A drug target interaction prediction method based on graph interaction and multi-granularity fusion

Abstract

This invention proposes a drug target interaction prediction method based on graph interaction and multi-granularity fusion, belonging to the field of bioinformatics. Addressing issues such as oversmoothing of graph neural networks in heterogeneous graphs, insufficient utilization of drug-target interactions and multi-scale information, and poor adaptability of multi-branch feature splicing, an improved method is proposed. First, a deep interactive graph neural network branch is constructed on the node adaptive local smoothing features and the heterogeneous graph composed of drug-drug, target-target, and drug-target edges, realizing message passing and fusion of classification edges and introducing residual connections. Second, a dual-tower interactive branch is constructed for multi-order interactions. Third, a multi-granularity fusion branch is constructed to achieve progressive fusion of multi-scale features. Finally, the multi-branch representations are weighted and fused through a gating network, combined with the output results of a perceptron and a sigmoid function, and optimized using binary cross-entropy. This invention improves prediction performance and interpretability, and is applicable to scenarios such as drug discovery and target screening.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and relates to a method for predicting drug-target interaction (DTI) based on graph interaction and multi-granularity fusion. Background Technology

[0002] Drug-target interactions are a core relationship connecting the chemical molecular space and the protein target space, and are fundamental for lead compound screening, drug retargeting, mechanism of action analysis, and preclinical evaluation. Binding or functional associations between drug molecules and protein targets can influence signaling pathways, metabolic processes, and disease phenotypes; conversely, accurate prediction of unknown drug target pairs helps narrow down experimental searches and reduce the cost of high-throughput screening. Researchers typically rely on in vitro binding experiments, cellular validation, and animal models to obtain interaction tags. While these methods are reliable, they generally suffer from long lead times, high costs, and limited throughput, making it difficult to systematically cover the combinatorial space between a vast number of candidate drugs and the massive proteome.

[0003] To overcome experimental bottlenecks, machine learning-based prediction of drug-target interactions has gradually become a research hotspot. Early methods were mostly based on heuristics such as chemical similarity, sequence similarity, or "similar drugs act on similar targets," predicting unknown edges through label propagation, kernel methods, or shallow classifiers. These methods are simple to implement, but have limited ability to characterize nonlinear relationships and high-order topological patterns. Representative works include an integrated framework combining drug similarity networks and target similarity networks with random walks and gradient boosting decision trees; a method based on Laplace regularized least squares to propagate labels on similarity networks; and a route that integrates multi-source biological information through network embedding to predict potential interactions. Subsequently, graph neural networks and graph convolutional networks have been widely used to learn low-dimensional embeddings of drug and target nodes, modeling known interactions as edges in a graph, thereby mining potential links at the topological level; some studies have also represented molecules as atomic graphs and used graph convolution to extract structural features and then fused them with protein sequence features. At the same time, a large number of paradigms have emerged that extract features from molecular maps, SMILES sequences and protein sequences, encode them through deep networks, and then splice them together for prediction. Typical approaches include dual-path convolutional networks, attention-weighted local structures, and multi-scale attention fusion, which have reported reproducible accuracy improvements on multiple public benchmarks.

[0004] However, existing technologies still have several shortcomings. First, graph neural networks are prone to oversmoothing of node representations when stacking multiple layers of convolution, meaning that the feature vectors of different nodes gradually converge in the embedding space, leading to blurred discrimination boundaries and limited generalization performance. Second, heterogeneous graphs contain drug-drug similarity edges, target-target similarity edges, and drug-target interaction edges simultaneously. If graph structures are only used in the feature extraction stage, and only drug embeddings and target embeddings are simply concatenated in the prediction stage, it is difficult to continuously distinguish between "similar propagation" and "known drug-target interaction" messages with different semantics within the prediction network, resulting in insufficient utilization of graph structure information. Third, the association between drugs and targets is non-linear and multi-order, and relying solely on a single vector concatenation or a single interaction form is insufficient to characterize complex hierarchical patterns. Fourth, different samples may rely more on graph structure cues, explicit hierarchical interaction cues, or multi-scale fusion cues, and fixed weights or simple averaging multi-branch fusion cannot adaptively adjust the contribution of each information channel. In addition, there is a certain proportion of difficult negative samples and inter-class overlap in publicly available benchmarks. If the model's expressive power is insufficient or the fusion strategy is rigid, ranking indicators such as the area under the receiver operating characteristic curve and the area under the precision and recall curves are difficult to improve simultaneously, which limits the applicability of the method in drug prioritization and drug repurposing scenarios.

