Protein-nucleic acid interaction prediction method and device based on cross-attention learning, storage medium and equipment

Abstract

The application discloses a protein nucleic acid interaction prediction method and device based on cross attention learning, a storage medium and equipment, and mainly aims to solve the problem of poor prediction accuracy of the existing protein nucleic acid interaction. The method comprises the following steps: obtaining a nucleic acid sequence and a protein sequence to be predicted; encoding the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid encoding features and protein encoding features; and performing cross prediction on the nucleic acid encoding features and the protein encoding features based on an interaction prediction model that has completed model training to obtain an interaction prediction result; and performing fusion based on the interaction prediction result to obtain an interaction prediction result.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method, apparatus, storage medium, and device for predicting protein-nucleic acid interactions based on cross-attention learning. Background Technology

[0002] Protein-nucleic acid interactions (PNIs) are a core molecular mechanism of life activities, permeating key biological processes such as gene transcription regulation, DNA (deoxyribonucleic acid) replication and repair, RNA (ribonucleic acid) splicing and processing, and chromatin assembly. The specific binding of transcription factors to DNA regulatory regions directly determines the spatiotemporal specificity of gene expression, while the interaction between RNA-binding proteins and mRNA precisely regulates mRNA stability and translation efficiency. Abnormal regulation of PNIs is closely related to various pathological states, including tumorigenesis, neurodegenerative diseases, and hereditary diseases. Therefore, accurately identifying PNI pairings and elucidating their interaction patterns is not only a core proposition in basic molecular biology research but also a crucial prerequisite for the discovery of disease diagnostic biomarkers, drug target screening, and new drug development.

[0003] Currently, deep learning has made groundbreaking progress in the field of bioinformatics due to its powerful capabilities in automatic feature extraction, nonlinear fitting, and large-scale data processing, providing a new paradigm for PNI prediction. Among them, IPMiner (Interaction Pattern Miner) extracts high-level abstract features from RNA and protein K-mer sequence vectors by stacking autoencoders, laying the foundation for RNA-protein interaction (RPI) prediction; diPaRIS (Dynamic and Interpretable Protein-RNA Interactions Prediction) integrates RNA dynamic structure coding and U-shaped networks, significantly improving the accuracy and interpretability of dynamic protein-RNA interaction prediction.

[0004] However, most existing models rely on prior knowledge from PNI networks and focus heavily on RNA-protein interactions, with very little research on predicting protein-DNA interactions. Furthermore, these methods are often single-function, lacking the ability to predict binding sites. For example, in binding site prediction, they lack accurate predictions of specific protein-nucleic acid pairings; in the field of PNI affinity prediction, traditional models heavily rely on the structural information of proteins and nucleic acids, making accurate predictions impossible when structural information is missing. Therefore, a protein-nucleic acid interaction prediction method based on cross-attention learning is urgently needed to address these technical problems. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, storage medium and device for predicting protein-nucleic acid interactions based on cross-attention learning, the main purpose of which is to solve the problem of poor accuracy in existing protein-nucleic acid interaction prediction.

[0006] According to one aspect of this application, a method for predicting protein-nucleic acid interactions based on cross-attention learning is provided, comprising: Obtain the nucleic acid and protein sequences to be predicted; The nucleic acid sequence and the protein sequence are encoded respectively to obtain nucleic acid coding features and protein coding features. Based on the action prediction model that has been trained, the nucleic acid coding features and the protein coding features are interactively predicted to obtain interactive prediction results. The action prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The interaction prediction results are fused to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships, and residue identification results.

[0007] Furthermore, before obtaining the interactive prediction result by performing interactive prediction on the nucleic acid-coding features and the protein-coding features based on the completed model training, the method further includes: Obtain a feature training sample set, which includes fused and aligned protein samples and nucleic acid samples; Construct a deep learning network, which includes a normalization layer, a connection layer, an interactive attention layer, an input layer, and an output layer; The deep learning network is trained based on the feature training sample set, and a predictive model with completed model training is obtained after meeting the preset prediction loss conditions.

[0008] Furthermore, the preset prediction loss condition is constructed based on affinity prediction loss, residue prediction loss, interaction prediction loss, and their respective weights.

[0009] Furthermore, before obtaining the feature training sample set, the method further includes: Obtain protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples; The protein structure sample and the nucleic acid structure sample are encoded to obtain protein structure features and nucleic acid structure features, and the protein sequence sample and the nucleic acid sequence sample are encoded to obtain protein sequence features and nucleic acid sequence features; Align the protein structural features with the protein sequence features to obtain a protein sample, and align the nucleic acid structural features with the nucleic acid sequence features to obtain a nucleic acid sample.

