Method and system for predicting binding affinity of protein and RNA (Ribonucleic Acid)

Abstract

The invention discloses a method and a system for predicting binding affinity of protein and RNA (Ribonucleic Acid), and belongs to the technical field of bioinformatics. The method comprises the following steps: firstly, obtaining high-dimensional embedding vectors and three-dimensional structures of original amino acid sequences and nucleotide sequences of protein and RNA; constructing a protein map by using a Graph Transform model, and integrating sequence and structural features to obtain a protein comprehensive feature vector; fusing the sequence and structural features of the RNA through an attention mechanism to generate an RNA nucleotide comprehensive structural feature vector; carrying out interactive learning on the protein and RNA characteristics by adopting a mutual attention model to obtain a global interactive affinity score matrix; determining the attention weight of each part of the protein and the RNA according to the interaction matrix, and constructing a joint feature vector; and finally, predicting the binding affinity of the protein and the RNA through a multi-layer perceptron. According to the method, multi-modal information is deeply fused, the prediction accuracy and generalization ability are remarkably improved, and the method is suitable for drug screening and biological mechanism research.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly to a protein and RNA binding affinity prediction method and system. BACKGROUND

[0002] Protein-RNA binding affinity (PRA) prediction is an important basis for understanding gene regulation mechanisms and developing targeted drugs. Currently, there are mainly two technical routes of experimental methods and computational methods. Experimental methods (such as X-ray crystallography, NMR, SPR, etc.) can provide high-precision structural information and binding parameters, but have inherent defects such as low throughput, long cycle, high cost, etc. Computational methods use machine learning models (such as SVM, RF, CNN, RNN, etc.) to combine sequence or structure features for prediction, which is more efficient, but still faces many technical bottlenecks in practical application. In recent years, although the application of deep learning technology (such as GNN, Transformer, etc.) and pre-trained language models (such as ProtTrans) has made some progress, but the existing methods are still difficult to meet the actual demand in key performance indicators.

[0003] The existing computational methods have the following significant technical defects: First, in the aspect of multi-modal information fusion, the existing methods are insufficient in integrating the sequence and structure features of proteins and RNAs, mostly using simple feature splicing or late decision fusion strategy, which is difficult to effectively capture the complex synergistic mechanism between sequence patterns and spatial conformation. Then, in the aspect of feature representation, traditional methods rely on hand-designed features (such as amino acid composition, contact area, etc.), while deep learning-based methods can automatically extract features, but the capture of key sequence motifs, long-range dependencies, and geometric topological features in three-dimensional structures that determine binding specificity is still insufficient. Secondly, in the aspect of model performance, the existing methods significantly decrease in prediction performance when facing new molecular pairs outside the training set, and lack the ability to distinguish subtle differences in affinity, which is difficult to meet the needs of high-precision application scenarios such as drug screening. In addition, in the aspect of model interpretability, the existing deep learning models generally have the "black box" problem, and the decision-making process lacks transparency, which is not conducive to the discovery of biological mechanisms and limits the optimization direction of the model. Finally, in the aspect of computational efficiency, the existing complex models are difficult to balance prediction accuracy and computational cost, and cannot meet the high-throughput screening needs of the genome scale.

[0004] These technical bottlenecks seriously restrict the application value of PRA prediction in actual scientific research and drug development. Specifically, the superficialization of information fusion leads to the inability of the model to fully understand the molecular recognition mechanism; the incompleteness of feature representation affects the specificity of prediction; the lack of generalization ability limits the application scope of the model; the absence of interpretability hinders the acquisition of biological insights; and the low computational efficiency is difficult to support large-scale application scenarios. Therefore, developing a new type of PRA prediction method that can deeply integrate multi-source information, realize accurate feature extraction, have strong generalization ability and good interpretability has become an urgent need in the field of computational biology and AI drug development. SUMMARY

[0005] In view of the problems in the prior art, the present application provides a method and system for predicting the binding affinity of proteins and RNA, which uses a pre-trained language model to extract sequence features of proteins and RNA, combines a geometric deep learning model to capture three-dimensional structure information, and effectively integrates these multi-modal information through an attention mechanism to achieve high-precision prediction of the binding affinity between proteins and RNA.

[0006] The present application is realized by the following technical solutions: A method for predicting the binding affinity of proteins and RNA, comprising the following steps: Step 1, obtaining high-dimensional embedding vectors and three-dimensional structures of the original amino acid sequence, and obtaining high-dimensional embedding vectors and three-dimensional structures of the original RNA nucleotide sequence; Step 2, establishing a protein graph according to the high-dimensional embedding vectors and three-dimensional structures of the amino acid sequence, aggregating the domain information of the nodes of the protein graph and transforming it, integrating the information of the edges of the protein graph, and making each node integrate the local and global context information in the protein graph to obtain a protein comprehensive feature vector; Step 3, taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and performing weighted fusion on the high-dimensional embedding vector and the three-dimensional structure through an attention mechanism to obtain an RNA nucleotide comprehensive sequence feature vector, and then combining the RNA nucleotide comprehensive sequence feature vector with the structure features of the RNA nucleotide after element-wise dot product to obtain an RNA nucleotide comprehensive structure feature vector; Step 4, performing interactive learning on the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector to obtain a global interactive affinity score matrix; According to the mutual influence of nucleotides and amino acids and combining the global interactive affinity score matrix, the attention weights of each part of the protein and the attention weights of each part of the RNA are determined respectively and ; According to the attention weights Determining a protein context vector , determining a protein context vector according to the attention weight respectively determining RNA context vectors ; Step 5, constructing a joint feature vector according to the protein context vector and the RNA context vector, and predicting the binding affinity of the protein and the RNA according to the joint feature vector.

[0007] Preferably, the protein comprehensive feature vector obtained in step 2 comprises: Taking the amino acid amino as a node, calculating the feature vector of adjacent nodes according to the three-dimensional structure coordinates of the amino acid sequence, splicing the feature vector with the high-dimensional embedding vector to obtain the protein residue node feature, and constructing a protein graph according to all residue node features. Inputting the protein graph into a Graph Transformer model, aggregating the nodes of the protein graph, simultaneously regulating the aggregation process according to the features of the edges of the protein graph, and performing nonlinear transformation on the aggregated information to obtain the protein comprehensive feature vector.

