IncRNA-protein interaction prediction method and system

Abstract

The invention relates to the technical field of new-generation information, and discloses an lncRNA-protein interaction prediction method and system, so as to improve the prediction accuracy. The method comprises the following steps: acquiring an initial protein feature representation fused with surface, structure and sequence features; obtaining an initial lncRNA feature map fused with a sequence feature kernel, an expression profile feature kernel, a sequence similarity kernel and a Gaussian interaction spectrum kernel of protein; obtaining a final lncRNA feature map and a final protein feature map from the initial protein feature representation and the initial lncRNA feature map based on a cross attention mechanism; carrying out downsampling on the final lncRNA feature map and the final protein feature map by using global maximum pool operation, and generating a combined feature representation vector consisting of a one-dimensional lncRNA feature vector and a one-dimensional protein feature vector; and inputting the joint feature representation vector into a decoder consisting of full-connection classification layers to obtain a prediction score of the LPI.

Description

Technical Field

[0001] The present invention relates to a data processing method and system suitable for prediction purposes in the field of new generation information technology, and in particular to a lncRNA-protein interaction prediction method and system. Background Art

[0002] With the development of high-throughput sequencing technology, a large number of lncRNAs (Long non-coding RNAs) have been discovered, but their functions and interaction mechanisms with proteins are not yet fully understood. lncRNA participates in the regulation of various biological processes through interaction with proteins (LPI, LncRNA--Protein interaction). Therefore, accurate prediction of LPI is crucial for further exploring the functions and effects of lncRNA, exploring the pathological mechanisms of lncRNA in disease development, mining potential drug targets, and drug screening and optimization. Although there is a lot of information about the association between lncRNA and diseases, their exact mechanism of action remains unknown. Traditional LPI prediction mainly relies on biochemical experiments and analysis, which is costly, time-consuming, and limited by experimental conditions and sample size. Methods based on sequence similarity, gene co-expression, and subcellular localization have been applied to LPI prediction, but these methods are limited by time and funds. Existing computational models have the following deficiencies in predicting LPI: less attention is paid to protein surface information, which is indispensable for LPI; when processing multimodal information of proteins, the heterogeneity between different modalities is not fully considered. Traditional methods directly connect the embeddings of different modalities, ignoring the differences between modalities and failing to effectively utilize the complementarity of different modalities. With the continuous growth of data volume, deep learning methods, as an important branch of machine learning, have great potential in the field of LPI prediction, but there are still deficiencies. Directly applying GNN (Graph Neural Network) to process protein 3D structure is difficult to fully capture its geometric information, and most methods only focus on single modality information and ignore other modalities. Therefore, it is particularly important to develop a new method that can efficiently and accurately predict the interaction between lncRNA and protein. Summary of the invention

[0003] The present invention aims to disclose a lncRNA-protein interaction prediction method and system to improve the accuracy of prediction.

[0004] To achieve the above purpose, the lncRNA-protein interaction prediction method disclosed in the present invention includes:

[0005] Step S1, extracting the structural features, surface features, and sequence features of the protein respectively.

[0006] Step S2: First, the protein surface features are processed through the Transforme architecture based on the self-attention mechanism to obtain a new surface embedding; then, the structural features and sequence features are aligned with the new surface embedding through the cross-attention mechanism of the Transformer architecture to obtain the structural embedding and sequence embedding respectively; finally, average pooling is applied to the new embeddings after surface, structure and sequence alignment, and then the pooled embeddings are merged to obtain the initial protein feature representation.

[0007] Step S3: extract the lncRNA sequence feature kernel, expression spectrum feature kernel, sequence similarity kernel and Gaussian interaction spectrum kernel with protein respectively.

[0008] Step S4: The lncRNA sequence feature kernel, the expression spectrum feature kernel, the sequence similarity kernel, and the Gaussian interaction spectrum kernel with the protein are fused into an integrated kernel; and then the integrated kernel is feature encoded through a graph convolutional network to obtain an initial lncRNA feature graph.

[0009] Step S5: The initial protein feature representation and the initial lncRNA feature map are combined based on a cross-attention mechanism to obtain a final lncRNA feature map and a final protein feature map; the final lncRNA feature map and the final protein feature map are downsampled using a global maximum pooling operation to generate a joint feature representation vector consisting of a one-dimensional lncRNA feature vector and a one-dimensional protein feature vector.

[0010] Step S6: input the joint feature representation vector into a decoder composed of a fully connected classification layer to obtain a prediction score of the LPI.

[0011] Preferably, before the joint feature representation vector is input into a decoder consisting of a fully connected classification layer, the method further comprises:

[0012] A cross-domain adaptive mechanism is used to adjust the distribution of samples, combining the cross entropy loss on the source domain and the adversarial loss of the domain discriminator to minimize the data distribution difference between the source domain and the target domain, thereby enhancing the generalization ability of cross-domain prediction.

[0013] Preferably, the step S1 comprises:

[0014] Step S11, input an atomic cloud containing 22 atomic types, obtain a directional point cloud representation of the protein surface by sampling, select the 512 surface points closest to the ligand center as surface pockets; for each selected point, select the nearest 16 atomic centers and their types; calculate the average curvature and Gaussian curvature of each point as geometric features, and splice them with the chemical feature vector calculated by the multi-layer perceptron into a complete feature vector, and use the geodesic convolution layer to obtain the final scalar embedding of the surface points.

