Post-translational modification site prediction method and system based on multi-modal fusion

Abstract

The application discloses a protein post-translational modification site prediction method and system based on multi-modal fusion, relates to the technical field of bioinformatics and artificial intelligence, and comprises the following steps: obtaining sequence data and structure coordinate data of a protein, performing double-flow feature extraction on the sequence data and the structure coordinate data of the protein to obtain a sequence feature vector and a structure feature vector, inputting the sequence feature vector and the structure feature vector into a pre-established leading flow fusion network model for feature enhancement to obtain an enhanced sequence feature, inputting the structure feature vector and a preset graph structure representation into a pre-established symmetric hierarchical flow graph network model, extracting local structure features and global structure features through isometric message passing and hierarchical structure modeling, and outputting multi-scale structure features; inputting the enhanced sequence feature and the multi-scale structure features into a pre-established KronHTR module for deep feature fusion, outputting multi-modal fusion features, predicting based on the multi-modal fusion features, and obtaining protein post-translational modification site prediction results, so that the accuracy, stability and generalization ability of post-translational modification site prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, specifically to a method and system for predicting protein post-translational modification sites based on multimodal fusion. Background Technology

[0002] Post-translational modifications (PTMs) are important regulatory mechanisms resulting from chemical modifications of proteins after translation, and they are widely involved in biological processes such as cell signal transduction, protein-protein interactions, metabolic regulation, and disease development. Common types of PTMs include phosphorylation, methylation, glycosylation, ubiquitination, and SUMOylation. Because PTMs can directly alter a protein's spatial conformation, active state, and binding ability to other molecules, accurate identification of modification sites is of great value in protein function research.

[0003] Existing post-translational modification site prediction methods mostly employ PTM site prediction models. These models are largely constructed based on protein sequence information and typically use traditional machine learning or deep learning methods to encode features of amino acid residues. Therefore, the single sequence features used in existing methods are difficult to reflect the true three-dimensional conformation of proteins and the spatial relationships between residues. There is a lack of effective methods to integrate protein sequence and structural information and to perform efficient learning while ensuring geometric symmetry, resulting in low prediction accuracy. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for predicting protein post-translational modification sites based on multimodal fusion, thereby solving the problems in the prior art where relying solely on sequence information is insufficient to reflect the true spatial conformation of proteins, there is a lack of unified characterization among different modification types, and PTM prediction performance is low with few samples.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a method for predicting protein post-translational modification sites based on multimodal fusion, the method comprising the following steps: The sequence data and structural coordinate data of the protein to be predicted are obtained. The sequence data and structural coordinate data of the protein are then subjected to dual-stream feature extraction to obtain sequence feature vectors and structural feature vectors. The dual-stream feature extraction includes sequence feature extraction based on the protein sequence and structural feature extraction based on the protein's three-dimensional structure. Sequence feature vectors and structural feature vectors are input into a pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses structural feature vectors as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vectors to obtain enhanced sequence features. The structural feature vectors and a preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. Local and global structural features are extracted through equivariant message passing and hierarchical structure modeling, and multi-scale structural features are output. The enhanced sequence features and multi-scale structural features are input into the pre-established KronHTR module for deep feature fusion, and the output is a multimodal fusion feature. Based on the multimodal fusion feature, prediction is performed to obtain the prediction results of protein post-translational modification sites.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing dual-stream feature extraction on the protein sequence data and structural coordinate data, as follows: For the sequence feature extraction process of dual-stream feature extraction, evolutionary and semantic features of protein sequence data are extracted based on the pre-trained protein language model ESM-C, and used as sequence feature vectors. For the structural feature extraction process of dual-stream feature extraction, the geometric and spatial features of the protein's structural coordinate data are extracted based on the structure-aware model SaProt, and used as structural feature vectors.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the sequence feature vector and the structural feature vector into a pre-established dominant flow fusion network model for feature enhancement, comprising: The offset and modulation parameters are obtained by modulating and extracting the structural feature vectors; By using structure-guided deformable convolution to achieve cross-modal geometric alignment of the sequence feature vectors with offsets and modulation parameters, enhanced sequence features are generated, thereby completing semantic-level content enhancement.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: inputting the structural feature vector and the preset graph structure representation into a pre-established symmetric hierarchical flow graph network model, the processing procedure is as follows: Based on structural feature vectors and predefined graph structure representations, this method uses the AttentiveEMAMPConv operator and hierarchical differentiable pooling mechanism to perform equivariant message passing, update node features and coordinates, and perform hierarchical coarsening of the graph through differentiable pooling to generate multi-scale structural features. This enables the simultaneous capture of local details and global topology of protein structures.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the preset graph structure is represented based on the construction of residues as nodes and spatial distance and sequence proximity as edges.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the enhanced sequence features and multi-scale structural features into a pre-established KronHTR module for deep feature fusion, comprising: By combining Gromov-Wasserstein optimal transport theory with dynamically generated Kronecker transform operators, we first achieve soft alignment of the internal geometric structures of the enhanced sequence features and multi-scale structural features at the non-parametric level. Then, we generate instance-specific structured transformations through conditional hypernetworks to iteratively refine the sequence features and achieve deep adaptive heterogeneous feature fusion.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the KronHTR module first calculates the Gromov-Wasserstein optimal transmission alignment matrix, then constructs a supernetwork conditioned on structural features, dynamically generates Kronecker factor pairs, and finally performs deep feature fusion through iterative refining blocks to obtain multimodal fused features.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the enhanced sequence features capturing long-range context dependencies through a Mamba module based on a selective state-space model.