[0008] Preferably, the constructing a protein graph according to all residue node features comprises: According to the three-dimensional structure coordinates of the amino acid sequence, the spatial proximity relationship between the amino acid residues is calculated, the adjacency matrix and the edge list of the protein are constructed, and then the corresponding feature vector of the node is calculated. Splice the feature vector and the high-dimensional embedding vector to obtain the protein residue node feature, embed the amino acid residue node feature into the corresponding node of the protein graph as the initial node feature, and calculate the corresponding edge feature according to the initial node feature to obtain the protein graph.

[0009] Preferably, the method for determining the RNA nucleotide comprehensive structure feature vector in step 3 is as follows: S3.1, determining the structure matrix of the RNA according to the three-dimensional structure of the RNA nucleotide sequence; S3.2, taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and linearly adding the obtained fusion vector to the structure matrix of the RNA to obtain the RNA nucleotide comprehensive sequence feature vector; S3.3, performing element-by-element multiplication on the RNA nucleotide comprehensive sequence feature vector using an attention mechanism, and then adding the RNA nucleotide comprehensive sequence feature vector to the structure matrix of the RNA to obtain the RNA nucleotide comprehensive structure feature vector.

[0010] Preferably, the method for determining the protein context vector according to the attention weight in step 4 comprises: , determining a protein context vector according to the attention weight respectively determining RNA context vectors comprises: The attention weights of different parts of a protein are determined by utilizing the influence of nucleotides on amino acids and combining this with a global interaction affinity score matrix. According to attention weight The protein context vector is obtained by weighted summation of amino acid features. ; The attention weights for different parts of RNA are determined by utilizing the influence of amino acids on nucleotides and combining this with a global interaction affinity score matrix. According to attention weight The RNA context vector is obtained by weighted summation of amino acid features. .

[0011] Preferably, the method for determining the global interaction affinity score matrix is ​​as follows: The RNA nucleotide integrated feature vector and the protein integrated feature vector are input into the mutual attention model for interactive learning. The interaction matrix C between protein feature P and RNA feature R is calculated according to the bilinear interaction method, as follows:

[0012] Where tanh is the activation function. This is a learnable weight matrix.

[0013] Preferably, in step 5, a multilayer perceptron is used to predict the binding affinity between the protein and RNA.

[0014] A predictive system for protein-RNA binding affinity, comprising: The feature extraction module is used to obtain the high-dimensional embedding vector and three-dimensional structure of the original amino acid sequence, as well as the high-dimensional embedding vector and three-dimensional structure of the original RNA nucleotide sequence. The amino acid feature module is used to build a protein graph based on the high-dimensional embedding vector and three-dimensional structure of the amino acid sequence. It aggregates and transforms the neighborhood information of the nodes in the protein graph and integrates the information of the edges of the protein graph, so that each node incorporates the local and global context information in the protein graph to obtain the comprehensive protein feature vector. The nucleotide feature module is used to take the high-dimensional embedding vector of the RNA nucleotide sequence as the query and the three-dimensional structure of the RNA nucleotide sequence as the bond and value. The high-dimensional embedding vector and the three-dimensional structure are weighted and fused through an attention mechanism to obtain the RNA nucleotide comprehensive sequence feature vector. The RNA nucleotide comprehensive sequence feature vector is then combined with the structural features of the RNA nucleotides after element-wise dot product to obtain the RNA nucleotide comprehensive structural feature vector. The fusion module is used to interactively learn the RNA nucleotide integrated feature vector and the protein integrated feature vector to obtain the global interaction affinity score matrix; determine the attention weight of each part of the protein according to the mutual influence of nucleotides and amino acids and combining the global interaction affinity score matrix and the attention weight of each part of the RNA ; determine the attention weight of each part of the protein according to the mutual influence of nucleotides and amino acids and combining the global interaction affinity score matrix determine the protein context vector , determine the RNA context vector according to the attention weight respectively ; an affinity module, for constructing a joint feature vector according to the protein context vector and predicting the binding affinity of the protein and the RNA according to the joint feature vector.

[0015] A drug screening method, comprising: determining the binding affinity of the protein and the RNA by using the prediction method of the binding affinity of the protein and the RNA; screening the drug according to the binding affinity.

[0016] An electronic device, comprising: a memory for storing a computer program; a processor for executing the computer program to realize the steps of the prediction method of the binding affinity of the protein and the RNA.

[0017] Compared with the prior art, the present application has the following beneficial technical effects: The application provides a protein and RNA binding affinity prediction method, which comprehensively uses pre-trained language models, geometric deep learning and attention mechanisms to accurately predict the protein and RNA binding affinity. The method first extracts high-dimensional embedding vectors of protein and RNA sequences through pre-trained models (such as ProtTrans and RiNaLMo), and simultaneously predicts their three-dimensional structures using ESMFold and RhoFold+ models to comprehensively capture sequence and structure information. Subsequently, the protein sequence and structure features are effectively integrated by constructing a protein graph and using a Graph Transformer model to obtain a protein comprehensive feature vector; at the same time, the high-dimensional embedding vectors and three-dimensional structures of the RNA sequence are weighted and fused using an attention mechanism to generate an RNA nucleotide comprehensive structure feature vector. Further, the mutual attention model is used to realize the interaction learning between protein and RNA features, capture the subtle dependency relationship between them, generate a global interaction affinity score matrix, and determine the attention weight of each part to finally construct a joint feature vector for predicting the binding affinity. The method has the advantages of deep fusion of multi-modal information, accurate feature extraction, strong generalization ability and good interpretability, which significantly improves the accuracy and reliability of PRA prediction, and provides strong support for drug screening and biological mechanism research.

[0018] The application also provides a protein and RNA binding affinity prediction system, an electronic device and a computer storage medium, which have all the advantages of the above-mentioned protein and RNA binding affinity prediction method. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0020] Fig. 1 The flowchart of the protein and RNA binding affinity prediction method of the application; Fig. 2 The structural block diagram of the sequence fusion module of the application; Fig. 3 The structural block diagram of the mutual attention module of the application. DETAILED DESCRIPTION

[0021] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0022] Therefore, the detailed description of the embodiments of the present application provided below in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by a person of ordinary skill in the art without creative labor based on the embodiments in the present application are within the scope of protection of the present application.

