional modules in the first four modules, three upsampling modules in the middle three modules, and a convolutional module in the last module.

[0077] The shape of the input data in the input layer is L*L*10.

[0078] The first module contains one convolutional layer, one normalization layer, and one max pooling layer. The convolutional layer has 96 kernels (3x3), a stride of 1, and padding of 1. After convolution, normalization is performed to aid model convergence. Following normalization, the model passes through the ReLU activation function, and finally, it undergoes pooling to output the final result. The pooling layer has a size of 2x2 and a stride of 2. The output shape of the first module is (L / 2)*(L / 2)*96. If both the length and width of the output matrix are odd numbers, this layer is recorded.

[0079] The second module contains one convolutional layer, one normalization layer, and one max pooling layer. The convolutional layer has 192 3x3 kernels, a stride of 1, and padding of 1. After convolution, normalization is performed, followed by ReLU activation. Finally, the result is pooled (2x2) with a stride of 2. The output shape of the second module is (L / 4)*(L / 4)*192. If both the length and width of the output matrix are odd numbers, this layer is recorded.

[0080] The third module contains three convolutional layers, three normalization layers, and one max pooling layer. In the first convolutional layer, the kernel size is 3x3, the number of kernels is 384, the stride is 1, and the padding is 1. After convolution, normalization, and the ReLU activation function, the output shape is (L / 4)*(L / 4)*384. The second convolutional layer follows the same process as the first, with an output shape of (L / 4)*(L / 4)*384. The third convolutional layer follows the same process as the first, with an output shape of (L / 4)*(L / 4)*384. Finally, a pooling operation is performed to output the final result. The pooling layer has a size of 2x2, a stride of 2, and an output shape of (L / 8)*(L / 8)*384. If both the length and width of the output matrix are odd numbers, this layer is recorded.

[0081] The fourth module contains one convolutional layer and one normalization layer. The convolutional layer has 384 kernels of 1x1 size and a stride of 1. After the convolution operation, the result is normalized and activated by the ReLU function before being output.

[0082] The fifth module has an input shape of (L / 8)*(L / 8)*384. This module contains one upsampling layer, one convolutional layer, one normalization layer, one padding layer, and one convolutional stacking layer. The data first passes through the upsampling layer, where bilinear interpolation is used to enlarge the matrix, resulting in an output shape of (L / 4)*(L / 4)*384. In the convolutional layer, the kernel size is 3*3, the number of kernels is 192, the stride is 1, and the padding is 1, resulting in an output shape of (L / 4)*(L / 4)*192. The output is then processed by convolution, normalization, and the ReLU activation function. If the convolutional output shape of the third module is odd, the output needs to be expanded to ensure proper stacking of the convolutional stacking layers. This expansion is done by adding one layer to the rightmost column and one layer to the bottom row of the matrix. If the shape is even, no expansion is required. After the padding layer, a convolution stacking operation is performed, adding the convolution output of the second module to the padding output of this module to supplement details. The final output shape is (L / 4)*(L / 4)*192.

[0083] The sixth module contains one upsampling layer, one convolutional layer, one normalization layer, and one padding layer. The input data passes through the upsampling layer, where bilinear interpolation is used to enlarge the matrix, resulting in an output shape of (L / 2)*(L / 2)*192. In the convolutional layer, the convolution sum is 3*3, with 96 convolutions, a stride of 1, and padding of 1, resulting in an output shape of (L / 2)*(L / 2)*96. The output is then processed by convolution, normalization, and the ReLU activation function. If the convolution output from the second module is an odd number, the output needs to be padded, with one layer added to the rightmost column and one layer to the bottom row of the matrix. If it is an even number, no padding is required. The output shape is (L / 2)*(L / 2)*96.

[0084] The seventh module contains one upsampling layer, one convolutional layer, one normalization layer, and one padding layer. The input data passes through the upsampling layer, resulting in an output shape of L*L*96. In the reconvolutional layer, the kernel size is 3*3, there is one kernel, the stride is 1, and the padding is 1, resulting in an output shape of (L)*(L)*1. The output is then processed by convolution, normalization, and the ReLU activation function. If the output of the first module is an odd number, it needs to be padded by adding one layer to the bottom row of the rightmost column of the kernel to restore it to its original size. If the output is an even number, no padding is required. The final output shape is L*L*1.

[0085] The eighth module contains one convolutional layer and one normalization layer. The convolutional layer has one 1x1 kernel, a stride of 1, and zero padding. The output is L*L*1. Using a 1x1 convolutional kernel improves the network's expressive power and feature extraction performance. The output is L*L*1 after convolution, normalization, and the ReLU activation function.

