Protein-dna binding site prediction method based on graph neural network and rbf-gated mechanism

Abstract

This invention discloses a protein-DNA binding site prediction method based on graph neural networks and RBF-gating mechanisms, belonging to the field of bioinformatics for predicting protein-DNA binding sites. The method includes S1, protein graph representation and feature generation; S2, input encoding and multi-scale distance representation; S3, learnable threshold gating and spatial bias attention modeling; S4, deep feature updating and model training; and S5, DNA binding site prediction output. Within the graph neural network framework, a graph Transformer module combining AlphaFold2 prediction structure and geometric inductive bias is used to perform structure-aware representation of the spatial relationships between residues, which is then used to predict the binding sites between protein residues and DNA. This framework organically integrates multi-scale radial basis function distance encoding, learnable threshold gating mechanisms, and spatial bias graph Transformer attention, enabling more comprehensive modeling of multi-scale geometric relationships and long-range dependencies in the three-dimensional structure of proteins. This results in improved accuracy in protein-DNA binding site prediction while exhibiting stronger generalization ability.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics prediction of protein-DNA binding sites, specifically a method for predicting protein-DNA binding sites based on graph neural networks and RBF-gated mechanisms. Background Technology

[0002] Early predictions of protein residue binding sites to DNA relied on experimental methods, which were typically time-consuming, technically demanding, and costly. When dealing with numerous variables or large datasets, machine learning and deep learning architectures are often combined to develop efficient and reliable computational models for high-precision DNA binding site prediction. The method entitled "A DNA-Protein Binding Site Prediction Method Based on Self-Attention Residual Network" (Publication No. CN112382338B) proceeds as follows: 1) Homology removal is performed on the ChIP-seq dataset, and the DNA sequences are encoded, converting the bases into feature vectors; 2) A global training dataset is constructed using random downsampling; 3) A deep learning network is established, consisting of convolutional layers, ELU activation function, max pooling layer, residual module, self-attention module, average pooling layer, fully connected layer, and sigmoid function. The residual structure enhances the deep feature transfer capability, and the self-attention mechanism improves the representation capability of key sequence information; 4) During training, the network is first trained using the global training dataset, and then the resulting model is transferred to the network for training; 5) The DNA sequence to be predicted is input into the transferred-trained network, and the probability of the DNA sequence binding to the protein is output through forward calculation.

[0003] The method entitled "A Protein DNA Binding Tendency Prediction Method Based on Multimodal Biomolecular Model" (Publication No. CN120356510A) is as follows: 1) Input the protein sequence information to be predicted for DNA binding tendency; 2) Use the AlphaFold3 tool to predict the three-dimensional structural information of the protein and generate relevant feature data; 3) Extract the functional annotation information and structural information of the protein; 4) Input the multimodal features of the protein into a multimodal protein language model to generate the protein embedding matrix; 5) Construct a residue-based protein graph; 6) Process the predicted feature embeddings to generate the predicted PWM matrix; 7) Use the trained model to predict the protein binding tendency to obtain the final PWM matrix.

[0004] Existing methods generally lack deep explicit modeling of the three-dimensional spatial structure of proteins, especially in their insufficient characterization of multi-scale spatial distance relationships and non-local geometric dependencies between residues, making it difficult to accurately reflect the structural basis of the real binding interface. Secondly, although some methods integrate structural prediction results or multimodal features, they mostly remain at the level of feature splicing or overall binding tendency modeling, and have not yet effectively integrated geometric information into the core prediction mechanism. Therefore, the accuracy of identifying residue-level binding sites is still limited. In addition, traditional convolution, residual, or fixed neighborhood graph modeling methods often fail to take into account both local key modes and long-range dependencies when dealing with complex protein folding structures, resulting in insufficient ability of the model to represent complex spatial interactions. Summary of the Invention

[0005] This invention provides a protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanisms, which can more fully model the multi-scale geometric relationships and long-range dependencies in the three-dimensional structure of proteins, thereby improving the accuracy of protein-DNA binding site prediction while exhibiting stronger generalization ability.

