Dual-view drug-disease interaction prediction method based on relationship graph guided semantic transformer

Abstract

This invention discloses a dual-view drug-disease association prediction method based on a relation graph-guided semantic Transformer. The method acquires drug data, disease data, and known drug-disease association data, constructing a multi-source drug similarity matrix, a multi-source disease similarity matrix, and a drug-disease association matrix. Two complementary heterogeneous graph views are constructed through Top-K local neighborhood sparsification. Node features are initialized based on truncated singular value decomposition features and node degree topological priors. Head-level relation biases are generated using a relation co-occurrence graph encoder and a relation bias generator, and a relation graph-guided Transformer encoder learns drug and disease representations. Further, node-level attention fusion of basic semantics, local semantics, and global semantics, as well as cross-view attention fusion and contrastive learning alignment, yields the final node representation. Finally, drug-disease association scores are output through interaction features and a prediction network. This invention improves the robustness and predictive performance of drug-disease association prediction under sparsity, noise, and class imbalance conditions, and can be used for priority screening of candidate associations for drug relocation.

Description

Technical Field

[0001] This invention relates to the fields of biomedical informatics, artificial intelligence-assisted drug relocation, drug-disease association prediction, heterogeneous graph representation learning, and graph Transformer, and particularly to a dual-view drug-disease association prediction method based on a relational graph-guided semantic Transformer. Background Technology

[0002] Drug retargeting aims to discover new indications for existing drugs or drug candidates. Compared to developing new drugs from scratch, drug retargeting can leverage prior knowledge of existing drugs in terms of safety, pharmacokinetics, and toxicology, thereby reducing development costs, shortening development cycles, and mitigating the risk of translational failure to some extent. Drug-disease association prediction is a key computational task in drug retargeting, aiming to identify potential drug-disease associations from drug, disease, and biomedical relationship data, providing prioritization references for candidate drug screening and experimental validation.

[0003] With the continuous accumulation of drug structure databases, disease knowledge bases, and drug-disease association databases, drug-disease association prediction based on computational models has developed rapidly. Existing methods mainly include matrix completion methods, similarity fusion methods, graph neural network methods, and heterogeneous graph learning methods. However, existing technologies still have the following shortcomings:

[0004] First, multi-source drug similarity and disease similarity are often dense and noisy. If the dense similarity matrix is ​​directly converted into a near-complete graph, weak similarity or pseudo-correlation connections are easily introduced into the graph structure, leading to neighborhood distortion and noise propagation.

[0005] Second, existing graph models often fail to adequately model relation types. Drug-drug relationships, disease-drug relationships, and drug-drug relationships have different semantics. If relation-independent or coarse-grained relation encoding methods are used, it is difficult to fully characterize the contribution of different relation types to association prediction.

[0006] Third, some graph neural networks rely mainly on local message passing, making it difficult to capture long-range dependencies implied by multi-hop connections; while the standard Transformer, although capable of global modeling, is difficult to directly adapt to heterogeneous drug-disease graphs without guidance on relation types.

[0007] Fourth, existing multi-source fusion methods often employ fixed weighting, average fusion, or shallow fusion, which makes it difficult to dynamically adjust semantic contributions based on the reliability of different nodes and different views, resulting in a weakening of source-specific information.

[0008] Fifth, drug-disease association data often suffers from severe class imbalance. Known positive samples are far fewer than unknown candidate samples, and the ordinary classification loss is easily affected by the majority class samples, thus reducing the ability to identify candidate associations.

[0009] Therefore, a novel drug-disease association prediction method is needed that can simultaneously achieve reliable graph neighborhood construction, relation type awareness, local and global semantic fusion, cross-view alignment, and class imbalance handling, thereby improving the stability and predictive performance of drug-disease association prediction. Summary of the Invention

[0010] This invention aims to provide a dual-view drug-disease association prediction method, system, device, and storage medium based on a relational graph-guided semantic Transformer, to address the problems in existing technologies such as high noise from multi-source similarity, insufficient modeling of heterogeneous relationship types, difficulty in unifying local and global dependencies, coarse view fusion, and the impact of class imbalance on prediction performance. The technical solution of this invention is as follows:

[0011] First, we obtain the drug set, disease set, and known drug-disease association set to construct a drug-disease association matrix. Then, we construct two types of drug similarity matrices and two types of disease similarity matrices. The drug similarity matrix includes a drug structure similarity matrix based on molecular fingerprints and a drug Gaussian interaction spectrum similarity matrix based on drug-disease association contours; the disease similarity matrix includes a disease similarity matrix based on textual or semantic information and a disease Gaussian interaction spectrum similarity matrix based on disease-drug association contours.