[0023] A method for predicting protein and RNA binding affinity, comprising the following steps: Step 1, obtaining a high-dimensional embedding vector and a three-dimensional structure of an original amino acid sequence, and obtaining a high-dimensional embedding vector and a three-dimensional structure of an original RNA nucleotide sequence; Step 2, establishing a protein graph according to the high-dimensional embedding vector and the three-dimensional structure of the amino acid sequence, aggregating the field information of the nodes of the protein graph and transforming the same, integrating the information of the edges of the protein graph, and making each node integrate the local and global context information in the protein graph to obtain a protein comprehensive feature vector.

[0024] Step 3, taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and obtaining an RNA nucleotide comprehensive sequence feature vector by weighted fusion of the high-dimensional embedding vector and the three-dimensional structure through an attention mechanism, and combining the RNA nucleotide comprehensive sequence feature vector with the structure features of the RNA nucleotide after element-by-element dot product to obtain an RNA nucleotide comprehensive structure feature vector.

[0025] Step 4, performing interactive learning on the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector to obtain a global interactive affinity score matrix; According to the mutual influence of nucleotides and amino acids and combining the global interactive affinity score matrix, attention weights of each part of the protein and attention weights of each part of the RNA are determined and RNA ; According to the attention weights , a protein context vector is determined , and according to the attention weights , RNA context vectors are determined ; Step 5, according to the protein context vector and construct a joint feature vector, and predict the binding affinity of the protein and RNA according to the joint feature vector.

[0026] The method focuses on the feature combination that contributes most to the binding affinity by introducing a mutual attention mechanism (Mutual-Attention Model), and effectively suppresses irrelevant noise, so as to dynamically and adaptively learn and fuse key information from two modalities of sequence and structure, thereby comprehensively understanding the interaction mode between protein and RNA, and realizing accurate prediction of the binding affinity of protein and RNA.

[0027] Embodiment 1 Reference Figs. 1-3 A method for predicting the binding affinity of protein and RNA, comprising the following steps: Step 1, obtaining high-dimensional embedding vectors and three-dimensional structures of the original amino acid sequence, and obtaining high-dimensional embedding vectors and three-dimensional structures of the original RNA nucleotide sequence; 1) Original amino acid sequence A. Input the original amino acid sequence into the pre-trained ProtTrans model to predict the high-dimensional embedding vector of the original amino acid sequence; The ProtTrans model learns deep context-dependent semantic representations from amino acid sequences, capturing local motifs, long-range dependencies, and evolutionary information in protein sequences.

[0028] The ProtTrans model is based on the Transformer architecture (such as BERT, T5, etc.), and is pre-trained on a large-scale unlabeled protein sequence database (such as UniRef, BFD). The model contains multiple layers of Transformer encoders, each of which iteratively processes and extracts information from the input sequence through self-attention mechanisms and feedforward neural networks.

[0029] The ProtTrans model output (high-dimensional embedding vector of the amino acid sequence) will be one of the important inputs for subsequent geometric deep learning modules to process protein structure and sequence-structure fusion.

[0030] B. Input the original amino acid sequence into the ESMFold model to predict the three-dimensional structure of the original amino acid sequence; The ESMFold model is an end-to-end protein structure prediction model, whose core advantage lies in the sequence understanding capability of ESM (Evolutionary Scale Modeling) language models. Although its main output is a three-dimensional structure, the internal amino acid sequence information is deeply processed and represented, and these intermediate sequence representations or sequence-dependent information in the structure prediction process are crucial for understanding protein function and interaction. The ESMFold model focuses on its deep encoding capability of sequence information and how the predicted structure can in turn guide or complement the understanding of sequence features.

[0031] The ESMFold model combines the sequence embeddings provided by ESM language models (such as ESM-2) and geometric modules similar to AlphaFold2 (such as folding trunk, structure module). Amino acid sequence information is first encoded by ESM models, and then these representations are sent to geometric modules for structure reasoning.

[0032] The protein structure predicted by the ESMFold model will be directly used for subsequent geometric feature extraction. The internal encoding results of the sequence can also be considered as a supplementary source of sequence features.

[0033] 2) Original RNA nucleotide sequence A. Input the original RNA nucleotide sequence into the pre-trained RiNaLMo model to predict the high-dimensional embedding vector of the original RNA nucleotide sequence. The RiNaLMo model is a pre-trained language model for RNA sequences. RiNaLMo is responsible for learning RNA sequence patterns, secondary structure tendencies, and possible binding sites and other related information from the input nucleotide sequence.

[0034] The RiNaLMo model is based on the Transformer architecture, but its pre-training data and tasks are designed for the characteristics of RNA, such as pre-training on a large RNA sequence database (such as RNACentral), which may include auxiliary tasks for secondary structure prediction.

[0035] The output of the RiNaLMo model (high-dimensional embedding of the RNA sequence) is a key input for subsequent geometric deep learning modules to process RNA structure and sequence-structure fusion.

[0036] B. Input the original RNA nucleotide sequence into the pre-trained RhoFold+ model to predict the three-dimensional structure of the original RNA nucleotide sequence.

[0037] RhoFold+ model, as a tool for RNA 3D structure prediction, is similar to ESMFold in that its main goal is structure prediction, but it necessarily relies on deep understanding and characterization of the input RNA sequence. Its internal processing of the sequence, as well as the sequence-structure constraints implied by the predicted structure, are valuable for subsequent affinity prediction.

[0038] RhoFold+ model adopts a deep learning architecture that combines sequence information (possibly processed through a language model-like module) and co-evolution information (if applicable) to predict RNA base pairing and 3D folding.

[0039] The RNA structure predicted by RhoFold+ model will be directly used for subsequent geometric feature extraction. Its internal encoding of the sequence can also serve as a kind of supplementary sequence feature.

[0040] The principles of ProtTrans model and RiNaLMo model are described as follows: ProtTrans model and RiNaLMo model, as pure sequence embedding models, treat amino acids or RNA as a token, and data flows through multiple layers of Transformer encoder in the model.

[0041] In each layer, the correlation (attention weight) between each token and all other tokens is calculated through the self-force mechanism, and the information is weighted and aggregated accordingly, so that the representation of each token can be integrated into its context information; the position encoding module of the model provides the position information of the token in the sequence to the model.

