Method and device for recognizing modification sites in RNA sequence, computer device and medium

Abstract

The application relates to a method and device for identifying a modification site in an RNA sequence, computer equipment and a medium. The method comprises the following steps: obtaining sequence features, structure features of the RNA sequence to be identified, and a pre-trained modification site identification model; inputting the structure features into a structure convolution layer for convolution to obtain convolution structure features; in each attention encoding unit: inputting the sequence features into a query weight matrix, a key weight matrix and a value weight matrix respectively for linear transformation, thereby obtaining query features, key features and value features; fusing the query features, the key features and the convolution structure features to obtain first fusion features; fusing the first fusion features and the value features to obtain second fusion features; splicing the second fusion features output by all the attention encoding units to obtain spliced features; and performing feature adjustment on the spliced features based on a feature adjustment module to obtain an identification result. The method can improve the accuracy of the identification result.

Description

Technical Field

[0001] This application relates to the field of interdisciplinary technology of biology and artificial intelligence, and in particular to a method, apparatus, computer equipment and medium for identifying modification sites in RNA sequences. Background Technology

[0002] Chemical modifications in RNA sequences play a crucial role in regulating RNA expression and function. Among them, N6-methyladenosylation (m6A modification) is one of the most abundant modification types in eukaryotic cells. Recent studies have highlighted the impact of m6A modification on RNA structure, RNA stability, and RNA translation. Therefore, accurately identifying m6A modification sites is one of the urgent needs for studying the pathogenesis of diseases related to the nervous system and tumors, as well as for studying the translational regulatory response to environmental changes.

[0003] In related technologies, the identification of m6A modification sites usually only considers RNA sequence information as an influencing factor, resulting in low accuracy of the identification results of modification sites in RNA sequences.

[0004] Therefore, how to identify modification sites in RNA sequences to improve the accuracy of identification results is a problem that needs to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, computer equipment, and medium for identifying modification sites in RNA sequences to address the aforementioned technical problems and improve the accuracy of the identification results.

[0006] In a first aspect, this application provides a method for identifying modification sites in an RNA sequence, comprising:

[0007] The sequence features and structural features of the RNA sequence to be identified are obtained, and a pre-trained modification site identification model is obtained. The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site identification model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix.

[0008] The structural features are input into the structural convolutional layer for convolution to obtain the convolutional structural features;

[0009] In each attention encoding unit, the following steps are performed: the sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain query features, key features, and value features; the query features, key features, and convolutional structure features are fused to obtain a first fused feature; the first fused feature is fused with the value features to obtain a second fused feature.

[0010] The second fusion features output by all the attention encoding units are concatenated to obtain the concatenated features;

[0011] Based on the feature adjustment module, the splicing features are adjusted to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0012] In one embodiment, the fusion of the query feature, the key feature, and the convolutional structure feature to obtain the first fused feature includes:

[0013] The query features and the key features are fused to obtain the fused features;

[0014] Based on the dimensional features corresponding to the fusion features and the key features, the scaled fusion features are obtained.

[0015] The scaled fusion feature is fused with the convolutional structure feature to obtain the first fusion feature.

[0016] In one embodiment, the attention encoder further includes sequential convolutional layers and sequential deconvolutional layers; the method further includes:

[0017] The sequence features are input into the sequence convolutional layer for convolution to obtain convolutional sequence features;

[0018] In each attention encoding unit, the convolutional sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain the corresponding second fusion feature;

[0019] The spliced ​​features are input into the sequence deconvolution layer for deconvolution to obtain deconvolution sequence features;

[0020] The step of adjusting the splicing features based on the feature adjustment module to obtain the identification result of the modification site in the RNA sequence to be identified includes:

[0021] Based on the feature adjustment module, the deconvolution sequence features are adjusted to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0022] In one embodiment, the pre-trained modification site recognition model includes multiple consecutive attention encoders and the feature adjustment module, each attention encoder further including a structural deconvolution layer; the method further includes:

[0023] In each of the attention encoders, the convolutional structure features are input into the corresponding deconvolutional layer to perform deconvolution, thereby obtaining the corresponding deconvolutional structure features.

[0024] The structural features in each of the input attention encoders are residually connected with the deconvolution structural features in the corresponding attention encoders to obtain the processed structural features.

[0025] The sequence features input to each attention encoder are residually concatenated with the deconvolution sequence features in the corresponding attention encoder to obtain the processed sequence features.

[0026] The processed structural feature output by the previous attention encoder is used as the structural feature input by the next attention encoder, and the processed sequence feature output by the previous attention encoder is used as the sequence feature input by the next attention encoder; the sequence feature input by the first attention encoder is the sequence feature of the RNA sequence to be identified, and the structural feature input by the first attention encoder is the structural feature of the RNA sequence to be identified;

[0027] The processed sequence features output by the last attention encoder are input into the feature adjustment module to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0028] In one embodiment, the feature adjustment module includes a max-pooling layer, a fully connected layer, and an output layer. The step of adjusting the spliced ​​features based on the feature adjustment module to obtain the modification site identification result of the RNA sequence to be identified includes:

[0029] The processed sequence features output by the last attention encoder are input into the max pooling layer for dimensionality reduction to obtain the dimensionality-reduced features.

[0030] Based on the fully connected layer, feature extraction is performed on the dimensionality-reduced features to obtain extracted feature information;

[0031] The extracted feature information is input into the output layer to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0032] In one embodiment, the output layer is a regression layer, and the identification result of the modification site in the RNA sequence to be identified includes the modification intensity of the modification site in the RNA sequence to be identified.

