A protein metal binding site prediction method and system based on saliency weighted graph transformer

Abstract

The present application relates to the technical field of protein analysis, and particularly relates to a protein metal binding site prediction method and system based on a saliency weighted graph transformer, which comprises the following steps: obtaining an amino acid sequence of a protein to be predicted; extracting sequence evolution features and protein spatial structure features, and constructing initial node features and graph edge features; performing local refining processing on the initial node features to obtain refined node features; constructing a protein residue topology graph by taking the refined node features as graph nodes and the graph edge features as graph edges, performing spatial topology coding, and obtaining coding features corresponding to each residue; calculating saliency weights of the corresponding coding features, weighting and aggregating all the coding features to obtain a global semantic vector, feeding back and fusing the global semantic vector into the coding features, obtaining residue enhanced features, and performing classification decoding to output metal ion binding site prediction results corresponding to each residue. The present application improves the prediction accuracy, stability and efficiency of metal ion binding sites.

Description

Technical Field

[0001] This invention relates to the field of protein analysis technology, and in particular to a method and system for predicting protein metal binding sites based on a saliency-weighted graph transformer. Background Technology

[0002] Metal ions are key cofactors in regulating the physiological functions of proteins, participating in core physiological processes such as stabilizing protein spatial conformation and mediating enzyme-catalyzed reactions. Protein metal ion binding sites refer to specific amino acid residue regions in the polypeptide chain of proteins that can form coordination bonds with metal ions and directly participate in the exercise of function. Accurate identification of binding sites is an important foundation for conducting protein functional annotation, pathogenic mechanism analysis, and drug target development.

[0003] Existing techniques for predicting protein metal ion binding sites primarily rely on early sequence-based methods that model one-dimensional features such as amino acid physicochemical properties and evolutionary conservatism spectra. While applicable to a wide range of scenarios, these methods struggle to capture three-dimensional spatial coordination patterns, resulting in inherent limitations in prediction accuracy. Alternatively, graph neural networks serve as the core architecture, treating amino acid residues as graph nodes and inter-residue spatial interactions as graph edges. They learn spatial coordination patterns through neighborhood message passing mechanisms. Some improved schemes further introduce pre-trained protein language models to extract co-evolutionary semantic features to supplement functional information. However, due to the inherent limitations of local neighborhood message passing mechanisms, these models lack global structural awareness and struggle to effectively capture long-range cooperative coordination patterns at the whole-protein scale. In metal ion prediction tasks with weak coordination signals and complex conformations, significant breakthroughs in prediction accuracy are difficult to achieve. Furthermore, directly constructing structural features using three-dimensional Cartesian coordinates is susceptible to geometric artifacts caused by coordinate system transformations. High-dimensional pre-trained sequence features generally suffer from redundant noise and low efficiency in extracting local functional motifs, further restricting improvements in prediction performance and model robustness.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] In view of at least one of the above technical problems, the present invention provides a protein metal binding site prediction method and system based on a saliency-weighted graph transformer, which adopts a collaborative architecture of local refinement of sequence features, spatial encoding of residue topology graph and global feedback of saliency weighting to improve the prediction accuracy and robustness of metal ion binding sites.

[0006] This invention provides a method for predicting protein metal binding sites based on a saliency-weighted graph transformer, comprising: Obtain the amino acid sequence of the protein to be predicted, wherein the amino acid sequence is composed of multiple amino acid residues in sequence; Extract the sequence evolution features and protein spatial structure features corresponding to the amino acid sequence, construct initial node features based on the sequence evolution features, and construct graph edge features based on the spatial structure features; The initial node features are locally refined to filter feature redundancy and extract local functional motifs to obtain refined node features corresponding to each residue. Using the refined node features as graph nodes and the graph edge features as graph edges, a protein residue topology graph is constructed. The graph is then input into a graph transformer encoder for spatial topology coding to obtain the coding features corresponding to each residue. Calculate the saliency weight of the coding feature corresponding to each residue, perform weighted aggregation on all coding features based on the saliency weight to obtain a global semantic vector, and feed the global semantic vector back into the coding feature to obtain residue-enhanced features; The enhanced features of the residues are classified and decoded, and the predicted results of the metal ion binding sites corresponding to each residue are output.