[0012] Secondly, Top-K local neighborhood sparsification is performed on each similarity matrix. For each node, only the K non-nearest nodes with the highest similarity are retained, and the weights of the retained edges are normalized. This step is used to reduce spurious connections in dense similarity matrices, forming a more reliable local neighborhood structure.

[0013] Next, two complementary drug-disease heterogeneous graph views are constructed. The first view consists of drug structural similarity, disease semantic similarity, and known drug-disease associations; the second view consists of drug Gaussian interaction spectrum similarity, disease Gaussian interaction spectrum similarity, and known drug-disease associations. Each view contains drug nodes, disease nodes, and three types of relationship edges: drug-drug, disease-disease, and drug-disease.

[0014] Furthermore, prior structural node features are constructed. The truncated singular value decomposition is performed on the sparsified similarity matrix to extract low-dimensional structural features, and the logarithmic transformation features of the local out-degrees of the nodes are calculated. These two features are concatenated to form the initial node features, which are then used as trainable parameters in model optimization.

[0015] Furthermore, a relation graph-guided semantic Transformer encoder is constructed. This encoder maintains learnable relation embeddings and constructs a relation co-occurrence graph based on the co-occurrence of different relation types on nodes. The relation embeddings are updated by the relation graph encoder and then mapped to relation bias terms for each attention head by a relation bias generator. During node message aggregation, the relation bias terms are added to self-attention scoring to achieve fine-grained head-level modeling of drug-drug, disease-drug, and drug-drug relationships.

[0016] Furthermore, three types of semantic representations are extracted within each heterogeneous graph view: basic semantics, local semantics, and global semantics. Basic semantics are obtained by linear mapping of the initial features of the nodes; local semantics are obtained by encoding the local subgraph while retaining some relation types; and global semantics are obtained by encoding the complete heterogeneous graph containing all relation types. Before semantic fusion, the basic semantics are added as initial residuals to the local and global semantics; subsequently, the three types of semantic representations are dynamically fused through node-level semantic attention.

[0017] Furthermore, cross-view attention fusion is performed on the representations of the same node in the two heterogeneous graph views to obtain the final drug and disease representations. Simultaneously, a cross-view contrastive learning loss is introduced to ensure that the representations of the same node remain consistent across different views, while preserving complementary information provided by different views.

[0018] Finally, for candidate drug-disease pairs, the final drug representation, the final disease representation, and their element-wise product are concatenated into an interaction feature, which is then input into the prediction network to output the association score. During the training phase, Focal Loss is used to alleviate the imbalance between positive and negative samples, and it, together with cross-view contrastive loss, constitutes the total loss function. Attached Figure Description

[0019] Figure 1 It is a two-view drug-disease interaction prediction process based on a relational graph-guided semantic Transformer; Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the drug-disease association prediction method of this invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0021] In the data preprocessing stage, drug, disease, and known drug-disease association data from publicly available biomedical databases are integrated to construct a drug-disease association matrix. This matrix indicates whether a known association exists between a drug and a disease; a known association is denoted as 1, and no known association is denoted as 0.

[0022] In the multi-source similarity construction stage, drug similarity networks and disease similarity networks are constructed separately. For drugs, on the one hand, drug structural similarity is calculated based on drug molecular fingerprints; on the other hand, drug Gaussian interaction spectrum similarity is constructed based on the association contours of drugs in the drug-disease association matrix. For diseases, on the one hand, disease semantic similarity is constructed based on disease text semantics or disease description information; on the other hand, disease Gaussian interaction spectrum similarity is constructed based on the association contours of diseases in the drug-disease association matrix.

