Large model and multi-mode fused lysine modification site prediction method and device

Abstract

The invention discloses a large model and multi-mode fused lysine modification site prediction method and device, relates to the technical field of biological information, and mainly aims to solve the problem of poor prediction effectiveness of lysine modification sites in existing proteins. Comprising the following steps: acquiring amino acid sequence information of cell tissue protein, and processing the amino acid sequence information based on a protein large language model which completes model training to obtain sequence representation; obtaining amino acid atom information of the cell tissue protein, and processing an amino acid contact graph structure constructed based on the amino acid atom coordinate information based on a graph structure model of which model training is completed to obtain structural characterization; and splicing the sequence representation and the structure representation, and performing feature fusion prediction on a representation vector obtained after splicing based on a multi-modal fusion model of which model training is completed to obtain a lysine modification site prediction result.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method and apparatus for predicting lysine modification sites that integrates large-scale models and multimodal approaches. Background Technology

[0002] Identification techniques for protein post-translational modification (PTM) sites primarily rely on two main approaches: experimental and computational methods. Early research relied heavily on artificially designed sequence features, such as amino acid composition, physicochemical properties, and evolutionary conservation information. In terms of experimental methods, high-throughput detection techniques based on mass spectrometry (MS), such as HPLC-MS / MS, have become the mainstream approach. These techniques can systematically locate PTM sites by accurately measuring the mass shift of modified peptides and combining this with high-resolution mass spectrometry analysis. To improve detection specificity, researchers often employ auxiliary methods such as antibody affinity enrichment or synthetic peptide co-elution verification, while also combining quantitative proteomics and isotope labeling techniques to analyze the dynamic changes in modifications. Although these methods provide highly reliable experimental results, their high cost, long experimental cycles, and dependence on specialized equipment limit their widespread application. Therefore, optimizing experimental design and reducing resource consumption are urgent problems to be solved in PTM research.

[0003] Currently, in terms of computational methods, the introduction of machine learning techniques provides an efficient and low-cost alternative for PTM site prediction. With the development of deep learning technology, multimodal fusion deep learning models, such as ProtFusion or Deepo-Glu, can be used. These models, by combining pre-trained language models and hand-designed features, have demonstrated excellent performance in PTM site prediction tasks. However, information redundancy between hand-designed features and pre-trained embeddings can lead to low computational efficiency. Existing models lack the ability to model the dynamic regulation mechanisms of PTM (such as modification synergy or temporal dependence), and their cross-species generalization performance is poor, especially in modification types or species not covered by the training data, resulting in poor accuracy and affecting the effectiveness of lysine modification site prediction in proteins. Summary of the Invention

[0004] In view of this, this application provides a method and apparatus for predicting lysine modification sites that integrates large-scale models and multimodal approaches, with the main purpose of solving the problem of poor effectiveness in predicting lysine modification sites in existing proteins.

[0005] According to one aspect of this application, a method for predicting lysine modification sites that integrates large-scale models and multimodal approaches is provided, comprising: The amino acid sequence information of proteins in cell tissues is obtained, and the amino acid sequence information is processed based on a protein big language model that has been trained to obtain sequence characterization. The amino acid atomic coordinate information of the protein in the cell tissue is obtained, and the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information is processed based on the graph structure model that has been trained to obtain the structural characterization. The sequence representation and the structural representation are spliced ​​together, and the spliced ​​representation vector is used for feature fusion prediction based on the multimodal fusion model that has been trained to obtain the prediction result of lysine modification site.

[0006] Furthermore, before processing the amino acid sequence information based on the protein large language model that has completed model training, the method further includes: Obtain training samples of protein amino acid sequences; A protein-based large language network is constructed, comprising a multilayer perceptron and a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network includes gated memory units. The protein big language network is trained using the protein amino acid sequence training samples to obtain the protein big language model.

[0007] Furthermore, before processing the amino acid sequence information based on the protein large language model that has completed model training, the method further includes: Obtain training samples of protein amino acid sequences; A protein-based large language network is constructed, comprising a three-layer convolutional neural network and a three-layer long short-term memory network. The one-dimensional convolutional layer of the convolutional neural network is used for feature extraction and dimensionality reduction, and the long short-term memory network is used to construct long-term dependencies and output sequence representations through mean pooling.

[0008] Furthermore, before the graph structure model based on the trained model processes the amino acid contact graph structure constructed based on the amino acid atom coordinate information, the method further includes: Obtain training samples of amino acid contact diagram structures; A three-layer graph convolutional network is constructed. The first layer of the three-layer graph convolutional network is used for feature expansion, the second layer of the three-layer graph convolutional network is used for primary feature compression, and the third layer of the three-layer graph convolutional network is used for secondary feature compression. The three-layer graph convolutional network is trained using the amino acid contact graph structure training samples to obtain the graph structure model.