[0007] Furthermore, the sequence evolution features are extracted through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

[0008] Furthermore, after obtaining the spatial coordinates of residues through a protein structure prediction model, the protein spatial structure features are transformed into a pairwise distance matrix between residues. The pairwise distance matrix has rotation and translation invariance and serves as a component of the graph edge features.

[0009] Furthermore, the local refinement process is implemented using a multi-layer gated convolutional network. The gated convolutional network adaptively filters redundant noise in high-dimensional features through a gating mechanism and captures local amino acid coordination motifs in the sequence through convolutional kernels, thereby completing feature dimensionality reduction and semantic refinement.

[0010] Furthermore, the protein residue topology graph is a K-nearest neighbor graph, where each residue node is connected to the K nearest neighbor residues in terms of spatial distance; the graph edge features are fused with three complementary features: relative position encoding, distance radial basis encoding, and local spatial orientation encoding.

[0011] Furthermore, the obtained residue enhancement feature includes: The L2 norm of the encoded feature corresponding to each residue is calculated as the original saliency score, and the saliency weight of each residue is obtained by normalization with masking. Using the saliency weight as a coefficient, the coding features of each residue are weighted and summed to obtain the global semantic vector that characterizes the overall functional context of the whole protein; The global semantic vector is copied along the sequence dimension and then concatenated with the encoded features of each residue along the feature dimension to obtain the residue-enhanced features.

[0012] Furthermore, the classifier used to classify and decode the residue enhancement features is obtained during the training phase using an adaptive weighted binary cross-entropy loss function; the loss function adjusts the loss weights of positive and negative samples according to the ratio of the number of bound residues to non-bound residues in the training set.

[0013] Furthermore, a ligand-independent prediction mode is adopted, which includes using only the amino acid sequence of the protein to be predicted as the sole input, without requiring input of ligand structure information.

[0014] This invention also provides a protein metal binding site prediction system based on a saliency-weighted graph transformer, comprising: The acquisition module acquires the amino acid sequence of the protein to be predicted, wherein the amino acid sequence is composed of multiple amino acid residues in sequence; The extraction and construction module extracts the sequence evolution features and protein spatial structure features corresponding to the amino acid sequence, constructs initial node features based on the sequence evolution features, and constructs graph edge features based on the spatial structure features; The preprocessing module performs local refinement on the initial node features, filters feature redundancy and extracts local functional motifs to obtain refined node features corresponding to each residue. The feature extraction module constructs a protein residue topology graph using the refined node features as graph nodes and the graph edge features as graph edges. The graph is then input into the graph transformer encoder for spatial topology encoding to obtain the encoded features corresponding to each residue. The feature enhancement module calculates the saliency weight of the encoded feature corresponding to each residue, performs weighted aggregation on all the encoded features based on the saliency weight to obtain a global semantic vector, and feeds back the global semantic vector into the encoded feature to obtain the residue-enhanced feature; The results output module classifies and decodes the residue enhancement features and outputs the predicted results of metal ion binding sites corresponding to each residue.

[0015] Furthermore, the extraction construction module includes: The training unit extracts the sequence evolution features through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

[0016] The technical solution of this invention can achieve the following technical effects: By employing a collaborative architecture of sequence feature refinement, spatial encoding of protein residue topology graphs, and saliency-weighted global feedback, this approach overcomes the core shortcomings of graph neural network prediction schemes, such as insufficient global structure perception, susceptibility to geometric artifacts, and redundancy of high-dimensional sequence features. It simultaneously captures local amino acid coordination motifs and long-range spatial coordination patterns at the whole-protein scale, improving the prediction accuracy of metal ion binding sites and exhibiting stronger generalization robustness for protein structure inputs of varying quality.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] 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, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the protein metal binding site prediction method based on a saliency-weighted graph transformer in an embodiment of the present invention; Figure 2 This is a logical diagram illustrating the sequence evolution features obtained in an embodiment of the present invention; Figure 3 This is a logical schematic diagram of the residue enhancement features obtained in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] This invention provides a method such as Figures 1 to 3The protein metal binding site prediction method based on the saliency-weighted graph transformer shown includes: Obtain the amino acid sequence of the protein to be predicted. The amino acid sequence is composed of multiple amino acid residues in sequence. Extract sequence evolution features and protein spatial structure features corresponding to amino acid sequences, construct initial node features based on sequence evolution features, and construct graph edge features based on spatial structure features; The initial node features are locally refined to filter feature redundancy and extract local functional motifs, resulting in refined node features corresponding to each residue. A protein residue topology graph is constructed using refined node features as graph nodes and graph edge features as graph edges. The graph is then input into a graph transformer encoder for spatial topology coding to obtain the coding features corresponding to each residue. Calculate the saliency weight of the coding feature corresponding to each residue, and aggregate all coding features based on the saliency weight to obtain a global semantic vector. Feed the global semantic vector back into the coding features to obtain residue-enhanced features. The residue enhancement features are classified and decoded, and the predicted results of metal ion binding sites corresponding to each residue are output.