[0023] In the local neighborhood sparsification stage, Top-K sparsification is performed on each drug similarity matrix and disease similarity matrix. Specifically, for each row of any similarity matrix, after removing its own similarity, only the K most similar neighbor nodes are retained, the remaining similarity values ​​are set to zero, and the retained neighborhood weights are row-normalized. This step reduces the interference of weak similarity relationships and spurious connections on subsequent graph propagation.

[0024] In the dual-view heterogeneous graph construction phase, two complementary heterogeneous graph views are constructed based on the sparsified similarity matrix and the drug-disease association matrix. The first heterogeneous graph view consists of drug structural similarity edges, disease semantic similarity edges, and drug-disease association edges; the second heterogeneous graph view consists of drug Gaussian interaction spectrum similarity edges, disease Gaussian interaction spectrum similarity edges, and drug-disease association edges. Each heterogeneous graph view contains three types of relationships: drug-drug relationships, disease-disease relationships, and drug-disease relationships. If multiple relationships exist between two nodes, they are represented by relationship edges with multiple band type identifiers.

[0025] In the node feature initialization phase, instead of completely random initialization, a structural prior is introduced. First, truncated singular value decomposition is performed on the sparsified global similarity matrix to extract low-dimensional structural features of drug and disease nodes. Second, the local out-degree of each node in the heterogeneous graph is calculated, and a log(1+deg) transformation is applied to the out-degree to obtain the topological degree feature. Finally, the truncated singular value decomposition feature and the topological degree feature are concatenated to form the initial node features, which are then set as trainable parameters.

[0026] During the relationship graph construction phase, a relationship co-occurrence graph is constructed based on the co-occurrence of three types of relationships on nodes: drug-drug, disease-drug, and drug-drug. If a node connects to two types of relationships simultaneously, these two relationship types are considered to co-occur. A relationship co-occurrence adjacency matrix is ​​generated based on this co-occurrence relationship, and self-loops and row normalization are added.

[0027] In the relation bias generation stage, a learnable relation embedding is set for each relation type. The relation embedding is input into a lightweight relation graph encoder, and statistical dependencies between different relation types are propagated through a relation co-occurrence adjacency matrix to obtain the updated relation embedding. Subsequently, the updated relation embedding is mapped to the relation bias term corresponding to each attention head through a bias mapping layer.

[0028] In the graph-guided Transformer encoding stage, node features are input into a multi-layer graph-guided Transformer encoder. In each layer, for any directed edge, a query vector is generated based on the target node representation, and a key vector and value vector are generated based on the source node representation. For the relation type to which the edge belongs, the corresponding head-level relation bias is obtained from the relation bias generator and added to the dot product attention score of the query vector and key vector. Subsequently, the incoming neighbors of the target node are normalized, and the neighbor value vectors are aggregated to obtain the relation bias attention output. Each layer also includes residual connections, Dropout, layer normalization, and a feedforward network to improve training stability and non-linear expressive power.

[0029] In the semantic representation extraction stage, basic semantics, local semantics, and global semantics are extracted for each heterogeneous graph view. Basic semantics are obtained by linear projection of the initial node features. Local semantics are obtained from two local subgraphs: the drug-view local subgraph preserves drug-drug relationships and drug-disease relationships; the disease-view local subgraph preserves disease-disease relationships and drug-disease relationships. Global semantics are obtained from the complete heterogeneous graph containing all relationship types.

[0030] In the semantic fusion stage, the basic semantics are used as initial residuals and added to the local and global semantics respectively to preserve the initial identity information of the nodes and promote gradient propagation. Subsequently, attention weights of the basic, local, and global semantics are calculated using shared projection vectors and query vectors, and then weighted and summed according to the attention weights to obtain the fused node representation under each view.

[0031] In the cross-view fusion stage, for the same drug node or disease node, its fused representation in the two heterogeneous graph views is obtained respectively. The weights of different views are calculated through node-level view attention, and the two view representations are weighted and fused according to the weights to obtain the final drug node representation and the final disease node representation.