[0009] Furthermore, before the graph structure model based on the trained model processes the amino acid contact graph structure constructed based on the amino acid atom coordinate information, the method further includes: Obtain training samples of amino acid contact diagram structures; A two-layer graph convolutional network is constructed. The first layer of the two-layer graph convolutional network is used for primary feature compression, and the second layer of the two-layer graph convolutional network is used for secondary feature compression. The two-layer graph convolutional network is trained using the amino acid contact graph structure training samples to obtain the graph structure model.

[0010] Furthermore, before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has already completed model training to obtain the structural characterization, the method further includes: Calculate the spatial distance between any two atoms based on the amino acid atom coordinate information; Contact edges are configured between target amino acids whose spatial distance is less than a preset distance threshold, and the edge attributes of the contact edges are configured. The amino acid contact diagram structure is constructed using the contact edges and the edge properties.

[0011] Furthermore, before the multimodal fusion model based on the completed model training performs feature fusion prediction on the spliced ​​representation vector to obtain the lysine modification site prediction result, the method further includes: Construct a multimodal fusion network and obtain multimodal samples tagged with lysine modification sites; The multimodal fusion network is trained based on the multimodal samples to obtain a multimodal fusion model; The multimodal fusion network comprises a three-layer network and optimizes the model parameters using an optimizer.

[0012] According to another aspect of this application, a lysine modification site prediction device integrating large-scale modeling and multimodal methods is provided, comprising: The first processing module is used to acquire the amino acid sequence information of proteins in cell tissues, and to process the amino acid sequence information based on the protein big language model that has been trained to obtain sequence characterization. The second processing module is used to obtain the amino acid atomic coordinate information of the cell tissue protein, and to process the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information based on the graph structure model that has been trained, so as to obtain the structural characterization. The prediction module is used to concatenate the sequence representation with the structural representation, and perform feature fusion prediction on the concatenated representation vector based on the multimodal fusion model that has been trained, to obtain the prediction result of lysine modification site.

[0013] Furthermore, the device also includes: The acquisition module is used to acquire protein amino acid sequence training samples; A construction module is used to construct a protein-based large language network, which includes a multilayer perceptron and a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network includes gated memory units. The training module is used to train the protein large language network model based on the protein amino acid sequence training samples to obtain the protein large language model.

[0014] Furthermore, The acquisition module is also used to acquire protein amino acid sequence training samples; The construction module is also used to construct a protein large language network, which includes a three-layer convolutional neural network and a three-layer long short-term memory network. The one-dimensional convolutional layer of the convolutional neural network is used for feature extraction and dimensionality reduction, and the long short-term memory network is used to construct long-term dependencies and output sequence representations through mean pooling. The training module is also used to train the protein big language network model based on the protein amino acid sequence training samples to obtain the protein big language model.

[0015] Furthermore, The acquisition module is also used to acquire amino acid contact diagram structure training samples; The construction module is also used to construct a three-layer graph convolutional network. The first layer of the three-layer graph convolutional network is used for feature expansion, the second layer of the three-layer graph convolutional network is used for primary feature compression, and the third layer of the three-layer graph convolutional network is used for secondary feature compression. The training module is also used to train the three-layer graph convolutional network based on the amino acid contact graph structure training samples to obtain the graph structure model.

[0016] Furthermore, The acquisition module is also used to acquire amino acid contact diagram structure training samples; The construction module is also used to construct a two-layer graph convolutional network. The first layer of the two-layer graph convolutional network is used for primary feature compression, and the second layer of the two-layer graph convolutional network is used for secondary feature compression. The training module is also used to train the two-layer graph convolutional network based on the amino acid contact graph structure training samples to obtain the graph structure model.

[0017] Furthermore, the device also includes: a computing module and a configuration module. The calculation module is used to calculate the spatial distance between any two atoms based on the amino acid atom coordinate information; The configuration module is used to configure contact edges between target amino acids whose spatial distance is less than a preset distance threshold, and to configure the edge attributes of the contact edges. The construction module is also used to construct the amino acid contact diagram structure through the contact edges and the edge attributes.

[0018] Furthermore, The building module is also used to build a multimodal fusion network and obtain multimodal samples with lysine modification site tags; The training module is also used to train the multimodal fusion network based on the multimodal samples to obtain a multimodal fusion model; The multimodal fusion network comprises a three-layer network and optimizes the model parameters using an optimizer.

[0019] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the above-described method for predicting lysine modification sites by fusing large models and multimodal approaches.

