Bi-directional intention based lncrna-protein interaction prediction method

Abstract

The application discloses a kind of lncRNA-protein interaction prediction methods based on bidirectional intention, belong to bioinformatics technical field, including the following steps: data acquisition and pre-processing, construct dataset;Through multilayer convolutional neural network, lncRNA feature extraction is carried out;Based on ACmix module, protein feature extraction is carried out;Through bidirectional intention network, the fusion of lncRNA and protein double mode features is carried out;LncRNA-protein interaction prediction is carried out to the fusion representation using three-layer multilayer perception machine network, and output is carried out in output layer;While designing loss function carries out model optimization.The model of the application is more comprehensive and accurate when capturing intersequence dependence, thereby effectively improving the reliability and generalization ability of prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bioinformatics, and particularly relates to a lncRNA-protein interaction prediction method based on bidirectional intention. BACKGROUND

[0002] The interaction between long non-coding RNA (lncRNA) and protein plays a key role in regulating cellular processes such as gene expression, and its disorder is related to various human diseases. Traditional experimental methods such as RNA immunoprecipitation (RIP) are highly reliable, but have problems such as long time-consuming, high cost, and low throughput. Computational prediction methods as a complementary solution have received widespread attention in recent years.

[0003] Current lncRNA-protein interaction prediction methods are mainly divided into three categories: methods based on traditional machine learning (such as LPI-EnEDT), methods based on matrix decomposition (such as LPI-WGRMF), and methods based on deep learning (such as Capsule-LPI, MHAM-NPI). The first two methods rely on hand-crafted features or are sensitive to data sparsity, making it difficult to model complex nonlinear dependencies between biological sequences. Although deep learning-based methods, especially graph neural networks and capsule networks, can automatically learn feature representations, they still have limitations in cross-modal feature fusion: most methods use relatively simple concatenation or one-way attention mechanisms, failing to fully exploit fine-grained bidirectional complementary information between lncRNA and protein modalities. For example, the existing technology MHAM-NPI uses a multi-head attention mechanism but lacks explicit cross-modal bidirectional guidance; Capsule-LPI integrates multi-modal features but the fusion strategy is relatively shallow. These methods are difficult to effectively model the complex reciprocal dependency between the two biological sequences, resulting in insufficient cross-modal information integration and limiting further improvement of prediction performance.

[0004] To overcome the above defects, the prior art proposes a LPILPI (lysine urea protein intolerance) prediction method based on a graph autoencoder, the process of which includes: constructing an lncRNA and protein similarity network, learning node embedding through a variational graph autoencoder, and finally predicting the association based on the embedding vector. The prior art also proposes a framework based on co-training and graph neural networks, the method of which includes: constructing the graph structure of lncRNA and protein respectively, jointly optimizing the representation of the two graphs through a co-training strategy, and predicting by using a label propagation mechanism. However, the above two methods have obvious limitations: first, the graph construction is heavily dependent on the pre-computed similarity matrix, and the accuracy of similarity calculation directly affects the performance of the model; second, the co-training strategy is sensitive to the quality of negative samples, and the training process may be unstable; third, most methods fail to fully utilize the local and global context information of the sequence itself, and lack an explicit and bidirectional modeling mechanism for cross-modal interaction. These problems restrict the generalization ability and application effect of the model in real biomedical scenarios. SUMMARY

[0005] To solve the above problems, the present application proposes a lncRNA-protein interaction prediction method based on bidirectional intention, which uses an end-to-end bidirectional intention learning framework to effectively maintain the semantic association of biological meaning while directly processing raw biological sequences. This framework has practical potential in real biological scenarios and can successfully predict multiple literature-supported interaction pairs. The present application has the advantages of strong cross-modal fusion capability, high model generalization, and stable prediction performance.

[0006] The technical scheme of the present application is as follows:

[0007] A lncRNA-protein interaction prediction method based on bidirectional intention, which constructs a lncRNA-protein interaction prediction model, the model includes a multi-layer convolutional neural network, an ACmix module, a bidirectional intention network, a three-layer multilayer perceptron network, and an output layer. The prediction method specifically includes the following steps:

[0008] Step 1, data acquisition and preprocessing, constructing a dataset;

[0009] Step 2, lncRNA feature extraction through a multi-layer convolutional neural network;

[0010] Step 3, protein feature extraction based on the ACmix module;

[0011] Step 4, fusion of lncRNA and protein dual-modal features through a bidirectional intention network;

[0012] Step 5: Use a three-layer multilayer perceptron network to predict lncRNA-protein interactions from the fused representations and output the results in the output layer; at the same time, design a loss function to optimize the model.

[0013] Furthermore, the specific process of step 1 is as follows:

[0014] Step 1.1: Obtain experimentally validated lncRNA-protein interaction pairs from the NPInter v2.0 database to form a positive sample set;

[0015] Step 1.2: Obtain lncRNA nucleotide sequences from the NONCODE v3.0 database and protein amino acid sequences from the UniProt database; use a random pairing preprocessing method to randomly pair lncRNAs with no known interactions in the lncRNA nucleotide sequences with proteins in the protein amino acid sequences to generate a negative sample set, and balance the number of positive and negative samples to form the final dataset. ;in, For the first One lncRNA; For the first One protein; For the first One tag; This refers to the size of the data; the number of lncRNAs, proteins, and tags corresponds accordingly.

[0016] Furthermore, the specific process of step 2 is as follows:

[0017] Step 2.1: Using the initialized first embedding matrix Mapping each token in the lncRNA nucleotide sequence to a 128-dimensional dense vector yields the lncRNA sequence embedding representation. ;

[0018] ;

[0019] in, To handle variable-length sequences, Indicated in the first embedding matrix Search for the first OK; It is the transpose symbol; The number of lncRNAs;

[0020] The initial values ​​of the first embedding matrix are generated by a random uniform distribution. Each element in the matrix is ​​a randomly initialized real number, representing the initial embedding feature weight of each nucleotide.

