Method and system for predicting post-translational modification sites of proteins under multi-modal cross-view

Abstract

The application relates to the technical field of post-translational modification site prediction, and discloses a multi-modal cross-view protein post-translational modification site prediction method and system, which comprises the following steps: acquiring protein peptide segments and constructing sequence modality-based feature representation, decomposing the protein into single amino acid units and capturing local structure information of each residue, constructing structure modality-based feature representation according to the local structure information, combining the sequence modality-based feature representation and the structure modality-based feature representation to construct initial cycle state features, using RIDCGA to perform cycle iteration fusion on the initial cycle state features to obtain fusion features, and performing protein post-translational modification site prediction according to the fusion features. The application can improve the prediction accuracy in a complex real biological data scene.

Description

Technical Field

[0001] This invention relates to the field of post-translational modification site prediction technology, and in particular to a method and system for predicting protein post-translational modification sites under a multimodal cross-field perspective. Background Technology

[0002] Post-translational modification (PTM) is a key mechanism regulating protein function, cell signaling, and disease development. In biological research, PTM can affect protein spatial structure, interaction networks, and metabolic pathways, thereby participating in the regulation of core life processes such as signal transduction, gene expression, and cell cycle. In the medical field, abnormal PTM is closely related to the occurrence and development of various diseases; therefore, accurately identifying PTM sites is of great significance for a deeper understanding of the laws governing life activities, the discovery of disease-related biomarkers, and potential drug targets. Traditional experimental methods, such as mass spectrometry, can directly identify PTM sites, but these processes are time-consuming, costly, and have limited throughput, making them unsuitable for large-scale, systematic research. Against this backdrop, developing efficient and accurate computational prediction methods has become a research hotspot in the intersection of bioinformatics and artificial intelligence.

[0003] Currently, computation-based post-translational modification site prediction methods can be mainly categorized into three types: machine learning methods based on manual feature engineering, end-to-end sequence modeling methods based on deep learning, and multimodal methods that integrate pre-trained models with structural information. Although these methods have advanced the field to varying degrees, they still have several significant limitations, including: 1. In terms of multimodal information utilization, existing methods often suffer from crude fusion mechanisms and insufficient information integration. Most methods rely solely on sequence information, or simply splice sequences, pre-trained embeddings, and structural features, failing to deeply explore and adaptively fuse the complementary and consistent relationships between different modalities (sequence local semantics, pre-trained global semantics, and three-dimensional spatial structure), resulting in insufficient information utilization.

[0004] 2. In terms of contextual information modeling capabilities, existing methods generally face the dilemma of insufficient long-range dependency capture or low computational efficiency. Convolutional neural network-based methods have limited receptive fields, making it difficult to model long-range amino acid dependencies; recurrent neural network-based methods suffer from the gradient vanishing problem; and directly applying the Transformer self-attention mechanism faces the problems of secondary computational complexity and suppression of local high-frequency features of the sequence.

[0005] 3. Existing models generally lack robustness to common challenges in real-world bioinformatics scenarios. Real-world post-translation modified data often face problems such as incomplete annotation, severe class imbalance, and significant performance degradation due to differences in "environmental dynamics" (changes in the distribution of sequence-structure-function relationships) when models generalize across species and protein families. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for predicting protein post-translational modification sites in a multimodal cross-field perspective, which can improve the prediction accuracy in complex real biological data scenarios.

[0007] To address the aforementioned technical problems, this invention provides a multimodal, cross-field method for predicting protein post-translational modification sites, comprising: The protein peptides are obtained and sequence modality-based feature representations are constructed. The protein is broken down into individual amino acid units and the local structural information of each residue is captured. Based on the local structural information, a structural modality-based feature representation is constructed. An initial cyclic state feature is constructed by combining the sequence-based feature representation and the structure-based feature representation. The initial cyclic state feature is then iteratively fused using RIDCGA to obtain a fused feature. Based on the fused feature, protein post-translational modification sites are predicted.