[0023] The working principle of this invention is as follows: First, the amino acid sequence of the protein to be predicted is obtained. The amino acid sequence consists of multiple amino acid residues linked sequentially along a peptide chain, serving as the basic input for the entire prediction process. Then, residue-level sequence evolution features and spatial structure features of the protein are extracted from the corresponding amino acid sequence. Sequence evolution features reflect the evolutionary conservation properties and functional semantic information of each amino acid residue, while spatial structure features reflect the three-dimensional spatial relative positions of different residues. Based on this, initial node features are constructed by integrating the sequence evolution features of each residue, and graph-edge features between residues are constructed by transforming the spatial structure features, thus completing the construction of multimodal input features. Next, the initial node features undergo local refinement through local feature filtering and extraction operations. To remove redundant noise from high-dimensional features and capture continuously distributed local functional motifs along the sequence, refined node features corresponding to each residue are obtained with reduced dimensions and focused semantics. Then, using the refined node features corresponding to each residue as graph nodes and the graph edge features between residues as graph edges, a protein residue topology graph representing the spatial topological relationships of the protein is constructed. This topology graph is input into a graph transformer encoder for multi-layer spatial topology encoding. Through a neighborhood message passing mechanism, the spatial neighborhood information around each residue is fused to learn the spatial coordination patterns between residues, ultimately obtaining the encoding features corresponding to each residue. Finally, a saliency-weighted global feedback processing is applied to the encoding features. Specifically, the saliency of the encoding feature corresponding to each residue is first calculated. Weights are used to quantify the contribution of corresponding residues to the overall function of the protein. Residues with richer feature information and stronger functional correlation correspond to higher significance weights. Then, based on the significance weights of each residue, the coding features of all residues in the protein are weighted and aggregated. This amplifies features in functionally critical regions and suppresses features in background regions, ultimately resulting in a global semantic vector that characterizes the overall functional context and macroscopic structural properties of the protein. Finally, this global semantic vector is copied along the peptide chain length dimension and fed back into the coding features of each residue. This allows local residue features to obtain supplementary global structural information at the whole protein scale, overcoming the inherent limitation of traditional graph neural networks that can only perceive local neighborhoods. The method identifies the residues and their corresponding residue enhancement features. Finally, it classifies and decodes these enhanced features, mapping the high-dimensional features to a metal ion binding tendency score for each residue. The final output is the predicted metal ion binding site for each residue; a higher score indicates a higher probability of the corresponding residue participating in metal ion coordination. For example, for typical globular proteins with hundreds of residues, the prediction method in this embodiment can complete the full sequence site prediction in a very short time. It can directly identify protein sequences with unknown functions, is suitable for predicting binding sites of various common metal ions, and maintains high computational efficiency while ensuring prediction accuracy. It can support applications such as large-scale protein functional annotation and drug target screening.

[0024] In some embodiments of the present invention, such as Figure 2 As shown, sequence evolution features are extracted through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

[0025] The pre-trained protein language model is pre-trained based on billions of protein sequences from different evolutionary branches through self-supervised learning. The training process does not require manual annotation of binding site information. The model implicitly learns the co-evolutionary patterns, secondary structure tendencies, and functional semantic rules contained in the amino acid sequence through a multi-layer attention mechanism. In specific implementation, the ESM-2 pre-trained protein language model with 3 billion parameters is preferred as the feature extraction tool. After inputting the complete amino acid sequence of the protein to be predicted into the model, the hidden state matrix output by the final attention layer of the model is retrieved to obtain the residue-level co-evolutionary semantic embedding corresponding to each amino acid residue. The high-dimensional embedding integrates global evolutionary context and local biochemical variation information, which can provide high-quality sequence functional prior features for subsequent local refinement and spatial topological coding, effectively making up for the lack of information at the chemical semantic level of single structural features.