[0086] Finally, the sigmoid activation function restricts each number in the matrix to between 0 and 1.

[0087] Training process

[0088] After the convolutional neural network model is built, it is first trained using the training set. All data undergoes a forward propagation training to obtain predicted values. Then, the loss value is calculated using the mean squared error as the loss function. Backpropagation is then performed based on the loss value to update the weights. This completes one training cycle, which is repeated for a total of 20 rounds. At the beginning of training, weight initialization is performed to ensure that the neural network can effectively transmit gradient signals and avoid abnormal activation value distributions, thereby improving the model's convergence speed and training stability. After training, the model's weights are saved for calculating the pairing probability matrix on the test set data.

[0089] Equation (1) Mean Squared Error Loss Function (MSE):

[0090]

[0091] Where y i For predicted values, is the actual value, and n is the amount of data.

[0092] Figure 3 The graph represents the loss function, with the horizontal axis representing the number of training iterations and the vertical axis representing the corresponding loss value.

[0093] Calculation of the pairing probability matrix

[0094] The test set is used for testing. Any RNA sequence in the test set is input, and the input to the model is determined based on the physicochemical properties of base pairing. Based on the weights trained by the model, a forward propagation is performed to calculate the base pairing probability matrix of the RNA sequence.

[0095] For example, given an RNA sequence GGCGAAGUCGAAAGAUGGCGCC with a length of 22 bases, since this sequence is less than 32 bases long, it needs to be padded into a 32*32 matrix. The input form of the processed matrix is ​​shown in the table below. Table 5 only shows the physicochemical properties of the first column; the processing method for the physicochemical properties of the remaining columns is similar.

[0096] Table 5

[0097]

[0098] After forward propagation, a 32x32 matrix is ​​calculated. Since the input data padded a 22x22 matrix to a 32x32 matrix, the original matrix size needs to be determined. Because the padded matrix is ​​located in the rightmost 10 columns and the bottom 10 rows, the top-left 22x22 matrix is ​​chosen as the final RNA sequence base pairing probability matrix. Table 6 shows the pairing probability matrix generated after forward propagation of the RNA sequence.

[0099] Table 6

[0100]

[0101] Constructing stem zone pools

[0102] The following rules are followed when predicting base pairings based on the pairing probability matrix:

[0103] (1) Starting from the top left corner, traverse each element of the matrix. When the value of an element is less than 0.9855, first check if the absolute value of the difference between the x-coordinate and y-coordinate is greater than 2. If it is less than or equal to 2, change the value of this element in the matrix to 0 and continue traversing; if it is greater than 2, check if base pairing occurs (only select 6 cases: A and U, U and A, C and G, G and C, G and U, U and G): if pairing fails, change the value of this element in the matrix to 0 and continue traversing; if pairing succeeds, change the value of this element in the matrix to 1.

[0104] (2) After changing all elements of the pairing probability matrix to 0 and 1, traverse the matrix again. For the upper triangular part, when traversing to the number 1, traverse along the upper right direction from this position and count the number of 1s until traversing to 0. When the number of 1s is greater than or equal to 3, generate the current stem region structure and add it to the stem region pool. The lower triangular part is processed in a similar way to the upper triangular part, the only difference being that it needs to traverse along the current 1 element to the lower left corner of the matrix.

[0105] (3) For sequences with a length of less than or equal to 32, select a structure with a number of consecutive 1s greater than or equal to 2 as a stem region and add it to the stem region pool; if the beginning and end are paired, when two consecutive pairs are detected, continue to check whether there are still base pairs that meet the Watson-Crick pairing requirements along the lower left or upper right. If the pairing requirements are met, change the value of the matrix to 1. At this time, it is not necessary to consider whether the value of the element is less than 0.9855.

[0106] (4) For RNA with more than 32 bases and less than or equal to 90 bases, select elements with a matrix element value less than or equal to 0.9855. For RNA with more than 90 bases and less than or equal to 300 bases, select elements with a matrix element value less than or equal to 0.9885, and select two or more consecutive paired hairpin loops to add to the stem region pool. For RNA with more than 300 bases and less than or equal to 500 bases, select elements with a matrix element value less than or equal to 0.989, and select two or more consecutive paired hairpin loops to add to the stem region pool.