[0042] The feedforward neural network of the model performs nonlinear transformation on the output of the self-attention layer, and finally outputs the high-dimensional embedding vector of the amino acid or RNA sequence.

[0043] The high-dimensional embedding vector of the protein sequence ( ) is output by ProtTrans, is the length of the protein sequence, is the dimension of the embedding vector.

[0044] The high-dimensional embedding vector of the RNA sequence ( ) is output by RiNaLMo, is the length of the RNA sequence, is the dimension of the embedding vector.

[0045] The ESMFold model and the RhoFold+ model are three-dimensional structure prediction models. The input sequence is deeply encoded to capture sequence characteristics and co-evolution signals (if using MSA). The encoded sequence information is sent to a special geometry / structure prediction module. Through iterative optimization and other methods, the three-dimensional coordinates of each amino acid residue / nucleotide are predicted. Finally, the protein three-dimensional structure and RNA three-dimensional structure are output, usually in PDB file format.

[0046] Step 2, according to the high-dimensional embedding vector of the amino acid sequence and the three-dimensional structure, a protein graph is established. The node information of the protein graph is aggregated and transformed. The information of the edges of the protein graph is integrated, so that each node is integrated into the local and global context information in the protein graph, and a protein comprehensive feature vector is obtained.

[0047] In this embodiment, the Graph Transformer network is used to perform iterative information propagation and aggregation on the graph representation of the protein, so as to learn a deep representation that can reflect the sequence characteristics and three-dimensional structure characteristics of the protein and their interaction. According to the high-dimensional embedding vector of the protein and the three-dimensional structure coordinates, a protein graph is constructed, and the Graph Transformer network is used for node (residue) level and graph level information aggregation and feature learning. Finally, a comprehensive feature vector that can effectively represent the protein binding potential is output.

[0048] S2.1, obtaining a high-dimensional embedding vector of an amino acid sequence and a three-dimensional structure of the amino acid sequence; S2.2, calculating the feature vector of the adjacent node according to the three-dimensional structure coordinates, splicing the feature vector and the high-dimensional embedding vector to obtain the protein residue node feature, and constructing the protein graph according to all the residue node features.

[0049] Each node represents an amino acid residue. The spatial proximity relationship between amino acid residues is calculated according to the three-dimensional structure coordinates, the adjacency matrix and the edge list of the protein are constructed, and then the corresponding feature vector (distance, direction, rotation angle, etc.) of the node is calculated. The feature vector and the high-dimensional embedding vector are spliced, the protein residue node feature is embedded into the corresponding node of the protein graph as the initial node feature, the corresponding edge feature (distance, direction, rotation angle, etc.) is calculated according to the initial node feature, and the protein graph is obtained.

[0050] S2.3, inputting the protein graph into the Graph Transformer model, aggregating the nodes of the protein graph, and simultaneously regulating the aggregation process according to the features of the edges of the protein graph, and nonlinearly transforming the aggregated information to obtain a protein comprehensive feature vector.

[0051] The protein graph is constructed using multiple Graph Transformer layers. In each layer, for each node (amino acid residue) in the protein graph, the node "queries" information from its neighboring nodes. The neighboring nodes "respond" to the query based on their own features and the edge features between them and the querying node. Through a graph attention mechanism, the querying node updates its feature representation by weighting and aggregating information from its neighbors and its own existing information according to the importance of different neighbor responses (attention weights). Edge features are used to modulate the attention weights or directly participate in the information aggregation process.

[0052] The aggregated information undergoes a nonlinear transformation via a feedforward neural network to further refine features. Through multi-layer stacking, the representation of each node can gradually perceive information from nodes further away in the graph, thereby capturing long-range interactions and global structural characteristics.

[0053] After the final Graph Transformer layer, each node (residue) possesses an updated feature vector rich in contextual information. These vector sequences... (in It is a residue The final representation after processing by the Graph Transformer can serve as a refined representation of the protein.

[0054] This comprehensive protein feature vector more fully integrates three-dimensional structural information and the interaction relationships between residues, providing a higher quality protein characterization for subsequent fusion with RNA information and final affinity prediction.

[0055] The working process of the Graph Transformer model is explained below: 1) The aim is to calculate the relationships between amino acids, using features such as the distance between amino acid atoms, angle bisectors, and dihedral angles as embeddings for the structure vector:

[0056]

[0057]

[0058]

[0059]

[0060] Finally, the embedding vector of the protein structure was obtained. .

[0061] 2) Construct a protein map.

[0062] Node Definition: Each amino acid residue in the protein is defined as a node in the protein graph.

[0063] The initial feature of each node is the concatenation of the protein sequence and structure embeddings. That is, the high-dimensional embedding vector of the corresponding amino acid residue serves as the initial feature of this node.

[0064]

[0065]

[0066] For the feature of the i th amino acid, it is derived from .

[0067] Edge Definition: The edges between nodes can be defined based on multiple criteria: Spatial Proximity: The distance between two amino acid residues is determined according to the protein three-dimensional structure coordinates. If two amino acid residues are close enough in space (e.g., the distance between their Cα atoms is less than a certain threshold (8 Å)), then there is an edge between them.

[0068] Sequence Proximity: Edges can be defined between sequentially adjacent amino acid residues (e.g., i and i+1) to capture the continuity of the peptide chain.

[0069] The features of the edges include spatial features and relative direction / angle.

[0070] Spatial Distance: The Euclidean distance between the specific atoms of the two connected residues.

[0071] Relative Direction / Angle: The relative spatial orientation between two amino acid residues, calculated using the dihedral angle.

[0072]

[0073] Multi-head Graph Attention Layer: This is the core of the Graph Transformer model, which adopts the self-attention mechanism in the standard Transformer, but is adapted to the graph structure.

[0074] (3) Attention Calculation: For each node i , it calculates the weight coefficients j and attention coefficients between itself and its neighbor nodes . These two coefficients measure the importance of the information of the neighbor nodes j to update the representation of the node i :

[0075]

[0076]

[0077]

[0078] where is a learnable query, key and value matrix.

[0079] Information aggregation: node i is computed as a weighted sum of its neighbor node representations, with the weights being attention coefficients.

[0080]

[0081] Edge i is computed as a weighted sum of its two neighbor nodes:

[0082] where, is a learnable weight function.

[0083] Layer Normalization & Residual Connections: Like the standard Transformer, used to stabilize training and facilitate information flow.

