A Method and System for Screening Anticancer Drug Targets Based on Hypergraph-Knowledge Graph Dual Channels

Abstract

This invention relates to the fields of bioinformatics and computational biology, and provides a method and system for screening anticancer drug targets based on a hypergraph-knowledge graph dual-channel architecture. The method constructs a hypergraph-knowledge graph dual-channel architecture through a feature extraction module, dynamically integrates multi-source heterogeneous data through an adaptive multimodal fusion module, adaptively fuses high-order gene interaction features and semantic association features, and utilizes a generative adversarial network to generate high-quality negative samples to optimize the training process. Finally, a discriminator optimizes the feature representation and calculates the synthetic lethal probability score between any gene pair. This application first normalizes the extracted multi-dimensional features through the feature extraction module, then dynamically weights the importance of different features based on an attention mechanism, and finally outputs the predicted probability of the synthetic lethal relationship of gene pairs through a multilayer perceptron. Within the theoretical framework of gene synthetic lethality, this provides an innovative solution for screening anticancer drug targets.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, and particularly relates to a method and system for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph approach. Background Technology

[0002] In the fields of bioinformatics and computational biology, the prediction of synthetic lethal (SL) gene pairs has been a research hotspot in the screening of anticancer drug targets. Traditionally, the identification of SL interactions mainly relies on high-throughput experimental techniques, such as RNA interference (RNAi) and CRISPR-Cas9. These experimental methods utilize gene knockout or gene editing to inhibit the function of target genes, and determine whether a synthetic lethal relationship exists between two genes by observing the survival status of cells under different gene inactivation combinations. Although experimental techniques have high reliability and realism, they have many inherent limitations in practical applications. First, the experimental costs are high, with huge investments in equipment and reagents; second, the experimental cycle is long, making it difficult to meet the needs of large-scale, systematic screening; in addition, due to large differences in cell lines and experimental conditions, experimental results may be subject to batch effects and off-target effects, resulting in poor consistency and reproducibility. These shortcomings directly restrict the application efficiency of traditional experimental methods in the screening of anticancer drug targets.

[0003] With the rapid development of high-throughput sequencing technology and the continuous accumulation of multi-omics data, computational model-based methods for predicting synthetic lethal interactions have gradually become a research hotspot. These methods integrate multi-dimensional omics data such as gene expression profiles, mutation characteristics, and protein-protein interaction networks to construct machine learning models to predict potential synthetic lethal gene pairs, providing new insights for the screening of anticancer drug targets.

[0004] The fusion of multimodal data provides a new perspective and approach for predicting SL gene pairs, reducing the high cost and low throughput of experimental screening to some extent, while also helping to uncover potential SL interactions from massive datasets. However, it still has the following shortcomings in data fusion and negative sample construction:

[0005] In terms of data fusion, current mainstream methods for multimodal data integration mainly rely on simple linear concatenation, which directly connects feature vectors from different modalities to form a unified input feature. This method fails to fully reflect the differences and specific contributions of each modality's data to the SL mechanism. For example, gene expression data reflects the gene activity state of cells under specific conditions, while protein-protein interaction networks mainly provide information on molecular-level physical connections and signal transduction; the two are fundamentally different in a biological sense. Simple linear concatenation does not assign dynamic weights to different modalities, making it difficult to effectively improve the quality of information integration and the overall performance of the prediction model. This can lead to key information being diluted or ignored by noisy data, thus affecting the overall performance of the model.

[0006] In constructing negative samples, most existing methods typically use unverified gene pairs directly as negative samples for model training. This approach carries potential risks because some unconfirmed gene pairs may actually possess synthetic lethal interactions, even if they haven't been detected experimentally. If these gene pairs with genuine SL relationships are incorrectly classified as negative, it will not only disrupt the model's training process and interfere with its ability to determine the boundary between positive and negative samples, but may also lead to a high misclassification rate and overfitting during prediction. Therefore, the lack of rigor in the negative sample selection mechanism becomes another key factor affecting the performance of prediction models based on multimodal data fusion.

