Method and system for predicting protein-rna interactions for structural enhancement and synergistic attention

Abstract

The present application relates to the field of biological information, and particularly relates to a protein-RNA interaction prediction method and system with structure enhancement and synergistic attention. The method comprises the following steps: fusing in-vivo structure data and in-vitro predicted base pairing probability matrix, and enhancing structure features through a graph convolutional neural network; adopting multi-scale convolution and an adaptive attention weighting mechanism, so that the model can capture sequence-structure context dependency; realizing deep synergy and mutual reinforcement of sequence and structure features through a bidirectional cross-attention mechanism, and making the model automatically focus on key binding regions through motif-aware attention pooling; and a hybrid expert classifier selects the most suitable expert network through gating adaptation to make a decision, so as to obtain a prediction result of protein-RNA interaction; and the accuracy, robustness and interpretability of protein-RNA interaction prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting protein-RNA interactions using structure enhancement and synergistic attention. Background Technology

[0002] RNA-binding proteins (RBPs) are a class of key functional proteins that specifically recognize and bind to specific RNA sequences or structural motifs. They often participate in and regulate various aspects of gene expression, encompassing the entire life cycle processes, including transcriptional regulation, mRNA splicing, translation, and RNA degradation. Determining the binding patterns of RBPs under different physiological conditions and cellular environments is crucial for elucidating their complex post-transcriptional regulatory networks and disease development mechanisms. Although elucidating the dynamic interactions between RBPs and RNA is essential for understanding their biological functions, given that RBPs constitute approximately 10% of the human proteome, establishing precise target relationships from such a vast number undoubtedly presents a tremendous workload and technical challenge. Therefore, in-depth exploration of the classification of RNA-protein interactions (RPIs) is vital for understanding biological processes and functional mechanisms.

[0003] Currently, RPI analysis is divided into two main technical routes: wet experimental high-throughput detection and computational prediction models. Among the biological methods based on high-throughput technology, sequencing methods such as RNAcompete and RNABind-Seq for in vitro characterization of RBP sequence preferences, and eCLIP and subsequent immunoprecipitation for in vivo identification of RBP binding sites are used to process and predict RBP binding. In terms of computational prediction, computational methods for RBP binding sites based on machine learning or deep learning have also been successfully applied.

[0004] However, traditional protein-RNA interaction prediction methods often suffer from low prediction accuracy and insufficient generalization ability due to the high cost, low throughput, and susceptibility to measurement bias in wet experiments, over-reliance on single sequence information, lack of adaptability to dynamic cellular environments, and simple feature fusion mechanisms. Summary of the Invention

[0005] Based on this, in order to solve the above-mentioned technical problems, a method and system for predicting protein-RNA interactions using structure enhancement and synergistic attention are provided, which can improve the prediction accuracy of protein-RNA interactions.

[0006] A method for predicting protein-RNA interactions using structure enhancement and collaborative attention, the method comprising:

[0007] A protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data is obtained, and the length of the RNA sequences is uniformly processed. The sequence embedding features of the RNA sequences are extracted in a sliding window manner.

[0008] The in vitro secondary structure base pairing probability matrix of the RNA sequence is obtained to obtain the adjacency matrix of the graph structure. The features of the in vivo structure data after multi-scale convolution processing are used as the initial features of the graph nodes. Message passing and feature aggregation of the graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain structure enhancement features.

[0009] Convolutional operations are performed on the sequence embedding features and the structural enhancement features using convolutional kernels of different scales to extract contextual dependencies at different spatial distances. An adaptive attention weighting mechanism is introduced to dynamically weight and fuse the outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0010] Learnable positional codes are added to the multi-scale sequence semantic features and the multi-scale structural semantic features respectively, and bidirectional cross-attention calculation is performed to obtain the interactively enhanced collaborative representation.

[0011] The collaborative representation is processed by motif-aware attention pooling to obtain an aggregated feature vector. The aggregated feature vector is then input into a hybrid expert classifier. A gating network assigns weights to each expert network and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions.

[0012] In one embodiment, a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data is obtained, and the length of the RNA sequences is uniformly processed. Sequence embedding features of the RNA sequences are extracted using a sliding window approach, including:

[0013] RNA sequence fragments corresponding to the RNA-binding protein region are extracted as positive samples, and negative samples are generated by random sampling or shuffling the order of the RNA sequences. The in vivo structural data corresponding to each RNA sequence are associated to construct a protein-RNA binding sample dataset.

[0014] Obtain the maximum length threshold, and use the maximum length threshold as a benchmark to perform length uniformization processing on the RNA sequence to obtain the length uniformized RNA sequence;

[0015] The RNA sequence with uniform length is segmented using a sliding window. The segmented RNA sequence is then input into a pre-trained RNA base model based on the Transformer architecture. A multi-layer self-attention mechanism is used to output feature vectors as sequence embedding features of the RNA sequence.

[0016] In one embodiment, the in vitro secondary structure base pairing probability matrix of the RNA sequence is obtained to obtain the adjacency matrix of the graph structure. The features of the in vivo structural data after multi-scale convolution processing are used as the initial features of the graph nodes. Message passing and feature aggregation of a graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain structural enhancement features, including:

[0017] The RNA sequence is input into an RNA secondary structure prediction tool, and the in vitro secondary structure of the RNA sequence is predicted based on a thermodynamic model. The probability of pairing between any two nucleotide bases in the RNA sequence is calculated to obtain a base pairing probability matrix, which is then used as the adjacency matrix of the graph structure.

[0018] The in vivo structural data is input into a multi-scale convolutional neural network. Convolution operations are performed through various multi-scale convolutional kernels to obtain the weights of features at each scale and complete weighted fusion, ultimately obtaining the initial features of each graph node.

[0019] The initial features are input into a graph convolutional neural network. The initial features are aggregated and nonlinearly transformed along the edges defined by the adjacency matrix. The updated feature vectors of each node are output. All the updated feature vectors are concatenated in sequence according to their position to obtain the structure-enhanced features.

[0020] In one embodiment, the sequence embedding features and the structural enhancement features are convolved using kernels of different scales to extract contextual dependencies at different spatial distances. An adaptive attention weighting mechanism is then introduced to dynamically weight and fuse the outputs at different scales, resulting in multi-scale sequence semantic features and multi-scale structural semantic features, including:

[0021] The sequence embedding feature and the structural enhancement feature are processed to have a uniform length, so that the sequence feature and the structural feature correspond in sequence position, resulting in the processed sequence embedding feature and structural enhancement feature.

[0022] The processed sequence embedding features and structural enhancement features are used as two independent inputs. Each input is subjected to convolution operation in parallel to obtain intermediate features at each scale corresponding to each feature, so as to capture the sequence and structural context dependencies at different spatial distances.

[0023] Each feature is assigned a corresponding attention weight for each scale intermediate feature, and the outputs of each convolution operation branch are weighted and summed according to their respective attention weights to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0024] In one embodiment, learnable positional codes are added to the multi-scale sequence semantic features and the multi-scale structural semantic features, respectively, and bidirectional cross-attention calculation is performed to obtain an interactively enhanced collaborative representation, including:

[0025] The first learnable positional encoding is superimposed on the multi-scale sequence semantic features to obtain sequence interaction features with positional information; the second learnable positional encoding is superimposed on the multi-scale structural semantic features to obtain structural interaction features with positional information; the parameters of the first learnable positional encoding and the second learnable positional encoding are independent of each other and are updated through backpropagation during model training.

[0026] Using the sequence interaction features with location information as queries and the structural interaction features with location information as keys and values, a sequence-to-structure cross-attention calculation is performed to obtain sequence features enhanced by structural information constraints.

[0027] Using the structural interaction features with location information as queries and the sequence interaction features with location information as keys and values, cross-attention calculation from structure to sequence is performed to obtain structural features enhanced by sequence information constraints.

[0028] By splicing and fusing sequence features enhanced by structural information constraints and structural features enhanced by sequence information constraints, an interactively enhanced collaborative representation is obtained.