[0084]

[0085]

[0086] where, is a learnable weight function Protein sequence embeddings and 3D structure coordinates are first input into the graph construction unit, which generates a protein graph with initial node features and edge features. This graph is then fed into a stack of multi-head graph attention layers and a position-wise feed-forward network. In each layer, the representation of a node (residue) is updated by aggregating its neighborhood information and transforming it. The information of an edge is integrated in the attention computation. After processing by multiple layers of Graph Transformer, the representation of each node incorporates its local and (through multi-layer propagation) global contextual information in the graph. Finally, the output can be the updated representation vector embedding of all nodes, which integrates the structure and sequence features of the protein with the contextual representation.

[0087] Step 3, the high-dimensional embedding vector of the RNA nucleotide sequence is taken as a query, the three-dimensional structure of the RNA nucleotide sequence is taken as a key and a value, the high-dimensional embedding vector and the three-dimensional structure are weighted and fused through an attention mechanism, the obtained fusion vector is linearly added to the structural features of the RNA nucleotide to obtain an RNA nucleotide comprehensive sequence feature vector, the RNA nucleotide comprehensive sequence feature vector is element-wise dot multiplied and added to the structural features of the RNA nucleotide to obtain an RNA nucleotide comprehensive structure feature vector.

[0088] S3.1, determining the structure matrix of the RNA according to the three-dimensional structure of the RNA nucleotide sequence ; Based on the three-dimensional structure coordinates of the RNA nucleotide sequence, geometric features related to the fusion of the RNA nucleotide sequence are calculated, including the distance, angle and dihedral angle between nucleotides, the adjacency information of the RNA graph based on spatial proximity is constructed according to the geometric features, and the structure matrix of the RNA is obtained .

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] S3.2, the high-dimensional embedding vector of the RNA nucleotide sequence is subjected to preliminary transformation and dimension adjustment through multiple linear layers and is aligned with the structure matrix of the RNA to realize preprocessing of the sequence information.

[0095]

[0096] S3.3, constructing a simplified graph neural network layer according to the structure matrix of the RNA .

[0097] S3.4, the high-dimensional embedding vector of the RNA nucleotide sequence is taken as a query, the three-dimensional structure of the RNA nucleotide sequence is taken as a key and a value, and the weighted fusion is performed through an attention mechanism to obtain an RNA nucleotide comprehensive sequence feature vector.

[0098] The weighted fusion mechanism serves as a core to deeply fuse sequence information and structure information .

[0099] Sequence features as queries, structure features (or node representations learned from the structure graph) as keys and values, are weightedly fused by attention computation, or vice versa.

[0100]

[0101]

[0102] where, is a learnable weight vector, is the adjacency matrix of the graph.

[0103] S3.5, element-wise product of the RNA nucleotide integrated sequence feature vector with the attention mechanism, forming an iterative or multi-step fusion process, to obtain the RNA nucleotide integrated structure feature vector.

[0104] Sequence information affects the representation of structure information, and structure information also affects the representation of sequence information. Using attention mechanism and element-wise product, an iterative or multi-step fusion process is formed:

[0105]

[0106]

[0107]

[0108] After n fusion, the final RNA nucleotide integrated structure feature vector is output

[0109] In this embodiment, the RNA sequence-structure fusion module is used to effectively integrate the sequence semantic information of RNA with its three-dimensional conformational features, to obtain the RNA nucleotide integrated structure feature vector.

[0110] The high-dimensional embedding vector of the RNA nucleotide sequence and the three-dimensional structure are input into the RNA sequence-structure fusion module.

[0111] The three-dimensional structure coordinates are input into the RNA structure feature extraction unit to calculate the key geometric features. The sequence embedding and the extracted structure features are processed through their respective processing paths for preliminary transformation and representation. The processed sequence and structure information are sent to the core fusion mechanism, and the information of the two modalities is deeply integrated and influenced each other, forming a unified representation. The fused representation finally generates the final RNA nucleotide integrated structure feature vector through the output layer.

[0112] Structure information extraction and representation: According to the three-dimensional coordinates of RNA, parameters describing its local and global geometric features are calculated, such as the distance between nucleotides, angles, or a graph representation based on proximity.

[0113] Sequence information preprocessing: necessary dimension adjustment or feature transformation is performed on the input sequence embedding to make it suitable for fusion with structure information.

[0114] Sequence-structure information fusion: the preprocessed sequence features and structure features are integrated through the core fusion mechanism designed in the module. Structure information guides the interpretation of sequence features, revealing important associations between nucleotides that are close in space but distant in sequence. The intrinsic patterns of sequences affect the weight or contribution of structure features, for example, conserved sequence motifs correspond to key domains. The fusion process is one-way (structure information modulates sequence information) or bidirectional iteration to fully exploit the synergistic effect between the two modalities.

[0115] Output representation generation: the fusion module outputs the updated feature vector of each nucleotide, which now encodes both the sequence context information of the nucleotide and its local environmental information in the three-dimensional structure. The final output is the RNA nucleotide comprehensive structure feature vector

[0116] Updated node (nucleotide) level feature sequence: a tensor of shape ( ), where is the length of the RNA sequence, is the nucleotide embedding dimension output by the fusion module. Each vector in this sequence integrates the sequence semantics of the corresponding nucleotide and its three-dimensional structure environmental information.

[0117] The RNA nucleotide comprehensive structure feature vector more closely combines the predicted three-dimensional structure information than the initial sequence embedding, providing more comprehensive and accurate RNA representation for subsequent interaction with protein information (through the mutual attention module) and final affinity prediction.

[0118] Step 4, interactive learning of RNA nucleotide comprehensive feature vectors and protein comprehensive feature vectors to obtain a global interactive affinity score matrix; Determine the attention weight of each part of the protein using the influence of nucleotides on amino acids and combining the global interactive affinity score matrix , and according to the attention weight , the amino acid features are weighted and summed to obtain the protein context vector .

[0119] Determine the attention weight of each part of the RNA using the influence of amino acids on nucleotides and combining the global interactive affinity score matrix , and according to the attention weight , the amino acid features are weighted and summed to obtain the RNA context vector .

[0120] In this step, the subtle dependence and mutual influence between protein and RNA features are captured by the cross-attention model, generating a highly integrated cross-modal representation that will be directly used for subsequent affinity prediction, as follows: S4.1, input the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector into the cross-attention model for interactive learning.

