Rice m based on feature extraction and fusion 6 A methylation site prediction method and system

Abstract

The present invention provides a rice m based on feature extraction and fusion 6 A methylation site prediction method and system, the method comprises: obtaining the rice m 6 A dataset, using label encoding and padding strategy to construct a maximum length padded sequence to obtain variable length m 6 The initial characteristics of the A sequence will become m long 6 The initial features of the A sequence are input to the trained rice m 6 A methylation site prediction model was used to obtain rice m 6 A methylation site prediction results. The present invention solves the existing rice m 6 A methylation site prediction method adopts a filling method for variable-length sequences, resulting in sparse and invalid features, and directly stacking features from different sources, which leads to rice m 6 A technical problem of low prediction accuracy and poor prediction effect of methylation sites.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for extracting and fusion-based rice m 6 A methylation site prediction method and system. Background Art

[0002] Epigenetics, first proposed by geneticist Conrad Hal Waddington, refers to stable, heritable changes in non-deoxyribonucleic acid sequences, such as chemical modifications. These mechanisms primarily include DNA methylation, RNA methylation, demethylation, chromatin remodeling, and histone modifications. Recently, epigenetic changes have been shown to play a crucial role in inducing phenotypic diversity and play an integral role in the development and biological processes of multicellular organisms.

[0003] There are more than 200 post-transcriptional epigenetic modifications in eukaryotic RNAs, such as N 6 -methyladenosine (m 6 A), N 1 -methyladenosine (m 1 A), 5-methylcytidine (m 5 C), 1-methylguanosine (m 1 G), pseudouridine (ψ), etc. The most common epigenetic modification is m 6 A, which is widely present in messenger RNAs, microRNAs, long non-coding RNAs, transfer RNAs and ribosomal RNAs. Methyltransferase complexes (Writers), demethylases (Erasers) and m 6 A binding protein (Readers) is m 6 A is the main component of the modification system. Writers and Erasers add and delete a methyl group (-CH3) on the amino group (-NH2) at the 6th position of adenosine in RNAs, respectively. Readers recognize m 6 A site and plays a specific regulatory role. In rice, m 6 A is involved in growth, development, and responses to biotic and abiotic stresses. 6 A regulates the early degeneration of microspores in the vacuolar pollen stage of rice. 6 A is dynamically regulated during rice-plant virus interactions. Cheng et al. found that rice m 6 A may be related to the abnormal root development induced by cadmium stress.

[0004] In plants, methyl-RNA immunoprecipitation and sequencing (MeRIP-seq, also known as m 6 A-seq) can detect m 6However, the MeRIP-seq method can only detect mA sites with a resolution of 100 to 200 nucleotides. 6 A transcript, cannot satisfy m 6 A base resolution requirement for editing. In order to achieve m 6 To detect the single-base resolution of A site, researchers have made some improvements to the MeRIP-seq technology, including photocrosslinking assisted m 6 A sequencing strategy (PA-m 6 A-seq), m 6 A single nucleotide resolution cross-linking and immunoprecipitation technique (miCLIP), m 6 A cross-linking immunoprecipitation sequencing (m 6 A-CLIP-seq) etc. However, m 6 A sites are mainly concentrated in organs and tissues with high RNA metabolism, and the sample size of these tissues is often small. Therefore, the above technology cannot be biologically replicated, so it is impossible to accurately detect m in tissues or cells. 6 The deamination adjacent RNA modification target sequencing (DART-seq) method only requires 10 ng of total RNAs, which is about 50 times less than that required by the MeRIP-seq method. The third generation sequencing technology has rapidly developed the detection of m 6 A site method, including single-molecule real-time sequencing and single-molecule nanopore sequencing. In 2020, Parker et al. successfully used nanopore direct RNA sequencing technology to map the Arabidopsis m 6 A. This technique is well suited for studying small samples and has the potential to accelerate the development of m 6 A research, and will have a great impact on future m 6 A single-base resolution modification and editing provides great assistance.

[0005] Although biological experiments 6 A site detection technology is constantly being improved and perfected, but it still requires a lot of manpower, material and financial resources, so the development of m 6 Computational methods related to A site identification are necessary.

[0006] At present, researchers have proposed some m 6 A site prediction method. For example, the existing invention patent application document with publication number CN111161793A is “N site prediction method based on stacking integration in RNA”. 6-Methyladenosine modification site prediction method" proposed the StackRAM method. The method includes: extracting RNA sequence features of three species, Saccharomyces cerevisiae, humans and Arabidopsis thaliana (sequence lengths are 101, 41 and 21, respectively, and trained separately) through 6 feature extraction methods, namely binary coding, chemical properties, nucleotide frequency, k-mer nucleotide frequency, pseudo-dinucleotide composition and position-specific trinucleotide tendency, and obtaining the initial feature space of the original data set through feature fusion; using the elastic network to reduce the dimension, eliminate redundant and noise features, retain important features related to model classification, and obtain the best feature set; input the optimal feature subset and the corresponding category label into the stacking ensemble for model training, and evaluate the prediction performance of the model in combination with the evaluation index to obtain the prediction model; input the RNA sequence to be predicted in the test set into the prediction model to predict m 6 A site and output. Existing public literature "m based on sequence features and graph embedding information 6 A site identification method", which also targets Saccharomyces cerevisiae, Arabidopsis thaliana and human species, using m 6 Seven sequence feature encoding methods are commonly used in the computational method of A modification site identification: nucleotide composition conversion distribution characteristics, k-spaced nucleotide pair frequency characteristics, nucleotide physicochemical property density characteristics, nucleotide pair position specificity characteristics, bispectral Bayesian characteristics, electron-ion pseudo potential energy characteristics and pseudo k-member nucleotide composition characteristics. Then, a fast neighborhood similarity calculation method was introduced, and a sample similarity network was constructed based on sequence characteristics. The graph embedding features of each sample in the network were learned through three graph node embedding algorithms, SocDim, node2vec and GraRep. Finally, the sequence features and graph embedding features were spliced ​​into an input vector, and the computational model was trained based on the CatBoost classifier. This computational model was named m6AGE. This method was first used in m 6 The computational method for identifying A modification sites combines the traditional sequence feature encoding method and the graph node embedding algorithm, and uses three graph node embedding algorithms in the sample similarity network to fully learn the graph embedding features of each sample in an unsupervised manner. The existing public document "Study on m6A(m) methylation site prediction based on integrated deep learning" is based on high-confidence data (sequence length 41) from multiple tissues of mammals, namely humans, mice and rats, and combines the one-hot feature encoding method, nucleotide chemical properties and nucleotide density (also known as cumulative nucleotide frequency, ANF) into a new feature, which is used as the input of the model. A new computational method, im6APred, was constructed using integrated deep learning to accurately identify m 6A site. This model, im6APred, is based on a comprehensive evaluation of multiple classification methods, including four traditional classification algorithms, three deep learning methods, and their ensembles. Five cross-validation tests were performed to select the best classifier combination to achieve an effective superposition model.