[0107] (5) For sequences longer than 32 and shorter than 100, for the upper triangular portion, when a short stem region containing two or more consecutive paired bases is found, the next two bases can be checked along the upper right to see if they meet the Watson-Crick pairing requirement. If they do, the value of the current matrix element is set to 1. Then, the same requirement is followed along the lower left to check for paired bases until an unpairable base is encountered. During this process, all matrix elements corresponding to paired bases are set to 1. In these cases, it is not necessary to consider whether the element value is less than 0.9855. The lower triangular portion is handled similarly to the upper triangular portion, except that it is necessary to traverse along this element towards the lower left corner of the matrix.

[0108] (6) For sequences with a length greater than 65 and less than or equal to 90, the method for generating stem regions is also different. Traverse the matrix. For the upper triangular part, if the current element has a value of 1, traverse along the upper right. When the last 1 is reached, calculate the length of the stem region. If the stem region meets the basic stem region conditions and the stem region length is short, add the stem region to the short stem region pool (the short stem region pool is for stem regions with a hairpin loop length less than or equal to 24 bases, and there is only a short stem region pool in the range of 65-90bp). For this stem region, change the outermost paired bases from brackets to dots each time and add them to the short stem region pool. The generation of a new stem region must ensure that the number of consecutive base pairs is greater than or equal to 3. When the stem region length is long, the processing method is the same as in 4. Add the stem region to the long stem region pool (the long stem region pool is for stem regions with a hairpin loop length greater than 24 bases, and there is only a long stem region pool in the range of 65-90bp). The difference between the processing method of the lower triangular part and the upper triangular part is that the traversal method is reversed.

[0109] The stem regions in the established stem region pool are sorted from top to bottom according to the number of consecutively paired bases. Within the same number of base pairs, they are sorted from top to bottom according to the number of unpaired bases in the hairpin loop, in order to facilitate subsequent stem region compatibility operations. Figure 4 This is the stem pool that can be generated for the pairing probability matrix in the previous section.

[0110] Stem region compatibility and specific RNA prediction

[0111] When performing stem region compatibility operations, different operations were performed for sequences of different lengths.

[0112] (1) For sequences with a length of 32 or less, select as many stem regions as possible at both ends as root-stem regions and perform compatibility operations sequentially.

[0113] For short sequences of length 32 or less, prioritize pairing as many stem regions as possible at both ends as possible as root-stem regions, and then sequentially make the stem regions compatible. The specific method for making stem regions compatible is to change the left and right parentheses to 1 and the dots to 0, which will result in a sequence containing only 0 and 1.

[0114] For example, ((((..............)))) can be processed as 1111000000000000001111, and ......((((.....))))... can be processed as 0000001111000001111000.

[0115] During the compatibility process, a bitwise "binary AND" operation is performed on the two 0-1 sequences. If the result is 1, the stem region is incompatible. Figure 5 This is a case where they are incompatible. If the result is all 0, proceed to the next step and perform a binary join between the first and second sequences, then save the result. Figure 6 This represents the compatible case. Equation (2) shows the compatibility result of the stem region, where R1 is the first sequence and R2 is the second sequence.

[0116]

[0117] Save the start and end positions of the stem regions of the two compatible sequences. Compare the start position of the stem region of the second sequence with the start and end positions of all already compatible sequences. If any of the following conditions are met, a stem region compatibility operation can be performed, merging the first and second sequences into the first sequence:

[0118] If the starting position of the stem region of the second sequence is less than the starting position of the stem region of the first sequence, and the ending position of the stem region of the second sequence is greater than the ending position of the stem region of the first sequence, Figure 7 The starting position of the second sequence is 8, which is greater than the starting position of the first sequence, and the ending position of the second sequence is 32, which is less than the ending position of the first sequence. Therefore, a compatibility operation can be performed.

[0119] b. If the starting position of the stem region of the second sequence is greater than the starting position of the stem region of the first sequence, and the ending position of the stem region of the second sequence is less than the ending position of the stem region of the first sequence, the compatibility case can be referred to... Figure 7 .

[0120] c. The start position of the stem region in the second sequence is greater than the end position of the stem region in the first sequence. Figure 8In addition to satisfying condition 1, the start and end positions of the second sequence are compatible with the first sequence. Furthermore, the start position 58 of the second sequence is greater than the end position 32 of a stem region in the first sequence, so a compatibility operation can be performed.

[0121] The end position of the stem region in the second sequence is less than the start position of the stem region in the first sequence. Figure 9 In addition to satisfying conditions 1 and 2, the start and end positions of the second sequence are compatible with the first sequence. Furthermore, the end position 55 of the second sequence is less than the start position 58 of a stem region in a sequence. Therefore, a compatibility operation can be performed.

