A method for predicting intrinsically disordered regions based on two-scale features of protein profiles
By using the DisoGraph method and combining it with a dual-scale feature model of protein maps, the problem of incomplete feature representation of IDRs is solved, and deep fusion of sequence and spatial features is achieved, which improves the prediction accuracy and reliability of IDRs and is applicable to biological research and drug design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN TECH UNIV
- Filing Date
- 2026-03-23
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, the characterization of intrinsic disordered regions (IDRs) of proteins is incomplete, the spatial structure information is not fully utilized, and the fusion effect of sequence and spatial features is poor, resulting in insufficient accuracy in IDR identification.
The DisoGraph method is adopted to construct a protein heterogeneous map by using a dual-scale feature model of protein maps, combined with a pre-trained protein language model and structure generator. The RGCN and GVP are used for feature fusion to achieve deep decoupling and efficient fusion of sequence and spatial features.
It improves the prediction accuracy of IDRs, reduces the false positive rate, and enhances the accuracy and reliability of model predictions in short/long disordered regions, providing a more reliable tool for biological research and drug design.
Smart Images
Figure CN122157798A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein structure prediction technology, and in particular to a method for predicting inherently disordered regions based on dual-scale features of protein maps. Background Technology
[0002] Intrinsically disordered proteins (IDPs) are a class of proteins with special properties, lacking a stable three-dimensional structure under physiological conditions. In most cases, the disordered nature of proteins exists only in certain segments, which are defined as intrinsically disordered regions (IDRs). The traditional structure-function paradigm of protein research holds that a stable three-dimensional structure is a prerequisite for proteins to perform their biological functions. However, research shows that IDPs / IDRs have challenged this traditional understanding. They are widely involved in key physiological processes such as cell signaling, DNA regulation, and post-translational modifications, and are closely related to the development and progression of various major diseases, including cancer, neurodegenerative diseases, and cardiovascular diseases. Therefore, accurate identification of IDPs / IDRs is not only a core step in elucidating the molecular mechanisms of life activities but also provides crucial support for the discovery of novel drug targets and drug design, possessing significant academic value and practical implications.
[0003] Currently, experimental methods for studying IDRs mainly include X-ray crystallography, nuclear magnetic resonance (NMR), and circular dichroism (CD). However, these experimental techniques have inherent drawbacks such as high cost, complex operation, and long processing time, making it difficult to meet the needs of large-scale protein sequence IDR analysis. To overcome the limitations of experimental methods, computational model-based IDR prediction tools have gradually become a research hotspot in this field. Feature extraction, as the core step of computational modeling, directly determines the performance ceiling of the prediction model. Traditional prediction methods mainly use two types of features: one is artificially designed features from protein sequences, and the other is sequence profile features obtained through multiple sequence alignment. For example, the first machine learning-based IDR prediction method, PONDR, uses a two-layer feedforward neural network architecture with amino acid composition as the core input feature; the classic method, DISOPRED, searches for target sequences in protein sequence databases using PSI-BLAST, generates a position-specific scoring matrix (PSSM) as a sequence profile feature, and then inputs it into the neural network to complete the prediction.
[0004] In recent years, the rapid development of protein language models has opened up a new paradigm for feature mining of protein sequences. These models analogize protein sequences to the natural language of living systems, using individual amino acids as basic semantic units. Through self-supervised learning strategies, they autonomously learn the combination rules, evolutionary conservation, and potential functional associations of amino acid residues from a massive database of billions of natural protein sequences, ultimately generating implicit protein representations with high information density, effectively capturing deep semantic information at the sequence level. Based on this technological breakthrough, IDR prediction methods have made a series of advances: LMDisorder directly uses the ProtTrans model to extract sequence embedding features, relying on its powerful sequence representation capabilities to achieve accurate IDR prediction; DeepDRP improves the model's predictive robustness by fusing embedding features from four different pre-trained protein language models, utilizing the complementarity of multivariate information; FusionEncoder further breaks through the limitations of single feature types, fusing sequence features extracted by pre-trained language models with traditional features, fully leveraging the advantages of both types of features, and further optimizing prediction results.