[0007] However, the prediction of rice m 6 There are few methods for calculating the A site. Until 2021, after collecting and processing the first rice m 6 A dataset, Wang and colleagues proposed the first rice m 6 A computational method for site A prediction is called SMEP. 6 Sequence A is a variable length sequence. SMEP first uses padding and label encoding to obtain a maximum length padded sequence. After the embedding layer, SMEP uses multiple concatenated convolutional layers to extract high-order features. Finally, after the flattening layer, SMEP uses a multi-layer perceptron for the final m 6 As an effective computational method, SMEP is a method for predicting the A site of rice. 6 The identification of A sites has laid the foundation and greatly promoted the development of algorithms in this field. Recently, by stacking multi-source features generated by statistics and deep learning to enhance feature representation, Song et al. proposed an integrated learning framework based on the gradient boosting decision tree algorithm, called PEA-m 6 A, for rice m 6 Improvements in A site prediction have achieved good performance.

[0008] Through the above method, it can be found that other species m 6 A data length is a fixed length short sequence. For other species m 6 The coding features used by the A site prediction methods StackRAM, m6AGE, and im6APred can be divided into two categories: feature coding methods based on nucleotide composition and feature coding methods based on nucleotide frequency. Feature coding methods based on nucleotide composition type include binary coding, chemical properties, pseudo-dinucleotide composition, and position-specific trinucleotide tendency. These feature coding methods are applied to m6AGE of fixed-length short sequences in these other species. 6 A is simple and convenient in terms of data, but cannot be directly applied to rice m 6 A variable length sequence dataset, because this will result in unequal feature dimensions that cannot be input into the model. Feature encoding methods based on nucleotide frequency types, such as nucleotide frequency and k-mer nucleotide frequency, can compress variable length sequence encoding into fixed length feature dimensions, but will result in a large amount of sequence information being lost. Therefore, feature encoding methods based on nucleotide composition and nucleotide frequency cannot be well applied to rice m 6A site prediction. SMEP uses the padding operation to obtain the maximum length of the padding sequence to deal with rice m 6 A variable length sequence. However, although the padding operation can retain the complete sequence information, it will also make the sequence features sparse and invalid, thereby affecting feature extraction and reducing prediction accuracy. 6 A enhances rice m by stacking multi-source features generated by statistics and deep learning 6 A site feature, but simply stacking these features from different sources has limitations for feature enhancement. Furthermore, traditional machine learning methods used in the above methods, such as stacking ensembles and CatBoost classifiers, are not end-to-end models and are weak in handling unstructured data.

[0009] In summary, the existing rice m 6 A methylation site prediction method adopts a filling method for variable-length sequences, resulting in sparse and invalid features, and directly stacking features from different sources, which leads to rice m 6 A technical problem of low prediction accuracy and poor prediction effect of methylation sites. Summary of the Invention

[0010] The technical problem to be solved by the present invention is: how to solve the existing rice m 6 A methylation site prediction method adopts a filling method for variable-length sequences, resulting in sparse and invalid features, and directly stacking features from different sources, which leads to rice m 6 A technical problem of low prediction accuracy and poor prediction effect of methylation sites.

[0011] The present invention adopts the following technical solutions to solve the above technical problems: rice m 6 A methylation site prediction method includes:

[0012] S1, from m 6 In A-seq sequence data, obtain rice m 6 A dataset;

[0013] S2, perform input representation operation, and change rice m 6 In the A dataset, all variable-length sequences are padded to the preset maximum length to obtain the maximum-length padded sequence. The maximum-length padded sequence is then label-encoded to obtain the input representation result F. input ;

[0014] S3. Based on the input representation result F input , extract multi-core features;

[0015] S4. During the model training operation, perform a global-local feature fusion operation on the multi-core features to obtain a global-local fusion feature. Use the Adam optimizer to set the batch size, learning rate, weight decay, and training generations of the model training operation, and use the warm-up learning rate adjustment strategy to control the model training operation.