[0010] Further, the encoding of the protein structure sample and the nucleic acid structure sample to obtain protein structure features and nucleic acid structure features includes: The protein structure sample and the nucleic acid structure sample are encoded by a structural encoder respectively to obtain protein isomorphism graph representation and nucleic acid isomorphism graph representation. Both the protein isomorphism graph representation and the nucleic acid isomorphism graph representation include node features and edge features. The protein isomer map representation and the nucleic acid isomer map representation are reconstructed by a decoder, and the reconstructed protein isomer map representation and the nucleic acid isomer map representation are reconstructed by masking residues based on a scaling reconstruction loss function to obtain the protein structural features and the nucleic acid structural features.

[0011] Furthermore, the interaction prediction results include protein prediction results and nucleic acid prediction results, and the fusion based on the interaction prediction results to obtain the interaction prediction results includes: The protein prediction results and nucleic acid prediction results are fused using an attention network to obtain the relevant interaction relationship and the affinity relationship; The protein prediction results are processed based on the convolutional neural network model to obtain the residue identification results.

[0012] Furthermore, the step of fusing the protein prediction results and nucleic acid prediction results through an attention network to obtain the relevant interaction relationship and the affinity relationship includes: The protein prediction results and the nucleic acid prediction results are reduced in dimensionality using a linear relationship, and attention weights are calculated for the reduced protein prediction results and the nucleic acid prediction results based on attention scores. The calculated attention weights are weighted to obtain a global vector for the protein and a global vector for the nucleic acid. The global vectors for the protein and the nucleic acid are then concatenated to generate the relevant interaction relationship and the affinity relationship.

[0013] According to another aspect of this application, a protein-nucleic acid interaction prediction device based on cross-attention learning is provided, comprising: The acquisition module is used to acquire the nucleic acid and protein sequences to be predicted. The prediction module is used to encode the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features, and to perform interactive prediction on the nucleic acid coding features and the protein coding features based on the completed model training of the effect prediction model to obtain interactive prediction results. The effect prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The fusion module is used to fuse the interaction prediction results to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships, and residue identification results.

[0014] Furthermore, the device also includes: a construction module and a training module. The acquisition module is also used to acquire a feature training sample set, which includes fused and aligned protein samples and nucleic acid samples. The building module is used to build a deep learning network, which includes a normalization layer, a connection layer, an interactive attention layer, an input layer, and an output layer. The training module is used to train the deep learning network based on the feature training sample set, and obtain the effective prediction model after the preset prediction loss condition is met.

[0015] Furthermore, the preset prediction loss condition is constructed based on affinity prediction loss, residue prediction loss, interaction prediction loss, and their respective weights.

[0016] Furthermore, the acquisition module is also used to acquire protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples; encode the protein structure samples and nucleic acid structure samples to obtain protein structure features and nucleic acid structure features, and encode the protein sequence samples and nucleic acid sequence samples to obtain protein sequence features and nucleic acid sequence features; align the protein structure features and protein sequence features to obtain protein samples, and align the nucleic acid structure features and nucleic acid sequence features to obtain nucleic acid samples.

[0017] Furthermore, the acquisition module is specifically used to encode the protein structure sample and the nucleic acid structure sample respectively through a structure encoder to obtain a protein isomorphism graph representation and a nucleic acid isomorphism graph representation, both of which include node features and edge features; to reconstruct the protein isomorphism graph representation and the nucleic acid isomorphism graph representation through a decoder, and to perform masked residue reconstruction on the reconstructed protein isomorphism graph representation and the nucleic acid isomorphism graph representation based on a scaling reconstruction loss function to obtain the protein structure features and the nucleic acid structure features.

[0018] Furthermore, the fusion module is specifically used to fuse the protein prediction results and nucleic acid prediction results through an attention network to obtain the correlation relationship and the affinity relationship; and to process the protein prediction results based on a convolutional neural model to obtain the residue recognition results.

[0019] Furthermore, the fusion module is specifically used to reduce the dimensionality of the protein prediction results and the nucleic acid prediction results through linear relationships, and to calculate attention weights for the dimensionality-reduced protein prediction results and the nucleic acid prediction results based on attention scores; to perform weighting based on the calculated attention weights to obtain a global vector of the protein and a global vector of the nucleic acid, and to concatenate the global vector of the protein and the global vector of the nucleic acid to generate the relevant interaction relationship and the affinity relationship.

[0020] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the protein-nucleic acid interaction prediction method based on cross-attention learning described above.

[0021] According to another aspect of this application, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the protein-nucleic acid interaction prediction method based on cross-attention learning described above.