[0122]

[0123] Equation (3) represents four compatible cases. In the first sequence, the stem region starts at Start1 and ends at End1. In the second sequence, the stem region starts at Start2 and ends at End2.

[0124] After comparing the first stem region of the first sequence, the comparison of the second stem region is performed, and so on. If all the above conditions are met, the two stem regions can be recombined. After recombination, the whole sequence is treated as a whole and then recombined with the next stem region for compatibility testing. All possible results can be obtained. For example, ((((..............)))) and ......(((....)))...... can finally be recombined as ((((..(((....)))..)))).

[0125] After the above operations, the following sequence can be obtained: (((((((((....)))))))) ((((.(((......))))))) (((((........)))))) ((((....((.....)))))) ((((.((....))....))))

[0131] Then, the free energy of each predicted sequence is calculated, and the sequence with the smallest free energy is selected as the final predicted structure. In the example above, the final selected sequence is ((((.((((....)))))))).

[0132] (2) For sequences with a length greater than or equal to 65 and less than or equal to 90, since most structures in this range are tRNAs, the tRNA is first determined. If six or more consecutive paired stem regions can be generated at both ends, the tRNA can be used as a template structure to construct the secondary structure. Specifically, stem regions with a length less than or equal to 24 are separated into stem region pool 1 (short stem region pool), and stem regions with a length greater than 24 are selected into stem region pool 2 (long stem region pool). Then, the first 20 from stem region pool 2 are selected as rootstock regions, and compatibility operations are performed with each stem region from stem region pool 1. When calculating the free energy, the free energy of the multi-branched loop needs to be considered. For each base appearing in the multi-branched loop, the free energy is increased by 0.2. Save the results with 2 or fewer internal stem regions and the results with more than 3 internal stem regions respectively. Select the secondary structure corresponding to the minimum free energy of the internal stem regions with more than 2 internal stem regions and the secondary structure corresponding to the minimum free energy of the internal stem regions with less than or equal to 2 internal stem regions. If the difference between the free energy of the selected secondary structure with more than 2 internal stem regions and the selected secondary structure with less than or equal to 2 internal stem regions is less than or equal to 4, then the secondary structure with more than 2 internal stem regions is used as the predicted secondary structure.

[0133] Equation (4) describes the process of dividing the stem region pool, where longRoot is the long stem region pool, shortRoot is the short stem region pool, L is the current stem region, and len(L) is the length of the stem region.

[0134]

[0135] Let LRoot be the structure corresponding to the minimum free energy in the stem pool where the internal stem region is greater than 2, SRoot be the structure corresponding to the minimum free energy in the stem pool where the internal stem region is less than or equal to 2, and finalPredict be the final prediction result, as shown in Equation (5).

[0136]

[0137] For example, consider the following 73bp RNA sequence: GGGGGUAUAGCUCAGUUGGUAGAGCGC UGCCUUUGCACGGCAGAUGUCAGGGGUUCGAGUCCCCUUACCUCCA. Based on the above rules, the following stem region pool can be generated. Figure 10 This is part of the No. 1 stem area pond. Figure 11 This is part of the No. 2 stem area pond.

[0138] The final secondary structure can be obtained as: (((((((..((((........)))).(((((.......))))).....(((((.......)))))))))))))).

[0139] (3) For sequences with a length greater than or equal to 111 and less than or equal to 125, RNA is mostly folded into 5S rRNA within this range. Therefore, we can first determine whether 5S rRNA can be generated based on the generated stem region. If the stem region pool can determine that 5S rRNA can be generated, then the secondary structure is generated using 5S rRNA as a template.

[0140] The specific characteristics of 5S rRNA are:

[0141] A ring can be formed if the number of paired bases at both ends of the ring is greater than or equal to 6 and the number of free bases in the hairpin ring is greater than or equal to 92.

[0142] b has two stem regions inside, and the left stem region is longer than the right stem region.

[0143] The distance from the left stem region to the innermost left-side paired base at both ends is 4, and the distance from the right stem region to the innermost right-side paired base at both ends is less than or equal to 2.

[0144] The left and right stem regions differ by one base.

[0145] The length of the left stem region is between 52 and 55 mm, and the length of the right stem region is between 30 and 45 mm.

[0146] The starting position of the left stem region is greater than or equal to 12, and the starting position of the right stem region is greater than or equal to 67.

[0147] The stem regions on the left and right sides of the stem region should maintain a symmetrical structure as much as possible.