[0005] Currently, some studies suggest that relying solely on sequence-level features is insufficient to fully characterize the disordered nature of IDPs. Spatial structural information of proteins (such as contact maps between residues and secondary structures) plays a crucial role in accurately determining the ordered / disordered state of residues. For example, Spot-Disorder, based on traditional PSSM features, introduces predicted protein structural features and experimentally validates the key role of predicted contact maps in IDP identification. Subsequent methods such as IDP-Seq2seq and DeepDRP also attempt to more comprehensively characterize disordered proteins by incorporating spatially relevant features. While these methods have achieved some success in IDP prediction, two common and critical limitations remain: first, most methods use simple feature splicing to integrate sequence and spatial features, making it difficult to capture the complex intrinsic dependencies between them, resulting in poor feature fusion performance; second, traditional structural features largely rely on PSSM information for generation, leading to significant information redundancy with sequence features, which may interfere with effective model learning. Therefore, how to overcome the shortcomings of existing methods in capturing spatial features and achieve deep and efficient fusion of sequence features and spatial features has become the core technical bottleneck for improving the quality of disordered protein characterization and the accuracy of IDR identification. Summary of the Invention
[0006] The purpose of this invention is to propose an intrinsic disorder region prediction method based on dual-scale features of protein maps, in order to solve the problems of incomplete feature representation, insufficient utilization of spatial structure information and poor fusion effect of sequence and spatial features in the prior art, and to achieve accurate identification of intrinsic disorder regions (IDRs) of proteins.
[0007] To achieve the above objectives, this invention proposes an intrinsically disordered region prediction method, DisoGraph, based on dual-scale features of protein maps. This method models IDPs as protein maps. The core module of DisoGraph includes four parts: a protein encoder, a protein map construction module, a joint feature learning module, and a classifier. The specific steps are as follows: Step S1: Obtain the original amino acid sequence of the protein to be predicted; Step S2: Use a protein encoder to encode the original amino acid sequence of the protein to obtain residue-level embedding matrix features and three-dimensional spatial coordinate features; the protein encoder includes a sequence encoder and a structure generator; Step S3: Based on the sequence embedding features and three-dimensional spatial coordinate features, construct a protein heterogeneous map containing scalar features and vector features using the protein map construction module; Step S4: Input the protein heterogeneity map into the joint feature learning module for feature fusion and learning to obtain the fused feature representation of each amino acid residue; the joint feature learning module is constructed based on the relational graph convolutional network RGCN and the geometric vector perceptron GVP. Step S5: Based on the fusion feature representation, predict the probability that each amino acid residue belongs to an ordered or disordered region using a classifier.
[0008] Preferably, in step S2, the residue-level embedding matrix features and three-dimensional spatial coordinate features are obtained, as follows: Step S21: Use the pre-trained protein language model ESM2 as a sequence encoder to encode the original amino acid sequence of the protein, generate a continuous dense vector for each amino acid residue, and form the residue-level embedding matrix feature of the protein sequence. Step S22: Using the structure generator of the protein structure prediction model ESM-Fold, a three-dimensional structure PDB file is generated based on the original amino acid sequence of the protein. The three-dimensional coordinates of the Cα atom of each amino acid residue are parsed from the three-dimensional structure PDB file as representative spatial coordinate features of the residue.
[0009] Preferably, the three-dimensional structure PDB file includes the three-dimensional coordinates, residue numbers, and amino acid types of all atoms in the protein.
[0010] Preferably, in step S3, the construction of the protein heterogeneity map is completed from two dimensions: scalar and vector levels, as detailed below: Step S31: Treat each amino acid residue as a node in the graph, and use the residue-level embedding matrix features as the initial scalar features of the node; Step S32: Based on the sequence position relationship and spatial distance relationship between amino acid residues, the edge set of the protein heterogeneity map is divided into three types: sequence proximity edge, spatial contact edge, and K-nearest edge. Step S33: Based on the three-dimensional spatial coordinate features, construct vector features for each node. The vector features include residue vectors, forward / backward vectors, and K-nearest residue vectors.