[0022] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a method, apparatus, storage medium, and device for predicting protein-nucleic acid interactions based on cross-attention learning. Compared with the prior art, the embodiments of this application obtain the nucleic acid sequence and protein sequence to be predicted; encode the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features; and perform interactive prediction on the nucleic acid coding features and the protein coding features based on a pre-trained interaction prediction model to obtain interactive prediction results. The interaction prediction model is constructed by introducing a deep learning network based on the cross-attention mechanism. The interaction layer in the deep learning network includes cross-attention sublayers in two directions. The interactive prediction results are fused to obtain interaction prediction results. The interaction prediction results include related interaction relationships, affinity relationships, and residue identification results, achieving the common prediction purpose of three dimensions, improving the adaptability of protein-nucleic acid prediction based on multiple molecular mechanisms, meeting the interaction prediction needs when structures are missing, and thus improving the effectiveness of protein-nucleic acid interaction prediction.

[0023] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The flowchart of a protein-nucleic acid interaction prediction method based on cross-attention learning provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of a protein-nucleic acid interaction prediction model provided in an embodiment of this application. Figure 3 This illustration shows a schematic diagram of a comprehensive evaluation result in an interaction and site prediction task provided by an embodiment of this application; Figure 4 This paper shows a block diagram of a protein-nucleic acid interaction prediction device based on cross-attention learning, provided in an embodiment of this application. Figure 5 A schematic diagram of the structure of a terminal provided in an embodiment of this application is shown. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0028] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0029] Based on this, in one embodiment, the present invention provides a protein-nucleic acid interaction prediction method based on cross-attention learning. Taking the application of this method to computer devices such as servers as an example, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0030] This application provides a method for predicting protein-nucleic acid interactions based on cross-attention learning, such as... Figure 1 As shown, the method includes: 101. Obtain the nucleic acid sequence and protein sequence to be predicted.

[0031] In this embodiment, the current executing entity, acting as the execution end for predicting protein-nucleic acid interactions, can be a terminal device or a server, to obtain the nucleic acid sequence and protein sequence to be predicted. Here, the nucleic acid sequence refers to the code sequence that characterizes a gene using text, and the protein sequence refers to the sequence of amino acids that characterizes the order of amino acids using text. The nucleic acid sequence and protein sequence to be predicted can be selected and entered from a protein-nucleic acid sequence database.

[0032] It should be noted that the proteins in the embodiments of this application can be obtained from the globally publicly available three-dimensional atomic structure database of biological macromolecules such as proteins and nucleic acids, and the embodiments of this application do not impose specific limitations.

[0033] 102. Encode the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features, and perform interactive prediction on the nucleic acid coding features and the protein coding features based on the completed model training prediction model to obtain interactive prediction results.

[0034] In this embodiment, nucleic acid and protein sequences can be encoded using a large language model. This large language model can be a Universal RNA Pre-trained Model (Uni-RNA) or an Evolutionary Scale Modeling Cambrian (ESMC) built based on Large Language Models (LLMs) to generate a 1024-dimensional RNA embedding vector as nucleic acid encoding features and a 960-dimensional protein embedding vector as protein encoding features. Then, interaction prediction is performed based on the nucleic acid and protein encoding features of the action prediction model to obtain interaction prediction results. The action prediction model is constructed using a deep learning network with a cross-attention mechanism. The interaction layer in the deep learning network includes cross-attention sublayers in two directions. This constructs a bidirectional interaction between protein sequence embeddings and nucleic acid sequence embeddings. The cross-attention mechanism replaces the self-attention mechanism in the transformer. Each interaction layer contains cross-attention sublayers in two directions, which are then input into a feedforward neural network (FFN) and residual connections for interaction prediction to obtain the interaction prediction results.

[0035] 103. Based on the interaction prediction results, perform fusion to obtain the interaction prediction results.

[0036] In this embodiment, to simultaneously predict interaction relationships, affinity relationships, and residue identification results, the interaction prediction results are fused to obtain an interaction prediction result that includes relevant interaction relationships, affinity relationships, and residue identification results. During fusion, attention fusion relaxation can be used to aggregate the interaction representations of proteins and nucleic acids into a fixed-length interaction vector to determine the corresponding interaction relationships, affinity relationships, and residue identification results. Here, residue identification results refer to the prediction of key residues in the protein, interaction relationships refer to the relationship between the protein and nucleic acid and whether they can adhere together, and affinity relationships refer to the strength and ease of binding between molecules.

[0037] In another embodiment of this application, for further definition and explanation, before the step of performing interactive prediction on the nucleic acid coding features and the protein coding features based on the completed model training prediction model to obtain the interactive prediction result, the method further includes: Obtain the feature training sample set; Build deep learning networks; The deep learning network is trained based on the feature training sample set, and a predictive model with completed model training is obtained after meeting the preset prediction loss conditions.