[0015] Step S12: In the process of capturing the 3D structural features of the protein, the GVP-GNN algorithm takes the scalar features and vector features of the node as input, and updates the node embedding through the message passing step. The specific calculation formula is:

[0016]

[0017] in, represents the embedding of node i, and represents the edge embedding, m ij represents the message transmitted from node j to node i, k′ is the number of incoming messages, GVPs is the geometric vector perceptron, LayerNorm is the normalization function, and Dropout is the regularization function to prevent overfitting.

[0018] Step S13: Process the input sequence through the protein sequence pre-training model ProtBert to obtain sequence features.

[0019] Preferably, the step S5 includes:

[0020] By performing two-way information interaction between the keys and values ​​that proteins focus on and the keys and values ​​that lncRNA focuses on, information exchange and association between lncRNA and proteins are achieved; protein features regulate their own expression through the attention weights of lncRNA features, and similarly, lncRNA features regulate their own expression through the attention weights of protein features; specifically, it includes:

[0021] lncRNA feature map F L is fed into a linear layer to calculate the query vector for the lncRNA Protein feature map F p Also obtained through the linear layer, the key vector of lncRNA can be calculated Sum value vector The acquisition process is as follows:

[0022]

[0023] in, are all different weight matrices in the linear layer, d head =D e / heads is the channel dimension, i = 1, 2, ..., heads, heads is the number of attention heads; D e is the embedding dimension.

[0024] Similarly, protein feature map F p Input to a linear layer to compute the protein query vector Then take the lncRNA feature map F L Generate protein bond vector and the protein value vector The query, key and value of a protein are calculated using the following formula:

[0025]

[0026] The weight matrix and The same weight as the lncRNA attention; by applying the softmax function, the attention matrix of lncRNA and protein The calculation is as follows:

[0027]

[0028] in, is the dimension K of lncRNA and protein.

[0029] The lncRNA / protein feature map of each head is obtained by multiplying the attention matrix of each attention head with the value matrix of the corresponding lncRNA / protein; subsequently, the lncRNA / protein feature maps of all attention heads are concatenated in the channel dimension and fed into the linear layer to obtain the attended lncRNA feature representation. and protein feature representation

[0030]

[0031] in, is the shared weight matrix.

[0032] Next, the feature map of interest is combined with the original feature map to obtain the final lncRNA feature map and protein feature maps

[0033] F ZL =0.5Z L +0.5F L ;

[0034] F ZP =0.5Z P +0.5F P ;

[0035] The lncRNA and protein feature maps were downsampled by using a global max pooling operation to generate a one-dimensional lncRNA feature vector and protein feature vector

[0036] l mp =Maxpooling(FZL );

[0037] p mp =Maxpooling(F ZP );

[0038] Finally, l mp and p mp Concatenate to obtain a one-dimensional joint feature representation vector

[0039] f = Concat(l mp ,p mp ).

[0040] Preferably, the cross-domain adaptation mechanism includes:

[0041] The feature extractor F(*) is a joint representation of the input domain data generated by the lncRNA and protein feature encoder together with the cross-attention module, i.e. and For the decoder G(*), a fully connected classification layer is used, and the Softmax function is followed to obtain the predicted classification results for G(*): and

[0042] Embed the joint feature representation vector f and classifier prediction g into the joint conditional representation In , it is defined as follows:

[0043]

[0044] in, is the outer product, and FLATTEN performs a flattening operation on the outer product of the f and g vectors.

[0045] The domain discriminator D(*) is used to align the joint representation F of the source domain and the target domain and the predicted classification distribution G. F(*) and G(*) are trained to minimize the cross entropy loss L of the source domain and the source label information, thereby generating a joint conditional representation C that confuses the discriminator D(*). In the cross-domain task, the function L based on the cross entropy loss on the labeled source domain S and the adversarial loss function L of the domain discriminator adν , the CDAN optimization problem is defined as a minimum value paradigm:

[0046]

[0047] Where ω is used to weight L adν Hyperparameters of .

[0048] To achieve the above objectives, the present invention also discloses a lncRNA-protein interaction prediction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above method.

[0049] The present invention has the following beneficial effects:

[0050] 1. Effective fusion of multimodal information: The multimodal feature encoder can effectively fuse multimodal information such as the surface, structure and sequence of the protein. It aligns the features through the Transformer technology, reduces the heterogeneity between different modal features, and improves the model's efficiency in utilizing multimodal information.

[0051] 2. The introduced cross-attention module not only retains the internal features of lncRNA and protein, but also deeply explores the interaction information between them, realizes feature fusion, and improves the LPI prediction performance.

[0052] 3. Improve prediction accuracy: Compared with existing baseline methods, the present invention achieves optimal performance in predicting lncRNA-protein interaction tasks.

[0053] 4. Broad application prospects: The method of the present invention can be widely used in many fields such as lncRNA function research, disease mechanism exploration and drug target discovery. Furthermore, the CDAN module can be used to enhance the performance and generalization ability of the model in cross-domain DTI prediction tasks, so that it can better adapt to the data distribution of new fields.