[0011] Preferably, in step S32, the three types of edges complement each other, comprehensively characterizing the interaction relationships between residues: Sequence proximity edges: constructed based on the positional relationships of residues in a sequence, used to capture local sequence dependencies; Spatial contact edge: Using residue spatial coordinates, a residue-to-residue distance matrix is constructed based on Euclidean distance to characterize the spatial interaction of spatially adjacent residues; K-nearest edges: Based on Euclidean distance, select the K nearest neighbors in space for each residue.
[0012] Preferably, in step S33, the vector features are as follows: Residue vector: Spatial coordinates of the target residue; Forward / backward vector: The coordinate offset vector between the current residue and its preceding and following adjacent residues; K-nearest residue vector: Based on Euclidean distance, select the K nearest neighbors in space for each residue and calculate the coordinate offset vector from the target residue to each nearest residue.
[0013] Preferably, in step S4, the joint feature learning module is composed of two stacked RGCN-GVP sub-layers; wherein, each RGCN-GVP sub-layer is used to receive scalar features and vector features from the previous layer, obtain the output vector after feature transformation and fusion through the scalar channel and vector channel of the geometric vector perceptron GVP, and then use the relational graph convolutional network to perform feature aggregation on all neighboring nodes of the target node and different types of edges, and finally sum the aggregated features of all edge types; No. l The formula for calculating the feature vector of the target residues in layer +1 is as follows: ; in, and target residues respectively i In the l +1 floor and the l The output features of the layer For target residues i Neighboring residues, For target residues j In the l The output features of the layer For target residues j vector features, For activation function, For geometric vector perceptron, R Let be the set of edge types.
[0014] Preferably, the geometric vector perceptron includes channels for performing linear transformations on scalar features and vector features respectively, and an operation in the scalar channel to concatenate the norm of the transformed vector features with the scalar features.
[0015] Preferably, in step S5, the classifier includes a fully connected layer, an activation function layer, a Dropout layer, and two fully connected network output layers connected in sequence.
[0016] Preferably, the probability of each amino acid residue belonging to an ordered or disordered region is predicted by a classifier, and the specific steps are as follows: First, a fully connected layer is used to refine the features output by the joint feature learning module and compress the feature dimensions to remove redundant features. Next, the ReLU activation function and Dropout layer are used to improve the model's generalization ability; Finally, a two-layer fully connected network is used to achieve binary classification prediction: the first layer of the fully connected network maps the extracted features into a 2-dimensional vector, corresponding to the ordered and unordered states respectively; the second layer uses the Softmax activation function to convert the 2-dimensional vector into a probability distribution.
[0017] Therefore, this invention proposes an intrinsically disordered region prediction method based on dual-scale features of protein maps, which has the following advantages: (1) This invention innovatively integrates the deep sequence semantic features extracted by the pre-trained protein language model with the high-precision spatial structure features generated end-to-end. By constructing a dual-scale map, it achieves comprehensive coverage of sequence information, spatial interaction and geometric vector information, which solves the problems of single feature dimension and insufficient mining of semantic and structural information in traditional methods.
[0018] (2) This invention, through the innovative combination of RGCN and GVP, not only retains the modeling advantages of heterogeneous graph topology, but also achieves deep decoupling and fusion of scalar and vector features by leveraging the invariance / equivariance learning capabilities of GVP, thus completely overcoming the defects of information redundancy and shallow fusion caused by simple splicing in traditional methods.
[0019] (3) This invention demonstrates advantages of low false positive rate and smooth prediction curve in both short / long disordered region prediction. Furthermore, its balanced accuracy, Matthews correlation coefficient and AUC value on independent test sets are superior to mainstream methods, providing a more reliable computational analysis tool for biological mechanism research and drug design.