[0026] In some embodiments of the present invention, after obtaining the spatial coordinates of residues through a protein structure prediction model, the protein spatial structure features are transformed into a pairwise distance matrix between residues. The pairwise distance matrix has rotation and translation invariance and serves as a component of the graph edge features.

[0027] End-to-end structural feature extraction can be performed using the ESMFold structure prediction model. During implementation, the amino acid sequence of the protein to be predicted is directly input into the model, which runs in fully inference mode with automatic mixed precision. The model extracts the three-dimensional spatial coordinates of each residue backbone atom from its internal tensor, reducing the computational overhead of file reading / writing and format conversion, and achieving seamless feature conversion from sequence semantics to spatial topology. Subsequently, based on the extracted spatial coordinates of each residue, the Euclidean distance between backbone atoms of all residue pairs is calculated, generating a pairwise distance matrix between residues. The specific formula is as follows: In the formula and These represent the three-dimensional spatial components of residues i and j generated internally by ESMFold, respectively. By preserving the relative positional relationship between residues, the geometric artifact problem can be eliminated, providing stable and unbiased geometric constraints for subsequent graph structure modeling. The pairwise distance matrix, as the core component of the graph edge features, can be used for the construction of subsequent protein residue topology graphs and spatial topology encoding.

[0028] In some embodiments of the present invention, the local refinement process is implemented using a multi-layer gated convolutional network. The gated convolutional network adaptively filters redundant noise in high-dimensional features through a gating mechanism and captures local amino acid coordination motifs in the sequence through convolutional kernels, thereby completing feature dimensionality reduction and semantic refinement.

[0029] A three-layer cascaded gated convolutional neural network structure is employed to perform layer-by-layer semantic extraction and dimensionality compression on the high-dimensional initial node features. In implementation, the high-dimensional co-evolutionary semantic components in the initial node features are first projected onto a set feature dimension space through a learnable linear layer, and then sequentially fed into three structurally identical gated convolutional blocks for processing. Each gated convolutional block has two parallel branches: a main convolutional path and a gated path. The main convolutional path uses a one-dimensional convolution operation with a kernel size of 5 to capture local amino acid arrangement patterns in the sequence, and employs zero-padding to maintain consistent sequence length before and after processing. The gated path generates element-wise dynamic weight coefficients through a sigmoid activation function, which are then used in conjunction with the input of the main convolutional path. The process involves element-wise multiplication to adaptively filter redundant noise unrelated to metal coordination function in high-dimensional features, while strengthening effective features related to local coordination motifs. Each gated convolutional block has a residual connection structure at its end, allowing input features to be directly superimposed onto output features, ensuring smooth gradient propagation during backpropagation and avoiding gradient decay issues in multi-layer networks. After progressive refinement through three gated convolutional blocks, the noise-filtered sequence features are obtained. These features are then concatenated along the feature dimension with the biophysical components in the initial node features. Finally, through joint linear projection and nonlinear activation processing, refined node features with dimension adaptation and semantic focus are obtained, completing the overall feature dimensionality reduction and semantic refinement process.

[0030] In some embodiments of the present invention, the protein residue topology graph is a K-nearest neighbor graph, where each residue node is connected to the K nearest neighbor residues in terms of spatial distance; the graph edge features are fused with three complementary features: sequence relative position encoding, distance radial basis encoding, and local spatial orientation encoding.