[0005] Existing work, such as NASNet-DTI, introduces Node-level Adaptive Local Smoothing (NDLS) on heterogeneous graphs. This approach mitigates oversmoothing by adaptively selecting the feature diffusion depth for each node and combines it with gradient boosting trees for classification, achieving good results on multiple benchmark datasets. While this approach fully utilizes the topology formed by drug-drug edges, target-target edges, and drug-target edges in the feature construction stage, if the prediction stage still relies primarily on node embedding and shallow classifiers, the mining of first- to higher-order interactions between drug-target pairs, multi-granularity scale information, and differentiated message passing on heterogeneous edges remains relatively limited. The structural prior emphasized by graph convolution—"interactive edges and edges of the same type should be treated differently"—is difficult to continuously and explicitly utilize within the classifier. At the algorithmic level, while a single binary cross-entropy loss can effectively drive binary classification convergence, its ranking ability and calibration can be further explored when class marginal distributions and hard-to-bear samples coexist. Fixed fusion strategies also struggle to adaptively adjust the contribution of each information channel based on the samples. In summary, the field of drug target interaction prediction urgently needs a framework that can integrate graph interaction modeling, hierarchical multi-level interactions, multi-granularity representation, and gating adaptive integration within the prediction network to improve accuracy, ranking metrics, and robustness, and provide gating and attention-level basis for subsequent interpretable analysis. Summary of the Invention

[0006] Addressing the core issues raised in the background art, such as excessive graph smoothing, insufficient utilization of heterogeneous edges, inadequate modeling of hierarchical interactions, simplistic multi-granularity fusion, and lack of adaptability in multi-branch fusion, this invention proposes a drug target interaction prediction method based on graph interaction and multi-granularity fusion. Its technical solution achieves breakthroughs through the following innovative designs: adaptive local smoothing of nodes and construction of heterogeneous graphs: On a heterogeneous graph composed of three types of edges—drug-drug, target-target, and drug-target—multi-hop diffusion is applied to node features, and the smoothing depth is adaptively determined based on the distance between the node and the global reference, resulting in more discriminative node representations and similar adjacencies, as well as drug target adjacencies; Deep Interactive Graph Neural Network (DIGNN): Within the graph convolutional layer, similar edges and drug target edges are explicitly distinguished, and message passing and fusion are performed separately. Residual connections are introduced to stack multiple layers, allowing graph structure information to continuously participate in the representation during the prediction stage. Novel; Dual-Tower Interaction Modeling (DTIM): After encoding drug and target embeddings separately, first-order element-wise interaction, second-order tensor flattening interaction, and high-order multilayer perceptron interaction are constructed, and then fused into a unified pairwise vector through multi-head self-attention; Hierarchical Multi-Granularity Fusion Network (HMFN): Drug and target sub-features are extracted across multiple hidden widths, and fusion is recursively performed from fine-grained to coarse-grained, introducing cross-granularity attention; Gated Adaptive Fusion: The above three pairwise representations are concatenated and passed through a gating network to obtain sample-level non-negative normalized weights. After weighted fusion, the interaction probabilities are output through a multilayer perceptron and a sigmoid function. These designs improve the accuracy, interpretability, and robustness of drug-target interaction prediction. The invention specifically includes the following steps: Step 1: Data Acquisition and Heterogeneous Graph Construction. Read the characteristics and interaction pairs of drugs and targets, establish drug-drug edges and target-target edges, and construct a heterogeneous graph by combining these with labeled edges.

[0007] Step 2: Node Feature and Adjacency Preprocessing. Adjacency matrix normalization; perform adaptive local smoothing of nodes; determine diffusion depth and fusion weights based on node and global reference distances; output smoothed features; and save adjacencies of the same type and drug target adjacencies.

[0008] Step 3: Deep Interactive Graph Neural Network Branch. Interact with node features and adjacent inputs via convolution, cluster similar message types and drug target interaction messages, fuse and stack the residuals to obtain node representations; extract drug and target vectors from sample indices, and concatenate and project them to obtain the first-way pairwise representation.