[0033] In one embodiment, the method further includes:

[0034] The structural information of the RNA sequence to be identified is perturbed to obtain the perturbed RNA sequence to be identified.

[0035] Based on the pre-trained modification site recognition model, the perturbed RNA sequence to be identified is used to identify modification sites, and the perturbation modification site recognition result corresponding to the RNA sequence to be identified is obtained.

[0036] Based on the identification results of the modification sites in the RNA sequence to be identified and the identification results of the perturbation modification sites, the structural sensitivity of the modification sites in the RNA sequence to be identified is determined.

[0037] Secondly, this application also provides a device for identifying modification sites in an RNA sequence, comprising:

[0038] The feature and model acquisition module is used to acquire the sequence features and structural features of the RNA sequence to be identified, and to acquire a pre-trained modification site recognition model. The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site recognition model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix.

[0039] The convolution module is used to input the structural features into the structural convolutional layer for convolution to obtain convolutional structural features;

[0040] The encoding module is used to perform the following steps in each of the attention encoding units:

[0041] The sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain query features, key features, and value features; the query features, key features, and convolutional structure features are fused to obtain a first fused feature; the first fused feature is fused with the value features to obtain a second fused feature;

[0042] The splicing module is used to splice the second fusion features output by all the attention encoding units to obtain spliced ​​features;

[0043] The feature adjustment module is used to adjust the splicing features based on the feature adjustment module to obtain the identification result of the modification site in the RNA sequence to be identified.

[0044] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method for identifying modification sites in RNA sequences in any of the above embodiments.

[0045] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for identifying modification sites in the RNA sequence in any of the above embodiments.

[0046] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the method for identifying modification sites in the RNA sequence in any of the above embodiments.

[0047] In the above implementation process, the sequence features and structural features of the RNA sequence to be identified are input into a pre-trained modification site identification model for modification site identification. This ensures that the identification process considers not only the sequence information of the RNA sequence but also its structural information, effectively improving the accuracy of modification site identification results. Furthermore, during modification site identification in the pre-trained modification site identification model, structural convolution modules are used to convolve the structural features, effectively extracting important information from the structural features. Further, the sequence features are input into each attention encoding unit, including the query weight matrix, key weight matrix, and value weight matrix, for linear transformation, thereby obtaining information corresponding to the sequence features in three different characteristics. Finally, the query features, key features, and convolutional structural features are fused to obtain a first fused feature, and this first fused feature is fused with the value features to obtain a second fused feature, effectively achieving the fusion of the structural features of the RNA sequence with the three different characteristics corresponding to the sequence features. Furthermore, the second fusion features output by multiple attention coding units are concatenated, thereby integrating the encoding information of structural and sequence features by multiple attention coding units, effectively enhancing the expressive power of the model. Finally, the concatenated features, which include rich structural and sequence features, are input into the feature fine-tuning module for feature fine-tuning, thereby effectively improving the accuracy of the modification site recognition results. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram illustrating the application environment of a method for identifying modification sites in an RNA sequence, as provided in an embodiment of this application.

[0050] Figure 2 This is a flowchart illustrating a method for identifying modification sites in an RNA sequence according to an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the structure of an optional pre-trained modification site recognition model provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the structure of another optional pre-trained modification site recognition model provided in the embodiments of this application;

[0053] Figure 5 This is a schematic diagram illustrating the prediction results of m6A modification intensity in a human embryonic kidney cell line, provided in an embodiment of this application.

[0054] Figure 6 This is a schematic diagram illustrating the predictive correlation of different sequence lengths in different cell lines, provided in an embodiment of this application.

[0055] Figure 7 This is a schematic diagram illustrating the perturbation processing of the structural information of an RNA sequence to be identified, provided in an embodiment of this application.

[0056] Figure 8 This is a schematic diagram illustrating the results of a structural perturbation experiment performed in a human embryonic kidney cell line, as provided in an embodiment of this application.

[0057] Figure 9 This is a schematic diagram illustrating the results of a structural perturbation experiment in a human cervical cancer cell line, as provided in an embodiment of this application.

[0058] Figure 10 This is a schematic diagram illustrating the accuracy results of different algorithms on different datasets according to an embodiment of this application.

[0059] Figure 11 This is a schematic diagram of the structure of a device for recognizing modification sites in an RNA sequence, provided in an embodiment of this application.

[0060] Figure 12This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0063] The method for identifying modification sites in RNA sequences provided in this application can be applied to, for example... Figure 1 In the application environment shown, Figure 1 This is a schematic diagram illustrating the application environment of a method for identifying modification sites in an RNA sequence according to an embodiment of this application. Terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located on a cloud or other network server. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0064] In one exemplary embodiment, Figure 2 This is a flowchart illustrating a method for identifying modification sites in an RNA sequence according to an embodiment of this application, which is applied to... Figure 1 Taking server 104 as an example, the explanation is as follows: Figure 2 As shown, the method may include the following steps:

[0065] Step 201: Obtain the sequence and structural features of the RNA sequence to be identified, and obtain a pre-trained modification site identification model.

[0066] The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site identification model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix.

[0067] For example, the sequence in the RNA fragment for which modification site identification is required is identified as the RNA sequence to be identified, and the base sequence in the RNA sequence to be identified is one-hot encoded to obtain the sequence characteristics of the RNA sequence to be identified.