[0121] S4.2, calculate the interaction matrix C between protein features P and RNA features R according to the bilinear interaction method:

[0122] Where tanh is the activation function. This interaction matrix C encodes the original association strength between each part of the protein and each part of the RNA, is a learnable weight matrix.

[0123] S4.3, determine the attention weight of each part of the protein by using the influence of nucleotides on amino acids and combining the global interaction affinity score matrix , and according to the attention weight , the amino acid features are weighted and summed to obtain the protein context vector .

[0124] The influence of RNA features on protein features, i.e. the attention score of each part of the protein ( ) is calculated as follows:

[0125]

[0126] Where, are all learnable weight matrices.

[0127] The protein context vector is calculated as follows:

[0128] S4.3, determine the attention weight of each part of the RNA by using the influence of amino acids on nucleotides and combining the global interaction affinity score matrix , and according to the attention weight , the amino acid features are weighted and summed to obtain the RNA context vector .

[0129] The influence of amino acid features on nucleotide features, i.e. the attention score of each part of the RNA ( ) is calculated as follows:

[0130]

[0131] wherein, are all learnable weight matrices.

[0132] RNA context vector as follows:

[0133] S4.4, concatenating the RNA context vector and the protein context vector to form a joint feature vector.

[0134] In this step, the mutual dependency between protein and RNA is constructed through attention mechanism. It not only considers the internal features of protein and RNA respectively, but more importantly, it calculates how the interaction between the two affects the respective representations, thus generating a joint feature vector that is more sensitive to the interaction between the two.

[0135] The joint feature vector of protein-RNA interaction is a fixed-dimensional vector that encodes the complex dependency and interaction information between protein and RNA learned through the bidirectional collaborative attention mechanism. It integrates the characteristics of protein and RNA and their mutual influence in the interaction process. This joint feature vector will be used as the input of the subsequent multi-layer perceptron (MLP) to predict the final protein-RNA binding affinity.

[0136] Step 5, predicting the binding affinity of protein and RNA according to the joint feature vector.

[0137] The final decision layer of the GPRM-PRA model is processed by one or more multi-layer perceptron (MLP) networks, and finally outputs a quantitative prediction value of the binding affinity between protein and RNA.

[0138] Using the nonlinear mapping capability of MLP, a complex function that can accurately predict the binding affinity value is learned from the input protein-RNA joint feature vector.

[0139] The input joint feature vector is first passed through the input layer to the first hidden layer. In each hidden layer, data flows from the previous layer to the current layer, and each neuron calculates its weighted input sum and passes it through an activation function. Information is propagated forward layer by layer, and each layer performs higher-level abstraction and combination of input features. Finally, data reaches the output layer, and the output layer neurons calculate the final prediction value.

[0140]

[0141] Loss Function Computation: During the model training phase, the predicted affinity values y* output by the MLP are compared with the true experimentally determined affinity values y. The Mean Squared Error (MSE) is used as the loss function to quantify the difference between the predicted and true values:

[0142] where n is the number of training samples.

[0143] Parameter Optimization: The gradients of the loss function with respect to all weights and biases in the MLP network are computed using the backpropagation algorithm. An optimizer (such as Adam, SGD, etc.) is used to update the network parameters based on the gradients to minimize the loss function.

[0144] This training process is performed end-to-end across the entire GPRM-PRA model (including the preceding modules), or the MLP module can be trained / fine-tuned separately based on fixed upstream module outputs.

[0145] The protein and RNA binding affinity prediction method innovatively integrates pre-trained language models, geometric deep learning, mutual attention mechanisms, and multi-layer perceptrons, and exhibits significant advantages and beneficial effects compared to existing technologies in protein-RNA binding affinity (PRA) prediction, mainly in the following aspects: 1. Significantly improving the accuracy and reliability of PRA prediction Precision improvement brought by deep feature fusion: The GPRM-PRA model can simultaneously extract deep and multi-modal features from the sequences and three-dimensional structures of proteins and RNAs. ProtTrans model, ESMFold model, RiNaLMo model, and RhoFold+ model provide high-quality initial sequence embeddings and structural information. The Graph Transformer model and the Structure-Sequence Fusion Module model perform deep fusion and representation of protein and RNA sequence-structure information, respectively. The Mutual-Attention Model further captures the cross-modal interaction between proteins and RNAs. This multi-level and multi-angle feature fusion and interaction modeling enables the model to more comprehensively understand the complex molecular mechanisms that determine binding affinity, thereby surpassing methods that rely on a single information source or simple feature concatenation in terms of prediction accuracy.

[0146] 2. By evaluating on the standard PRA benchmark dataset, GPRM-PRA model significantly outperforms current mainstream PRA prediction tools (such as CoPRA, IPMiner, GraphBind, etc.) on key evaluation metrics (such as Pearson Correlation Coefficient PCC, Root Mean Square Error RMSE, Area Under Receiver Operating Characteristic Curve AUC).

[0147] 3. Enhanced discrimination of subtle affinity differences: With more refined feature representation and interaction modeling, GPRM-PRA model is expected to better distinguish protein-RNA pairs with similar but subtle differences in affinity, which is of great significance for optimizing lead compounds in drug screening.

[0148] 4. Prediction ability for novel molecular pairs: The introduction of pre-trained language models and geometric deep learning enables the model to learn more general molecular representations and interaction patterns, rather than just fitting specific patterns in the training set. Therefore, GPRM-PRA model is expected to maintain high prediction accuracy when facing protein-RNA pairs that have not appeared in the training set, with novel structures or sequences, showing stronger generalization ability.

[0149] 5. Cross-species applicability: Since the model learns universal molecular features and interaction principles, it is expected to perform well in predicting protein-RNA interactions across different species, reducing over-reliance on data from specific species.

[0150] 6. Accelerating candidate drug / target screening: High-precision computational prediction can massively and quickly screen potential RNA-targeting drugs or protein / RNA-based biological therapy candidates, significantly reducing the scope of experimental verification, thereby significantly shortening the drug development cycle.

[0151] 7. Guiding experimental design and reducing trial and error: The prediction results of the model can provide valuable clues for experimental scientists, such as predicting key binding residues / nucleotides or predicting the effect of mutations on affinity, thereby guiding subsequent experimental design, reducing blind trial and error, and saving expensive experimental reagents and time costs.