[0031] The protein residue topology graph is constructed as a K-nearest neighbor graph, where K can be set to 30. During construction, the refined node features corresponding to each residue are used as graph nodes. Based on the pairwise distance information between residues, the 30 spatially closest neighboring residues are selected for each target residue node, and topological edges are established to form a residue topology graph with controllable sparsity, covering the three-dimensional spatial coordination environment around the residues and avoiding the computational surge problem caused by fully connected graphs. The graph edge features are integrated with three complementary topological features: sequence relative position encoding, distance radial basis encoding, and local spatial orientation encoding. Among them, the sequence relative position encoding is converted into a fixed-dimensional feature vector by using sine and cosine frequency mapping, introducing sequence relative proximity into the graph structure. Information assists the model in perceiving the positional associations of residues on the peptide chain. Distance radial basis encoding expands the spatial distance between residues at multiple scales using multiple equally spaced Gaussian kernels, characterizing the local spatial density changes within the pocket and enhancing the model's ability to perceive fine distance differences related to metal coordination. Local spatial orientation encoding constructs a local orthogonal coordinate system for each residue and parameters the relative spatial orientation between two residue coordinate systems with quaternions, preserving side chain spatial orientation information to assist in the identification of complex coordination patterns. After the three types of features are spliced ​​and aligned along the feature dimensions, they are projected onto the set hidden dimension space after linear mapping and layer normalization to obtain parameterized graph edge tensors, providing high-fidelity geometric constraints for the spatial topology encoding of the subsequent graph transformer encoder.

[0032] In some embodiments of the present invention, such as Figure 3 As shown, residue enhancement features are obtained, including: The L2 norm of the coding feature corresponding to each residue is calculated as the original saliency score, and the saliency weight of each residue is obtained by normalization with masking. Using significance weights as coefficients, the coding features of each residue are weighted and summed to obtain a global semantic vector that characterizes the overall functional context of the whole protein. The global semantic vector is copied along the sequence dimension and then concatenated with the encoding features of each residue along the feature dimension to obtain the residue-enhanced features.

[0033] Residue enhancement features are generated through a saliency-weighted global feedback approach, specifically including three steps: saliency weight calculation, global semantic aggregation, and feature feedback fusion. First, the L2 norm of the encoded feature corresponding to each residue is calculated as the original saliency score. The specific formula for the original saliency score is as follows: ,in This represents the original significance score, pre-defined. This represents the output feature vector of residue node i after the final refinement layer of the GraphTransformer encoder. The score is used to quantify the information energy density of the corresponding residue features. Residues in functionally critical regions contain richer geometric coordination and evolutionary semantic information, resulting in higher original saliency scores, thus enabling automatic identification of functionally significant regions of the whole protein. Subsequently, the original saliency scores of all residues are normalized with a mask. The sequence mask is used to hide zero-padding residues at the end of the sequence to avoid invalid padding terms interfering with the weight distribution. After normalization, the saliency weight corresponding to each residue is obtained. The specific calculation formula is as follows: ,in This constitutes the normalized structural significance weight of residue i, satisfying... M i and M j Let L represent the sequence mask moments corresponding to residues i and j, and L represent the total length of the protein sequence, thus completing the normalization mapping of the original scores. Next, using the saliency weight of each residue as the aggregation coefficient, the encoding features of all residues in the whole protein are summed element-wise with weights. This amplifies the features of functionally significant regions during aggregation, while suppressing the features of non-functional background regions, ultimately yielding a global semantic vector that characterizes the overall functional context and macroscopic structural topological properties of the whole protein. , where g represents the global semantic vector; finally, the obtained global semantic vector is copied along the sequence length dimension so that each residue position corresponds to a complete set of global semantic information. Then, the copied global semantic vector and the coding features corresponding to each residue are spliced ​​and fused along the feature dimension so that the local features of each residue can supplement the global structural context at the whole protein scale. Finally, residue enhancement features with both local and global perspectives are obtained, which can overcome the short-sightedness of traditional graph neural networks that can only perceive information in local neighborhoods, effectively transmit long-range cooperative coordination information across structural domains, improve the recognition accuracy of metal ion binding sites with weak coordination signals and complex conformations, and suppress false positive prediction results in non-functional regions.

[0034] In some embodiments of the present invention, the classifier used for classifying and decoding residue enhancement features is obtained during the training phase using an adaptive weighted binary cross-entropy loss function; the loss function adjusts the loss weights of positive and negative samples according to the ratio of the number of bound residues to the number of unbound residues in the training set.