[0009] Step 4: Dual-Tower Interactive Modeling Branch. Drug and target features are encoded in the same dimension using dual-tower encoding. A first-order element-wise product, a second-order outer product flattening, and a high-order splicing multilayer perceptron interaction are constructed. The results are fused through self-attention and layer normalization and mapped to a second-path pair-level representation.

[0010] Step 5: Hierarchical multi-granularity fusion network branch. Extract drug and target sub-features, interact and fuse the upper-level projection results, and normalize the recursion; input the vector sequence into cross-granularity attention and pooling, and map it into a third-way pairwise representation.

[0011] Step 6: Gated fusion and classification prediction. Align and concatenate the three hierarchical representations, output weights from the gating network, and sum them to obtain the fusion vector. Then, pass it through a perceptron and a sigmoid function to output the interaction probability.

[0012] Step 7: Model Training and Validation. The model is trained using binary cross-entropy loss and the Adam optimizer, and hierarchical K-fold cross-validation is employed. The mean and standard deviation of the AUC are calculated to evaluate stability and generalization.

[0013] Accurate prediction of drug target interactions is beneficial for cost reduction and efficiency improvement, drug repositioning, and mechanism understanding; gating weights can reflect the relative dependence of samples on graph structure, hierarchical interactions, and multi-granularity information, and have a certain degree of interpretability. This invention unifies heterogeneous graphs, hierarchical interactions, and multi-scale priors into an end-to-end framework, facilitating deployment and expansion. The above describes typical beneficial effects and does not constitute a limitation on the scope or effect of protection. Attached Figure Description

[0014] Figure 1 This is a flowchart of a drug target interaction prediction method based on graph interaction and multi-granularity fusion.

[0015] Figure 2 This is a schematic diagram of a deep interactive graph neural network.

[0016] Figure 3 This is a schematic diagram of a dual-tower interactive modeling system.

[0017] Figure 4 This is a schematic diagram of a multi-layered, multi-granularity fusion network. Detailed Implementation

[0018] The implementation steps of this invention are described in detail below with reference to the accompanying drawings and embodiments. This method is implemented in a Python environment based on the PyTorch deep learning framework and can run on workstations or servers equipped with a central processing unit and a graphics processing unit. The memory and video memory capacity should meet the requirements of batch matrix operations and adjacency storage. Software dependencies include numerical computation libraries, sparse matrix libraries, and corresponding versions of the deep learning framework. The specific process includes seven steps: data acquisition and heterogeneous graph construction, node feature and adjacency preprocessing, deep interactive graph neural network branch, dual-tower interactive modeling branch, hierarchical multi-granularity fusion network branch, gated fusion and classification prediction, and model training and validation. Figure 1 A flowchart of a drug target interaction prediction method based on graph interaction and multi-granularity fusion is shown.

[0019] Step 1: Data Acquisition and Heterogeneous Graph Construction. This step is used to construct the unified input space required for all subsequent computations and to determine the supervised learning samples. Let the set of drug nodes be... The target node set is ,in , Total node set Total number of nodes Each node corresponds to 3D real-valued features, forming a node feature matrix ,in, Represents the feature matrix of all nodes; Indicates the first The feature vector of each node; Indicates the total number of nodes; Indicates the feature dimension of a single node; Indicates transpose; express Real matrix space. Define the set of labeled drug target interaction edges as follows: The negative sample edge set is obtained by sampling from the set of unlabeled drug target pairs. and satisfy To avoid index confusion, a node renumbering mapping is established. This ensures that the drug node is located within the numbering interval. The target node is located in the numbering range. The supervised sample set is defined as follows: in, Represents the set of supervised samples; For drug indexing; For target indexing; For index pairs Tags; and These represent the sets of positive and negative sample edges, respectively. Represents the union of sets; Let represent the binary label space. And satisfy . ,in, Indicates drug With target There are known interactions; This indicates that the sample is defined as a negative sample; and These represent whether the sample belongs to the positive edge set or the negative edge set, respectively. To ensure consistency between the training and evaluation distributions, [the following is used:] ... Divide into 10 tiers and record. in, Indicates the first Collection of excerpts; For indexes; Indicates intersection; Represents the empty set; Indicates any two distinct fold indices; This indicates that different folds do not overlap. And the ratio of positive to negative samples within each fold remains consistent with the overall ratio. The training-test set is defined as follows: in, Indicates the first Folding training set; Indicates the first Fold the test set; Indicates the current test index; This means that all folds except the test fold are included in the training set. This yields four types of inputs that are consistently used in subsequent steps: node features. Candidate edge set of the same type, drug target interaction edge set, and supervision sample set .