[0020] According to another aspect of this application, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the above-described method for predicting lysine modification sites by fusing large models and multimodal approaches.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a method and apparatus for predicting lysine modification sites by integrating a large model and multimodal approaches. Compared with existing technologies, the embodiments of this application obtain the amino acid sequence information of cellular tissue proteins and process the amino acid sequence information based on a protein large language model that has completed model training to obtain sequence characterization; obtain the amino acid atomic coordinate information of the cellular tissue proteins and process the amino acid contact graph structure constructed based on the amino acid atomic coordinate information based on a graph structure model that has completed model training to obtain structural characterization; concatenate the sequence characterization and the structural characterization, and perform feature fusion prediction on the concatenated characterization vector based on a multimodal fusion model that has completed model training to obtain the lysine modification site prediction result. This achieves the purpose of multimodal data fusion processing, improves cross-species generalization performance, improves the computational efficiency and learning ability of the learning model, and thus improves the effectiveness of lysine modification site prediction in proteins.

[0022] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper presents a flowchart of a method for predicting lysine modification sites that integrates large-scale models and multimodal approaches, as provided in an embodiment of this application. Figure 2 This illustration shows a flowchart of a multimodal characterization process provided in an embodiment of this application; Figure 3 A flowchart of a sequence characterization process provided in an embodiment of this application is shown; Figure 4 A flowchart of a structural characterization process provided in an embodiment of this application is shown; Figure 5 This paper illustrates a flowchart of an amino acid contact diagram structure processing method provided in an embodiment of this application. Figure 6 A flowchart of an overall method for predicting lysine modification sites provided in an embodiment of this application is shown; Figure 7 This illustration shows a block diagram of a lysine modification site prediction device that integrates large-scale models and multimodal approaches, as provided in an embodiment of this application. Figure 8 A schematic diagram of the structure of a terminal provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] This application provides a method for predicting lysine modification sites that integrates large-scale models and multimodal approaches, such as... Figure 1 As shown, the method includes: 101. Obtain the amino acid sequence information of proteins in cell tissues, and process the amino acid sequence information based on the protein big language model that has been trained to obtain sequence characterization.

[0026] In this embodiment, the current execution entity, acting as the processing end for predicting lysine modification sites by fusing large-scale models and multimodal approaches, can be a terminal device or a cloud server, etc., to acquire amino acid sequence information of proteins from cell tissues. The cell tissues can be obtained from different species within a large-scale spatiotemporal omics dataset after being segmented along different dimensions, or they can be cell tissues collected in real time. Correspondingly, the amino acid sequence information is the site-related amino acid sequence obtained from a protein database, and semantic representations can be extracted based on a large-scale protein language model ESM-2 (such as Evolutionary Scale Modeling-2). This embodiment does not impose specific limitations on this. In some embodiments, the current execution end can load the esm2_t33_650M_UR50D model as a tool for extracting amino acid sequence information. In this case, the esm2_t33_650M_UR50D model is a 33-layer attention mechanism transformer architecture with 650 million parameters. By cooperating with the corresponding alphabetic system and simultaneously using a batch conversion tool for sequence transformation, site-related amino acid sequences in the protein are extracted to obtain amino acid sequence information. The esm2_t33_650M_UR50D model can be set to evaluation mode to ensure deterministic encoding, and random operations such as information inactivation are disabled. When the amino acid sequence information is used as model input, it can be formatted as tuples with identifiers, and then automatically segmented by the batch conversion tool for encoding in conjunction with the esm2_t33_650M_UR50D model to obtain sequence representation. During the amino acid sequence information representation process, special CLS (start) and EOS (end) tokens can also be added to map amino acid residues to their corresponding token indices and convert them into normalized tensors compatible with the model. During the forward propagation process in the esm2_t33_650M_UR50D model, hidden representations can also be extracted from the last layer of the attention mechanism transformer. This layer outputs a 1280-dimensional context embedding vector for each token position (including special tokens) as amino acid sequence information.

[0027] It should be noted that when the current execution end processes the amino acid sequence information based on the protein big language model that has completed model training to obtain the sequence representation, the protein big language model can select a multilayer perceptron and a bidirectional long short-term memory network for construction in order to edit it into a sequence representation. At this time, the sequence representation is described in matrix form, which is a form of description of amino acids in tissue genes using matrix representation in the field of mathematics. This application embodiment does not make specific limitations.

[0028] In this embodiment, protein data can be obtained from three electronic databases: Compendium of Prontein Lysine Modifications (CPLM), UniProt, and AlphaFold. All data in these databases is in computer-readable format. The protein data in this embodiment primarily originates from CPLM, encompassing various post-translational modification types and proteins from multiple species. No wet-scale experiments were performed to obtain the data. This embodiment does not impose any specific limitations on these data.