[0054] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0056] Figure 1 It is the MMF-LPI disclosed in the embodiment of the present invention CDAN Schematic diagram of the overall framework of the model. DETAILED DESCRIPTION

[0057] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0058] Example 1

[0059] This embodiment discloses a lncRNA-protein interaction prediction method. The method is applied to a lncRNA-protein interaction prediction system that aligns different modal features by combining protein surface, three-dimensional structure and sequence information through multimodal feature fusion (MMF, Multimodal Framework), and using a cross-attention mechanism.

[0060] In this embodiment, the system construction of the lncRNA and protein interaction prediction system can refer to the following configuration:

[0061] A. Hardware environment: Choose a server or high-performance computer, ensure that the GPU memory is >= 8GB, the CPU memory is >= 16GB, and has sufficient computing power and storage resources. Software environment: Install the required software and libraries, including Python, PyTorch, DGL, etc. Ensure that the software versions are compatible and configure the relevant environment variables. This system can use the source code developed in Python3.8 with PyTorch 1.7.1.

[0062] B. Input: lncRNA sequence data, protein amino acid sequence, and confirmed LPI obtained from the database.

[0063] C. Heterogeneous network construction: In order to better predict LPI, this example constructs a heterogeneous graph of lncRNA-protein interaction G = (V l ,V p ,E). Among them, V l represents the set of lncRNA nodes, V p Represents a collection of protein nodes. If lncRNA With protein If there is an interaction between ij =1, otherwise e ij =0.

[0064] For ease of description, the model used in this system can be called MMF-LPI CDAN Among them, CDAN (Conditional Domain Adversarial Network) is used to improve the cross-domain generalization ability of the MFF-LPI model.

[0065] MMF-LPI CDAN The overall framework of the model can be found in Figure 1 , including protein multimodal feature fusion module, lncRNA multimodal feature kernel fusion module, cross-attention module (fusing lncRNA and protein features), decoder (for predicting LPI) and conditional domain adversarial network CDAN module.

[0066] In this embodiment, refer to Figure 1 ,MMF-LPI CDAN The algorithm flow is described as follows:

[0067] 1. Protein multimodal feature extraction and fusion

[0068] From a microscopic perspective, the basic structure of a protein is a continuous sequence of amino acids, which undergo complex folding and conformational changes in three-dimensional space, ultimately shaping the protein's unique three-dimensional structure. In view of this, in order to fully capture the overall characteristics of a protein, this embodiment uses its complete amino acid sequence as a representative carrier of global information.

[0069] 1.1: Surface feature extraction

[0070] This embodiment can use the existing dMaSIF (differentiable molecular surface interaction fingerprinting) sampling algorithm to instantly calculate and generate protein surfaces from the underlying atomic point cloud. Input an atomic cloud containing 22 types of atoms, and obtain a directional point cloud representation of the protein surface by sampling. Then, this embodiment selects the 512 surface points closest to the ligand center as surface pockets. For each selected point, the 16 nearest atomic centers (the closer the distance, the greater the correlation and the more similar) and their types are selected, and the chemical feature vector is calculated by a multi-layer perceptron. The average curvature and Gaussian curvature of each point are calculated as geometric features, and spliced ​​into a complete feature vector with the chemical features. The final scalar embedding of the surface points is obtained using a geodesic convolution layer. The convolution operation is invariant to 3D rotation and translation, allowing the model to be predicted based on the local chemical and geometric properties of the protein surface. The residue graph of the protein is used as input, where the nodes represent amino acid residues and the edges represent the interactions between the residues.

[0071] 1.2: Structural feature extraction

[0072] This embodiment can use the geometric vector perceptron (GVP) GVP-GNN combined with graph neural networks (GNNs) to capture the 3D structural features of proteins. GVP-GNN takes the scalar features and vector features of the nodes as input. The scalar features are used to ensure the rotational invariance of the molecule, and the vector features represent the absolute direction of each node. The node embedding (as a vector) is updated through the message passing step. The specific calculation formula is:

[0073]

[0074] in, represents the embedding of node i, and represents the edge embedding, m ijrepresents the message transmitted from node j to node i, k′ is the number of incoming messages, GVPs is the geometric vector perceptron, LayerNorm is the normalization function, and Dropout is the regularization function to prevent overfitting.

[0075] 1.3: Sequence feature extraction

[0076] This embodiment can process the input sequence through an existing protein sequence pre-trained model ProtBert using a masked language modeling (MLM) objective to obtain the initial sequence embedding of the model of this embodiment. ProtBert is pre-trained on a public dataset and can be used for protein sequence analysis and prediction. This embodiment only uses ProtBert to obtain the initial sequence embedding, and its network weights do not participate in subsequent multimodal training.

[0077] 1.4: Multimodal Feature Alignment

[0078] The effective fusion of protein multimodality is affected by the heterogeneity and scale differences of its different modalities. Commonly used techniques, such as concatenated embedding, often ignore the heterogeneity between different modalities when processing protein data, which may lead to the loss of modality-specific features. This example first uses Transformer Encoder to process protein surface embeddings, and deeply learns protein surface details and features through the self-attention mechanism; then, the protein structure and sequence embeddings are globally aligned with the new surface embedding processed by the encoder through Transformer Decoder, retaining the unique information and complementarity of each modality and promoting effective communication between different modalities; finally, average pooling is applied to the new embeddings of the surface, structure, and sequence, respectively, and the pooled embeddings are merged to create a unified and information-rich feature representation F. p .

