A model training and viral RNA editing site prediction method, device and product

Abstract

This specification provides a method, apparatus, and product for model training and prediction of viral RNA editing sites. The model training method includes: acquiring target domain sample data; inputting the target domain sample data into a pre-trained human genome editing site prediction model; extracting biological semantic features, local motif features, and spatial structural features from the target domain sample data; fusing the extracted features to obtain fused features; training each point in the target domain sample data as a predicted viral RNA editing site based on the fused features; and adjusting the model parameters of the editing site prediction model with the goal of minimizing the predicted probability of negative samples as viral RNA editing sites and maximizing the predicted probability of positive samples as viral RNA editing sites, thereby obtaining a viral RNA editing site prediction model.

Description

Technical Field

[0001] This specification relates to the field of computer technology in one or more embodiments, and more particularly to a method, apparatus, and product for model training and prediction of viral RNA editing sites. Background Technology

[0002] RNA editing is an important post-transcriptional modification process that plays a crucial role in host antiviral immunity and viral adaptive evolution. Adenosine deaminases acting on RNA (ADARs) mediate A-to-I editing, altering the base information of viral RNA sequences post-transcriptionally, thereby affecting viral replication, protein expression, immune recognition, and the formation of genetic diversity. For RNA viruses such as SARS-CoV-2 and HIV, accurate identification of RNA editing sites is essential for revealing host-virus interaction mechanisms, viral mutation patterns, and potential intervention targets. Therefore, establishing methods for predicting viral RNA editing sites has become a vital requirement in virology analysis and bioinformatics research.

[0003] However, existing RNA editing site prediction methods are mostly built based on human RNA data and are primarily geared towards the human genome. When these methods are applied to viral RNA, due to the significant differences between the host and the virus, existing models often struggle to maintain stable predictive performance and fail to meet the practical needs of identifying viral RNA editing sites. Summary of the Invention

[0004] In view of the above, one or more embodiments of this specification provide the following technical solutions: According to a first aspect of one or more embodiments of this specification, a model training method is proposed, comprising: Obtain target domain sample data; wherein, the target domain sample data is viral RNA sequence fragments, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites; The target domain sample data is input into a pre-trained human RNA editing site prediction model to extract biological semantic features, local motif features, and spatial structure features of the target domain sample data. The extracted features are then fused to obtain fused features. Based on the fused features, the predicted probability of each point in the target domain sample data as a viral RNA editing site is determined. With the goal of minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site, the model parameters of the editing site prediction model are adjusted to obtain an editing site prediction model for viral RNA.

[0005] Optionally, the positive viral RNA editing site sample includes: a viral RNA sequence fragment that satisfies the ADAR editing preference sequence motif for RNA-acting adenosine deaminase; The negative samples of non-edited viral RNA sites include: viral RNA sequence fragments that are rejected by ADAR, excluding the ADAR editing-preferred sequence motif, within the context window length corresponding to the positive samples of edited viral RNA sites.

[0006] Optionally, the prediction model for editing sites on human RNA is obtained by training the prediction model with source domain sample data; The source domain sample data includes: positive samples of human RNA editing sites and negative samples of human RNA non-editing sites; The positive samples of human RNA editing sites include: human RNA sequence fragments corresponding to the edited sites; The negative samples of human RNA non-edited sites include: human RNA sequence fragments corresponding to non-edited sites obtained through spatial isolation sampling, wherein the distance between the non-edited site and any labeled edited site is not less than a preset threshold, and the negative samples of human RNA non-edited sites do not contain adenosine deaminase ADAR editing preference sequence motifs that act on RNA.

[0007] Optionally, the extracted features are fused to obtain fused features, specifically including: The biological semantic features, the local motif features, and the spatial structure features are respectively subjected to dimensionality-upgrading processing to obtain a unified dimension for each upgraded feature representation; The various upgraded feature representations are concatenated to obtain a high-dimensional joint feature matrix; The high-dimensional joint feature matrix is ​​compressed to obtain a compressed feature representation; The compressed feature representation is mapped to an input dimension that is compatible with the backbone network of the prediction model through a preset linear projection layer.

[0008] Optionally, the biological semantic features, local motif features, and spatial structure features of the target domain sample data are extracted using the edit site prediction model, specifically including: Context encoding is performed on viral RNA sequence fragments in the target domain sample data to obtain the biological semantic features; The local motif features are obtained by extracting the k-mer composition features of the candidate site neighborhood in the viral RNA sequence fragment; The viral RNA sequence fragment is subjected to RNA secondary structure prediction, and the predicted dotted bracket structure representation and minimum free energy are encoded to obtain the spatial structure features.

[0009] Optionally, the edit site prediction model includes a multimodal feature extraction base, a feature adaptation module, and a backbone network; The multimodal feature extraction base is used to extract biological semantic features, local motif features, and spatial structure features from the target domain sample data, and to fuse the extracted features to obtain fused features. The backbone network is a Conformer network, and the feature adaptation module is used to map the fused features into an input representation that is adapted to the Conformer network. The Conformer network extracts local motif patterns related to viral RNA editing through a convolutional module and extracts long-range dependencies in the neighborhood of candidate sites through a multi-head self-attention module.

[0010] According to a second aspect of one or more embodiments of this specification, a method for predicting viral RNA editing sites is proposed, comprising: Obtain the viral RNA sequence to be predicted; The viral RNA sequence is input into a pre-trained viral RNA editing site prediction model to obtain the probability that each point in the viral RNA sequence is an editing site; wherein, the viral RNA editing site prediction model is trained by the above method; Based on the stated probability, the editing sites in the viral RNA sequence are determined.

[0011] According to a third aspect of one or more embodiments of this specification, a model training apparatus is provided, comprising: The acquisition module is used to acquire target domain sample data; wherein, the target domain sample data is viral RNA sequence fragments, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites; The prediction module is used to input the target domain sample data into a pre-trained human RNA editing site prediction model, so as to extract the biological semantic features, local motif features and spatial structure features of the target domain sample data through the editing site prediction model, and to perform feature fusion on the extracted features to obtain fused features; and, based on the fused features, to determine the predicted probability of each point in the target domain sample data as a viral RNA editing site. The training module is used to adjust the model parameters of the editing site prediction model by minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site, thereby obtaining an editing site prediction model for viral RNA.

[0012] According to a fourth aspect of one or more embodiments of this specification, a device for predicting viral RNA editing sites is provided, comprising: The acquisition module is used to acquire the viral RNA sequence to be predicted; The prediction module is used to input the viral RNA sequence into a pre-trained prediction model for editing sites of viral RNA to obtain the probability that each point in the viral RNA sequence is an editing site. The localization module is used to determine the editing site in the viral RNA sequence based on the probability.