[0007] Patent publication CN113299338B discloses a method, system, terminal, and medium for predicting synthetic lethal gene pairs based on knowledge graphs. This method utilizes various biological information, such as biological processes, diseases, and pathways, from knowledge graphs to construct a network of relationships between genes, and achieves prediction of SL gene pairs through subgraph extraction and feature aggregation. In this method, by integrating subgraph features containing gene neighborhood relationships extracted from the knowledge graph, and using deep learning techniques such as graph neural networks to update and fuse gene features, it considers, to some extent, the complex biological relationships and common mechanisms between genes. However, this existing technology still faces certain limitations in multimodal information integration: on the one hand, it lacks dynamic adaptive weight allocation for the specific biological significance contained in different data modalities, resulting in some key information not being fully utilized during the splicing process; on the other hand, the practice of using gene pairs that have not been sufficiently experimentally verified as negative classes in the negative sample construction method may introduce a large amount of potential noise, further affecting the accuracy and stability of the model's prediction results.

[0008] Although existing technologies have made some progress in predicting SL gene pairs using multi-omics data and knowledge graphs, the high cost and low throughput of experimental methods, the simplicity of data fusion methods, and the lack of rigor in negative sample construction strategies still restrict the further development of this technology and its application efficiency in anticancer drug target screening. Summary of the Invention

[0009] In view of this, the present invention aims to address the problems of high experimental cost, low throughput, simplistic data fusion methods, and unrigorous negative sample construction strategies in the prediction of synthetic lethal gene pairs during anticancer drug target screening. This application discloses an anticancer drug target screening method and system based on a hypergraph-knowledge graph dual-channel architecture. By constructing a hypergraph-knowledge graph dual-channel architecture, it adaptively fuses high-order gene interaction features and semantic association features, and introduces a generative adversarial network to generate high-quality negative samples to optimize the training process, significantly improving the accuracy of synthetic lethal relationship prediction and the efficiency of target screening.

[0010] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0011] The first objective of this application is to disclose a method for screening anticancer drug targets based on a dual-channel approach of hypergraph and knowledge graph, comprising four processes: gene semantic feature extraction and hypergraph structure learning based on knowledge graph, feature optimization based on adaptive multimodal fusion, negative sample enhancement based on adversarial learning, and feature discrimination optimization based on adversarial learning.

[0012] Step S1: Gene semantic feature extraction and hypergraph structure learning based on knowledge graphs can effectively capture the complex semantic relationships and topological structures between genes, thereby providing feature representations for gene function analysis and drug target prediction.

[0013] First, the TransE knowledge representation learning model is used to embed the gene-related biomedical knowledge graph SynLethKG, extracting deep semantic feature vectors of genes. Simultaneously, based on known gene synthetic lethality relation data, a gene hypergraph network G = (V, E) is constructed, where V represents the set of gene nodes and E represents the set of hyperedges. Each hyperedge consists of a group (two or more) of gene nodes with synthetic lethality relations, used to model higher-order interactions between genes. This hypergraph data is input into a feature extraction module based on a hypergraph neural network, which aggregates neighbor node information through a multi-layer message passing mechanism, ultimately obtaining a comprehensive feature representation containing both higher-order gene interaction features and lower-order structural information.

[0014] Step S2: Feature optimization based on adaptive multimodal fusion can dynamically integrate multi-source heterogeneous gene data, thereby improving the robustness of feature representation and prediction performance.

[0015] The gene structure feature representation and semantic feature representation obtained in step S1 are input into the adaptive multimodal fusion module. This module uses an attention mechanism to dynamically learn the weight distribution of different modal features, evaluates the importance of features through a trainable parameter matrix W, and performs weighted fusion of features based on attention scores. In this way, the model can adaptively adjust the contribution of different modal features according to the needs of downstream tasks, thereby improving the discriminative ability of feature representations.

[0016] Step S3: Negative sample augmentation method based on adversarial learning. Through the adversarial training mechanism between the generator and the discriminator, high-quality negative samples with a distribution similar to the real samples are generated, thereby expanding the training dataset and improving the generalization ability of the model.