[0029] In one embodiment, the sequence interaction features with location information are used as queries, and the structural interaction features with location information are used as keys and values. Sequence-to-structure cross-attention calculation is performed to obtain sequence features enhanced by structural information constraints, including:

[0030] Calculate the attention similarity between each position in the sequence interaction feature with location information and all positions in the structural interaction feature with location information;

[0031] The attention similarity is scaled and then normalized to attention weights using the Softmax function. The structural interaction features are then weighted and aggregated using these attention weights to obtain sequence features enhanced with structural information.

[0032] In one embodiment, the collaborative representation is processed by motif-aware attention pooling to obtain an aggregated feature vector, including:

[0033] The attention scores at each position in the collaborative representation are calculated using a multilayer perceptron, and the attention scores are normalized using a Softmax function to obtain the attention weights corresponding to each position.

[0034] The collaborative representation is weighted and summed according to the attention weights, and the regional features containing the binding motif are compressed into an aggregated feature vector.

[0035] In one embodiment, the aggregated feature vector is input into a hybrid expert classifier, where a gating network assigns weights to each expert network and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions, including:

[0036] The aggregated feature vector is input into a gated network consisting of two linear transformation layers to obtain the original weight scores, and the original weight scores are normalized to obtain adaptive weights.

[0037] Each of the adaptive weights is assigned to a separate independent expert network, and the aggregated feature vector is input into each of the separate independent expert networks. Each group of the separate expert networks outputs a prediction score.

[0038] The predicted scores are weighted and summed with the adaptive weights respectively to obtain the final prediction result of protein-RNA interaction.

[0039] A protein-RNA interaction prediction system based on structure enhancement and collaborative attention, the system comprising:

[0040] The data acquisition module is used to acquire a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and to perform length uniform processing on the RNA sequences and extract the sequence embedding features of the RNA sequences in a sliding window manner.

[0041] The structure enhancement module is used to obtain the base pairing probability matrix of the in vitro secondary structure of the RNA sequence, obtain the adjacency matrix of the graph structure, and use the features of the in vivo structure data after multi-scale convolution processing as the initial features of the graph nodes. Message passing and feature aggregation of the graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain the structure enhancement features.

[0042] The multi-scale feature acquisition module is used to perform convolution operations on the sequence embedding features and the structural enhancement features using convolution kernels of different scales, extract contextual dependencies at different spatial distances, and introduce an adaptive attention weighting mechanism to dynamically weight and fuse the outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0043] The bidirectional collaborative interaction module is used to add learnable positional codes to the multi-scale sequence semantic features and the multi-scale structural semantic features respectively, perform bidirectional cross-attention calculation, and obtain the collaborative representation after interaction enhancement.

[0044] The pooling and prediction module is used to obtain an aggregated feature vector by motif-aware attention pooling of the collaborative representation, and input the aggregated feature vector into a hybrid expert classifier. The gating network assigns weights to each expert network and performs a weighted summation of the outputs of each expert network to obtain the prediction result of protein-RNA interaction.

[0045] The aforementioned protein-RNA interaction prediction method and system based on structure enhancement and synergistic attention addresses the shortcomings of incomplete information from a single data source by fusing in vivo structural data with in vitro predicted base pairing probability matrices and enhancing structural features through graph convolutional neural networks. Employing multi-scale convolution and adaptive attention weighting mechanisms enables the model to capture sequence-structure contextual dependencies, enhancing its adaptability. A bidirectional cross-attention mechanism achieves deep synergy and mutual reinforcement of sequence and structural features, while motif-aware attention pooling allows the model to automatically focus on key binding regions. A hybrid expert classifier adaptively selects the most suitable expert network for decision-making through gating, improving the model's generalization ability to handle diverse binding patterns and enhancing the accuracy, robustness, and interpretability of protein-RNA interaction prediction. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the application environment of a protein-RNA interaction prediction method based on structural enhancement and collaborative attention in one embodiment.

[0047] Figure 2 This is a flowchart illustrating a protein-RNA interaction prediction method based on structure enhancement and collaborative attention in one embodiment.

[0048] Figure 3 This is a structural block diagram of a protein-RNA interaction prediction system for structural enhancement and collaborative attention in one embodiment;

[0049] Figure 4 , Figure 5 This is a schematic diagram illustrating the ablation study results of MCoMoE on the structure and modules in one embodiment;

[0050] Figure 6 This is a schematic diagram illustrating the results of interpretability analysis in one embodiment.

[0051] Figure 7 , Figure 8 This is a schematic diagram of the model architecture of MCoMoE in one embodiment;

[0052] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] The protein-RNA interaction prediction method based on structure enhancement and cooperative attention provided in this application embodiment is called the MCoMoE method, and can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer device 110. Computer device 110 can acquire a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and perform length uniformization processing on the RNA sequences, extracting sequence embedding features of the RNA sequences using a sliding window method. Computer device 110 can acquire the base pairing probability matrix of the in vitro secondary structure of the RNA sequences, obtaining an adjacency matrix of the graph structure. Using the features of the in vivo structural data after multi-scale convolution processing as the initial features of the graph nodes, message passing and feature aggregation are performed along the edges defined by the adjacency matrix using a graph convolutional neural network to obtain structural enhancement features. Computer device 110 can perform convolution operations on the sequence embedding features and structural enhancement features using convolution kernels of different scales. The algorithm extracts contextual dependencies at different spatial distances and introduces an adaptive attention weighting mechanism to dynamically weight and fuse outputs at different scales, obtaining multi-scale sequence semantic features and multi-scale structural semantic features. Computer device 110 can add learnable positional encodings to the multi-scale sequence semantic features and multi-scale structural semantic features respectively, and perform bidirectional cross-attention computation to obtain interactively enhanced collaborative representations. Computer device 110 can then use motif-aware attention pooling to obtain aggregated feature vectors from the collaborative representations. These aggregated feature vectors are input into a hybrid expert classifier, where a gating network assigns weights to each expert network and performs a weighted sum of the outputs of each expert network to obtain the prediction results of protein-RNA interactions. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, and other devices.

[0055] In one embodiment, such as Figure 2 As shown, a protein-RNA interaction prediction method based on structure enhancement and collaborative attention is provided, comprising the following steps:

[0056] Step 202: Obtain a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and perform length uniform processing on the RNA sequences, extracting the sequence embedding features of the RNA sequences in a sliding window manner.

[0057] The computer equipment can first complete the construction of the protein-RNA binding sample dataset and the standardization of RNA sequence length, and then complete the extraction of RNA sequence embedding features through a sliding window and a pre-trained RNA base model.

[0058] In one embodiment, the provided protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include data acquisition and feature extraction processes. Specifically, the process includes: extracting corresponding RNA sequence fragments from RNA-binding protein regions as positive samples, generating negative samples by random sampling or shuffling the RNA sequence order, associating in vivo structural data corresponding to each RNA sequence, and constructing a protein-RNA binding sample dataset; obtaining a maximum length threshold, and using the maximum length threshold as a benchmark, performing length unification processing on the RNA sequences to obtain length-unified RNA sequences; segmenting the length-unified RNA sequences using a sliding window, inputting the segmented RNA sequences into a pre-trained RNA base model based on the Transformer architecture, and using a multi-layer self-attention mechanism to output feature vectors as sequence embedding features of the RNA sequences.

[0059] The protein-RNA binding sample dataset was derived from eCLIP-seq experimental data in the ENCODE database. Specifically, to achieve high dataset quality and mitigate batch effects, information on 844,854 HepG2 and K562 human RNA binding sites was collected from the same batch of eCLIP-seq data in the ENCODE database. The eCLIP-seq sequencing data used the human genome GRCh37 / HG19 as a reference genome, and all binding sites were processed using a unified pipeline. Next, the computer system used a peak region generated by each RNA-binding protein (RBP) in the eCLIP-seq experiment as a positive sample binding site. That is, the peak generated by each human RBP was considered a binding site, and the corresponding RNA sequence fragment was extracted as a positive sample. For each RNA-binding protein, the computer system preferentially selected the top 5000 peaks with the highest confidence as positive samples, and the number of samples for each RNA-binding protein needed to be greater than 2000.

[0060] In this configuration, each RNA sequence is fixed to a length of 101 nucleotides. Regions longer than 101 nucleotides are truncated from both sides, while regions shorter than 101 nucleotides are extended to both sides. In other words, the computer device can obtain a preset maximum length threshold, which in this embodiment is set to 101 nucleotides. Based on this maximum length threshold, all RNA sequences are processed to achieve a uniform length, resulting in RNA sequences of 101 nucleotides through truncation or extension operations.