[0068] As another example, the sequence information of the RNA sequence to be identified is sequentially input into an embedding layer and a linear mapping layer to obtain sequence features. The embedding layer can convert discrete base symbols into sequence vectors, and the linear mapping layer can convert sequence vectors into high-level sequence features.

[0069] It should be noted that the method for identifying modification sites in the RNA sequence of this application can be applied to any of the various modification sites, such as 5-methylcytosine (m5C), N1-methyladenine (m1A), pseudouracil (ψ), 5-hydroxymethylcytosine (hm5C), and N6-methyladenylation (m6A). This application only identifies m6A modification in the RNA sequence.

[0070] The base sequences in the RNA sequence include adenine (A), guanine (G), cytosine (C), and thymine (T).

[0071] The structural information of the RNA sequence to be identified is extracted. This structural information may include at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. In this embodiment, the structural information including the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified is used as an example for illustration, and no limitation is made herein.

[0072] Furthermore, the base pairing probability matrix and relative distance matrix of the RNA sequence to be identified are encoded to obtain the structural features of the RNA sequence. This encoding method may include at least one of size normalization, numerical normalization, and channel fusion.

[0073] And obtain a pre-trained modification site recognition model, which can be obtained by pre-training based on RNA sequence samples. Figure 3 This is a schematic diagram of the structure of an optional pre-trained modification site recognition model provided in an embodiment of this application, as shown below. Figure 3 As shown, the pre-trained modification site recognition model includes an attention encoder and a feature adjustment module. The attention encoder includes N attention encoding units and a structured convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix. N is greater than or equal to 2. In this embodiment, N equals 32 as an example.

[0074] As an example, multiple attention encoding units can be multiple attention heads of a Transformer model, each attention head including three linear transformation matrices Q, K, and V, which correspond to the query weight matrix, key weight matrix, and value weight matrix in this application, respectively.

[0075] Step 202: Input the structural features into the structural convolutional layer for convolution to obtain the convolutional structural features.

[0076] Furthermore, when inputting the structural and sequence features of the RNA sequence to be identified into the pre-trained modification site identification model for modification site identification, the structural and sequence features can be input into the attention encoder for encoding, and the result encoded by the attention encoder can be input into the feature adjustment module for feature adjustment, thereby obtaining the identification result of the modification site in the RNA sequence to be identified.

[0077] Specifically, the structural features can be input into a structural convolutional layer for convolution to obtain convolutional structural features. As an example, the structural convolutional layer can be a 2D convolutional network or a 1D convolutional network. In this embodiment, a 2D convolutional network is used as an example for illustration, and no limitation is made here.

[0078] Step 203: Perform the following steps in each attention encoding unit: input the sequence features into the query weight matrix, key weight matrix and value weight matrix respectively for linear transformation to obtain the query features, key features and value features respectively; fuse the query features, key features and convolutional structure features to obtain the first fused feature; fuse the first fused feature with the value feature to obtain the second fused feature.

[0079] Furthermore, the sequence features of the RNA sequence to be identified are input into each attention coding unit for encoding. Specifically, the following operations can be performed in each attention coding unit:

[0080] The sequence features are input into the query weight matrix, key weight matrix, and value weight matrix respectively for linear transformation to obtain the query feature, key feature, and value feature corresponding to each attention encoding unit.

[0081] Then, the query features, key features, and convolutional structure features output by the structured convolutional layer are fused to obtain the first fused feature. This first fused feature is then fused with the value features from the same attention encoding unit to obtain the second fused feature output by that attention encoding unit.

[0082] Step 204: Concatenate the second fusion features output by all attention encoding units to obtain concatenated features.

[0083] Furthermore, the second fusion features output by all attention coding units are determined in step 203, and the second fusion features output by all attention coding units are concatenated to obtain the concatenated features.

[0084] Step 205: Based on the feature adjustment module, perform feature adjustment on the splicing features to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0085] For example, the splicing features are input into the feature adjustment module for feature adjustment, thereby obtaining the identification results of the modification sites in the RNA sequence to be identified.

[0086] As an example, the feature adjustment module may include a max pooling layer, a fully connected layer, and an output layer. The spliced ​​features can be input into the max pooling layer for dimensionality reduction to obtain dimensionality-reduced features. The dimensionality-reduced features can then be input into the fully connected layer for feature extraction to obtain extracted feature information. Finally, the extracted feature information can be input into the output layer, thereby enabling the output layer to output the identification results of the modification sites in the RNA sequence to be identified.

[0087] As an example, the output layer could be a classifier that can output a result indicating whether the RNA sequence to be identified contains a modification site.

[0088] In the above implementation process, the sequence features and structural features of the RNA sequence to be identified are input into a pre-trained modification site identification model for modification site identification. This ensures that the identification process considers not only the sequence information of the RNA sequence but also its structural information, effectively improving the accuracy of modification site identification results. Furthermore, during modification site identification in the pre-trained modification site identification model, structural convolution modules are used to convolve the structural features, effectively extracting important information from the structural features. Further, the sequence features are input into each attention encoding unit, including the query weight matrix, key weight matrix, and value weight matrix, for linear transformation, thereby obtaining information corresponding to the sequence features in three different characteristics. Finally, the query features, key features, and convolutional structural features are fused to obtain a first fused feature, and this first fused feature is fused with the value features to obtain a second fused feature, effectively achieving the fusion of the structural features of the RNA sequence with the three different characteristics corresponding to the sequence features. Furthermore, the second fusion features output by multiple attention coding units are concatenated, thereby integrating the encoding information of structural and sequence features by multiple attention coding units, effectively enhancing the expressive power of the model. Finally, the concatenated features, which include rich structural and sequence features, are input into the feature fine-tuning module for feature fine-tuning, thereby effectively improving the accuracy of the modification site recognition results.