[0006] This invention provides the following technical solution: a method for predicting protein-DNA binding sites based on graph neural networks and RBF-gated mechanisms, comprising: S1. Obtain the amino acid sequence and three-dimensional structural coordinates of the target protein, and construct a protein residue layer map with each amino acid residue as a node and the spatial contact relationship between residues as an edge; at the same time, extract and fuse multi-source residue features, and calculate the Cα-Cα Euclidean distance matrix between residues, and form a structure weight matrix after analytical distance decay mapping; S2. Input the residue feature matrix obtained in S1 into the coding layer and map the original multimodal features to a unified latent space. At the same time, perform multi-scale RBF distance encoding on the distance matrix between residues, divide the distance between residue pairs into at least two spatial scales with overlapping intervals, and perform weighted fusion on the distance responses at each scale to obtain a multi-scale structural representation that can characterize the geometric proximity relationship in different spatial ranges. S3. Based on the multi-scale structural representation generated in S2, a learnable threshold gating mechanism is constructed to adaptively screen the structural weights between residues and dynamically suppress irrelevant or weakly contributing distant residue pairs. Then, the gating results and distance bias information are introduced into the multi-head self-attention module to perform weighted modeling of the relationship between residues. S4. The attention output obtained in S3 is further fed into the feedforward neural network, and the residue representation is nonlinearly transformed and iteratively updated in combination with the residual connection to obtain high-level semantic features for binding site discrimination. S5. Input the protein sample to be predicted into the trained model, and calculate the probability value of each residue belonging to the DNA binding site through the output layer and the sigmoid function.

[0007] The present invention has the following beneficial effects: This invention introduces a multi-scale RBF distance encoding method, mapping Cα-Cα Euclidean distance to learnable continuous spatial features and combining it with an analytical distance decay term to generate distance-aware structural weights. A learnable threshold gating mechanism is implemented, enabling differentiable neighborhood selection for different attention heads, thus avoiding the use of manually set distance truncation. An improved multi-head self-attention layer is used, where distance-derived biases and adaptive gates only reweight the attention weights, prioritizing spatially proximate and functionally related interactions while reducing the influence of low-information long-range residue pairs. This allows the model framework to more effectively capture DNA recognition patterns in structurally complex proteins. This significantly enhances the protein's three-dimensional geometric feature expression, neighborhood selection adaptability, and key spatial interaction recognition capabilities, enabling more accurate prediction of protein-DNA binding sites. It exhibits high prediction accuracy, robustness, and significant potential for widespread application. Attached Figure Description

[0008] Figure 1 This is a diagram illustrating the operational steps of the model of the present invention.

[0009] Figure 2 This is a performance comparison chart of the precision-recall curves of the model of this invention and GraphSite.

[0010] Figure 3 A visualization of residue features for the Test_181 dataset.

[0011] Figure 4 A visualization of the prediction results for proteins 5kbj_A and 7ad8_B. Detailed Implementation

[0012] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0013] To facilitate understanding of this invention, the following terms are now explained: Residue: A residue refers to an incomplete amino acid. A complete amino acid includes a carboxyl group (-COOH), an amino group (-NH2), an H group, and an R group. The absence of any one of these parts constitutes an amino acid residue.

[0014] RBF: Radial Basis Function, is a mathematical function whose value depends only on the distance from the input point to a certain center point.

[0015] This invention discloses a protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanisms. It is a graph Transformer model framework that combines AlphaFold2 predicted structure with geometric inductive bias to predict the binding sites of protein residues to DNA. This framework organically integrates multi-scale RBF distance encoding, a learnable threshold gating mechanism, and spatially biased graph Transformer attention, enabling more comprehensive modeling of multi-scale geometric relationships and long-range dependencies in the three-dimensional structure of proteins. This results in improved accuracy in protein-DNA binding site prediction while exhibiting stronger generalization ability. The specific steps of this method include: Step 1: Select the dataset.