[0013] According to a fifth aspect of one or more embodiments of this specification, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor implements the steps of the method described above by executing the executable instructions.

[0014] According to a sixth aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method described above.

[0015] According to a seventh aspect of one or more embodiments of this specification, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0016] As can be seen from the above embodiments, this specification acquires target domain sample data and inputs it into a pre-trained human RNA editing site prediction model. While retaining the existing editing site identification capabilities of the source domain model, it adjusts the parameters of the target domain viral RNA samples, thereby effectively migrating the model from the human RNA scenario to the viral RNA scenario and reducing the adverse effects of differences between different species scenarios on the prediction results. Furthermore, this application extracts biological semantic features, local motif features, and spatial structural features, and fuses multiple features to improve the model's ability to represent viral RNA editing-related information, enabling the trained model to more comprehensively characterize the criteria for identifying viral RNA editing sites. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a model training method provided in an exemplary embodiment; Figure 2 This is a schematic diagram of the model architecture of an edit site prediction model provided in an exemplary embodiment; Figure 3 This is a flowchart illustrating a method for predicting viral RNA editing sites, provided in an exemplary embodiment. Figure 4 This is a schematic structural diagram of a device provided in an exemplary embodiment; Figure 5 This is a block diagram of a model training apparatus provided in an exemplary embodiment; Figure 6 This is a block diagram of a device for predicting viral RNA editing sites, provided in an exemplary embodiment. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0019] RNA editing is an important nucleic acid modification process that occurs in the post-transcriptional stage. For viral RNA, editing events alter local nucleotide information and subsequent translation, replication, or host recognition processes. Therefore, the identification of viral RNA editing sites can be used for tasks such as viral mutation analysis, host-virus interaction studies, virulence change assessment, vaccine candidate sequence screening, site selection in targeted drug development, and viral surveillance and control analysis. In practical applications, researchers often need to rapidly scan large amounts of viral RNA sequences to determine which sites are more likely to be edited, thereby narrowing the scope of experimental validation and reducing wet laboratory costs.

[0020] While existing RNA editing site prediction methods can output editing probabilities in some human data scenarios, they still have several shortcomings in viral RNA scenarios. First, some schemes rely primarily on single sequence encoding, capturing mainly local base arrangement information, making it difficult to simultaneously characterize the deep contextual semantics, local motif patterns, and spatial structure information associated with viral RNA editing. This results in an insufficient portrayal of the real editing mechanism. Second, most existing schemes are trained on human RNA data. However, human RNA and viral RNA differ significantly in sequence composition, contextual distribution, and structural patterns. Direct transfer of these differences can easily lead to a decline in predictive performance, making it difficult to meet the needs of cross-scenario applications. Third, viral RNA editing sites typically account for a low percentage in real-world data. If the negative sample construction method is coarse, the model is prone to learning unstable or biased discrimination boundaries, further affecting the accuracy and generalization ability of the prediction results.

[0021] Based on this, this specification provides a model training method. While preserving the basic representational capabilities of the source domain model, it gradually adapts the model to the viral RNA scenario, thereby improving the accuracy and generalization ability of viral RNA editing site prediction.

[0022] To facilitate understanding of the technical solution of this application, the key concepts involved in this application are explained below.

[0023] Viral RNA sequence fragments: These refer to RNA sequence data with a preset context window length that is truncated around candidate sites. In the context of this application, the model does not necessarily have to use the entire viral RNA sequence as the smallest processing unit. Instead, the entire sequence can be divided into multiple viral RNA sequence fragments according to candidate sites, and then each fragment can be input into the model for training or prediction.

[0024] Target domain sample data: refers to the dataset used to transfer a pre-trained human RNA editing site prediction model to a viral RNA scenario. Target domain sample data includes at least positive viral RNA editing site samples and negative viral RNA non-editing site samples, which can be derived from sequencing data, database annotation results, experimental validation results, or a combination of the above.

[0025] Biological semantic features refer to high-level representations obtained by contextual encoding of viral RNA sequence fragments. These features are not limited to simple single-base encoding but are used to characterize longer-range contextual dependencies, sequence semantic relationships, and potential biological patterns within the site's neighborhood. In some embodiments, these features can be generated by a pre-trained biological sequence language model.

[0026] Local motif features: These refer to the patterns of base combinations in the vicinity of candidate sites. These features can reflect the presence of statistical patterns of k-mer combinations and motif preferences related to RNA editing around the candidate site.

[0027] Spatial structure features: These refer to features constructed based on the secondary structure prediction results of viral RNA sequence fragments. These features may include bracket structures representing the corresponding coding results, minimum free energy, and other information characterizing base pairing states, loop structures, or stem structures. These features enhance the model's ability to understand information related to RNA spatial conformation.

[0028] Conformer network: refers to a network structure that combines convolutional processing capabilities with self-attention processing capabilities. In the context of this application, the convolutional module in this network is mainly used to extract local motif patterns related to viral RNA editing, while the multi-head self-attention module is mainly used to extract long-range dependencies in the neighborhood of candidate sites, thereby enabling the model to take into account both local and global information.

[0029] ADAR-preferred sequence motifs refer to the sequence patterns in the vicinity of candidate sites where ADAR, an RNA-acting adenosine deaminase, is more likely to cause editing. Preferred motifs can be set based on existing research findings, experimental statistics, or the distribution of the target task dataset. ADAR-rejected sequence motifs represent sequence patterns in the vicinity that are unfavorable to the occurrence of the corresponding editing event. This application utilizes the difference between these two types of motifs for reverse selection during the construction of negative samples in the target domain.

[0030] Spatial isolation sampling refers to imposing distance constraints between non-edited sites and labeled edited sites during source domain negative sample construction to reduce potential conflicts between samples. This distance constraint can be achieved through genomic coordinate distance, relative positional distance within transcripts, or other measurable methods.

[0031] In this specification, the entity used to perform model training and the prediction method for viral RNA editing sites can be a designated device such as a server. In the scenario of viral RNA editing site identification, the server can perform two types of tasks. The first type of task is the model training task, which involves adjusting the parameters of an editing site prediction model that has been pre-trained for human RNA based on target domain sample data to obtain an editing site prediction model for viral RNA. The second type of task is the prediction task, which involves inputting the viral RNA sequence to be predicted into the trained model, outputting the probability of each bit as an editing site, and further determining the editing site. The former mainly corresponds to the training phase, while the latter mainly corresponds to the actual application phase after model deployment. The two types of tasks can be performed on the same server or distributed across different servers.

