Method and related device for detecting N4-acetylcytosine sites in mRNA

Through the dual neural network model of integer encoding and PseKNC encoding combined with self-attention mechanism, the high cost and low efficiency of the existing RNAac4C site detection method is solved, and higher recognition accuracy and efficiency are achieved, proving that DPNN- The superiority of the ac4C model in identifying ac4C loci.

CN118314956BActive Publication Date: 2024-11-08HAINAN UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202410509514.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-26
Publication Date
2024-11-08
Estimated Expiration
2044-04-26

AI Technical Summary

Technical Problem

The existing RNAac4C site detection methods are time-consuming, costly and limited in accuracy, and have limitations in feature extraction depth and efficiency.

Method used

Integer encoding and PseKNC encoding are used to convert mRNA sequences, and a dual-path neural network model based on self-attention mechanism is constructed. Combined with integer encoding and PseKNC feature vectors, sequence and local feature information are extracted through the first path network and the second path network. Fully connecting layer output of the classification module is performed to identify the N4-acetylcytosine site.

Benefits of technology

The recognition accuracy of ac4C sites was improved, and the ACC increased by 2.07% in the existing model comparison, showing better performance and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118314956B_ABST
    Figure CN118314956B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for detecting N4-acetylcytosine sites in mRNA and a related device. The method uses integer coding and PseKNC coding to convert mRNA sequences to obtain sequence integer coding and PseKNC feature vectors; constructs a two-way neural network model based on a self-attention mechanism, and the two-way neural network model based on the self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer coding to obtain a tensor V1; the second path network processes the PseKNC feature vector and outputs a tensor V2; the classification module is used to splice the tensor V1 with the tensor V2 and input them into a fully connected layer, and output a classification result. The present invention can effectively identify RNA ac4C sites and has low detection cost, high efficiency and accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics analysis, and in particular to a method for detecting N4-acetylcytosine sites in mRNA and related devices. Background Art

[0002] RNA ac4C modification is a highly conserved RNA modification in prokaryotes and eukaryotes. Ac4C modification causes acetylation of N4-acetylcytosine under the action of ac4C modification enzymes. Ac4C modification is not only present in eukaryotic tRNA and 18SrRNA, but also in mammalian mRNA. It improves translation efficiency, enhances mRNA stability, and regulates gene expression. Currently, NAT10 (N-acetyltransferase 10) is the only known eukaryotic ac4C modification enzyme. N4-acetylcytosine (ac4C) mRNA catalyzed by NAT10 (N-acetyltransferase 10) has been implicated in a variety of human diseases, especially cancer. Recognition of ac4C modification on mRNA is very important in disease detection, in-depth understanding of its translation mode, and early prevention.

[0003] Existing RNA ac4C modification detection methods rely on complex biotechnology methods, such as ac4C-seq, LC-MS, and acRIP-seq, which are time-consuming and costly. Although traditional machine models can sequence the ac4C sites of mRNA more conveniently and quickly, such as using two random forest classifiers, extreme gradient boosting classifiers, three encoding methods based on nucleotide composition, nucleotide chemical property, and accumulated nucleotide frequency, and mRMR feature selection combined with GBDT classifiers. However, most of them rely on manual feature extraction and traditional machine learning models, which have limitations in the depth and efficiency of feature extraction. Summary of the invention

[0004] In order to solve the problems of high cost, low efficiency and limited accuracy of RNAac4C site detection in the prior art, the present invention provides a method for detecting N4-acetylcytosine sites in mRNA and related devices.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] The method for detecting N4-acetylcytosine sites in mRNA comprises the following steps:

[0007] The mRNA sequence is converted using integer coding and PseKNC coding to obtain the sequence integer coding and PseKNC feature vector;

[0008] A two-path neural network model based on a self-attention mechanism is constructed, wherein the two-path neural network model based on a self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer encoding to obtain a tensor V 1 ; The second path network processes the PseKNC feature vector and outputs a tensor V 2 ; The classification module is used to convert the tensor V 1 With tensor V 2 Concatenate and input into the fully connected layer to output the classification result.