[0038] To achieve accurate prediction of protein and nucleic acid sequences, the current execution end pre-trains a predictive model for interaction prediction. This is then based on a deep learning network. The deep learning network includes normalization layers, connection layers, interaction attention layers, input layers, and output layers. In this case, the deep learning network is built upon a transformer network, replacing the self-attention mechanism in the transformer with a cross-attention mechanism. Each interaction layer contains two-directional cross-attention sublayers, which are then input into a feedforward neural network (FFN) and residual connections. Simultaneously, a feature training sample set is acquired, including fused and aligned protein and nucleic acid samples, to train the constructed deep learning network model based on these samples.

[0039] It should be noted that during the model training process, the embodiments of this application pre-set the prediction loss conditions as training constraints in the model learning process. At this time, the preset prediction loss conditions are constructed based on affinity prediction loss, residue prediction loss, interaction prediction loss, and corresponding weights.

[0040] In the model training process, this embodiment first employs a multi-objective training strategy, that is, simultaneously optimizing two tasks during training: binary interaction prediction and nucleic acid binding residue identification. Interaction prediction score This indicates that an interaction exists between a given protein-nucleic acid pair, otherwise there is no interaction. Similarly, the residue prediction score... This is represented by residues that are binding sites. The corresponding loss functions are expressed as follows: ; in, To predict the loss for interactions, For predicting loss of residues, For multi-image comparison learning loss,

[0041] at this time, For true labels of interactions in protein and nucleic acid samples, For the true label of the binding site, The weighting parameters used to balance the prediction losses of interaction and affinity are denoted as . Therefore, by summing these losses together and simultaneously optimizing the total loss, we obtain . .

[0042] Furthermore, the loss in affinity prediction can be further optimized and expressed as follows:

[0043] The expression for the Huber piecewise function is: ; in, This represents the true affinity value of the i-th sample. To predict affinity values, This represents the set of sample pairs with unequal true labels. Indicates the desired sorting direction. The threshold for Huber's piecewise function.

[0044] In a specific example of a model training process, protein-to-nucleic acid attention is used. For example, the calculation process includes: ;

[0045] in, For protein sequence characteristics, The sequence characteristics of nucleic acids, , , For learnable projection matrices, For the feature dimension, This refers to the number of attention heads. In the interaction layer, it is symmetrical, and the attention from nucleic acids to proteins can be calculated. The outputs from both directions are passed through residual connections and layer normalization, and then input into a feedforward fully connected FFN layer. After stacking N layers, fully interactive protein and nucleic acid representations are obtained. , This enables cross-modal information interaction, allowing protein representation to focus on important nucleotide positions in nucleic acids from multiple representation subspaces. At the same time, nucleic acid representation can also focus on key residues in proteins from multiple perspectives, thereby more comprehensively capturing the long-range dependencies between the two and enhancing the discriminative ability of fusion features.

[0046] In another embodiment of this application, for further definition and explanation, before the step of obtaining the feature training sample set, the method further includes: Obtain protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples; The protein structure sample and the nucleic acid structure sample are encoded to obtain protein structure features and nucleic acid structure features, and the protein sequence sample and the nucleic acid sequence sample are encoded to obtain protein sequence features and nucleic acid sequence features; Align the protein structural features with the protein sequence features to obtain a protein sample, and align the nucleic acid structural features with the nucleic acid sequence features to obtain a nucleic acid sample.

[0047] To improve the accuracy of model training, when the current execution end obtains nucleic acid samples and protein samples, it first acquires protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples. These samples can be obtained from the Protein-Nucleic Acid Thermodynamics Database (PNATDB) or the BioLip2 database. This application embodiment does not make specific limitations.

[0048] Furthermore, the protein structure samples and nucleic acid structure samples are encoded to obtain protein structure features and nucleic acid structure features. At this point, structural embedding can be used to encode the protein structure samples and nucleic acid samples respectively. Simultaneously, the protein sequence samples and nucleic acid sequence samples are encoded to obtain protein sequence features and nucleic acid sequence features. Large Language Models (LLMs) can be used for encoding at this point. In a specific embodiment, the ESMC model can be used for encoding the protein sequence, and the Uni-RNA model can be used for encoding the nucleic acid sequence. During encoding, a 1024-dimensional RNA embedding vector and a 960-dimensional protein embedding vector can be generated. The embedding model uses an unsupervised learning approach, i.e., it is trained independently of subsequent interaction prediction tasks. Given a nucleotide sequence... and amino acid sequence At that time, we used Uni-RNA and ESMC for encoding: , ; in, as well as Represented as an embedding vector of LLMs , R is all real numbers.

[0049] It should be noted that after encoding, the protein structural features are aligned with the protein sequence features to obtain protein samples, and the nucleic acid structural features are aligned with the nucleic acid sequence features to obtain nucleic acid samples. At this point, the structural features obtained from structural encoding and the sequence features obtained from sequence encoding are input into a simple self-attention module to unify the dimension and perform comparative learning. This makes the sequence embeddings and structural embeddings of the same protein or nucleic acid closer, further rejecting all embeddings of different proteins or nucleic acids in the latent space. In this embodiment, an improved loss function based on SimCLR can be used as the comparative loss; however, this embodiment does not impose specific limitations.

