Low-confidence or disordered protein structure searching method based on representation algorithm
By employing self-supervised pre-training and supervised fine-tuning of deep graph neural networks and large protein language models, combined with a self-supervised learning framework and attention score loss function, the high cost and disordered structure exploration challenges in existing technologies are addressed, achieving high-precision protein structure search.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-04-03
AI Technical Summary
Existing characterization-based protein structure search methods rely on experimentally resolved ordered structures, resulting in high acquisition costs and difficulty in effectively exploring disordered structures, and there is a lack of dedicated disordered structure databases as benchmarks.
We employ a representation-based approach, using a deep graph neural network and a large protein language model for self-supervised pre-training and supervised fine-tuning of the structure. By combining a self-supervised learning framework and an attention score loss function, we extract protein sequence and structural representations, and perform feature fusion and normalization to improve the accuracy of structural representations.
It significantly improves the structure search accuracy and retrieval ranking performance of low-confidence or unordered structures, and can more robustly handle low-confidence or unordered structures, thereby increasing the proportion of correctly similar structures hit in the search results.
Smart Images
Figure CN121789754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of bioengineering, specifically a method for searching low-confidence or disordered protein structures based on characterization algorithms. Background Technology
[0002] Existing characterization-based protein structure search methods typically rely on experimentally resolved ordered structures as training labels, which has two limitations: firstly, acquiring experimental data is time-consuming and costly; secondly, exploring protein structures containing inherently disordered regions still faces challenges, as there is a lack of specialized databases of disordered structures as benchmarks. Summary of the Invention
[0003] This invention addresses the shortcomings of existing technologies in effectively searching for structures in low-confidence or disordered regions. It proposes a protein structure search method based on a representation algorithm. Under the contrastive learning framework, it learns protein structure representations based on deep graph neural networks and a large protein language model. Through self-supervised structure pre-training and supervised fine-tuning, it significantly improves the accuracy of structure representation and search precision.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a method for searching low-confidence or disordered protein structures based on a representation algorithm. The method comprises: ① extracting protein sequences from protein structure files, extracting general sequence embeddings using a protein language model pre-trained on a large sequence database, and obtaining fine-tuned sequence representations through an additional network; ② extracting protein tertiary structures from protein structure files, constructing protein graphs through preprocessing, and extracting structural representations using a graph neural network-based structural encoder; then, performing feature fusion and normalization on the sequence and structural representations to obtain the final sequence-enhanced structural representation; comparing the final structural representation with a structural representation database, and outputting the found similar structures.
[0006] The acquisition of sequence representation includes: standardizing and encoding the extracted protein sequence and inputting it into a protein sequence pre-training model to obtain a full-sequence-level sequence embedding; inputting the sequence embedding into a sequence-side auxiliary network for fine-tuning and dimensionality reduction;
[0007] The sequence-side supplemental network is implemented using, but is not limited to, Transformer networks and multilayer perceptron networks.
[0008] The aforementioned preprocessing for constructing protein maps specifically involves: [the process is described in the original text, but the translation is incomplete]. A protein map was constructed using Cartesian coordinates of carbon atoms, and original node features were extracted. This protein map uses amino acid residues as nodes and residue coordinates as features. Carbon atom coordinates represent node positions; edges are established based on the Euclidean distance between any two residues and a preset distance threshold to form adjacency relationships;
[0009] The original node features include: distance-based geometric features and angle-based geometric features that remain unchanged under rigid body transformations of rotation and translation.
[0010] The distance-based geometric features include: calculating the distance between a target residue and at least one set of reference points in three-dimensional space, wherein each set of reference points refers to a predetermined residue within a continuous segment of the protein structure. Cartesian coordinates of a carbon atom.
[0011] The angle-based geometric features include: based on three consecutive residues in the protein structure. The middle residue of the carbon atom corresponds to The carbon atom forms the vertex, which determines the included angle.
[0012] The structural encoder includes a context extraction module, a feature extraction module, and a readout module. The context extraction module performs local pattern extraction and long-range dependency modeling based on the original node features obtained from the preprocessing module to obtain features rich in sequence context information. The feature extraction module uses these context-rich features as the original node features of the protein graph and performs multiple rounds of message passing on the graph structure to obtain node features with structural context. The readout module summarizes the embeddings of all nodes in the protein graph into a graph embedding based on the node features with structural context, thus obtaining a representation of the entire protein structure.