[0017] The joint features fused through multimodal analysis are embedded into the input gene feature generator, while a random noise vector z ~ N(0, I) is injected to enhance the diversity of generation. The generator learns the distribution characteristics of real samples through a multi-layer neural network to generate high-quality synthetic negative sample features. These synthetic negative samples have the following advantages: (1) they approximate the distribution boundary of real samples through an adversarial training mechanism, increasing the difficulty of identifying negative samples; (2) they avoid using unverified gene pairs as negative samples, which may introduce noise; and (3) they generate feature representations that are similar to the distribution of positive samples but have discriminative power.

[0018] Step S4: Feature discrimination optimization based on adversarial learning. Through the adversarial training mechanism between the generator and the discriminator, the discrimination ability of the feature representation is dynamically improved, thereby enhancing the generalization performance and prediction accuracy of the model.

[0019] The high-quality synthetic features h′ generated by the generator and the real sample features h are input together into the discriminator D for adversarial training. The discriminator learns the decision boundary that distinguishes between real features and generated features through a multi-layer neural network, and its objective function O can be expressed as:

[0020]

[0021] Where G represents the generator, D represents the discriminator, D(x) represents the discriminator's judgment result on the real sample x, G(z) represents the fake sample generated by the generator, z represents the random noise vector, and E x～p(x) Let E represent the expectation of the true sample x, where x is sampled from the true data distribution p(x). z～q(z) Let represent the expectation of random noise z, where z is sampled from the noise distribution q(z). Through this adversarial training mechanism, the model can achieve the following optimization objective:

[0022] (1) Accurately learn the data distribution characteristics of real samples; (2) Improve the accuracy of the discriminator in the prediction task of synthetic lethal gene pairs; (3) Enhance the robustness of the model to noise and outliers; (4) Improve the model's generalization ability on unseen data; (5) Promote the collaborative optimization of the generator and discriminator.

[0023] After the discriminator is trained, the probability P that genes x and y are synthetic lethal partners is obtained by linearly splicing the modality fusion features h_joint_x and h_joint_y of genes x and y and passing them through the discriminator.

[0024] Another objective of this invention is to propose an anticancer drug target screening system based on a hypergraph-knowledge graph dual-channel approach, comprising:

[0025] The knowledge graph feature extraction module is used to embed gene-related biomedical knowledge graphs using a knowledge representation learning model, capture semantic relationships between genes, and extract gene semantic feature representations with biological significance.

[0026] The hypergraph structure learning module is used to integrate known gene synthetic lethality data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethality relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on the hypergraph neural network, thereby obtaining a comprehensive feature representation that includes high-order gene interaction features and low-order structural information.

[0027] An adaptive multimodal fusion module is used to input the gene semantic feature representation obtained by the knowledge graph feature extraction module and the comprehensive feature representation obtained by the hypergraph structure learning module, and adopts a dynamic weighting strategy based on attention mechanism to learn the weights of each modality information through unbiased linear transformation and SoftMax function, and fuses the information of each modality to obtain joint embedding;

[0028] The generative adversarial network training module includes a generator and a discriminator. The generator adopts a feedforward neural network structure, which includes multiple linear layers, batch normalization layers and activation functions. It combines random noise of the same dimension as the input features to generate high-quality negative samples that are similar to the distribution of real samples. The discriminator is used to perform adversarial training on real samples and generated samples to optimize the dataset.

[0029] The synthetic lethal relationship prediction module is used to calculate the synthetic lethal probability score between any gene pairs by using the joint embedding output by the adaptive multimodal fusion module after the discriminator training is completed, thereby predicting the probability of potential synthetic lethal partner relationships.

[0030] This application first normalizes the extracted multi-dimensional features through a feature extraction module, then dynamically weights the importance of different features based on an attention mechanism, and finally outputs the probability prediction of the synthetic lethal relationship of gene pairs through a multilayer perceptron. Under the theoretical framework of gene synthetic lethality, it provides an innovative solution for the screening of anticancer drug targets. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0032] Figure 1 This is a flowchart of the anticancer drug target screening method based on a hypergraph-knowledge graph dual-channel according to an embodiment of the present invention;

[0033] Figure 2 This is an example diagram of a hypergraph constructed from the gene synthesis lethal data described in the embodiments of the present invention;

[0034] Figure 3 This is a diagram of the generator structure described in an embodiment of the present invention;

[0035] Figure 4 This is a structural diagram of the discriminator described in an embodiment of the present invention. Detailed Implementation

[0036] To make the technical means and objectives and effects of the present invention easier to understand, the embodiments of the present invention will be described in detail below with reference to specific illustrations.