[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 This is a flowchart of an intrinsic disordered region prediction method based on dual-scale features of protein maps according to the present invention. Figure 2 A visualization comparison of the prediction results for the P31946 protein (including short disordered regions); Figure 3 A visualization comparison of the prediction results for the DP04242 protein (including long disordered regions). Detailed Implementation
[0022] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.
[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0024] Example like Figure 1 As shown, this invention provides a method for predicting inherently disordered regions (IDPs) based on dual-scale features of protein maps, called DisoGraph. IDPs are modeled as protein maps. The core module of DisoGraph includes four parts: a protein encoder, a protein map construction module, a joint feature learning module, and a classifier. The specific steps are as follows: Step S1: Obtain the original amino acid sequence of the protein to be predicted; Step S2: Use a protein encoder to encode the original amino acid sequence of the protein to obtain residue-level embedding matrix features and three-dimensional spatial coordinate features; the protein encoder includes a sequence encoder and a structure generator; The residue-level embedding matrix features and three-dimensional spatial coordinate features are obtained as follows: Step S21: Use the pre-trained protein language model ESM2 as a sequence encoder to encode the original amino acid sequence of the protein, generate a continuous dense vector for each amino acid residue, and form the residue-level embedding matrix feature of the protein sequence. Step S22: Using the structure generator of the protein structure prediction model ESM-Fold, a three-dimensional structure PDB file is generated based on the original amino acid sequence of the protein. The three-dimensional coordinates of the Cα atom of each amino acid residue are parsed from the three-dimensional structure PDB file as representative spatial coordinate features of the residue. The three-dimensional structure PDB file includes the three-dimensional coordinates of all atoms in the protein, residue number, and amino acid type.
[0025] Step S3: Based on sequence embedding features and three-dimensional spatial coordinate features, construct a protein heterogeneous map containing scalar and vector features using the protein map construction module; The construction of protein heterogeneity maps was accomplished from two dimensions: scalar and vector levels, as detailed below: Step S31: Treat each amino acid residue as a node in the graph, and use the residue-level embedding matrix features as the initial scalar features of the node; Step S32: Based on the sequence position and spatial distance relationships between amino acid residues, the edge set of the protein heterogeneity map is divided into three types: sequence proximity edges, spatial contact edges, and K-nearest edges. These three types of edges complement each other, comprehensively characterizing the interaction relationships between residues. Sequence proximity edges: constructed based on the positional relationships of residues in a sequence, used to capture local sequence dependencies; Spatial contact edge: Using residue spatial coordinates, a residue-to-residue distance matrix is constructed based on Euclidean distance to characterize the spatial interaction of spatially adjacent residues; K-nearest edges: Based on Euclidean distance, select the K nearest neighbors in space for each residue.
[0026] Step S33: Based on the three-dimensional spatial coordinate features, construct vector features for each node; the vector features include residue vectors, forward / backward vectors, and K-nearest residue vectors, as detailed below: Residue vector: Spatial coordinates of the target residue; Forward / backward vector: The coordinate offset vector between the current residue and its preceding and following adjacent residues; K-nearest residue vector: Based on Euclidean distance, select the K nearest neighbors in space for each residue and calculate the coordinate offset vector from the target residue to each nearest residue.
[0027] Step S4: Input the protein heterogeneity map into the joint feature learning module for feature fusion and learning to obtain the fusion feature representation of each amino acid residue; The joint feature learning module is built on the relational graph convolutional network RGCN and the geometric vector perceptron GVP, and consists of two stacked RGCN-GVP sub-layers. Each RGCN-GVP sub-layer receives scalar and vector features from the previous layer. The output vector after feature transformation and fusion is obtained through the scalar and vector channels of the geometric vector perceptron GVP. Then, the relational graph convolutional network is used to aggregate features of the target node and all neighboring nodes of different edge types. Finally, the aggregated features of all edge types are summed. No. l The formula for calculating the feature vector of the target residues in layer +1 is as follows: ; in, and target residues respectively i In the l +1 floor and the l The output features of the layer For target residues i Neighboring residues, For target residues j In the l The output features of the layer For target residues j vector features, For activation function, For geometric vector perceptron, R Let be the set of edge types.