[0061] The computer device can randomly generate candidate sequences within genomic regions outside the positive sample regions. The number of candidate sequences is twice the number of positive samples. Sequences of the same number as the number of positive samples are randomly selected from the candidate sequences as negative samples, ensuring a balanced ratio of positive and negative samples. Simultaneously, the in vivo structural data corresponding to each RNA sequence is correlated. Specifically, the computer device can acquire in vivo structural data, i.e., icSHAPE-Seq data. icSHAPE-Seq data provides in vivo RNA structural information of the entire transcriptome at single-nucleotide resolution. Each nucleotide position corresponds to an icSHAPE structure score, reflecting the structural flexibility or pairing state of nucleotides in the cellular native environment. In this embodiment, since the ENCODE data uses the GRCh37 / HG19 reference genome, while the icSHAPE-Seq data uses the GRCh38 / HG38 reference genome, the CrossMap tool and the hg19ToHg38 conversion package are used to convert the genomic coordinates of all binding sites in the ENCODE dataset from HG19 to HG38, ensuring that the sequence data and structural data are aligned under the same reference genome.

[0062] In other words, the negative sample acquisition method involves randomly generating twice the number of positive samples from the positive sample region, resulting in 1000 negative samples of length 101 nt. Finally, for each RBPs dataset, models were built for the HepG2 and K562 cell lines respectively. The training and test sets were randomly divided in a 4:1 ratio, and 20% of the training set was randomly selected as the test set. Given the potential for literature bias in public datasets and the approximation limitations of negative sample generation methods, extensive validation strategies and systematic testing and comparisons were conducted on 75 different protein datasets (e.g., AARS_K562) containing the HepG2 and K562 cell lines in the ENCODE database to mitigate potential data bias, thereby ensuring the robustness and generalization ability of the model evaluation.

[0063] Computer devices can use a sliding window approach to segment and encode RNA sequences of uniform length. Specifically, the sliding window size is set to 3-mer (i.e., three consecutive nucleotides as a basic unit), and the step size is set to 1. For an RNA sequence of 101 nucleotides in length, a total of 99 window positions are generated after scanning with a 3-mer sliding window. The computer device can input the three consecutive nucleotides within each 3-mer window into a pre-trained RNA Foundation Model (RNA-FM) based on the Transformer architecture. RNA-FM is a large-scale language model specifically pre-trained for RNA sequences, and its core component is a multi-head self-attention mechanism. During the encoding process, each 3-mer window is first mapped to an initial embedding vector, and then processed through multiple layers of self-attention layers of the Transformer. Specifically, the multi-head self-attention mechanism uses parallel computation of multiple attention heads to interact with each position in the input sequence with all other positions, capturing long-distance dependencies and global contextual information within the sequence. Each self-attention layer calculates attention weights through query-key-value projection operations, and the output is a context-aware weighted sum of value vectors.

[0064] In one embodiment, a core component of the Transformer-based RNA pre-trained large model is the multi-head attention mechanism. By computing multiple attention heads in parallel, multidimensional features and long-range dependencies of the input sequence can be captured, significantly enhancing the model's ability to parse and understand RNA sequences. Specifically, each self-attention layer essentially operates through query-key-value projection, in which the output is a context-aware weighted sum of value vectors, determined by the affinity between queries and keys. The basic process of attention computation can be described as follows: ; ; ;in, Represents the query matrix, used in conjunction with the key matrix. Calculate similarity to determine which locations' information needs attention; The key matrix represents the matrix that is related to the query matrix. The basis for calculating attention weights; The value matrix stores the actual information to be extracted, and is a weighted aggregation of values ​​through attention weights. This represents the weight matrix of the output layer, which maps the result of concatenating multiple heads to the final output dimension. This represents an attention head, where each head projects the input to the first attention head. The query matrix consists of a query space, a key space, and a value space. AND key matrix The dot product generates the affinity matrix, which is then normalized by softmax to obtain the attention weights. These weights quantify the relevance of each query-key vector pair and reflect the contextual dependencies in the attention mechanism. The dimension of the key vector, scaling factor This is used to prevent gradient vanishing or exploding during optimization; finally, the attention weights are applied to the value matrix through a weighted summation. Each position represents a synthesis context-enhanced embedding.

[0065] After processing by the multi-layer self-attention mechanism of RNA-FM, each 3-mer window is encoded as a 640-dimensional dense feature vector. A 101-nucleotide RNA sequence, after sliding window and RNA-FM encoding, outputs a sequence embedding feature matrix H_seq with dimensions of 99×640, where 99 represents the sequence length dimension and 640 represents the feature channel dimension. Based on this, to preserve the absolute positional order information of each nucleotide in the RNA sequence, a learnable positional encoding PE with the same dimension as H_seq is superimposed on the sequence embedding feature matrix H_seq, resulting in a sequence feature representation with positional information, X_seq = H_seq + PE. This learnable positional encoding is updated through backpropagation during model training, enabling the model to adaptively learn the importance distribution of RBP binding sites at different sequence positions.

[0066] Step 204: Obtain the base pairing probability matrix of the in vitro secondary structure of the RNA sequence to obtain the adjacency matrix of the graph structure. Use the features of the in vivo structural data after multi-scale convolution processing as the initial features of the graph nodes. Perform message passing and feature aggregation of the graph convolutional neural network along the edges defined by the adjacency matrix to obtain the structure enhancement features.

[0067] Computer equipment can construct a graph adjacency matrix based on the prediction results of RNA secondary structure in vitro, combine it with in vivo structural data to generate initial features of graph nodes through multi-scale convolution, and then complete the feature transfer and aggregation between nodes through a graph convolutional network, finally outputting structural enhancement features that integrate in vivo and in vitro structural information.

[0068] In one embodiment, the provided protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include a feature enhancement process, specifically including: inputting the RNA sequence into an RNA secondary structure prediction tool, predicting the in vitro secondary structure of the RNA sequence based on a thermodynamic model, calculating the probability of pairing between any two nucleotide bases in the RNA sequence to obtain a base pairing probability matrix, and using the base pairing probability matrix as the adjacency matrix of the graph structure; inputting the in vivo structure data into a multi-scale convolutional neural network, performing convolution operations through various multi-scale convolutional kernels to obtain the weights of features at each scale and completing weighted fusion, finally obtaining the initial features of each graph node; inputting the initial features into the graph convolutional neural network, performing feature aggregation and nonlinear transformation along the edges defined by the adjacency matrix, outputting the updated feature vectors of each node, and concatenating all the updated feature vectors in sequence position order to obtain the structure-enhanced features.

[0069] RNA structural features are divided into two categories: icSHAPE structure scores and the secondary structure pairing probability matrix predicted by RNAplfold. The secondary structure pairing probability matrix integrates the spatial correlation features of one-dimensional sequence adjacency relationships and RNA secondary base pairing. To enhance structural feature information, in this embodiment, the RNAplfold tool can be used to predict the local secondary structure features at each position in the input RNA sequence, ultimately converting them into graph structure information as a base pairing probability matrix. RNAplfold uses the thermodynamic nearest neighbor method to return the secondary structures associated with the minimum free energy of the RNA sequence. The obtained secondary structures are represented by dots and parentheses; dots represent unpaired bases, and a pair represented by left and right parentheses represents a single base pair.

[0070] Specifically, computer equipment can extract RNA sequences using a 3-mer sliding window RNA basic model, obtaining a 640-dimensional feature vector of length 99, to more comprehensively acquire local and global semantic feature information of RNA. Here, the nucleotide base ri in each RNA sequence is represented as the RNA basic model (ri). RNA structural features include two representations: one is represented by icSHAPE structure scores, where the secondary structure information of each nucleotide residue is represented by the pairing probability icSHAPE(ri, cell); the other is represented by the base pairing probability matrix of the RNA secondary structure predicted by the RNAplfold model. This matrix includes not only the linear sequence neighbor relationships but also the base pairing relationships between two nucleotides in the secondary structure. If nucleotide residues ri and rj are adjacent in the primary sequence (i.e., rj...), the pairing probability matrix represents the secondary structure of RNA. If two nucleotides pair up in a secondary structure, then an edge exists (E(ri,rj)). If two nucleotides pair up in a secondary structure, then an edge exists, and the weight is determined by the probability of base pairing. This weight is mainly used to specify the weight value of the secondary structure pairing edge S(ri,rj). Ultimately... Aij = E(ri,rj) + S(ri,rj); where hi represents the embedding vector of the i-th nucleotide; || This merges sequence and structural features along the channel dimension. It is The adjacency matrix (L is the sequence length).