[0035] The classifier used for classifying and decoding residue enhancement features preferably employs a linear readout layer structure. During model training, the classifier is optimized end-to-end using an adaptive weighted binary cross-entropy loss function. Due to the extremely high spatial sparsity of metal ion binding sites in natural proteins, the number of non-binding residues in the training samples is order of magnitude greater than that of binding residues. The ordinary binary cross-entropy loss function tends to bias the model towards predicting the higher proportion of non-binding residues, making it difficult to effectively learn sparse coordination features. Therefore, an adaptive weighting mechanism is used to specifically correct the loss function. Before training begins, the total number of binding and non-binding residues in the current training set is fully counted. Based on the ratio of their numbers, the loss scaling factor corresponding to positive samples is automatically calculated. The lower the proportion of binding residue samples, the higher the corresponding loss weight, thus adaptively balancing the contribution ratio of positive and negative samples in the loss calculation. The specific formula for the loss scaling factor is: Where γ represents the loss scaling factor for positive samples, and N neg N represents the total number of non-binding residues in the training set. pos This represents the total number of binding residues in the training set. To avoid excessive bias towards positive samples due to excessively high weights under extreme imbalance conditions, a safe upper limit is set for the scaling factor, balancing recall and precision. A sequence position mask is also introduced during loss calculation to mask invalid loss gradients caused by zero-padding residues at the end of the sequence, ensuring that loss calculation only applies to the true protein residue sequence. This improves the numerical stability of the training process, effectively preventing the model from converging to a trivial solution that predicts all non-binding residues, and enhancing the learning sensitivity to highly localized coordination pocket geometry. The overall loss function L is formulated as follows:

[0036] in, y represents the experimental true label of residue i. i This represents the actual tag corresponding to residue i, with a value of 1 for bound residues and a value of 0 for unbound residues; This forms the sequence position mask corresponding to residue i, used to eliminate the loss gradient caused by the tail zero-padding token, p i This represents the prediction score corresponding to residue i.

[0037] In some embodiments of the present invention, a ligand-independent prediction mode is adopted, including: using only the amino acid sequence of the protein to be predicted as the only input, without inputting ligand structure information.

[0038] The prediction method employs a ligand-independent prediction mode, where the entire prediction process uses only the amino acid sequence of the protein to be predicted as the sole input. No pre-input of any ligand structural, chemical, or identifier information is required, nor is it necessary to specify the type of metal ion to be predicted. During implementation, after inputting the amino acid sequence, the model automatically completes the entire process of sequence evolution feature extraction, protein spatial structure prediction, topology map construction and encoding, saliency-weighted feedback, and classification decoding. Ultimately, it outputs prediction results for binding sites corresponding to various common metal ions, eliminating reliance on known ligand information and enabling universal binding site screening even in scenarios with unknown binding ligand types, significantly expanding the method's applicability.

[0039] Based on the same inventive concept as the protein metal binding site prediction method based on a saliency-weighted graph transformer in the foregoing embodiments, the present invention also provides a protein metal binding site prediction system based on a saliency-weighted graph transformer, comprising: The acquisition module acquires the amino acid sequence of the protein to be predicted. The amino acid sequence is composed of multiple amino acid residues arranged sequentially. Extract the building blocks, extract the sequence evolution features and protein spatial structure features corresponding to the amino acid sequences, construct the initial node features based on the sequence evolution features, and construct the graph edge features based on the spatial structure features; The preprocessing module performs local refinement on the initial node features, filters feature redundancy and extracts local functional motifs to obtain the refined node features corresponding to each residue. The feature extraction module uses refined node features as graph nodes and graph edge features as graph edges to construct a protein residue topology graph. The graph transformer encoder performs spatial topology encoding to obtain the encoded features corresponding to each residue. The feature enhancement module calculates the saliency weight of the coding feature corresponding to each residue, performs weighted aggregation on all coding features based on the saliency weight to obtain a global semantic vector, and feeds the global semantic vector back into the coding feature to obtain the residue-enhanced feature; The results output module classifies and decodes the residue enhancement features and outputs the predicted results of the metal ion binding sites corresponding to each residue.

[0040] The prediction system described above in this invention can effectively realize the protein metal binding site prediction method based on the saliency weighted graph transformer, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0041] In some embodiments of the present invention, the extraction building module includes: The training unit extracts sequence evolution features through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

[0042] Similarly, the above-mentioned optimization schemes for the system can also respectively achieve the optimization effects corresponding to the methods in the embodiments, which will not be repeated here.