[0020] Step 2: Node Feature and Adjacency Preprocessing. This step transforms discrete edge relationships into stable graph propagation operators and constructs adaptive smoothing features for nodes. First, a similar adjacency matrix is ​​constructed. It only contains drug-drug edges and target-target edges; construct an interactive adjacency matrix. It only contains the drug target edge and performs bidirectional symmetry. For Adding a self-loop, we get in, Represents an adjacency matrix of the same type; This represents the adjacency matrix of the same type after adding self-loops; Represents the identity matrix. Defines the row-degree matrix. ,in in, Degree matrix The One diagonal element; express The Line number Column elements; This represents the total number of nodes. Based on this, a random walk normalized propagation operator is constructed. in, Represents the random walk normalized propagation operator Degree matrix The inverse. To eliminate feature scale differences, row normalization is performed on node features. For any node... : in, Representation matrix No. OK; This indicates an assignment update; Denotes the norm 1; This represents a small positive constant that avoids a denominator of zero. This is used to avoid a denominator of zero. Then, node-adaptive local smoothing is performed. The smoothing parameter is set to a fixed value. Constructing diffusion characteristic sequences: in, Indicates the first The feature matrix after the second diffusion; Represents the initial characteristic matrix; Indicates the diffusion step index; This represents the maximum number of diffusion steps. Calculate the global reference coefficient vector. : in, Represents a node The corresponding global reference coefficient; Represents a node The degree; Indicates the number of sides; This represents the total number of nodes. A reference feature is also defined. in, Indicates global reference features; This represents a row vector composed of the reference coefficients of each node; This represents the node feature matrix. For each node... Define the minimum convergent step size in, Represents a node Adaptive smoothing step size; Indicates the minimum value; Indicates the search range; Represents the L2 norm; Represents a node In the Characteristics of the step; Represents a node Reference features; This represents the convergence threshold. If the condition is never met, then let... Node smoothing output is defined as... ,in, Represents a node Smoothed features; The smoothing step size for this node; Indicates the original feature retention coefficient; For the step index in the weighted average; Indicates the previous The results of each step are averaged. All nodes are then stacked to obtain... This step outputs three items: smoothing features. Same type propagation operator Interactive adjacency These three terms remain consistent in definition throughout the entire training process.

[0021] Step 3: Deep Interactive Graph Neural Network Branch. This step performs multi-layer graph interactive updates on the nodes and extracts the graph structure representation of drug-target pairs. Let the number of network layers be... , No. Layer input is initial value In each layer, similar messages and interactive messages are computed in parallel. Similar messages: in, Indicates the first Messages of the same type at the same level; Indicates the first Layer node representation; Indicates the first Layer-specific linear weights. For interaction edges. First, construct the edge representation: in Indicates the first Layer interaction edge Side message; and Representing nodes respectively and In the Layer representation; This indicates vector concatenation; and They represent the first Weights and biases for edge message transformation. The edge representations are then written back to the endpoints and normalized according to interaction degree. For nodes... : , in, Represents a node In the Layered interaction and aggregation messages; Represents a node Degree in the interaction diagram; Indicates the first Layer interaction edge The message from the side, Indicates the first Layer interaction edge Side message; It is a stable term that prevents division by zero; Represents the set of interactive edges; the sum of the two terms represents the edges from... Departure and destination Accumulated edge messages. Merged and updated to... in, Indicates the first Candidate representation after layer fusion; and These represent the fusion transformation weights and biases, respectively. and These are messages of the same type and interactive messages; This represents the activation function. in, Indicates the first The final node representation of the layer; Represents a random deactivation regularization mapping; This indicates a dimension matching indicator function; it returns 1 if the dimensions are the same, otherwise it returns 0. These represent the feature dimensions of two adjacent layers. (Through...) Layer iterations obtained For any sample Extract the corresponding node vectors and map them to the graph branch pair level representation: in, Indicates sample The branching hierarchy of the graph; Representing drug nodes and target nodes In the Layer representation; These are the projection weights and the bias, respectively; This represents a 64-dimensional real vector space. This branch explicitly encodes both similarity propagation of the same type and drug target interaction propagation, and the residual term ensures that information is not excessively attenuated under deep propagation.