[0032] In the cross-view alignment stage, a contrastive learning loss is introduced. Representations of the same node in two views are treated as positive sample pairs; representations of different nodes in different views are treated as negative sample pairs. Through the contrastive learning loss in the form of InfoNCE, the cross-view representations of the same node become closer, while simultaneously enhancing the separability between representations of different nodes.

[0033] In the association prediction stage, for each candidate drug-disease pair, the final representations of the drug and the disease are taken, and an interaction feature is constructed. This interaction feature is composed of the drug representation, the disease representation, and their element-wise product. This interaction feature is input into a two-layer multilayer perceptron prediction network, which outputs the association logit or probability score of the candidate drug-disease pair. The unknown candidate drug-disease pairs are then ranked according to the association score to obtain the drug relocation candidate results.

[0034] During model training, positive samples are constructed from known drug-disease associations, while negative samples are sampled from unobserved drug-disease pairs. To alleviate the severe imbalance between positive and negative samples, Focal Loss is used as the classification loss, making the model pay more attention to hard-to-classify samples and minority class positive samples. Simultaneously, the cross-view contrastive learning loss and the classification loss are weighted and summed to form the overall training objective. By optimizing the overall training objective, the model achieves both stable association prediction ability and consistent cross-view representation ability.

Claims

1. A dual-view drug-disease association prediction method based on a relational graph-guided semantic Transformer, characterized in that, Includes the following steps: A. Acquire drug data, disease data, and known drug-disease association data to construct a drug-disease association matrix; B. Construct a multi-source drug similarity matrix and a multi-source disease similarity matrix based on the aforementioned drug data and disease data; C. Perform Top-K local neighborhood sparsification on the multi-source drug similarity matrix and the multi-source disease similarity matrix respectively, delete low-confidence similarity connections, and retain high-confidence neighborhood structures; D. Based on the sparsed drug similarity matrix, disease similarity matrix, and drug-disease association matrix, construct two complementary drug-disease heterogeneity graph views; E. Initialize drug node and disease node features based on global structural similarity and topological graph priors; F. Construct a relation graph to guide the semantic Transformer encoder, generate head-level attention biases for different relation types through a relation bias generator, and use the relation bias self-attention mechanism to learn the representations of drug nodes and disease nodes; G. Extract basic semantics, local semantics, and global semantics from each heterogeneous graph view, and fuse them through node-level semantic attention to obtain a fused representation within the view; H. Perform cross-view attention fusion on the node representations obtained from two heterogeneous graph views, and constrain the consistency of the same node representation in different views through cross-view comparative learning; I. For candidate drug-disease pairs, construct interaction features based on the final drug representation and the final disease representation, and output drug-disease association scores through a prediction network; J. Sort the candidate drug-disease pairs according to the drug-disease association score to obtain the candidate drug repositioning results.

2. The method according to claim 1, characterized in that, In step B, the multi-source drug similarity matrix includes a drug structure similarity matrix based on drug molecular fingerprints and a drug Gaussian interaction spectrum similarity matrix based on drug-disease association contours; wherein, the drug structure similarity matrix is ​​calculated through drug molecular fingerprints, and the similarity between any two drugs is measured using the Tanimoto coefficient; the drug Gaussian interaction spectrum similarity matrix is ​​constructed based on the row vectors of drugs in the drug-disease association matrix, and the similarity between any two drug association contours is calculated using a Gaussian kernel function.

3. The method according to claim 1, characterized in that, In steps B and C, the multi-source disease similarity matrix includes a disease similarity matrix based on text mining or disease semantic information and a disease Gaussian interaction spectrum similarity matrix based on disease-drug association contours. The Top-K local neighborhood sparsification includes: for each row of any similarity matrix, retaining the K elements with the highest similarity (excluding the diagonal elements), setting the other elements to zero, and performing row normalization on the retained elements.

4. The method according to claim 1, characterized in that, In step D, the two complementary drug-disease heterogeneous graph views include a first heterogeneous graph view and a second heterogeneous graph view; wherein, the first heterogeneous graph view is composed of drug structural similarity, disease semantic similarity, and known drug-disease associations, and the second heterogeneous graph view is composed of drug Gaussian interaction spectrum similarity, disease Gaussian interaction spectrum similarity, and known drug-disease associations; the heterogeneous graph view includes drug nodes, disease nodes, and three types of directed relation edges, namely drug-drug relation edges, disease-disease relation edges, and drug-disease relation edges.