[0016] S5, downsample the global-local fusion features and embed them to obtain applicable features; among them, alternately use convolutional layers, pooling layers and residual connections to perform layer-by-layer downsampling and feature embedding to retain high-frequency important information, and use the contrastive learning loss function L cl Adjust the distance between similar samples and heterogeneous samples to optimize the feature optimization results;

[0017] S6. In the prediction stage, the feature optimization results are sent to the fully connected layer to obtain rice m 6 A methylation site prediction probability to obtain rice sequence m 6 Site A contains the judgment result, wherein the fully connected layer includes: σ function.

[0018] This paper introduces multi-core feature extraction and fusion to obtain dimensional information from different kernel spaces. By effectively fusing multi-core features, the paper enhances feature representation. Furthermore, by proposing a downsampling feature embedding method, the fused multi-core features are effectively compressed, ensuring effective information transfer and high computational efficiency.

[0019] The present invention realizes the rice m 6 Classification of A methylation sites. Experimental results show that compared with other methods, the present invention achieves better results on multiple data sets. At the same time, the model of the present invention has better robustness and generalization.

[0020] Specifically, compared with other existing species m 6 A prediction method is different, in order to be applied to rice m 6 For variable-length sequence data, the present invention does not use a feature encoding method based on nucleotide composition and nucleotide frequency, but instead uses category label encoding and padding operations to obtain fixed-dimensional features while retaining all complete sequence information.

[0021] In a more specific technical solution, in S3, the following logic is used to extract multi-core features:

[0022] O ebd =Cv1(Embed(F input ))

[0023] K1=Cv1(O ebd )

[0024] K2=Cv3(Cv1( ebd ))

[0025] K3=Cv5(Cv1( ebd ))

[0026] K4=Cv1(MP3( ebd ))

[0027] MK=concat(K1,K2,K3,K4)

[0028] Where, Embed represents the embedding layer, Cv1, Cv3, and Cv5 represent the convolution layers with kernel sizes of 1, 3, and 5, respectively, MP3 represents the maximum pooling layer with kernel size of 3, concat represents the concatenation operation, and ebd is the initial embedding vector, K1, K2, K3 and K4 are the features of different kernels obtained, and MK is the multi-kernel feature obtained.

[0029] The present invention extracts and fuses multi-core features to enhance feature representation, which means enhancing key information and suppressing sparse and invalid information. The present invention introduces multi-core feature extraction and fusion to extract key features across multiple receptive fields and reduce sparsity.

[0030] Specifically, with rice m 6 Different from the prediction method A, the present invention takes into account and alleviates the feature sparsity caused by the padding method, and correspondingly proposes a multi-core feature extraction function in the multi-core feature fusion module, which extracts multi-core features of different kernel sizes on the maximum length padding sequence, instead of using only one maximum length padding sequence like SMEP.

[0031] In more specific technical solutions, S3 also includes:

[0032] S31, input the result F input The initial embedding vector O is obtained through the embedding layer Embed and the first convolutional layer Cv1 ebd ;

[0033] S32, for the initial embedding vector O ebd , respectively use the first convolution layer Cv1, the second convolution layer Cv3 and the third convolution layer Cv5 to extract information of different spatial dimensions, and obtain the first kernel feature K1, the second kernel feature K2, the third kernel feature K3 and the fourth kernel feature K4;

[0034] S33 , performing concat processing on the first kernel feature K1 , the second kernel feature K2 , the third kernel feature K3 , and the fourth kernel feature K4 to obtain a multi-kernel feature MK.

[0035] In a more specific technical solution, S32 includes:

[0036] S321, in the initial acquisition stage of the first kernel feature K2 and the second kernel feature K3, executing the first convolutional layer Cv1 to adjust the number of channels;

[0037] S322, the fourth kernel feature K4 uses MP3 to obtain feature information of the channel dimension, and uses the first convolutional layer Cv1 to change the number of channels.

[0038] In a more specific technical solution, in S4, the following logic is used to perform a global-local feature fusion operation:

[0039] MK=Cv(MK)

[0040]

[0041]

[0042] GL=Cv(concat(G,L))

[0043]

[0044] Where σ represents the sigmoid function and AP represents the average pooling layer.

[0045] In a more specific technical solution, S4 includes:

[0046] S41, obtain the global attention weight of the multi-core feature MK through the average pooling layer;

[0047] S42, using the convolutional layer, obtain the local attention weight of the multi-core feature MK;

[0048] S43, multiplying the global attention weight and the local attention weight by the elements of the multi-core feature MK respectively to obtain the global feature G and the local feature L of the multi-core feature MK;

[0049] S44, splicing the global feature G and the local feature L, and the initial embedding vector O ebd Perform residual connection operation Processing to obtain global-local fusion feature O GL .

[0050] This paper focuses on the global-local features of multiple cores for feature fusion, ensuring efficient feature migration and reducing information loss. Specifically, to better integrate features from different sources, the present invention proposes a global-local fusion function in the multi-core feature fusion module, rather than directly stacking features from different sources like PEA-m6A.

[0051] In a more specific technical solution, in S5, the following contrastive learning loss function L is usedcl Optimize to obtain feature space:

[0052]

[0053] Where Z indicates whether two samples i and j belong to the same class, with 0 assigned to the same class and 1 to different classes; D represents the Euclidean distance; and M represents the optimized distance threshold.

[0054] The present invention proposes a downsampling feature embedding module, which uses downsampling to embed features, while reducing feature maps and improving computational memory efficiency, while ensuring efficient and effective transmission of key features, thereby improving rice m 6 A site prediction accuracy. The present invention introduces downsampling feature embedding for effective feature compression to alleviate the invalid features caused by padding and the complexity and low computational efficiency of high-dimensional features.