[0022] Step 4: Dual-Tower Interaction Modeling Branch. This step symmetrically encodes drug and target features and constructs first- to higher-order interaction representations. For the samples... Let the drug input vector be... The target input vector is The result obtained through the twin-tower mapping is... in, Indicates drug Input features; Indicate target point Input features; and These represent the mappings of the drug tower and the target tower, respectively. This represents the embedding of the two tower outputs. Three types of interaction features are constructed: in, Represents first-order interaction features; Indicates element-wise multiplication; These are the first-order interactive mapping weights and biases, respectively. in, Represents second-order interaction features; Indicates the outer product; Vectorization; These are the weights and biases of the second-order interactive mapping, respectively. in, Represents higher-order interaction features; Represents a higher-order nonlinear interaction mapping; This indicates the splicing of drug embedding and target embedding. Among them, Depicting the co-occurrence of dimensions, Characterizing second-order bilinear coupling, Characterizing nonlinear high-order combinations. Stacking the three elements sequentially in order of dimension: in, This represents an interaction sequence tensor composed of third-order interaction features; the semicolon ";" indicates stacking by sequence dimension. Perform self-attention fusion: in, This represents the interaction sequence after attention fusion; This represents the attention operator, where three... Each as a query ,key Sum The fusion result is then vectorized and mapped to a unified dimension. in, This represents a hierarchical representation of dual-tower interactive branches; These are the output mapping weights and biases, respectively; This indicates that the fusion sequence will be vectorized. The output of this branch is an explicit interaction representation for each drug target pair, which serves as the second input for subsequent gated fusion.

[0023] Step 5: Hierarchical Multi-Granularity Fusion Network Branch. This step is used to characterize drug target matching relationships at different semantic granularities. Assume the granularity set is fixed. in, Represents a multi-granularity dimensional set; These represent the three feature granularity widths, respectively. For any sample Extract multi-granularity features of the drug and target respectively: in, and These represent the particle size of the drug and the target, respectively. The following characteristics; and These represent feature extraction mappings at corresponding granularities; Granularity from set The value is taken from the middle. The first layer is merged into element-wise interaction: in, This represents the first layer of fused features; and These represent the drug and target characteristics at a 64-dimensional particle size, respectively. This indicates element-by-element interaction. The second and third layers use a recursive fusion method. in, Indicates the first Layer fusion features; Indicates the first Layer fusion mapping; Indicates the first Layer dimension alignment mapping; Indicates the first The granularity value corresponding to the layer; This indicates that the recursion is used for the second and third layers. This constructs the granularity sequence. For each particle size ,by For queries, perform cross-granularity attention using the full-granularity sequence as the key: in, Indicates the first Features of granularity enhanced by cross-granularity attention; Represents the query vector; Represents a key sequence; Represents a sequence of values; This represents a sequence composed of three layers of fused features. The mid-granularity attention output is taken and linearly projected onto a unified dimension to obtain a multi-granularity branch-level representation: in, This represents a multi-granularity branching hierarchy. Indicates medium-grained enhancement features; These represent the output mapping weights and biases, respectively. This branch allows local and global granular information to enter the prediction process together through recursion and attention mechanisms.