[0008] Furthermore, the acquisition of protein peptides and the construction of sequence modality-based feature representations include: Let any peptide segment be denoted as In the formula, Let L represent the i-th amino acid and L represent the peptide length; in the supervised embedding layer, Mapping to a high-dimensional word embedding space yields word embedding vectors, denoted as... Simultaneously, the position embedding vector at position i is learned, denoted as... ,Will and Superposition as the initial amino acid sequence, denoted as ; A ProtRecBert model is constructed, which is a model based on ProtBert that incorporates regression loops to enhance feature iteration and purification, focusing on sequence conservation information. Each amino acid in the initial amino acid sequence is mapped to a label, and special labels are added to the beginning and end of the sequence to obtain a labeled sequence. The labeled sequence is input into the ProtRecBert model, and the encoder completes the first round of representation learning. The initial pre-trained representation is extracted from the last hidden layer, denoted as . ; then The encoder is re-inputted for secondary representation learning. After T iterations, the final converged iterative representation is taken as the model output, denoted as [model output]. ; The BLOSUM62 scoring matrix is ​​used to assign a score to any peptide segment, which is then converted into a vector representation, denoted as […]. ; Will The input is fed into a gated bidirectional long short-term memory network and a single-layer Transformer to deeply mine complex patterns and dependencies in the sequence, resulting in a dense representation, denoted as... ; A cross-field-of-view large-kernel attention module is constructed. This module serves as an attention mechanism for amino acid sequence tasks. While maintaining linear computational complexity, it achieves equivalent large receptive field coverage by decomposing the large-kernel convolution into depthwise convolution and depthwise dilated convolution. The decomposed depthwise convolution captures the relationships between adjacent amino acids within a local region by sliding small convolution kernels. The depthwise dilated convolution increases the receptive field by expanding the spacing between convolution kernels, capturing features of distant amino acids. Input the cross-view large kernel attention module and obtain the output of the cross-view large kernel attention module, denoted as R; The learned feature representation is obtained by learning R through DenseNet, denoted as . ,in, Represents the number of convolutional layers. Represents the first in a dense convolutional block The output of each convolutional layer; The feature representation based on sequence modes is obtained through average pooling layers and linear layers.

[0009] Furthermore, when decomposing the large kernel convolution into depthwise convolution and depthwise dilated convolution, the method for calculating the kernel size of the depthwise convolution and depthwise dilated convolution is as follows: , ; In the formula, This represents the kernel size of a depthwise convolution. The kernel size represents the depthwise dilated convolution, K represents the kernel size of the large kernel convolution, and d represents the dilation rate. Indicates rounding up; The output of the cross-field large kernel attention module is: , In the formula, R is the output of the cross-view large kernel attention module. Indicates that the convolution kernel is Convolution operation, Represents element-wise multiplication. For attention features, According to The obtained features and The calculation method is as follows: , ; In the formula, This represents the depthwise dilation convolution operation. This represents a depthwise convolution operation. This represents the activation function GELU. This represents a one-dimensional convolution operation.

[0010] Furthermore, the step of breaking down the protein into individual amino acid units and capturing the local structural information of each residue, and constructing a feature representation based on structural modality according to the local structural information, includes: The target amino acid unit is denoted as The main chain dihedral angle, solvent-accessible surface area, secondary structure, and residue type of the target amino acid unit are extracted as amino acid node features, and these amino acid node features are denoted as... ; A contact diagram is constructed based on the spatial distance between residues, and any two amino acid units are denoted as... and ,like and of If the Euclidean distance between atoms is less than a preset threshold, then it is considered... and There is spatial contact, and an edge is created in the graph, denoted as . ; Construct a graph, and denote the constructed graph as , ,Will As input to the GAT layer For the nodes in the graph and its neighboring nodes Calculate the unnormalized attention score, denoted as Use the softmax function to Normalization is performed to obtain the normalized attention score, denoted as... ; according to compute nodes The new feature representation, based on the node The new feature representation of the final amino acid granularity of the target amino acid unit is as follows: , In the formula, The final representation of the amino acid granularity of the target amino acid unit. Represents a node The new feature representation is i=1,2,…,N, where N represents the number of amino acid units; Will The input is a stacked autoencoder, which is used to learn a low-dimensional representation of the structural representation layer by layer, resulting in a feature representation based on the structural modality.

[0011] Furthermore, the unnormalized attention score is: , In the formula, Represents a node and its neighboring nodes Unnormalized attention score This represents the LeakyReLU activation function. and It is a learnable weight matrix. This indicates a splicing operation. It is Mapping to Simple networks in the same dimension; The node The new feature representation is calculated as follows: , In the formula, Represents a node The new feature representation, This represents the activation function GELU. Represents a node All neighboring nodes.

[0012] Furthermore, when using RIDCGA to iteratively fuse the initial cyclic state features, each iteration includes bidirectional guided attention and modal gating cyclic unit state propagation; in the t-th iteration, the feature representations based on sequence modality and feature representations based on structure modality extracted in the previous round are split using a learnable modal splitting matrix, and residual connections are added to avoid gradient vanishing. The dual-path bidirectional guided attention and modal gating loop unit state transfer includes a first path and a second path. The first path guides the purification of the feature representation based on the structure modality, while the second path guides the optimization of the feature representation based on the structure modality, with the purified feature representation based on the sequence modality as the main guide, forming a bidirectional closed loop.