[0037] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0038] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0039] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0040] like Figure 1 As shown, this application discloses a method for screening anticancer drug targets based on a hypergraph-knowledge graph dual-channel approach, comprising the following steps:

[0041] S1: Gene semantic feature extraction and hypergraph structure learning of knowledge graphs;

[0042] By using a knowledge representation learning model, semantic relationships between genes are captured to obtain gene semantic feature representations with biological significance. Simultaneously, known gene synthetic lethality data are integrated to construct a gene hypergraph network G = (V, E), where V represents the set of gene nodes and E represents the set of hyperedges. Each hyperedge consists of a set of gene nodes with synthetic lethality relationships. High-order interaction relationships between genes are modeled. The hypergraph data is input into a feature extraction module based on a hypergraph neural network. Neighbor node information is aggregated through a multi-layer message passing mechanism to obtain a comprehensive feature representation that includes high-order gene interaction features and low-order structural information.

[0043] S2: Adaptive Multimodal Fusion

[0044] The gene semantic feature representation and comprehensive feature representation obtained in step S1 are input into the adaptive multimodal fusion module. The adaptive multimodal fusion module adopts a dynamic weighting strategy based on the attention mechanism and dynamically adjusts the weight of each modality information through the learned parameters.

[0045] S3: Generative Adversarial Network Training

[0046] A generator is constructed to generate high-quality negative samples that are similar to the distribution of real samples by introducing random noise vectors; at the same time, a discriminator is constructed to distinguish real samples from samples synthesized by the generator; through an adversarial learning mechanism between the generator and the discriminator, the training dataset is expanded and the data distribution characteristics of real samples are learned.

[0047] S4: Synthetic Lethality Prediction

[0048] After the discriminator is trained, the structural and semantic feature representations of genes are integrated using an adaptive fusion mechanism to calculate the synthetic lethality probability score between any pair of genes.

[0049] This application discloses a dual-channel anticancer drug target screening method based on hypergraphs and knowledge graphs. By combining knowledge graphs and hypergraph neural networks, it achieves comprehensive capture of complex relationships between genes. First, a knowledge representation learning model is used to embed gene-related biomedical knowledge graphs, extracting biologically significant gene semantic features. These features reflect the associations of genes in biological processes, diseases, and signaling pathways. Simultaneously, a gene hypergraph network is constructed using known synthetic lethal data. Each hyperedge in this network consists of a group of gene nodes with synthetic lethal relationships, used to model higher-order interactions between genes. Next, the hypergraph data is input into a feature extraction module based on a hypergraph neural network. Through a multi-layer message passing mechanism, information from neighboring nodes is aggregated to obtain a comprehensive feature representation that reflects both local structure and higher-order interaction features. Finally, the semantic feature representation extracted from the knowledge graph is combined with… The comprehensive feature representation generated by the hypergraph neural network is fed into the adaptive multimodal fusion module. This module adopts a dynamic weighting strategy based on an attention mechanism to learn the importance of information from different modalities, achieving effective integration of features from each modality. Then, a generator and a discriminator are constructed using a generative adversarial network. By introducing a random noise vector z ~ N(0,I), high-quality negative samples with a distribution similar to the real samples are generated, enhancing the diversity of generation. The discriminator uses adversarial learning between real samples and generated samples to optimize the training dataset, increase the difficulty of identifying negative samples, avoid using unverified gene pairs as noise that may be introduced by negative samples, and generate feature representations with a distribution similar to positive samples but with discriminative power. Finally, after the discriminator is trained, the synthetic lethal probability score between any gene pairs is calculated using the adaptively fused feature representation, thereby achieving the prediction of potential synthetic lethal relationships and providing accurate data support for the screening of anticancer drug targets.