[0009] Preferably, the mRNA sequence is converted using integer coding to obtain the sequence integer coding, comprising the following steps:

[0010] Define a base dictionary in which each unique base symbol corresponds to a discrete integer value, that is, map A to 0, C to 1, G to 2, and U to 3 one by one to obtain the final sequence integer encoding.

[0011] Preferably, the mRNA sequence is converted using PseKNC encoding to obtain a PseKNC feature vector, comprising the following steps:

[0012] For a given RNA sequence S, which contains L bases, it is divided into non-overlapping subsequences of length k. For N types of k-mers, the number of times they appear in the sequence is counted, and a frequency distribution vector is constructed, which is recorded as:

[0013] {X 1 , X 2 , ..., X N}

[0014] Calculate the frequency of each k-mer in the sequence and get the frequency vector:

[0015]

[0016] Where P i is the frequency of the i-th k-mer, i∈{1,2,…,N}. Secondly, PseKNC improves the expressiveness of sequence features by introducing “pseudo-components”, which include two aspects of pseudo-components: one is the weight factor ω of each k-mer frequency. i We get, i∈{1,2,…,N}:

[0017]

[0018] The second is the correlation vector obtained by the correlation between different k-mers:

[0019]

[0020] The frequency vector, weight factor and correlation vector of KNC are combined to obtain the PseKNC feature vector:

[0021]

[0022] Preferably, the processing of the first path network comprises the following steps:

[0023] Linearly mapping the sequence integer code to a high-dimensional representation through an embedding layer to obtain a 512-dimensional high-dimensional tensor;

[0024] Concatenate the high-dimensional tensor with the position encoding matrix to obtain a concatenated matrix;

[0025] The concatenated matrix is ​​input into the bidirectional GRU network to extract contextual semantic information and calculate based on multi-head attention to obtain the tensor V 1 .

[0026] Preferably, the hidden layer of the bidirectional GRU network is 512-dimensional and has 3 layers.

[0027] Preferably, the processing of the second path network comprises the following steps:

[0028] A convolutional neural network is used to extract local feature information of the PseKNC feature vector, wherein the convolutional neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a third convolutional layer, a third activation layer, and a third maximum pooling layer connected in sequence, wherein the padding parameters of the first convolutional layer, the second convolutional layer, and the third convolutional layer are all 2, the convolution kernel size is 4, and the number of channels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are 16, 32, and 16, respectively;

[0029] Use multi-head attention to calculate the local feature information and output the tensor V 2 .

[0030] Preferably, the first activation layer, the second activation layer and the third activation layer are all ReLU activation functions.

[0031] Based on the above content, the present invention also discloses a system for detecting N4-acetylcytosine sites in mRNA, comprising:

[0032] A conversion module, used to convert the mRNA sequence using integer coding and PseKNC coding to obtain sequence integer coding and PseKNC feature vector;

[0033] A classification module is used to construct a two-way neural network model based on a self-attention mechanism, wherein the two-way neural network model based on a self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer encoding to obtain a tensor V 1 ; The second path network processes the PseKNC feature vector and outputs a tensor V 2 ; The classification module is used to convert the tensor V 1 With tensor V 2 Concatenate and input into the fully connected layer to output the classification result.

[0034] Based on the above content, the present invention further discloses a computer device, including: a memory for storing a computer program; and a processor for implementing any of the above methods when executing the computer program.

[0035] Based on the above content, the present invention further discloses a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any of the methods described above is implemented.

[0036] Based on the above technical scheme, the beneficial effects of the present invention are: N4-acetylcytosine (ac4C) is a type of RNA modification that occurs naturally in cells, which plays an important role in post-transcriptional regulation, mRNA stability and translation efficiency. Understanding the distribution pattern of ac4C helps to reveal new mechanisms of RNA function and how this modification affects the global post-transcriptional regulatory network. More and more studies have shown that abnormal RNA modification is closely related to the occurrence and development of various diseases (such as cancer, etc.). Therefore, accurate identification of ac4C sites helps to discover potential biomarkers and therapeutic targets. In the comparison of all existing models, the ACC of the DPNN-ac4C model of the present invention is improved by 2.07%, which strongly proves that the DPNN-ac4C model has better performance and reliability in identifying ac4C sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of a method for detecting N4-acetylcytosine sites in mRNA in one embodiment;