[0055] In a more specific technical solution, S6 includes:

[0056] S61, determine rice m 6 A. Whether the predicted probability of the methylation site is less than the preset threshold;

[0057] S62, if yes, then determine that the current rice sequence does not contain m 6 A site;

[0058] S63, if not, then determine that the current rice sequence contains m 6 A site.

[0059] In a more specific technical solution, in S6, the following binary cross entropy loss function is used as the classification objective function to perform the minimization optimization operation:

[0060]

[0061] Where N represents the batch size, y n is the true label, is the predicted probability.

[0062] In a more specific technical solution, rice m 6 A methylation site prediction system includes:

[0063] Dataset acquisition module, used to obtain data from m 6 In A-seq sequence data, obtain rice m 6 A dataset;

[0064] Input representation module, used to perform input representation operation, rice m 6In the A dataset, all variable-length sequences are padded to the preset maximum length to obtain the maximum-length padded sequence. The maximum-length padded sequence is then label-encoded to obtain the input representation result F. input ,The input representation module is connected to the dataset acquisition module;

[0065] Multi-core feature extraction module, used to represent the result F based on the input input ,Extract multi-core features, the multi-core feature extraction module is connected to the input representation module;

[0066] The feature fusion module is used to perform global-local feature fusion operations on multi-core features during model training operations to obtain global-local fusion features. The Adam optimizer is used to set the batch size, learning rate, weight decay, and training generations of the model training operation, and a warm-up learning rate adjustment strategy is used to control the model training operation. The feature fusion module is connected to the multi-core feature extraction module.

[0067] The downsampling feature embedding module is used to downsample the global-local fusion features and embed them to obtain applicable features. Convolutional layers, pooling layers, and residual connections are used alternately to perform layer-by-layer downsampling and feature embedding to retain high-frequency important information. The contrastive learning loss function L is used to cl Adjust the distance between similar samples and heterogeneous samples to optimize the feature optimization results, and connect the downsampling feature embedding module with the feature fusion module;

[0068] The site contains a judgment module, which is used to send the feature optimization results to the fully connected layer in the prediction stage to obtain rice m 6 A methylation site prediction probability to obtain rice sequence m 6 A site contains the judgment result, wherein the fully connected layer includes: a σ function, and the site contains the judgment module is connected with the downsampling feature embedding module.

[0069] Compared with the prior art, the present invention has the following advantages:

[0070] This paper introduces multi-core feature extraction and fusion to obtain dimensional information from different kernel spaces. By effectively fusing multi-core features, the paper enhances feature representation. Furthermore, by proposing a downsampling feature embedding method, the fused multi-core features are effectively compressed, ensuring effective information transfer and high computational efficiency.

[0071] The present invention realizes the rice m 6 Classification of A methylation sites. Experimental results show that compared with other methods, the present invention achieves better results on multiple data sets. At the same time, the model of the present invention has better robustness and generalization.

[0072] The present invention extracts and fuses multi-core features to enhance feature representation, which means enhancing key information and suppressing sparse and invalid information. The present invention introduces multi-core feature extraction and fusion to extract key features across multiple receptive fields and reduce sparsity.

[0073] The present invention performs feature fusion by focusing on the global-local features of multiple cores, ensuring effective feature migration and reducing information loss.

[0074] The present invention uses downsampling to embed features, which reduces the feature map and improves the computational memory efficiency while ensuring the efficient and effective transmission of key features. 6 A site prediction accuracy. The present invention introduces downsampling feature embedding for effective feature compression to alleviate the high-dimensional features and low computational efficiency caused by invalid padding.

[0075] Based on the defects of the existing technology, the present invention proposes a rice m 6 A prediction method for methylation sites. In order to alleviate the feature sparsity caused by the filling operation, the rice m 6 After the initial features of A, the patent builds a multi-core feature fusion module, which includes a multi-core feature extraction function and a global-local fusion function. The multi-core feature extraction function mines the maximum length filling m by different kernel sizes. 6 The essential information of different scales in the A sequence enriches the extracted features. The global-local fusion function performs effective feature information fusion and transmission by focusing on the global channel and local space of the extracted multi-core features, thereby improving the rice m 6 A. The accuracy of methylation site prediction. Considering the complexity and computational efficiency of high-dimensional features caused by invalid padding, the patent designs a downsampling feature embedding module. By alternating between convolutional layers, pooling layers, and residual connections, it performs layer-by-layer downsampling and feature embedding, optimizing the compression of the feature space while retaining high-frequency important information, achieving accurate feature expression and efficient computational performance.

[0076] The present invention solves the existing rice m 6 A methylation site prediction method adopts a filling method for variable-length sequences, resulting in sparse and invalid features, and directly stacking features from different sources, which leads to rice m 6 A technical problem of low prediction accuracy and poor prediction effect of methylation sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 The rice m based on feature extraction and fusion of Example 1 of the present invention 6 A. Schematic diagram of the basic steps of the methylation site prediction method;

[0078] Figure 2This is a schematic diagram of the multi-core feature extraction structure of Example 2 of the present invention;

[0079] Figure 3 Schematic diagram of the global-local feature fusion structure of Example 2 of the present invention;

[0080] Figure 4 Schematic diagram of the downsampling feature embedding structure of embodiment 2 of the present invention;

[0081] Figure 5 Schematic diagram of the prediction process of Example 2 of the present invention;

[0082] Figure 6 The figure shows the ROC and PR curves of the experimental prediction results of Example 2 of the present invention. DETAILED DESCRIPTION

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0084] Example 1

[0085] like Figure 1 As shown, the rice m based on feature extraction and fusion provided by the present invention 6 A methylation site prediction method includes the following basic steps:

[0086] Get the rice m to be predicted 6 A dataset, using label encoding and padding strategy to construct a maximum length padded sequence to obtain variable length m 6 The initial characteristics of the A sequence will become m long 6 The initial features of the A sequence are input to the trained rice m 6 A methylation site prediction model was used to obtain rice m 6 A. Methylation site prediction results;

[0087] The trained rice m 6 A methylation site prediction model is obtained by the following method:

[0088] S1. Obtain rice m 6 A's dataset;

[0089] In this embodiment, rice m 6 A sequence collected from m 6 A-seq sequence data was obtained after removing homology deviation and redundancy through CD-HIT.