[0071] In this embodiment, the computer device can input a standardized RNA sequence that has undergone length uniformization into the RNAplfold secondary structure prediction tool. Based on the RNA nearest neighbor thermodynamic model, the tool predicts the in vitro secondary structure of the entire sequence. It iterates through any two nucleotide bases in the sequence and calculates the probability values ​​of stem-loop pairing between them. Based on all pairing probability values, a two-dimensional base pairing probability matrix is ​​generated. The row and column indices of this matrix correspond to nucleotides at different positions in the RNA sequence, and the matrix elements represent the pairing confidence of the corresponding two bases. The computer device can directly use this base pairing probability matrix as a graph structure adjacency matrix to characterize the topological association relationship of RNA. The values ​​in the matrix correspond to the weights of the edges connecting the nodes in the graph, thereby simultaneously characterizing the linear adjacency relationship of the primary sequence and the secondary spatial base pairing relationship of RNA.

[0072] Computer equipment can input in vivo icSHAPE single-base structure data that is matched to RNA sequences base by base into a multi-scale convolutional neural network. The network can perform convolution operations in parallel using one-dimensional convolutional kernels of various sizes to extract local features of in vivo structures in short-range, medium-range, and long-range intervals. The computer equipment can calculate adaptive weights for the scale features output by each type of convolutional kernel, and perform weighted fusion of the multi-scale features based on the weights to obtain an integrated feature vector corresponding to each nucleotide site. Each nucleotide is used as an independent node in the graph, and the fused site feature vector is used as the initial feature of the corresponding graph node.

[0073] Specifically, to comprehensively capture the semantic information and hierarchical relationships of sequence and structural features, computer devices can use multi-scale local window convolutions to flexibly capture short-range, medium-range, and long-range semantic information. Furthermore, adaptive attention weighting is integrated, which can automatically learn the importance of feature information at different scales. Compared to simple feature averaging methods, adaptive attention weighting is better able to identify the most important information in both sequence and structural information, focusing on the most useful local features. The formulas for extracting local structural features and calculating attention weights are as follows: ; ;in, These represent composite operations with scales of 3, 5, and 7 branches, respectively. ; Indicates attention weights. ;in, This represents the different unnormalized attention scores, i.e., the original weight coefficients corresponding to each branch.

[0074] Previously, computer equipment could use linear interpolation to align RNA sequences and RNA structures, ensuring absolute alignment in position to maintain the integrity of the secondary structure matrix and avoid edge effect problems. Then, relying on multi-scale mechanisms, the diverse contextual dependencies of sequence structural features at different spatial distances can be captured to better characterize the specific binding preferences of RNA molecules.

[0075] The computer device can input the initial features of all graph nodes into a graph convolutional neural network. The network uses a pre-constructed base pairing probability adjacency matrix as the pathway for information transmission between nodes. Cross-node feature message passing is performed along the node connection edges defined by the adjacency matrix. All feature information carried by each node itself, its linearly adjacent bases, and its secondary paired bases is aggregated, and feature updates are completed through nonlinear transformations. After completing multi-layer graph convolution iterations, the updated feature vector of each nucleotide node is output. According to the order of the base positions in the original RNA sequence, the updated feature vectors of all nodes are concatenated to form a complete two-dimensional feature matrix, which is the structure enhancement feature that integrates in vivo measured dynamic structure and in vitro predicted secondary topological information.

[0076] Step 206: Convolution operations are performed on sequence embedding features and structural enhancement features using convolution kernels of different scales to extract contextual dependencies at different spatial distances. An adaptive attention weighting mechanism is introduced to dynamically weight and fuse outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0077] After acquiring sequence embedding features and structural enhancement features, computer devices can use multi-scale convolution operations combined with an adaptive attention weighting mechanism to perform multi-level contextual feature mining and dynamic fusion on the two types of features respectively, fully capturing the contextual dependencies of RNA sequences and structures at different spatial distances, thereby obtaining multi-scale sequence semantic features and multi-scale structural semantic features with stronger representation capabilities.

[0078] In one embodiment, the protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include convolution operations and feature weighted fusion processes. Specifically, the process includes: uniformly processing the lengths of sequence embedding features and structure enhancement features to ensure that sequence features and structural features correspond at sequence positions, resulting in processed sequence embedding features and structure enhancement features; using the processed sequence embedding features and structure enhancement features as two independent inputs, with each input undergoing parallel convolution operations to obtain intermediate features at each scale corresponding to each feature, thereby capturing sequence and structural contextual dependencies at different spatial distances; assigning corresponding attention weights to the intermediate features at each scale corresponding to each feature, and weighting and summing the outputs of each convolution operation branch according to their respective attention weights to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0079] Computer devices can leverage graph convolutional neural networks (GCNs) to enhance structural information features. Specifically, the icSHAPE structural features, further processed through multi-scale local window convolutions, are integrated with the base probability pairing matrix predicted by the RNAplfold model. A GCN module guides the transfer and aggregation of node features between edges, thus enhancing local information through global structural information. Here, G = (V, E, A) represents the graph structure of RNA. E(ri,rj) represents the feature embedding of each nucleotide ri, E(ri,rj) represents the probability of base interaction between two nucleotides, and A is the adjacency matrix corresponding to the weights connecting the two points.

[0080] In this embodiment, the computer device can perform length-uniform alignment processing on the extracted sequence embedding features and structural enhancement features. Since sequence features and structural features are prone to feature position offsets and inconsistent dimensional lengths during extraction and enhancement, by unifying feature lengths and calibrating feature arrangement positions, the sequence embedding features and structural enhancement features achieve a precise one-to-one correspondence in both the base sequence dimension and spatial position dimension. This ensures the consistency and effectiveness of subsequent feature convolution operations and feature fusion, ultimately resulting in processed sequence embedding features and structural enhancement features with aligned positions and unified dimensions.

[0081] Next, the computer equipment can perform dual-path parallel multi-scale convolution operations to mine multi-distance contextual dependencies. Specifically, the aligned sequence embedding features and structure enhancement features can be set as two independent input branches, which are processed in parallel using the same convolutional operation structure. Each input uses convolutional kernels of various scales to perform convolution operations, respectively adapting to capture short-distance, medium-distance, and long-distance base contextual association information, and obtaining multiple sets of intermediate features at each scale for each feature, thereby comprehensively extracting the base arrangement dependencies of RNA sequences and the spatial conformation dependencies of RNA structures at different spatial distances, overcoming the shortcomings of single-scale convolution which can only capture fixed interval features and has incomplete feature extraction. Finally, multi-scale feature dynamic fusion is completed based on an adaptive attention weighting mechanism. For each set of intermediate features at different scales output by each feature, the model autonomously learns and assigns corresponding adaptive attention weights. The weight size corresponds to the contribution of different scale features to the protein-RNA interaction prediction task, without the need for manual parameter preset. Within each feature path, the intermediate scale features output from each convolutional branch are weighted and fused according to their corresponding attention weights, automatically strengthening high-contribution scale features and weakening ineffective and redundant features. Specifically, sequence embedding features are fused through multi-scale convolution and adaptive weighting to output multi-scale sequence semantic features; structural enhancement features are processed through the same procedure to output multi-scale structural semantic features. Both types of features are subsequently used for bidirectional cross-attention interaction enhancement processing.

[0082] Step 208: Add learnable positional codes to the multi-scale sequence semantic features and multi-scale structural semantic features respectively, and perform bidirectional cross-attention calculation to obtain the interactively enhanced collaborative representation.