[0024] Step 6: Gated Fusion and Classification Prediction. This step performs sample-level adaptive weighting on the pairwise representations of the three parallel branches. The three features are spliced ​​together: in, Indicates sample The three-branch splicing feature; These represent the outputs of graph branches, interaction branches, and multi-granularity branches, respectively. Calculate the gating weights: in, Indicates sample The gate weight vector; Represents the normalization function; These are the gating mapping weights and biases, respectively. in, These correspond to the weights of the three branches respectively; This indicates that the weights of each branch are non-negative; This indicates that the sum of the weights of the three branches is 1. Branch merging is performed based on the weights: in, Represents the peer vector after gated fusion; The three-branch features are weighted separately. Then, the interaction probabilities are obtained through multiple layers of nonlinear mapping and a Sigmoid function. in, Indicates sample The predicted probability; Represents the Sigmoid function; Indicates the activation function of the hidden layer; These represent the weights of the three-layer linear mapping; These represent the corresponding biases. The training objective is the binary cross-entropy: in, This represents the binary cross-entropy loss; This indicates the number of samples in the current loss statistics; Indicates the sample sequence number; Indicates the true label; Indicates the predicted probability; Represented by natural constant The natural logarithm function is given by base . Parameter updates are optimized using weighted decay adaptive moment estimation, denoted as . in, This represents all learnable parameters of the model; Indicates the learning rate; This represents the adaptive moment estimation update operator; This represents the gradient of the loss with respect to the parameters; This represents the weight decay coefficient. This update is performed synchronously on all learnable parameters in each training round.

[0025] Step 7: Model Training and Validation. For each fold, update the parameters forward and backward on the training set, and calculate the AUC on the test fold; The results are calculated using the mean and standard deviation. Training loss and training set AUC curves over epochs can be recorded to observe convergence and overfitting signs. Histograms of the gating weights are used to check for degradation to a constant or excessive weighting by a single path. The number of epochs for early stopping and the validation split ratio can be set according to the data size. Repeating experiments with different random seeds can improve the stability of the results report. Output includes comma-separated files of epoch metrics, mean and variance, confusion matrix, and predicted probabilities for easy comparison with the baseline model.

[0026] Experiments have shown that the proposed method achieves an AUC of 0.9497 on the DrugBank dataset with 10-fold cross-validation, which is 1.57% higher than the NASNet-DTI method.

[0027] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting drug target interactions based on graph interaction and multi-granularity fusion, characterized in that, Includes the following steps: Step 1: Data Acquisition and Heterogeneous Graph Construction; Read drug and target features and interaction pairs, establish drug-drug edges and target-target edges, and construct a heterogeneous graph by combining labeled edges; Step 2: Node feature and adjacency preprocessing; adjacency matrix normalization; perform adaptive local smoothing of nodes, determine diffusion depth and fusion weight by the distance between nodes and global reference, output smoothing features, and save adjacency of the same type and drug target adjacency; Step 3: Deep interactive graph neural network branch; interactive convolution of node features and adjacent inputs, clustering similar type messages and drug target interaction messages, fusing and stacking residuals to obtain node representations; extract drug and target vectors from sample index, and splice and project them to obtain the first-way pairwise representation; Step 4, Dual-tower interactive modeling branch; Drug and target features are encoded in the same dimension through dual-tower encoding, and a first-order element-wise product, a second-order outer product flattening, and a high-order splicing multilayer perceptron interaction are constructed. The results are fused through self-attention and layer normalization and mapped to the second-way pair level representation. Step 5: Hierarchical multi-granularity fusion network branching; Extract drug and target features, interact and fuse the upper-level projection results, and normalize the recursion; input the vector sequence into cross-granularity attention and pooling, and map it into a third-way pair-level representation; Step 6: Gating fusion and classification prediction; The three hierarchical representations are aligned and concatenated. The weights output by the gated network are then summed to obtain the fusion vector, which is then processed by a perceptron and a sigmoid to output the interaction probability. Step 7: Model training and validation; The model is trained using binary cross-entropy loss and Adam optimizer, and hierarchical K-fold cross-validation is used. The mean and standard deviation of AUC are calculated to evaluate stability and generalization.