[0032] Of course, the aforementioned execution subject can also be devices such as tablet computers, laptops, and desktop computers, as well as clients installed on these devices. For ease of description, the following will use a server as the execution subject to illustrate the method provided in this manual.

[0033] The technical solutions described in the embodiments of this specification will be explained in detail below with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a model training method provided in an exemplary embodiment, including the following steps: S101: Obtain target domain sample data; wherein, the target domain sample data is viral RNA sequence fragments, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites.

[0034] In the scenario of identifying viral RNA editing sites, the server first acquires sample data of the target domain.

[0035] In practical applications, target domain sample data can be viral RNA sequence fragments, specifically including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites.

[0036] The input format for target domain sample data can be FASTA files, CSV files, database records, structured sample files in object storage, or data objects directly output from upstream sequencing analysis processes.

[0037] The server can first acquire a batch of raw viral RNA sequences, and then locate candidate sites within each raw sequence based on preset candidate site types. Next, it extracts viral RNA sequence fragments of a preset window length around each candidate site. The window length can be configured according to task requirements, such as extracting a fixed length of context centered on the candidate site. The window length can be an odd number to ensure the candidate site is centered; it can also be an asymmetric window to highlight upstream or downstream sequence information. The specific value of the window length can be set based on training resources, the average length of the target sequence, and empirical test results; this manual does not impose any limitations on this.

[0038] In addition, positive samples of viral RNA editing sites can also be derived from experimentally validated editing sites, high-confidence sites reported in the literature, labeled sites compiled from databases, or results obtained through cross-validation of information from multiple sources. The server can use these known editing sites as central sites to extract viral RNA sequence fragments from the corresponding viral RNA sequences to construct positive samples of viral RNA editing sites.

[0039] Furthermore, to enhance the biological reliability of the target domain sample data, positive samples of viral RNA editing sites can include viral RNA sequence fragments that satisfy ADAR editing preference sequence motifs. In other words, when constructing positive samples, the server can retain only sequence fragments in the neighborhood of candidate sites that conform to the preference motif rules, or assign higher sampling priority to sequence fragments that conform to the preference motif rules. The preference motifs can be pre-configured as a set of motifs, and the server can determine whether a viral RNA sequence fragment satisfies the preference motifs through string matching, position weight matrix matching, rule engine comparison, and other methods.

[0040] Negative samples of viral RNA non-edited sites can be obtained by randomly selecting fragments corresponding to non-edited sites. To better distinguish between real positive samples and target domain negative samples, the server can use reverse selection (such as Motif reverse selection) to construct target domain negative samples. Specifically, negative samples of viral RNA non-edited sites can include: viral RNA sequence fragments of the same context window length as the positive samples of viral RNA edited sites, excluding ADAR-rejected viral RNA sequence motifs that are preferred by ADAR editing.

[0041] In other words, when constructing negative samples, the server does not simply treat unlabeled sites as negative samples. Instead, it prioritizes viral RNA sequence fragments that are consistent with positive samples in terms of window length, central site organization, etc., but exhibit opposite motif preferences. This allows the model to focus more on edit-related boundaries during training, rather than relying solely on coarse differences in class distribution for judgment.

[0042] Optionally, the server can first determine the window length corresponding to the positive sample, for example, uniformly set to L nucleotides. Then, it scans all candidate windows that meet the center site type requirements in the viral RNA sequence, removes windows containing ADAR editing preference sequence motifs, and selects windows containing ADAR rejection motifs from the remaining windows as negative samples. The selection method for negative samples can be full inclusion, random sampling, stratified sampling by sequence source, or sampling controlled by category proportion.

[0043] Furthermore, the ratio of positive to negative samples can be organized using a 1:1, 1:n, or dynamic ratio. If the number of positive samples in the target domain is small, the server can perform resampling, data augmentation, or weighted training on the positive samples; if the number of negative samples is much greater than the number of positive samples, the server can use undersampling, hard negative sample screening, or loss weighting to control the training balance. This application does not limit the precise ratio of positive to negative samples.

[0044] To better understand the process of acquiring target domain sample data, an example is given below. Assume the server acquires a viral RNA sequence of length M and extracts all window fragments centered on candidate bases. For fragments that are labeled as edit sites by experimental results and conform to ADAR edit preference sequence motifs (such as UAG, UAC, UAU, UAA, AAG, AAC, AAU, AAA), the server classifies them as positive viral RNA edit sites. For fragments that do not belong to known edit sites, have the same window length as the positive samples, and contain ADAR exclusion motifs (such as CAC, GAC, CAA, GAA, etc., in the avoid50 negative sample set), the server classifies them as negative viral RNA non-edit sites. Finally, the server obtains target domain sample data for subsequent transfer learning training.

[0045] S102: Input the target domain sample data into a pre-trained human RNA editing site prediction model to extract biological semantic features, local motif features, and spatial structure features of the target domain sample data through the editing site prediction model, and perform feature fusion on the extracted features to obtain fused features; and, based on the fused features, determine the predicted probability of each point in the target domain sample data as a viral RNA editing site.

[0046] In this step, the server can use a pre-trained source domain edit site prediction model to perform forward computation on the target domain sample data. This edit site prediction model can be trained using source domain sample data. The source domain sample data includes positive human RNA edit sites and negative human RNA non-edit sites.

[0047] It should be noted that, in this manual, the pre-training and fine-tuning processes of the site prediction model are essentially the same in terms of model execution flow. The main differences lie in the training sample data, parameter adjustment range, and training objective settings. To avoid repetition, this manual will not elaborate on the technical content that is the same as or similar to the pre-training process during fine-tuning; relevant content can be found in the description of the pre-training process.

[0048] Positive samples of human RNA editing sites can include human RNA sequence fragments corresponding to the labeled editing sites. The server can obtain labeled editing sites from public databases, existing annotation sets, or experimental records, and then extract the corresponding fragments as positive samples of the source domain.

[0049] Negative samples of non-edited human RNA sites can include human RNA sequence fragments corresponding to non-edited sites obtained through spatial isolation sampling. The server can first determine the coordinates of the labeled edited sites, then select non-edited sites from the human RNA sequence whose distance from any labeled edited site is not less than a preset threshold, and extract corresponding fragments around these sites. Furthermore, to reduce potential editing bias interference in the negative samples, the server can also require that the negative samples of non-edited human RNA sites do not contain ADAR editing-biased sequence motifs. This results in fewer conflicts between the source domain negative samples and positive samples, which is beneficial for improving the basic discriminative ability learned by the pre-trained model.