[0083] After acquiring multi-scale sequence semantic features and multi-scale structural semantic features, the computer device can supplement the base position information of the two types of features by independently learnable position encoding, and realize the mutual constraint and bidirectional enhancement of sequence features and structural features through a bidirectional cross-attention mechanism, fully explore the deep correlation between RNA sequence information and spatial structural information, and finally obtain the interactively enhanced collaborative representation.

[0084] In one embodiment, the provided protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include encoding and feature enhancement processes. Specifically, the process includes: superimposing a first learnable positional encoding on multi-scale sequence semantic features to obtain sequence interaction features with positional information; superimposing a second learnable positional encoding on multi-scale structural semantic features to obtain structural interaction features with positional information; the parameters of the first and second learnable positional encodings are independent of each other and are updated through backpropagation during model training; using the sequence interaction features with positional information as queries and the structural interaction features with positional information as keys and values, performing sequence-to-structure cross-attention calculation to obtain sequence features enhanced by structural information constraints; using the structural interaction features with positional information as queries and the sequence interaction features with positional information as keys and values, performing structure-to-sequence cross-attention calculation to obtain structural features enhanced by sequence information constraints; and concatenating and fusing the sequence features enhanced by structural information constraints and the structural features enhanced by sequence information constraints to obtain a collaboratively enhanced representation.

[0085] Computer devices can superimpose independent learnable positional codes onto multi-scale sequence semantic features and multi-scale structural semantic features, respectively. Specifically, a first learnable positional code and a second learnable positional code with independent parameters can be set. These two types of positional codes correspond to the sequence feature branch and the structural feature branch, respectively, with no parameter sharing. The computer device can superimpose the first learnable positional code onto the multi-scale sequence semantic features to obtain sequence interaction features carrying base position arrangement information; and superimpose the second learnable positional code onto the multi-scale structural semantic features to obtain structural interaction features carrying base position arrangement information. All learnable positional code parameters are not fixed and can be continuously iteratively updated during the overall model training process using the backpropagation algorithm.

[0086] Next, the computer device can perform a one-way cross-attention calculation from sequence to structure, completing the constraint enhancement of sequence features by structural information. Using the sequence interaction features with added positional encoding as query features, and the structural interaction features carrying complete positional information as key and value features, cross-attention operations are performed. The computer device can adaptively allocate attention weights to different structural regions by calculating the correlation similarity between sequence features and structural features, correct the feature weights of sequence motifs using the topological constraints of RNA spatial structure, filter and strengthen sequence feature information that matches the real RNA structure, suppress interference from invalid sequence features without structural support, and finally obtain sequence features enhanced by structural information constraints.

[0087] In a symmetrical configuration to the aforementioned cross-attention process, the computer device can use structural interaction features with added positional encoding as query features, and sequence interaction features carrying complete positional information as key and value features, to perform reverse cross-attention operations. Based on the conserved arrangement preference of RNA base sequences, the weight distribution of RNA secondary structure features is corrected, ineffective structural features without sequence conservation support are weakened, and structural representations conforming to real protein binding rules are strengthened, ultimately yielding structural features enhanced by sequence information constraints.

[0088] Finally, the computer device can splice and fuse the sequence features enhanced by structural information constraints with the structural features enhanced by sequence information constraints in the feature dimension, integrating the bidirectional constraints and complementary information of sequence pairs and structure pairs, to obtain global features that simultaneously contain RNA base sequence semantics, spatial topology, base position information, and bidirectional interactive associations. This is the interactively enhanced collaborative representation, providing high-dimensional and robust fused feature inputs for subsequent feature aggregation and classification prediction.

[0089] In one embodiment, a computer device can calculate the attention similarity between each position in the sequence interaction feature with location information and all positions in the structural interaction feature with location information; after scaling the attention similarity, it is normalized to attention weights using the Softmax function; and the structural interaction features are weighted and aggregated using the attention weights to obtain the sequence features enhanced with structural information.

[0090] In one embodiment, a computer device can calculate the attention similarity between each position in the structural interaction feature with location information and all positions in the sequence interaction feature with location information; after scaling the attention similarity, it is normalized to attention weights using the Softmax function; and the sequence interaction features are weighted and aggregated using the attention weights to obtain sequence features enhanced with sequence information.

[0091] Traditional models simply splice together sequence and structural information, failing to allow for mutual guidance and reinforcement between the sequences and structures. In this embodiment, a collaborative attention mechanism is used to enable adaptive learning of RNA sequence information and structural feature information, achieving bidirectional dynamic interaction and enhancement. This accurately captures sequence motifs under structural constraints and structural conformations under sequence regulation, thereby significantly improving the model's predictive performance.

[0092] ; ;

[0093] ; ; ;in, , These represent the initial features of the sequence mode and the structure mode, respectively; , These represent the final input representations of the sequence mode and the structure mode, respectively. A collaborative attention function representing the fusion of sequence and structural modes; The original embedding vector represents the interaction between the sequence and structural information input to a specific attention module; The corresponding embedding vector, which interacts with sequence and structural information, serves as the generation key. Sum The raw materials. , These represent the query, key, and value vectors resulting from the interaction of sequence and structural information, respectively. This represents the scaling factor for the vector dimension within each attention head; This represents a multilayer perceptron function module; , This represents the first-layer feedforward fusion function, and the inputs are the sequence query structure and the structure query sequence, respectively. The second-layer fusion function represents the interaction of cross-attention information between sequences and structures; The representation layer normalization function; cross-attention represents the cross-attention method between sequence and structural information; , (i=1,2) represents the collaborative attention module, where L represents the sequence length and C represents the number of feature channels; the addition of positional encoding (PE) is to obtain learnable positional encodings of the positional information of each nucleotide sequence, and finally, dynamic gating is used to adaptively output feature importance for further training.

[0094] Step 210: The collaborative representation is processed by motif-aware attention pooling to obtain an aggregated feature vector. The aggregated feature vector is then input into a hybrid expert classifier. The gating network assigns weights to each expert network and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions.

[0095] After obtaining the bidirectional interactive enhanced collaborative representation, the computer device can adaptively filter and compress high-dimensional global features through motif-aware attention pooling, focusing on the effective feature information of key RNA binding motif regions, discarding irrelevant and redundant features, and compressing the long-dimensional sequence-level collaborative representation into a fixed-dimensional aggregated feature vector.

[0096] In one embodiment, a protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include a pooling process, specifically comprising: calculating the attention score at each position in the collaborative representation using a multilayer perceptron; normalizing each attention score using a Softmax function to obtain the attention weight corresponding to each position; and performing a weighted summation of the collaborative representation based on the attention weights to compress the regional features containing binding motifs into an aggregated feature vector.

[0097] The computer device can calculate the attention score of the co-representation position by position and perform normalization processing. Specifically, in this embodiment, a multilayer perceptron is used to perform nonlinear mapping calculation on the feature vector corresponding to each base position in the co-representation, and output the attention score corresponding to each position one by one. This attention score can quantify the contribution of different base regions to the protein-RNA interaction event, effectively representing the confidence level of each position belonging to the RBP binding motif. The computer device can further use the Softmax function to perform overall normalization processing on the attention scores of all base positions, so that the attention weight values ​​of all positions are normalized to the effective range and satisfy the global weight distribution constraint, obtaining the standardized attention weight uniquely corresponding to each nucleotide position, realizing the model's autonomous distinction between key motif regions and invalid background regions.

[0098] Next, the computer device can perform global feature weighting aggregation based on normalized attention weights to generate an aggregated feature vector. Specifically, based on the adaptive attention weights at each position, the computer device can use each set of attention weights to perform element-wise weighted operations on the feature vectors at corresponding positions in the collaborative representation, and perform global summation and aggregation of the weighted features at all positions. This compresses the long sequence of collaborative representations containing scattered binding motif information into a single fixed-dimensional global feature vector, ultimately obtaining an aggregated feature vector focusing on key binding motifs.

[0099] In one embodiment, the provided protein-RNA interaction prediction method based on structure enhancement and collaborative attention may further include a process for obtaining the final prediction result. Specifically, this process includes: inputting the aggregated feature vector into a gated network composed of two linear transformation layers to obtain the original weight scores, and normalizing the original weight scores to obtain adaptive weights; assigning each adaptive weight to a separate independent expert network, inputting the aggregated feature vector into each independent expert network, and having each independent expert network output its prediction score; and weighting and summing each prediction score with its respective adaptive weight to obtain the final prediction result of the protein-RNA interaction.