[0050] This application discloses a dual-channel anticancer drug target screening method based on hypergraph and knowledge graph. By fully integrating two data sources, knowledge graph and hypergraph neural network, it achieves efficient fusion of gene semantic information and structural information, effectively overcoming the limitations of traditional simple linear concatenation in the multimodal data fusion process. It utilizes an adaptive multimodal fusion module to dynamically allocate the weights of each modality, thereby ensuring that key information is not diluted and improving the accuracy of feature expression. On this basis, high-quality negative samples are constructed by generative adversarial networks, effectively reducing the noise problem introduced by the traditional direct negative sample construction method, enhancing the training stability and generalization ability of the model. This not only improves the accuracy and discrimination ability of synthetic lethal relationship prediction, but also shortens the screening cycle of anticancer drug targets and reduces the screening cost, demonstrating high practicality and promotional value.

[0051] As a preferred example of this application, in step S1, the TransE knowledge representation learning model is used to embed the gene-related biomedical knowledge graph SynLethKG, and the energy function is optimized:

[0052] E(h,r,t)=||h+rt|| (1)

[0053] Map entities and relationships in the knowledge graph to a low-dimensional vector space to extract deep semantic feature vectors of genes;

[0054] Where h represents the head entity, r represents the relation, t represents the tail entity, and |||| represents the Euclidean distance between vectors.

[0055] In the specific implementation process, the gene-related knowledge graph SynLethKG is input into TransE to extract semantic information. TransE treats relations as translation vectors from head entities to tail entities. Given a triple (head entity h, relation r, tail entity t), the goal of TransE is to make h + r ≈ t. The learning process is completed by minimizing the distance between h + r and t, ultimately obtaining a deep semantic feature vector that accurately reflects the semantic information of the genes. This deep semantic feature vector of genes serves as the input to the subsequent hypergraph structure learning and adaptive multimodal fusion module, thereby enabling the hypergraph network composed of known gene synthetic lethal data and knowledge graph data to achieve effective fusion at a higher level and providing reliable basic data support for the training of generative adversarial networks.

[0056] This approach fully leverages the advantages of the TransE model to transform the complex semantic relationships in the SynLethKG biomedical knowledge graph into a computer-processable vector form. The extracted deep semantic feature vectors can more accurately capture the semantic relationships between genes. Compared with traditional methods, it can more comprehensively reflect the function and mechanism of genes in the biomedical environment. Through continuous iterative optimization, the head entity vector plus the relation vector is made as close as possible to the tail entity vector. While ensuring that the complex semantic relationships between genes in the knowledge graph are fully captured, the model extracts deep semantic feature vectors that can accurately reflect the semantic information of genes.

[0057] As a preferred example of this application, in step S1, the combined feature of higher-order gene interaction features and lower-order structural information is represented as follows:

[0058]

[0059] Where β represents the attention weight, H ij Let h represent the hypergraph incidence matrix. i (0) This represents the node feature matrix.

[0060] In the examples of this application, synthetic lethal data from multiple data sources are integrated, and a synthetic lethal hypergraph is constructed based on the synthetic lethal data. Figure 2 The constructed hypergraph is used as an example, and then it is input into the feature extraction module based on the hypergraph neural network to extract structural information. The adjacency matrix of the hypergraph can be defined as:

[0061] A = HWH T -D v (3)

[0062] Where H represents the hypergraph's incidence matrix, W represents the hyperedge weight matrix, and D... v The degree matrix represents the nodes.

[0063] Hyperedges can connect multiple nodes, aggregate local information of nodes into the hyperedge, and then pass global information back to the nodes through the hyperedge, directly modeling the relationship between multiple nodes to extract higher-order structural information.

[0064] For a given gene pair (g i ,g j The initial node features are h i ,h j ∈R d×1 , where d is the dimension of the initial features. To extract low-order structural information from the hypergraph, an attention mechanism is applied to aggregate the features of neighboring nodes, that is:

[0065]

[0066] Where N i Let w1 represent the set of neighboring nodes of node i, w1 be the learnable parameter matrix, ReLU() represent the activation function, and h be the set of neighboring nodes of node i. i (0) a represents the aggregated node features. ij This represents the attention weight between node i and node j.