[0090] S2. In the input representation process, rice m 6 All variable length sequences of A are padded to the maximum length to obtain the maximum length padded sequence, and then label encoding is performed to obtain the input representation F input ;

[0091] In this embodiment, padding and label encoding are used to obtain the maximum length padded sequence, so that rice m 6 The characteristic of the A sequence indicates that the maximum integrity is retained. Specifically, the length of the aforementioned maximum filling sequence can be set to, for example, 800 as required.

[0092] S3, based on the input representation result F obtained in the above step S2 input Extract multi-core features;

[0093] In this embodiment, the number of output channels of the first convolutional neural network is 128, and the number of output channels of the remaining convolutional neural networks is 32.

[0094] In this embodiment, the formula for extracting multi-core features is as follows:

[0095] O ebd =Cv1(Embed(F input ))

[0096] K1=Cv1(O ebd )

[0097] K2=Cv3(Cv1( ebd ))

[0098] K3=Cv5(Cv1( ebd ))

[0099] K4=Cv1(MP3( ebd ))

[0100] MK=concat(K1,K2,K3,K4)

[0101] Among them, Embed represents the embedding layer, Cv1, Cv3 and Cv5 represent the convolution layers with the convolution kernel size set to 1, 3 and 5 respectively, MP3 represents the maximum pooling layer with the kernel size set to 3, concat represents the splicing operation, ebd is the initial embedding vector, K1, K2, K3 and K4 are the features of different kernels obtained, and MK is the multi-kernel feature obtained;

[0102] Specifically, in F input The initial embedding vector O is obtained by Embed and Cv1 ebd . For O ebdCv1, Cv3, and Cv5 are used to extract information from different spatial dimensions to obtain K1, K2, and K3, respectively. K2 and K3 initially execute Cv1 to reduce the number of channels, thereby reducing the number of parameters and model complexity. K4 uses MP3 to obtain feature information in the channel dimension and uses Cv1 to change the number of channels. The features of these different kernels are concatenated to obtain the final MK.

[0103] S4. During the training process, the multi-core features obtained in the above step S3 are subjected to global-local feature fusion;

[0104] During the training process of this embodiment, the Adam optimizer was used. The batch size, learning rate, weight decay, and number of training generations were set as needed, and a warm-up learning rate adjustment strategy was used. Specifically, the batch size was set to 128, the learning rate was set to 0.001, the weight decay was set to 1e-6, and a warm-up learning rate adjustment strategy was used to prevent the model from converging too quickly in the early stages of training and thus falling into a local optimal solution. The number of training generations was set to 100.

[0105] The convolution kernel sizes of the convolutional neural network of this embodiment can be set to, for example, 1, 5, 5, 5, 1 respectively; the number of output channels can be set to, for example, 32, 8, 32, 1, 128 respectively.

[0106] In this embodiment, the formula for performing the aforementioned global-local feature fusion operation is as follows:

[0107] MK=Cv(MK)

[0108]

[0109]

[0110] GL=Cv(concat(G,L))

[0111]

[0112] Among them, σ represents the sigmoid function, AP represents the average pooling layer;

[0113] Specifically, the global attention weight and local attention weight of MK are obtained by the average pooling layer and the convolution layer respectively, and then these weights are multiplied with the MK elements to obtain the global features and local features of MK, namely G and L. After G and L are concatenated, they are combined with O ebd Perform residual connection Reduce distortion to obtain the final global-local fusion feature O GL .

[0114] S5. During the training process, the fusion features obtained in the above step S4 are downsampled and embedded to obtain high-level features;

[0115] In this embodiment, in order to solve the problem of high feature dimension and low computational efficiency caused by invalid padding, convolutional layers, pooling layers and residual connections are used alternately to perform layer-by-layer downsampling and feature embedding; while reducing the size of the feature map, the consistency of the feature map space is ensured, and high-frequency important information is retained. In addition, the contrastive learning loss function L is used cl The feature space obtained by optimization is as follows:

[0116]

[0117] Among them, Z indicates whether two samples i and j belong to the same class, the same class is assigned 0, and different classes are assigned 1; D represents the Euclidean distance; M represents the optimized distance threshold;

[0118] Specifically, L cl The function is to achieve feature optimization during the training process by making the distances between similar samples closer and the distances between heterogeneous samples farther apart.

[0119] In this embodiment, the optimized distance threshold M may be set to, for example, 2.

[0120] S6. In the prediction stage, the feature optimization results obtained in the above step S5 are fed into the fully connected layer with the σ function to obtain the rice m 6 A methylation site prediction probability; when the prediction probability is less than the threshold, it is determined that the rice sequence does not contain m 6 A site, otherwise the rice sequence contains m 6 A site. And, the binary cross entropy loss function is used as the classification objective function for minimization optimization:

[0121]

[0122] Among them, N represents the batch size, y n is the true label, is the predicted probability.

[0123] In this embodiment, the number of neurons in the fully connected layer can be set to, for example, 12800. The threshold of the prediction probability can be set to, for example, 0.5.