[0079] In summary, protein features, including protein surface, 3D structure and sequence information, are processed through a multimodal feature extraction framework. First, the geometric and chemical features of the protein surface are extracted; second, the GVP model is used to capture the spatial and topological structural features of the protein; finally, the ProtBert model is used to extract protein sequence features, and the different modalities are aligned and fused through the cross-attention mechanism.

[0080] 2. lncRNA multimodal features and nuclear fusion

[0081] 2.1: lncRNA sequence signature core

[0082] The sequence of lncRNA is represented by a conjoint triad, and the lncRNA sequence feature kernel The radial basis function (RBF) can be used to calculate:

[0083]

[0084] in, lncRNAl i and l j The sequence feature kernel, CT(l i ) and CT(l j ) represent lncRNAl i and l j , γ is a hyperparameter, and ||·|| represents the norm of two vectors (or the modulus of the vectors).

[0085] 2.2: lncRNA expression profile feature core

[0086] The expression profile of lncRNA involves 24 cell types from the NONCODE database. After expressing each lncRNA as a 24-dimensional expression profile vector, the lncRNA expression profile kernel is generated according to the radial basis function (RBF) above. The kernel bandwidth is set to 1.

[0087] 2.3: Sequence similarity kernel of lncRNA

[0088] lncRNA sequence similarity kernel It can be calculated by the normalized Smith-Waterman (SW) score:

[0089]

[0090] in, Represents lncRNAl i and lncRNAl j , SW(·) represents the SW score calculated by the Smith-Watermark algorithm.

[0091] 2.4: Gaussian interaction kernel of lncRNA and protein

[0092] The Gaussian interaction spectrum kernel is actually a measurement method that can reflect the network topology similarity, which can be obtained by calculating the interaction spectrum. For example, the Gaussian interaction spectrum kernel of lncRNA It can be calculated by the following formula:

[0093]

[0094] Among them, the vector IP(l i ) represents lncRNAl i The interaction spectrum of lncRNAli and whether there are interactions between all proteins. Similarly, vector IP(l i ) is the i-th column vector of the interaction matrix E, representing protein p i Whether there are interactions with all lncRNAs. l is the bandwidth parameter, which can be calculated by the following formula:

[0095]

[0096] Typically, setting γ l ′ is equal to 1. It should be noted that since the calculation of the Gaussian interaction spectrum kernel requires the use of the known lncRNA-protein interaction LPI, the interaction vector IP(l i ) will change during the cross-validation (there are many types of cross-validation, generally 5-fold or 10-fold; for example: divide the data into 5 / 10 equal parts, take one of them for validation, and use the rest for training, alternate the validation set and the training set one by one, and average the results of the performance parameters in each round; average the results of multiple validations as the cross-validation performance indicator of the model on the training set; then select the model with the best performance and its corresponding hyperparameter combination based on the performance indicator of the validation set obtained during the cross-validation process), therefore, in each cross-validation, the Gaussian interaction spectrum kernel GIP needs to be recalculated according to the changed training samples.

[0097] 2.5: LncRNA multimodal feature fusion

[0098] In this embodiment, the FKL fast core fusion method can be used to fuse multiple lncRNA multimodal feature cores into an integrated core. Assume that m l is the maximum number of nodes in the heterogeneous graph, D l is the embedding dimension of the lncRNA representation vector. Then the feature map of lncRNA can be obtained by encoding the feature of lncRNA through the graph convolutional network GCN. The encoder works as follows:

[0099]

[0100] in, and are the weight matrix and bias vector of the i-th layer in GCN respectively. is the adjacency matrix with self-connection added. represents the hidden node representation of the i-th layer,

[0101] In summary, the method of extracting and processing lncRNA features includes: first, representing lncRNA sequence features by joint triplets, and calculating the sequence kernel using radial basis function (RBF); second, extracting the expression profile and sequence similarity kernel of lncRNA, and fusion of multimodal features into an integrated kernel by the fast kernel fusion (FKL) method, and then encoding the features using a graph convolutional network (GCN) so as to finally capture the interaction features between lncRNA and protein through a cross-attention module.

[0102] 3. Cross-Attention Module

[0103] After obtaining the feature maps of lncRNA and protein through the feature encoder, this embodiment introduces a cross-attention module to effectively simulate the interaction between lncRNA and protein, thereby capturing the enhanced representation of their interaction and providing a more reliable feature representation for LPI prediction. By performing bidirectional information interaction between the keys and values ​​of protein attention and the keys and values ​​of lncRNA attention, information exchange and association between lncRNA and protein are achieved, thereby capturing the interaction characteristics between lncRNA and protein targets. In this process, protein features can regulate their own expression through the attention weights of lncRNA features, and vice versa. This interaction and adjustment mechanism enables the cross-attention module to promote the flow of information between feature maps, effectively fuse lncRNA and protein features, and extract a more comprehensive LPI feature representation.