[0013] Furthermore, the first path, guided primarily by the structural modality-based feature representation, leads to the purification of sequence modality-based feature representations, including: The query, key, and value constructed in the first path are: , ; In the formula, This represents the query constructed using the first path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality; This represents the key constructed by the first path in the t-th iteration. This represents the value constructed by the first path in the t-th iteration. This represents the feature representation based on sequence modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on sequence modality; The first path, after gated scaling dot product attention calculation, yields the attention features and output as follows: , ; In the formula, This represents the attention features of the first path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the first path in the t-th iteration. Let represent the learnable gating coefficients of the first path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate the feature components that contain only sequence modality information as the final sequence modality-based feature representation extracted in the t-th iteration.

[0014] Furthermore, the second path primarily guides the optimization of structure-based modality feature representations by using the purified sequence-based feature representations as the main driver, including: The query, key, and value constructed by the second path are: , ; In the formula, This represents the query constructed using the second path in the t-th iteration. This indicates a splicing operation. This represents the feature representation based on sequence mode in round t. This represents the global features learned by the ESM2 model in the t-th iteration. This represents the positional encoding vector corresponding to the feature representation based on sequence modality. This represents the key constructed by the second path in the t-th iteration. This represents the value constructed by the second path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality; The second path, after gated scaling dot product attention calculation, yields the attention features and output as follows: , ; In the formula, This represents the attention feature of the second path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the second path in the t-th iteration. Let represent the learnable gating coefficients of the second path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate the feature components that contain only structural modality information as the final structural modality-based feature representation extracted in the t-th iteration.

[0015] Furthermore, the calculation method for the fusion feature is as follows: The concatenated features are obtained by concatenating the sequence-based feature representation extracted in the t-th iteration and the structure-based feature representation extracted in the t-th iteration. After layer normalization, the concatenated features are input into the modality-gated recurrent unit to update the loop state. , In the formula, This represents the loop state in the t-th iteration. This represents a modal gated loop unit. This represents the splicing features after layer normalization; The last loop state is taken as the final loop state, and the features extracted from the final loop state by global adaptive pooling and fully connected layers are taken as the final fused features.

[0016] This invention also provides a multimodal, cross-field-of-view protein post-translational modification site prediction system, comprising: The data acquisition module is used to acquire protein peptides; The sequence feature construction module is used to construct feature representations based on sequence modes; The structural feature construction module is used to break down proteins into individual amino acid units and capture the local structural information of each residue, and construct a feature representation based on structural modality based on the local structural information; The feature fusion module is used to combine the feature representation based on sequence mode and the feature representation based on structure mode to construct an initial cyclic state feature, and to use RIDCGA to perform iterative fusion of the initial cyclic state feature to obtain the fused feature; The prediction module is used to predict protein post-translational modification sites based on the fusion features.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention constructs sequence-based and structure-based feature representations, extracting complementary feature representations from the protein sequence and three-dimensional spatial structure, respectively. Based on this, RIDCGA is used to iteratively fuse the initial loop state features and predict protein post-translational modification sites. This can deeply fuse multimodal information while maintaining linear computational complexity, efficiently model long-distance amino acid dependencies, enhance feature reuse and information flow, and improve the prediction accuracy of protein post-translational modification sites for imbalanced and incompletely labeled data. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0019] Figure 2 This is a process framework diagram of the method in a preferred embodiment of the present invention.

[0020] Figure 3 This is a framework diagram of the cross-field large kernel attention module in a preferred embodiment of the present invention.

[0021] Figure 4 This is a diagram illustrating the Transformer encoding process of the ProtRecBert model in a preferred embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] like Figure 1 and Figure 2 As shown, this invention discloses a method for predicting protein post-translational modification sites under multimodal cross-field perspectives, comprising the following steps: S1: Obtain protein peptides and construct sequence modality-based feature representations.

[0024] S1-1: Let any peptide segment be denoted as... In the formula, Let represent the i-th amino acid, and L represent the peptide length; each amino acid belongs to {X, A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y}, where 'X' is a pseudo-amino acid. In the supervised embedding layer, Mapping to a high-dimensional word embedding space yields word embedding vectors, denoted as... Simultaneously, the position embedding vector at position i is learned, denoted as... ,Will and Superposition as the initial amino acid sequence, denoted as ,Right now: , In the formula, This indicates an overlay operation.