[0013] Secondly, in order to achieve the above objectives, this invention discloses a protein post-translational modification site prediction system based on multimodal fusion, comprising: A primary processing module is used to acquire the sequence data and structural coordinate data of the protein to be predicted, and to perform dual-stream feature extraction on the sequence data and structural coordinate data of the protein to obtain sequence feature vectors and structural feature vectors. The dual-stream feature extraction includes sequence feature extraction based on the protein sequence and structural feature extraction based on the three-dimensional structure of the protein. The secondary processing module is used to input the sequence feature vector and the structural feature vector into a pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses the structural feature vector as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vector to obtain enhanced sequence features. The structural feature vector and the preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. Local structural features and global structural features are extracted through equivariant message passing and hierarchical structure modeling, and multi-scale structural features are output. The prediction output module is used to input the enhanced sequence features and multi-scale structural features into the pre-established KronHTR module for deep feature fusion, and output multi-modal fusion features. Based on the multi-modal fusion features, prediction is performed to obtain the prediction results of protein post-translational modification sites.

[0014] The beneficial effects of this invention are: This invention constructs a dual-stream collaborative multimodal modeling architecture to achieve efficient alignment and deep fusion of sequence and structural features while ensuring geometric symmetry. At the same time, it combines a generative evolutionary data augmentation strategy to effectively alleviate the problems of data scarcity and class imbalance for few-sample post-translational modification types, which is conducive to improving the accuracy, stability and generalization ability of post-translational modification site prediction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the structure of the dual-stream feature extraction model of the present invention; Figure 3 This is a schematic diagram comparing the MCC performance of different models on various PTM types within the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the AUC performance of UniProNet with existing models under different PTM types in the embodiments of the present invention; Figure 5 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1 As shown, a method for predicting protein post-translational modification sites based on multimodal fusion includes the following steps: S101: Obtain the sequence data and structural coordinate data of the protein to be predicted, and perform dual-stream feature extraction on the sequence data and structural coordinate data of the protein to obtain sequence feature vector and structural feature vector. The dual-stream feature extraction includes sequence feature extraction based on protein sequence and structural feature extraction based on protein three-dimensional structure. The process of performing two-stream feature extraction on protein sequence data and structural coordinate data is as follows: For the sequence feature extraction process of dual-stream feature extraction, evolutionary and semantic features of protein sequence data are extracted based on the pre-trained protein language model ESM-C, and used as sequence feature vectors. For the structural feature extraction process of dual-stream feature extraction, the geometric and spatial features of the protein's structural coordinate data are extracted based on the structure-aware model SaProt, and used as structural feature vectors.