[0029] 102. Obtain the amino acid atomic coordinate information of the cell tissue protein, and process the amino acid contact graph structure constructed based on the amino acid atomic coordinate information based on the graph structure model that has been trained to obtain the structural characterization.

[0030] In this embodiment, the current execution end can obtain protein structure data, mainly including the three-dimensional atomic coordinates of each amino acid residue, from the AlphaFold database based on the UniProt unique identifier of each protein and stored in .pdb format. This extracts the three-dimensional atomic coordinates of the amino acid residues in the protein at their corresponding positions, i.e., the amino acid atomic coordinate information. Simultaneously, an amino acid contact graph structure is constructed using the amino acid atomic coordinate information as input to a graph structure model. This graph structure model can be constructed using a three-layer or two-layer graph convolutional network to process the amino acid contact graph structure and obtain structural characterization.

[0031] 103. The sequence representation and the structural representation are spliced ​​together, and the spliced ​​representation vector is used for feature fusion prediction based on the multimodal fusion model that has been trained to obtain the prediction result of lysine modification site.

[0032] In this embodiment, after obtaining the sequence representation and structural representation, the sequence representation and structural representation, which are in one-dimensional vector form, can be concatenated as input to a multimodal fusion model to perform feature fusion prediction and obtain the final lysine modification site prediction result, such as... Figure 2 As shown in the figure, the splicing of sequence characterization and structural characterization achieves the goal of sequence-structure fusion, thereby meeting the requirements for accurate prediction of amino acid resting sites.

[0033] In another embodiment of this application, for further definition and explanation, before processing the amino acid sequence information based on the protein large language model that has completed model training, the method further includes: Obtain training samples of protein amino acid sequences; A protein-based large language network is constructed, comprising a multilayer perceptron and a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network includes gated memory units. The protein big language network is trained using the protein amino acid sequence training samples to obtain the protein big language model.

[0034] To achieve the goal of learning sequence representations in amino acids using neural networks, the current execution end pre-trains the protein large language network. Specifically, it first acquires protein amino acid sequence training samples and constructs the protein large language network. The protein large language network includes a multilayer perceptron and a bidirectional long short-term memory network. The bidirectional long short-term memory network includes gated memory units. The multilayer perceptron can map the sequence training samples used as model input from 1280 dimensions to 512 dimensions through a linear transformation, which can be represented as: ; Where x is the sequence representation or sequence training sample used as input. Subsequently, the dimensionality-reduced sequence data... The process is performed using a 3-layer bidirectional Long Short-Term Memory (LSTM) network with 120-dimensional hidden states. Since each LSM layer contains both forward and backward directions, the initial hidden state and cell state tensor have a shape of 6 × batch × 120 (where 6 = 2 directions × 3 layers), which can be represented as: ; in, Initially hidden state, The shape of the initial cell state tensor. Additionally, the bidirectional long short-term memory network module employs gated memory units to process protein sequences; these memory units include: input gates. Forgotten Gate Output gate The σ activation function is used. and Transform the input and hidden states separately. Update the cell state as follows: , ,in, This represents the Hadamard product. This protein integrates dynamic regulatory features of the large language network. Memory retention and status output This generates 240-dimensional bidirectional features and captures sequence patterns at different scales through mean pooling. Finally, in the 3-layer bidirectional long short-term memory network architecture, the computation process can be represented as follows: ; Among them, the output , where 240 = 120 × 2 comes from bidirectional concatenation, followed by mean pooling along the sequence dimension (dim=1) to generate the global sequence representation z: .

[0035] In another embodiment of this application, for further definition and explanation, before processing the amino acid sequence information based on the protein large language model that has completed model training, the method further includes: Obtain training samples of protein amino acid sequences; Constructing a large protein language network; The protein big language network is trained using the protein amino acid sequence training samples to obtain the protein big language model.

[0036] To achieve an alternative compilation method for amino acid sequence characterization and meet the demand for efficient sequence representation processing, the current execution end can construct another network during the training of the Protein Large Language Network (PLL). This PLAL consists of a three-layer convolutional neural network and a three-layer long short-term memory (LSTM) network. In this case, the one-dimensional convolutional layer of the convolutional neural network performs feature extraction and dimensionality reduction, while the LSM network is used to construct long-term dependencies and output the sequence representation through mean pooling. Specifically, during training, the PLAL adopts a cascaded feature extraction architecture. First, a one-dimensional convolutional layer performs local pattern extraction and channel dimensionality reduction on the input 1280-dimensional protein sequence features, reducing it to 640 dimensions. Then, large-scale max pooling is used to compress features and enhance translation invariance. Finally, after dimensionality transposition, the three-layer LSM network models the long-term dependencies of the sequence, and finally, mean pooling along the sequence dimension outputs the global feature representation of the sequence, i.e., the sequence representation. Figure 3 As shown.