[0050] The aforementioned preset thresholds can be set to a fixed distance according to task requirements, such as a distance threshold based on the number of nucleotides; alternatively, stratified thresholds, dynamic thresholds, or different thresholds can be set according to sequence region type. Spatial isolation sampling can also be further combined with weighted sampling strategies, such as weighting by chromosome length, transcript length, region coverage, or sample source ratio, to make the source domain sample distribution more balanced.

[0051] After completing the source domain pre-training, the server inputs the target domain sample data into the pre-trained model. Specifically, the edit site prediction model can first perform context encoding on the viral RNA sequence fragments in the target domain sample data to obtain biological semantic features. Context encoding can be implemented using pre-trained biological sequence language model encoding, Transformer encoding, bidirectional recurrent encoding, convolutional encoding, or a combination of the above methods.

[0052] Preferably, the server can employ a pre-trained language model oriented towards nucleic acid sequences to tokenize viral RNA sequence fragments and generate contextual embedding representations. Alternatively, the server can encode the sequence into a fixed-dimensional vector before inputting it into the upstream encoding network to obtain biological semantic features.

[0053] The server can also extract k-mer composition features from the neighborhood of candidate sites in viral RNA sequence fragments to obtain local motif features. The k value of the k-mer can be 1, 2, 3, 4, or other integers; it can also extract statistical features under multiple k values ​​simultaneously and splice them. The pseudo-k-nucleotide composition can be constructed using nucleotide frequencies, physicochemical property mapping values ​​of adjacent nucleotides, sequence order-related parameters, etc. For different sequence lengths or different virus types, the server can use a uniform parameter extraction method or an extraction method configured according to the dataset.

[0054] The server can also predict the secondary structure of viral RNA sequence fragments and encode the predicted bracket structure representation and minimum free energy to obtain spatial structure features. Secondary structure prediction can be implemented based on the minimum free energy principle, dynamic programming algorithms, or other RNA structure prediction methods. The bracket structure representation can be further mapped to numerical codes, for example, mapping unpaired positions, left bracket positions, and right bracket positions to different integers or vectors; the minimum free energy can be directly used as a scalar feature or, after normalization, as input as a structural sub-feature.

[0055] In some embodiments, the spatial structural features may further include features such as stem length, ring length, pairing ratio, local hairpin structure markers, and neighborhood unpairing density.

[0056] After obtaining biological semantic features, local motif features, and spatial structure features, the server performs feature fusion on multiple types of features to obtain fused features.

[0057] A specific implementation of feature fusion for the extracted features is as follows: Biological semantic features, local motif features, and spatial structure features are respectively subjected to dimensionality upscaling to obtain unified dimensionality feature representations; the dimensionality upscaling feature representations are concatenated to obtain a high-dimensional joint feature matrix; the high-dimensional joint feature matrix is ​​compressed to obtain a compressed feature representation; and the compressed feature representation is mapped to an input dimension that is compatible with the backbone network of the prediction model through a preset linear projection layer.

[0058] The dimensionality upscaling can be accomplished by independent fully connected layers, convolutional mapping layers, embedding mapping layers, or linear layers with activation functions. The purpose of unifying the dimensionality is to reduce scale differences between features of different modalities, facilitating subsequent concatenation and joint modeling. Concatenation can be performed along the feature dimension or along the sequence position dimension, depending on the input organization of the backbone network. After the high-dimensional joint feature matrix is ​​formed, the server can compress it using bottleneck layers, low-rank mappings, linear compression layers, or nonlinear dimensionality reduction layers to reduce redundant information and control the computational load of the model. A pre-defined linear projection layer further maps the compressed feature representation to an input representation that the backbone network can accept.

[0059] In a more specific embodiment, the server can perform unified dimensionality upscaling on biological semantic features, local motif features, and spatial structure features, respectively, so that the three types of features are each represented by a 2560-dimensional (2560-dim) feature vector. Specifically, biological semantic features can be extracted from biological sequence models, local motif features can be constructed from k-mer composition features and pseudo-K nucleotide composition, and spatial structure features can be obtained from RNA secondary structure prediction results and minimum free energy encoding. Subsequently, the server concatenates and splices the three types of 2560-dimensional feature vectors to form a 7680-dimensional (7680-dim) high-dimensional joint feature vector. Then, the server uses a compression layer, preferably a bottleneck layer or a nonlinear projection layer (Linear Projection), to perform high-fidelity compression on this 7680-dimensional joint feature vector, transforming it into a 1024-dimensional (1024-dim) compressed feature representation. Finally, the server uses a preset linear projection layer to map this 1024-dimensional compressed feature representation into an input representation adapted to the backbone network's input requirements, and uses it as the input token representation of the backbone network.

[0060] Furthermore, the editing site prediction model can include a multimodal feature extraction platform, a feature adaptation module, and a backbone network. The multimodal feature extraction platform extracts biological semantic features, local motif features, and spatial structure features from the target domain sample data, and fuses these extracted features to obtain fused features. The backbone network is a Conformer network. The feature adaptation module maps the fused features to an input representation adapted to the Conformer network. The Conformer network extracts local motif patterns related to viral RNA editing through convolutional modules and extracts long-range dependencies in the neighborhood of candidate sites through a multi-head self-attention module.

[0061] In practical applications, the server can load pre-trained Conformer model weights obtained externally and use them as initialization parameters for the backbone network. By reusing the parameters of the pre-trained Conformer model, the backbone network can possess certain prior knowledge of feature modeling in the early stages of training, thus serving as a foundation of prior knowledge for multimodal fusion feature processing. This reduces the parameter search overhead of training the model from scratch and improves model convergence efficiency and training stability.

[0062] The Conformer network can include feedforward modules, convolutional modules, multi-head self-attention modules, and normalization units. The convolutional modules can employ one-dimensional convolution, depthwise separable convolution, or other local convolutional structures to enhance local pattern recognition capabilities. The multi-head self-attention module can be used to calculate correlation weights between different locations, thereby capturing a longer range of contextual dependencies. The model output can be connected to one or more fully connected layers or a classification head (such as a 9-layer MLP network) to output the predicted probability of each point as a viral RNA editing site.

[0063] The predicted probabilities can be output using a sigmoid function to implement binary probability mapping. For each viral RNA sequence fragment, the server can output the predicted probability of its central candidate site as an edit site. For batch samples, the server can output a probability vector or a probability list. For the entire viral RNA sequence, the server can iteratively execute the above process on multiple candidate sites to ultimately obtain the site-level predicted probability distribution of the entire sequence.

[0064] Figure 2 This is a schematic diagram of the model architecture of an edit site prediction model provided in an exemplary embodiment.