2. The drug target interaction prediction method based on graph interaction and multi-granularity fusion according to claim 1, characterized in that, Deep Interactive Graph Neural Network Branch: This step performs multi-layer graph interactive updates on nodes and extracts the graph structure representation of drug-target pairs; let the number of network layers be... , No. Layer input is initial value ; Messages of the same type and interactive messages are computed in parallel at each layer; messages of the same type: in, Indicates the first Messages of the same type at the same level; Indicates the first Layer node representation; Indicates the first Layer-specific linear weights; for interaction edges First, construct the edge representation: in Indicates the first Layer interaction edge Side message; and Representing nodes respectively and In the Layer representation; This indicates vector concatenation; and They represent the first Weights and biases of edge message transformations; then write the edge representation back to the endpoints and normalize it according to the interaction degree; for nodes : ,in, Represents a node In the Layered interaction and aggregation messages; Represents a node Degree in the interaction diagram; Indicates the first Layer interaction edge The message from the side, Indicates the first Layer interaction edge Side message; It is a stable term that prevents division by zero; Represents the set of interactive edges; the sum of the two terms represents the edges from... Departure and destination Accumulated edge messages; merged and updated into ,in, Indicates the first Candidate representation after layer fusion; and These represent the fusion transformation weights and biases, respectively. and These are messages of the same type and interactive messages; Indicates the activation function; in, Indicates the first The final node representation of the layer; Represents a random deactivation regularization mapping; This indicates a dimension matching indicator function; it returns 1 if the dimensions are the same, otherwise it returns 0. These represent the feature dimensions of two adjacent layers respectively; through Layer iterations obtained For any sample Extract the corresponding node vectors and map them to the graph branch pair level representation: in, Indicates sample The branching hierarchy of the graph; Representing drug nodes and target nodes In the Layer representation; These are the projection weights and the bias, respectively; This represents a 64-dimensional real vector space.

3. The drug target interaction prediction method based on graph interaction and multi-granularity fusion according to claim 1, characterized in that, Dual-tower interactive modeling branch: This step symmetrically encodes drug and target features and constructs first- to higher-order interactive representations; for samples... Let the drug input vector be... The target input vector is ; Obtained through twin-tower mapping in, Indicates drug Input features; Indicate target point Input features; and These represent the mappings of the drug tower and the target tower, respectively. Represents the two-tower output embedding; constructs three types of interaction features: in, Represents first-order interaction features; Indicates element-wise multiplication; These are the first-order interaction mapping weights and biases, respectively; in, Represents second-order interaction features; Indicates the outer product; Vectorization; These are the weights and biases of the second-order interactive mapping, respectively; in, Represents higher-order interaction features; Represents a higher-order nonlinear interaction mapping; This indicates the splicing of drug embedding and target embedding; among which, Depicting the co-occurrence of dimensions, Characterizing second-order bilinear coupling, Characterizing nonlinear high-order combinations; stacking the three elements sequentially in order of dimension: in, This represents an interaction sequence tensor composed of third-order interaction features; the semicolon ";" indicates stacking by sequence dimension; self-attention fusion is performed. in, This represents the interaction sequence after attention fusion; This represents the attention operator, where three... Each as a query ,key Sum The fusion result is then vectorized and mapped to a unified dimension. in, This represents a hierarchical representation of dual-tower interactive branches; These are the output mapping weights and biases, respectively; This indicates that the fused sequence will be vectorized; the output of this branch serves as the second input for subsequent gated fusion.

4. The drug target interaction prediction method based on graph interaction and multi-granularity fusion according to claim 1, characterized in that, Hierarchical multi-granularity fusion network branch: This step is used to characterize drug target matching relationships at different semantic granularities; Assume the particle size set is fixed as in, Represents a multi-granularity dimensional set; These represent the three feature granularity widths respectively; for any sample Extract multi-granularity features of the drug and target respectively: in, and These represent the particle size of the drug and the target, respectively. The following characteristics; and These represent feature extraction mappings at corresponding granularities; Granularity from set Values ​​are taken from the middle; the first layer is merged into element-by-element interaction: in, This represents the first layer of fused features; and These represent the drug and target characteristics at a 64-dimensional particle size, respectively. This indicates element-by-element interaction; the second and third layers use a recursive fusion method. in, Indicates the first Layer fusion features; Indicates the first Layer fusion mapping; Indicates the first Layer dimension alignment mapping; Indicates the first The granularity value corresponding to the layer; This indicates that the recursion is used for the second and third layers; thus constructing the granularity sequence. For each particle size ,by For queries, perform cross-granularity attention using the full-granularity sequence as the key: in, Indicates the first Features of granularity enhanced by cross-granularity attention; Represents the query vector; Represents a key sequence; Represents a sequence of values; This represents a sequence composed of three layers of fused features; the medium-granularity attention output is taken and linearly projected onto a unified dimension to obtain a multi-granularity branch-level representation: in, This represents a multi-granularity branching hierarchy. Indicates medium-grained enhancement features; These are the output mapping weights and biases, respectively.

Images