[0148] If a stem region satisfying the first four characteristics can be found in the stem region pool, then 5S rRNA can be used as a template to construct the secondary structure.

[0149] Based on the above characteristics, the stem regions paired at both ends can be grouped into separate root-stem region pools and sorted according to the number of base pairs and stem region length. Stem regions with a starting position greater than or equal to 12 and a length between 52 and 55 are designated as the left frame stem region pool and sorted according to stem region length. Stem regions with a starting position greater than or equal to 67 and a length between 30 and 45 are designated as the right frame stem region pool and sorted according to stem region length. Stem regions with a starting position greater than or equal to 12, a length less than 52, and an ending position less than 67 are designated as the left stem region pool. Stem regions with a starting position greater than or equal to 67 and a length less than 30 are designated as the right stem region pool. The left and right stem region pools are then sorted according to the number of base pairs and stem region length.

[0150] After setting up the stem region pool, first determine the first stem region in the root-stem region pool. If the number of paired bases is greater than or equal to 6, it is directly designated as a root-stem region. If the number of paired bases is less than 6, the first stem region is sequentially matched with the following stem regions. When the number of paired bases is greater than or equal to 6, the stem region matching is stopped. If the number of paired bases is still less than 6 after the stem region matching operation, it can be determined that it is not 5S rRNA.

[0151] In the left frame stem region pool, select stem regions from bottom to top where the distance from the innermost left-side paired bases at both ends is 4. In the right frame stem region pool, select stem regions from top to bottom that differ from the left frame stem region by 1 base and whose distance from the innermost right-side paired bases at both ends is less than or equal to 2. Then, reconcile the rootstock region, left frame stem region, and right frame stem region.

[0152] After establishing the overall framework of the secondary structure of 5S rRNA, compatibility operations are performed with the left and right stem region pools. During the stem region compatibility operation, the stem region structures in the left and right stem regions should be kept as symmetrical as possible.

[0153] Equation (6) describes the compatibility process of the root-stem region, where the stem region in the root-stem region pool is rootStem, its internal start position is rtSS, and its internal end position is rtSE. The first stem region of the root-stem region is sequentially compatible with the remaining stem regions. This continues until the number of paired stem regions after fusion is greater than or equal to 6. During fusion, the overall internal start position (NrtSS) and the overall internal end position (NrtSE) are continuously updated.

[0154]

[0155] Equation (7) represents the compatibility process of the left frame stem region, where the stem region in the left frame stem region pool is leftRootStem, with an external start position of lRSS and an external end position of lRSE. The internal start position is lRSIS and the internal end position is lRSIE. If the left frame stem region can be found, the internal start position lRSIS, the internal end position lRSIE, and the external end position lRSE of this stem region are recorded. If it cannot be found, it is not 5S rRNA.

[0156]

[0157] Equation (8) represents the compatibility process of the right frame stem region, where the stem region in the right frame stem region pool is rightRootStem, with an external start position of rRSS and an external end position of rRSE. The internal start position is rRSIS and the internal end position is rRSIE. If the left frame stem region can be found, the internal start position rRSIS and the internal end position rRSIE of this stem region are recorded; otherwise, it is not a 5S rRNA.

[0158]

[0159] Equation (9) represents the compatibility process for the left stem region, where the stem region in the left stem region pool is lefStem, with an external start position of lSS and an external end position of lSE. The internal start position is lSIS and the internal end position is lSIE. Let lSS - lRSIS = L and lRSIE - lSIE = R. If this condition is met, the process is compatible, and lRSIS and lRSIE are updated.

[0160]

[0161] Equation (10) represents the compatibility process of the right stem region, where the stem region in the right stem region pool is rightStem, with an external start position of rSS and an external end position of rSE. The internal start position is rSIS and the internal end position is rSIE.

[0162] Let rSS - rRSIS = L, and rRSIE - rSE = R. If this condition is satisfied, then it is compatible, and rRSIS and rRSIE are updated.

[0163]

[0164] For example, given a 121bp RNA sequence: UAUUCUGGUGUCCUAGGCGUAGAGGAACCACACCAAUCCAUCCCGAACUUGGUGGUUAAACUCUACUGCGGUGACGAUACUGUAGGGGAGGUCCUGCGGAAAAAUAGCUCGACGCCAGAAU, the following can be generated according to the rules above: Figures 5-9 Stem region pool. Figure 12 This is a partial rhizome area pond. Figure 13 This is part of the left-side frame stem area pool. Figure 14 This is part of the right-side frame stem area pool. Figure 15 This is part of the left stem region pool. Figure 16 This is part of the right stem region pool.