[0065] The edit site prediction model may include biological semantic feature embeddings (Sequence Embeddings), local motif feature embeddings (Trinucleotide Embeddings), spatial structure feature embeddings (Structure Embeddings), concatenation, a dimension reduction block, a Conformer encoder, global mean pooling, and a classifier head. In one example, the feature dimensions of the biological semantic feature embeddings, local motif feature embeddings, and spatial structure feature embeddings can be uniformly set to 2560 dimensions. The three features are concatenated to form a joint feature representation with a dimension of 7680 dimensions. The dimension reduction block may include a linear layer, layer normalization, ReLU activation function, and dropout layer to map the joint feature representation to an input representation adapted to the Conformer encoder; in one example, the mapped feature dimension can be 1024 dimensions. Furthermore, the Conformer encoder may include multiple encoder layers, such as two encoder layers. Each encoder layer may include a feed-forward network, layer normalization, a multi-head self-attention module, and a convolution module. In one example, the model dimension d_model can be set to 1024, the number of attention heads nhead can be set to 8, the feed-forward network dimension ffn_dim can be set to 2048, the dropout rate can be set to 0.1, and the convolution kernel size k of the convolution module can be set to 31. The features output by the Conformer encoder can be further processed by global mean pooling to obtain sample-level representations, which are then input into a classification head for classification prediction. The classification head may be a multi-layer perceptron (MLP), which may include multiple layers of linear mapping and activation operations, such as mapping layers from 1024 dimensions to 512 dimensions, from 512 dimensions to 256 dimensions, from 256 dimensions to 128 dimensions, and from 128 dimensions to 2 dimensions, to output classification output values ​​(Logits). During the training phase, the classification output values ​​and ground truth labels can be input together into a binary cross-entropy loss function with logits to optimize the model parameters.

[0066] S103: With the goal of minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site, the model parameters of the editing site prediction model are adjusted to obtain an editing site prediction model for viral RNA.

[0067] In this step, the server trains or fine-tunes the model based on the prediction results corresponding to the target domain sample data. Specifically, the server aims to maximize the prediction probability of positive samples and minimize the prediction probability of negative samples. Therefore, an optimization objective can be constructed based on the prediction probabilities of positive and negative samples, and the model parameters of the editing site prediction model can be adjusted.

[0068] It should be added that minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site can be an objective for training the editing site prediction model. That is, the editing site prediction model needs to simultaneously satisfy the minimization of the predicted probability of the negative sample as a viral RNA editing site and the maximization of the predicted probability of the positive sample as a viral RNA editing site.

[0069] In one embodiment, the server may employ a binary cross-entropy loss function to calculate the loss value based on the deviation between the positive and negative sample labels and the predicted probabilities, and update the model parameters through a backpropagation algorithm.

[0070] Considering that the target domain sample data may have class imbalance, the server can also use the focus loss function, the weighted cross-entropy loss function, or a combination of both to increase the model's attention to minority class positive samples and hard negative samples.

[0071] Specifically, for the edit sites marked in the positive samples, the server can determine the first loss value based on the predicted probability that the site is a viral RNA editing site. The first loss value is negatively correlated with the predicted probability, that is, the higher the predicted probability, the smaller the first loss value. For negative samples, the server can determine a second loss value based on the predicted probability that the corresponding site is a viral RNA editing site. The second loss value is positively correlated with this predicted probability; that is, the higher the predicted probability, the larger the second loss value. Subsequently, the server can determine a target loss value based on the first and second loss values ​​(e.g., by summing, weighted summing, or other loss aggregation methods), and optimize by minimizing this target loss value. The server uses a backpropagation algorithm to iteratively update the model parameters of the editing site prediction model until a preset optimization condition is met. This preset optimization condition may include at least one of the following: the target loss value converges to a preset range, the loss decrease is less than a preset threshold in multiple consecutive iterations, or a preset number of training iterations is reached.

[0072] Of course, the server can also choose to minimize negative samples as the prediction probability of viral RNA editing sites or maximize positive samples as the prediction probability of viral RNA editing sites as the training objective of the editing site prediction model.

[0073] During parameter tuning, the server can update all parameters of the entire edit site prediction model, or it can update only the parameters of certain modules. For example, the server can freeze some source domain pre-trained layers and only update the feature adaptation module, backbone network high-level parameters, and classification head parameters to reduce the risk of overfitting and preserve source domain knowledge. For scenarios with limited computational resources, a phased training strategy can also be adopted, such as first training only the feature adaptation module and classification head, and then gradually unfreezing the backbone network high-level layers for joint fine-tuning.

[0074] In some embodiments, the server can organize training samples in batches, setting batch size, learning rate, number of training epochs, early stopping strategy, and validation set monitoring metrics. The learning rate can be a fixed value, piecewise decay, cosine annealing, or adaptive optimization. The optimizer can employ stochastic gradient descent, Adam, AdamW, or other optimization algorithms. Validation metrics may include accuracy, recall, precision, F1 score, AUC, or other classification task evaluation metrics. This application does not limit the specific training hyperparameters.

[0075] After training, the server obtains a prediction model for editing sites on viral RNA. This model can be stored in local storage, a database, a model repository, or a remote object storage, and can be called by subsequent inference services.

[0076] To facilitate understanding of the training process described above, an exemplary implementation is provided below. The server first trains a source domain editing site prediction model using positive human RNA editing site samples and negative human RNA non-editing site samples obtained through spatially isolated sampling. Then, the server acquires target domain sample data for a certain type of viral RNA, where positive samples consist of fragments with known editing sites, and negative samples consist of fragments with consistent window lengths containing ADAR exclusion motifs. The server inputs the target domain sample data into the source domain editing site prediction model, extracts and fuses three types of features, and then outputs the prediction probability for each point through a Conformer network. Finally, the server updates the model parameters by minimizing the prediction probability of negative samples and maximizing the prediction probability of positive samples, thus obtaining an editing site prediction model for this type of viral RNA.

[0077] In some alternative embodiments, model evaluation and threshold determination steps can be performed after model training is complete. For example, the server can statistically analyze precision and recall at different probability thresholds on the validation set to select a target threshold suitable for subsequent prediction tasks. The threshold can be a fixed threshold or dynamically set according to virus type, data source, or task objective. For example, in scenarios emphasizing no missed detections, a lower threshold can be used to improve recall; in scenarios emphasizing high-confidence results, a higher threshold can be used to improve precision.

[0078] Furthermore, this specification also provides a method for predicting editing sites applied to the aforementioned editing site prediction model for viral RNA, such as... Figure 3 As shown.