[0124] Example 2

[0125] In this embodiment, the trained rice m is obtained. 6A methylation site prediction model. In the input representation process, we first use the rice m 6 A public dataset. The positive samples of this dataset are extracted from the leaves of Japanese naked seedlings of rice. 6 A-seq peak sequence data, sample length is 20 to 800. Negative samples are obtained by cutting off the upper and lower parts of the corresponding positive samples without m 6 They also used CD-HIT[2]FuL,Niu B,Zhu Z,et al.CD-HIT:accelerated for clustering the next-generation sequencing data[J].Bioinformatics,2012,28(23):3150-3152 to reduce homology bias and redundancy in the dataset. After data cleaning and deduplication, 80% of the positive and negative samples were randomly selected from the dataset as a benchmark dataset to evaluate the model performance. In addition, in order to verify the generalization ability of the model, the remaining samples were used as an independent test set. The information of the aforementioned dataset is shown in Table 1 below:

[0126] dataset positive samples negative samples total samples Benchmark datasets 19,844 39,692 59,536 Independent test set 4,963 9,923 14,886

[0127] Table 1

[0128] In this embodiment, the data set is divided into a training set and a validation set according to a ratio of 4:1, and a 5-fold cross validation is performed during the experiment.

[0129] Specifically, rice m 6 A sequence is variable length data. In order to retain the complete sequence information and incorporate it into the model, this embodiment fills all sequences to a maximum length of 800. At the same time, in order to adapt to the model, this embodiment uses label encoding to convert rice m 6 A sequence is converted to a numeric representation, which assigns a unique numeric identifier to each character class.

[0130] In this example, assuming the maximum sequence length is 10, the sequence "ATTCG" consists of four base types: A, T, G, and C. "P" is used to represent padding, so the padded sequence is "PPPPPATTCG." The five base types (P, A, T, G, and C) are assigned numbers 0, 1, 2, 3, and 4, respectively, according to the label encoding. Therefore, this padded sequence can be represented as "0,0,0,0,0,1,2,2,4,3."

[0131] like Figure 2 As shown, during the training process, the obtained rice m 6 A sequence input representation for multi-core feature extraction.

[0132] Specifically, rice m 6 A sequence input represents F input After that, the initial embedding vector is obtained through the embedding layer and the convolution layer. For this initial embedding vector, convolution layers with different kernel sizes and maximum pooling layers are used respectively to extract information in different spatial dimensions to obtain multi-kernel features. The convolution layer with a kernel size of 1 is used to reduce the number of channels, thereby reducing the number of parameters and model complexity. The relevant calculation formula is as follows:

[0133] O ebd =Cv1(Embed(F input ))

[0134] K1=Cv1(O ebd )

[0135] K2=Cv3(Cv1( ebd ))

[0136] K3=Cv5(Cv1( ebd ))

[0137] K4=Cv1(MP3( ebd ))

[0138] MK=concat(K1,K2,K3,K4)

[0139] Among them, Embed represents the embedding layer, Cv1, Cv3 and Cv5 represent the convolution layers with the convolution kernel size set to 1, 3 and 5 respectively, MP3 represents the maximum pooling layer with the kernel size set to 3, concat represents the splicing operation, ebd is the initial embedding vector, K1, K2, K3 and K4 are the features of different kernels obtained, and MK is the multi-kernel feature obtained.

[0140] like Figure 3 As shown, in this embodiment, the obtained multi-core features are subjected to global-local feature fusion.

[0141] Specifically, the global attention weights and local attention weights of the multi-core features are obtained through the average pooling layer and the convolution layer, respectively. These weights are then element-wise multiplied with the multi-core features to obtain their global and local features. After concatenating the global and local features, a residual connection is performed with the initial embedding vector obtained previously to reduce distortion and obtain the final global-local fusion feature. The relevant calculation formula is as follows:

[0142] MK=Cv(MK)

[0143]

[0144]

[0145] GL=Cv(concat(G,L))

[0146]

[0147] Among them, σ represents the sigmoid function, AP represents the average pooling layer; Indicates element-wise multiplication; Represents residual connection. G and L represent global features and local features respectively. GL It is a global-local fusion feature.

[0148] like Figure 4 As shown, in this embodiment, downsampling feature embedding is performed on the obtained fusion features to obtain high-level features.

[0149] Specifically, to address the issues of high feature dimensionality and low computational efficiency caused by ineffective padding, we alternately use convolutional layers, pooling layers, and residual connections to perform layer-by-layer downsampling and feature embedding. This achieves the goal of reducing the size of feature maps while maintaining spatial consistency and preserving high-frequency important information. Furthermore, we optimize the obtained feature space using the contrastive learning loss function, as follows:

[0150]

[0151] Among them, Z indicates whether two samples i and j belong to the same type, the same type is assigned a value of 0, and different types are assigned a value of 1; D represents the Euclidean distance; M represents the optimized distance threshold.

[0152] The contrastive learning loss function L used in this embodiment is cl , can achieve feature optimization during the training process by making the distances between similar samples closer and the distances between heterogeneous samples farther apart.

[0153] like Figure 5As shown, in the prediction process steps of this embodiment, first, for the input sequence feature representation obtained by feature encoding, the multi-core feature MK is extracted; MK is subjected to global and local feature fusion to obtain GL ; for O GL , by downsampling feature embedding, using contrastive learning loss function optimization to obtain high-level features; flattening the optimized features and sending them to the fully connected layer with σ function to obtain rice m 6 A methylation site prediction probability; when the prediction probability is less than the threshold, it is determined that the rice sequence does not contain m 6 A site, otherwise the rice sequence contains m 6 A site.