[0021] Step 2.2, the obtained lncRNA sequence embedding representation is processed by using a 4-layer convolutional neural network in series, each layer of the convolutional neural network comprising a convolution operation, batch normalization and a ReLU activation function;

[0022] ;

[0023] wherein, , are outputs of the first layer, the second layer, the third layer and the fourth layer of the convolutional neural network, respectively; is a ReLU activation function; is batch normalization; is a convolution operation; , , are weight matrices and bias vectors of the first layer, the second layer, the third layer and the fourth layer of the convolutional neural network, respectively; The convolution kernel size of the 4-layer convolutional neural network is [3, 6, 9, 9] in turn, and the output channel number is 128; the final lncRNA feature representation is obtained after the 4-layer convolutional neural network

[0024] .

[0025] Further, in the step 3, the ACmix module comprises a convolution branch and a self-attention branch; the specific process of step 3 is as follows:

[0026] Step 3.1, each residue in the protein amino acid sequence is mapped to a 128-dimensional vector by using the initialized second embedding matrix to obtain a protein sequence embedding representation ; wherein, is the number of proteins; the initial value of the second embedding matrix is also generated by a random uniform distribution, each element in the matrix is a randomly initialized real number, representing the initial feature weight of each protein embedding;

[0027] Step 3.2, the obtained protein sequence embedding representation is encoded by using a 2-layer ACmix module in series, each layer of the ACmix module synchronously performing convolution branch and self-attention branch operations, the convolution branch using a filter with kernel size [3, 6, 9], and the self-attention branch using 8 heads of attention mechanism, and fusing the outputs of the two branches through learnable parameters to finally obtain a protein feature representation .

[0028] Further, the specific process of step 3.2 is as follows:

[0029] Step 3.2.1, the convolution branch uses convolution layers with different convolution kernel sizes to process the input protein sequence embedding representation ​​Multi-scale convolutional feature extraction is performed; the convolutional layer calculates local contextual features through a sliding window, the first... Convolution output at each position Defined as:

[0030] ;

[0031] in, This is the stride index of the convolution kernel; The kernel weight matrix; The kernel size; For the input sequence at position Protein sequence embedding representation at the location; This is the index of the center position of the current convolution operation;

[0032] The convolutional feature representation is formed by concatenating the outputs of multi-scale convolutions. ;

[0033] Step 3.2.2: Simultaneously, a multi-head attention mechanism is introduced from the self-attention branch to model long-distance sequence dependencies;

[0034] First, the query matrix is ​​generated through linear transformation. Key matrix Sum matrix :

[0035] ;

[0036] ;

[0037] ;

[0038] in, , , It is a trainable projection matrix;

[0039] Then, the attention weights are calculated, and the value matrix is ​​weighted and summed to obtain the attention output of an attention head:

[0040] ;

[0041] in, For the first The attention output of each attention head; For attention mechanisms; It is the softmax function; For square root operations; The dimension of the vector in each attention head;

[0042] For multi-head attention mechanisms, it will come from The attention output of the attention head is spliced to obtain the final attention representation ;

[0043] Step 3.2.3, finally, the outputs of the convolution branch and the attention branch are linearly fused, and the protein feature representation output after fusion is defined as:

[0044] .

[0045] Further, in the step 4, the bidirectional intention network introduces a bidirectional intention mechanism for bidirectional intention fusion, and the cross-modal representation learning is strengthened through multi-head attention and intention operation; the specific process is as follows:

[0046] Step 4.1, the and are input into the multi-head self-attention module to enhance the internal dependency of each modality, and the lncRNA enhanced representation and the protein enhanced representation are obtained.

[0047] ;

[0048] ;

[0049] wherein, is multi-head self-attention; , , are the query matrix, the key matrix and the value matrix of the lncRNA respectively; , , are the query matrix, the key matrix and the value matrix of the protein respectively;

[0050] Step 4.2, the bidirectional intention mechanism is introduced for bidirectional intention fusion to obtain the feature representation after interaction:

[0051] ;

[0052] wherein, is the feature representation after interaction; , are the query feature and the key feature; is a regularization parameter; is an identity matrix; is a value feature;

[0053] Step 4.3, the bidirectional intention mechanism is used to realize the mutual guidance and refinement of the lncRNA and protein features:

[0054] ;

[0055] ;

[0056] wherein, , are respectively the features of lncRNA and protein after refining; , are learnable parameters; is a softmax function;

[0057] Step 4.4, the refined features and are max-pooled and spliced to obtain the final fusion representation ;

[0058] ;

[0059] wherein, is splicing; is max-pooling.

[0060] Further, the specific process of step 5 is:

[0061] Step 5.1, input the fusion representation into a three-layer multi-layer perception network; the three-layer multi-layer perception network is composed of three fully connected layers in series, and the hidden layer dimensions are 512, 256, and 128 respectively;

[0062] Step 5.2, after each fully connected layer, ReLU is used as an activation function; at the same time, a Dropout regularization operation is added between each fully connected layer;

[0063] ;

[0064] ;

[0065] ;

[0066] wherein, , , are respectively the results of the first, second, and third fully connected layer regularization processing; is a ReLU activation function; is a Dropout regularization operation; , respectively represent the weight matrix and bias vector corresponding to each layer;

[0067] Step 5.3, the result of the last fully connected layer regularization processing The output layer obtains the prediction probability of the lncRNA-protein interaction through a Sigmoid function ;

[0068] ;

[0069] wherein, is a Sigmoid function;

[0070] Step 5.4, a binary cross-entropy loss function and L2 regularization are used to construct a loss function for joint optimization of the model:

[0071] ;

[0072] wherein, is the label of the i-th sample; is the prediction probability of the i-th sample; is an L2 regularization coefficient; is a learnable parameter in the model; is L2 regularization; is the total number of samples; The optimizer during training adopts AdamW.

[0073] The beneficial technical effects brought by the present application are as follows.