[0100] The computer device can input the preprocessed aggregated feature vector into a gated network consisting of two linear transformation layers. Relying on these two linear transformation layers, the network performs deep feature mapping on the aggregated feature vector, outputting the original weight scores for each independent expert network. Each adaptive weight quantifies the degree to which different expert networks fit the current RNA-RBP binding prediction task, achieving dynamic adaptive weight allocation. The computer device can construct multiple sets of structurally independent expert networks with non-shared parameters, inputting the same aggregated feature vector into each of these independent expert networks. Each set of independent expert networks can autonomously learn different RNA sequence motif features, secondary structure conformation features, and protein binding preference patterns. They perform feature parsing and inference on the input aggregated feature vector, outputting independent prediction scores adapted to their own learning patterns, achieving differentiated fitting for diverse protein-RNA binding scenarios.

[0101] The computer device can match the predicted scores output by each independent expert network with the adaptive weights generated by the gating network, and then fuse and integrate multiple sets of differentiated predicted scores through weighted summation. Utilizing the dynamic weights learned by the gating network, it automatically strengthens expert predictions that fit the current sample combination pattern, weakens interference from expert outputs with lower matching degrees, and integrates the advantageous features of all expert networks, ultimately outputting a more accurate and robust protein-RNA interaction prediction result.

[0102] Before implementing a classifier, motif-aware attention pooling is often used. This method allows the model to automatically learn regions rich in motifs within a sequence while suppressing irrelevant regions, thereby extracting relatively pure binding features and achieving intelligent pooling that weights important regions and weakens irrelevant regions. Faced with different RBP binding sequences, this embodiment employs a hybrid expert classifier. Based on different RNA structures or sequence morphologies, it automatically assigns the prediction to the most proficient expert, thus enhancing the model's diversity, robustness, and generalization ability. The hybrid expert classifier consists of multiple independent expert networks. and a gated network The model is composed of experts who learn a specific subset of features. The introduced gating network adaptively assigns expert weights to each sample based on the input features, enabling the model to adopt the optimal decision mode for RNAs with different topologies. ; = =Linear(ReLU(Linear( ))); ;in, It is to satisfy The expert weight vector; y represents the final classification prediction result, which is the weighted average of all expert outputs; These are the variables input to the expert network after passing through the gating network. Additionally, Softmax represents the normalized exponential function; ReLU represents the modified linear unit activation function; and bn represents batch normalization. This represents the weight matrix of a linear transformation layer. This represents the bias term for gating at a certain layer; This represents the i-th expert network. Represents the individual prediction output of the i-th expert; Linear refers to a linear layer.

[0103] Meanwhile, the model employs a binary cross-entropy loss function in its binary classification task, measuring the difference between the model's predicted probabilities and the actual labels to achieve more accurate classification. The formula for the binary cross-entropy loss function is: Where N represents the mini-batch size, Represents the true label, The combination probability predicted by the model; It is an L2 regularization term that applies to all model parameters, mitigating overfitting during training by reducing weights.

[0104] This application presents a protein-RNA interaction prediction method based on structure enhancement and collaborative attention. MCoMoE demonstrates superior performance compared to baseline methods under both static and dynamic cellular conditions. Specifically, under static cellular conditions, it is compared with five other state-of-the-art computational methods—diPaRIS, HDRNet, PrismNet, GraphProt, and iDeep—on a dataset of 75 static RBP binding sites. The RBP dataset's binding site partitioning for both training and testing sets follows the PrismNet dataset preprocessing method to ensure data accuracy and consistency. MCoMoE outperforms all other models in both AUC and ACC. On the 75 RBP dataset, MCoMoE surpasses the best-performing HDRNet model, with an average ACC increase of 3.1%. Furthermore, to further evaluate the stability of MCoMoE on the 75 datasets, AUC, ACC, F1, and MCC performance metrics are analyzed and compared.

[0105] Since RBP binding is significantly influenced by different cellular environments, this study evaluates the performance of MCoMoE in predicting RNA-protein interaction classification tasks under dynamic cellular conditions. Using a dataset of 32 RBPs from K562 and HepG2 cell lines, the performance of the same protein on the HepG2 cell line with training weights from K562 was dynamically evaluated, as well as the performance changes on the K562 cell line with training weights from HepG2. The dynamic test results of MCoMoE were also compared with those of diPaRIS, HDRNet, Prismnet, GraphProt, and iDeep. Whether using HepG2 or K562 as the training set for dynamic testing of another cell, the MCoMoE model outperformed the other models. When predicting RBP sites in K562 cells after training on HepG2 cells, MCoMoE achieved average AUPRC and ACC 1.8% and 2.1% higher than the best performance metrics, respectively. Then, when predicting RBP sites in HepG2 cells after training on K562 cells, MCoMoE's average AUPRC was 0.9% lower than the HDRNet method, while its ACC was 1.2% higher. Furthermore, regardless of the cell type used as the benchmark for AQR protein performance testing, MCoMoE clearly outperformed all other models and could intuitively distinguish between positive and negative samples, highlighting its superior feature learning ability and robustness in cross-prediction.

[0106] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0107] In one embodiment, such as Figure 3 As shown, a protein-RNA interaction prediction system based on structure enhancement and collaborative attention is provided, comprising: a data acquisition module 310, a structure enhancement module 320, a multi-scale feature acquisition module 330, a bidirectional collaborative interaction module 340, and a pooling and prediction module 350, wherein:

[0108] The data acquisition module 310 is used to acquire a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and to perform length uniform processing on the RNA sequences and extract the sequence embedding features of the RNA sequences in a sliding window manner.

[0109] The structure enhancement module 320 is used to obtain the base pairing probability matrix of the in vitro secondary structure of the RNA sequence, obtain the adjacency matrix of the graph structure, and use the features of the in vivo structure data after multi-scale convolution processing as the initial features of the graph nodes. Message passing and feature aggregation of the graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain the structure enhancement features.

[0110] The multi-scale feature acquisition module 330 is used to perform convolution operations on sequence embedding features and structural enhancement features using convolution kernels of different scales, extract contextual dependencies under different spatial distances, and introduce an adaptive attention weighting mechanism to dynamically weight and fuse outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0111] The bidirectional collaborative interaction module 340 is used to add learnable positional codes to multi-scale sequence semantic features and multi-scale structural semantic features respectively, perform bidirectional cross-attention calculation, and obtain the collaborative representation after interaction enhancement.

[0112] The pooling and prediction module 350 is used to obtain an aggregated feature vector by motif-aware attention pooling of the collaborative representation. The aggregated feature vector is then input into a hybrid expert classifier, where a gating network assigns weights to each expert network and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions.

[0113] In one embodiment, the data acquisition module 310 is further configured to extract the corresponding RNA sequence fragments of the RNA-binding protein region as positive samples, and generate negative samples by random sampling or shuffling the RNA sequence order, associate the in vivo structural data corresponding to each RNA sequence, and construct a protein-RNA binding sample dataset; obtain the maximum length threshold, and perform length unification processing on the RNA sequence based on the maximum length threshold to obtain the length-unified RNA sequence; use a sliding window to segment the length-unified RNA sequence, input the segmented RNA sequence into a pre-trained RNA basic model based on the Transformer architecture, and use a multi-layer self-attention mechanism to output feature vectors as sequence embedding features of the RNA sequence.

[0114] In one embodiment, the structure enhancement module 320 is further configured to input the RNA sequence into an RNA secondary structure prediction tool, predict the in vitro secondary structure of the RNA sequence based on a thermodynamic model, calculate the probability of pairing between any two nucleotide bases in the RNA sequence to obtain a base pairing probability matrix, and use the base pairing probability matrix as the adjacency matrix of the graph structure; input the in vivo structure data into a multi-scale convolutional neural network, perform convolution operations through various multi-scale convolution kernels to obtain the weights of the features at each scale and complete weighted fusion, and finally obtain the initial features of each graph node; input the initial features into the graph convolutional neural network, perform feature aggregation and nonlinear transformation along the edges defined by the adjacency matrix, output the updated feature vectors of each node, and concatenate all the updated feature vectors in sequence position order to obtain the structure enhancement features.