[0018] Specifically, in this embodiment, to address the challenge of scarce PTM-type data, the present invention integrates the PTM-AGEN data augmentation strategy. This strategy combines the sequence generation capabilities of the protein language model with a multi-objective evolutionary algorithm. Under strict adherence to biological evolutionary constraints and structural feasibility assessments, it synthesizes high-fidelity training samples and actively mines difficult negative samples, effectively expanding and balancing the training set. At the feature representation level, the present invention simultaneously utilizes the ESM-C universal protein language model to extract sequence embeddings containing deep evolutionary information, and the SaProt structure-aware protein language model to extract structural embeddings that explicitly encode local three-dimensional geometry, providing rich and complementary input representations for subsequent fusion.

[0019] The construction and training process of the pre-trained protein language model ESM-C and the structure-aware model SaProt includes: The sequence feature model construction and training process based on the pre-trained protein language model ESM-C includes: First, the protein sequence is preprocessed, candidate site sequence fragments are constructed with candidate modification sites as the center, and input into the pre-trained ESM-C model. The sequence embedding representation corresponding to each residue position is extracted as the sequence feature input of the UniProNet model.

[0020] In the model construction process, optional basic network structures include multilayer perceptrons, convolutional neural networks, recurrent neural networks, Transformer coding structures, and Mamba long sequence modeling structures. Considering that protein post-translational modification site prediction in this application requires the simultaneous utilization of local sequence patterns and long-range contextual dependencies, the final selected basic structure uses ESM-C embeddings as input and combines DSFN and Mamba modules for sequence feature enhancement. Specifically, the DSFN module is used for structure-guided feature correction and semantic modulation of ESM-C sequence features, while the Mamba module is used to capture long-range dependencies in protein sequences.

[0021] During model training, the input data includes candidate site sequence fragments and their sequence embedding representations extracted by ESM-C, and the output data is the predicted labels corresponding to the candidate modification sites. The AdamW optimizer is used during training, with a learning rate set to 1×10⁻⁶. -4 When fine-tuning the last few layers of the ESM-C encoder, the fine-tuning learning rate is set to 1×10. -5 The batch size was set to 32, the number of training epochs to 100, the Dropout ratio to 0.3, the hidden layer dimension to 512, and the activation function to GELU. The cross-entropy loss function was used, combined with a hierarchical contrastive loss function for joint optimization.

[0022] Construction and training process of structure-aware model based on SaProt The process of building and training a structural feature model based on the structure-aware model SaProt includes: First, the protein sequence fragments corresponding to candidate modification sites and their three-dimensional structural information are obtained. The three-dimensional structural information comes from experimentally resolved structures or protein structure prediction models. Then, the amino acid sequence markers and structural markers are combined to obtain a structure-aware input sequence, which is then input into a pre-trained SaProt model. The structure-aware embedding representation corresponding to each residue position is extracted and used as the structural feature input of the UniProNet model.

[0023] In the model construction process, the optional basic network structures include multilayer perceptron, convolutional neural network, graph neural network, equivariant graph neural network, and structure-aware attention network. Considering the characteristics of modification site prediction in this application, which are related to the spatial neighborhood, local conformation, and multi-scale structural relationships of residues, the basic structure chosen is SaProt structure-aware embedding as input, combined with the SHFGNN module for structural feature modeling. Among them, the SHFGNN module is built based on E(3) equivariant graph neural network, which is used to model the local structural features and multi-scale spatial relationships in the protein structure map, and output shallow structural features and deep structural features.

[0024] During model training, the input data includes candidate site sequence fragments, the corresponding structure-aware input sequences, and the structure-aware embedding representations extracted by SaProt. The output data consists of the predicted labels corresponding to the candidate modification sites. The AdamW optimizer is used during training, with a learning rate of 1×10⁻⁶. -4 When fine-tuning the last few layers of the SaProt encoder, the fine-tuning learning rate is set to 1×10. -5 The batch size was set to 32, the number of training epochs to 100, the Dropout ratio to 0.3, the hidden layer dimension to 512, and the activation function to GELU. The cross-entropy loss function was used, combined with a hierarchical contrastive loss function for joint optimization, where the weight parameter β in the hierarchical contrastive loss was set to 0.72.