[0152] 8. Promoting understanding of complex biological processes: By analyzing model predictions and (possibly through explainability analysis in the future) model decision-making basis, it helps to deeply understand protein-RNA interaction networks related to specific diseases, or reveals the molecular mechanisms of complex life processes such as gene expression regulation.

[0153] 9. Predicting the impact of disease-related mutations: The GPRM-PRA model can be used to assess how disease-related mutations (causing changes in protein or RNA sequence / structure) affect their binding affinity with other molecules, helping to understand the molecular mechanisms of disease occurrence and providing the basis for the development of personalized treatment plans.

[0154] 10. Assistance in biomarker discovery: By analyzing large-scale protein-RNA interaction data combined with model predictions, it helps to discover biomarkers related to specific disease states or drug responses.

[0155] 11. Insights into attention mechanisms: Attention weights can be visualized and analyzed to reveal which regions of the protein and RNA contribute most to the binding process, providing clues to understanding the key features of the binding interface and the mode of action, enhancing the interpretability of the model.

[0156] 12. Guidance for model optimization and new hypothesis generation: Understanding the internal features and attention patterns of the model not only helps to further optimize the model structure, but also inspires new biological hypotheses about protein-RNA recognition mechanisms, and in turn guides experimental verification.

[0157] In summary, the GPRM-PRA model is expected to become a powerful and practical computational tool in the field of protein-RNA binding affinity research, bringing substantial impetus to life science research and drug development.

[0158] Example 2 Correspondingly, the present application also provides a protein and RNA binding affinity prediction system, comprising: a feature extraction module for obtaining high-dimensional embedding vectors and three-dimensional structures of original amino acid sequences, and obtaining high-dimensional embedding vectors and three-dimensional structures of original RNA nucleotide sequences; an amino acid feature module for establishing a protein graph according to the high-dimensional embedding vectors and three-dimensional structures of the amino acid sequences, aggregating and transforming the field information of the nodes of the protein graph, integrating the information of the edges of the protein graph, and making each node integrate the local and global context information in the protein graph to obtain a protein comprehensive feature vector.

[0159] a nucleotide feature module for taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and obtaining an RNA nucleotide comprehensive sequence feature vector by weighted fusion of the high-dimensional embedding vector and the three-dimensional structure through an attention mechanism, and obtaining an RNA nucleotide comprehensive structure feature vector by element-wise dot product of the RNA nucleotide comprehensive sequence feature vector and the structure features of the RNA nucleotide.

[0160] The fusion module is used to interactively learn the RNA nucleotide integrated feature vector and the protein integrated feature vector to obtain the global interaction affinity score matrix; Based on the interaction influence between nucleotides and amino acids and combined with the global interaction affinity score matrix, attention weights for each part of the protein are determined. Attention weights for different parts of RNA ; Based on attention weight Determine the protein context vector According to attention weight Determine the RNA context vector respectively ; Affinity module, used to determine protein context vectors A joint feature vector is constructed, and the binding affinity between proteins and RNA is predicted based on the joint feature vector.

[0161] Example 3 This application also provides a method for screening drugs, comprising the following steps: Based on the protein-RNA binding affinity prediction method in Example 1, the affinity between proteins and RNA was determined, and drugs were screened based on the affinity.

[0162] This drug is an RNA-targeting drug, or a protein / RNA-based biotherapy candidate.

[0163] Example 4 This application also provides a drug screening system, comprising: The affinity module is used to determine the affinity between proteins and RNA based on methods for predicting protein-RNA binding affinity. The screening module is used to screen drugs based on affinity.

[0164] This drug is an RNA-targeting drug, or a protein / RNA-based biotherapy candidate.

[0165] It should be noted that in the several embodiments provided in the present application, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, another division manner can be used. For example, a plurality of modules can be combined or integrated into another device, or some features can be omitted or not executed. The modules described as separate components can be or can not be physically separated. The components displayed as modules can be one physical unit or a plurality of physical units, that is, can be located in one place or can be distributed in a plurality of different places. According to actual needs, part or all of the modules can be selected to achieve the purpose of the embodiment.

[0166] In addition, each module in each embodiment of the present application can be integrated in one processing unit, or each module can be physically present alone, or two or more modules can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0167] The electronic device provided in the embodiment of the present application includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the protein and RNA binding affinity prediction method described in any of the above embodiments are realized.

[0168] The electronic device provided in another embodiment of the present application can further include: an input port connected to the processor, used to transmit the multi-modal data collected by an external collection device to the processor; a display unit connected to the processor, used to display the processing result of the processor to the outside world; and a communication module connected to the processor, used to realize the communication between the electronic device and the outside world. The display unit can be a display panel, a laser scanning display, etc. The communication mode used by the communication module includes but is not limited to mobile high-definition link technology (HML), universal serial bus (USB), high-definition multimedia interface (HDMI), wireless connection (including wireless fidelity technology (WiFi), Bluetooth communication technology, low-power Bluetooth communication technology, and IEEE 802.11s-based communication technology).

[0169] The computer readable storage medium provided in the embodiment of the present application stores a computer program. When the computer program is executed by the processor, the steps of the protein and RNA binding affinity prediction method described in any of the above embodiments are realized.

[0170] The related parts in the protein and RNA binding affinity prediction system, the electronic device and the computer readable storage medium provided in the embodiments of the present application are described in detail in the corresponding part of the protein and RNA binding affinity prediction method provided in the embodiments of the present application, and will not be described here. In addition, the part of the above technical solution provided in the embodiments of the present application which is consistent with the implementation principle of the corresponding technical solution in the prior art is not described in detail, so as not to be too redundant.

[0171] The above is only to illustrate the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical solution falls within the protection scope of the claims of the present application.