[0165] The final prediction result of the secondary structure is: .(((((((((((....(((((((((((.((((((..(((((......))..) ...

[0166] (4) If an RNA with more than 32 bases and less than 140 bases is neither tRNA nor 5SrRNA according to the previous judgment, then normal prediction shall be performed.

[0167] If the length of the stem region pool is less than or equal to 20, then each stem region in the stem region pool is treated as a root-stem region and sequentially subjected to stem region compatibility operations with the remaining stem regions. If the length of the stem region pool is greater than 20, then the first 20 stem regions are selected as root-stem regions and sequentially subjected to stem region compatibility operations with the remaining stem regions.

[0168] b. Based on the compatibility results, the secondary structure with the minimum free energy is selected as the predicted secondary structure.

[0169] For example: There is an RNA sequence of length 46bp: UGUGGAGAACUUUUCUUGCAUUAAUUUGUAGGGAUUGGGACCCGCA.

[0170] Figure 17 For part of the stem area pool, Figure 18 This is a partial prediction result.

[0171] Selecting the prediction structure with the minimum free energy from all prediction results, the final prediction result is: (((((....(((((((((((......)))))))))..)))))))

[0172] Heuristic algorithms predict RNA secondary structure

[0173] For RNA sequences longer than 140 bases, due to their length and the large number of stem regions generated, selecting the first 20 stem regions as root-stem regions and sequentially performing stem region compatibility operations with the remaining stem regions fails to find the secondary structure corresponding to the minimum free energy. Heuristic algorithms can enumerate numerous stem region compatibility cases and quickly find the secondary structure corresponding to the minimum free energy. The specific process is as follows.

[0174] (1) Formation of stem pools

[0175] Special processing is performed when generating the stem region. Starting from the top left corner, each element of the matrix is ​​traversed. When an element's value is less than 0.9876 and satisfies the Watson-Crick pairing requirement, the traversal continues from that position to the upper right and lower left. If the Watson-Crick pairing requirement is met, the matrix value is changed to 1. If the pairing requirement is not met, or the absolute value of the x-coordinate minus the y-coordinate is less than or equal to 2, the traversal stops. After changing the pairing matrix to 0s and 1s, the matrix is ​​traversed again. For the upper triangular part, when a number 1 is encountered, the traversal continues from that position to the lower left, counting the number of 1s until a 0 is encountered. When the number of 1s is greater than or equal to 3, the current stem region structure is generated and added to the stem region pool. During the traversal, all traversed 1s are changed to 0.

[0176] The stem regions in the established stem region pool are sorted from top to bottom according to the number of consecutively paired bases. Within the same number of base pairs, they are sorted from top to bottom according to the number of unpaired bases in the hairpin loop, in order to facilitate subsequent stem region compatibility operations.

[0177] (2) Heuristic Algorithm

[0178] A copy of the stem region pool is created as a backup for subsequent stem region pool filling operations. The first stem region that has not been deleted is selected from the top down in the stem region pool as the root-stem region. A stem region compatibility operation is performed between the root-stem region and the remaining stem regions in the stem region pool to obtain a predicted structure, which is then saved. If any deleted stem regions exist, they are restored to their original positions. If the free energy of the currently predicted secondary structure is less than the currently predicted minimum free energy, it is recorded. Otherwise, the current structure is deleted, and the secondary structure corresponding to the minimum free energy is processed. If the iteration condition is met, iteration is performed. If the free energy of the secondary structure in the last iteration is less than the currently predicted minimum free energy, it is recorded; otherwise, it is deleted, and the secondary structure corresponding to the minimum free energy is processed. After further replacement and mutation operations, this round ends, for a total of 1000 rounds.

[0179] a Replacement Operation

[0180] For each compatible stem region, a random number is set. If the value of the random number is greater than 0.5, the corresponding stem region is deleted, and the corresponding deleted stem region in the spare stem region pool is set to 0. The deleted stem region is restored in the next replacement. Figure 19 This refers to the specific replacement operation.

[0181] b mutation operation

[0182] The secondary structure after deleting the stem region is mutated according to probability. If the mutation probability is greater than 0.98, the mutation operation is performed; otherwise, it is not performed. The mutation operation is achieved by randomly selecting a stem region and canceling either the innermost or outermost pairing within the stem region. When the number of unpaired bases in the stem region is less than or equal to 8, only the outermost pairing is canceled; otherwise, either the innermost or outermost pairing is canceled with a probability of 0.5. Figure 20 For specific mutation operations.