[0037] In another embodiment of this application, for further definition and explanation, before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has completed model training, the method further includes: Obtain training samples of amino acid contact diagram structures; Construct a three-layer graph convolutional network; The three-layer graph convolutional network is trained using the amino acid contact graph structure training samples to obtain the graph structure model.

[0038] To learn the structural relationships between atoms in amino acids and improve the accuracy of multimodal prediction, a graph structure model is pre-constructed at the current execution end for training. Specifically, when constructing a three-layer graph convolutional network, the first layer is used for feature expansion, the second layer for primary feature compression, and the third layer for secondary feature compression. For example... Figure 4 As shown, the input layer receives contact map node features with a dimension of 45, corresponding to the feature representations of amino acid residues in the amino acid contact map structure. The first layer of the graph convolutional neural network expands the feature dimension to 120, and a ReLU activation function is used to introduce a non-linear transformation. At this point, 20% information can be randomly deactivated to prevent overfitting. The second layer of the graph convolutional neural network further compresses the feature dimension to 60. Here, the ReLU activation function can still be used for non-linear mapping. The final three-layer graph convolutional network (GCN) reduces the feature dimension to 30. At this point, the activation function is no longer needed to preserve the original feature space information.

[0039] It should be noted that during the network output processing phase of the training process, the node features of the last layer output are subjected to mean pooling along the node dimension to generate a graph-level global representation. Subsequently, the output dimension can be adjusted through decompression to adapt to the requirements of downstream tasks. At this point, the architecture of the three-layer graph convolutional network can achieve hierarchical extraction from local amino acid interactions to global protein structure features through layer-by-layer feature transformation. The introduction of the information random inactivation layer effectively enhances the model's generalization ability, enabling the entire network to learn the topological structure information of the contact graph in an end-to-end manner, providing effective feature encoding for subsequent protein characteristic prediction tasks.

[0040] In another embodiment of this application, for further definition and explanation, before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has completed model training, the method further includes: Obtain training samples of amino acid contact diagram structures; Construct a two-layer graph convolutional network; The two-layer graph convolutional network is trained using training samples based on the amino acid contact graph structure to obtain the graph structure model. The first layer of the two-layer graph convolutional network is used for primary feature compression, and the second layer is used for secondary feature compression. Specifically, the first layer maps the input 45-dimensional node features (i.e., the feature representations of amino acid residues in the amino acid contact graph structure) to a 40-dimensional feature space, which is further compressed to 30 dimensions by the second layer. After each layer operation, a ReLU activation function is used to introduce a non-linear transformation, and a 20% random deactivation regularization is applied after the first layer to prevent overfitting. Furthermore, the two-layer graph convolutional network can propagate information between nodes through topological relationships defined by edge indices. Finally, mean pooling along the node dimensions generates a graph-level global feature representation, i.e., a structural representation.

[0041] In another embodiment of this application, for further definition and explanation, before the step of processing the amino acid contact graph structure constructed based on the amino acid atomic coordinate information based on the graph structure model that has completed model training to obtain the structural characterization, the method further includes: Calculate the spatial distance between any two atoms based on the amino acid atom coordinate information; Contact edges are configured between target amino acids whose spatial distance is less than a preset distance threshold, and the edge attributes of the contact edges are configured. The amino acid contact diagram structure is constructed using the contact edges and the edge attributes.

[0042] To implement multimodal prediction based on amino acid contact map structure, in a specific amino acid atomic structure, to construct the contact map between residues in a protein, multiple contact standards are first configured. Specifically, the spatial distance between any two atoms is calculated based on the amino acid atomic coordinate information. For example, for any two atoms... and Their spatial distance can be calculated using Euclidean distance in three-dimensional space: .

[0043] During the construction of the contact graph, since peptide bonds exist between amino acid residues in the protein molecule, the current execution end configures contact edges between target amino acids whose spatial distance is less than a preset distance threshold, and configures the edge attributes of the contact edges; the amino acid contact graph structure is constructed through the contact edges and the edge attributes. In some embodiments, the Euclidean distance between the corresponding "C" and "N" atoms in three-dimensional space is used as the edge attribute. Then, if the corresponding residues... If the distance between atoms is less than a preset distance threshold, preferably 10 Å, then a contact edge is added between these two residues, with the edge properties being two in three-dimensional space. Euclidean distances between atoms, such as Figure 5 As shown. Simultaneously, due to the contact between charged residues glutamic acid (E), aspartic acid (D), lysine (K), arginine (R), and histidine (H), when the Euclidean distance between residues with opposite charges is less than a preset distance threshold (preferably configured as 16 Å), we add contact edges between them and set the corresponding distance as an edge attribute. In this case, the preset distance threshold between different atomic residues can be configured to different values; this embodiment does not impose specific limitations. Finally, after constructing contact maps between residues for each protein, an independent map is formed, namely the amino acid contact map structure.