[0016] First, download three publicly available benchmark datasets: Train_573 as the training set, Test_129 as the standard test set, and the independent dataset Test_181 as the external validation set. Statistical information about the datasets, including the number of proteins and the distribution of binding sites, is shown in Table 1.

[0017] Table 1. Dataset Information Description Step 2: Protein diagram representation and feature generation.

[0018] Each protein chain is represented as a residue-level graph, where residues act as nodes and the spatial relationships between residues are encoded as edges. To construct biologically meaningful node and edge representations, this invention integrates evolutionary, structural, and geometric features derived from both sequence and structural information sources.

[0019] The protein structure predicted by AlphaFold2 provides the structural basis for feature generation. It provides atomic coordinates and residue identity information, enabling the calculation of Euclidean distances between residues and the extraction of geometric descriptors. Based on these atomic coordinates, the feature generation is achieved by measuring the distances between each pair of residues. The distances were used to construct a residue-residue distance map. This matrix encodes the spatial topology of the protein and was subsequently transformed into an edge weight matrix using a differentiable analytical mapping function: , This function ensures the smooth decay of the effects of long-distance interactions. Simultaneously, this invention runs DSSP (Secondary Structure Assignment Program) on the main chain atoms (N, Cα, C, and O) to obtain the secondary structure state and main chain twist angle. In addition, solvent accessibility-related metrics were included. These features extracted by DSSP were integrated into a 14-dimensional vector for each residue to simultaneously characterize the local geometric conformation and environmental exposure properties of the residue.

[0020] To incorporate evolutionary information into residue features, this invention employs PSI-BLAST (Position-Specific Iterative BLAST) for iterative searching of protein sequences. The first round of searching is identical to BLASTp. Subsequently, a position-specific scoring matrix (PSSM) is constructed based on the hit sequences, and Hidden Markov Model (HHM) features are built using HHblits. Each residue is represented by a 20-dimensional PSSM vector and a 20-dimensional HHM vector. All features are Min-Max normalized using global extrema obtained statistically from the training set and scaled to the interval [0,1] to ensure numerical comparability between different proteins.

[0021] In addition to traditional sequence-based features, this invention introduces AlphaFold2's residue-level single representation to characterize the deep contextual embedding information of residues learned during structure prediction. Specifically, the first row of MSA embeddings is extracted, further optimized by the Evoformer module, and projected onto each residue as a 384-dimensional vector. This feature naturally encodes evolutionary covariance information and learned geometric dependencies, thus serving as a bridge connecting the sequence domain and the structural domain. To ensure consistency among different protein samples, all single representations are first standardized using pre-computed global statistics; subsequently, they are normalized to the [0,1] interval based on pre-computed global minimum and maximum values ​​to ensure scale consistency among different proteins. Finally, the normalized representations are directly incorporated into the node feature set.

[0022] After feature extraction, residue-level descriptors from four main sources—single representation, PSSM, HHM, and DSSP—are concatenated in a fixed order to form the final node feature matrix. ,in Indicates the sequence length. This represents the total feature dimension. Under the complete feature configuration, each residue ultimately corresponds to a 438-dimensional feature vector, where 384 dimensions come from the single representation, 20 dimensions from the PSSM, 20 dimensions from the HHM, and 14 dimensions from the DSSP. Correspondingly, the distance graph obtained from the protein structure defines the edge matrix. This is used to characterize the spatial geometry of proteins. Both node and edge features undergo integrity and dimensionality consistency checks to ensure they are correctly aligned with residue indexes in the sequence and tag files.

[0023] In addition to residue-level features, the three-dimensional structural information of proteins is also introduced through pairwise distance matrices. Let the... Among the residues The Cartesian coordinates of an atom are Then the residues and residues The Euclidean distance between them is defined as: , The resulting distance matrix The spatial relationships between residues were characterized and will be used to construct structure-aware interaction weights in subsequent attention-based network layers.

[0024] Step 3: Input encoding layer.