[0043] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for predicting protein metal binding sites based on a saliency-weighted graph transformer, characterized in that, Includes the following steps: Obtain the amino acid sequence of the protein to be predicted, wherein the amino acid sequence is composed of multiple amino acid residues in sequence; Extract the sequence evolution features and protein spatial structure features corresponding to the amino acid sequence, construct initial node features based on the sequence evolution features, and construct graph edge features based on the spatial structure features; The initial node features are locally refined to filter feature redundancy and extract local functional motifs to obtain refined node features corresponding to each residue. Using the refined node features as graph nodes and the graph edge features as graph edges, a protein residue topology graph is constructed. The graph is then input into a graph transformer encoder for spatial topology coding to obtain the coding features corresponding to each residue. Calculate the saliency weight of the coding feature corresponding to each residue, perform weighted aggregation on all coding features based on the saliency weight to obtain a global semantic vector, and feed the global semantic vector back into the coding feature to obtain residue-enhanced features; The enhanced features of the residues are classified and decoded, and the predicted results of the metal ion binding sites corresponding to each residue are output.

2. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The sequence evolution features are extracted through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

3. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The protein spatial structure features are obtained by using a protein structure prediction model to obtain the spatial coordinates of residues, and then transformed into a pairwise distance matrix between residues. The pairwise distance matrix has rotation and translation invariance and serves as a component of the graph edge features.

4. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The local refinement process is implemented using a multi-layer gated convolutional network. The gated convolutional network adaptively filters redundant noise in high-dimensional features through a gating mechanism and captures local amino acid coordination motifs in the sequence through convolutional kernels, thereby completing feature dimensionality reduction and semantic refinement.

5. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The protein residue topology graph is a K-nearest neighbor graph, where each residue node is connected to the K nearest neighbor residues in terms of spatial distance; the graph edge features are fused with three complementary features: relative position encoding, distance radial basis encoding, and local spatial orientation encoding.

6. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The obtained residue enhancement features include: The L2 norm of the encoded feature corresponding to each residue is calculated as the original saliency score, and the saliency weight of each residue is obtained by normalization with masking. Using the saliency weight as a coefficient, the coding features of each residue are weighted and summed to obtain the global semantic vector that characterizes the overall functional context of the whole protein; The global semantic vector is copied along the sequence dimension and then concatenated with the encoded features of each residue along the feature dimension to obtain the residue-enhanced features.

7. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The classifier used to classify and decode the residue enhancement features is obtained during the training phase using an adaptive weighted binary cross-entropy loss function; the loss function adjusts the loss weights of positive and negative samples according to the ratio of the number of bound residues to non-bound residues in the training set.

8. The protein metal binding site prediction method based on a saliency-weighted graph transformer according to claim 1, characterized in that, The prediction model employs ligand-independent methods, including: The only input is the amino acid sequence of the protein to be predicted; no ligand structure information is required.

9. A protein metal binding site prediction system based on a saliency-weighted graph transformer, characterized in that, include: The acquisition module acquires the amino acid sequence of the protein to be predicted, wherein the amino acid sequence is composed of multiple amino acid residues in sequence; The extraction and construction module extracts the sequence evolution features and protein spatial structure features corresponding to the amino acid sequence, constructs initial node features based on the sequence evolution features, and constructs graph edge features based on the spatial structure features; The preprocessing module performs local refinement on the initial node features, filters feature redundancy and extracts local functional motifs to obtain refined node features corresponding to each residue. The feature extraction module constructs a protein residue topology graph using the refined node features as graph nodes and the graph edge features as graph edges. The graph is then input into the graph transformer encoder for spatial topology encoding to obtain the encoded features corresponding to each residue. The feature enhancement module calculates the saliency weight of the encoded feature corresponding to each residue, performs weighted aggregation on all the encoded features based on the saliency weight to obtain a global semantic vector, and feeds back the global semantic vector into the encoded feature to obtain the residue-enhanced feature; The results output module classifies and decodes the residue enhancement features and outputs the predicted results of metal ion binding sites corresponding to each residue.

10. The protein metal binding site prediction system based on a saliency-weighted graph transformer according to claim 9, characterized in that, The extraction construction module includes: The training unit extracts the sequence evolution features through a pre-trained protein language model, which is obtained based on large-scale protein sequence self-supervised training and is used to output residue-level co-evolutionary semantic embeddings.

Images