[0074] The present application introduces a bidirectional intention network mechanism between lncRNA and protein feature representation, realizes dynamic interaction and semantic complementarity of cross-modal information, and fundamentally improves the effectiveness and prediction performance of feature fusion. Compared with traditional one-way attention or simple splicing method, the bidirectional intention structure of the present application can model the attention of lncRNA to protein features and the feedback of protein features to lncRNA features at the same time, so that the model is more comprehensive and accurate when capturing inter-sequence dependency, thereby effectively improving the reliability and generalization ability of prediction.

[0075] The present application combines convolutional neural network and ACmix module to realize collaborative modeling of local and global features. Convolutional neural network can fully mine local structure patterns in lncRNA sequence, and ACmix module is good at capturing long-range dependency in protein sequence, the combination of the two can consider both fine-grained structural features and high-level semantic information when processing biological sequence data, significantly enhancing the richness and biological relevance of feature expression.

[0076] The present application combines convolutional neural network and ACmix module to realize collaborative modeling of local and global features. Convolutional neural network can fully mine local structure patterns in lncRNA sequence, and ACmix module is good at capturing long-range dependency in protein sequence, the combination of the two can consider both fine-grained structural features and high-level semantic information when processing biological sequence data, significantly enhancing the richness and biological relevance of feature expression.

[0077] ​Through the nonlinear mapping of the three-layer multilayer perception and the Dropout regularization mechanism, the application effectively prevents overfitting while improving the discriminant ability of the classifier. The model is significantly better than existing mainstream methods in various performance indicators (including AUC, AUPR, ACC and F1 value), verifying the superiority of the application in the lncRNA-protein interaction prediction task.

[0078] In addition, the intent network structure of the application has good interpretability, and the attention weight of the model can be used to locate the key area in the sequence, providing valuable clues for subsequent experimental verification and biological mechanism analysis. This method not only can accurately identify known interactions, but also can predict potential candidate lncRNA-protein relationships, providing new computational tools and theoretical basis for disease mechanism research, drug target screening and biomarker discovery.

[0079] In summary, the lncRNA-protein interaction prediction method based on the bidirectional intent network proposed by the application has high accuracy, strong generalization and good interpretability, and has wide application prospect and significant scientific research and industrialization value in the field of bioinformatics and computational biology. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The flowchart of the lncRNA-protein interaction prediction method based on the bidirectional intent of the application.

[0081] Figure 2 The effect comparison chart of different model variants in the ablation experiment of the application; wherein (a), (b) correspond to the comparison results of AUC and AUPR when different model variants.

[0082] Figure 3 The effect comparison chart of different combinations of encoder configurations in the encoder effect analysis experiment of the application; wherein (a), (b) correspond to the comparison results of AUC and AUPR when different combinations of encoder configurations.

[0083] Figure 4 The AUC and AUPR comparison result chart of ACmix and CNN with different layers in the parameter sensitivity analysis experiment of the application; wherein (a), (b) correspond to the comparison results of ACmix and CNN with different layers.

[0084] Figure 5 The visualization comparison result chart of the real label and the model prediction label in the feature visualization experiment of the application; wherein (a) is the visualization result of the real label; (b) is the visualization result of the model prediction label of the application. DETAILED DESCRIPTION

[0085] The application will be described in further detail below with reference to the drawings and specific embodiments.

[0086] The application discloses a long-chain non-coding RNA and protein interaction prediction method based on a bidirectional intention network. The method first extracts features from lncRNA and protein sequences. Local sequence structure features of lncRNA sequences are extracted through a convolutional neural network, and global dependency relationships of protein sequences are modeled through an improved ACmix module, thereby obtaining feature vectors with higher expression capabilities. Subsequently, a bidirectional intention interaction mechanism is introduced, enabling lncRNA and protein features to guide and dynamically fuse with each other in a semantic space. The lncRNA feature can adjust the attention weight of the protein feature according to information feedback, and the protein feature can also be adaptively updated according to the lncRNA feature, realizing bidirectional coupling and deep interaction of cross-modal information. The fused features are subjected to nonlinear mapping through a three-layer multilayer perceptron, with hidden layer dimensions of 512, 256, and 128, respectively. ReLU activation functions are used in each layer, and Dropout regularization (dropout rate 0.2) is combined to suppress overfitting. Finally, a Sigmoid function is used to output the probability of whether there is interaction between lncRNA and protein. A binary cross-entropy loss function is used in the model training process, and an L2 regularization term is introduced in the objective function to prevent the model from being unstable due to excessively large parameters. The method realizes the joint optimization of feature extraction, interaction fusion, and prediction through end-to-end training, and can significantly enhance the generalization performance and biological interpretability of the model while maintaining high accuracy.

[0087] The application constructs an lncRNA-protein interaction prediction model BiIN-LPI, which includes a multilayer convolutional neural network, an ACmix module, a bidirectional intention network, a three-layer multilayer perceptron (MLP) network, and an output layer. Figure 1 As shown in the figure, the method specifically includes the following steps:

[0088] Step 1, data acquisition and preprocessing, constructing a data set; the specific process is as follows:

[0089] Step 1.1, obtaining experimentally verified lncRNA-protein interaction pairs from the NPInter v2.0 database to constitute a positive sample set;

[0090] Step 1.2, obtaining lncRNA nucleotide sequences from the NONCODE v3.0 database and protein amino acid sequences from the UniProt database; using a random pairing preprocessing method, the lncRNA in the lncRNA nucleotide sequence without known interaction is randomly paired with the protein in the protein amino acid sequence to generate a negative sample set, and the number of positive and negative samples is balanced to form a final data set . is the first lncRNA; is the first lncRNA; is the first protein; is the first protein; is the first label; is the first label; is the size of the data volume;the number of lncRNAs, proteins and labels correspond;

[0091] In the present application, the constructed data set contains 8112 lncRNA-protein interaction pairs, involving 3046 lncRNAs and 136 proteins.