[0050] In an example of protein sequence-structure alignment, the sequence of the same protein is embedded... and structural embedding Defined as correct, correct The multi-view contrast loss function is expressed as:

[0051] in, For indicator functions, when the condition is met The value is 1 at time. and Temperature parameter, batch size is , for The Sim function is used to quantify the similarity between embedded vectors x and y, and is defined as follows:

[0052] At this point, the loss function helps to maximize the protein. In both views Embedding alignment in the protein and minimizing protein... Other proteins The alignment between them. The final contrastive loss is calculated on all positive sample pairs within a minibatch. Similarly, the contrastive learning loss for nucleic acids... The same definition can also be used, and the formula for calculating the total contrast loss is expressed as: This contrastive loss can serve as the basis for constructing the loss function during model training.

[0053] In another embodiment of this application, for further definition and explanation, the step of encoding the protein structure sample and the nucleic acid structure sample to obtain protein structure features and nucleic acid structure features includes: The protein structure sample and the nucleic acid structure sample are encoded respectively using a structure encoder to obtain protein isomorphism map representations and nucleic acid isomorphism map representations; The protein isomer map representation and the nucleic acid isomer map representation are reconstructed by a decoder, and the reconstructed protein isomer map representation and the nucleic acid isomer map representation are reconstructed by masking residues based on a scaling reconstruction loss function to obtain the protein structural features and the nucleic acid structural features.

[0054] To improve the effectiveness of structure-based prediction of protein-nucleic acid interactions, the current execution end first encodes the protein structure sample and the nucleic acid structure sample separately using a structure encoder, obtaining protein heteromorphic graph representations and nucleic acid heteromorphic graph representations, respectively. Both protein and nucleic acid heteromorphic graph representations include node features and edge features. In a specific example, the structure encoder in JmcPPI is used to model the protein or nucleic acid as a heteromorphic graph, constructed based on the HetGNN framework. The encoding process can be represented as follows:

[0055] in, For the first The node features of the layer are initially one-hot encoded residue types. Let be the set of edges. Indicates batch normalization, It is a fully connected layer. It is a non-linear activation function.

[0056] In a specific embodiment, to improve the accuracy of the coding structure, this embodiment reconstructs the coding structure. First, the protein isomorphism graph representation and the nucleic acid isomorphism graph representation are reconstructed using a decoder. Specifically, a reconstruction task that adds edges while minimizing the difference between the reconstructed features and the original features can be represented by the standard reconstruction task loss function as follows:

[0057] in, These are the original node features. For decoding and reconstructing values Total number of nodes For edge type The adjacency matrix, To predict the adjacency matrix.

[0058] In this embodiment, the reconstructed protein isomorphism graph representation and nucleic acid isomorphism graph representation are further reconstructed using a scaling reconstruction loss function, resulting in masked residue reconstruction of the protein structural features and nucleic acid structural features. This scaling reconstruction loss function further enhances the model's expressive power. We also added an edge reconstruction task, and the masked reconstruction task loss becomes:

[0059] in, This represents the set of indices of the masked nodes. This represents the reconstruction features of the mask nodes by the decoder. For scaling index, Predict the adjacency matrix under masked conditions. Therefore, the final total loss is expressed as: The structural features of proteins are obtained by encoding them using the HetGNN model. and nucleic acid structural features .

[0060] In another embodiment of this application, for further definition and explanation, the step of fusing the interaction prediction results to obtain the interaction prediction results includes: The protein prediction results and nucleic acid prediction results are fused using an attention network to obtain the relevant interaction relationship and the affinity relationship; The protein prediction results are processed based on the convolutional neural network model to obtain the residue identification results.

[0061] In order to simultaneously obtain the identification results of relevant interaction relationships, affinity relationships, and residue identification results, and improve the efficiency of interaction prediction, the current execution end, after obtaining the interaction prediction results including protein prediction results and nucleic acid prediction results, fuses the protein prediction results and nucleic acid prediction results through an attention network to obtain the relevant interaction relationships and the affinity relationships.

[0062] In one specific embodiment, attention fusion can be used to aggregate the interaction representations of protein prediction results and nucleic acid prediction results into a fixed-length interaction vector. Specifically, for the protein representation... Using protein-protein interaction representation The binding sites are predicted using a Convolutional Neural Network (CNN). In this embodiment, the training process of the CNN model is not specifically limited.

[0063] In another embodiment of this application, for further definition and explanation, the step of fusing the protein prediction results and nucleic acid prediction results through an attention network to obtain the correlation relationship and the affinity relationship includes: The protein prediction results and the nucleic acid prediction results are reduced in dimensionality using a linear relationship, and attention weights are calculated for the reduced protein prediction results and the nucleic acid prediction results based on attention scores. The calculated attention weights are weighted to obtain a global vector for the protein and a global vector for the nucleic acid. The global vectors for the protein and the nucleic acid are then concatenated to generate the relevant interaction relationship and the affinity relationship.