[0079] Figure 3 This is a flowchart illustrating a method for predicting viral RNA editing sites, provided in an exemplary embodiment, including the following steps: S301: Obtain the viral RNA sequence to be predicted.

[0080] In practical applications, the server can obtain the viral RNA sequence to be predicted from sequencing results, public databases, files uploaded to experimental platforms, or upstream analysis systems. The input can be a single viral RNA sequence or a set of multiple viral RNA sequences. The viral RNA sequence to be predicted can be a complete sequence, a fragment sequence, a spliced ​​sequence, or a denoised and normalized sequence.

[0081] After obtaining the viral RNA sequence to be predicted, the server can perform preprocessing. Preprocessing may include removing illegal characters, standardizing base representation, handling missing values, length normalization, and screening candidate sites. For example, if there are uncertain base markers in the sequence, the server can replace them using preset rules, skip windows containing that position, or retain and mark that position. If the sequence to be predicted is long, the server can first locate all candidate sites and then construct viral RNA sequence fragments around each candidate site.

[0082] S302: Input the viral RNA sequence into a pre-trained prediction model for editing sites on viral RNA to obtain the probability that each point in the viral RNA sequence is an editing site.

[0083] The viral RNA editing site prediction model is trained using the aforementioned model training method. In one embodiment, the server can split the entire viral RNA sequence into multiple candidate site fragments, and then input these fragments into the model in batches to obtain the predicted probability of each candidate site. In another embodiment, the server can also use a sliding window approach to scan along the viral RNA sequence position by position, calculating the probability of each candidate site being an editing site.

[0084] The feature extraction, feature fusion, feature adaptation, and backbone network processing methods used during model inference can remain consistent with those used in the training phase. In other words, the server can still extract biological semantic features, local motif features, and spatial structure features based on viral RNA sequence fragments, convert them into fused features, and output probability results through the Conformer network.

[0085] S303: Determine the editing sites in the viral RNA sequence based on probability.

[0086] The server can compare the predicted probability of each site with a preset threshold. When the predicted probability of a site is not less than the threshold, the site is determined to be an edit site; otherwise, the site is determined to be a non-edit site. In another embodiment, the server can also select the top K sites as high-confidence edit sites after sorting them according to their predicted probabilities, or use a dual-threshold mechanism to divide sites into high-confidence edit sites, sites to be verified, and non-edit sites.

[0087] In addition, the server's output prediction results include not only the location information of the edit site, but also the corresponding probability value, window segment, local motif description, structural feature summary, or model confidence level. This result can be used by downstream experimental platforms, data analysis platforms, or reporting systems.

[0088] Once the editing sites on the viral RNA are determined, the server can perform subsequent application analyses based on the prediction results.

[0089] For example, the server can perform virus analysis based on the predicted location of the editing site and its predicted probability, statistically analyze the differences in the editing distribution of viral RNA in different samples, at different time points, from different host sources or under different treatment conditions, and then analyze the hot spots of viral RNA editing, the trend of editing frequency changes and their association with viral mutation patterns, which can be used to assist in the study of viral evolution paths and host adaptation mechanisms.

[0090] The server can also perform control and prevention analyses based on high-confidence edit sites. For example, it can identify potential edit hotspots associated with immune escape, changes in replication efficiency, changes in transmissibility, or changes in virulence, and continuously track these areas to provide data support for virus surveillance, risk assessment, and the development of control and prevention strategies.

[0091] In vaccine screening scenarios, servers can assess the stability of candidate viral RNA sequences based on predicted high-probability edited regions. For example, during candidate sequence design or screening, high-probability edited regions can be avoided, or candidate sequences containing high-probability edited regions can be flagged for risk, thereby reducing the risk of sequence instability or antigenic characteristic changes caused by RNA editing.

[0092] In targeted drug development scenarios, servers can use high-confidence editing sites as key areas for structural and functional analysis. By combining protein coding region information, domain information, or other functional annotation results, editing sites that may affect key viral functions can be prioritized for analysis, thereby assisting in target selection, drug mechanism research, and candidate intervention site screening.

[0093] In addition, the server can sort candidate editing sites based on the prediction results, and prioritize outputting sites with a prediction probability higher than a preset threshold to form a list of sites to be verified, thereby narrowing the scope of subsequent experimental verification, reducing the workload of blind screening, and improving the efficiency of wet experimental resource utilization and overall research efficiency.

[0094] As can be seen from the above methods, the technical solution provided in this specification can adjust the model parameters based on a human RNA pre-trained editing site prediction model, combined with target domain viral RNA sample data, thereby achieving effective model transfer from human RNA scenarios to viral RNA scenarios. This approach can reduce the adverse impact of data distribution differences between different species scenarios on prediction performance, and improve the accuracy and generalization ability of predicting viral RNA editing sites.

[0095] Furthermore, this specification extracts biological semantic features, local motif features, and spatial structural features, and fuses these features to enable the model to simultaneously consider sequence context information, local motif patterns, and RNA secondary structure information, thereby improving its ability to characterize information related to viral RNA editing sites. Compared to schemes that rely solely on single sequence information, this technical approach helps improve the model's ability to recognize complex editing patterns.

[0096] Furthermore, in the process of constructing target domain samples, this manual introduces positive samples that conform to the ADAR editing preference sequence motif and ADAR-rejected negative samples obtained based on reverse selection. This helps to enhance the model's ability to learn the boundaries between positive and negative samples and reduce the impact of negative sample definition bias on model training.

[0097] Therefore, the obtained prediction model for editing sites on viral RNA can be further used for tasks such as virus analysis, prevention and control analysis, vaccine screening, targeted drug development, and experimental site optimization, and has good application value.

[0098] Figure 4 This is a schematic structural diagram of a device provided in an exemplary embodiment. For example... Figure 4As shown, device 401 mainly consists of a communication interface 402, a user interface 403, a processor 404, and a data storage 405. These components are interconnected and communicate with each other via a system bus, network, or other connection mechanism 406. The communication interface 402 enables device 401 to communicate with other devices, access networks, and transmission networks via analog or digital modulation. For example, the communication interface 402 may include a chipset and antenna for wireless communication with a radio access network or access point. Furthermore, the communication interface 402 can be a wired interface such as Ethernet, Token Ring, or a USB port, or a wireless interface such as Wi-Fi, Bluetooth, Global Positioning System (GPS), or a wide-area wireless interface (e.g., WiMAX or LTE). Of course, the communication interface 402 can also support other forms of physical layer interfaces and standard or proprietary communication protocols. The communication interface 402 may also include multiple physical communication interfaces, such as Wi-Fi, Bluetooth, and wide-area wireless interfaces.