[0092] Step 2, lncRNA feature extraction is performed by a multi-layer one-dimensional convolutional neural network, and hierarchical local sequence pattern features are extracted from the lncRNA nucleotide sequence;The specific process is as follows:

[0093] Step 2.1, the initialized first embedding matrix is used, the initial value of which is generated by random uniform distribution, specifically, each element in the matrix is a randomly initialized real number, representing the initial embedding feature weight of each nucleotide, that is, each token in the lncRNA nucleotide sequence is mapped to a 128-dimensional dense vector to obtain the lncRNA sequence embedding representation ; Wherein, is the number of lncRNAs;

[0094] ;

[0095] Wherein, to process variable-length sequences, zero-padding operation is adopted for all lncRNA samples to ensure that they are represented as matrices of the same dimension, represents the first in the embedding matrix find the row (i.e. the corresponding token index); is the transpose symbol.

[0096] Step 2.2, the obtained lncRNA sequence embedding representation is processed by a 4-layer one-dimensional convolutional neural network CNN, and each layer of convolutional neural network contains convolution operation, batch normalization and ReLU activation function;

[0097] ;

[0098] Wherein, , are the outputs of the layer and the layer of convolutional neural network respectively; is the ReLU activation function. is batch normalization; is convolution operation; , are weight matrix and bias vector of the first layer convolutional neural network respectively;

[0099] The convolution kernel size of the 4-layer convolutional neural network is [3, 6, 9, 9] in turn, and the output channel number is 128; after passing through the 4-layer convolutional neural network, the final lncRNA feature representation is obtained; in the application, the effective receptive field of the one-dimensional convolutional neural network covers about 24 nucleotides, which is used to capture local sequence motifs and medium-range dependencies.

[0100] Step 3, protein feature extraction based on the ACmix module, modeling space and channel dependency simultaneously from the protein sequence to obtain semantic features; the ACmix module includes a convolution branch and a self-attention branch; the specific process of step 3 is as follows:

[0101] Step 3.1, using the initialized second embedding matrix to map each residue in the protein amino acid sequence to a 128-dimensional vector to obtain a protein sequence embedding representation ; wherein, is the number of proteins; the initial value of the second embedding matrix is also generated by a random uniform distribution, and each element in the matrix is a randomly initialized real number, representing the initial feature weight of each protein embedding;

[0102] Step 3.2, using a 2-layer ACmix module in series to encode the obtained protein sequence embedding representation, and each layer of the ACmix module simultaneously performs convolution branch and self-attention branch operations, the convolution branch uses filters with kernel sizes of [3, 6, 9], and the self-attention branch uses 8 heads of attention mechanism, and the final protein feature representation is obtained by fusing the outputs of the two branches through a learnable parameter ; in the application, the fusion parameter has an initial value of 0.5 and is dynamically updated during training; specifically, first, the convolution branch uses convolution layers with different kernel sizes to perform multi-scale convolution feature extraction on the input protein sequence embedding representation . The convolution layer calculates the local context features through a sliding window, and the convolution output at the i-th position is defined as:

[0103] ;

[0104] wherein, is the sliding step index of the convolution kernel; ​is the weight matrix of the convolution kernel; is the size of the convolution kernel; is the protein sequence embedding representation of the input sequence at position for participating in the convolution calculation; is the center position index of the current convolution operation.

[0105] When using multiple sets of convolution kernels of different scales (e.g., sizes of 3, 6, 9), context-dependent information of different ranges can be extracted. The multi-scale convolution outputs are concatenated to form the convolution feature representation .

[0106] At the same time, the self-attention branch introduces a multi-head attention mechanism to model long-range sequence dependencies. First, the query matrix , the key matrix and the value matrix are generated by linear transformation:

[0107] ;

[0108] ;

[0109] ;

[0110] where , , are trainable projection matrices. Then the attention weights are calculated, and the weighted sum of the value matrix is obtained to get the attention output of an attention head:

[0111] ;

[0112] where is the attention output of the th attention head; is the attention mechanism; is the softmax function; is the square root operation; is the dimension of the vector in each attention head;

[0113] For the multi-head attention mechanism (Multi-head Attention), the attention outputs from attention heads are concatenated to get the final attention representation .

[0114] Finally, the outputs of the convolution branch and the attention branch are linearly fused to integrate local and global features. The protein feature representation output after fusion is defined as:

[0115] ;

[0116] Step 4, the bidirectional intention network introduces a bidirectional intention mechanism for bidirectional intention fusion, realizes mutual guidance and refinement of lncRNA and protein double-modal features, and strengthens cross-modal representation learning through multi-head attention and intention operation; the specific process is as follows:

[0117] Step 4.1, input and into the multi-head self-attention module to enhance the internal dependency of each modality, to obtain lncRNA enhanced representation and protein enhanced representation ; in the present application, the number of multi-head self-attention heads is set to 8;

[0118] ;

[0119] ;

[0120] wherein, is multi-head self-attention; , , are the query matrix, key matrix and value matrix of lncRNA respectively; , , are the query matrix, key matrix and value matrix of protein respectively;

[0121] Step 4.2, introduce a bidirectional intention mechanism for bidirectional intention fusion to obtain the feature representation after interaction, wherein the operation definition of the bidirectional intention mechanism is as follows:

[0122] ;

[0123] wherein, is the feature representation after interaction; , are the query feature and key feature; is transpose; is a regularization parameter, set to 1e-5; is an identity matrix; is a value feature;

[0124] Step 4.3, utilize the bidirectional intention mechanism to realize mutual guidance and refinement of lncRNA and protein features:

[0125] ;

[0126] ;

[0127] wherein, , Refined features of lncRNA and protein, respectively; , are learnable parameters; is a softmax function;

[0128] Step 4.4, performing max-pooling on the refined features and concatenating to obtain the final fusion representation ;

[0129] ;

[0130] wherein, is concatenation; is max-pooling.