[0089] In one embodiment, fusing query features, key features, and convolutional structure features to obtain a first fused feature may include the following steps:

[0090] Step 1: Fuse the query features with the key features to obtain the fused features.

[0091] Step 2: Scaling is performed based on the dimensional features corresponding to the fusion features and key features to obtain the scaled fusion features.

[0092] Step 3: Fuse the scaled and fused features with the convolutional structure features to obtain the first fused feature.

[0093] For example, if the sequence characteristics of the RNA sequence to be identified are In the Transformer's self-attention mechanism, each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix, respectively. , as well as The query features, key features, and value features are as follows: .

[0094] Then you can use query features transpose of key features Multiply to obtain the fused feature, and then divide the fused feature by the square root of the dimension corresponding to the key feature. This allows for scaling of the fused features, resulting in scaled fused features. .

[0095] Finally, the scaled and fused features are fused with the convolutional structure features to obtain the first fused feature.

[0096] As an example, if the convolutional structure features are The first fusion feature can be:

[0097]

[0098] in, These are query features, key features, and value features; Structural features of the RNA sequence to be identified; The dimension of the key feature; is a learnable balancing coefficient used to control the proportion of the structure-aware bias contribution relative to sequence similarity-based bias.

[0099] In the above implementation process, by fusing query features and key features, and scaling the fused features using the dimensions of the construction features, the similarity of sequence features is effectively determined. The similarity determination result is then fused with the convolutional structure features, effectively fusing the sequence features and structural features of the RNA sequence. This facilitates further fusion of the fused features with the value features, thereby achieving effective encoding of convolutional structure features, query features, key features, and value features in the attention encoding unit.

[0100] In one embodiment, the attention encoder further includes sequential convolutional layers and sequential deconvolutional layers; the method may also include the following steps:

[0101] Step 1: Input the sequence features into the sequence convolutional layer for convolution to obtain the convolutional sequence features;

[0102] Step 2: In each attention encoding unit, the convolutional sequence features are input into the query weight matrix, key weight matrix, and value weight matrix respectively for linear transformation to obtain the corresponding second fusion feature;

[0103] Step 3: Input the concatenated features into the sequence deconvolution layer for deconvolution to obtain the deconvolution sequence features;

[0104] Step 4: Based on the feature adjustment module, perform feature adjustment on the splicing features to obtain the identification results of the modification sites in the RNA sequence to be identified, including: based on the feature adjustment module, perform feature adjustment on the deconvolution sequence features to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0105] For example, the attention encoder may also include sequential convolutional layers and sequential deconvolutional layers. As an example, the sequential convolutional layers and sequential deconvolutional layers may both be one-dimensional or both are two-dimensional. In this embodiment of the application, the sequential convolutional layers and sequential deconvolutional layers are described as one-dimensional, and no limitation is made here.

[0106] Sequence information is input into a sequence convolutional layer for convolution, thereby effectively extracting local k-mer features, i.e., convolutional sequence features, from the sequence features. These local k-mer features are then input into each attention coding unit for processing. The processing method for local k-mer features and convolutional structure features in each attention coding unit is the same as the processing method for sequence features and convolutional structure features in each attention coding unit. This results in each attention coding unit outputting a corresponding second fusion feature, enhancing the model's ability to identify positions in the RNA sequence and facilitating accurate identification of whether there are modification sites in the RNA sequence.

[0107] Furthermore, the second fusion features output by all attention encoding units are concatenated to obtain concatenated features, which are then input into a sequence deconvolution layer for deconvolution to obtain deconvolution sequence features, facilitating further processing of the encoded features.

[0108] In the above implementation process, by convolving the sequence features through the sequence convolutional layer, local sequence features can be effectively extracted and input into the multi-head attention coding unit. This combines the two characteristics of the convolutional layer being able to efficiently extract local information with the subsequent attention mechanism being able to capture long-distance information. This makes it easier for the subsequent attention coding unit to accurately identify the positional information in the RNA sequence, and further facilitate the identification of modification sites based on the precise positional information in the sequence, thereby improving the accuracy of the modification site identification results.

[0109] In one embodiment, the pre-trained modification site recognition model includes multiple consecutive attention encoders and a feature adjustment module, each attention encoder further including a structural deconvolution layer; the method may also include the following steps:

[0110] Step 1: In each attention encoder, the convolutional structure features are input into the corresponding deconvolutional layer to perform deconvolution, thereby obtaining the corresponding deconvolutional structure features;

[0111] Step 2: Perform residual connections between the structural features in each input attention encoder and the deconvolutional structural features in the corresponding attention encoder to obtain the processed structural features;

[0112] Step 3: Perform residual concatenation between the sequence features in each attention encoder and the deconvolution sequence features in the corresponding attention encoder to obtain the processed sequence features;

[0113] Step 4: Use the processed structural features output by the previous attention encoder as the structural features input by the next attention encoder, and use the processed sequence features output by the previous attention encoder as the sequence features input by the next attention encoder; the sequence features input by the first attention encoder are the sequence features of the RNA sequence to be identified, and the structural features input by the first attention encoder are the structural features of the RNA sequence to be identified.