[0028] The geometric vector perceptron includes channels that perform linear transformations on scalar and vector features respectively, as well as an operation in the scalar channel that concatenates the norm of the transformed vector features with the scalar features.
[0029] Step S5: Based on the fusion feature representation, predict the probability of each amino acid residue belonging to an ordered or disordered region using a classifier; wherein, the classifier includes a fully connected layer, an activation function layer, a Dropout layer, and two fully connected network output layers connected in sequence. The probability of each amino acid residue belonging to an ordered or disordered region is predicted using a classifier. The specific steps are as follows: First, a fully connected layer is used to refine the features output by the joint feature learning module and compress the feature dimensions to remove redundant features. Next, the ReLU activation function and Dropout layer are used to improve the model's generalization ability; Finally, a two-layer fully connected network is used to achieve binary classification prediction: the first layer of the fully connected network maps the extracted features into a 2-dimensional vector, corresponding to the ordered and unordered states respectively; the second layer uses the Softmax activation function to convert the 2-dimensional vector into a probability distribution.
[0030] The invention will be further illustrated below through specific implementation examples.
[0031] 1. Dataset.
[0032] 1.1 Training dataset.
[0033] The training and optimization of the model in this embodiment uses the DM4229 dataset, which contains 4229 data points. The selection criteria are as follows: The protein sequence similarity is less than 25%; the protein resolution is less than or equal to 2A; the sequence length is greater than or equal to 30 residues; the information of each residue is fully recorded in the PDB database; non-standard amino acid residues in the dataset have been removed; the dataset is divided into two parts, DM3000 as the training set and DM1229 as the validation set.
[0034] 1.2 Test dataset.
[0035] The DISORDER723 dataset is a protein sequence database containing inherently disordered regions, containing 723 proteins with a total of 121,212 amino acid residues. The DISORDER723 dataset is divided into two subsets: the first subset contains 345 protein sequences with a single disordered region, and the second subset contains 378 protein sequences with multiple disordered regions. Detailed statistical information about the dataset is shown in Table 1.
[0036] Table 1. Detailed statistics of the dataset
[0037] 2. Specific implementation.
[0038] First, the sequence encoder was initialized using the pre-trained protein language model weight file esm2_t33_600M_UR50D. The model hyperparameters were configured as follows: the optimizer was AdamW, with an initial learning rate of 3e^(-1 / 2). -4 A Dropout rate of 0.2 is introduced in the second fully connected layer; the batch size is set to 8; the cross-entropy loss function is used for training, and the maximum number of training epochs is 20; to prevent model overfitting, an early stopping strategy is introduced in this embodiment. When the validation set loss does not decrease for two consecutive epochs, the training process is terminated, and the optimal model weight file selected based on the minimum loss of the validation set is automatically saved.
[0039] Based on the actual annotation results of the residues, the model's prediction results for each residue can be divided into four categories: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Sensitivity (Sn) measures the model's ability to correctly identify disordered residues, while specificity (Sp) measures the model's ability to correctly identify ordered residues. Balanced accuracy (BACC) and Matthews correlation coefficient (MCC) can comprehensively reflect the model's overall classification performance on both disordered and ordered samples, effectively avoiding evaluation bias caused by sample imbalance. The area under the ROC curve (AUC) can quantitatively evaluate the model's ability to distinguish the predictive tendency of residue disorder; a higher value indicates better predictive discrimination.
[0040] 3. Results.
[0041] In this embodiment, DP04242 protein containing a long disordered region (LDR, ≥30 consecutive disordered residues) and P31946 protein containing a short disordered region (SDR, <30 consecutive disordered residues) were selected from the MobiDB and Disprot databases, respectively. Figure 2 and Figure 3 The visualization details show the prediction results of disordered regions for these two proteins using three methods: DisoGraph, flDPnn, and IUPred3.