[0099] User interface 403 includes receiving user input and providing output to the user. Therefore, user interface 403 may include input components such as a keypad, keyboard, touch-sensitive or presence-sensitive panel, computer mouse, trackball, joystick, microphone, still camera, and video camera, and output components such as a display screen (which may be combined with a touch-sensitive panel), CRT, LCD, LED, display using DLP technology, printer, and other similar devices known or developed in the future. User interface 403 may also generate auditory output via speakers, speaker jacks, audio output ports, audio output devices, headphones, and other similar devices known or developed in the future. In some embodiments, user interface 403 may include software, circuitry, or other forms of logic capable of transmitting and receiving data from external user input / output devices. Additionally or alternatively, device 401 may support remote access from other devices via communication interface 402 or another physical interface (not shown). User interface 403 may be configured to receive user input, the position and movement of which may be indicated by indicators or cursors described herein. User interface 403 may also be configured as a display device for rendering or displaying text fragments.

[0100] Processor 404 may contain one or more general-purpose processors and / or special-purpose processors.

[0101] Data storage 405 may include one or more volatile and / or non-volatile storage components and may be integrated wholly or partially with processor 404. Data storage 405 may include removable and non-removable components.

[0102] Processor 404 is capable of executing program instructions 410 (e.g., compiled or uncompiled program logic and / or machine code) stored in data storage 405 to perform the various functions described herein. Data storage 405 may contain a non-transitory computer-readable medium on which program instructions are stored, which, when executed by device 401, enable device 401 to perform any methods, processes, or functions disclosed in this specification and / or the accompanying drawings. Execution of program instructions 410 by processor 404 may result in processor 404 using data 407.

[0103] For example, program instructions 410 may include an operating system 412 (e.g., an operating system kernel, device drivers, and / or other modules) installed on device 401 and one or more applications 411 (e.g., a browser, social media application, or game application). Similarly, data 407 may include operating system data 409 and application data 408. Operating system data 409 is primarily accessible to the operating system 412, while application data 408 is primarily accessible to one or more applications 411. Application data 408 may reside in a file system visible or hidden from the user of device 401.

[0104] Application 411 can communicate with operating system 412 through one or more application programming interfaces (APIs). These APIs help application 411 read and / or write application data 408, transmit or receive information via communication interface 402, receive or display information on user interface 403, etc.

[0105] In some terminology, application 411 may be simply referred to as "app". Furthermore, application 411 can be downloaded to device 401 through one or more online app stores or app markets. However, the application can also be installed on device 401 in other ways, such as through a web browser or a physical interface on device 401 (e.g., a USB port).

[0106] Please refer to Figure 5 Model training devices can be applied to, for example Figure 5 The device shown is used to implement the technical solution of this specification. The model training device may include: The first acquisition module 501 is used to acquire target domain sample data; wherein, the target domain sample data is a viral RNA sequence fragment, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites; The first prediction module 502 is used to input the target domain sample data into a pre-trained human RNA editing site prediction model, so as to extract the biological semantic features, local motif features and spatial structure features of the target domain sample data through the editing site prediction model, and to perform feature fusion on the extracted features to obtain fused features; and, based on the fused features, to determine the prediction probability of each point in the target domain sample data as a viral RNA editing site. The training module 503 is used to adjust the model parameters of the editing site prediction model by minimizing the prediction probability of the negative sample as a viral RNA editing site and maximizing the prediction probability of the positive sample as a viral RNA editing site, so as to obtain an editing site prediction model for viral RNA.

[0107] Optionally, the positive viral RNA editing site sample includes: a viral RNA sequence fragment that satisfies the ADAR editing preference sequence motif acting on RNA; the negative viral RNA non-editing site sample includes: a viral RNA sequence fragment that is ADAR rejected, excluding the ADAR editing preference sequence motif, among viral RNA sequence fragments with the same context window length as the positive viral RNA editing site sample.

[0108] Optionally, the prediction model for human RNA editing sites is obtained by training the prediction model with source domain sample data; the source domain sample data includes: positive human RNA editing site samples and negative human RNA non-editing site samples; the positive human RNA editing site samples include: human RNA sequence fragments corresponding to the labeled editing sites; the negative human RNA non-editing site samples include: human RNA sequence fragments corresponding to the non-editing sites obtained through spatial isolation sampling, wherein the distance between the non-editing site and any labeled editing site is not less than a preset threshold, and the negative human RNA non-editing site samples do not contain adenosine deaminase ADAR editing preference sequence motifs that act on RNA.

[0109] Optionally, the first prediction module 502 is specifically used to perform dimensionality-upgrading processing on the biological semantic features, the local motif features, and the spatial structure features respectively to obtain dimensionality-upgraded feature representations of a unified dimension; to concatenate the dimensionality-upgraded feature representations to obtain a high-dimensional joint feature matrix; to compress the high-dimensional joint feature matrix to obtain a compressed feature representation; and to map the compressed feature representation to an input dimension adapted to the backbone network of the prediction model through a preset linear projection layer.

[0110] Optionally, the first prediction module 502 is specifically used to: perform context encoding on viral RNA sequence fragments in the target domain sample data to obtain the biological semantic features; extract k-mer composition features of candidate site neighborhoods in the viral RNA sequence fragments to obtain the local motif features; perform RNA secondary structure prediction on the viral RNA sequence fragments, and encode the predicted dotted bracket structure representation and minimum free energy to obtain the spatial structure features.

[0111] Optionally, the editing site prediction model includes a multimodal feature extraction base, a feature adaptation module, and a backbone network. The multimodal feature extraction base is used to extract biological semantic features, local motif features, and spatial structure features of the target domain sample data, and to fuse the extracted features to obtain fused features. The backbone network is a Conformer network, and the feature adaptation module is used to map the fused features to an input representation that is compatible with the Conformer network. The Conformer network extracts local motif patterns related to viral RNA editing through a convolution module and extracts long-range dependencies in the neighborhood of candidate sites through a multi-head self-attention module.

[0112] Please refer to Figure 6 Devices for editing viral RNA sites can be applied to, for example... Figure 4 The device shown is used to implement the technical solution described in this specification. The apparatus for the viral RNA editing site may include: The second acquisition module 601 is used to acquire the viral RNA sequence to be predicted; The second prediction module 602 is used to input the viral RNA sequence into a pre-trained prediction model for editing sites of viral RNA to obtain the probability that each point in the viral RNA sequence is an editing site. The determination module 603 is used to determine the editing site in the viral RNA sequence based on the probability.