[0131] Step 5, using a three-layer multilayer perceptron network to predict lncRNA-protein interactions on the fusion representation, and designing a loss function for model optimization; the specific process is as follows:

[0132] Step 5.1, inputting the fused representation into a three-layer multilayer perceptron (MLP) network. The network is composed of three fully connected layers (Fully Connected Layers) in series, with hidden layer dimensions of 512, 256, and 128, respectively. This structure is used to gradually refine high-dimensional feature information, thereby obtaining a more compact and task-related representation;

[0133] Step 5.2, using ReLU (Rectified Linear Unit) as the activation function after each fully connected layer to enhance the non-linear representation ability of the model. At the same time, in order to prevent overfitting, a Dropout regularization operation is added between each fully connected layer, with a dropout rate of 0.2, to improve the generalization performance of the model.

[0134] ;

[0135] ;

[0136] ;

[0137] wherein, , , are the results of the first, second, and third fully connected layers after regularization, respectively; is a ReLU activation function; is a Dropout regularization operation; , respectively represent the weight matrix and bias vector corresponding to each layer, used to control the parameters of linear transformation;

[0138] Step 5.3, the result of the last fully connected layer after regularization processing is input into the output layer, and the output layer obtains the prediction probability of the lncRNA-protein interaction through a Sigmoid function ;

[0139] ;

[0140] wherein, is a Sigmoid function;

[0141] Step 5.4, a binary cross-entropy loss function and L2 regularization are used to construct a loss function , which is used to jointly optimize the model:

[0142] ;

[0143] wherein, is the label of the i-th sample; is the prediction probability of the i-th sample; is an L2 regularization coefficient, set to 1e-5; is a learnable parameter in the model; is L2 regularization; is the total number of samples; The optimizer uses AdamW, the initial learning rate is set to 1e-4, and the weight decay coefficient is 1e-5.

[0144] The prediction threshold is set to 0.5, if

[0145] , the sample label is set to 1. The application also provides an lncRNA-protein interaction prediction system based on bidirectional intention, comprising: a data preprocessing module for extracting sequence data from a multi-source biological database and constructing a data set; an lncRNA feature extraction module for learning nucleotide sequence features through a multi-layer convolutional neural network; a protein feature extraction module for learning amino acid sequence features through an Acmix module; a bidirectional intention fusion module for realizing mutual guidance and enhancement of double-mode features; and an interaction prediction module for predicting lncRNA-protein interaction probability.

[0146] The application also provides an lncRNA-protein interaction prediction system based on bidirectional intention, comprising: a data preprocessing module for extracting sequence data from a multi-source biological database and constructing a data set; an lncRNA feature extraction module for learning nucleotide sequence features through a multi-layer convolutional neural network; a protein feature extraction module for learning amino acid sequence features through an Acmix module; a bidirectional intention fusion module for realizing mutual guidance and enhancement of double-mode features; and an interaction prediction module for predicting lncRNA-protein interaction probability.

[0147] ​The application effectively improves the accuracy and interpretability of interaction prediction by bidirectional intention network explicitly modeling the reciprocal dependence relationship between lncRNA and protein modalities, and synchronously capturing local and global features of protein sequences using the ACmix module. Compared with existing single-modal or weakly coupled fusion methods, the application has the technical advantages of cross-modal alignment precision, complete biological semantics preservation, and strong generalization performance, and is particularly suitable for large-scale biological sequence interaction prediction tasks.

[0148] In order to prove the feasibility and superiority of the application, the following experiments were carried out.

[0149] Experiment 1: Five-fold cross-validation;

[0150] In order to comprehensively evaluate the performance of the BiIN-LPI model proposed in the application, it is compared with several representative existing methods, which include traditional ensemble learning models, capsule networks and prediction methods based on graph neural networks. The existing methods are as follows:

[0151] LPI-EnEDT uses an ensemble strategy to combine Extra Tree and Decision Tree classifiers. The model uses tools such as Pyfeat and BioTriangle to extract features of lncRNA and protein, then splices them into paired representations, and classifies them through an ensemble model.

[0152] EnANDeep is an ensemble framework that combines three basic learners: an adaptive k- nearest neighbor classifier, a deep neural network, and a deep forest. The outputs of the three are fused through a soft voting mechanism to improve the stability of the prediction.

[0153] MHAM-NPI introduces a multi-head attention mechanism to capture complex interactions between non-coding RNA and proteins. The model projects sequence features into multiple representation spaces and models diverse interaction patterns through stacked attention layers with residual connections.

[0154] LPIGAC is an end-to-end graph autoencoder model that uses a co-training strategy. It constructs independent graph structures for lncRNA and protein, respectively, and integrates information through joint training, using graph representation learning and label propagation to predict interactions.

[0155] LPIDF is a deep forest-based LPI prediction method. The model extracts sequence features of lncRNA and protein (such as tetranucleotide composition and BioSeq2vec vector representation), then inputs them into a cascading forest model to identify potential interactions.

[0156] LPICGAE uses a combined graph autoencoder to predict LPI. This method employs a variational graph autoencoder to learn node embeddings and utilizes another graph autoencoder to reconstruct the adjacency matrix, thereby optimizing the association prediction.

[0157] GANLDA utilizes a graph attention network (GAT) to predict lncRNA-disease associations. After denoising features through principal component analysis (PCA), the model extracts important features using a multi-layer attention mechanism and classifies them through a multi-layer perceptron (MLP).

[0158] RGCNCDA is a model based on a relational graph convolutional network (R-GCN) for circular RNA-disease prediction. It integrates multiple similarity and association networks into a heterogeneous graph and uses an R-GCN encoder and a DistMult decoder to infer latent associations.

[0159] HPN (heterogeneous propagation network) alleviates the semantic confusion problem in graph neural networks through a semantic propagation mechanism. It introduces a semantic propagation module in node-level aggregation and combines different meta-paths through a semantic fusion mechanism.

[0160] Capsule-LPI integrates multi-modal features of lncRNA and proteins, including sequence, motif, physicochemical properties, and structural information. Subsequently, these features are fused through a capsule network to generate a more comprehensive interaction representation.