Claims

1. A method for predicting protein and RNA binding affinity, characterized by, The method comprises the following steps: Step 1, obtaining a high-dimensional embedding vector and a three-dimensional structure of an original amino acid sequence, and obtaining a high-dimensional embedding vector and a three-dimensional structure of an original RNA nucleotide sequence; Step 2, establishing a protein graph according to the high-dimensional embedding vector and the three-dimensional structure of the amino acid sequence, aggregating field information of nodes of the protein graph and transforming the nodes, integrating information of edges of the protein graph, and making each node integrate local and global context information in the protein graph to obtain a protein comprehensive feature vector; Step 3, taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and obtaining an RNA nucleotide comprehensive sequence feature vector by weighting and fusing the high-dimensional embedding vector and the three-dimensional structure through an attention mechanism, and combining the RNA nucleotide comprehensive sequence feature vector with the structure feature of the RNA nucleotide after element-by-element dot product to obtain an RNA nucleotide comprehensive structure feature vector; Step 4, performing interactive learning on the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector to obtain a global interactive affinity score matrix; According to the mutual influence of nucleotides and amino acids and combined with the global interaction affinity score matrix, the attention weights of each part of the protein and each part of the RNA are determined respectively and RNA ​ According to the attention weight Determining a protein context vector , according to the attention weight Determining RNA context vectors, respectively ; Step 5, predicting the binding affinity of the protein and RNA according to the joint feature vector. and constructing a joint feature vector, predicting the binding affinity of the protein and RNA according to the joint feature vector.

2. The method of predicting protein and RNA binding affinity according to claim 1, wherein, The protein comprehensive feature vector obtained in step 2 comprises: taking an amino acid as a node, calculating a feature vector of adjacent nodes according to three-dimensional structure coordinates of the amino acid sequence, splicing the feature vector and the high-dimensional embedding vector to obtain a protein residue node feature, and constructing a protein graph according to all residue node features; inputting the protein graph into a Graph Transformer model, aggregating nodes of the protein graph, simultaneously regulating the aggregation process according to features of edges of the protein graph, and nonlinearly changing aggregated information to obtain a protein comprehensive feature vector.

3. The method for predicting protein-RNA binding affinity according to claim 2, characterized in that, The constructing of the protein graph according to all residue node features comprises: calculating a spatial proximity relationship between amino acid residues according to three-dimensional structure coordinates of the amino acid sequence, constructing an adjacency matrix and an edge list of the protein, then calculating a corresponding feature vector of a node, splicing the feature vector and the high-dimensional embedding vector to obtain a protein residue node feature, embedding the amino acid residue node feature into a corresponding node of the protein graph as an initial node feature, and calculating corresponding edge features according to the initial node feature to obtain the protein graph.

4. The method for predicting protein-RNA binding affinity according to claim 1, characterized in that, The determination method of the RNA nucleotide comprehensive structure feature vector in step 3 is as follows: S3.1, determining a structure matrix of the RNA according to the three-dimensional structure of the RNA nucleotide sequence; S3.2, taking the high-dimensional embedding vector of the RNA nucleotide sequence as a query, taking the three-dimensional structure of the RNA nucleotide sequence as a key and a value, linearly adding the obtained fusion vector and the structure matrix of the RNA to obtain an RNA nucleotide comprehensive sequence feature vector; S3.3, performing element-by-element multiplication on the RNA nucleotide comprehensive sequence feature vector through an attention mechanism, and adding the structure matrix of the RNA to obtain an RNA nucleotide comprehensive structure feature vector.

5. The method for predicting protein-RNA binding affinity according to claim 1, characterized in that, determining the attention weights determining a protein context vector determining the attention weights determining RNA context vectors comprising: The attention weights of different parts of a protein are determined by utilizing the influence of nucleotides on amino acids and combining this with a global interaction affinity score matrix. According to attention weight The protein context vector is obtained by weighted summation of amino acid features. ; The attention weights for each part of the RNA are determined using the effect of the amino acids on the nucleotides in combination with a global cross-interaction affinity score matrix , according to the attention weights The amino acid features are weighted summed to obtain the RNA context vector .

6. The method for predicting protein-RNA binding affinity according to claim 5, characterized in that, The determination method of the global interactive affinity score matrix is as follows: inputting the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector into a mutual attention model for interactive learning, and calculating an interaction matrix C between protein features P and RNA features R according to a bilinear interaction method, as follows: where tanh is the activation function, are learnable weight matrices.

7. The method for predicting protein-RNA binding affinity according to claim 1, characterized in that, In step 5, the multi-layer perception is used to predict the binding affinity between the protein and the RNA.

8. A system for predicting protein and RNA binding affinity, characterized by, The method comprises the following steps: The feature extraction module is used to obtain the high-dimensional embedding vector and the three-dimensional structure of the original amino acid sequence, and obtain the high-dimensional embedding vector and the three-dimensional structure of the original RNA nucleotide sequence; The amino acid feature module is used to establish a protein graph according to the high-dimensional embedding vector and the three-dimensional structure of the amino acid sequence, aggregate the field information of the nodes of the protein graph and transform the nodes, integrate the information of the edges of the protein graph, and make each node integrate the local and global context information in the protein graph to obtain a protein comprehensive feature vector; The nucleotide feature module is used to take the high-dimensional embedding vector of the RNA nucleotide sequence as a query, take the three-dimensional structure of the RNA nucleotide sequence as a key and a value, and obtain an RNA nucleotide comprehensive sequence feature vector by weighting and fusing the high-dimensional embedding vector and the three-dimensional structure through an attention mechanism, and obtain an RNA nucleotide comprehensive structure feature vector by element-wise dot product of the RNA nucleotide comprehensive sequence feature vector and combination of the structure features of the RNA nucleotide; The fusion module is used to perform interactive learning on the RNA nucleotide comprehensive feature vector and the protein comprehensive feature vector to obtain a global interactive affinity score matrix. According to the mutual influence of nucleotides and amino acids and combined with the global interaction affinity score matrix, the attention weights of each part of the protein and each part of the RNA are determined respectively and RNA ​ According to the attention weight Determining a protein context vector , according to the attention weight Determining RNA context vectors, respectively ; Affinity module, used to determine protein context vectors A joint feature vector is constructed, and the binding affinity between proteins and RNA is predicted based on the joint feature vector.

9. A method of screening for a drug, characterized by, The method comprises the following steps: The method for predicting the binding affinity between the protein and the RNA according to any one of claims 1-7 is used to determine the binding affinity between the protein and the RNA; The drug is screened according to the binding affinity.

10. An electronic device, comprising: The method comprises the following steps: The memory is used to store a computer program; The processor is used to execute the computer program to realize the steps of the method for predicting the binding affinity between the protein and the RNA according to any one of claims 1-7.

Images