[0025] During the joint training phase, ESM-C sequence features, enhanced by DSFN and Mamba modules, are fused with shallow and deep structural features obtained from SaProt structural features via SHFGNN. The shallow structural features participate in the intermediate fusion of DSFN, while the deep structural features are fused with the sequence features in the later stage via the KronHTR module. Finally, the fused comprehensive features are input into a multilayer perceptron, outputting the final prediction results for candidate modification sites. The total training loss function consists of a classification loss function and a hierarchical contrastive loss function.

[0026] S102: Input the sequence feature vector and structural feature vector into the pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses the structural feature vector as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vector to obtain enhanced sequence features. Input the structural feature vector and the preset graph structure representation into the pre-established symmetric hierarchical flow graph network model. Extract local structural features and global structural features through equivariant message passing and hierarchical structure modeling, and output multi-scale structural features. The process of inputting sequence feature vectors and structural feature vectors into a pre-established dominant flow fusion network model for feature enhancement includes: The offset and modulation parameters are obtained by modulating and extracting the structural feature vectors; By using structure-guided deformable convolution to achieve cross-modal geometric alignment of the sequence feature vectors with offsets and modulation parameters, enhanced sequence features are generated, thereby completing semantic-level content enhancement.

[0027] The enhanced sequence features capture long-range contextual dependencies using a Mamba module based on a selective state-space model.

[0028] The pre-established dominant flow fusion network model DSFN is implemented as follows: Taking sequence feature vectors and structural feature vectors as input, LoRA modulation units are first used to generate offset ΔP and FiLM parameters (γ, β) from the structural features; then, structure-guided deformable convolution (SGDC) is applied to spatially align the sequence features; finally, channel-level affine transformation is performed through the FiLM layer to output the enhanced sequence features. .

[0029] Specifically, in one embodiment, to accurately capture the long-range sequence context information affecting PTM sites, this invention introduces a Mamba module based on a selective state-space model. This module can process full-length protein sequences with linear computational complexity, efficiently modeling long-range dependencies and overcoming the computational bottleneck of traditional Transformer models in long sequence processing. Furthermore, this invention designs a hierarchical contrastive loss function. This function, by constructing contrastive learning tasks at different levels of the model, forces multimodal features to remain compact within layers and semantically coherent between layers, which helps improve the consistency and discriminativeness of the learned feature representations, further optimizing the overall performance and generalization ability of the model. The structural feature vectors and the preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. The processing procedure is as follows: Based on structural feature vectors and predefined graph structure representations, this method uses the AttentiveEMAMPConv operator and hierarchical differentiable pooling mechanism to perform equivariant message passing, update node features and coordinates, and perform hierarchical coarsening of the graph through differentiable pooling to generate multi-scale structural features. This enables the simultaneous capture of local details and global topology of protein structures.

[0030] The preset graph structure is constructed based on residues as nodes and spatial distance and sequence proximity as edges.

[0031] Specifically, the implementation process of the Symmetric Hierarchical Flow Graph Network (SHFGNN) is as follows: A graph structure is constructed using residues as nodes and spatial distance and sequence proximity as edges. First, AttentiveEMAMPConv is used for equivariant message passing to update node features and coordinates. Then, differentiable pooling is used to hierarchically coarsen the graph, generating a multi-scale structural representation. Finally, shallow features are output. (Local details) and deep features (Global context).

[0032] The construction and training processes of the dominant flow fusion network model DSFN and the symmetric hierarchical flow graph network model SHFGNN are as follows: The dominant flow fusion network model DSFN is used to perform intermediate fusion of the dominant sequence features extracted by ESM-C and the shallow structural features output by SHFGNN. During model construction, selectable basic fusion structures include feature concatenation, weighted summation, cross-attention, gated fusion, and deformable convolution-guided fusion. Considering the actual situation in this application where ESM-C sequence features are the dominant features and SaProt structure-aware features are the auxiliary features, a two-stage fusion structure based on structure-guided deformable convolution and FiLM feature modulation is ultimately chosen as the basic network structure of DSFN.