[0161] In the experiment, negative samples were generated by randomly pairing lncRNAs with proteins to ensure a balanced number of positive and negative samples. To obtain a fair and reliable performance evaluation, the methods described in the present invention were tested on the NPInter benchmark dataset using a five-fold cross-validation strategy. Specifically, the dataset was randomly divided into five equal parts, and each time four parts were used for training, and the remaining one part was used for testing. This process was repeated five times, so that each subset was used as a test set once. The final reported results are the average of five experiments.

[0162] Four commonly used performance evaluation indicators were used: accuracy (ACC), area under the ROC curve (AUC), area under the precision-recall curve (AUPR), and F1-score. These indicators comprehensively reflect the classification ability of the model from different angles.

[0163] The detailed results of each model under five-fold cross-validation are shown in Table 1.

[0164] Table 1 Comparison of results of each model under five-fold cross-validation

[0165] .

[0166] The experimental results show that the AUC reaches 0.9633±0.0048 (corresponding to 0.9633 (0.0048) in Table 1), the AUPR reaches 0.9473±0.0072 (corresponding to 0.9473 (0.0072) in Table 1), the ACC reaches 0.9246±0.0082 (corresponding to 0.9246 (0.0082) in Table 1), and the F1-score reaches 0.9283±0.0080 (corresponding to 0.9283 (0.0080) in Table 1), that is, the best results are obtained in all four indicators, which proves the effectiveness of the proposed framework. Especially, the AUC and AUPR are significantly higher than those of other models, indicating that the model has stronger ability to distinguish interactive and non-interactive samples.

[0167] Compared with the ensemble learning method (such as LPI-EnEDT and EnANDeep), BiIN-LPI achieves higher AUC and AUPR, indicating that simple feature splicing and voting mechanism cannot fully capture complex interaction patterns.

[0168] Compared with Capsule-LPI based on capsule network, BiIN-LPI still has significant performance improvement.

[0169] At the same time, compared with the graph neural network method (such as LPIGAC, LPICGAE, RGCNCDA and HPN), BiIN-LPI also shows better prediction ability. This further verifies the advantage of introducing the bidirectional intention mechanism in representation learning.

[0170] Overall, BiIN-LPI not only retains the expression ability of the deep learning model, but also effectively enhances the cross-modal feature fusion, making it have higher accuracy and reliability in lncRNA-protein interaction prediction.

[0171] Experiment 2: Ablation experiment

[0172] In order to quantitatively evaluate the role of the intention mechanism in the model, four different model variants are designed, and the AUC and AUPR results are reported under the same five-fold cross-validation protocol, and the specific performance distribution results are shown in Table 2. Figure 2

[0173] The four different model variants are as follows:

[0174] The whole model BiIN-LPI uses a bidirectional intention network, which can refine and strengthen the features of lncRNA and protein.

[0175] The Linear version removes the bidirectional intention network and only splices the features of the two modalities.

[0176] ​L-Int only retains the one-way intention mechanism from lncRNA to protein;

[0177] P-Int only retains the one-way intention mechanism from protein to lncRNA.

[0178] From Figure 2 The results show that BiIN-LPI has the highest discriminant performance, with an AUC of 0.9633±0.0048 and an AUPR of 0.9473±0.0072, significantly better than the other three variants.

[0179] Compared to the Linear version (AUC = 0.9387±0.0146, AUPR = 0.9108±0.0227), BiIN-LPI improves by 0.0246 in AUC and 0.0365 in AUPR.

[0180] Compared to the one-way intention models L-Int (AUC = 0.9459±0.0072, AUPR = 0.9227±0.0097) and P-Int (AUC = 0.9479±0.0069, AUPR = 0.9223±0.0098), BiIN-LPI still maintains an advantage (AUC improvement +0.0154, AUPR improvement +0.0246).

[0181] Figure 2 The box plots in Figure 6 further show that BiIN-LPI not only has a higher median value, but also presents a more compact distribution, indicating that the bidirectional intention mechanism can produce more informative and stable fused features. Compared to linear concatenation or one-way guidance, the bidirectional interaction mechanism makes feature fusion more complete and robust, thereby significantly improving prediction performance.

[0182] The above ablation experiments confirm that removing the bidirectional intention network (Linear variant) causes AUC to drop to 0.9387±0.0146 and AUPR to drop to 0.9108±0.0227; using one-way intention (L-Int or P-Int variant) performance is also lower than the complete model, proving the effectiveness of the bidirectional intention mechanism.

[0183] Experiment 3: Encoder Effect Analysis;

[0184] To study the impact of different encoder combinations on model performance, four encoder configurations based on CNN and ACmix modules were tested for lncRNA and protein sequence feature extraction.

[0185] The four combinations are as follows (abbreviations are used for ease of explanation):

[0186] LCNN-PCNN: both lncRNA and protein use CNN as encoder;

[0187] LAC-PAC: both lncRNA and protein use ACmix as encoder;

[0188] LAC-PCNN: lncRNA uses ACmix as encoder, protein uses CNN as encoder;

[0189] LCNN-PAC: lncRNA uses CNN as encoder, protein uses ACmix as encoder.

[0190] The experimental results of the above combinations are shown in Figure 3 . The results show that the combination of lncRNA using convolutional neural network and protein using ACmix (LCNN-PAC) has the best performance, with AUC reaching 0.9633±0.0048 and AUPR reaching 0.9473±0.0072, both of which are better than the other three combinations.

[0191] LCNN-PCNN and LAC-PAC have slightly lower performance (AUC is about 0.9609 and 0.9587 respectively), while the results of LAC-PCNN are similar to them but still not optimal (AUC = 0.9599, AUPR = 0.9406).

[0192] These results show that CNN is more suitable for capturing hierarchical local patterns in lncRNA sequences, while ACmix is better at modeling long-range dependencies in protein sequences.

[0193] Therefore, the complementary design of lncRNA using CNN and protein using ACmix can simultaneously utilize the advantages of both encoders, thereby achieving the best prediction performance.