[0025] S1-2: Construct the ProtRecBert model. The ProtRecBert model is based on ProtBert and incorporates regression loops to enhance feature iteration and purification, focusing on sequence conservation information. Specifically, each amino acid in the initial amino acid sequence is mapped to a label, and special labels (e.g., ...) are added to the beginning and end of the sequence. <cls>and <eos>The tokenized sequence is obtained; the tokenized sequence is input into the ProtRecBert model, and the encoder completes the first round of representation learning, extracting the initial pre-trained representation from the last hidden layer, denoted as . ; then The encoder is re-inputted for secondary representation learning. After T iterations, the final iterative representation after convergence is taken (denoted as...). Let the model output be denoted as . ,Right now: , In the formula, This represents the Transformer encoder of the ProtRecBert model. This indicates the regression correction module. This represents the current characterization and sequence conservation score.

[0026] S1-3: To reflect protein evolution information, the BLOSUM62 score matrix was used for any peptide segment ( The score is assigned and converted into a vector representation. ,Right now: , In the formula, This indicates the operation of the BLOSUM62 score matrix.

[0027] S1-4: Encode the three sequence features The input is fed into an innovatively designed gated bidirectional long short-term memory network (GBiLSTM) and a single-layer Transformer, such as... Figure 4 As shown, this method delves deeper into the complex patterns and dependencies within a sequence to obtain a dense representation, denoted as... ; The width is Channel size is GBiLSTM introduces a gated collaboration mechanism on top of the traditional BiLSTM. Through the coordinated control of the forget and update gates, it dynamically filters redundant sequence information, enhancing the bidirectional transmission of effective features. Simultaneously, it incorporates cross-timestep residual connections to alleviate gradient decay issues during long sequence training. A single-layer Transformer further captures global contextual information, dynamically focusing on important information in the sequence and transforming the feature vector into a dense representation. ,Right now: , In the formula, This indicates a splicing operation. This represents a single-layer Transformer operation. This represents a gated bidirectional long short-term memory network.

[0028] S1-5: Construct a cross-view large kernel attention module (1D-LKA), such as Figure 3 As shown, the cross-field large kernel attention module is an attention mechanism for amino acid sequence tasks. While maintaining linear computational complexity, it achieves equivalent large receptive field coverage by decomposing the large kernel convolution into depthwise convolution and depthwise dilated convolution. This decomposition strategy avoids the efficiency bottleneck of self-attention and traditional large kernel convolution in long sequence tasks, and is particularly suitable for ultra-long sequences. Compared with existing 1D attention mechanisms, 1D-LKA achieves a better balance between long-range dependency modeling and computational efficiency, making it a superior choice. The decomposed depthwise convolution captures the relationship between adjacent amino acids in local regions by sliding small convolution kernels; the depthwise dilated convolution increases the receptive field by expanding the spacing between convolution kernels, capturing features of distant amino acids. The decomposed large kernel attention can efficiently model long-range dependencies on long sequences while maintaining low computational overhead.

[0029] When decomposing a large kernel convolution into depthwise convolution and depthwise dilated convolution, the kernel size of depthwise convolution and depthwise dilated convolution is calculated as follows: , ; In the formula, This represents the kernel size of a depthwise convolution. The kernel size represents the depthwise dilated convolution, K represents the kernel size of the large kernel convolution, and d represents the dilation rate. This indicates rounding up to the nearest integer.

[0030] Will Input the cross-view large kernel attention module and obtain its output, denoted as R. The output of the cross-view large kernel attention module is: , In the formula, R is the output of the cross-view large kernel attention module. Indicates that the convolution kernel is Convolution operation, Represents element-wise multiplication. For attention features, According to The obtained features and The calculation method is as follows: , ; In the formula, This represents the depthwise dilation convolution operation. This represents a depthwise convolution operation. This represents the activation function GELU. This represents a one-dimensional convolution operation.

[0031] S1-6: Learn the learned feature representation in R using DenseNet (Densely Connected Convolutional Networks), denoted as... ,in, Represents the number of convolutional layers. Represents the first in a dense convolutional block The output of each convolutional layer; DenseNet ensures that features from both shallow and deep layers are effectively transferred and fused through a feature reuse mechanism, enhancing gradient flow, improving the model's ability to learn complex sequence patterns, and mitigating training instability that may be caused by data imbalance or annotation noise. , This is the weight matrix.

[0032] S1-7: A high-level representation of the sequence, namely a feature representation based on sequence modes, is obtained through an average pooling layer and a linear layer. , In the formula, This represents a feature representation based on sequence modes. This represents the activation function GELU. These are weighting coefficients, in this embodiment... 0.2, This represents the average pooling layer.

[0033] S2: The protein is broken down into individual amino acid units and the local structural information of each residue is captured. Based on the local structural information, a feature representation based on structural modes is constructed.