[0033] Specifically, the sequence features output by ESM-C are denoted as dominant features. The shallow structural features output by SHFGNN are recorded as auxiliary features. First, auxiliary features Input is fed into a LoRA-based parameter-efficient injection unit to generate deformable convolution offsets. and FiLM modulation parameters and ,Right now: Then, the offset The input structure guides the deformable convolutional module, which uses auxiliary structural features to adjust the sampling positions of the dominant sequence features, resulting in corrected dominant features. The calculation process is expressed as follows: in, Indicates the current sampling position. This represents the sampling grid of the convolution kernel. Represents the first in the sampling grid One sampling point, Indicates sampling point The corresponding convolution weights, Indicates by offset The extracted first Each sampling offset, This represents the bilinear interpolation function. Subsequently, the FiLM modulation parameters are used... and For the corrected dominant features Channel-level affine modulation is performed to obtain the fusion characteristics of the DSFN output. ,Right now: in, This represents element-wise multiplication. Through the above process, DSFN first uses shallow structural features to perform structure-guided sampling of sequence features, and then completes semantic enhancement through FiLM modulation to obtain enhanced sequence fusion features.

[0034] During training, DSFN is not trained as an independent classification model, but rather as an intermediate fusion module within the overall UniProNet network for end-to-end joint training. The input data for DSFN includes sequence embedding features extracted by ESM-C and shallow structural features output by SHFGNN; the output data is the enhanced sequence fusion feature. This enhanced sequence fusion feature is further input into the Mamba module for long-range dependency modeling and participates in the subsequent deep fusion of the KronHTR module. During training, the LoRA parameters, structure-guided deformable convolution parameters, and FiLM modulation parameters in DSFN are updated together with the overall UniProNet loss function. S103: The enhanced sequence features and multi-scale structural features are input into the pre-established KronHTR module for deep feature fusion, and the multi-modal fusion features are output. Based on the multi-modal fusion features, predictions are made to obtain the prediction results of protein post-translational modification sites.

[0035] The process of inputting the enhanced sequence features and multi-scale structural features into a pre-built KronHTR module for deep feature fusion includes: By combining Gromov-Wasserstein optimal transport theory with dynamically generated Kronecker transform operators, we first achieve soft alignment of the internal geometric structures of the enhanced sequence features and multi-scale structural features at the non-parametric level. Then, we generate instance-specific structured transformations through conditional hypernetworks to iteratively refine the sequence features and achieve deep adaptive heterogeneous feature fusion.

[0036] The KronHTR module first calculates the Gromov-Wasserstein optimal transfer alignment matrix, then constructs a supernetwork conditioned on structural features to dynamically generate Kronecker factor pairs, and finally performs deep feature fusion through iterative refinement blocks to obtain multimodal fused features.

[0037] When making predictions based on multimodal fusion features, the multimodal fusion features are input into a three-layer MLP classifier, which outputs the PTM probability of each residue. A joint loss function is used during training. ,in For cross-entropy loss, The hierarchical contrast loss is defined as follows: This loss promotes feature consistency among similar samples at the same level and semantic alignment between features at different levels.

[0038] The construction and training process of the KronHTR module includes: The KronHTR module is used for post-fusion of sequence augmentation features output by the Mamba module and deep structural features output by SHFGNN. Its input data includes sequence augmentation features. and deep structural features The output data is multimodal fusion features. .

[0039] During module construction, modal intra-modal distance matrices are first constructed based on sequence enhancement features and deep structural features, respectively. Specifically, the intra-modal distance matrix of sequence features is calculated based on cosine distance. and the internal distance matrix of structural features Then, the alignment matrix between the two modes is calculated using Gromov-Wasserstein optimal transport. Its optimization objective is expressed as: in, This represents the set of transfer matrices that satisfy the marginal distribution constraints. Represents the first feature in the sequence The position and structural features of the first The geometric correspondence between the positions. The alignment matrix As a nonparametric structure alignment prior, it is used to guide the subsequent feature fusion process.

[0040] Subsequently, a hypernetwork conditioned on structural features is constructed. Specifically, deep structural features are considered. Pooling is performed to obtain the global structure condition vector. The global structural condition vector is then input into the multilayer perceptron to dynamically generate the first... The Kronecker factor corresponding to each refined block and ,Right now: in, and This is a low-dimensional transformation matrix used to construct structured feature transformation operators. For the input features... The Kronecker transform process is represented as follows: in, This means reshaping the vector into a matrix. This indicates that the matrix is ​​revectorized. Using this method, the module can dynamically generate feature transformation operators based on structural features without directly constructing a high-dimensional parameter matrix.