[0154] In this embodiment, a binary cross entropy loss function is used as the classification objective function for minimization optimization:

[0155]

[0156] Among them, N represents the batch size, y n is the true label, is the predicted probability.

[0157] In this embodiment, the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPR) are used to evaluate the performance indicators of the model.

[0158] In this example, in order to verify the present invention to predict rice m 6 The effect of A methylation sites was studied by using the two types of datasets mentioned above, namely the benchmark dataset and the independent test set, and comparing them with the prediction of m 6 The main methods of site A, SMEP and PEA-m6A, were compared. For the benchmark dataset, the experiment set 5-fold cross validation, and the average of the 5-fold cross validation results was taken as the final result (see Figure 6 shown). Figure 6 The experimental results show that the comprehensive performance of the present invention is better than other existing methods, indicating that the present invention can more effectively predict rice m 6 A methylation site. Specifically, the present invention achieved an AUC of 0.9038 and an AUPR of 0.8201, representing a 9.55% improvement in overall performance compared to the suboptimal method, SMEP. To further demonstrate the superiority and generalizability of the present invention, we compared it with existing prediction methods on an independent test set. (See Table 2 below for the results, where the best results are bolded and the suboptimal results are underlined.)

[0159] method AUC AUPR The present invention 0.8941 0.8068 SMEP <![CDATA[ 0.8687 ]]> <![CDATA[ 0.7602 ]]> PEA-m6A 0.8400 0.7074

[0160] Table 2

[0161] The experimental results in Table 2 above demonstrate that our method demonstrates greater effectiveness than the comparison methods on an independent test set, demonstrating good generalization capabilities. Specifically, our method achieves 2.54% and 4.66% improvements in AUC and AUPR, respectively, over the next-best method, SMEP. Notably, the results on the independent test set are comparable to the 5-fold cross-validation results, demonstrating that all methods exhibit robustness and transferability.

[0162] The present invention has achieved better results on the aforementioned multiple data sets, with higher prediction accuracy. At the same time, the experimental results also show that the model used in the present invention has better robustness and generalization ability, and can be extended to rice m 6 A site prediction is a similar problem.

[0163] The present invention uses a global-local fusion of multi-core feature extraction and downsampling feature embedding method to reduce the sparsity, high-dimensional features and low computational efficiency problems caused by padded sequence encoding. Specifically, since variable-length sequences contain a large amount of information, label encoding and padding strategies are used to construct maximum-length padded sequences to promote feature learning. However, lengthy and invalid sequence padding will lead to sparse and redundant features and weaken the model's ability to capture meaningful sequence-level information. Therefore, the multi-core feature extraction method we proposed can extract key features across multiple receptive fields to reduce the sparsity problem caused by maximum-length padded sequence encoding. At the same time, through the fusion of global and local features, the features extracted from the multi-core can be effectively migrated. In addition, in order to solve the problems of high feature dimension and low computational efficiency caused by invalid sequence padding, a downsampling feature embedding method is introduced to effectively compress the features and improve computational performance. Experimental results show that the present invention has good predictive performance, robustness and generalization ability, which is conducive to enhancing feature representation and extraction and transmitting efficient information.

[0164] The prediction method of the present invention has a strong feature enhancement effect on data. The prediction method of the present invention can be used in the future rice m 6 It plays a beneficial role in the prediction of A methylation sites and has application prospects in more similar fields.

[0165] In summary, this paper introduces multi-core feature extraction and fusion to obtain dimensional information from different kernel spaces, while also enhancing feature representation by effectively fusing multi-core features. Furthermore, this paper proposes a downsampling feature embedding method to effectively compress the fused multi-core features, ensuring effective information transfer and high computational efficiency.

[0166] The present invention realizes the rice m 6Classification of A methylation sites. Experimental results show that compared with other methods, the present invention achieves better results on multiple data sets. At the same time, the model of the present invention has better robustness and generalization.

[0167] The present invention extracts and fuses multi-core features to enhance feature representation, which means enhancing key information and suppressing sparse and invalid information. The present invention introduces multi-core feature extraction and fusion to extract key features across multiple receptive fields and reduce sparsity.

[0168] The present invention performs feature fusion by focusing on the global-local features of multiple cores, ensuring effective feature migration and reducing information loss.

[0169] The present invention uses downsampling to embed features, which reduces the feature map and improves the computational memory efficiency while ensuring the efficient and effective transmission of key features. 6 A site prediction accuracy. The present invention introduces downsampling feature embedding for effective feature compression to alleviate the high-dimensional features and low computational efficiency caused by invalid padding.

[0170] The present invention solves the existing rice m 6 A methylation site prediction method adopts a filling method for variable-length sequences, resulting in sparse and invalid features, and directly stacking features from different sources, which leads to rice m 6 A technical problem of low prediction accuracy and poor prediction effect of methylation sites.