[0064] To obtain more effective affinity and interaction relationships, during fusion using an attention network, the protein and nucleic acid prediction results are dimensionality-reduced through a linear relationship. Here, the linear relationship can be a linear mapping within a linear layer for dimensionality reduction. Furthermore, attention weights are calculated for the dimensionality-reduced protein and nucleic acid prediction results based on attention scores. These calculated attention weights are then weighted to obtain global vectors for the protein and nucleic acid. Finally, these global vectors are concatenated to generate the relevant interaction and affinity relationships. Specifically, for protein representation... and nucleic acid Dimensionality reduction to linear layers respectively Attention score matrix The calculation is as follows:

[0065] in, For learnable parameter vectors, Indicates taking the first element of the matrix. Row vectors, for The attention weights for proteins and nucleic acids are obtained by averaging the rows and columns and applying softmax. , .

[0066] in, Global vectors of proteins and nucleic acids are obtained through weighted aggregation. The spliced ​​data is then processed through a linear layer to predict interactions.

[0067] In a specific example, such as Figure 2 The network architecture shown in the training phase is as follows: In part a, the sequences and structures of proteins and nucleic acids are processed by a large language model and a structure encoding module, respectively, to extract features. After processing by the self-attention module, the sequence and structure features are aligned in the latent space through contrastive learning. Subsequently, the sequence information of proteins and nucleic acids is jointly input into the CA-Former module for information exchange. Next, the protein sequence features are input into the CNN module to predict binding sites, and simultaneously combined with the nucleic acid sequence information in the fusion module. The final generated fused embedding vector is used for interaction and affinity prediction. In part b, the architecture of the structure encoding module is shown. This module constructs a residue-level heterogeneous graph based on HetGNN and optimizes the total loss function, which includes standard node reconstruction, masked residue reconstruction, and edge reconstruction, to extract the structural features of proteins and nucleic acids. In part c, the workflow of the fusion module is shown. First, the protein and nucleic acid representations are linearly projected; then, the attention score matrix is ​​calculated, and the attention weights are obtained by combining the row and column averages through the softmax function; finally, weighted aggregation is performed to obtain global vectors of the two molecules, which are then concatenated through a linear layer. In part d, the workflow of the prediction phase PANIP is shown. In the prediction phase, only the sequence information of the protein and nucleic acid is needed to predict the interaction / affinity and binding residues.

[0068] In a specific example, such as Figure 3As shown, sections a and b display the statistical results of the PRI dataset under different parameter settings, as well as the average performance of the prediction method of this application and other methods in the interaction task using five-fold cross-validation. Data points represent the validation results of each fold, and n=5. Section c displays the statistical and comparative results of the interaction prediction performance on the PDI dataset. Data points represent the validation results of each fold, and n=5. Scatter plots in sections d and f compare the performance of PANIP and AlphaFold3, and ProRNA3D single models on the PDI and PRI datasets using AUROC and MCC metrics, respectively, for the binding site prediction task. The line graph in section e shows the predicted probability of PANIP (blue) and AlphaFold3 (gray) for the two groups of protein-RNA pairing binding sites, with the actual binding sites marked by red dashed lines.

[0069] In a specific example, the training dataset in this application embodiment can be obtained from the BioLip2 database. The BioLip2 database defines each interacting protein-nucleic acid chain pair in the complex as an entry, collecting protein-RNA interaction datasets and protein-DNA interaction datasets from it. Then, entries with completely identical protein and nucleic acid sequences are removed, and the length of the protein and nucleic acid is limited to less than 750. To reduce the impact of positive and negative sample imbalance in site identification, samples with fewer than 3 sites are removed. Finally, PRI and PDI data are obtained for training interaction prediction and site prediction tasks. To avoid data leakage, the data are clustered according to 40% protein sequence similarity and 80% nucleic acid similarity, and the training set and test set are divided into classes. Negative samples are randomly generated. At the same time, the affinity dataset can be collected from CoPRA, including the structurally information-rich protein-RNA dataset PRA310 and the blind test set mCSM. PRA310 contains 310 protein-RNA complexes, while mCSM contains 79 non-redundant single-point mutations from 14 protein-RNA complexes. Two protein-nucleic acid affinity datasets without structural information were collected from the PNATDB database. Due to the large number of similar sequences, redundancy removal was performed, ultimately yielding 745 entries for protein-RNA affinity dataset PNATDB_R and 280 entries for protein-DNA affinity dataset PNATDB_D, all of which are single-stranded DNA.