[0183] c Iteration operation

[0184] Fitness function:

[0185]

[0186] Equation (11) is the fitness function, where Len is the length of the sequence, k is the kth round, E1 is the minimum free energy of the second-order structure predicted up to the kth round, and E2 is the free energy of the second-order structure predicted in the kth round. If the above conditions are met, the second-order structure predicted in the kth round is used to enter the iteration process. The iteration process is consistent with the process of the heuristic algorithm, and a total of 100 rounds are iterated.

[0187] d Prediction effect

[0188] To prevent the algorithm from failing to converge, the above process is repeated 40 times. The second-order structure that appears most frequently in these 40 iterations is selected as the final predicted second-order structure, and then sorted in descending order of minimum free energy. If each structure appears almost the same number of times, or appears less than or equal to 3 times, then the structure with the lowest free energy among these 40 structures is selected as the current predicted structure.

[0189] For example, given an RNA sequence of length 164 bp: GCUUGUUCUCGGGGCGGGGUGAAACUCCCCACCGGCGGUAUGAAAAGCAAUUUUCA AGCCCGCGAGCGCCUGAAAUGGAAGCCGAUUCGCAUGCCAUUUCAGGGUCAGCAGA UCCGGUGAGAUGCCGGAGCCGACGGUUAAAGUCCGGAUGGAAGAGAGCGAAU, the following result can be generated according to the above rules.

[0190] Figure 21To determine the predicted result, the second-order structure with the most predictions was selected. The final predicted result is: ...(((((((((((((.......))))(((((((((((....)))))......(((((((((...............)))))))))).........((((((((...)))))))))))))).)))))........))))))).., with a free energy of -51.1.

[0191] This invention obtains a pairing probability matrix through a fully convolutional neural network and then uses this matrix to obtain a stem pooling pool. To minimize the number of stem regions in the pooling pool, this invention sets appropriate parameters for the pairing probability matrix to retain correct stem regions to the greatest extent possible while filtering out erroneous ones, thereby increasing the accuracy of the algorithm's predictions. Through extensive testing and analysis, this invention found that for sequences with lengths between 140 and 200, the pairing probability matrix parameter is set to 0.9876; for sequences with lengths between 200 and 300, the parameter is set to 0.9912. Since the number of stem regions with sequence lengths between 0 and 100 is small, they are not included in this statistical analysis. Figure 22 This is a comparison of the number of stem regions before and after some stem regions were deleted.

[0192] By comparing the average number of stem regions generated in each sequence with the number of stem regions calculated using the pairing probability matrix, it can be seen that the algorithm of this invention can significantly filter out a large number of stem regions with a sequence length between 100 and 200, showing a significant advantage. While some stem regions can be filtered out with a sequence length between 200 and 300, the number of stem regions filtered out is relatively small compared to the total number of stem regions, possibly due to significant feature loss in longer sequences caused by model training.

[0193] Predictive performance is primarily based on the number of correctly predicted base pairs (TP), the number of incorrectly predicted base pairs (FP), the number of correctly predicted free bases (TN), the number of incorrectly predicted free bases (FN), accuracy (ACC), sensitivity (SEN), and positive predictive value (PPV). In addition to these basic metrics, the harmonicity score F1 is also used for performance evaluation and serves as a primary performance metric. Sequences are divided into three length groups: less than 100 bp, 100–200 bp, and 200–300 bp. The prediction results from the ANfold algorithm of this invention are compared with those from the thermodynamic algorithms RNAstructure and Mfold. [ RNAfold, CONTRAfold[ CentroidH omfold

[24] and the convolutional neural network algorithm REDfold ] E2Efold [ Compare with CNNfold.

[0194]

[0195] Table 7

[0196]

[0197] Table 8

[0198]

[0199]

[0200] Table 9

[0201]

[0202] As shown in Tables 7, 8, and 9, the ANfold algorithm has a significant advantage in predicting RNA sequences within 200 bp, with prediction results significantly higher than other models and algorithms. When predicting sequences of 100 bp to 200 bp, its F1 score is slightly higher than other models and algorithms, but its accuracy is slightly lower than the REDfold model. When predicting sequences of 200 bp to 300 bp, it shows significant shortcomings compared to RNAstructure and Mfold, but in terms of F1 score, its prediction performance is almost identical to that of convolutional neural network algorithms.