[0013] The fine-tuning involves connecting a multilayer perceptron to the output of the pre-trained structural encoder to map general structural features to structural representations required for downstream tasks. A supervised fine-tuning dataset is then constructed to minimize the distance between structurally similar samples (positive samples) and representations, and maximize the distance between structurally dissimilar samples (negative samples), within a contrastive learning framework. During backpropagation, only the parameters of the multilayer perceptron and the additional network are updated; the parameters of the pre-trained structural encoder remain fixed and are not updated.
[0014] This invention relates to a protein structure search system based on a representation algorithm for implementing the above-mentioned method, characterized by low confidence or disorder. The system comprises: a structure file parsing unit, a sequence representation generation unit, a structural representation generation unit, and a fusion retrieval output unit. Specifically: the structure file parsing unit parses the input protein structure file information to obtain protein sequence information and protein tertiary structure information; the sequence representation generation unit uses a protein language model pre-trained on a large sequence database to extract sequence information based on the protein sequence information, and performs feature mapping through an additional network to obtain a sequence representation; the structural representation generation unit performs graph construction and node feature extraction processing in a preprocessing module based on the protein tertiary structure information, and performs structural feature extraction processing through a graph neural network-based structural encoder to obtain a structural representation; the fusion retrieval output unit performs feature fusion and normalization processing based on the sequence representation and structural representation to obtain a final structural representation enhanced with sequence representation, and compares and sorts the final structural representation with a structural representation database to obtain and output search results for similar structures.
[0015] Technical effect
[0016] This invention constructs high-confidence substructures based on prediction confidence and forms training constraints: Starting from the residue-level confidence index (pLDDT) of the predicted structure, it automatically extracts local high-confidence substructures and constructs positive sample pairs between the overall structure and local high-confidence substructures in contrastive learning, enabling the model to learn a structural consistency representation that is insensitive to prediction noise; Confidence-guided attention constraint loss: In addition to the contrastive learning objective, an additional loss based on attention score and pLDDT is introduced, enabling the structural encoder to actively increase attention to high-confidence regions and suppress interference from low-confidence regions during pre-training, thereby obtaining a more robust structural representation for low-confidence or disordered structures; Lightweight fine-tuning strategy of freezing the pre-trained encoder and updating only the additional network: During the fine-tuning stage, the parameters of the pre-trained structural encoder and the pre-trained protein language model are fixed, and the adaptation to the target task is completed only by updating the parameters of the additional network. The sequence representation obtained from the sequence channel and the structural representation obtained from the structure channel are fused and normalized to obtain the final sequence-enhanced structural representation for structural similarity comparison and retrieval. Compared with existing technologies, this invention can achieve better retrieval ranking performance and a higher proportion of top retrieval results hitting the correct similar structures in structure search tasks with low confidence prediction or disordered structures. Attached Figure Description
[0017] Figure 1 This is a flowchart of the present invention;
[0018] Figure 2 A diagram illustrating the comparative learning framework;
[0019] Figure 3 This is a schematic diagram of the preprocessing module of the present invention;
[0020] Figure 4 This is a schematic diagram illustrating the construction of a local high-confidence region substructure for this invention;
[0021] Figure 5 This is a schematic diagram of the structure encoder in the pre-training stage of the present invention;
[0022] Figure 6 This is a schematic diagram of the structural encoder for the fine-tuning stage of the present invention; Detailed Implementation
[0023] like Figure 1 As shown in this embodiment, a method for searching low-confidence or disordered protein structures based on a characterization algorithm is provided, including:
[0024] S1. Under the framework of contrastive learning, the self-supervised pre-training of the graph neural network-based structural encoder is carried out with the goal of minimizing the distance between similar sample representations and maximizing the distance between dissimilar sample representations. Specifically, a dataset of substructures in local high-confidence regions is constructed, and the overall structure and substructures in local high-confidence regions are regarded as positive sample pairs. The graph neural network-based structural encoder is trained with an additional loss function based on attention score and predicted structure confidence pLDDT to achieve the effect of paying more attention to the high-confidence regions of the predicted structure.