[0038] Figure 2 is a schematic diagram of the structure of a dual-path neural network based on a self-attention mechanism in a method for detecting N4-acetylcytosine sites in mRNA in an embodiment;

[0039] Figure 3 is a schematic diagram of a dual-path neural network process based on a self-attention mechanism in a method for detecting N4-acetylcytosine sites in mRNA in an embodiment;

[0040] Figure 4 is a bar chart comparing the performance of the DPNN-ac4C model and the existing model on an independent test set in an embodiment;

[0041] Figure 5 The figure is a radar chart comparing the performance of the DPNN-ac4C model and the existing model on an independent test set in an embodiment. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0043] like Figure 1 As shown, this embodiment provides a method for detecting N4-acetylcytosine sites in mRNA. First, the RNA sequence text is converted into integer coding and PseKNC coding. Because the neural network cannot directly accept text input. Then the integer-coded and PseKNC-coded vectors are input into the first network path and the second network path respectively. In the first network path, the integer-coded vector will pass through the embedding layer, the bidirectional GRU network, and the multi-head self-attention mechanism in sequence; in the second network path, the PseKNC vector will pass through the convolutional neural network and the multi-head self-attention mechanism in sequence, as described below:

[0044] Step 201, convert the mRNA sequence using integer coding and PseKNC coding to obtain the sequence integer coding and PseKNC feature vector.

[0045] In this embodiment, the computer device retrieves the mRNA sequence data from the database and converts the mRNA sequence using integer coding and PseKNC coding as follows:

[0046] 1) Convert mRNA sequences using integer encoding

[0047] Sequence integer encoding (IntegerEncoding) is used as the first encoding method of mRNA sequence. Integer encoding is achieved by mapping the bases of RNA sequence to integer domain. A base dictionary is first defined, in which each unique base symbol corresponds to a discrete integer value. Specifically, A is mapped to 0, C is mapped to 1, G is mapped to 2, and U is mapped to 3 one by one to obtain the final sequence encoding.

[0048] 2) Use PseKNC coding to convert mRNA sequences

[0049] PseKNC (Pseudo K-tuple Nucleotide Composition) is a feature extraction method for DNA, RNA or protein sequences. It is mainly used to describe the distribution of k-mers in biological sequences in order to better capture the local structural information of the sequence. First, for a given RNA sequence S, which contains L bases, it is divided into non-overlapping subsequences of length k, called k-mers. K-tuple Nucleotide Composition (KNC) is a method to describe the combination and frequency of these k-mers. For N types of k-mers, count the number of times it appears in the sequence and construct a frequency distribution vector, recorded as:

[0050] {X 1 , X 2 , ..., X N}

[0051] Calculate the frequency of each k-mer in the sequence and get the frequency vector:

[0052]

[0053] Where P i is the frequency of the i-th k-mer, i∈{1,2,…,N}. Secondly, PseKNC improves the expressiveness of sequence features by introducing “pseudo-components”, which include two aspects of pseudo-components: one is the weight factor ω of each k-mer frequency. i We get, i∈{1,2,…,N}:

[0054]

[0055] The second is the correlation vector obtained by the correlation between different k-mers:

[0056]

[0057] The frequency vector, weight factor and correlation vector of KNC are combined to obtain the PseKNC feature vector:

[0058]

[0059] Step 202: construct a dual-path neural network model based on the self-attention mechanism (DPNN-ac4C model), wherein the dual-path neural network model based on the self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer encoding to obtain a tensor V 1 ; The second path network processes the PseKNC feature vector and outputs a tensor V 2; The classification module is used to convert the tensor V 1 With tensor V 2 Concatenate and input into the fully connected layer to output the classification result. Figure 2 , 3 . Specifically,

[0060] 1) The first path network (left branch), for integer encoding:

[0061] a. Take data from the database as input;

[0062] b. Input the sequence integer encoding into the embedding layer, and set the embedding dimension to 512;

[0063] c. Add the position encoding matrix to the embedded tensor;

[0064] d. Input the position-encoded tensor into the bidirectional GRU network. Since the hidden dimension is 512, the output dimension is 1024.