[0203] The function of non-coding RNA is usually closely related to its structure, and accurate prediction of RNA secondary structure helps to reveal RNA function. Fold-based algorithms have stagnated in prediction performance, and machine learning-based models are limited in prediction accuracy by the amount of real RNA structure data. Therefore, this invention proposes a method for predicting RNA secondary structure based on a combination of convolutional neural network models and heuristic algorithms. The convolutional network uses AlexNet as the base network, and subsequent networks use deconvolution followed by the corresponding positive feature matrices for superposition and output. During training, 10 typical physicochemical properties of RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix, with base pairing probability matrices ranging from [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the RNA secondary structure according to the RNA type or the principle of minimum free energy. This method demonstrates a significant advantage in predicting RNA sequences within 200 bp, and its results are largely consistent with other machine learning methods in RNA sequences between 200 bp and 300 bp. It provides technical support and new ideas for further predicting RNA structure, inferring functional mechanisms, and designing novel experiments.

[0204] Example 2:

[0205] This invention also provides an RNA secondary structure prediction device, comprising:

[0206] The first processing module is used to acquire the RNA dataset;

[0207] The second processing module is used to predict the secondary structure of RNA based on the RNA dataset using a convolutional neural network model and heuristic algorithms.

[0208] As one embodiment of the present invention, the convolutional network part of the convolutional neural network model uses AlexNet as the front-end base network. In terms of usage, deconvolution is followed by the corresponding positive feature matrix for superposition and output. During training, 10 physicochemical properties in the RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix with base pairing probability matrix values ​​between [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the RNA secondary structure according to the RNA type or the principle of minimum free energy.

[0209] As one embodiment of the present invention, 10 physicochemical properties are studied as objects of study, namely PC1: rise; PC2: roll; PC3: shift; PC4: slide; PC5: tilt; PC6: twist; PC7: enthalpy; PC8: entropy; PC9: stack energy; PC10: free energy.

[0210] Example 3:

[0211] This invention also provides an RNA secondary structure prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs an RNA secondary structure prediction method when executed by the processor.

[0212] Example 4:

[0213] This invention also provides a storage medium storing a computer program that executes an RNA secondary structure prediction method during runtime.

[0214] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting RNA secondary structure, characterized in that, include: Step S1: Obtain the RNA dataset; Step S2: Based on the RNA dataset, predict the secondary structure of RNA using a convolutional neural network model and heuristic algorithm.

2. The RNA secondary structure prediction method as described in claim 1, characterized in that, In step S2, the convolutional network part of the convolutional neural network model uses AlexNet as the front-end base network. In terms of usage, deconvolution is followed by the corresponding positive feature matrix for superposition and output. During training, 10 physicochemical properties in the RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix with a base pairing probability matrix with a value between [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the secondary structure of RNA based on the type of RNA or the principle of minimum free energy.

3. The RNA secondary structure prediction method as described in claim 2, characterized in that, Ten physicochemical properties were used as the research objects: PC1: rising; PC2: rolling; PC3: displacement; PC4: sliding. PC5: Tilt; PC6: Twist; PC7: Enthalpy; PC8: Entropy; PC9: Stacked energy; PC10: Free energy.

4. An RNA secondary structure prediction device, characterized in that, include: The first processing module is used to acquire the RNA dataset; The second processing module is used to predict the secondary structure of RNA based on the RNA dataset using a convolutional neural network model and heuristic algorithms.

5. The RNA secondary structure prediction device as described in claim 4, characterized in that, The convolutional network part of the convolutional neural network model uses AlexNet as the front-end base network. In terms of usage, deconvolution is followed by the corresponding positive feature matrix for superposition and output. During training, 10 physicochemical properties in the RNA structure are selected as research objects. The input data is an L*L*10 matrix, and the output is an L*L*1 matrix, which is a base pairing probability matrix with values ​​between [0, 1]. The stem pool is calculated based on the pairing probability matrix, and then a heuristic algorithm is used to search the stem pool to predict the secondary structure of RNA based on the type of RNA or the principle of minimum free energy.

6. The RNA secondary structure prediction device as described in claim 5, characterized in that, Ten physicochemical properties were used as the research objects: PC1: rising; PC2: rolling; PC3: displacement; PC4: sliding. PC5: Tilt; PC6: Twist; PC7: Enthalpy; PC8: Entropy; PC9: Stacked energy; PC10: Free energy.

7. An RNA secondary structure prediction system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the RNA secondary structure prediction method as described in any one of claims 1-3 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the RNA secondary structure prediction method as described in any one of claims 1-3.

Images