[0041] Finally, the KronHTR module completes deep feature fusion through multiple iterative refining blocks. In the... In each refined block, the output features of the previous layer will be... As input, features are first aggregated through a self-attention mechanism modulated by the Gromov-Wasserstein alignment matrix: in, , and They represent the first The query matrix, key matrix, and value matrix in each refined block Representing feature dimension, This represents the modulation coefficients of the alignment matrix. Then, the attention output is... The input consists of a structured transformation operator composed of Kronecker factor pairs, which, combined with residual connections and layer normalization, yields the features for the next layer. in, Indicates the first The learnable scaling factor for each refined block Representation layer normalization. After... After several iterations of the refined blocks, the multimodal fusion features are obtained: During training, the KronHTR module is not trained as an independent classification model, but rather as a late-stage fusion module within the UniProNet overall network for end-to-end joint training. During training, the sequence enhancement features output by the Mamba module and the deep structural features output by SHFGNN are input into the KronHTR module to obtain multimodal fusion features. These multimodal fusion features are then input into a three-layer MLP classifier, which outputs the predicted probability of each residue belonging to the target PTM category.

[0042] The overall training loss function of UniProNet consists of cross-entropy loss and hierarchical contrastive loss, expressed as: in, This represents the cross-entropy loss for PTM site classification. Indicates the layered contrast loss. These represent the weighting coefficients of the hierarchical contrastive loss. The hierarchical contrastive loss is expressed as: in, and These represent the same-level contrast loss within different levels. This represents the cross-layer contrast loss between shallow and deep features. This represents the weight coefficients of the cross-layer contrastive loss. Through this joint loss function, the model optimizes the PTM classification results while maintaining feature consistency among similar samples within the same layer and enhancing the semantic alignment between features at different layers.

[0043] Specifically, the present invention will be further illustrated below through embodiments: The process of acquiring protein sequence data and structural coordinate data is as follows: Data preparation and enhancement implementation In this embodiment, labeled datasets containing ten common PTM types were collected to construct model training and testing data. These include O-glycosylation, N-phosphorylation, succinylation, S-carboxyethylation, S-sulfohydration, S-nitrosylation, O-phosphorylation, hydroxylation, SUMOylation, and methylation. Each dataset was divided into training, validation, and test sets in a 7:1:2 ratio to ensure a reasonable ratio of positive to negative samples. Detailed information for each dataset is shown in Table 1.

[0044] Table 1. Dataset Overview For PTM types with a small number of samples in the training set, the PTM-AGEN data augmentation module was enabled. First, the protein language model ESM-3 was used, with real PTM site sequences as context, to generate diverse candidate positive sample sequences through a sequence repair task. Second, the NSGA-II algorithm was used to perform Pareto optimization on the candidate sequences, with optimization objectives including PTM motif conservation (assessed through sequence alignment), structural accessibility (assessed through solvent accessibility predicted by ESM-3), and sequence biological plausibility (assessed through pseudo-perplexity calculated by ESM-3). Then, a mutation operator based on the BLOSUM62 matrix was introduced during the evolutionary process to ensure that the generated amino acid substitutions conform to natural evolutionary laws. Finally, ESM-3 was used to extract the semantic embeddings of the sequences, and cosine similarity with positive samples was calculated. Sequences with high similarity but not labeled as PTM were selected as difficult negative samples and added to the training set. The augmented data volume did not exceed 25% of the original training data, and the test set retained the original real data to ensure unbiased evaluation. Detailed statistical information is shown in Table 2.

[0045] Table 2 New datasets augmented with data. Sequence features were extracted using a pre-trained protein language model, ESM-C. The input was the amino acid sequence of the protein (length ≤ 1024), and a feature vector with a dimension of 1280 was extracted for each residue position. This feature contains information on evolutionary conservation, physicochemical properties, and implicit structural preferences.