[0065] e. Input the tensor output by Bi-GRU into the multi-head self-attention, and the number of heads is set to 8.

[0066] 2) The second path network (right branch) for PseKNC coding:

[0067] a. Take data from the database as input;

[0068] b. Input the PseKNC feature vector into the first convolutional layer. The sequence is originally 1 channel and the output channel dimension is 16;

[0069] c. After ReLU activation and maximum pooling processing;

[0070] d. Input the tensor output from 3) into the second convolutional layer, with 16 input channels and 32 output channels;

[0071] e. After ReLU activation and maximum pooling processing;

[0072] f. Input the tensor output from 5) into the third convolutional layer, with 32 input channels and 16 output channels;

[0073] g. After ReLU activation and maximum pooling processing;

[0074] h. The output tensor is input into the multi-head self-attention, and the number of heads is set to 8.

[0075] 3. Classification Module

[0076] Encode the sequence integer into the tensor V after processing by the first path network 1 and the tensor V output by the second path network 2The concatenated input is fed into the fully connected layer to obtain the classification result. In the classification result, samples with a score lower than 0.5 are classified as negative samples, and samples with a score higher than 0.5 are classified as positive samples.

[0077] In this embodiment, a multi-head self-attention mechanism is designed and used on both paths of the dual-path neural network model (DPNN-ac4C model) based on the self-attention mechanism, which is connected to the convolutional neural network (CNN) and the bidirectional GRU network (Bi-GRU) respectively. Unlike a single attention mechanism that only performs weighted processing on the input sequence from one perspective, the multi-head self-attention designs multiple attention "heads" running in parallel, each of which focuses on different subspace features of the input sequence through different linear transformations. By independently calculating these attention weights and then splicing the results of each head, the model can simultaneously capture multiple association patterns and contextual information of the sequence data. Unlike independent Bi-GRU or CNN, which rely on fixed-length context windows or local receptive fields to capture sequence information, the self-attention mechanism can process sequence data globally, allowing the model to directly establish connections between elements at any two positions. Specifically, CNN is good at extracting continuous nucleotide local pattern features, Bi-GRU can capture bidirectional time series dependencies, and the self-attention module further enhances the capabilities of these two models, so that the final V 1 and V 2 (like Figure 2 ) not only retains local features, but also supplements global context information, which is of great significance for understanding complex RNA secondary structures and predicting their potential biological functions. The number of multi-heads in the present invention is set to 8.

[0078] In one embodiment, the method for detecting N4-acetylcytosine sites in mRNA also provides a processing process of the first pathway network, comprising the following steps:

[0079] Linearly mapping the sequence integer code to a high-dimensional representation through an embedding layer to obtain a 512-dimensional high-dimensional tensor;

[0080] Concatenate the high-dimensional tensor with the position encoding matrix to obtain the concatenated matrix v embedding ;

[0081] The concatenated matrix v embedding Input the bidirectional GRU network to extract contextual semantic information and calculate based on multi-head attention to obtain the tensor V 1 .

[0082] In this embodiment, the embedding layer (Embedding Module): the RNA sequence is integer encoded and then input into the embedding layer, which converts the features into higher-dimensional feature vector representations, and these vectors are learnable and will gradually change with the training of the model, which will further strengthen the expression of the original features. After becoming a higher-dimensional vector, in order to further enhance the representation of its position information, positional encoding (Positional Encoding) is further used to strengthen its representation. The embedding size of the embedding layer is set to 512 dimensions. Specifically, the high-dimensional tensor after the output of the embedding layer is The position encoding matrix is Then for the position encoding matrix PE, its odd columns and even columns are:

[0083] Odd columns:

[0084] Even columns: The position-encoded vector:

[0085] v embedding =v+PE

[0086] here a∈N.