[0025] The training described, such as Figure 2 As shown, specifically, the Momentum Comparison (MoCo) method is used, employing two encoders sharing the same architecture but with different parameter sets. and The implementation involves: First, the protein and bond representations are input into a preprocessing module for protein graph construction and raw node feature extraction. These features are then input into the subsequent encoder to obtain the corresponding query and bond representations. The first encoder... Update by optimizing the loss function through backpropagation. Second encoder Then through To perform momentum update Loss functions of the two encoders ,in: For query representation, For the dynamic queue One representation, This represents the positive samples corresponding to the query in the current mini-batch. For temperature coefficient, This is the size of the queue. All hyperparameters are set to [value]. The default value given.
[0026] The preprocessing module, such as Figure 3 As shown, this preprocessing module, used to encode the tertiary structure of a protein into a graph structure and extract the original node features, employs a residue-level graph construction approach to model the protein structure, that is, based on the graph of each residue... A protein graph is constructed using Cartesian coordinates of carbon atoms, where the set of nodes represents all residues in the protein structure. Carbon atoms; the construction of edge sets is first based on the residues. The Cartesian coordinates of carbon atoms are used to calculate the distance matrix between residues. Then, when the Euclidean distance between residues is less than a preset distance threshold, the distance matrix is calculated. (e.g. 10) When a residue is present, an edge is constructed between it; otherwise, no edge exists, thus obtaining the adjacency matrix between residues. The extraction of original node features is also based on each residue's... The Cartesian coordinates of a carbon atom include two types of features: distance-based and angle-based. The final dimension is 32, and these coordinates are invariant to rotation and translation. The core of distance-based feature derivation lies in the selection of reference points, which are determined based on the distances between the target residue and a certain number of reference points in three-dimensional space. The coordinates of these reference points are... According to the formula Calculations show that , Represents the first protein residues Cartesian coordinates of atoms. Original node feature vectors based on distance. The dimension is ,here This is a hyperparameter that controls the number of reference points, with a default value of 5. Its core idea is to divide the reference points into... Group, for the first Group, wherein: the number of reference points contained is This represents a short, continuous segment within a protein. Angle-based features are based on three consecutive residues in the protein structure. It is derived from the angle formed by the atoms, and the specific calculation formula is as follows: ,in: The coordinates of the current residue. The coordinates of the previous residue. This is the coordinate of the next residue. By definition, the first and last residues of a protein sequence have no left (right) adjacent residues, therefore... The value is directly assigned to 0.
[0027] The construction of the local high-confidence region substructure dataset involves the following steps: AlphaFoldDB (e.g., v4 version) is selected as the baseline dataset. Specifically, AlphaFold2 predicted structures corresponding to the proteomes of some common model organism species in UniProtKB are used, as shown in Table 1. This dataset does not include isomers and sequences with unknown or non-standard amino acids. To control quality and reduce dataset size, the residue-level prediction confidence index pLDDT for each predicted structure is averaged to obtain the average confidence index. And filter out The predicted structure is located between 60-80 (or other preset intervals). Then, the corresponding local high-confidence region substructure is constructed, specifically as follows: Figure 4 The process is as follows: 1. For each predicted protein structure, extract a set of residue indices with a pLDDT (prediction confidence level) of at least 80. If no residues satisfy pLDDT ≥ 80 or no continuous fragments exist, the fragment is considered invalid. 2. Determine if any continuous fragments exist with a length at least half the total length of the protein. If so, select the longest fragment from the continuous fragments that meet the condition as the local high-confidence region substructure and output it. 3. If no continuous fragments meet the length condition, splicing is allowed as long as the interval between adjacent fragments does not exceed a preset maximum interval threshold, resulting in a spliced candidate fragment. The candidate fragments are then checked for validity, requiring that the average pLDDT of the spliced local protein substructure is still at least 80, and the length of the spliced fragment is at least half the total length of the protein. If these conditions are not met, the fragment is considered invalid. 4. Add the candidate fragments that meet the conditions to the fragment library and introduce randomness in the fragment selection stage: either random selection or selection of the fragment with the highest average pLDDT as the final local high-confidence region substructure is used for output. Based on the above selection criteria, after filtering the original 1,874,834 predicted structures, the remaining 196,570 predicted structures corresponded to local high-confidence region substructures, which were used as the pre-training dataset for the structures.
[0028] Table 1: Structural Predictions in Selected AlphaFoldDB Samples
[0029] like Figure 5As shown, the pre-trained structural encoder based on graph neural networks includes a context extraction module, a feature extraction module, and a readout module. Specifically: the context extraction module performs local pattern extraction and long-range dependency modeling based on the original node features obtained from the preprocessing module to obtain features rich in sequence context information; the feature extraction module uses these context-rich features as the original node features of the protein graph and performs multiple rounds of message passing on the graph structure to obtain node features with structural context; the readout module summarizes the embeddings of all nodes in the protein graph into a graph embedding based on the node features with structural context to obtain a representation of the entire protein structure.