[0114] Step 5: Input the processed sequence features output by the last attention encoder into the feature adjustment module to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0115] For example, the pre-trained modification site recognition model includes M consecutive attention encoders and a feature adjustment module, where M is greater than or equal to 2. In this embodiment, M is 5 as an example.

[0116] The following processing is performed in each attention encoder:

[0117] The input structural features are passed through a structural convolutional layer to obtain corresponding convolutional structural features, and the input sequence features are passed through a sequence convolutional layer to obtain corresponding convolutional sequence features. These convolutional structural features and convolutional sequence features are then input into multiple attention encoding units, resulting in concatenated features output by these units. The concatenated features are then passed through a sequence deconvolutional layer to obtain deconvolutional sequence features, and the convolutional structural features are input into the structural deconvolutional layer to obtain deconvolutional structural features. Finally, the input structural features and deconvolutional structural features are residually concatenated to obtain processed structural features, and the input sequence features are residually concatenated with the deconvolutional sequence features to obtain processed sequence features. These processed structural features and processed sequence features are then used as the output features of an attention encoder.

[0118] Then, the features output by the previous attention encoder are input into the next attention encoder. Specifically, the processed structural features output by the previous attention encoder are used as the structural features input into the next attention encoder, and the processed sequence features output by the previous attention encoder are used as the sequence features input into the next attention encoder. The same operation as the previous attention encoder is performed in each attention encoder until the last attention encoder has been completed.

[0119] Because the attention encoder in the pre-trained modification site recognition model consists of multiple consecutive attention encoders, when the first attention encoder is used, the input structural features are the structural features of the RNA sequence to be recognized, and the input sequence features are the sequence features of the RNA sequence to be recognized. This allows the structural features of the RNA sequence to be recognized and the sequence features to be encoded consecutively by multiple attention encoders, effectively coupling the structural and sequence features. The coupled features then identify modification sites in the RNA sequence, significantly improving the accuracy of the recognition results.

[0120] Then, the processed sequence features output by the last attention encoder are input into the feature adjustment module for feature adjustment to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0121] In the above implementation process, multiple consecutive attention encoders are set in the trained modification site recognition model, so that the input structural features and sequence features can be coupled through multiple consecutive attention encoders, which enhances the expressive power of the model and effectively extracts the coupled features, which facilitates the improvement of the accuracy of subsequent identification of modification sites in RNA sequences.

[0122] In one embodiment, the feature adjustment module includes a max pooling layer, a fully connected layer, and an output layer. Based on the feature adjustment module, the spliced ​​features are adjusted to obtain the modification site identification result of the RNA sequence to be identified. This may include the following steps:

[0123] Step 1: Input the processed sequence features output by the last attention encoder into the max pooling layer for dimensionality reduction to obtain the dimensionality-reduced features.

[0124] Step 2: Extract features from the dimensionality-reduced features using a fully connected layer to obtain the extracted feature information.

[0125] Step 3: Input the extracted feature information into the output layer to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0126] For example, Figure 4This is a schematic diagram of the structure of another optional pre-trained modification site recognition model provided in the embodiments of this application, as shown below. Figure 4 The feature adjustment module shown includes a max pooling layer, a fully connected layer, and an output layer.

[0127] In the feature adjustment module, the input features are sequentially processed through a max pooling layer, a fully connected layer, and an output layer. If the input features are concatenated features, they are processed through a max pooling layer for dimensionality reduction, resulting in dimensionality-reduced features. If the input features are the processed sequence features output from the last attention encoder, they are input into a max pooling layer for dimensionality reduction, resulting in dimensionality-reduced features. As an example, the length dimension of the input features can be compressed to 1.

[0128] Furthermore, the dimensionality-reduced features are processed through a fully connected layer for feature extraction, yielding extracted feature information. This extracted feature information is then input into the output layer to obtain the identification result of modification sites in the RNA sequence to be identified. As an example, this output layer is a classification layer, which allows the identification result to include whether modification sites exist in the RNA sequence to be identified.

[0129] As another example, the feature adjustment module may also include a position fusion multi-head attention layer. The features input to the feature adjustment module are first processed by the position fusion multi-head attention layer to obtain position-fused features. Then, the position-fused features are passed through a max pooling layer, a fully connected layer, and an output layer to obtain the identification results of modification sites in the RNA sequence to be identified. By processing the features input to the feature adjustment module through the position fusion multi-head attention layer, the final output identification results of modification sites can be intuitively represented in the RNA sequence.

[0130] In the above implementation process, the features of the input feature adjustment module are processed by the max pooling layer for dimensionality reduction, thereby effectively ensuring the dimensionality of the output recognition result. The dimensionality-reduced features are then passed through the fully connected layer and the output layer to effectively obtain the recognition result of the modification site in the RNA sequence to be identified.

[0131] In one embodiment, the output layer is a regression layer, and the identification results of the modification sites in the RNA sequence to be identified include the modification intensity of the modification sites in the RNA sequence to be identified.

[0132] For example, the output layer in the feature adjustment module can be set as a regression layer so that when training the model, the output can include the modification intensity of the modification site in the RNA sequence to be identified, which can be characterized by the probability that the RNA sequence to be identified includes the modification site.