[0067] It can be calculated using the following formula:

[0068] a ij =softmax(Φ(h) i ,h j ,A))(5)

[0069] Where Φ is a single-layer feedforward neural network that can capture the non-linear relationships between node features; A represents the hypergraph adjacency matrix; and softmax() represents the activation function.

[0070] To learn the higher-order structural information of a hypergraph, it is necessary to consider the importance of different hyperedges to node features. By combining the feature representation matrix of the nodes, a comprehensive representation containing both higher-order and lower-order structural information of the hypergraph can be obtained, that is:

[0071]

[0072] Where ε represents the set of hyperedges containing node j. The attention score for the hyperedge can be calculated using the following formula:

[0073]

[0074] The parameters (w2,b2) and (w3,b3) are the weights and biases of the first and second fully connected neural networks, respectively.

[0075] In the process of extracting structural information using a hypergraph neural network, this application further applies an attention mechanism based on a single-layer feedforward neural network to dynamically assign weights to the features of the node's neighbors. This is achieved by aggregating the local information of multiple nodes within the hyperedge to the hyperedge and transmitting global information from the hyperedge back to each node. By combining the hypergraph adjacency matrix, association matrix, weight matrix, and node degree matrix, the relative importance of different hyperedges in node feature learning is modeled, ultimately forming a comprehensive representation that includes both high-order and low-order structural information. Based on this comprehensive representation, the prediction of synthetic lethal gene pairs is performed. Based on the improved scheme of hypergraph construction and attention aggregation mechanism, it can fully explore the multi-node interaction information contained in gene synthetic lethality data, avoiding information loss or aggregation bias that may be caused by relying solely on simple adjacency methods. By introducing the attention mechanism of a single-layer feedforward neural network into the hypergraph neural network to calculate the relative importance between nodes and hyperedges, it captures the gain effect brought by higher-order associations while preserving local structural features. It can better reflect the potential interactions between genes and significantly enhance the attention to key hyperedges in complex biological networks. Therefore, when the model deals with synthetic lethality prediction tasks, it can take into account both the overall impact of multi-gene joint inactivation and the differential information within each hyperedge, improving the accuracy of judging potential synthetic lethality relationships between genes.

[0076] As a preferred example of this application, in step S1, the knowledge graph is input into the TransE model to extract semantic features, and then the semantic features are input into a semantic mapper consisting of linear layers with input and output feature dimensions of 32 to obtain the semantic information of the gene. At the same time, the gene synthesis lethality relation data is input into the hypergraph neural network, and the gene synthesis lethality relation data comes from SynLethDB-v2.0.

[0077] This approach utilizes the TransE model to process knowledge graphs, extracting their semantic features. These features contain rich information about entities and relationships within the knowledge graph. These semantic features are then input into a semantic mapper composed of linear layers. The linear layers perform linear transformations on the features, mapping the high-dimensional semantic features to a specific dimensional space, preferably a 32-dimensional space, thus obtaining more representative gene semantic information. Simultaneously, gene synthesis lethality relationship data obtained from SynLethDB-v2.0 is fed into a hypergraph neural network. This network utilizes a multi-layer message passing mechanism to model and aggregate information about complex high-order interactions between genes, achieving comprehensive extraction of both local gene structure and global interaction features. By integrating gene semantic information and hypergraph structural information, a more comprehensive and accurate basis is provided for subsequent drug target screening.

[0078] As a preferred example of this application, in step S2, the importance of semantic feature vectors and structural feature vectors is evaluated using a trainable parameter matrix W, and the features are weighted and fused based on the attention score to calculate the modal attention score a. m :

[0079]

[0080] Where a m It is the weight of modal information. It's a dot product operation, W m W n These are the learnable parameters for the corresponding modes m and n, respectively. M = {structure, semantics} represents structural information and semantic information.