[0087] Bi-GRU network: Bidirectional GRU network is a variant of recurrent neural network (RNN), which is often used to process sequence tasks. "Bidirectional" means that the network considers past and future information at each time step. GRU is a gated recurrent unit that can capture long-term dependencies in sequences well, and because of its relatively small number of parameters, it has higher training performance than LSTM. Specifically, the hidden layer of the Bi-GRU network is 512-dimensional and has 3 layers.

[0088] In one embodiment, the method for detecting N4-acetylcytosine sites in mRNA also provides a processing process of the second pathway network, comprising the following steps:

[0089] A convolutional neural network is used to extract local feature information of the PseKNC feature vector, wherein the convolutional neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a third convolutional layer, a third activation layer, and a third maximum pooling layer connected in sequence, wherein the padding parameters of the first convolutional layer, the second convolutional layer, and the third convolutional layer are all 2, the convolution kernel size is 4, and the number of channels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are 16, 32, and 16, respectively;

[0090] Use multi-head attention to calculate the local feature information and output the tensor V 2 .

[0091] In the present embodiment, the convolutional neural network is a type of deep learning neural network, which is mainly used to process and analyze data such as images and sequences. The main feature of the convolutional neural network is to use convolutional layers to efficiently capture local features, thereby achieving efficient learning and representation of data. In addition, in order to make the network more expressive, robust and reduce the number of parameters, activation layers and pooling layers are interspersed among multiple convolutional layers in the present embodiment. These convolutional layers, activation layers and pooling layers are stacked together to form the convolutional neural network module of this article. Specifically, there are 3 convolutional layers in the convolutional neural network, and their convolution output layers are 16, 32, and 16 respectively, the padding is 2, and the convolution kernel size is 4. The first activation layer, the second activation layer, and the third activation layer are all ReLU activation functions.

[0092] The DPNN-ac4C model showed excellent performance in multiple evaluation indicators (SEN, SPE, ACC, MCC, AUROC). In the training validation set test, ACC reached 83.89%, MCC was 66.55%, AUC was 92.72%, SPE was 85.61%, and SEN was 82.18%. In the independent test set data, ACC reached 82.78%, MCC was 65.78%, AUC was 91.03%, SPE was 84.78%, and SEN was 81.78%, see Figure 4 , 5 Except for SEN, the other four indicators are higher than those of the existing models (PACES model, XG-ac4C model, and iRNA-ac4C model), which strongly proves that the DPNN-ac4C model has better performance and reliability in identifying ac4C sites.

[0093] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0094] The above is only a preferred embodiment of the method for detecting N4-acetylcytosine sites in mRNA and related devices disclosed in the present invention, and is not intended to limit the protection scope of the embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of this specification should be included in the protection scope of the embodiments of this specification.

Claims

1. Method for detecting N4-acetylcytosine sites in mRNA, It is characterized in that The steps include: The mRNA sequence is converted using integer coding and PseKNC coding to obtain the sequence integer coding and PseKNC feature vector; A two-path neural network model based on a self-attention mechanism is constructed, wherein the two-path neural network model based on a self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer encoding to obtain a tensor V 1 ; The second path network processes the PseKNC feature vector and outputs a tensor V 2 ; The classification module is used to convert the tensor V 1 With tensor V 2 Concatenate and input into the fully connected layer to output the classification results. Specifically, The processing process of the first path network includes the following steps: linearly mapping the sequence integer code to a high-dimensional representation through an embedding layer to obtain a 512-dimensional high-dimensional tensor; splicing the high-dimensional tensor with the position encoding matrix to obtain a splicing matrix; inputting the splicing matrix into a bidirectional GRU network to extract contextual semantic information and calculate based on multi-head attention to obtain a tensor V 1 ; The processing process of the second path network includes the following steps: using a convolutional neural network to extract local feature information of the PseKNC feature vector, the convolutional neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a third convolutional layer, a third activation layer and a third maximum pooling layer connected in sequence, the padding parameters of the first convolutional layer, the second convolutional layer and the third convolutional layer are all 2, the convolution kernel size is 4, and the number of channels of the first convolutional layer, the second convolutional layer and the third convolutional layer are 16, 32 and 16 respectively; using multi-head attention to calculate the local feature information and output a tensor V 2 .