[0030] The context extraction module includes two convolutional layers and a bidirectional long short-term memory network. The first convolutional layer performs convolution operations and nonlinear mapping on features at adjacent positions based on local neighborhood information of the input feature sequence to obtain a first convolutional feature map. The bidirectional long short-term memory network performs bidirectional temporal modeling and context dependency learning on the feature vectors at each position based on the sequence feature information of the first convolutional feature map to obtain a sequence context representation containing forward and backward context information. The second convolutional layer further performs convolutional fusion and feature reconstruction on the context features within the local neighborhood based on the sequence context representation to obtain a context-rich feature representation as the original feature of the residue nodes in the protein map.
[0031] The feature extraction module includes multiple graph neural network blocks, wherein the graph neural network blocks perform layer-by-layer graph attention calculation, graph convolution feature aggregation, and batch normalization processing on each residue feature according to the residue node features and the adjacency relationship information between residues in the protein graph. Through the cascaded stacking of multiple graph neural network blocks, an updated residue feature representation that characterizes the multi-level topological structure and multi-scale context of the protein is obtained.
[0032] The graph neural network block includes: a graph attention layer (GAT), a graph convolutional layer (GCN), and a batch normalized (BN) layer, the output of which is: ,in: Specifically: The Graph Attention Layer (GAT) adaptively learns the attention weights between each node and its neighbors based on the initial features of each node in the input graph and the adjacency information between nodes, and performs weighted aggregation on the features of the neighbors to obtain a first node feature representation representing the first-order neighborhood relationship; the Graph Convolutional Layer (GCN) performs graph convolution operations and feature smoothing aggregation on the features of each node based on the first node feature representation and the adjacency information of the graph to obtain a second node feature representation encoding the graph structure information and contextual dependencies; the Batch Normalization Layer (BN) normalizes the second node feature representation to improve the stability of network training and accelerate the convergence speed.
[0033] The graph attention layer GAT has 8 attention heads, each with an independent weight matrix and output. The average of all outputs is then used as the output of the graph attention layer GAT and subsequently input into the graph convolutional layer GCN.
[0034] The graph neural network blocks described herein are constructed by introducing skip connections and identity mappings to create residual blocks. Each residual block contains two graph neural network blocks, specifically: ,in: And if and When the dimensions are inconsistent, It can be used to match feature dimensions. This represents the LeakyReLU linear unit.
[0035] The readout module is used to aggregate the node features output by stacking multiple graph neural network blocks into a graph-level representation. It includes a global pooling unit and a mapping unit. The global pooling unit performs global max pooling on the node features in the protein graph to obtain a graph-level feature vector. The mapping unit performs a dimensionality transformation on the graph-level feature vector and then performs a nonlinear transformation through a feedforward mapping network composed of several convolutional layers, outputting a fixed-dimensional graph embedding to represent the structural embedding corresponding to the protein. The feedforward mapping network includes batch normalization, activation functions, and regularization operations.
[0036] The additional loss function is used to encourage the model to assign higher attention weights to neighboring nodes with higher prediction confidence estimates (pLDDT) than the center node when aggregating features of surrounding nodes in the protein graph. Specifically: Where: self represents the pLDDT of the central node, and nbr represents the pLDDT of the neighboring nodes. The attention weights of neighboring nodes to the central node generated by the graph attention layer GAT are... It is a hyperparameter used to scale the attention score. According to the principle of minimizing the loss function: when the pLDDT of a neighboring node is higher than that of the center node, the corresponding attention score is forced to increase; if the pLDDT of a neighboring node is lower than that of the center node, the corresponding attention score is forced to decrease.
[0037] S2. Fine-tune the pre-trained structural encoder obtained in step S1 using supervised and sequence augmentation methods. Specifically, construct a supervised fine-tuning dataset and fine-tune the pre-trained structural encoder from step S1 by enhancing the structural representation through sequence representation based on a protein language model.