[0113] For ease of description, the above devices are described by dividing them into various modules or units based on their functions. Of course, when implementing one or more of these specifications, the functions of each module or unit can be implemented in the same or different software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0114] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0115] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0116] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0117] What those skilled in the art will understand is: In this specification, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of additional identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded.

[0118] In this specification, “a,” “an,” and “the” do not specifically refer to the singular, but may also include the plural.

[0119] In this specification, ordinal numbers such as "first," "second," etc., do not necessarily indicate order; they are often used to distinguish between objects. For example, "first server" and "second server" usually refer to two servers. To differentiate between these two servers, they are described as "first server" and "second server." Of course, sometimes these two servers may be the same server.

[0120] In this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can also mean indirect receiving and sending. For example, A receiving data sent by B can be understood as A directly receiving the data sent by B, or it can be understood as A indirectly receiving the data sent by B through other entities such as C. Similarly, B sending data to A can be understood as B sending the data directly to A, or it can be understood as B indirectly sending the data to A through other entities such as C. Here, C can be one entity, or it can be two or more entities.

[0121] In this specification, unless explicitly stated otherwise, the relationships between structures can be direct or indirect. For example, when describing "A is connected to B," unless it is explicitly stated that A and B are directly connected, it should be understood that A can be directly connected to B or indirectly connected to B. Similarly, when describing "A is on top of B," unless it is explicitly stated that A is directly above B (AB is adjacent and A is above B), it should be understood that A can be directly above B or indirectly above B (AB is separated by other elements, and A is above B). And so on.

[0122] This specification uses specific terms to describe embodiments thereof. Terms such as "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of those different embodiments or examples, without contradiction.

[0123] Although one or more embodiments of this specification provide method steps as described in the embodiments or flowcharts, it is understood that the order of steps listed in the embodiments or flowcharts is only one of many possible execution orders and does not represent the only execution order. Therefore, when the claims involve method steps, any changes or adjustments to the order of such steps, or the parallelism between steps, are also within the scope of protection of the claims.

Claims

1. A model training method, characterized in that, include: Obtain target domain sample data; wherein, the target domain sample data is viral RNA sequence fragments, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites; The target domain sample data is input into a pre-trained human RNA editing site prediction model to extract biological semantic features, local motif features, and spatial structure features of the target domain sample data. The extracted features are then fused to obtain fused features. Furthermore, based on the fusion features, the predicted probability of each point in the target domain sample data as a viral RNA editing site is determined; With the goal of minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site, the model parameters of the editing site prediction model are adjusted to obtain an editing site prediction model for viral RNA.

2. The method as described in claim 1, characterized in that, The positive viral RNA editing site samples include viral RNA sequence fragments that satisfy the ADAR editing preference sequence motif for RNA-acting adenosine deaminase; The negative samples of non-edited viral RNA sites include: viral RNA sequence fragments that are rejected by ADAR, excluding the ADAR editing-preferred sequence motif, within the context window length corresponding to the positive samples of edited viral RNA sites.

3. The method as described in claim 1, characterized in that, The prediction model for editing sites on human RNA is obtained by training the prediction model with source domain sample data; The source domain sample data includes: positive samples of human RNA editing sites and negative samples of human RNA non-editing sites; The positive samples of human RNA editing sites include: human RNA sequence fragments corresponding to the edited sites; The negative samples of human RNA non-edited sites include: human RNA sequence fragments corresponding to non-edited sites obtained through spatial isolation sampling, wherein the distance between the non-edited site and any labeled edited site is not less than a preset threshold, and the negative samples of human RNA non-edited sites do not contain adenosine deaminase ADAR editing preference sequence motifs that act on RNA.

4. The method as described in claim 1, characterized in that, The extracted features are fused to obtain fused features, which specifically include: The biological semantic features, the local motif features, and the spatial structure features are respectively subjected to dimensionality-upgrading processing to obtain a unified dimension for each upgraded feature representation; The various upgraded feature representations are concatenated to obtain a high-dimensional joint feature matrix; The high-dimensional joint feature matrix is ​​compressed to obtain a compressed feature representation; The compressed feature representation is mapped to an input dimension that is compatible with the backbone network of the prediction model through a preset linear projection layer.

5. The method as described in claim 1, characterized in that, The biological semantic features, local motif features, and spatial structure features of the target domain sample data are extracted using the edit site prediction model, specifically including: Context encoding is performed on viral RNA sequence fragments in the target domain sample data to obtain the biological semantic features; The local motif features are obtained by extracting the k-mer composition features of the candidate site neighborhood in the viral RNA sequence fragment; The viral RNA sequence fragment is subjected to RNA secondary structure prediction, and the predicted dotted bracket structure representation and minimum free energy are encoded to obtain the spatial structure features.

6. The method as described in claim 1, characterized in that, The edit site prediction model includes a multimodal feature extraction base, a feature adaptation module, and a backbone network; The multimodal feature extraction base is used to extract biological semantic features, local motif features, and spatial structure features from the target domain sample data, and to fuse the extracted features to obtain fused features. The backbone network is a Conformer network, and the feature adaptation module is used to map the fused features into an input representation that is adapted to the Conformer network. The Conformer network extracts local motif patterns related to viral RNA editing through a convolutional module and extracts long-range dependencies in the neighborhood of candidate sites through a multi-head self-attention module.

7. A method for predicting viral RNA editing sites, characterized in that, include: Obtain the viral RNA sequence to be predicted; The viral RNA sequence is input into a pre-trained viral RNA editing site prediction model to obtain the probability that each point in the viral RNA sequence is an editing site; wherein the viral RNA editing site prediction model is trained by the method described in any one of claims 1 to 6 above. Based on the stated probability, the editing sites in the viral RNA sequence are determined.

8. A model training device, characterized in that, include: The acquisition module is used to acquire target domain sample data; wherein, the target domain sample data is viral RNA sequence fragments, including positive samples of viral RNA editing sites and negative samples of viral RNA non-editing sites; The prediction module is used to input the target domain sample data into a pre-trained human RNA editing site prediction model, so as to extract the biological semantic features, local motif features and spatial structure features of the target domain sample data through the editing site prediction model, and to perform feature fusion on the extracted features to obtain fused features; and, based on the fused features, to determine the predicted probability of each point in the target domain sample data as a viral RNA editing site. The training module is used to adjust the model parameters of the editing site prediction model with the goal of minimizing the predicted probability of the negative sample as a viral RNA editing site and maximizing the predicted probability of the positive sample as a viral RNA editing site, so as to obtain an editing site prediction model for viral RNA.

9. An electronic device, comprising: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as described in any one of claims 1-7 by executing the executable instructions.

10. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1-7.

Images