[0025] This module aims to convert raw residue-level sequence and structural information into a continuous representation suitable for downstream structure-aware modeling. Assume the target protein consists of L residues. For each residue, a multi-source continuous feature representation is constructed by splicing heterogeneous descriptors from different information sources. Specifically, the residue feature matrix is ​​defined as: , Each row corresponds to a residue, and the feature dimension is determined by default. The final feature vector integrates AlphaFold2's single representation, evolutionary feature spectrum (PSSM and HHM), and structural descriptors extracted by DSSP.

[0026] To obtain the initial potential representation of each residue, the concatenated feature matrix is ​​first normalized, then projected into a lower-dimensional embedding space via a linear mapping, followed by a nonlinear activation, as follows: , in Representation layer normalization, This represents the LeakyReLU activation function. Let be a learnable projection matrix. In this implementation, the hidden dimension is set to . .

[0027] The resulting matrix: , As the initial residue embedding matrix, each row Indicates the first The encoding feature vector of each residue.

[0028] By leveraging pre-trained representations and evolutionary feature spectra, the proposed input encoding scheme implicitly captures amino acid identity information and its contextual dependencies, thereby alleviating the need for explicit one-hot encoding. This design enables the model to start from semantically richer residue representations, supporting more effective structure-aware attention modeling in downstream layers.

[0029] Step 4: Introduce a multi-scale RBF distance coding scheme.

[0030] To incorporate three-dimensional structural information of the protein at different spatial scales, a multi-scale RBF distance encoding scheme was employed. Specifically, for each pair of residues, its position in three-dimensional space was calculated. The Euclidean distance between coordinates. Unlike directly using these raw, continuous distance values, this invention uses each... The distance is mapped to a nonlinear RBF response vector, thus providing a smooth basis function expansion for geometric proximity and facilitating the modeling of interaction patterns across multiple distance ranges.

[0031] set up This is a residue-residue distance matrix, where elements Represents residues and residues The Euclidean distance between them. Then, based on the magnitude of the distance, it is divided into three overlapping spatial intervals: short-range distance... Å, mid-range distance Å, and long-distance Å. Correspondingly, encoding is performed using 6, 8, and 6 RBF kernels respectively, resulting in dimensions of Å. , and The multi-scale distance feature maps were then constructed. Finally, the outputs of the three encoders were fused to obtain a unified distance-aware representation for each pair of residues.

[0032] For each distance interval, an independent RBF encoder is used, and a dedicated set of Gaussian kernels is configured for it. Specifically, the distance... In the The response on an RBF core is defined as: , in, Indicates the first The center of a Gaussian kernel, Used to control the bandwidth of kernel functions. Each kernel center... They are evenly distributed within the corresponding distance interval.

[0033] For a given distance scale The RBF distance encoder proposed in this invention will use paired distance tensors Convert to a scalar structure encoding Specifically, the present invention operates within a predefined distance range. Inner instantiation A Gaussian kernel, with its center uniformly distributed: And allocate one bandwidth to each core This bandwidth can be fixed (i.e., shared by multiple cores) or learnable. For a distance of... residue pairs The corresponding Gaussian RBF activation is defined as: Unlike direct use Dimensional expansion indicates that this invention further constructs a distance-dependent soft allocation mechanism with respect to the kernel function to highlight closer... The core center is defined as: , Ultimately, the encoding at a single scale is obtained by weighted aggregation of the responses of each kernel: , and all corresponding Stack them up to get the output matrix This matrix provides a smooth and scale-specific representation of the geometric proximity relationships between residue pairs.

[0034] In the implementation of this module, the short-range, medium-range, and long-range branches are defined in intervals. Å、 Å and On Å, this generates three scale-specific matrices. , and These three branches are connected by a learnable weight vector. The mixture is then fused and obtained through softmax operation. , and The final multi-scale RBF representation is obtained by weighted summation: , This representation is then integrated into the backbone network of the main model, based on the original Euclidean distance. Calculate an analytical distance decay prior: , When the RBF module is enabled, a multiplicative enhancement is applied: , in, It is a learnable scaling factor. Finally, for Masking is performed to exclude padding positions and the matrix is ​​cropped to a range of stable values, thus obtaining the distance-aware structure weight matrix used by subsequent attention mechanisms and gating operations.