[0038] The construction of the supervised fine-tuning dataset specifically involves using cd-hit to filter out structures from the pre-trained dataset that have... Protein structures with sequence identity are selected to ensure non-redundancy. The fine-tuning dataset contains 36,111 structures. The TM-align structure alignment algorithm is used to perform all structure alignments on the fine-tuning dataset to determine positive and negative sample pairs. That is, for a single query, it identifies which other protein structures in the dataset are its structural neighbors (positive samples) and which are structurally dissimilar samples (negative samples).
[0039] The structure encoder in the fine-tuning stage, such as Figure 6 As shown, the specific steps are: structural channel representation extraction, sequence channel representation extraction, fusion, and normalization. Specifically, structural channel representation extraction involves connecting a multilayer perceptron to the output of the pre-trained structural encoder obtained in step S1 to map general structural features into structural representations required for downstream tasks. Sequence channel representation extraction involves connecting an additional network to the output of the pre-trained protein language model to further encode general sequence embeddings into sequence representations. The structural representations obtained from the structural channels and the sequence representations obtained from the sequence channels are added and fused to obtain a fused representation. Subsequently, the fused representation is normalized to obtain the final sequence-enhanced structural representation.
[0040] In the fine-tuning phase, within the framework of contrastive learning, the distance between structurally similar samples (positive sample pairs) is minimized, while the distance between structurally dissimilar samples (negative sample pairs) is maximized. During backpropagation, only the parameters of the multilayer perceptron and the additional network are updated, while the parameters of the pre-trained model remain fixed and are not updated. The pre-trained model includes a pre-trained structural encoder and a pre-trained protein language model.
[0041] The sequence channel characterization extraction specifically involves: extracting sequence information from protein sequences using a large-scale protein sequence pre-trained model, preferably the ESM-2 model, and more preferably a 33-layer model pre-trained on UniRef50. The parameter version is used to obtain the sequence embedding at the full sequence level (1280 dimensions). The sequence embedding is then input into an additional network to obtain the sequence representation of the sequence channel. The additional network includes a Transformer model and a multilayer perceptron; the Transformer model employs a multi-head attention mechanism (quad-head attention), several encoder layers (2 layers), and receives word embedding dimensions (1280 dimensions) consistent with the sequence embedding dimensions.
[0042] S3. The fine-tuned structure encoder from step S2 is used to characterize the protein structure. The similarity between protein structures is compared by calculating the similarity between structural representations. Further, to accelerate the structure search, a target protein structure database can be pre-constructed, storing the final structural representation vectors corresponding to the proteins. Specifically, the representation vector of the structure to be queried is compared with the representation vectors in the structural representation database using cosine similarity calculation, and the top similarity scores are ranked. The output is a similar structure, where: The number of similar structures required by the user.
[0043] The structural representation database is derived from the supervised dataset during the fine-tuning phase, and a training dataset containing 28,888 structures was selected using 5-fold cross-validation.
[0044] Through specific practical experiments, this embodiment downloaded the protein structure file numbered AF-B4F8R0-F1 (https: / / alphafold.ebi.ac.uk / entry / B4F8R0) from the AlphaFoldDB website as input, selected the search database as the target database for the fine-tuning stage, and obtained the results shown in Table 2 using the aforementioned protein structure search method based on the representation algorithm for low confidence or disorder. The results include: the number of similar structures returned. It is 5.
[0045] Table 2: Search results for protein structures numbered AF-B4F8R0-F1
[0046] Note: The TM-score value ranges from 0 to 1, with the closer to 1 indicating higher structural similarity.
[0047] The species for which the protein structure was queried in the table is Zeamays, with a length of 431, an average pLDDT of 78.6 (high), and Pfam annotation for Proteinphosphatase. The search results show that although the fourth target structure found has a high TM-score with the query structure, it was not chosen as the primary target due to its low confidence in the predicted structure. This is because the core of this invention is to make the structure encoder pay more attention to the high-confidence region of the predicted structure throughout the entire process of structure representation and its downstream task of structure search.
[0048] The evaluation metrics used in this embodiment are Area Under Precision-Recall (AUPRC) and the previous... The top-K hit rate for each search result, where true and false positives refer to the probability of finding a protein. The TM-score was calculated using the structure alignment algorithm TM-align with all other proteins in the dataset, where the largest TM-score was denoted as: When the target protein is found With query protein The TM-score between them meets the condition ,but regarded as True positive, also known as The structure of the neighboring structures must be valid; otherwise, it is a false positive.