[0194] Experiment 4: Parameter sensitivity analysis;

[0195] To study the effect of model depth on performance, the number of layers of ACmix encoder and CNN encoder was analyzed, and evaluated under the same five-fold cross-validation conditions. The number of ACmix layers varies from 1 to 4 (corresponding to Acmix1 to Acmix4 in Figure 4 ), and the number of CNN layers varies from 1 to 6 (corresponding to CNN1 to CNN6 in Figure 4 ). AUC and AUPR were used as evaluation indicators, and the experimental results are shown in Figure 4 .

[0196] For the ACmix module, the model performance significantly improves when the number of layers increases from 1 to 2 (AUC = 0.9633, AUPR = 0.9473); but the performance slightly decreases when the number of layers continues to increase to 3 and above.

[0197] For the CNN module, the model performance reaches the best when the number of layers is 4 (AUC = 0.9633, AUPR = 0.9473), and there is no further improvement or even a slight decrease when the depth exceeds this value.

[0198] These results show that a moderate network depth (2 layers for ACmix and 4 layers for CNN) achieves a good balance between feature representation ability and overfitting risk. Therefore, this structure is adopted as the default configuration of BiIN-LPI. In summary, the parameter sensitivity analysis determines that the model performance is best when the number of ACmix layers is 2 and the number of convolutional neural network layers is 4.

[0199] Experiment 5: Feature visualization;

[0200] To further investigate the distribution of the learned fusion features by BiIN-LPI, 1200 lncRNA-protein sample pairs are randomly selected from the test set, and t-SNE is used to map their high-dimensional representations to two-dimensional space. Figure 5 The visualization results are shown, where Fig. (a) is colored by the true label; Fig. (b) is colored by the model prediction label.

[0201] From the figure, it can be observed that class 0 (non-interaction samples) is clustered as a dense cluster, while class 1 (interaction samples) is distributed as several small islands.

[0202] The distribution of model prediction is highly consistent with the distribution of true labels, with only a small amount of misclassification in the boundary region, indicating that the learned fusion features can accurately reflect the class structure, and the decision boundary of the classifier is highly consistent with the intrinsic geometric distribution of the data.

[0203] In addition, the obvious separation of positive and negative samples in the feature space shows that the bidirectional intention network significantly enhances the integration ability of cross-modal features, making the generated representation both discriminative and robust to noise.

[0204] The clear boundary between clusters indicates that the learned embedding space can preserve biologically meaningful differences, thus exhibiting good interpretability and reliability in prediction.

[0205] In summary, this visualization result provides intuitive evidence for the high performance of BiIN-LPI: the model not only fits the training data, but also learns a representation space that is naturally separable between classes.

[0206] Experiment 6: Case Study;

[0207] To verify the effectiveness of BiIN-LPI in real biological applications, a case study was conducted on three representative proteins, namely:

[0208] HNRNPL (P14866) - involved in RNA splicing and mRNA stability regulation;

[0209] EWSR1 (Q01844) - plays a key role in transcription regulation and cancer chromosomal translocation;

[0210] CPSF6 (Q16630) - involved in mRNA cleavage and polyadenylation processes.

[0211] In this experiment, the complete dataset was used as the training set, and all potential lncRNA-protein pairs involving the above three proteins were used as the test set (excluding known interaction pairs).

[0212] Subsequently, the top 10 lncRNAs with the highest prediction scores for each protein were sorted and verified, and the results are shown in Table 2.

[0213] The results show that for P14866 (HNRNPL), 9 out of the top 10 predictions are supported by literature; for Q01844 (EWSR1), 8 out of the top 10 have been experimentally verified; and for Q16630 (CPSF6), also 8 have literature support.

[0214] For example, BiIN-LPI successfully identified the interaction between P14866 and NONHSAG039112 (LINC02273), which has been experimentally confirmed.

[0215] Studies have shown that HNRNPL can recognize the CA repeat sequence at the 3' end of LINC02273, thereby enhancing its transcription stability and promoting the activation of oncogene AGR2, driving the occurrence and metastasis of breast cancer.

[0216] Table 2 Top 10 lncRNAs with the highest prediction scores for each protein

[0217] .

[0218] In addition, the model also accurately predicted the interaction between Q16630 and NONHSAG099120 (CCAT2). Studies have shown that CCAT2 can regulate alternative splicing and polyadenylation processes by interacting with the CFIm25 / CFIm68 complex, while Q16630 is functionally closely related to CFIm68.

[0219] Meanwhile, some predictions such as P14866-NONHSAG029985, Q01844-NONHSAG012484 have not been supported by experiments, which may represent potential new lncRNA-protein interactions, providing valuable hypotheses for future biological experiments.

[0220] Overall, BiIN-LPI can not only effectively identify a large number of known interactions, but also mine new candidate pairs with biological significance, proving the application potential of the model in real biological systems.

[0221] Of course, the above description is not a limitation on the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present application should also be within the scope of the present application.