[0035] Step 5: Introduce a gating mechanism based on learnable thresholds.

[0036] Not all pairwise interactions are equally important for binding site prediction, as spatially distant residues are generally unlikely to directly participate in binding. Therefore, this invention introduces a learnable threshold gating mechanism to weaken or mask contributions from less relevant, typically long-range, interactions.

[0037] The gating object of this mechanism is not the original Euclidean distance. Instead, it is the distance-aware structure weight calculated in the previous small point. .

[0038] Given First, its dynamic range is stabilized through logarithmic transformation: , Subsequently, in the effective residue pair set Perform masked normalization on top: , in Normalized matrix It will be broadcast to all attention heads, and each attention head Each is equipped with an independent learnable threshold. and a temperature parameter The default value is fixed as a buffer parameter. Subsequently, the soft gating value is defined as: , in, This represents the Sigmoid function. To prevent subsequent Numerical issues arise during computation. To encourage controllable sparsity, this invention further introduces a sparsity regularization term, making the average gated activation value approximate the target sparsity rate. : in, This is used to control the regularization strength. Finally, the gating value is injected into the self-attention's logits as an additive bias: This is completely equivalent to applying a multiplicative gate inside the softmax normalization, i.e.: The gating value can be viewed as a soft mask applied to the distance matrix to adjust the attention scores in subsequent modules. By suppressing the contributions of distant residues, this gating mechanism enables the model to focus more on relevant local and mid-range interactions, thereby reducing noise and improving prediction accuracy. This design implements a numerically stable, head-specific, and learnable neighborhood selection mechanism driven by normalized structural distance weights.

[0039] Step 6: Form a feedforward neural network with residual connections.

[0040] Following the attention sublayer, the model further introduces a position-wise feedforward neural network to independently transform the representation of each residue. This sublayer enhances the feature transformation capability at each residue position by introducing nonlinear mapping, thereby improving the model's expressive power. This feedforward module consists of two linear transformation layers separated by a LeakyReLU nonlinear activation function.

[0041] For a given residue ,set up Let represent the output vector of the attention sublayer. Then, the feedforward process is defined as follows: in, and For the first layer parameters, and These are the parameters for the second layer. This represents the dimension of the hidden layers in the feedforward network. The LeakyReLU activation function is defined as follows: ,in With a relatively small negative slope, it can preserve non-zero gradients when the input is negative, thereby improving optimization stability.

[0042] This two-stage transformation first expands the feature dimension and then maps it back to the original space, enabling each residue location to learn richer nonlinear feature interactions. Subsequently, the feedforward network output is used to... Adding the original attention representation to introduce a residual connection: This output serves as a residue in this layer. The final representation. Residual connections, while preserving information from the attention sublayer, introduce new transformation features, which helps alleviate the vanishing gradient problem and promotes stable training of deep networks.

[0043] Each layer consists of a multi-head attention module and a feedforward module, with residual connections added outside both modules. This stacked design allows the model to progressively refine the residue representation through alternating context aggregation and nonlinear transformations, while maintaining efficient gradient propagation during training.

[0044] Step 7: Set the output layer and loss function.

[0045] The output header assigns a binding site probability to each residue. After the layers are stacked, the residues The representation of is denoted as Subsequently, through a linear projection followed by a sigmoid mapping, the... Convert to predicted probabilities: in, and For output layer parameters, Represents the sigmoid function. Scalar. Represents residues This represents the predicted probability of the binding site; a higher value indicates higher model confidence. The model parameters are optimized using a binary cross-entropy objective function, which is to calculate the predicted value... With the true label of each residue Comparison: in, This represents the total number of residues, and the loss is averaged over all residues within each protein. In this formula, if a certain residue... In fact, it is the binding site (i.e. However, the model predicts its probability. If it's lower, then the first item This will result in significant penalties. Similarly, if the residues... Not a binding site (i.e.) ),but If it is higher, then the second item This will also contribute significantly to the overall loss. Therefore, minimizing the loss function... The process essentially encourages the model to output a higher probability for truly bound residues and a lower probability for non-bound residues.