[0044] In another embodiment of this application, for further definition and explanation, before the step of performing feature fusion prediction on the spliced ​​representation vector based on the multimodal fusion model that has completed model training to obtain the lysine modification site prediction result, the method further includes: Construct a multimodal fusion network and obtain multimodal samples tagged with lysine modification sites; The multimodal fusion network is trained based on the multimodal samples to obtain a multimodal fusion model.

[0045] To achieve the goal of fusing multimodal features and improve the prediction accuracy of lysine modification sites, a multimodal fusion network is pre-constructed at the current execution end, and multimodal samples labeled with lysine modification sites are acquired for training the multimodal fusion network. This multimodal fusion network consists of three layers, and its model parameters are optimized using an optimizer. Specifically, the structural representations extracted by the graph convolutional network are concatenated with the sequence representations along the feature dimension to form a 320-dimensional joint feature representation. The multimodal fusion network then first applies random inactivation regularization (e.g., setting the regularization parameter p=0.2) to the concatenated joint feature representation to suppress neuronal co-adaptation, enhance the model's generalization ability, and prevent overfitting during the multimodal fusion process. Furthermore, the multimodal fusion network can progressively achieve nonlinear feature transformation and dimensionality reduction through three fully connected layers. The specific steps may include: the first layer maps the 320-dimensional joint feature representation to a 128-dimensional space and introduces nonlinearity using the ReLU activation function; the second layer further compresses the dimensionality-reduced joint feature representation to 32 dimensions, also using ReLU activation; the final layer outputs the unnormalized log probability for 2-dimensional classification, adapting to binary classification tasks, such as... Figure 6 As shown. Finally, the network performs log-probability normalization on the output feature representation to ensure numerical stability and adapt to the cross-entropy loss calculation. At the same time, it eliminates redundant batch dimensions through compression operations to maintain the shape consistency of the output tensor, thus obtaining the final output modification sites.

[0046] It should be noted that the architecture of the multimodal fusion network in this application embodiment is characterized by adopting a feature-level fusion strategy, combining information random deactivation regularization and ReLU activation function, so as to achieve effective multimodal information integration while maintaining model simplicity, and can achieve efficient collaborative learning of cross-modal features with low computational complexity.

[0047] In a specific implementation scenario, for a multimodal fusion model, end-to-end supervised learning can be achieved by co-processing protein sequence coding data and structural data. The multimodal fusion model can also employ the Adam optimizer (learning rate 0.0001) for parameter optimization. During training, negative log-likelihood loss is used as the objective function, and several regularization measures are introduced: weight initialization uses a Xavier uniform distribution strategy, and the feature fusion layer is set with a random inactivation rate of 0.2 to prevent overfitting. In this case, the total number of parameters in the multimodal fusion model is approximately 1.8M, including 16,200 parameters in the GCN model, 655,360 parameters in the ESM projection layer, 1,140,000 parameters in the bidirectional long short-term memory network, and 36,000 parameters in the classifier module. During training, an iterative optimization strategy can also be used, with each training round involving a complete data traversal. The main training steps include: 1) Zeroing gradients (zero_grad) prevents accumulation; 2) Forward propagation generates a log-probability output that has been normalized to log-probability. 3) Calculate the negative log-likelihood loss; 4) Automatic differentiation during backpropagation; 5) Parameter update.

[0048] At the end of each training round, the average loss is calculated to monitor the training progress. A feature-level fusion strategy is used to achieve collaborative processing of graph structure and sequence data. By implementing a combination of log probability normalization and negative log likelihood calculation in stages, numerical stability is maintained while enhancing flexibility. Through device management, a hybrid strategy combining GPU-accelerated computation and CPU result return is adopted, which can effectively achieve joint learning of multimodal features and provide a reliable framework for modeling complex data.

[0049] This application provides a method for predicting lysine modification sites by integrating a large model and multimodal approaches. Compared with existing technologies, this application obtains the amino acid sequence information of cellular tissue proteins and processes the amino acid sequence information based on a protein large language model that has been trained to obtain sequence representation; it obtains the amino acid atomic coordinate information of the cellular tissue proteins and processes the amino acid contact graph structure constructed based on the amino acid atomic coordinate information based on a graph structure model that has been trained to obtain structural representation; it concatenates the sequence representation and the structural representation, and performs feature fusion prediction on the concatenated representation vector based on a multimodal fusion model that has been trained to obtain the lysine modification site prediction result. This achieves the purpose of multimodal data fusion processing, improves cross-species generalization performance, improves the computational efficiency and learning ability of the learning model, and thus improves the effectiveness of lysine modification site prediction in proteins.