[0115] In one embodiment, the multi-scale feature acquisition module 330 is further configured to perform length unification processing on the sequence embedding features and structural enhancement features, so that the sequence features and structural features correspond in sequence position, to obtain the processed sequence embedding features and structural enhancement features; the processed sequence embedding features and structural enhancement features are used as two independent inputs, and each input is subjected to convolution operation in parallel to obtain intermediate features of each scale corresponding to each feature, so as to capture the sequence and structural context dependencies at different spatial distances; corresponding attention weights are assigned to the intermediate features of each scale corresponding to each feature, and the outputs of each convolution operation branch are weighted and summed according to their respective attention weights to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

[0116] In one embodiment, the bidirectional collaborative interaction module 340 is further configured to superimpose a first learnable positional encoding on multi-scale sequence semantic features to obtain sequence interaction features with positional information; superimpose a second learnable positional encoding on multi-scale structural semantic features to obtain structural interaction features with positional information; the parameters of the first and second learnable positional encodings are independent of each other and are updated through backpropagation during model training; using the sequence interaction features with positional information as a query and the structural interaction features with positional information as keys and values, perform sequence-to-structure cross-attention calculation to obtain sequence features enhanced by structural information constraints; using the structural interaction features with positional information as a query and the sequence interaction features with positional information as keys and values, perform structure-to-sequence cross-attention calculation to obtain structural features enhanced by sequence information constraints; and concatenate and fuse the sequence features enhanced by structural information constraints and the structural features enhanced by sequence information constraints to obtain a collaborative representation with enhanced interaction.

[0117] In one embodiment, the bidirectional collaborative interaction module 340 is further configured to calculate the attention similarity between each position in the sequence interaction feature with location information and all positions in the structural interaction feature with location information; after scaling the attention similarity, it is normalized to attention weights through the Softmax function; and the structural interaction features are weighted and aggregated through the attention weights to obtain the sequence features enhanced with structural information.

[0118] In one embodiment, the pooling and prediction module 350 is further configured to use a multilayer perceptron to calculate the attention score at each position in the collaborative representation, use the Softmax function to normalize each attention score to obtain the attention weight corresponding to each position, and perform a weighted summation on the collaborative representation according to the attention weight to compress the regional features containing the binding motif into an aggregated feature vector.

[0119] In one embodiment, the pooling and prediction module 350 is further configured to input the aggregated feature vector into a gated network composed of two linear transformations to obtain the original weight scores, and to normalize the original weight scores to obtain adaptive weights; to assign each adaptive weight to an independent expert network, and to input the aggregated feature vector into each independent expert network, with each group of independent expert networks outputting each predicted score; and to perform a weighted summation of each predicted score with each adaptive weight to obtain the final prediction result of protein-RNA interaction.

[0120] In one embodiment, to evaluate the superiority of the RNA base model feature extraction method, the Bert large model used by the HDRNet method was selected for ablation experiments. Figure 4 The radar chart in the figure shows that the RNA-FM model significantly outperforms the BERT model in terms of AUC, ACC, AUPRC, F1, and MCC metrics. Specifically, ACC is improved by 2.22%, and AUPRC by 3.01%. The superior performance of the RNA-FM model is likely due to its specific focus on RNA-specific features, which are highly aligned with downstream tasks, making it more suitable for RBP-based prediction scenarios. (Figure caption: n=75)

[0121] Structural information can capture the spatial conformational features of RNA, compensating for the inability of single sequence information to reflect key information such as long-range base pairing and stem-loop structures. Therefore, a comparative analysis of sequence and structural information was conducted here, such as... Figure 4As shown in the violin diagram of B, a significant performance difference can be observed with and without structural information, further illustrating the importance of structural information. Furthermore, since structural information can be divided into in vivo and in vitro structures, fusing in vitro structural information after obtaining icSHAPE structural information becomes a promising direction. Therefore, here we fuse the base probability pairing matrix predicted by RNAplfold based on icSHAPE structural information to enhance structural feature information. (See diagram for example.) Figure 4 As shown in the camel-cup plot of CD, in the structural information of the enhanced GCN (Graph Convolutional Neural Network), the AUC and ACC indices have been improved to a certain extent, especially the AUC and ACC values, which have increased by 3.95% and 1.4%, respectively.

[0122] To assess the contribution of some components in MCoMoE, ablation experiments were conducted. For example... Figure 5 As shown in the GH umap plot, umap plots of ZNF800_HepG2 and ZNF800_K562 before and after the bidirectional sequence structure collaborative attention module are plotted. It can be observed that the distribution of positive and negative samples after fusion is very clear and visible, which is in stark contrast to the umap before fusion where positive and negative samples are mixed together. Figure 4 As shown in EF, the box plots drawn for the ablation study of the MoE module demonstrate the effectiveness and feasibility of the MCoMoE method.

[0123] In summary, MCoMoE demonstrates outstanding robustness and rationality in various ablation studies on feature extraction and structural design, significantly improving the model's ability to learn and summarize from data.

[0124] By visualizing the contribution of key bases and structural signals, interpretability analysis helps verify whether the model captures the true binding motifs and spatial conformation features of RBPs, revealing the RNA sequences and structural regions that the model truly focuses on during prediction. This transforms the model from a "black box" into an understandable decision-making process, providing biologically sound support for the prediction results. Figure 6 The circular evolution diagram shown in Figure A uses hierarchical clustering achieved by calculating the Pearson correlation coefficients among motifs in 75 datasets. The outermost circle displays the most significantly enriched binding motif sequence (Logo) for each protein, visually demonstrating the conservation of binding sequence preferences across the 75 datasets. To visualize the extracted feature distribution, the RNA base model is fine-tuned, and backpropagation is used to calculate the contribution of input features to the final prediction, thereby obtaining the importance score for each RNA base position. This outputs the feature map gradient significance value, which is then plotted. Figure 6The interpretable spectrum of BC in the graph shows the base importance distribution in AATF_K562 or EXOSC5_HepG2 in the first layer. The saliency heatmap in the second layer is largely consistent with the base enrichment distribution of the icSHAPE values ​​in the first and last layers. Therefore, the feasibility of the MCoMoE model in accurately identifying and analyzing relevant sequence features is further validated.

[0125] In one embodiment, MCoMoE is a structure-enhanced model employing bidirectional sequence structure-coordinated attention to predict protein binding sites in K562 and HepG2 cell lines, such as... Figure 7 , Figure 8 As shown, Figure 7 The network architecture of part D in the middle is as follows: Figure 8 As shown, the RNA-FM model extracts RNA sequences using a 3-mer sliding window and combines in vivo RNA secondary structure information to characterize the sequence and structural properties of RNA. These are then flexibly captured through a multi-scale local window convolution module to extract semantic information and hierarchical relationships of the sequence and structural features. RNAplfold can predict the in vitro secondary structure of RNA based on the RNA sequence. To further capture more feature information, a graph convolutional neural network (GCNA) module is used to enhance the base probability matrix of the secondary structure information predicted by the RNAplfold model. For this purpose, the newly enhanced structural and sequence information are mutually constrained and enhanced through a bidirectional sequence structure synergistic cross-attention module to further strengthen the capture of feature information. Dynamic gating adaptation and motif-aware attention pooling are then used to obtain the final important feature information. Finally, four hybrid expert classifiers are used to achieve the RPI prediction task, and the downstream analysis task of the corresponding binding sites is analyzed. Furthermore, cross-validation experiments using different protein families revealed that there may be a base at a certain position in the DDX3X protein that inhibits the protein's interaction with RNA. Therefore, MCoMoE performed information content logo analysis to further explain the feasibility and effectiveness of the protein cross-testing experiment. It also specifically addresses whether single nucleotide mutations in the DDX3X protein at the STAT3 gene level might enhance or weaken the protein's RNA binding effect.

[0126] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a protein-RNA interaction prediction method based on structure enhancement and collaborative attention. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0127] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0128] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a protein-RNA interaction prediction method for structure enhancement and collaborative attention.

[0129] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement steps of a protein-RNA interaction prediction method for structure enhancement and collaborative attention.