[0046] In actual training, the binary cross-entropy loss is calculated and averaged over all residues, and additional regularization terms can be added if necessary. By minimizing this loss, the model parameters are continuously optimized to achieve the final residue representation. It has stronger discriminative power between binding and non-binding residues, thus enabling accurate prediction of binding sites on unseen protein structures.

[0047] The aforementioned prediction method, further enhanced by incorporating a geometric inductive bias graph Transformer modeling framework based on the AlphaFold2 prediction structure, improves the protein's 3D geometric feature expression, neighborhood selection adaptation, and key spatial interaction recognition capabilities through multi-scale RBF distance encoding, learnable threshold gating, and attention reweighting strategies. This enables more accurate prediction of protein-DNA binding sites, demonstrating high prediction accuracy, robustness, and widespread application value. These innovative components allow the model framework to more effectively capture DNA recognition patterns in structurally complex proteins. In numerous benchmark tests, this model framework outperforms other representative sequence-based and structure-based baseline methods.

[0048] The comparative experiments and results are as follows: The experiments of this invention were conducted on a GPU platform equipped with 32 GB of video memory. The model used the AdamW optimizer for parameter updates, the maximum number of training epochs was set to 15, an early stopping mechanism (patience=4) was introduced, the batch size was set to 4, and other relevant key hyperparameter configurations are shown in Table 2.

[0049] Table 2 Model Hyperparameter Settings To analyze the benefits of geometry modeling within the AlphaFold2-based model framework, this invention compares the model with GraphSite. GraphSite is a related method also built on AlphaFold2, employing a graph Transformer architecture and using the same residue features as the model in this invention. However, GraphSite uses a k-nearest neighbor masking strategy, which retains only spatially adjacent residues in attention computation and fixes the number of neighbors, making it unable to dynamically adjust the scope of attention based on the geometric characteristics of different proteins.

[0050] like Figure 2 As shown, on the three datasets, the PR curves of the model framework of this invention are generally higher than those of GraphSite, especially in high-recall regions where the advantage is more obvious. This indicates that the model of this invention, through learnable multi-scale RBF distance encoding and head-specific threshold gating, can more accurately capture medium- and long-range geometric dependencies, while effectively suppressing irrelevant long-range noise, thereby achieving stronger discriminative ability on complex proteins, especially in multi-domain long sequences in Test_181, further verifying the effectiveness of the innovative method proposed in this invention.

[0051] Next, the input features and the learned embedding representations were visualized and compared on the Test_181 dataset to evaluate the ability of the model to distinguish between bound and non-bound residues. For each target residue, its initial node representation (including AlphaFold2 single-sequence representation, PSSM, HHM, and DSSP, with a total dimension of 438) was defined as the original feature vector. The latent embedding representation was then concatenated from the outputs of each layer of the graph Transformer, resulting in a 128-dimensional vector. t-distributed random neighborhood embedding (t-SNE) was used to map the high-dimensional features to a two-dimensional space, thus allowing for a direct comparison of the distribution of bound and non-bound residues. Specific visualization results are shown below. Figure 3 As shown.

[0052] Figure 3 The A subplot shows significant overlap between DNA-binding residues (red) and non-binding residues (blue), with blurred inter-class boundaries, indicating weak discriminative power of the input feature space alone. In contrast, Figure 3 As shown in the B-subgraph, the embedded representation obtained after the model processing of this invention exhibits a clearer clustering structure, in which most binding residues can be clearly separated from a large number of non-binding residues.