Claims

1. A method for predicting lncRNA-protein interaction based on bidirectional intention, characterized in that, A lncRNA-protein interaction prediction model is constructed, which comprises a multi-layer convolutional neural network, an ACmix module, a bidirectional intention network, a three-layer multilayer perception network, and an output layer; the prediction method specifically comprises the following steps: Step 1, data acquisition and preprocessing, constructing a data set; Step 2, lncRNA feature extraction by a multi-layer convolutional neural network; Step 3, protein feature extraction based on the ACmix module; Step 4, fusion of lncRNA and protein double-modal features by a bidirectional intention network; the bidirectional intention network introduces a bidirectional intention mechanism for bidirectional intention fusion, and strengthens cross-modal representation learning through multi-head attention and intention operations; The specific process is as follows: Step 4.1, representing lncRNA features with protein feature representation Respectively input multi-head self-attention module to enhance the internal dependence of each modal, and obtain lncRNA enhanced representation with protein enhanced representation ; ； ； wherein, is multi-head self-attention; , , are respectively a query matrix, a key matrix and a value matrix of the lncRNA; , , are respectively a query matrix, a key matrix and a value matrix of the protein; Step 4.2, introducing a bidirectional intention mechanism for bidirectional intention fusion to obtain the feature representation after interaction: ； wherein, is the feature representation after interaction; , is the query feature, the key feature; is a regularization parameter; is an identity matrix; is the value feature; Step 4.3, using the bidirectional intention mechanism to realize mutual guidance and refinement of lncRNA and protein features: ； ； wherein, , are respectively the features after refining the lncRNA and the protein; , are learnable parameters; is a softmax function; Step 4.4, post-refinement feature with max-pooling and concatenation to get the final fused representation ; ； wherein, is concatenation; is max pooling; Step 5, using a three-layer multilayer perception network to predict lncRNA-protein interaction for the fused representation, and outputting in the output layer; meanwhile, a loss function is designed for model optimization. 2.The method of claim 1, wherein, The specific process of step 1 is as follows: Step 1.1, obtaining experimentally verified lncRNA-protein interaction pairs from the NPInter v2.0 database to constitute a positive sample set; Step 1.2, obtain lncRNA nucleotide sequences from NONCODE v3.0 database and protein amino acid sequences from UniProt database; adopt random pairing pre-processing mode to randomly pair lncRNAs without known interactions in the lncRNA nucleotide sequences with proteins in the protein amino acid sequences to generate a negative sample set, and balance the number of positive and negative samples to form a final data set ; wherein, is the th lncRNA; is the th protein; is the th label; is the size of the data amount; the number of lncRNAs, proteins and labels correspond. 3.The method of claim 2, wherein the method comprises: The specific process of step 2 is as follows: Step 2.1, utilizing the initialized first embedding matrix Each token in the lncRNA nucleotide sequence is mapped to a 128-dimensional dense vector, resulting in a lncRNA sequence embedding representation ; ； wherein, to handle variable length sequences, denotes looking up the first row in the first embedding matrix ; is a transpose symbol; is the number of lncRNAs; The first embedding matrix initial value is generated by random uniform distribution, and each element in the matrix is a randomly initialized real number, representing the initial embedding feature weight of each nucleotide; Step 2.2, using a 4-layer convolutional neural network in series to process the obtained lncRNA sequence embedding representation, and each layer of convolutional neural network comprises convolution operation, batch normalization and ReLU activation function; ； wherein, , are the output of the first layer, the output of the first layer convolutional neural network, respectively; is a ReLU activation function; is a batch normalization; is a convolution operation; , are the weight matrix, the bias vector of the first layer convolutional neural network, respectively; The convolution kernel size of the 4-layer convolutional neural network is [3, 6, 9, 9] in turn, and the output channel number is 128; the lncRNA feature representation is finally obtained after the 4-layer convolutional neural network . 4.The method of claim 3, wherein, In step 3, the ACmix module comprises a convolution branch and a self-attention branch; the specific process of step 3 is as follows: Step 3.1, utilizing the initialized second embedding matrix mapping each residue in a protein amino acid sequence into a 128-dimensional vector to obtain a protein sequence embedding representation ; wherein, is the number of proteins; the initial value of the second embedding matrix is also generated by a random uniform distribution, each element in the matrix is a randomly initialized real number, representing the initial feature weight of the embedding of each protein; Step 3.2, the obtained protein sequence embedding representation is feature encoded by using 2 layers of ACmix modules in series, the convolution branch and the self-attention branch operations are synchronously performed by each layer of the ACmix modules, the convolution branch uses filters with kernel sizes of [3, 6, 9], the self-attention branch uses 8 heads of attention mechanisms, and the outputs of the convolution branch and the self-attention branch are fused through learnable parameters to obtain the final protein feature representation .

5. The method of claim 4, wherein the method is based on a bidirectional intention of the lncRNA-protein interaction. The specific process of step 3.2 is as follows: Step 3.2.1, the convolution branch adopts convolution layers with different kernel sizes to the input protein sequence embedding representation Multi-scale convolution feature extraction is performed; the convolution layer calculates the local context feature through the sliding window, and the convolution output of the first position is defined as: ​ ； in, This is the stride index of the convolution kernel; The kernel weight matrix; The kernel size; For the input sequence at position Protein sequence embedding representation at the location; This is the index of the center position of the current convolution operation; concatenating the multi-scale convolution outputs to form a convolution feature representation ; Step 3.2.2, at the same time, the self-attention branch introduces a multi-head attention mechanism to model long-distance sequence dependency; First, a query matrix is generated by a linear transformation , a key matrix , and a value matrix : ； ； ； wherein , , is a trainable projection matrix; Then, the attention weight is calculated, and the value matrix is weighted and summed to obtain the attention output of one attention head: ； wherein, is the attention output for the th attention head; is the attention mechanism; is the softmax function; is the square root operation; is the dimension of the vector in each attention head. For multi-headed attention mechanism, the attention outputs from the attention heads are concatenated to obtain the final attention representation ; Step 3.2.

3. Finally, the outputs of the convolution branch and the attention branch are linearly fused, and the protein feature representation output after fusion is defined as: 。 6.The method of claim 5, wherein the method is characterized by, The specific process of step 5 is as follows: Step 5.1, fusing the characterizations Input a three-layer multi-layer perceptron network; the three-layer multi-layer perceptron network is composed of three fully connected layers in series, and the hidden layer dimensions are 512, 256, 128, respectively; Step 5.2, ReLU is used as the activation function after each fully connected layer; at the same time, a Dropout regularization operation is added between each fully connected layer; ； ； ； wherein, , , are the results of the first, second, and third fully connected layer regularization processing, respectively; is a ReLU activation function; is a Dropout regularization operation; , respectively represent the weight matrix and bias vector corresponding to each layer. Step 5.3, the result of the last fully connected layer after regularization The output layer obtains the prediction probability of the lncRNA-protein interaction through a Sigmoid function ; ； wherein, is a sigmoid function; Step 5.4, Constructing the loss function with binary cross-entropy loss function and L2 regularization for joint optimization of the model: ； wherein, is the label of the th sample; is the predicted probability of the th sample; is the L2 regularization coefficient; is the learnable parameter in the model; is the L2 regularization; is the total number of samples; The optimizer during training is AdamW.

Images