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting protein-RNA interactions using structure enhancement and collaborative attention, characterized in that, The method includes: A protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data is obtained, and the length of the RNA sequences is uniformly processed. The sequence embedding features of the RNA sequences are extracted in a sliding window manner. The in vitro secondary structure base pairing probability matrix of the RNA sequence is obtained to obtain the adjacency matrix of the graph structure. The features of the in vivo structure data after multi-scale convolution processing are used as the initial features of the graph nodes. Message passing and feature aggregation of the graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain structure enhancement features. Convolutional operations are performed on the sequence embedding features and the structural enhancement features using convolutional kernels of different scales to extract contextual dependencies at different spatial distances. An adaptive attention weighting mechanism is introduced to dynamically weight and fuse the outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features. Learnable positional codes are added to the multi-scale sequence semantic features and the multi-scale structural semantic features respectively, and bidirectional cross-attention calculation is performed to obtain the interactively enhanced collaborative representation. The collaborative representation is processed by motif-aware attention pooling to obtain an aggregated feature vector. The aggregated feature vector is then input into a hybrid expert classifier. A gating network assigns weights to each expert network and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions.

2. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 1, characterized in that, Obtain a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and perform length uniformization on the RNA sequences. Extract sequence embedding features of the RNA sequences using a sliding window approach, including: RNA sequence fragments corresponding to the RNA-binding protein region are extracted as positive samples, and negative samples are generated by random sampling or shuffling the order of the RNA sequences. The in vivo structural data corresponding to each RNA sequence are associated to construct a protein-RNA binding sample dataset. Obtain the maximum length threshold, and use the maximum length threshold as a benchmark to perform length uniformization processing on the RNA sequence to obtain the length uniformized RNA sequence; The RNA sequence with uniform length is segmented using a sliding window. The segmented RNA sequence is then input into a pre-trained RNA base model based on the Transformer architecture. A multi-layer self-attention mechanism is used to output feature vectors as sequence embedding features of the RNA sequence.

3. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 1, characterized in that, The in vitro secondary structure base pairing probability matrix of the RNA sequence is obtained to obtain the adjacency matrix of the graph structure. Using the features of the in vivo structural data after multi-scale convolution processing as the initial features of the graph nodes, message passing and feature aggregation of a graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain structural enhancement features, including: The RNA sequence is input into an RNA secondary structure prediction tool, and the in vitro secondary structure of the RNA sequence is predicted based on a thermodynamic model. The probability of pairing between any two nucleotide bases in the RNA sequence is calculated to obtain a base pairing probability matrix, which is then used as the adjacency matrix of the graph structure. The in vivo structural data is input into a multi-scale convolutional neural network. Convolution operations are performed through various multi-scale convolutional kernels to obtain the weights of features at each scale and complete weighted fusion, ultimately obtaining the initial features of each graph node. The initial features are input into a graph convolutional neural network. The initial features are aggregated and nonlinearly transformed along the edges defined by the adjacency matrix. The updated feature vectors of each node are output. All the updated feature vectors are concatenated in sequence according to their position to obtain the structure-enhanced features.

4. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 1, characterized in that, The sequence embedding features and the structural enhancement features are convolved using kernels of different scales to extract contextual dependencies at different spatial distances. An adaptive attention weighting mechanism is then introduced to dynamically weight and fuse the outputs at different scales, resulting in multi-scale sequence semantic features and multi-scale structural semantic features, including: The sequence embedding feature and the structural enhancement feature are processed to have a uniform length, so that the sequence feature and the structural feature correspond in sequence position, resulting in the processed sequence embedding feature and structural enhancement feature. The processed sequence embedding features and structural enhancement features are used as two independent inputs. Each input is subjected to convolution operation in parallel to obtain intermediate features at each scale corresponding to each feature, so as to capture the sequence and structural context dependencies at different spatial distances. Each feature is assigned a corresponding attention weight for each scale intermediate feature, and the outputs of each convolution operation branch are weighted and summed according to their respective attention weights to obtain multi-scale sequence semantic features and multi-scale structural semantic features.

5. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 1, characterized in that, Learnable positional codes are added to the multi-scale sequence semantic features and the multi-scale structural semantic features respectively, and bidirectional cross-attention calculation is performed to obtain the interactively enhanced collaborative representation, including: The first learnable positional encoding is superimposed on the multi-scale sequence semantic features to obtain sequence interaction features with positional information; the second learnable positional encoding is superimposed on the multi-scale structural semantic features to obtain structural interaction features with positional information; the parameters of the first learnable positional encoding and the second learnable positional encoding are independent of each other and are updated through backpropagation during model training. Using the sequence interaction features with location information as queries and the structural interaction features with location information as keys and values, a sequence-to-structure cross-attention calculation is performed to obtain sequence features enhanced by structural information constraints. Using the structural interaction features with location information as queries and the sequence interaction features with location information as keys and values, cross-attention calculation from structure to sequence is performed to obtain structural features enhanced by sequence information constraints. By splicing and fusing sequence features enhanced by structural information constraints and structural features enhanced by sequence information constraints, an interactively enhanced collaborative representation is obtained.

6. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 5, characterized in that, Using the sequence interaction features with location information as queries and the structural interaction features with location information as keys and values, a sequence-to-structure cross-attention calculation is performed to obtain sequence features enhanced by structural information constraints, including: Calculate the attention similarity between each position in the sequence interaction feature with location information and all positions in the structural interaction feature with location information; The attention similarity is scaled and then normalized to attention weights using the Softmax function. The structural interaction features are then weighted and aggregated using these attention weights to obtain sequence features enhanced with structural information.

7. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 1, characterized in that, The collaborative representation is processed by motif-aware attention pooling to obtain an aggregated feature vector, including: The attention scores at each position in the collaborative representation are calculated using a multilayer perceptron, and the attention scores are normalized using a Softmax function to obtain the attention weights corresponding to each position. The collaborative representation is weighted and summed according to the attention weights, and the regional features containing the binding motif are compressed into an aggregated feature vector.

8. The protein-RNA interaction prediction method based on structure enhancement and collaborative attention according to claim 7, characterized in that, The aggregated feature vector is input into a hybrid expert classifier. A gating network assigns weights to each expert network, and the outputs of each expert network are weighted and summed to obtain the prediction results of protein-RNA interactions, including: The aggregated feature vector is input into a gated network consisting of two linear transformation layers to obtain the original weight scores, and the original weight scores are normalized to obtain adaptive weights. Each of the adaptive weights is assigned to a separate independent expert network, and the aggregated feature vector is input into each of the separate independent expert networks. Each group of the separate expert networks outputs a prediction score. The predicted scores are weighted and summed with the adaptive weights respectively to obtain the final prediction result of protein-RNA interaction.

9. A protein-RNA interaction prediction system based on structure enhancement and collaborative attention, characterized in that, The system includes: The data acquisition module is used to acquire a protein-RNA binding sample dataset containing RNA sequences and their corresponding in vivo structural data, and to perform length uniform processing on the RNA sequences and extract the sequence embedding features of the RNA sequences in a sliding window manner. The structure enhancement module is used to obtain the base pairing probability matrix of the in vitro secondary structure of the RNA sequence, obtain the adjacency matrix of the graph structure, and use the features of the in vivo structure data after multi-scale convolution processing as the initial features of the graph nodes. Message passing and feature aggregation of the graph convolutional neural network are performed along the edges defined by the adjacency matrix to obtain the structure enhancement features. The multi-scale feature acquisition module is used to perform convolution operations on the sequence embedding features and the structural enhancement features using convolution kernels of different scales, extract contextual dependencies at different spatial distances, and introduce an adaptive attention weighting mechanism to dynamically weight and fuse the outputs at different scales to obtain multi-scale sequence semantic features and multi-scale structural semantic features. The bidirectional collaborative interaction module is used to add learnable positional codes to the multi-scale sequence semantic features and the multi-scale structural semantic features respectively, perform bidirectional cross-attention calculation, and obtain the collaborative representation after interaction enhancement. The pooling and prediction module is used to obtain an aggregated feature vector by motif-aware attention pooling of the collaborative representation, and input the aggregated feature vector into a hybrid expert classifier. The gating network assigns weights to each expert network and performs a weighted summation of the outputs of each expert network to obtain the prediction result of protein-RNA interaction.

Images