[0034] S2-1: The target amino acid unit is denoted as... Extract the main chain dihedral angle of the target amino acid unit ( horn, Amino acid node features are defined by their angle, solvent accessible surface area (ASA), secondary structure (SS), and residue type. These amino acid node features are denoted as... .

[0035] S2-2: Construct a contact diagram based on the spatial distance between residues, denoting any two amino acid units as... and ,like and of The Euclidean distance between atoms is less than a preset threshold (in this embodiment, the preset threshold is set to...). ), then it is believed and There is spatial contact, and an edge is created in the graph, denoted as . ,Right now , express and of The Euclidean distance between atoms.

[0036] S2-3: Construct the graph, and denote the constructed graph as... , ,Will As input to the GAT (Graph Attention Layer), the GAT layer uses an attention mechanism to assign different importance weights to the neighbors of each node, thereby aggregating information.

[0037] S2-4: For the nodes in the graph and its neighboring nodes The unnormalized attention score is calculated as follows: , In the formula, Represents a node and its neighboring nodes Unnormalized attention score This represents the LeakyReLU activation function. and It is a learnable weight matrix. This indicates a splicing operation. It is Mapping to Simple networks in the same dimension.

[0038] S2-5: Using the softmax function to... Normalization is performed to obtain the normalized attention score, denoted as... .

[0039] S2-6: According to compute nodes The new feature is represented as: , In the formula, Represents a node The new feature representation, This represents the activation function GELU. Represents a node All neighboring nodes.

[0040] S2-7: Based on nodes The new feature representation of the final amino acid granularity of the target amino acid unit is as follows: , In the formula, The final representation of the amino acid granularity of the target amino acid unit. Represents a node The new feature is represented as i=1,2,…,N, where N represents the number of amino acid units.

[0041] S2-8: In order to extract more effective features, The input is a stacked autoencoder (SAE), which learns low-dimensional representations of the structure layer by layer. The stacked autoencoder uses unsupervised learning to progressively map each granularity of structural representation to a low-dimensional latent space. In this process, the SAE model encodes the input data and learns more informative high-level features through layer-by-layer training. Each layer of the autoencoder obtains a more concise and information-rich low-dimensional representation through compression and reconstruction. This low-dimensional representation not only effectively captures the core information of the protein structure but also enhances the model's ability to represent complex structural features. These three granularity high-level representations are then stacked and fused and optimized through the last linear layer of the module to obtain the final high-order structural feature vector of the protein structure, i.e., the feature representation based on structural modality, denoted as . , can be represented as: , In the formula, This indicates a stacked autoencoder.

[0042] S3: Combine the feature representations based on sequence modality and feature representations based on structure modality to construct initial cyclic state features, and use RIDCGA to iteratively fuse the initial cyclic state features to obtain fused features; RIDCGA is a residual graph attention convolutional module (Residual-Interaction-Dual-Channel Graph Attention) for protein structure and interaction modeling. Its core is to add residual connections, interaction modeling, and dual-channel attention on the basis of graph attention (GAT) to solve the problems of long-range dependence and local geometry modeling of protein residue graphs.

[0043] S3-1: The initial cyclic state features are obtained by concatenating the two aligned features and mapping them through a lightweight fully connected layer, specifically: , In the formula, Indicates the characteristics of the initial loop state. This indicates a fully connected layer.

[0044] S3-2: Use RIDCGA to iteratively fuse the initial cyclic state features to obtain fused features.

[0045] When using RIDCGA to iteratively fuse the initial recurrent state features, each iteration includes bidirectional guided attention along a dual-path path and state propagation of the Modal Gated Recurrent Unit (MGRU). In the t-th iteration, the feature representation based on sequence modality extracted in the previous iteration (i.e., the feature representation based on sequence modality in the (t-1)-th iteration, denoted as...) is extracted using a learnable modality splitting matrix. ) and feature representation based on structure mode (i.e., the feature representation based on structure mode in round t-1, denoted as The method adds residual connections to avoid gradient vanishing, where t∈[1,T'] and T' represents the maximum number of iterations. The bidirectional guided attention and modal gated recurrent unit (MGRU) state transfer includes a first path and a second path. The first path guides the purification of the sequence modality-based feature representation with the structure modality-based feature representation as the dominant guide, while the second path guides the optimization of the structure modality-based feature representation with the purified sequence modality-based feature representation as the dominant guide, forming a bidirectional closed loop.