[0133] Specifically, in the embodiments of this application, the structural features and sequence features of the RNA sequence to be identified can be processed through multiple attention encoders to extract the coupling information between the sequence features and structural features. The global feature aggregation is performed using a max pooling layer, and the feature is perceived through a fully connected layer of a multilayer perceptron (MLP). The result is output through a regression layer, thereby predicting the modification intensity of each modification site in different cell lines.

[0134] Figure 5 This is a schematic diagram illustrating the prediction results of m6A modification intensity in a human embryonic kidney cell line, as provided in an embodiment of this application. Figure 5 As shown, in the human embryonic kidney cell line, the predicted modification intensity was significantly positively correlated with the actual measured modification level (Pearson correlation coefficient r=0.56).

[0135] Figure 6 This is a schematic diagram illustrating the predictive correlation of different sequence lengths in different cell lines, provided in an embodiment of this application. For example... Figure 6 As shown, the model's performance changes in human embryonic kidney cell lines and human cervical cancer cell lines under different input sequence lengths (101, 201, 301 nt). The longer the input sequence, the stronger the model's ability to capture contextual information, and the predictive relevance gradually increases. Specifically, the relevance of the model increases from 0.50 to 0.56 in human embryonic kidney cell lines and from 0.48 to 0.51 in human cervical cancer cell lines.

[0136] In the above implementation process, by setting the output layer of the pre-trained modification site recognition model as a regression layer, the pre-trained modification site recognition model can quantitatively identify the modification intensity of modification sites in RNA sequences.

[0137] In one embodiment, the method may further include the following steps:

[0138] Step 1: Perturb the structural information of the RNA sequence to be identified to obtain the perturbed RNA sequence to be identified.

[0139] Step 2: Based on the pre-trained modification site recognition model, the modification sites of the perturbated RNA sequence to be identified are identified to obtain the perturbation modification site recognition results corresponding to the RNA sequence to be identified.

[0140] Step 3: Based on the identification results of modification sites in the RNA sequence to be identified and the identification results of perturbation modification sites, determine the structural sensitivity of modification sites in the RNA sequence to be identified.

[0141] For example, in order to verify the structural sensitivity of modification sites in RNA structural features, the structural information of the RNA sequence to be identified can be perturbed to obtain the perturbed RNA sequence to be identified.

[0142] As an example, the secondary structure annotations (dot-brackets) of the RNA sequence can be locally perturbed while keeping the nucleotide sequence of the RNA sequence unchanged. A series of consecutive nucleotides (e.g., 15) with the highest distribution during recognition in the attention encoder of a pre-trained modification site recognition model can be selected as structural key sites, and their paired symbols “()” can be interchanged with their unpaired symbols “.”, thereby simulating the opening or re-pairing of local RNA regions.

[0143] Figure 7 This is a schematic diagram illustrating the perturbation processing of the structural information of an RNA sequence to be identified, provided in an embodiment of this application. Figure 7 As shown, the structural features and sequence features of the RNA sequence to be identified before perturbation are input into the model for modification site identification, and the identification results of modification sites in the RNA sequence to be identified are obtained.

[0144] Furthermore, the structural features and sequence features of the perturbated RNA sequence to be identified are input into the model for modification site identification, thereby obtaining the identification results of the perturbated modification sites corresponding to the RNA sequence to be identified. This model is the pre-trained modification site identification model in the embodiments of this application.

[0145] Furthermore, the identification results of the modification sites in the RNA sequence to be identified are compared with the identification results of the perturbation modification sites. If the identification results of the modification sites and the perturbation modification sites do not change, it indicates that the perturbation site is a structurally insensitive site; if the identification results of the modification sites and the perturbation modification sites change, it indicates that the perturbation site is a structurally sensitive site.

[0146] Figure 8 This is a schematic diagram illustrating the results of a structural perturbation experiment performed in a human embryonic kidney cell line, as provided in an embodiment of this application. Figure 9 This is a schematic diagram illustrating the results of a structural perturbation experiment in a human cervical cancer cell line, as provided in an embodiment of this application. Figure 8 as well as Figure 9 As shown, the icSHAPE values ​​of structure-sensitive modification sites in embryonic kidney cell lines and human cervical cancer cell lines were significantly higher than those of structure-insensitive modification sites (KS test, p<0.05), indicating that the modification of these sites is closely related to local RNA conformational changes, and the model captures the regulatory relationship between modification sites and structure.

[0147] In the above implementation process, the structural information of the RNA sequence to be identified is perturbed, and the perturbed RNA sequence is used to identify modification sites using a pre-trained modification site identification model to obtain the perturbed modification site identification result. Furthermore, if the perturbed modification site identification result changes from the modification site identification result of the RNA sequence before perturbing, it indicates that the corresponding site is a structurally sensitive site; if the perturbed modification site identification result does not change from the modification site identification result of the RNA sequence before perturbing, it indicates that the corresponding site is a structurally insensitive site, thereby accurately determining the structural sensitivity of sites in the RNA sequence.