[0104] The cross-attention module mainly consists of lncRNA and protein attention. Assume that the embedding dimensions are equal and let D e =D l =D p , lncRNA feature map F L is fed into a linear layer to calculate the query vector for the lncRNA Then, the protein feature map F p It can also be obtained through the linear layer, and further calculation can be performed to obtain the key vector of lncRNA Sum value vector The acquisition process is as follows:

[0105]

[0106] in, are all different weight matrices in the linear layer, d head =D e / heads is the channel dimension, i=1,2,…,heads, heads is the number of attention heads.

[0107] Protein attention follows a process similar to lncRNA attention. Protein feature map Fp Input to a linear layer to compute the protein query vector Then take the lncRNA feature map F L Generate protein bond vector and the protein value vector The query, key and value of a protein are calculated using the following formula:

[0108]

[0109] Among them, the weight matrix and The attention matrix of lncRNA and protein has the same weight as lncRNA attention. The calculation is as follows:

[0110]

[0111] in, is the dimension K of lncRNA and protein. The lncRNA / protein feature map of each head is obtained by multiplying the attention matrix of lncRNA / protein of each attention head with the value matrix of the corresponding lncRNA / protein. Subsequently, the lncRNA / protein feature maps of all attention heads are concatenated in the channel dimension and fed into a linear layer to obtain the attended lncRNA feature representation. and protein feature representation

[0112]

[0113] Among them, and is the shared weight matrix.

[0114] Next, the feature map of interest is combined with the original feature map to obtain the final lncRNA feature map and protein feature maps

[0115] F ZL =0.5Z L +0.5F L ;

[0116] F ZP =0.5Z P +0.5F P ;

[0117] The lncRNA and protein feature maps were downsampled by using a global max pooling operation to generate a one-dimensional lncRNA feature vector and protein feature vector

[0118] l mp =Maxpooling(F ZL );

[0119] p mp =Maxpooling(F ZP );

[0120] Finally, l mp and p mp Concatenate to obtain a one-dimensional joint feature representation vector

[0121] f = Concat(l mp ,p mp ).

[0122] 4. Prediction of potential lncRNA-protein interactions

[0123] To predict LPI, this embodiment inputs the joint feature representation vector f into a decoder consisting of a fully connected classification layer, and obtains the predicted score p of LPI:

[0124] p = σ(Wf + b);

[0125] Where W and b are the learnable weight matrix and bias vector.

[0126] In the model training process, this embodiment uses back propagation to optimize the learnable parameters simultaneously. The training goal of this embodiment is to minimize the cross entropy of the loss function:

[0127]

[0128] Among them, y i represents the true value label of the i-th LPI pair. i represents the score predicted by the model. θ is a set of learnable weight matrices and bias vectors, and λ is a hyperparameter of L2 regularization to prevent overfitting.

[0129] In this embodiment, the classifier is divided into two categories: associated or unassociated. The relationship between the score and the classification is set as needed, such as setting the score to 1 when it exceeds a certain value, which can be flexibly determined by adjusting the parameters. Among them, the known association is the positive sample of the training sample, and the existing LPI used for the training set will participate in the feature extraction in the above association step in addition to labeling and other processing.

[0130] 5. Adaptive Enhancement of Cross-Domain Generalization Capabilities

[0131] Deep learning models perform well on similar data (i.e., within the domain) that is consistent with the distribution of the training data. However, the performance on different data with different distributions (i.e., across domains) is not satisfactory. To this end, this embodiment uses the CDAN module to improve the generalization ability of the MFF-LPI model from a source domain rich in labeled data to a target domain containing only unlabeled data. The CDAN module is integrated into the framework after MFF-LPI (i.e., MMF-LPICDAN), which consists of three key components: feature extractor F(*), decoder G(*), and discriminator D(*).

[0132] CDAN is a domain adaptation technique that has been used for drug-target interaction prediction. It aims to address the challenge of domain migration with different distributions. This embodiment uses CDAN to integrate the joint feature representation vector f of the source domain and the target domain and the classifier prediction g into the joint conditional representation distinguished by the discriminator. The structure of the discriminator is a three-layer fully connected network, and its specific goal is to distinguish the target domain from the source domain by minimizing the domain classification error.

[0133] In the cross-domain task, this embodiment relies on CDAN to adjust the distribution of samples to optimize the cross-domain prediction performance. The feature extractor is the lncRNA and protein feature encoder that together with the cross-attention module generates a joint representation of the input domain data, i.e. and For the decoder G(*), this embodiment uses a fully connected classification layer and follows the Softmax function to obtain the predicted classification result for G(*): and Subsequently, the joint feature representation vector f and the classifier prediction g are embedded into the joint conditional representation Its definition is as follows:

[0134]

[0135] in, is the outer product, and FLATTEN performs a flattening operation on the outer product of the f and g vectors.

[0136] Adhering to the CDAN principle, this embodiment uses a domain discriminator D(*) to align the joint representation F of the source domain and the target domain and the predicted classification distribution G. D(*) is a domain discriminator consisting of a three-layer FCN, which learns to distinguish whether the joint conditional representation C is the source domain or the target domain. F(*) and G(*) are trained to minimize the cross entropy loss L of the source domain and the source label information, thereby generating a joint conditional representation C that confuses the discriminator D(*). In the cross-domain task, this embodiment uses two loss functions: one for optimizing classification prediction and the other for optimizing the distribution alignment of the source domain and the target domain:

[0137]

[0138] Among them, L S is the cross entropy loss on the labeled source domain, L adν is the adversarial loss for the domain discriminator. The optimization problem is formulated in the minimization form:

[0139]

[0140] Among them, ω>0 is used for weighting L adν Hyperparameters of .