[0049] The TM-score specifically refers to: ,in: Let be the distance between the i-th pair of aligned residues. This is a distance parameter used to standardize the distance. The number of aligned residues, The length of the target protein.
[0050] The Top-K hit rate is specifically as follows: ,in: For the algorithm targeting the first The number of positive samples found in the Top-K for each query. For the first The number of actual structural neighbors for each query. This represents the total number of queries.
[0051] As shown in Table 3, this invention demonstrates significant advantages over other methods across various evaluation metrics. Regarding the AUPRC metric, this invention achieves a 5.82% improvement over the state-of-the-art TM-Vec method. Furthermore, in evaluating Top-K hit rates (K=1, 5, 10), this invention achieves improvements of 53.43%, 58.69%, and 56.61% respectively compared to FoldExplorer, and 3.36%, 3.32%, and 3.56% respectively compared to the state-of-the-art TM-Vec method.
[0052] Table 3: Comparison of results between this embodiment and other methods
[0053] Compared with existing technologies, the performance improvement of this device / method (i.e., the effect achieved by adopting new technologies in which of the above-mentioned stages) lies in the following: In the pre-training stage, by introducing a contrastive learning framework, not only is the problem of insufficient protein structure labels solved, but the overall protein structure and local high-confidence substructures are also used as positive sample pairs. An additional loss function based on attention scores and the confidence index pLDDT of predicted structures is proposed to encourage the model to actively focus on high-confidence regions of predicted structures, thereby improving the accuracy of structural representation. In the fine-tuning stage, the powerful functions of the pre-trained protein language model and Transformer are used to extract protein sequence features, further enhancing the structural representation generated by the graph encoder pre-trained on the structure. Specifically, this includes two channels: sequence representation extraction and structural representation extraction. For the sequence representation extraction channel, protein sequences are first extracted from the protein structure file, and then a general sequence embedding is extracted using a protein language model pre-trained on a large sequence database. Finally, a fine-tuned sequence representation is obtained through an additional network. For the structural representation extraction channel, the tertiary structure of the protein is first extracted from the protein structure file, and then a protein graph is constructed through preprocessing. Finally, a structural representation is extracted using a graph neural network-based structural encoder. The sequence representation and structural representation are processed by feature fusion and normalization to obtain the final sequence representation-enhanced structural representation. The final structural representation is compared with the structural representation database to output the similar structures found.
[0054] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for searching low-confidence or disordered protein structures based on a representation algorithm, characterized in that, ① After extracting protein sequences from protein structure files, a general sequence embedding is extracted using a protein language model pre-trained on a large sequence database, and then a fine-tuned sequence representation is obtained through a sequence-side attachment network; ② After extracting protein tertiary structures from protein structure files, a protein graph is constructed through preprocessing, and then a structural representation is extracted using a graph neural network-based structural encoder. The sequence representation and structural representation are then processed by feature fusion and normalization to obtain the final sequence representation-enhanced structural representation. The final structural representation is then compared with the structural representation database to output the similar structures found.
2. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 1, characterized in that, The acquisition of sequence representation includes: standardizing and encoding the extracted protein sequence and inputting it into a protein sequence pre-training model to obtain a full-sequence-level sequence embedding; and inputting the sequence embedding into a sequence-side auxiliary network for fine-tuning and dimensionality reduction.
3. The method for searching low-confidence or disordered protein structures based on representation algorithms according to claim 1, characterized in that, The aforementioned preprocessing for constructing protein maps specifically involves: [the process is described in the original text, but the translation is incomplete]. A protein map was constructed using Cartesian coordinates of carbon atoms, and original node features were extracted. This protein map uses amino acid residues as nodes and residue coordinates as features. Carbon atom coordinates represent node positions; edges are established based on the Euclidean distance between any two residues and a preset distance threshold to form adjacency relationships; The original node features include distance-based geometric features and angle-based geometric features, which remain unchanged under rigid body transformations of rotation and translation.
4. The method for searching low-confidence or disordered protein structures based on representation algorithms according to claim 3, characterized in that, The distance-based geometric features include: calculating the distance between a target residue and at least one set of reference points in three-dimensional space, wherein each set of reference points refers to a predetermined residue within a continuous segment of the protein structure. Cartesian coordinates of a carbon atom; the angle-based geometric features include: based on three consecutive residues in the protein structure. The middle residue of the carbon atom corresponds to The carbon atom forms the vertex, which determines the included angle.
5. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 1, characterized in that, The structural encoder includes a context extraction module, a feature extraction module, and a readout module. The context extraction module performs local pattern extraction and long-range dependency modeling based on the original node features obtained from the preprocessing module to obtain features rich in sequence context information. The feature extraction module uses these context-rich features as the original node features of the protein graph and performs multiple rounds of message passing on the graph structure to obtain node features with structural context. The readout module summarizes the embeddings of all nodes in the protein graph into a graph embedding based on the node features with structural context, thus obtaining a representation of the entire protein structure.
6. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 5, characterized in that, The structure encoder described herein is obtained through self-supervised contrastive learning pre-training, specifically including: constructing positive sample pairs and negative sample sets, and optimizing them with contrastive loss and confidence-guided additional loss, so that the related structural representations of the same protein are close to each other in the embedding space, and the structural representations of different proteins are far apart from each other. The confidence-guided additional loss is constrained by residue-level prediction confidence information (pLDDT) and model attention weights. Where: self is the pLDDT of the central node, and nbr is the pLDDT of the neighboring nodes. The attention weights of neighboring nodes to the center node generated by the graph attention mechanism. This is a hyperparameter used to scale the attention score. This enhances the model's focus on high-confidence predicted structural regions and suppresses interference from low-confidence regions, thereby improving the accuracy and robustness of structural representation.
7. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 6, characterized in that, The construction of positive sample pairs and negative sample sets refers to: constructing positive sample pairs based on the overall structure of the same protein and its local high-confidence substructures, and constructing negative sample sets based on other protein structures or historical batch structures; wherein, the construction of the local high-confidence substructures includes: screening a set of residues with a confidence level not less than a preset threshold from residue-level predicted confidence information (pLDDT), and forming one or more continuous fragments according to the residue index; when the interval between adjacent continuous fragments is not greater than a preset maximum interval threshold, it is allowed to splice the adjacent continuous fragments into candidate fragments; retaining a set of candidate fragments that meet the requirements of an average confidence level not less than a preset threshold and a length not less than a preset proportion of the total length, and determining the local high-confidence substructures from the set of candidate fragments, wherein the determination method can be random selection or selection of the candidate fragment with the highest average confidence level.
8. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 1, characterized in that, The fine-tuning involves connecting a multilayer perceptron to the output of the pre-trained structural encoder to map general structural features to structural representations required for downstream tasks. A supervised fine-tuning dataset is then constructed to minimize the distance between structurally similar samples (positive samples) and representations, and maximize the distance between structurally dissimilar samples (negative samples), within a contrastive learning framework. During backpropagation, only the parameters of the multilayer perceptron and the additional network are updated; the parameters of the pre-trained structural encoder remain fixed and are not updated.
9. The method for searching low-confidence or disordered protein structures based on a representation algorithm according to claim 1, characterized in that, The aforementioned structural representation database is constructed as follows: AlphaFold2 predicted structures corresponding to the proteomes of some common model organisms in UniProtKB are subjected to the sequence representation extraction, structural representation extraction, and feature fusion normalization processing described in claim 1 to obtain the corresponding final structural representations. These final structural representations and corresponding structural identification information are then stored to form the structural representation database. The search includes: calculating the similarity between the final structural representation of the structure to be queried and the final structural representations in the structural representation database, preferably using cosine similarity; sorting the structures according to their similarity scores; and outputting the top k similar structures as search results, where k is a preset value or is set by the user.
10. A protein structure search system based on a representation algorithm for low-confidence or disordered structures, implementing the method of any one of claims 1-9, characterized in that, include: The system comprises a structure file parsing unit, a sequence representation generation unit, a structural representation generation unit, and a fusion retrieval output unit. Specifically: the structure file parsing unit parses the input protein structure file information to obtain protein sequence information and protein tertiary structure information; the sequence representation generation unit uses the protein sequence information to extract sequence information by calling a protein language model pre-trained on a large sequence database and performs feature mapping through an additional network to obtain a sequence representation; the structural representation generation unit performs graph construction and node feature extraction processing in the preprocessing module based on the protein tertiary structure information, and performs structural feature extraction processing through a graph neural network-based structural encoder to obtain a structural representation; and the fusion retrieval output unit performs feature fusion and normalization processing based on the sequence representation and structural representation to obtain a final structural representation enhanced with sequence representation, and compares and ranks the final structural representation with a structural representation database to obtain search results for similar structures and output them.