[0050] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this application provides a lysine modification site prediction device that integrates large-scale models and multimodal approaches, such as... Figure 7 As shown, the device includes: The first processing module 21 is used to acquire the amino acid sequence information of cell tissue proteins and process the amino acid sequence information based on the protein big language model that has been trained to obtain sequence characterization. The second processing module 22 is used to obtain the amino acid atomic coordinate information of the cell tissue protein, and process the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information based on the graph structure model that has been trained to obtain the structural characterization. The prediction module 23 is used to splice the sequence representation with the structural representation, and perform feature fusion prediction on the spliced ​​representation vector based on the multimodal fusion model that has completed model training, so as to obtain the prediction result of lysine modification site.

[0051] Furthermore, the device also includes: The acquisition module is used to acquire protein amino acid sequence training samples; A construction module is used to construct a protein-based large language network, which includes a multilayer perceptron and a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network includes gated memory units. The training module is used to train the protein large language network model based on the protein amino acid sequence training samples to obtain the protein large language model.

[0052] Furthermore, The acquisition module is also used to acquire protein amino acid sequence training samples; The construction module is also used to construct a protein large language network, which includes a three-layer convolutional neural network and a three-layer long short-term memory network. The one-dimensional convolutional layer of the convolutional neural network is used for feature extraction and dimensionality reduction, and the long short-term memory network is used to construct long-term dependencies and output sequence representations through mean pooling. The training module is also used to train the protein big language network model based on the protein amino acid sequence training samples to obtain the protein big language model.

[0053] Furthermore, The acquisition module is also used to acquire amino acid contact diagram structure training samples; The construction module is also used to construct a three-layer graph convolutional network. The first layer of the three-layer graph convolutional network is used for feature expansion, the second layer of the three-layer graph convolutional network is used for primary feature compression, and the third layer of the three-layer graph convolutional network is used for secondary feature compression. The training module is also used to train the three-layer graph convolutional network based on the amino acid contact graph structure training samples to obtain the graph structure model.

[0054] Furthermore, The acquisition module is also used to acquire amino acid contact diagram structure training samples; The construction module is also used to construct a two-layer graph convolutional network. The first layer of the two-layer graph convolutional network is used for primary feature compression, and the second layer of the two-layer graph convolutional network is used for secondary feature compression. The training module is also used to train the two-layer graph convolutional network based on the amino acid contact graph structure training samples to obtain the graph structure model.

[0055] Furthermore, the device also includes: a computing module and a configuration module. The calculation module is used to calculate the spatial distance between any two atoms based on the amino acid atom coordinate information; The configuration module is used to configure contact edges between target amino acids whose spatial distance is less than a preset distance threshold, and to configure the edge attributes of the contact edges. The construction module is also used to construct the amino acid contact diagram structure through the contact edges and the edge attributes.

[0056] Furthermore, The building module is also used to build a multimodal fusion network and obtain multimodal samples with lysine modification site tags; The training module is also used to train the multimodal fusion network based on the multimodal samples to obtain a multimodal fusion model; The multimodal fusion network comprises a three-layer network and optimizes the model parameters using an optimizer.

[0057] This application provides a method and apparatus for predicting lysine modification sites by integrating a large model and multimodal approaches. Compared with existing technologies, the embodiments of this application obtain the amino acid sequence information of cellular tissue proteins and process the amino acid sequence information based on a protein large language model that has completed model training to obtain sequence characterization; obtain the amino acid atomic coordinate information of the cellular tissue proteins and process the amino acid contact graph structure constructed based on the amino acid atomic coordinate information based on a graph structure model that has completed model training to obtain structural characterization; concatenate the sequence characterization and the structural characterization, and perform feature fusion prediction on the concatenated characterization vector based on a multimodal fusion model that has completed model training to obtain the lysine modification site prediction result. This achieves the purpose of multimodal data fusion processing, improves cross-species generalization performance, improves the computational efficiency and learning ability of the learning model, and thus improves the effectiveness of lysine modification site prediction in proteins.

[0058] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the lysine modification site prediction method that integrates large models and multimodal approaches in any of the above method embodiments.

[0059] Figure 8 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0060] like Figure 8 As shown, the terminal may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0061] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0062] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0063] The processor 302 is used to execute program 310, which can specifically execute the relevant steps in the above-described embodiment of the method for predicting lysine modification sites by fusing large models and multimodal approaches.

[0064] Specifically, program 310 may include program code that includes computer operation instructions.