[0141] Thus, through L adν By introducing adversarial training, the difference in data distribution between the source domain and the target domain is reduced, thereby enhancing the generalization ability of cross-domain prediction.

[0142] Furthermore, this embodiment constructs a specific data set to evaluate the performance of the above method, which is described as follows:

[0143] [Dataset]: The model is evaluated on three datasets (RP11807, NPInter, and RP11168), with different segmentation strategies for the datasets for intra-domain and cross-domain tasks.

[0144] As shown in Table 1, RPI1807 and NPInter are from the PDB database and the NPInter v2.0 database. RPI1807 is determined by measuring the distance between RNA and protein atoms, and it consists of 1807 RPI pairs and 1436 non-RPI pairs, represented by 1078 RNAs and 3131 proteins. NPInter collects records of lncRNA and protein interaction experiments in the NPInter v2.0 database. The database contains 10412 RPI pairs and 0 non-RPI pairs, represented by 4636 RNAs and 449 proteins. Since non-RPI pairs do not exist in NPInter, which leads to a lack of negative samples in the training model, the same number of non-RPI pairs as RPI pairs are randomly selected from the proteins and RNAs that exclude RPI pairs. The RPI1168 dataset consists of 1168 RPIs and 1168 non-RPI pairs. The positive pairs are from RPI2241. RPI pairs with RNA sequences (<200nt) were removed and filtered in RPI2241, and finally 1168 RPI pairs and 0 non-RPI pairs were obtained, represented by 421 RNAs and 1035 proteins, respectively. Since RPI2241 did not have non-RPI pairs like NPInter, which led to a lack of negative samples in the training model, the same number of non-RPI pairs as RPI pairs were randomly selected from the proteins and RNAs excluded from the RPI.

[0145] Table 1 describes the datasets used in this study

[0146] Dataset RPI Non-RPI Pair RNAs protein RP11807 1807 1436 1078 3131 NPInter 10412 10412 4636 449 RP11168 1168 1168 421 1035

[0147] In-domain evaluation: For each dataset, it is randomly split into training set, validation set, and test set with a ratio of 7:1:2.

[0148] Cross-domain tasks: Since the sample size of RP11807 and RP11168 datasets is relatively small, the cross-domain performance evaluation is mainly focused on the NPInter dataset. A clustering-based bisection strategy is used to construct a cross-domain scenario, as follows:

[0149] 1. Use the single linkage clustering method to cluster the RNA and target proteins in the NPInter dataset separately and measure the pairwise distances.

[0150] 2. The distance threshold is set to 0.5 to ensure that the clusters are not too tight while maximizing the separation of different samples.

[0151] 3. Use all labeled source domain data and 80% unlabeled target domain data as the training set, and the remaining 20% ​​labeled target domain data as the test set.

[0152] [Evaluation indicators]: To show MMF-LPI CDAN To show the superiority of the framework, we use the evaluation indicators AUROC (i.e., area under the ROC curve) and AUPRC (i.e., area under the PR curve) for classification tasks as our main evaluation criteria.

[0153] Among them, true positives (TP) and true negatives (TN) are the numbers of successfully identified lncRNA-protein pairs with interactions and lncRNA-protein pairs without interactions, respectively. False positives (FP) and false negatives (FN) represent the numbers of incorrectly identified lncRNA-protein pairs with interactions and lncRNA-protein pairs without interactions.

[0154] Performance comparison: In-domain performance evaluation: In the in-domain scenario, the conventional MFF-LPI model is used without embedding the CDAN module. MFF-LPI is compared with six baseline models under a random split setting, including SVM, RF, GraphDTA, TransformerCPI, MolTrans, and DrugBAN. The comparison results on three datasets show that MFF-LPI performs competitively in all evaluation indicators in the in-domain scenario. Cross-domain performance evaluation: In order to better simulate real-world situations, we focus on more challenging cross-domain predictions, that is, situations where training data and test data have different distribution characteristics. The CDAN module is embedded into the MFF-LPI model and the MFF-LPI is used. CDANThe results show that all methods have a significant decrease compared to the previous in-domain prediction results because the information overlap between the training and test datasets is reduced. However, MFF-LPI CDAN The model significantly outperforms other state-of-the-art models on both datasets.

[0155] Ablation experiment: In order to explore the influence of CNN combined with Transformer for lncRNA feature extraction, LPI features captured by the cross attention module, and CDAN, ablation experiment was conducted. CDAN Comparison with three variant models, including a model that removes the Transformer and uses only CNN for lncRNA feature encoder, a model that removes CNN and uses only Transformer for lncRNA feature encoder, and a model that removes the cross-attention module. CDAN Compared with MFF-LPI without CDAN module and DrugBAN with CDAN (i.e. DrugBAN CDAN The experimental results show that when the model integrates the complete module, the prediction performance of MFF-LPI is enhanced, which confirms the effectiveness of combining CNN and Transformer for lncRNA feature extraction, the efficient interaction ability of the cross-attention module, and the improvement of the cross-domain performance by the CDAN module.