[0070] This application provides a protein-nucleic acid interaction prediction method based on cross-attention learning. Compared with the prior art, this application obtains the nucleic acid sequence and protein sequence to be predicted; encodes the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features; and performs interactive prediction on the nucleic acid coding features and the protein coding features based on a pre-trained interaction prediction model to obtain interactive prediction results. The interaction prediction model is constructed by introducing a deep learning network based on the cross-attention mechanism. The interaction layer in the deep learning network includes cross-attention sublayers in two directions. The interactive prediction results are fused to obtain the interaction prediction results, which include related interaction relationships, affinity relationships, and residue identification results, achieving the common prediction purpose of three dimensions, improving the adaptability of protein-nucleic acid prediction based on multi-molecular mechanisms, meeting the interaction prediction needs when structures are missing, and thus improving the effectiveness of protein-nucleic acid interaction prediction.

[0071] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this application provides a protein-nucleic acid interaction prediction device based on cross-attention learning, such as... Figure 4 As shown, the device includes: The acquisition module 21 is used to acquire the nucleic acid sequence and protein sequence to be predicted; The prediction module 22 is used to encode the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features, and to perform interactive prediction on the nucleic acid coding features and protein coding features based on the completed model training of the effect prediction model to obtain interactive prediction results. The effect prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The fusion module 23 is used to fuse the interaction prediction results to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships and residue identification results.

[0072] Furthermore, the device also includes: a construction module and a training module. The acquisition module is also used to acquire a feature training sample set, which includes fused and aligned protein samples and nucleic acid samples. The building module is used to build a deep learning network, which includes a normalization layer, a connection layer, an interactive attention layer, an input layer, and an output layer. The training module is used to train the deep learning network based on the feature training sample set, and obtain the effective prediction model after the preset prediction loss condition is met.

[0073] Furthermore, the preset prediction loss condition is constructed based on affinity prediction loss, residue prediction loss, interaction prediction loss, and their respective weights.

[0074] Furthermore, the acquisition module is also used to acquire protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples; encode the protein structure samples and nucleic acid structure samples to obtain protein structure features and nucleic acid structure features, and encode the protein sequence samples and nucleic acid sequence samples to obtain protein sequence features and nucleic acid sequence features; align the protein structure features and protein sequence features to obtain protein samples, and align the nucleic acid structure features and nucleic acid sequence features to obtain nucleic acid samples.

[0075] Furthermore, the acquisition module is specifically used to encode the protein structure sample and the nucleic acid structure sample respectively through a structure encoder to obtain a protein isomorphism graph representation and a nucleic acid isomorphism graph representation, both of which include node features and edge features; to reconstruct the protein isomorphism graph representation and the nucleic acid isomorphism graph representation through a decoder, and to perform masked residue reconstruction on the reconstructed protein isomorphism graph representation and the nucleic acid isomorphism graph representation based on a scaling reconstruction loss function to obtain the protein structure features and the nucleic acid structure features.

[0076] Furthermore, the fusion module is specifically used to fuse the protein prediction results and nucleic acid prediction results through an attention network to obtain the correlation relationship and the affinity relationship; and to process the protein prediction results based on a convolutional neural model to obtain the residue recognition results.

[0077] Furthermore, the fusion module is specifically used to reduce the dimensionality of the protein prediction results and the nucleic acid prediction results through linear relationships, and to calculate attention weights for the dimensionality-reduced protein prediction results and the nucleic acid prediction results based on attention scores; to perform weighting based on the calculated attention weights to obtain a global vector of the protein and a global vector of the nucleic acid, and to concatenate the global vector of the protein and the global vector of the nucleic acid to generate the relevant interaction relationship and the affinity relationship.

[0078] This application provides a protein-nucleic acid interaction prediction device based on cross-attention learning. Compared with the prior art, this application obtains the nucleic acid sequence and protein sequence to be predicted; encodes the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features; and performs interactive prediction on the nucleic acid coding features and the protein coding features based on a pre-trained interaction prediction model to obtain interactive prediction results. The interaction prediction model is constructed by introducing a deep learning network based on the cross-attention mechanism. The interaction layer in the deep learning network includes cross-attention sublayers in two directions. The interaction prediction results are fused to obtain interaction prediction results, which include related interaction relationships, affinity relationships, and residue identification results, achieving the common prediction purpose of three dimensions, improving the adaptability of protein-nucleic acid prediction based on multi-molecular mechanisms, meeting the interaction prediction needs when structures are missing, and thus improving the effectiveness of protein-nucleic acid interaction prediction.

[0079] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the protein-nucleic acid interaction prediction method based on cross-attention learning in any of the above method embodiments.

[0080] Figure 5 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0081] like Figure 5 As shown, the terminal may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0082] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0083] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0084] The processor 302 is used to execute program 310, specifically to execute the relevant steps in the above-described embodiments of the protein-nucleic acid interaction prediction method based on cross-attention learning.