[0065] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The terminal includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0066] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0067] Specifically, program 310 can be used to cause processor 302 to perform the following operations: The amino acid sequence information of proteins in cell tissues is obtained, and the amino acid sequence information is processed based on a protein big language model that has been trained to obtain sequence characterization. The amino acid atomic coordinate information of the protein in the cell tissue is obtained, and the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information is processed based on the graph structure model that has been trained to obtain the structural characterization. The sequence representation and the structural representation are spliced ​​together, and the spliced ​​representation vector is used for feature fusion prediction based on the multimodal fusion model that has been trained to obtain the prediction result of lysine modification site.

[0068] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0069] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting lysine modification sites that integrates large-scale models and multimodal approaches, characterized in that, include: The amino acid sequence information of proteins in cell tissues is obtained, and the amino acid sequence information is processed based on a protein big language model that has been trained to obtain sequence characterization. The amino acid atomic coordinate information of the protein in the cell tissue is obtained, and the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information is processed based on the graph structure model that has been trained to obtain the structural characterization. The sequence representation and the structural representation are spliced ​​together, and the spliced ​​representation vector is used for feature fusion prediction based on the multimodal fusion model that has been trained to obtain the prediction result of lysine modification site.

2. The method according to claim 1, characterized in that, Before processing the amino acid sequence information using the protein big language model that has already completed model training, the method further includes: Obtain training samples of protein amino acid sequences; A protein-based large language network is constructed, comprising a multilayer perceptron and a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network includes gated memory units. The protein big language network is trained using the protein amino acid sequence training samples to obtain the protein big language model.

3. The method according to claim 1, characterized in that, Before processing the amino acid sequence information using the protein big language model that has already completed model training, the method further includes: Obtain training samples of protein amino acid sequences; A protein-based large language network is constructed, comprising a three-layer convolutional neural network and a three-layer long short-term memory network. The one-dimensional convolutional layer of the convolutional neural network is used for feature extraction and dimensionality reduction, and the long short-term memory network is used to construct long-term dependencies and output sequence representations through mean pooling. The protein big language network is trained using the protein amino acid sequence training samples to obtain the protein big language model.

4. The method according to claim 1, characterized in that, Before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has already completed model training, the method further includes: Obtain training samples of amino acid contact diagram structures; A three-layer graph convolutional network is constructed. The first layer of the three-layer graph convolutional network is used for feature expansion, the second layer of the three-layer graph convolutional network is used for primary feature compression, and the third layer of the three-layer graph convolutional network is used for secondary feature compression. The three-layer graph convolutional network is trained using the amino acid contact graph structure training samples to obtain the graph structure model.

5. The method according to claim 1, characterized in that, Before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has already completed model training, the method further includes: Obtain training samples of amino acid contact diagram structures; A two-layer graph convolutional network is constructed. The first layer of the two-layer graph convolutional network is used for primary feature compression, and the second layer of the two-layer graph convolutional network is used for secondary feature compression. The two-layer graph convolutional network is trained using the amino acid contact graph structure training samples to obtain the graph structure model.

6. The method according to claim 1, characterized in that, Before processing the amino acid contact graph structure constructed based on the amino acid atom coordinate information using the graph structure model that has already completed model training to obtain the structural characterization, the method further includes: Calculate the spatial distance between any two atoms based on the amino acid atom coordinate information; Contact edges are configured between target amino acids whose spatial distance is less than a preset distance threshold, and the edge attributes of the contact edges are configured. The amino acid contact diagram structure is constructed using the contact edges and the edge properties.

7. The method according to claim 1, characterized in that, Before the multimodal fusion model based on the completed model training performs feature fusion prediction on the spliced ​​representation vector to obtain the lysine modification site prediction result, the method further includes: Construct a multimodal fusion network and obtain multimodal samples tagged with lysine modification sites; The multimodal fusion network is trained based on the multimodal samples to obtain a multimodal fusion model; The multimodal fusion network comprises a three-layer network and optimizes the model parameters using an optimizer.

8. A device for predicting lysine modification sites that integrates large-scale models and multimodal approaches, characterized in that, include: The first processing module is used to acquire the amino acid sequence information of proteins in cell tissues, and to process the amino acid sequence information based on the protein big language model that has been trained to obtain sequence characterization. The second processing module is used to obtain the amino acid atomic coordinate information of the cell tissue protein, and to process the amino acid contact diagram structure constructed based on the amino acid atomic coordinate information based on the graph structure model that has been trained, so as to obtain the structural characterization. The prediction module is used to concatenate the sequence representation with the structural representation, and perform feature fusion prediction on the concatenated representation vector based on the multimodal fusion model that has been trained, to obtain the prediction result of lysine modification site.

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

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

Images