[0156] It is worth noting that, in order to facilitate understanding by persons outside the technical field, some of the concepts described in this embodiment can refer to the following existing papers:

[0157] 1. "Protein-ligand binding affinity prediction based on LSTM and attention mechanism", May 2024, Journal of Shaanxi Normal University (Natural Science Edition), Vol. 52, No. 3, authors are Wang Wei, Wu Shiyu, etc.

[0158] 2. “Research on Protein-Ligand Affinity Prediction Method Based on Spatial Attention Mechanism”, Master’s thesis of Shandong University, author is Wang Yuxiao.

[0159] 3. “Prediction of Protein-Ligand Interactions Based on Deep Learning”, a master’s thesis from Shandong University, written by Wang Zechen.

[0160] Example 2

[0161] Corresponding to the above method embodiment, this embodiment discloses a lncRNA-protein interaction prediction system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method of the above embodiment. The method includes at least the following steps:

[0162] Step S1, extracting the structural features, surface features, and sequence features of the protein respectively.

[0163] Step S2: First, the protein surface features are processed through the Transforme architecture based on the self-attention mechanism to obtain a new surface embedding; then, the structural features and sequence features are aligned with the new surface embedding through the cross-attention mechanism of the Transformer architecture to obtain the structural embedding and sequence embedding respectively; finally, average pooling is applied to the new embeddings after surface, structure and sequence alignment, and then the pooled embeddings are merged to obtain the initial protein feature representation.

[0164] Step S3: extract the lncRNA sequence feature kernel, expression spectrum feature kernel, sequence similarity kernel and Gaussian interaction spectrum kernel with protein respectively.

[0165] Step S4: The lncRNA sequence feature kernel, the expression spectrum feature kernel, the sequence similarity kernel, and the Gaussian interaction spectrum kernel with the protein are fused into an integrated kernel; and then the integrated kernel is feature encoded through a graph convolutional network to obtain an initial lncRNA feature graph.

[0166] Step S5: The initial protein feature representation and the initial lncRNA feature map are combined based on a cross-attention mechanism to obtain a final lncRNA feature map and a final protein feature map; the final lncRNA feature map and the final protein feature map are downsampled using a global maximum pooling operation to generate a joint feature representation vector consisting of a one-dimensional lncRNA feature vector and a one-dimensional protein feature vector.

[0167] Step S6: input the joint feature representation vector into a decoder composed of a fully connected classification layer to obtain a prediction score of the LPI.

[0168] Preferably, before the joint feature representation vector is input into a decoder consisting of a fully connected classification layer, the method further comprises:

[0169] A cross-domain adaptive mechanism is used to adjust the distribution of samples, combining the cross entropy loss on the source domain and the adversarial loss of the domain discriminator to minimize the data distribution difference between the source domain and the target domain, thereby enhancing the generalization ability of cross-domain prediction.

[0170] For specific details of the above steps, please refer to the above embodiment 1 and will not be repeated here.

[0171] In summary, the lncRNA-protein interaction prediction method and system disclosed in the embodiments of the present invention have at least the following features:

[0172] Beneficial effects:

[0173] 1. Effective fusion of multimodal information: The multimodal feature encoder can effectively fuse multimodal information such as the surface, structure and sequence of the protein. It aligns the features through the Transformer technology, reduces the heterogeneity between different modal features, and improves the model's efficiency in utilizing multimodal information.

[0174] 2. The introduced cross-attention module not only retains the internal features of lncRNA and protein, but also deeply explores the interaction information between them, realizes feature fusion, and improves the LPI prediction performance.

[0175] 3. Improve prediction accuracy: Compared with existing baseline methods, the present invention achieves optimal performance in predicting lncRNA-protein interaction tasks.

[0176] 4. Broad application prospects: The method of the present invention can be widely used in many fields such as lncRNA function research, disease mechanism exploration and drug target discovery. Furthermore, the CDAN module can be used to enhance the performance and generalization ability of the model in cross-domain DTI prediction tasks, so that it can better adapt to the data distribution of new fields.

[0177] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A lncRNA-protein interaction prediction method, characterized in that: include: Step S1, extracting the structural features, surface features, and sequence features of the protein respectively; Step S2: First, the protein surface features are processed by the Transformer architecture based on the self-attention mechanism to obtain a new surface embedding; then, the structural features and sequence features are aligned with the new surface embedding through the cross-attention mechanism of the Transformer architecture to obtain the structural embedding and sequence embedding respectively; finally, average pooling is applied to the new embeddings after surface, structure and sequence alignment, and then the pooled embeddings are merged to obtain the initial protein feature representation; Step S3, extracting lncRNA sequence feature kernel, expression spectrum feature kernel, sequence similarity kernel and Gaussian interaction spectrum kernel with protein respectively; Step S4, fusing the lncRNA sequence feature kernel, the expression spectrum feature kernel, the sequence similarity kernel, and the Gaussian interaction spectrum kernel with the protein into an integrated kernel; and then performing feature encoding on the integrated kernel through a graph convolutional network to obtain an initial lncRNA feature graph; Step S5, obtaining a final lncRNA feature map and a final protein feature map based on the cross attention mechanism by combining the initial protein feature representation and the initial lncRNA feature map; Downsampling the final lncRNA feature map and the final protein feature map using a global maximum pooling operation to generate a joint feature representation vector consisting of a one-dimensional lncRNA feature vector and a one-dimensional protein feature vector; Step S6: input the joint feature representation vector into a decoder composed of a fully connected classification layer to obtain a prediction score of the LPI.