[0171] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. Rice m based on feature extraction and fusion 6 A methylation site prediction method, characterized in that The method comprises: S1, from m 6 In A-seq sequence data, obtain rice m 6 A dataset; S2, perform input representation operation, and change rice m 6 In the A dataset, all variable-length sequences are padded to the preset maximum length to obtain the maximum-length padded sequence. The maximum-length padded sequence is then label-encoded to obtain the input representation result F. input ; S3. Based on the input representation result F input , extract multi-core features; In S3, the following logic is used to extract multi-core features: O ebd =Cv1(Embed(F input )) K1=Cv1(O ebd ) K2=Cv3(Cv1(O ebd )) K3=Cv5(Cv1(O ebd )) K4=Cv1(MP3(O ebd )) MK=concat(K1,K2,K3,K4) Where, Embed represents the embedding layer, Cv1, Cv3, and Cv5 represent the convolution layers with kernel sizes of 1, 3, and 5, respectively, MP3 represents the maximum pooling layer with kernel size of 3, concat represents the concatenation operation, and ebd is the initial embedding vector, K1, K2, K3 and K4 are the features of different kernels obtained, and MK is the multi-kernel feature obtained; S3 also includes: S31, input the result F input The initial embedding vector O is obtained through the embedding layer Embed and the first convolutional layer Cv1 ebd ; S32, for the initial embedding vector O ebd , respectively use the first convolution layer Cv1, the second convolution layer Cv3 and the third convolution layer Cv5 to extract information of different spatial dimensions, and obtain the first kernel feature K1, the second kernel feature K2, the third kernel feature K3 and the fourth kernel feature K4; S32 includes: S321, in the initial acquisition stage of the first kernel feature K2 and the second kernel feature K3, executing the first convolution layer Cv1 to adjust the number of channels; S322, the fourth kernel feature K4 uses MP3 to obtain feature information of the channel dimension, and uses the first convolutional layer Cv1 to change the number of channels; S33, performing concat processing on the first core feature K1, the second core feature K2, the third core feature K3 and the fourth core feature K4 to obtain a multi-core feature MK; S4. During the model training operation, perform a global-local feature fusion operation on the multi-core features to obtain a global-local fusion feature. Use the Adam optimizer to set the batch size, learning rate, weight decay, and training generations of the model training operation, and use the warm-up learning rate adjustment strategy to control the model training operation. S4 includes: S41, obtain the global attention weight of the multi-core feature MK through the average pooling layer; S42. Use the convolutional layer to obtain the local attention weight of the multi-core feature MK; S43, multiplying the global attention weight and the local attention weight by the elements of the multi-core feature MK respectively to obtain the global feature G and the local feature L of the multi-core feature MK; S44, splicing the global feature G and the local feature L, and the initial embedding vector O ebd Perform residual connection operation ⊕ to obtain global-local fusion feature O GL ; S5, downsample the global-local fusion features and embed them to obtain applicable features; among them, alternately use convolutional layers, pooling layers and residual connections to perform layer-by-layer downsampling and feature embedding to retain high-frequency important information, and use the contrastive learning loss function L cl Adjust the distance between similar samples and heterogeneous samples to optimize the feature optimization results; S6. In the prediction stage, the feature optimization results are sent to the fully connected layer to obtain rice m 6 A methylation site prediction probability to obtain rice sequence m 6 Site A contains the judgment result, wherein the fully connected layer includes: σ function.

2. The rice m based on feature extraction and fusion according to claim 1 6 A methylation site prediction method, characterized in that In S4, the global-local feature fusion operation is performed using the following logic: MK=Cv(MK) GL=Cv(concat(G,L)) Where σ represents the sigmoid function and AP represents the average pooling layer.

3. The rice m based on feature extraction and fusion according to claim 1 6 A methylation site prediction method, characterized in that In S5, the following contrastive learning loss function L is used cl Optimize to obtain feature space: Where Z indicates whether two samples i and j belong to the same class, with 0 assigned to the same class and 1 to different classes; D represents the Euclidean distance; and M represents the optimized distance threshold.

4. The rice m based on feature extraction and fusion according to claim 1 6 A methylation site prediction method, characterized in that The S6 includes: S61, determining the rice m 6 A. Whether the predicted probability of the methylation site is less than the preset threshold; S62, if yes, then determine that the current rice sequence does not contain m 6 A site; S63, if not, then determine that the current rice sequence contains the m 6 A site.

5. The rice m based on feature extraction and fusion according to claim 1 6 A methylation site prediction method, characterized in that In S6, the following binary cross entropy loss function is used as the classification objective function to perform the minimization optimization operation: Where N represents the batch size, y n is the true label, is the predicted probability.

6. Rice m based on feature extraction and fusion 6 A methylation site prediction system for performing the rice m 6 A methylation site prediction method, characterized in that The system comprises: Dataset acquisition module, used to obtain data from m 6 In A-seq sequence data, obtain rice m 6 A dataset; Input representation module, used for performing input representation operation, the rice m 6 In the A dataset, all variable-length sequences are padded to the preset maximum length to obtain the maximum-length padded sequence. The maximum-length padded sequence is label-encoded to obtain the input representation result F. input , the input representation module is connected to the data set acquisition module; Multi-core feature extraction module, used to represent the result F based on the input input , extracting multi-core features, the multi-core feature extraction module is connected to the input representation module; A feature fusion module is used to perform a global-local feature fusion operation on the multi-core features during the model training operation to obtain a global-local fusion feature; an Adam optimizer is used to set the batch size, learning rate, weight decay, and training generations of the model training operation, and a warm-up learning rate adjustment strategy is used to control the model training operation. The feature fusion module is connected to the multi-core feature extraction module; The downsampling feature embedding module is used to downsample the global-local fusion features and embed them to obtain applicable features. The convolutional layer, pooling layer and residual connection are used alternately to perform layer-by-layer downsampling and feature embedding to retain high-frequency important information. The contrastive learning loss function L is used to cl Adjusting the distance between similar samples and the distance between heterogeneous samples to optimize the feature optimization result, the downsampling feature embedding module is connected to the feature fusion module; The site includes a judgment module, which is used to send the feature optimization results to the fully connected layer in the prediction stage to obtain rice m 6 A methylation site prediction probability to obtain rice sequence m 6 A site contains a judgment result, wherein the fully connected layer includes: a σ function, and the site inclusion judgment module is connected to the downsampling feature embedding module.

Images