[0085] Specifically, program 310 may include program code that includes computer operation instructions.

[0086] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The terminal includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0087] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0088] Specifically, program 310 can be used to cause processor 302 to perform the following operations: Obtain the nucleic acid and protein sequences to be predicted; The nucleic acid sequence and the protein sequence are encoded respectively to obtain nucleic acid coding features and protein coding features. Based on the action prediction model that has been trained, the nucleic acid coding features and the protein coding features are interactively predicted to obtain interactive prediction results. The action prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The interaction prediction results are fused to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships, and residue identification results.

[0089] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0090] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting protein-nucleic acid interactions based on cross-attention learning, characterized in that, include: Obtain the nucleic acid and protein sequences to be predicted; The nucleic acid sequence and the protein sequence are encoded respectively to obtain nucleic acid coding features and protein coding features. Based on the action prediction model that has been trained, the nucleic acid coding features and the protein coding features are interactively predicted to obtain interactive prediction results. The action prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The interaction prediction results are fused to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships, and residue identification results.

2. The method according to claim 1, characterized in that, Before obtaining the interactive prediction result by performing interactive prediction on the nucleic acid coding features and the protein coding features based on the completed model training, the method further includes: Obtain a feature training sample set, which includes fused and aligned protein samples and nucleic acid samples; Construct a deep learning network, which includes a normalization layer, a connection layer, an interactive attention layer, an input layer, and an output layer; The deep learning network is trained based on the feature training sample set, and a predictive model with completed model training is obtained after meeting the preset prediction loss conditions.

3. The method according to claim 2, characterized in that, The preset prediction loss conditions are constructed based on affinity prediction loss, residue prediction loss, interaction prediction loss, and their corresponding weights.

4. The method according to claim 2, characterized in that, Before obtaining the feature training sample set, the method further includes: Obtain protein structure samples, protein sequence samples, nucleic acid structure samples, and nucleic acid sequence samples; The protein structure sample and the nucleic acid structure sample are encoded to obtain protein structure features and nucleic acid structure features, and the protein sequence sample and the nucleic acid sequence sample are encoded to obtain protein sequence features and nucleic acid sequence features; Align the protein structural features with the protein sequence features to obtain a protein sample, and align the nucleic acid structural features with the nucleic acid sequence features to obtain a nucleic acid sample.

5. The method according to claim 4, characterized in that, The process of encoding the protein structure sample and the nucleic acid structure sample to obtain protein structure features and nucleic acid structure features includes: The protein structure sample and the nucleic acid structure sample are encoded by a structural encoder respectively to obtain a protein isomorphism graph representation and a nucleic acid isomorphism graph representation. Both the protein isomorphism graph representation and the nucleic acid isomorphism graph representation include node features and edge features. The protein isomer map representation and the nucleic acid isomer map representation are reconstructed by a decoder, and the reconstructed protein isomer map representation and the nucleic acid isomer map representation are reconstructed by masking residues based on a scaling reconstruction loss function to obtain the protein structural features and the nucleic acid structural features.

6. The method according to claim 1, characterized in that, The interaction prediction results include protein prediction results and nucleic acid prediction results. The fusion of these interaction prediction results to obtain the interaction prediction results includes: The protein prediction results and nucleic acid prediction results are fused using an attention network to obtain the relevant interaction relationship and the affinity relationship; The protein prediction results are processed based on the convolutional neural network model to obtain the residue identification results.

7. The method according to claim 6, characterized in that, The step of fusing the protein prediction results and nucleic acid prediction results through an attention network to obtain the relevant interaction relationship and the affinity relationship includes: The protein prediction results and the nucleic acid prediction results are reduced in dimensionality using a linear relationship, and attention weights are calculated for the reduced protein prediction results and the nucleic acid prediction results based on attention scores. The calculated attention weights are weighted to obtain a global vector for the protein and a global vector for the nucleic acid. The global vectors for the protein and the nucleic acid are then concatenated to generate the relevant interaction relationship and the affinity relationship.

8. A protein-nucleic acid interaction prediction device based on cross-attention learning, characterized in that, include: The acquisition module is used to acquire the nucleic acid and protein sequences to be predicted. The prediction module is used to encode the nucleic acid sequence and the protein sequence respectively to obtain nucleic acid coding features and protein coding features, and to perform interactive prediction on the nucleic acid coding features and the protein coding features based on the completed model training of the effect prediction model to obtain interactive prediction results. The effect prediction model is constructed by introducing a deep learning network based on the cross attention mechanism. The interaction layer in the deep learning network includes cross attention sublayers in two directions. The fusion module is used to fuse the interaction prediction results to obtain the interaction prediction results, which include the relevant interaction relationships, affinity relationships, and residue identification results.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images