[0046] The first path, guided by the structural modality-based feature representation, leads to the purification of sequence modality-based feature representations, including: The query, key, and value constructed in the first path are: , ; In the formula, This represents the query constructed using the first path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality. Used to perceive the order and spatial relationships of structures; This represents the key constructed by the first path in the t-th iteration. This represents the value constructed by the first path in the t-th iteration. This represents the feature representation based on sequence modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on sequence modality. Used to capture the order dependency of a sequence;

[0047] The first path, after gated scaling dot product attention calculation, yields the attention features and output as follows: , ; In the formula, This represents the attention features of the first path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the first path in the t-th iteration. Let represent the learnable gating coefficients of the first path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate the feature components that contain only sequence mode information, which are then used as the final sequence mode-based feature representation extracted in the t-th iteration. .

[0048] The second approach, guided by the refined sequence-based feature representation, optimizes the structure-based feature representation, including: The query, key, and value constructed by the second path are: , ; In the formula, This represents the query constructed using the second path in the t-th iteration. This indicates a splicing operation. This represents the feature representation based on sequence mode in round t. This represents the global features learned by the ESM2 model in the t-th iteration (i.e., the aggregate vector of all valid residue cycle states in the t-th layer). This represents the positional encoding vector corresponding to the feature representation based on sequence modality. This represents the key constructed by the second path in the t-th iteration. This represents the value constructed by the second path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality; The second path, after gated scaling dot product attention calculation, yields the attention features and output as follows: , ; In the formula, This represents the attention feature of the second path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the second path in the t-th iteration. Let represent the learnable gating coefficients of the second path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate feature components containing only structural modality information, which are then used as the final structural modality-based feature representation extracted in the t-th iteration. .

[0049] The calculation method for fused features is as follows: The concatenated feature is obtained by splicing the sequence mode-based feature representation extracted in the t-th iteration and the structure mode-based feature representation extracted in the t-th iteration. ,Right now The spliced ​​features are normalized by layers and then input into the modal gating recurrent unit to update the recurrent state. , In the formula, This represents the loop state in the t-th iteration. This represents a modal gated loop unit. This represents the spliced ​​features after layer normalization; MGRU prioritizes retaining valid information by adding modal weights to the update and reset gates.

[0050] The state of the last cycle is taken as the final cycle state (i.e., the cycle state of the T'th cycle, denoted as...). The features extracted from the final recurrent state through global adaptive pooling and fully connected layers are used as the final fused features, i.e.: , In the formula, Indicates fusion characteristics, This indicates global adaptive pooling. This indicates a fully connected layer.

[0051] S4: Predict post-translational modification sites of proteins based on the fusion features.

[0052] The fused features are input into the MLP decoder, and the probability that the target residue is a post-translational modification site is output through the sigmoid function, thus achieving protein post-translational modification site prediction. The probability that the target residue is a post-translational modification site is: , In the formula, This indicates the probability that the target residue is a post-translational modification site. This represents the Sigmoid function. This indicates an MLP decoder.

[0053] In this embodiment, the MLP decoder is trained before prediction, and the loss function during training is: , In the formula, Represents the loss function. Represents the total number of protein samples. Represents the true label value of the sample. This represents the predicted probability of the i-th sample.

[0054] This invention also discloses a multimodal, cross-field-of-view protein post-translational modification site prediction system, comprising: The data acquisition module is used to acquire protein peptides; The sequence feature construction module is used to construct feature representations based on sequence modes; The structural feature construction module is used to break down proteins into individual amino acid units and capture the local structural information of each residue, and construct a feature representation based on structural modality based on the local structural information; The feature fusion module is used to combine the feature representation based on sequence mode and the feature representation based on structure mode to construct an initial cyclic state feature, and to use RIDCGA to perform iterative fusion of the initial cyclic state feature to obtain the fused feature; The prediction module is used to predict protein post-translational modification sites based on the fusion features.

[0055] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for predicting protein post-translational modification sites in a multimodal, cross-field perspective.

[0056] The present invention also discloses a multimodal cross-field protein post-translational modification site prediction device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a multimodal cross-field protein post-translational modification site prediction method.

[0057] This invention constructs sequence-based and structure-based feature representations, extracting complementary feature representations from the protein sequence and three-dimensional spatial structure, respectively. Based on this, RIDCGA is used to iteratively fuse the initial loop state features and predict protein post-translational modification sites. This can deeply fuse multimodal information while maintaining linear computational complexity, efficiently model long-distance amino acid dependencies, enhance feature reuse and information flow, and improve the prediction accuracy of protein post-translational modification sites for imbalanced and incompletely labeled data.

[0058] This invention designs a multimodal adaptive fusion mechanism and an efficient cross-field big kernel attention module. Based on the deep fusion of heterogeneous information, the cross-field attention mechanism in the cross-field big kernel attention module accurately captures long-range contextual dependencies, which can improve the prediction accuracy, stability and generalization ability of the model in complex real biological data scenarios, thus providing an effective solution for more accurate and robust post-translational modification site calculation and prediction.