[0148] Figure 10 This is a schematic diagram illustrating the accuracy results of different algorithms on different datasets provided in this application embodiment. Figure 10 As shown, the modification site identification method of this application, along with MASS, iM6A, DeepM6ASeq-CNN, DeepM6ASeq-RNN, XGBoost, RandomForest, and SVM, were used to identify modification sites on the SRAMP, DeepM6ASeq, SMART-m6A, Cerebellum7, and Lung4 datasets, respectively. The accuracy of the corresponding identification results was obtained. The modification site identification method of this application achieved the highest accuracy across all datasets. Therefore, this application can effectively improve the accuracy of modification site identification results in existing technologies.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0150] Based on the same inventive concept, this application also provides an RNA sequence modification site identification device for implementing the above-described method for identifying modification sites in RNA sequences. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more RNA sequence modification site identification device embodiments provided below can be found in the limitations of the RNA sequence modification site identification method described above, and will not be repeated here.

[0151] In one exemplary embodiment, Figure 11 This is a schematic diagram of the structure of a device for recognizing modification sites in an RNA sequence, as provided in an embodiment of this application. Figure 11 As shown, the device includes:

[0152] The feature and model acquisition module 111 is used to acquire the sequence features and structural features of the RNA sequence to be identified, and to acquire a pre-trained modification site recognition model. The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site recognition model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix.

[0153] Convolution module 112 is used to input structural features into the structural convolutional layer for convolution to obtain convolutional structural features;

[0154] Encoding module 113 is used to perform the following steps in each attention encoding unit:

[0155] The sequence features are input into the query weight matrix, key weight matrix, and value weight matrix respectively for linear transformation to obtain the query features, key features, and value features. The query features, key features, and convolutional structure features are then fused to obtain the first fused feature. The first fused feature is then fused with the value feature to obtain the second fused feature.

[0156] The splicing module 114 is used to splice the second fusion features output by all attention encoding units to obtain spliced ​​features;

[0157] The feature adjustment module 115 is used to adjust the splicing features based on the feature adjustment module to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0158] In one embodiment, the encoding module 113 is specifically used for:

[0159] The query features and key features are fused to obtain the fused features;

[0160] The scaled-down fusion features are obtained by scaling the dimensional features corresponding to the fusion features and key features.

[0161] The scaled and fused features are then fused with the convolutional structure features to obtain the first fused feature.

[0162] In one embodiment, the attention encoder further includes a sequence convolutional layer and a sequence deconvolutional layer; the encoding module 113 is also configured to:

[0163] The sequence features are input into the sequence convolutional layer for convolution to obtain the convolutional sequence features;

[0164] In each attention encoding unit, the convolutional sequence features are input into the query weight matrix, key weight matrix, and value weight matrix respectively for linear transformation to obtain the corresponding second fusion features;

[0165] The concatenated features are input into the sequence deconvolution layer for deconvolution to obtain the deconvolution sequence features;

[0166] The feature adjustment module 115 is specifically used to: perform feature adjustment on the deconvolution sequence features based on the feature adjustment module, and obtain the identification results of the modification sites in the RNA sequence to be identified.

[0167] In one embodiment, the pre-trained modification site recognition model includes multiple consecutive attention encoders and a feature adjustment module, each attention encoder further including a structural deconvolution layer; the encoding module 113 is also used for:

[0168] In each attention encoder, the convolutional structural features are input into the corresponding deconvolutional layer to perform deconvolution, thereby obtaining the corresponding deconvolutional structural features.

[0169] The structural features in each input attention encoder are residually concatenated with the deconvolutional structural features in the corresponding attention encoder to obtain the processed structural features.

[0170] The sequence features input to each attention encoder are residually concatenated with the deconvolutional sequence features in the corresponding attention encoder to obtain the processed sequence features.

[0171] The processed structural features output by the previous attention encoder are used as the structural features input by the next attention encoder, and the processed sequence features output by the previous attention encoder are used as the sequence features input by the next attention encoder; the sequence features input by the first attention encoder are the sequence features of the RNA sequence to be identified, and the structural features input by the first attention encoder are the structural features of the RNA sequence to be identified.

[0172] The processed sequence features output from the last attention encoder are input into the feature adjustment module to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0173] In one embodiment, the feature adjustment module includes a max pooling layer, a fully connected layer, and an output layer. The feature adjustment module 115 is specifically used for:

[0174] The processed sequence features output by the last attention encoder are input into the max pooling layer for dimensionality reduction to obtain the dimensionality-reduced features.

[0175] Feature extraction is performed on the dimensionality-reduced features based on a fully connected layer to obtain the extracted feature information;

[0176] The extracted feature information is input into the output layer to obtain the identification results of the modification sites in the RNA sequence to be identified.

[0177] In one embodiment, the output layer is a regression layer, and the identification results of the modification sites in the RNA sequence to be identified include the modification intensity of the modification sites in the RNA sequence to be identified.

[0178] In one embodiment, the encoding module 113 is further configured to:

[0179] The structural information of the RNA sequence to be identified is perturbed to obtain the perturbed RNA sequence to be identified.

[0180] Based on a pre-trained modification site recognition model, the modification sites of the perturbated RNA sequence to be identified are identified, and the perturbation modification site recognition results of the RNA sequence to be identified are obtained.

[0181] Based on the identification results of modification sites in the RNA sequence to be identified and the identification results of perturbation modification sites, the structural sensitivity of modification sites in the RNA sequence to be identified is determined.

[0182] Each module in the RNA sequence modification site recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the operations corresponding to each module.

[0183] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 12 As shown, Figure 12This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data of RNA sequences to be identified. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying modification sites in the RNA sequence.

[0184] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0185] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for identifying modification sites in an RNA sequence in any of the above embodiments.

[0186] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements steps in a method for identifying modification sites in an RNA sequence.