[0042] like Figure 2 As shown, for proteins containing short disordered regions, the DisoGraph method exhibits superior prediction performance: its false positive (FP) residue count is only 3, significantly lower than the other two methods (13 and 43, respectively); simultaneously, DisoGraph's false negative (FN) residue count is only 1. Figure 3 As shown, when predicting proteins containing long disordered regions, DisoGraph's prediction curves are smoother compared to the other two methods, and no mutations occur in the disorder probability between consecutive residues. In the prediction experiments for these two proteins, the DisoGraph method exhibited the lowest false positive rates (FPR), at 1.3 and 5.4, respectively.
[0043] The results show that the DisoGraph method significantly reduces the false positive rate during prediction, providing a more reliable computational analysis tool for the accurate analysis of disordered protein regions and subsequent functional studies.
[0044] To verify the effectiveness of the proposed method, its performance was compared with 10 current mainstream state-of-the-art methods on the independent test set DISORDER723. These methods included DISOPRED, DeepCNF, AUCpreD, IDP-Seq2Seq, RFPR-IDP, flDPnn, IUPred3, and DeepDRP. Among these methods, except for DeepDRP, all rely on multi-source information fusion strategies and use the position-specific scoring matrix (PSSM) as the core input feature. Further analysis revealed that DeepCNF, AUCpreD, IDP-Seq2Seq, and flDPnn also integrate structural and functional information predicted by various tools to enhance the model's ability to represent disordered proteins, such as solvent accessibility predicted by SPIDER2, secondary structure inferred by PSIPRED, and other auxiliary features. Unlike the aforementioned methods, DeepDRP uses static embedding vectors generated by a pre-trained protein language model as input features. To ensure the fairness of the comparative experiments, sequences with a similarity exceeding 25% between the training set and the DISORDER723 test set were first removed. The DisoGraph method was then retrained and its performance evaluated, with results shown in Table 2. As can be seen from the data in Table 2, the protein language model-based methods (DeepDRP and DisoGraph) outperformed other sequence profile-based methods. DisoGraph achieved the best performance, with its BACC, MCC, and AUC exceeding those of other methods by 2.7%–16.5%, 1.9%–36.4%, and 1.6%–21.5%, respectively.
[0045] Table 2 Comparison of different methods on the DISORDER723 test set
[0046] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.
[0047] Therefore, this invention provides a method for predicting intrinsically disordered regions (IDRs) based on dual-scale features of protein maps. By comprehensively covering protein sequence and multi-dimensional spatial features, it achieves deep and efficient fusion of scalar and vector features, significantly improving the prediction accuracy and reliability of intrinsically disordered regions (IDRs) of proteins, and providing a better computational analysis tool for related biological mechanism research and drug design.
[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for predicting inherently disordered regions based on dual-scale features of protein maps, characterized in that, The specific steps are as follows: Step S1: Obtain the original amino acid sequence of the protein to be predicted; Step S2: Use a protein encoder to encode the original amino acid sequence of the protein to obtain residue-level embedding matrix features and three-dimensional spatial coordinate features; the protein encoder includes a sequence encoder and a structure generator; Step S3: Based on the sequence embedding features and three-dimensional spatial coordinate features, construct a protein heterogeneous map containing scalar features and vector features using the protein map construction module; Step S4: Input the protein heterogeneity map into the joint feature learning module for feature fusion and learning to obtain the fusion feature representation of each amino acid residue; The joint feature learning module is constructed based on the relational graph convolutional network RGCN and the geometric vector perceptron GVP; Step S5: Based on the fusion feature representation, predict the probability that each amino acid residue belongs to an ordered or disordered region using a classifier.
2. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 1, characterized in that, In step S2, the residue-level embedding matrix features and three-dimensional spatial coordinate features are obtained, as follows: Step S21: Use the pre-trained protein language model ESM2 as a sequence encoder to encode the original amino acid sequence of the protein, generate a continuous dense vector for each amino acid residue, and form the residue-level embedding matrix feature of the protein sequence. Step S22: Using the structure generator of the protein structure prediction model ESM-Fold, a three-dimensional structure PDB file is generated based on the original amino acid sequence of the protein. The three-dimensional coordinates of the Cα atom of each amino acid residue are parsed from the three-dimensional structure PDB file as representative spatial coordinate features of the residue.
3. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 2, characterized in that, The three-dimensional structure PDB file includes the three-dimensional coordinates, residue numbers, and amino acid types of all atoms in the protein.
4. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 1, characterized in that, In step S3, the protein heterogeneity map is constructed from two dimensions: scalar and vector, as detailed below: Step S31: Treat each amino acid residue as a node in the graph, and use the residue-level embedding matrix features as the initial scalar features of the node; Step S32: Based on the sequence position relationship and spatial distance relationship between amino acid residues, the edge set of the protein heterogeneity map is divided into three types: sequence proximity edge, spatial contact edge, and K-nearest edge. Step S33: Based on the three-dimensional spatial coordinate features, construct vector features for each node. The vector features include residue vectors, forward / backward vectors, and K-nearest residue vectors.
5. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 4, characterized in that, In step S32, the three types of edges complement each other, comprehensively characterizing the interaction relationships between residues: Sequence proximity edges: constructed based on the positional relationships of residues in a sequence, used to capture local sequence dependencies; Spatial contact edge: Using residue spatial coordinates, a residue-to-residue distance matrix is constructed based on Euclidean distance to characterize the spatial interaction of spatially adjacent residues; K-nearest edges: Based on Euclidean distance, select the K nearest neighbors in space for each residue.
6. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 5, characterized in that, In step S33, the vector features are as follows: Residue vector: Spatial coordinates of the target residue; Forward / backward vector: The coordinate offset vector between the current residue and its preceding and following adjacent residues; K-nearest residue vector: Based on Euclidean distance, select the K nearest neighbors in space for each residue and calculate the coordinate offset vector from the target residue to each nearest residue.
7. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 6, characterized in that, In step S4, the joint feature learning module is composed of two stacked RGCN-GVP sub-layers. Each RGCN-GVP sub-layer receives scalar and vector features from the previous layer, obtains the output vector after feature transformation and fusion through the scalar and vector channels of the geometric vector perceptron (GVP), and then uses a relational graph convolutional network to aggregate features of the target node and all neighboring nodes of different types of edges. Finally, the aggregated features of all edge types are summed. No. l The formula for calculating the feature vector of the target residues in layer +1 is as follows: ； in, and target residues respectively i In the l +1 floor and the l The output features of the layer For target residues i Neighboring residues, For target residues j In the l The output features of the layer For target residues j vector features, For activation function, For geometric vector perceptron, R Let be the set of edge types.
8. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 7, characterized in that, The geometric vector perceptron includes channels that perform linear transformations on scalar features and vector features respectively, and an operation that concatenates the norm of the transformed vector features with the scalar features in the scalar channel.
9. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 8, characterized in that, In step S5, the classifier includes a fully connected layer, an activation function layer, a Dropout layer, and two fully connected network output layers connected in sequence.
10. The method for predicting inherently disordered regions based on dual-scale features of protein maps according to claim 9, characterized in that, The probability of each amino acid residue belonging to an ordered or disordered region is predicted using a classifier. The specific steps are as follows: First, a fully connected layer is used to refine the features output by the joint feature learning module and compress the feature dimensions to remove redundant features. Next, the ReLU activation function and Dropout layer are used to improve the model's generalization ability; Finally, a two-layer fully connected network is used to achieve binary classification prediction: the first layer of the fully connected network maps the extracted features into a 2-dimensional vector, corresponding to the ordered and unordered states respectively; the second layer uses the Softmax activation function to convert the 2-dimensional vector into a probability distribution.