[0046] If a protein has an experimentally resolved or predicted 3D structure, the structure-aware protein language model SaProt is used to extract structural features. The input is the protein's PDB file or equivalent structural representation, and the output is a 1280-dimensional feature vector for each residue, explicitly encoding its local 3D geometric conformation. For proteins without structural data, tools such as AlphaFold2 can be used to predict their 3D structure before extracting SaProt features.

[0047] In this embodiment, to comprehensively evaluate the performance of the proposed UniProNet model in the PTM site prediction task, we adopted a series of evaluation metrics widely used in bioinformatics and machine learning. These metrics measure the model's predictive ability from different perspectives: Accuracy reflects the overall classification accuracy; Sensitivity (SEN) and Specificity (SPEC) evaluate the model's ability to identify the positive class (modified sites) and the negative class (unmodified sites), respectively; Mean Accuracy (AP) and F1 score provide more reliable evaluations in imbalanced scenarios; and the Matthews correlation coefficient (MCC) comprehensively considers all confusion matrix elements and is particularly sensitive to imbalanced datasets. Their definitions are as follows.

[0048] In the formula, TP is the number of true positives, representing the number of samples that are actually modified sites and are predicted as modified sites; TN is the number of true negatives, representing the number of samples that are actually unmodified sites and are predicted as unmodified sites; FP is the number of false positives, representing the number of samples that are actually unmodified sites but are predicted as modified sites; FN is the number of false negatives, representing the number of samples that are actually modified sites but are predicted as unmodified sites; p(r) represents the precision at a recall rate of r, where r represents the recall rate.

[0049] Experimental comparison results between UniProNet and existing methods As shown in Table 3, Figure 3 and Figure 4As shown, a comprehensive comparative analysis of UniProNet and several state-of-the-art post-translational modification (PTM) prediction models is presented. This embodiment selects several existing PTM site prediction models as comparison models and compares their performance on datasets of ten different PTM types. Empirical results on a large benchmark dataset containing 10 different PTM types demonstrate that UniProNet outperforms existing state-of-the-art models in the PTM site prediction task. Specifically, UniProNet achieves average improvements of 4.32%, 8.11%, and 5.94% in area under the curve (AUC), Matthews correlation coefficient (MCC), and average precision (AP) compared to the previous best-performing models. Detailed comparisons of these performance metrics are shown in Table 3 and [Table data would be inserted here]. Figure 3 and Figure 4 As shown.

[0050] Table 3 Performance comparison of existing models Example 2: To achieve the above objective, such as Figure 5 As shown, based on Example 1, this invention discloses a protein post-translational modification site prediction system based on multimodal fusion, comprising: The primary processing module 11 is used to acquire the sequence data and structural coordinate data of the protein to be predicted, and to perform dual-stream feature extraction on the sequence data and structural coordinate data of the protein to obtain sequence feature vectors and structural feature vectors. The dual-stream feature extraction includes sequence feature extraction based on the protein sequence and structural feature extraction based on the three-dimensional structure of the protein. The secondary processing module 12 is used to input the sequence feature vector and the structural feature vector into a pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses the structural feature vector as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vector to obtain enhanced sequence features. The structural feature vector and the preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. Local structural features and global structural features are extracted through equivariant message passing and hierarchical structure modeling, and multi-scale structural features are output. The prediction output module 13 is used to input the enhanced sequence features and multi-scale structural features into the pre-established KronHTR module for deep feature fusion, and output multi-modal fusion features. Based on the multi-modal fusion features, prediction is performed to obtain the prediction results of protein post-translational modification sites.

[0051] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0052] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, electronic, magnetic, optical, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0053] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0054] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for predicting protein post-translational modification sites based on multimodal fusion, characterized in that, The method includes the following steps: The sequence data and structural coordinate data of the protein to be predicted are obtained. The sequence data and structural coordinate data of the protein are then subjected to dual-stream feature extraction to obtain sequence feature vectors and structural feature vectors. The dual-stream feature extraction includes sequence feature extraction based on the protein sequence and structural feature extraction based on the protein's three-dimensional structure. Sequence feature vectors and structural feature vectors are input into a pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses structural feature vectors as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vectors to obtain enhanced sequence features. The structural feature vectors and a preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. Local and global structural features are extracted through equivariant message passing and hierarchical structure modeling, and multi-scale structural features are output. The enhanced sequence features and multi-scale structural features are input into the pre-established KronHTR module for deep feature fusion, and the output is a multimodal fusion feature. Based on the multimodal fusion feature, prediction is performed to obtain the prediction results of protein post-translational modification sites.

2. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 1, characterized in that, The dual-stream feature extraction includes: The protein sequence data is input into a sequence feature extraction network based on the pre-trained protein language model ESM-C to obtain a sequence feature vector. Based on the structural coordinate data of the protein, structure-aware input data is generated, and the structure-aware input data is input into the structure feature extraction network based on the structure-aware model SaProt to obtain the structure feature vector.

3. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 1, characterized in that, The process of inputting sequence feature vectors and structural feature vectors into a pre-established dominant flow fusion network model for feature enhancement includes: The offset and modulation parameters are obtained by modulating and extracting the structural feature vectors; By using structure-guided deformable convolution to achieve cross-modal geometric alignment of the sequence feature vectors with offsets and modulation parameters, enhanced sequence features are generated, thereby completing semantic-level content enhancement.

4. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 1, characterized in that, The process of inputting the structural feature vector and the preset graph structure representation into the pre-established symmetric hierarchical flow graph network model is as follows: Based on structural feature vectors and preset graph structure representations, the AttentiveEMAMPConv is used to perform equivariant message passing through the equivariant attention message passing operator and hierarchical differentiable pooling mechanism to update node features and coordinates. Hierarchical coarsening of the graph is performed through differentiable pooling to generate multi-scale structural features.

5. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 4, characterized in that, The preset graph structure is constructed based on residues as nodes and spatial distance and sequence proximity as edges.

6. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 1, characterized in that, The process of inputting the enhanced sequence features and multi-scale structural features into a pre-established KronHTR module for deep feature fusion includes: Calculate the Gromov-Wasserstein optimal transmission alignment matrix based on the enhanced sequence features and multi-scale structural features; Kronecker factor pairs are generated based on the aforementioned multi-scale structural features; Based on the Gromov-Wasserstein optimal transmission alignment matrix and Kronecker factor pair, the enhanced sequence features and multi-scale structural features are iteratively fused to obtain multimodal fused features.

7. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 6, characterized in that, The KronHTR module first calculates the Gromov-Wasserstein optimal transmission alignment matrix, then constructs a supernetwork conditioned on structural features, dynamically generates Kronecker factor pairs, and finally performs deep feature fusion through iterative refining blocks to obtain multimodal fused features.

8. The method for predicting protein post-translational modification sites based on multimodal fusion according to claim 1, characterized in that, The enhanced sequence features capture long-range context dependencies using a Mamba module based on a selective state-space model.

9. A protein post-translational modification site prediction system based on multimodal fusion, employing the protein post-translational modification site prediction method based on multimodal fusion as described in any one of claims 1 to 8, characterized in that, include: A primary processing module is used to acquire the sequence data and structural coordinate data of the protein to be predicted, and to perform dual-stream feature extraction on the sequence data and structural coordinate data of the protein to obtain sequence feature vectors and structural feature vectors. The dual-stream feature extraction includes sequence feature extraction based on the protein sequence and structural feature extraction based on the three-dimensional structure of the protein. The secondary processing module is used to input the sequence feature vector and the structural feature vector into a pre-established dominant flow fusion network model for feature enhancement. The dominant flow fusion network model uses the structural feature vector as guiding information to perform geometric alignment and semantic enhancement on the sequence feature vector to obtain enhanced sequence features. The structural feature vector and the preset graph structure representation are input into a pre-established symmetric hierarchical flow graph network model. Local structural features and global structural features are extracted through equivariant message passing and hierarchical structure modeling, and multi-scale structural features are output. The prediction output module is used to input the enhanced sequence features and multi-scale structural features into the pre-established KronHTR module for deep feature fusion, and output multi-modal fusion features. Based on the multi-modal fusion features, prediction is performed to obtain the prediction results of protein post-translational modification sites.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the protein post-translational modification site prediction method based on multimodal fusion as described in any one of claims 1 to 8.

Images