2. The method according to claim 1, characterized in that Before the joint feature representation vector is input into a decoder composed of a fully connected classification layer, it also includes: A cross-domain adaptive mechanism is used to adjust the distribution of samples, combining the cross entropy loss on the source domain and the adversarial loss of the domain discriminator to minimize the data distribution difference between the source domain and the target domain, thereby enhancing the generalization ability of cross-domain prediction.

3. The method according to claim 1 or 2, characterized in that: The step S1 comprises: Step S11, input an atomic cloud containing 22 atomic types, obtain a directional point cloud representation of the protein surface by sampling, select 512 surface points closest to the ligand center as surface pockets; for each selected point, select the nearest 16 atomic centers and their types; calculate the average curvature and Gaussian curvature of each point as geometric features, and splice them with the chemical feature vector calculated by the multi-layer perceptron into a complete feature vector, and use the geodesic convolution layer to obtain the final scalar embedding of the surface points; Step S12: In the process of capturing the 3D structural features of the protein, the GVP-GNN algorithm takes the scalar features and vector features of the node as input, and updates the node embedding through the message passing step. The specific calculation formula is: in, represents the embedding of node i, and represents the edge embedding, m ij represents the message transmitted from node j to node i, k′ is the number of incoming messages, GVPs is the geometric vector perceptron, LayerNorm is the normalization function, and Dropout is the regularization function to prevent overfitting; Step S13: Process the input sequence through the protein sequence pre-training model ProtBert to obtain sequence features.

4. The method according to claim 1 or 2, characterized in that: In the step S5, it includes: By performing two-way information interaction between the keys and values ​​that proteins focus on and the keys and values ​​that lncRNA focuses on, information exchange and association between lncRNA and proteins are achieved; protein features regulate their own expression through the attention weights of lncRNA features, and similarly, lncRNA features regulate their own expression through the attention weights of protein features; specifically, it includes: lncRNA feature map F L is fed into a linear layer to calculate the query vector for the lncRNA Protein feature map F p Also obtained through the linear layer, the key vector of lncRNA can be calculated Sum value vector The acquisition process is as follows: in, are all different weight matrices in the linear layer, d head =D e / heads is the channel dimension, i = 1, 2, ..., heads, heads is the number of attention heads; D e is the embedding dimension; Similarly, protein feature map F p Input to a linear layer to compute the protein query vector Then take the lncRNA feature map F L Generate protein bond vector and the protein value vector The query, key and value of a protein are calculated using the following formula: The weight matrix and The same weight as the lncRNA attention; by applying the softmax function, the attention matrix of lncRNA and protein The calculation is as follows: in, is the dimension K of lncRNA and protein; The lncRNA / protein feature map of each head is obtained by multiplying the attention matrix of each attention head with the value matrix of the corresponding lncRNA / protein; subsequently, the lncRNA / protein feature maps of all attention heads are concatenated in the channel dimension and fed into the linear layer to obtain the attended lncRNA feature representation. and protein feature representation in, is the shared weight matrix; Next, the feature map of interest is combined with the original feature map to obtain the final lncRNA feature map and protein feature maps F ZL =0.5Z L +0.5F L ; F ZP =0.5Z P +0.5F P ; The lncRNA and protein feature maps were downsampled by using a global max pooling operation to generate a one-dimensional lncRNA feature vector and protein feature vector l mp =Maxpooling(F ZL ); p mp =Maxpooling(F ZP ); Finally, l mp and p mp Concatenate to obtain a one-dimensional joint feature representation vector f=Concat(l mp ,p mp )。 5. The method according to claim 4, characterized in that The cross-domain adaptive mechanism includes: The feature extractor F(*) is a joint representation of the input domain data generated by the lncRNA and protein feature encoder together with the cross-attention module, i.e. and For the decoder G(*), a fully connected classification layer is used, and the Softmax function is followed to obtain the predicted classification results for G(*): and Embed the joint feature representation vector f and classifier prediction g into the joint conditional representation In , it is defined as follows: in, is the outer product, FLATTEN performs a flattening operation on the outer product of the f and g vectors; The domain discriminator D(*) is used to align the joint representation F of the source domain and the target domain and the predicted classification distribution G. F(*) and G(*) are trained to minimize the cross entropy loss L of the source domain and the source label information, thereby generating a joint conditional representation C that confuses the discriminator D(*). In the cross-domain task, the function L based on the cross entropy loss on the labeled source domain S and the adversarial loss function L of the domain discriminator adν , the CDAN optimization problem is defined as a minimum value paradigm: Where ω is used to weight L adν Hyperparameters of .

6. A lncRNA-protein interaction prediction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to implement the method according to any one of claims 1 to 5.

Images