[0053] This separation phenomenon demonstrates that the model of this invention can effectively learn information-rich and discriminative latent representations, thereby achieving clearer distinction between different residue classes. Through the combined effect of multi-scale RBF distance encoding and adaptive gating mechanisms, the model can enhance spatially proximate and functionally related geometric signals into key patterns in high-dimensional representations, while suppressing irrelevant long-range interference. Quantitatively, the inter-class distance in the latent embedding space is significantly increased, confirming that the model of this invention successfully transforms the input "weakly separable" features into "strongly discriminative" representations, providing a more reliable decision basis for subsequent residue classification. The visualization results further confirm the superiority of the model architecture from a representation learning perspective. This indicates that the model successfully transforms complex sequence and structural information into a latent representation space that more realistically reflects potential biochemical interactions.

[0054] This invention selected two proteins as case studies: 5kbj_A from Test_129 and 7ad8_B from Test_181, to compare representative protein-DNA binding site prediction methods. 5kbj_A is relatively short, containing only 130 residues, while 7ad8_B contains 480 residues, representing a significantly longer protein sequence. For each protein, predictions were performed using the model and the comparison method GraphSite of this invention for comparative analysis. Figure 4 The corresponding prediction results were visualized using the PYMOL tool, with the colors representing: DNA (orange), true positive (blue), false positive (red), and false negative (green).

[0055] For 5kbj_A, the model of this invention predicts 19 binding residues (TP=12, FP=7, FN=2, TN=109), with corresponding performance metrics: Spe=0.940, Rec=0.857, Pre=0.632, F1=0.727, MCC=0.699, Acc=0.931. GraphSite predicts 20 binding residues (TP=11, FP=9, FN=3, TN=107), with performance metrics: Spe=0.922, Rec=0.786, Pre=0.550, F1=0.647, MCC=0.608, Acc=0.908.

[0056] For 7ad8_B, the model of this invention predicts 12 binding residues (TP=8, FP=4, FN=22, TN=641), with corresponding performance metrics: Spe=0.994, Rec=0.267, Pre=0.667, F1=0.381, MCC=0.406, Acc=0.962. GraphSite predicts 16 binding residues (TP=6, FP=10, FN=24, TN=635), with performance metrics: Spe=0.985, Rec=0.200, Pre=0.375, F1=0.261, MCC=0.250, Acc=0.950.

[0057] Data comparison and visualization show that, compared to GraphSite, the model framework of this invention exhibits higher accuracy and better overall prediction performance for proteins of different lengths and complexities.

[0058] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A method for predicting protein-DNA binding sites based on graph neural networks and RBF-gated mechanisms, characterized in that, include: S1. Obtain the amino acid sequence and three-dimensional structural coordinates of the target protein, and construct a protein residue layer map with each amino acid residue as a node and the spatial contact relationship between residues as an edge; at the same time, extract and fuse multi-source residue features, and calculate the Cα-Cα Euclidean distance matrix between residues, and form a structure weight matrix after analytical distance decay mapping; S2. Input the residue feature matrix obtained in S1 into the coding layer and map the original multimodal features to a unified latent space. At the same time, perform multi-scale radial basis function distance encoding on the distance matrix between residues, divide the distance between residue pairs into at least two spatial scales with overlapping intervals, and perform weighted fusion on the distance responses at each scale to obtain a multi-scale structural representation that can characterize the geometric proximity relationship in different spatial ranges. S3. Based on the multi-scale structural representation generated in S2, a learnable threshold gating mechanism is constructed to adaptively screen the structural weights between residues and dynamically suppress irrelevant or weakly contributing distant residue pairs. Then, the gating results and distance bias information are introduced into the multi-head self-attention module to perform weighted modeling of the relationship between residues. S4. The attention output obtained in S3 is further fed into the feedforward neural network, and the residue representation is nonlinearly transformed and iteratively updated in combination with the residual connection to obtain high-level semantic features for binding site discrimination. S5. Input the protein sample to be predicted into the trained model, and calculate the probability value of each residue belonging to the DNA binding site through the output layer and the sigmoid function.

2. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 1, characterized in that, The extraction and processing process of the multi-source residue features described in S1 is as follows: A position-specific scoring matrix (PSSM) and a hidden Markov model (HHM) were constructed; each residue was represented by a 20-dimensional PSSM vector and a 20-dimensional HHM vector, respectively. The first row of MSA embeddings is extracted, optimized by the Evoformer module, and projected onto a 384-dimensional vector for each residue. DSSP was run on the N, Cα, C and O atoms of the protein backbone to extract secondary structure states, backbone twist angles and solvent accessibility measures to obtain a 14-dimensional vector; All features are normalized to the [0,1] interval by Min-Max based on the global extrema of the training set and then concatenated in a fixed order to form a 438-dimensional residue feature vector.

3. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 1, characterized in that, The function for the analytical distance decay mapping described in S1 is: , where d ij Let be the Euclidean distance between residue i and residue j.

4. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 1, characterized in that, The multi-scale RBF distance coding process described in S2 is as follows: The Euclidean distance between each pair of residues is divided into three overlapping spatial intervals: short-range, medium-range, and long-range. Six, eight, and six RBF Gaussian kernels are configured for encoding, respectively. The outputs of these three encoders are fused to obtain a unified distance-aware representation for each pair of residues. distance In the The response on an RBF core is defined as: ,in Indicates the first The center of a Gaussian kernel, Used to control the bandwidth of kernel functions; A distance-dependent soft allocation mechanism is constructed for the RBF kernel response at each scale and weighted aggregation is performed to obtain the distance feature map at each scale. The weights at each scale are obtained by performing a softmax operation on the learnable weight vectors, and the distance feature maps at each scale are then weighted and fused. Finally, the weights are multiplied by the analytical distance decay prior and a learnable scaling coefficient is introduced to obtain the final distance-aware structure weight matrix. .

5. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 1, characterized in that, The processing procedure of the learnable threshold gating mechanism described in S3 is as follows: Distance-aware structural weight matrix Perform a logarithmic transformation: , where ε is the minimum value to prevent numerical problems; A masked normalization process is performed on the set of effective residue pairs to obtain the normalization matrix Z. ij ; Normalize the matrix Z ij Broadcast to all attention heads, each configured with an independent learnable threshold. and fixed temperature parameters Calculate the soft gating value: ;in, This represents the Sigmoid function. To attract attention, ; Introducing a sparsity regularization term makes the average gated activation value approximate the target sparsity. The regularization term is: ,in, Used to control the strength of regularization The target sparsity; By injecting the gating value into the logits of self-attention in the form of additive bias, dynamic suppression of interactions between irrelevant residues is achieved.

6. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 1, characterized in that, The feedforward neural network described in S4 consists of two linear transformation layers separated by LeakyReLU nonlinear activation. The feedforward process is defined as follows: ， ， in This represents the output vector of the attention sublayer. and For the first layer parameters, and For the second layer parameters; Indicates the dimension of the hidden layer in the feedforward network; It has a relatively small negative slope.

7. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 6, characterized in that, Each layer of the S4 model consists of a multi-head attention module and a feedforward neural network module, with residual connections added outside these two modules. The calculation process for the residual connections is as follows: , This is the output vector of the feedforward network.

8. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 7, characterized in that, In S5, the model output layer assigns a binding site probability to each residue, after which... After the layers are stacked, the residues The representation of is denoted as ; by mapping through a linear projection followed by a sigmoid function, Convert to predicted probabilities.

9. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 8, characterized in that, The formula for calculating the predicted probability of binding sites is: ;in, and For output layer parameters, This represents the sigmoid function.

10. The protein-DNA binding site prediction method based on graph neural networks and RBF-gated mechanism according to claim 9, characterized in that, The model parameters are optimized using a binary cross-entropy objective function, which is to predict the values. With the true label of each residue Comparison: ， in, This represents the total number of residues, and the loss is the average of all residues within each protein.

Images