[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.< / eos> < / cls>

Claims

1. A method for predicting protein post-translational modification sites under multimodal cross-field perspectives, characterized in that, include: The protein peptides are obtained and sequence modality-based feature representations are constructed. The protein is broken down into individual amino acid units and the local structural information of each residue is captured. Based on the local structural information, a structural modality-based feature representation is constructed. An initial cyclic state feature is constructed by combining the sequence-based feature representation and the structure-based feature representation. The initial cyclic state feature is then iteratively fused using RIDCGA to obtain a fused feature. Based on the fused feature, protein post-translational modification sites are predicted.

2. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 1, characterized in that: The acquisition of protein peptides and the construction of sequence modality-based feature representations include: Let any peptide segment be denoted as In the formula, Let L represent the i-th amino acid and L represent the peptide length; in the supervised embedding layer, Mapping to a high-dimensional word embedding space yields word embedding vectors, denoted as... Simultaneously, the position embedding vector at position i is learned, denoted as... ,Will and Superposition as the initial amino acid sequence, denoted as ; A ProtRecBert model is constructed, which is a model based on ProtBert that incorporates regression loops to enhance feature iteration and purification, focusing on sequence conservation information. Each amino acid in the initial amino acid sequence is mapped to a label, and special labels are added to the beginning and end of the sequence to obtain a labeled sequence. The labeled sequence is input into the ProtRecBert model, and the encoder completes the first round of representation learning. The initial pre-trained representation is extracted from the last hidden layer, denoted as . ; then The encoder is re-inputted for secondary representation learning. After T iterations, the final converged iterative representation is taken as the model output, denoted as [model output]. ; The BLOSUM62 scoring matrix is ​​used to assign a score to any peptide segment, which is then converted into a vector representation, denoted as […]. ; Will The input is fed into a gated bidirectional long short-term memory network and a single-layer Transformer to deeply mine complex patterns and dependencies in the sequence, resulting in a dense representation, denoted as... ; A cross-field-of-view large-kernel attention module is constructed. This module serves as an attention mechanism for amino acid sequence tasks. While maintaining linear computational complexity, it achieves equivalent large receptive field coverage by decomposing the large-kernel convolution into depthwise convolution and depthwise dilated convolution. The decomposed depthwise convolution captures the relationships between adjacent amino acids within a local region by sliding small convolution kernels. The depthwise dilated convolution increases the receptive field by expanding the spacing between convolution kernels, capturing features of distant amino acids. Input the cross-view large kernel attention module and obtain the output of the cross-view large kernel attention module, denoted as R; The learned feature representation is obtained by learning R through DenseNet, denoted as . ,in, Represents the number of convolutional layers. Represents the first in a dense convolutional block The output of each convolutional layer; The feature representation based on sequence modes is obtained through average pooling layers and linear layers.

3. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 2, characterized in that: When decomposing large-kernel convolution into depthwise convolution and depthwise dilated convolution, the method for calculating the kernel size of depthwise convolution and depthwise dilated convolution is as follows: ， ； In the formula, This represents the kernel size of a depthwise convolution. The kernel size represents the depthwise dilated convolution, K represents the kernel size of the large kernel convolution, and d represents the dilation rate. Indicates rounding up; The output of the cross-field large kernel attention module is: ， In the formula, R is the output of the cross-view large kernel attention module. Indicates that the convolution kernel is Convolution operation, Represents element-wise multiplication. For attention features, According to The obtained features and The calculation method is as follows: ， ； In the formula, This represents the depthwise dilation convolution operation. This represents a depthwise convolution operation. This represents the activation function GELU. This represents a one-dimensional convolution operation.

4. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 1, characterized in that: The process of breaking down proteins into individual amino acid units and capturing the local structural information of each residue, and constructing a feature representation based on structural modalities based on the local structural information, includes: The target amino acid unit is denoted as The main chain dihedral angle, solvent-accessible surface area, secondary structure, and residue type of the target amino acid unit are extracted as amino acid node features, and these amino acid node features are denoted as... ; A contact diagram is constructed based on the spatial distance between residues, and any two amino acid units are denoted as... and ,like and of If the Euclidean distance between atoms is less than a preset threshold, then it is considered... and There is spatial contact, and an edge is created in the graph, denoted as . ; Construct a graph, and denote the constructed graph as , ,Will As input to the GAT layer For the nodes in the graph and its neighboring nodes Calculate the unnormalized attention score, denoted as Use the softmax function to... Normalization is performed to obtain the normalized attention score, denoted as... ; according to compute nodes The new feature representation, based on the node The new feature representation of the final amino acid granularity of the target amino acid unit is as follows: ， In the formula, The final representation of the amino acid granularity of the target amino acid unit. Represents a node The new feature representation is i=1,2,…,N, where N represents the number of amino acid units; Will The input is a stacked autoencoder, which is used to learn a low-dimensional representation of the structural representation layer by layer, resulting in a feature representation based on the structural modality.

5. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 4, characterized in that: The unnormalized attention score is: ， In the formula, Represents a node and its neighboring nodes Unnormalized attention score This represents the LeakyReLU activation function. and It is a learnable weight matrix. This indicates a splicing operation. It is Mapping to Simple networks in the same dimension; The node The new feature representation is calculated as follows: ， In the formula, Represents a node The new feature representation, This represents the activation function GELU. Represents a node All neighboring nodes.

6. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 1, characterized in that: When using RIDCGA to iteratively fuse the initial cyclic state features, each iteration includes bidirectional guided attention and modal gating cyclic unit state propagation. In the t-th iteration, the feature representations based on sequence modality and feature representations based on structure modality extracted in the previous round are split using a learnable modality splitting matrix, and residual connections are added to avoid gradient vanishing. The dual-path bidirectional guided attention and modal gating loop unit state transfer includes a first path and a second path. The first path guides the purification of the feature representation based on the structure modality, while the second path guides the optimization of the feature representation based on the structure modality, with the purified feature representation based on the sequence modality as the main guide, forming a bidirectional closed loop.

7. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 6, characterized in that: The first path, guided by the structural modality-based feature representation, leads to the purification of sequence modality-based feature representations, including: The query, key, and value constructed in the first path are: ， ； In the formula, This represents the query constructed using the first path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality; This represents the key constructed by the first path in the t-th iteration. This represents the value constructed by the first path in the t-th iteration. This represents the feature representation based on sequence modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on sequence modality; The first path, after gated scaling dot product attention calculation, yields the attention features and output as follows: ， ； In the formula, This represents the attention features of the first path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the first path in the t-th iteration. Let represent the learnable gating coefficients of the first path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate the feature components that contain only sequence modality information as the final sequence modality-based feature representation extracted in the t-th iteration.

8. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 7, characterized in that: The second path, guided by the purified sequence-based feature representation, optimizes the structure-based feature representation, including: The query, key, and value constructed by the second path are: ， ； In the formula, This represents the query constructed using the second path in the t-th iteration. This indicates a splicing operation. This represents the feature representation based on sequence mode in round t. This represents the global features learned by the ESM2 model in the t-th iteration. This represents the positional encoding vector corresponding to the feature representation based on sequence modality. This represents the key constructed by the second path in the t-th iteration. This represents the value constructed by the second path in the t-th iteration. This represents the feature representation based on structural modes in the (t-1)th round. This represents the positional encoding vector corresponding to the feature representation based on structural modality; The second path, after gated scaling dot product attention calculation, yields the attention features and output as follows: ， ； In the formula, This represents the attention feature of the second path in the t-th iteration. This represents the softmax function, and T represents the transpose operation. Represents the vector dimension. This represents the output of the second path in the t-th iteration. Let represent the learnable gating coefficients of the second path in the t-th iteration; Learnable modality splitting matrix pairs Implicit feature decomposition is performed to separate the feature components that contain only structural modality information as the final structural modality-based feature representation extracted in the t-th iteration.

9. The method for predicting protein post-translational modification sites under multimodal cross-field perspectives according to claim 8, characterized in that: The calculation method for the fusion feature is as follows: The concatenated features are obtained by concatenating the sequence-based feature representation extracted in the t-th iteration and the structure-based feature representation extracted in the t-th iteration. After layer normalization, the concatenated features are input into the modality-gated recurrent unit to update the loop state. ， In the formula, This represents the loop state in the t-th iteration. This represents a modal gated loop unit. This represents the splicing features after layer normalization; The last loop state is taken as the final loop state, and the features extracted from the final loop state by global adaptive pooling and fully connected layers are taken as the final fused features.

10. A multimodal, cross-field-of-view protein post-translational modification site prediction system, characterized in that, include: The data acquisition module is used to acquire protein peptides; The sequence feature construction module is used to construct feature representations based on sequence modes; The structural feature construction module is used to break down proteins into individual amino acid units and capture the local structural information of each residue, and construct a feature representation based on structural modality based on the local structural information; The feature fusion module is used to combine the feature representation based on sequence mode and the feature representation based on structure mode to construct an initial cyclic state feature, and to use RIDCGA to perform iterative fusion of the initial cyclic state feature to obtain the fused feature; The prediction module is used to predict protein post-translational modification sites based on the fusion features.

Images