2. The method for detecting N4-acetylcytosine sites in mRNA according to claim 1, It is characterized in that The mRNA sequence is converted using integer coding to obtain the sequence integer coding, including the following steps: Define a base dictionary in which each unique base symbol corresponds to a discrete integer value, that is, map A to 0, C to 1, G to 2, and U to 3 one by one to obtain the final sequence integer encoding.

3. The method for detecting N4-acetylcytosine sites in mRNA according to claim 1, It is characterized in that The mRNA sequence is converted using PseKNC encoding to obtain the PseKNC feature vector, including the following steps: For a given RNA sequence S, which contains L bases, it is divided into non-overlapping subsequences of length k. For N types of k-mers, the number of times they appear in the sequence is counted, and a frequency distribution vector is constructed, which is recorded as: {X 1 ,X 2 ,...,X N } Calculate the frequency of each k-mer in the sequence and get the frequency vector: Where P i is the frequency of the i-th k-mer, i∈{1,2,…,N}. Secondly, PseKNC improves the expressiveness of sequence features by introducing "pseudo-components", which include two aspects of pseudo-components: one is the weight factor ω of each k-mer frequency. i We get, i∈{1,2,…,N}: The second is the correlation vector obtained by the correlation between different k-mers: The frequency vector, weight factor and correlation vector of KNC are combined to obtain the PseKNC feature vector:

4. The method for detecting N4-acetylcytosine sites in mRNA according to claim 1, It is characterized in that The hidden layer of the bidirectional GRU network is 512-dimensional and has 3 layers.

5. The method for detecting N4-acetylcytosine sites in mRNA according to claim 1, It is characterized in that The first activation layer, the second activation layer, and the third activation layer are all ReLU activation functions. 6.N4-acetylcytosine site detection system in mRNA, It is characterized in that include: A conversion module, used to convert the mRNA sequence using integer coding and PseKNC coding to obtain sequence integer coding and PseKNC feature vector; A classification module is used to construct a two-way neural network model based on a self-attention mechanism, wherein the two-way neural network model based on a self-attention mechanism includes a first path network, a second path network and a classification module, wherein the first path network processes the sequence integer encoding to obtain a tensor V 1 ; The second path network processes the PseKNC feature vector and outputs a tensor V 2 ; The classification module is used to convert the tensor V 1 With tensor V 2 Concatenate and input into the fully connected layer to output the classification results. Specifically, The processing process of the first path network includes the following steps: linearly mapping the sequence integer code to a high-dimensional representation through an embedding layer to obtain a 512-dimensional high-dimensional tensor; concatenating the high-dimensional tensor with a position encoding matrix to obtain a concatenated matrix; The concatenated matrix is ​​input into the bidirectional GRU network to extract contextual semantic information and calculate based on multi-head attention to obtain the tensor V 1 ; The processing process of the second path network includes the following steps: using a convolutional neural network to extract local feature information of the PseKNC feature vector, the convolutional neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a third convolutional layer, a third activation layer and a third maximum pooling layer connected in sequence, the padding parameters of the first convolutional layer, the second convolutional layer and the third convolutional layer are all 2, the convolution kernel size is 4, and the number of channels of the first convolutional layer, the second convolutional layer and the third convolutional layer are 16, 32 and 16 respectively; using multi-head attention to calculate the local feature information and output a tensor V 2 .

7. A computer device, It is characterized in that include: Memory for storing computer programs; A processor, configured to implement the method according to any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, It is characterized in that The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Citation Information

Patent Citations

  • Hybrid model for predicting N4-acetylcytidine

    CN113362900A

  • Multi-label image classification method fusing strong correlation between labels

    CN114648635A

  • RNA modification site prediction method of two-way representation model based on attention

    CN115424663A

  • RNA positioning prediction method and device and storage medium

    CN117501374A

  • Detection method using fusion network based on attention mechanism, and terminal device

    US11222217B1