5. The method according to claim 1, characterized in that, In step E, node feature initialization includes: performing truncated singular value decomposition on the sparsed drug similarity matrix and disease similarity matrix to extract low-dimensional structural features; calculating the local out-degree of each node and performing a logarithmic transformation on the local out-degree; and concatenating the low-dimensional structural features with the logarithmic out-degree features to form the initial node features of the drug node and the disease node.

6. The method according to claim 1, characterized in that, In step F, the relation bias generator includes a relation embedding matrix, a relation co-occurrence graph encoder, and a bias mapping layer; the relation co-occurrence graph encoder constructs a relation co-occurrence adjacency matrix based on the co-occurrence of different relation types on the same node, and updates the relation embedding through a lightweight graph encoder with residual connections; The relation bias self-attention mechanism adds the relation bias term generated by the relation bias generator to the dot product attention score of the query vector and the key vector in each attention head to guide the model to distinguish different types of drug-drug, disease-drug, and drug-drug relationships. The relation graph-guided semantic Transformer encoder includes a multi-layer relation graph-guided Transformer layer, each layer including a relation bias self-attention sub-layer, a feedforward network sub-layer, residual connections, Dropout, and a layer normalization module.

7. The method according to claim 1, characterized in that, In step G, the basic semantics are obtained by linear mapping of the initial features of the nodes; the local semantics include drug-perspective local semantics and disease-perspective local semantics, wherein the drug-perspective local semantics are obtained based on subgraphs constructed from drug-drug relationships and drug-disease relationships, and the disease-perspective local semantics are obtained based on subgraphs constructed from disease-disease relationships and drug-disease relationships; the global semantics are obtained based on a complete heterogeneous graph containing drug-drug relationships, disease-disease relationships, and drug-disease relationships; before semantic fusion, the basic semantics are added to the local semantics and global semantics as initial residuals to preserve the initial identity information of the nodes and enhance gradient propagation; the node-level semantic attention calculates semantic weights based on the representations of the basic semantics, local semantics, and global semantics, and performs a weighted summation on different semantic channels according to the semantic weights.

8. The method according to claim 1, characterized in that, In steps H, I, and J, the cross-view attention fusion calculates view weights for the representations of the same drug node or disease node in two heterogeneous graph views, and generates the final node representation based on the view weights. The cross-view contrastive learning employs InfoNCE loss to bring the representations of the same node closer together in two heterogeneous graph views and to distinguish the representations of different nodes. The interaction features are obtained by concatenating the final drug representation, the final disease representation, and the element-wise product of the final drug representation and the final disease representation. The prediction network is a two-layer multilayer perceptron, including layer normalization and GELU activation function, and outputs the association logit or association probability score of candidate drug-disease pairs. Focal Loss was used as the classification loss during model training to reduce the impact of class imbalance in positive and negative samples in drug-disease association data on model training. The total model loss was composed of a weighted average of Focal Loss and cross-view comparison loss.

9. A dual-view drug-disease association prediction system based on a relational graph-guided semantic Transformer, characterized in that, include: The system comprises the following modules: a data construction module for acquiring drug data, disease data, and known drug-disease association data, and constructing a drug-disease association matrix; a similarity construction module for constructing multi-source drug similarity matrices and multi-source disease similarity matrices; a Top-K sparsification module for performing local neighborhood sparsification on the multi-source drug similarity matrix and multi-source disease similarity matrix; a dual-view heterogeneous graph construction module for constructing two complementary heterogeneous graph views based on the sparsified similarity matrix and drug-disease association matrix; a structural prior feature initialization module for initializing drug nodes and disease nodes based on truncated singular value decomposition features and topological degree features; and a relation graph-guided Transformer encoding module for learning node representations through a relation-biased self-attention mechanism. The semantic fusion module is used to fuse basic semantics, local semantics, and global semantics; the cross-view fusion and alignment module is used to fuse dual-view representations and align representations through contrastive learning. The association prediction module is used to output the candidate drug-disease association score.

Images