[0187] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0188] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0189] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0190] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying modification sites in an RNA sequence, characterized in that, The method includes: The sequence features and structural features of the RNA sequence to be identified are obtained, and a pre-trained modification site identification model is obtained. The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site identification model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix. The structural features are input into the structural convolutional layer for convolution to obtain the convolutional structural features; In each attention encoding unit, the following steps are performed: the sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain query features, key features, and value features; the query features, key features, and convolutional structure features are fused to obtain a first fused feature; the first fused feature is fused with the value features to obtain a second fused feature. The second fusion features output by all the attention encoding units are concatenated to obtain the concatenated features; Based on the feature adjustment module, the splicing features are adjusted to obtain the identification results of the modification sites in the RNA sequence to be identified.

2. The method according to claim 1, characterized in that, The first fused feature is obtained by fusing the query feature, the key feature, and the convolutional structure feature, including: The query features and the key features are fused to obtain the fused features; Based on the dimensional features corresponding to the fusion features and the key features, the scaled fusion features are obtained. The scaled fusion feature is fused with the convolutional structure feature to obtain the first fusion feature.

3. The method according to claim 1, characterized in that, The attention encoder further includes sequential convolutional layers and sequential deconvolutional layers; the method further includes: The sequence features are input into the sequence convolutional layer for convolution to obtain convolutional sequence features; In each attention encoding unit, the convolutional sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain the corresponding second fusion feature; The spliced ​​features are input into the sequence deconvolution layer for deconvolution to obtain deconvolution sequence features; The step of adjusting the splicing features based on the feature adjustment module to obtain the identification result of the modification site in the RNA sequence to be identified includes: Based on the feature adjustment module, the deconvolution sequence features are adjusted to obtain the identification results of the modification sites in the RNA sequence to be identified.

4. The method according to claim 3, characterized in that, The pre-trained modification site recognition model includes multiple consecutive attention encoders and the feature adjustment module, each attention encoder further including a structural deconvolution layer; the method further includes: In each of the attention encoders, the convolutional structure features are input into the corresponding deconvolutional layer to perform deconvolution, thereby obtaining the corresponding deconvolutional structure features. The structural features in each of the input attention encoders are residually connected with the deconvolution structural features in the corresponding attention encoders to obtain the processed structural features. The sequence features input to each attention encoder are residually concatenated with the deconvolution sequence features in the corresponding attention encoder to obtain the processed sequence features. The processed structural feature output by the previous attention encoder is used as the structural feature input by the next attention encoder, and the processed sequence feature output by the previous attention encoder is used as the sequence feature input by the next attention encoder; the sequence feature input by the first attention encoder is the sequence feature of the RNA sequence to be identified, and the structural feature input by the first attention encoder is the structural feature of the RNA sequence to be identified; The processed sequence features output by the last attention encoder are input into the feature adjustment module to obtain the identification results of the modification sites in the RNA sequence to be identified.

5. The method according to claim 4, characterized in that, The feature adjustment module includes a max pooling layer, a fully connected layer, and an output layer. The feature adjustment based on the splicing features, to obtain the modification site identification result of the RNA sequence to be identified, includes: The processed sequence features output by the last attention encoder are input into the max pooling layer for dimensionality reduction to obtain the dimensionality-reduced features. Based on the fully connected layer, feature extraction is performed on the dimensionality-reduced features to obtain extracted feature information; The extracted feature information is input into the output layer to obtain the identification results of the modification sites in the RNA sequence to be identified.

6. The method according to claim 5, characterized in that, The output layer is a regression layer, and the identification result of the modification sites in the RNA sequence to be identified includes the modification intensity of the modification sites in the RNA sequence to be identified.

7. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The structural information of the RNA sequence to be identified is perturbed to obtain the perturbed RNA sequence to be identified. Based on the pre-trained modification site recognition model, the perturbed RNA sequence to be identified is used to identify modification sites, and the perturbation modification site recognition result corresponding to the RNA sequence to be identified is obtained. Based on the identification results of the modification sites in the RNA sequence to be identified and the identification results of the perturbation modification sites, the structural sensitivity of the modification sites in the RNA sequence to be identified is determined.

8. A device for recognizing modification sites in an RNA sequence, characterized in that, The device includes: The feature and model acquisition module is used to acquire the sequence features and structural features of the RNA sequence to be identified, and to acquire a pre-trained modification site recognition model. The structural features are generated by encoding at least one of the base pairing probability matrix and the relative distance matrix of the RNA sequence to be identified. The pre-trained modification site recognition model includes an attention encoder and a feature adjustment module. The attention encoder includes multiple attention encoding units and a structural convolutional layer. Each attention encoding unit includes a query weight matrix, a key weight matrix, and a value weight matrix. The convolution module is used to input the structural features into the structural convolutional layer for convolution to obtain convolutional structural features; The encoding module is used to perform the following steps in each of the attention encoding units: The sequence features are respectively input into the query weight matrix, the key weight matrix, and the value weight matrix for linear transformation to obtain query features, key features, and value features; the query features, key features, and convolutional structure features are fused to obtain a first fused feature; the first fused feature is fused with the value features to obtain a second fused feature; The splicing module is used to splice the second fusion features output by all the attention encoding units to obtain spliced ​​features; The feature adjustment module is used to adjust the splicing features based on the feature adjustment module to obtain the identification result of the modification site in the RNA sequence to be identified.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images