[0081] In the example of this application, when the gene structure feature representation and semantic feature representation are input into the adaptive multimodal fusion module in step S2, the adaptive multimodal fusion module uses an attention mechanism to dynamically learn the weight distribution of different modal features. Assume s p It is the semantic information of gene p, for gene embedding h of m∈M m ,have:

[0082]

[0083] h merge This represents the node embedding after modality fusion, used for accurate prediction of a given target gene g. p Potential synthetic lethal partner genes provide key decision support for the screening of anticancer drug targets.

[0084] Specifically, in the design of the adaptive multimodal fusion module, a trainable parameter matrix W is used to evaluate feature importance, and features are weighted and fused based on attention scores. This allows the model to adaptively adjust the contribution of different modal features according to the needs of downstream tasks. When calculating modal attention scores, corresponding formulas are used for semantic and structural feature vectors to determine the weights of different modal features in the fusion process, making the fused feature representation more robust and discriminative, thereby improving prediction performance. By adopting the above scheme, multi-source heterogeneous gene data can be dynamically integrated, effectively improving the robustness and prediction performance of feature representation. Dynamically learning the weight distribution and weighting and fusing features through the attention mechanism makes the model more flexible in adapting to downstream task requirements, further enhancing the prediction accuracy of synthetic lethal gene pairs.

[0085] As a preferred example of this application, in step S2, the gene structure information h_structure and semantic information h_semantics are input into the adaptive multimodal fusion module. The importance of different modal information is learned through an unbiased linear layer Linear, and its weights are obtained through the SoftMax function. Then, the joint embedding h_joint that automatically adapts to different modal information is obtained, which is used by the discriminator to predict the probability of synthetic lethality between a pair of genes and for model inference. In the example of this application, during adaptive multimodal fusion in step S2, input data containing gene structure information and semantic information is fed into the adaptive multimodal fusion module. This module first performs a linear transformation on the input gene structure information h_structure and gene semantic information h_semantics using an unbiased linear layer, aiming to capture key information features in both types of information. The intermediate output results reflect the importance of each modality. Then, the output of the linear layer is input into the SoftMax function, which normalizes the values ​​into a probability distribution, thereby automatically determining the weight allocation of each modality in the final fusion. This linear transformation then performs a weighted fusion of the original structure information and semantic information based on the obtained weights, thereby obtaining a unified joint embedding h_joint. This embedding simultaneously contains the physical structure features and semantic features of the gene, comprehensively reflecting the multidimensional information of the gene under different data sources, providing accurate and representative feature inputs for the subsequent discriminator to predict the probability of gene-synthetic lethality.

[0086] This method effectively overcomes the problems of information dilution or overemphasis on single information in traditional fixed-weight concatenation methods by dynamically adjusting weights, and ensures that multimodal information can be optimized and combined according to the actual situation during the fusion process, thus laying a solid data foundation and feature expression advantage for subsequent model inference.

[0087] As a preferred example of this application, in step S3, the gene structure information and semantic information extracted by the model are respectively input into a gene-based generator, and random noise with the same dimension as the information is simultaneously injected into the generator. The generator learns the distribution characteristics of real samples through a multi-layer neural network. In the example of this application, the generator adopts a feedforward neural network structure consisting of four linear layers, with a batch normalization layer and a GELU activation function between each layer, and the last layer uses a Sigmoid activation function to restrict the output to the [0, 1] interval, thereby generating high-quality negative samples that are similar to the distribution of real samples and have diversity. Specifically, Figure 3The diagram shows the structure of the generator, which consists of a sequential model containing four fully connected (Linear) layers, three batch normalization (BatchNorm1d) layers, a GELU activation function, and a Sigmoid activation function. Gene structure, semantic information, and random noise are input into the generator. After passing through the first fully connected layer, batch normalization layer, and GELU activation function, the output dimension of the features becomes 256. After passing through the second fully connected layer, batch normalization layer, and GELU activation function, the output dimension of the features becomes 128. After passing through the third fully connected layer, batch normalization layer, and GELU activation function, the output dimension of the features becomes 64. Finally, after being input into the last fully connected layer and Sigmoid activation function, the output feature dimension is the same as the feature dimension extracted by the model, thereby enhancing the diversity of the generated samples. This generator structure enables the generation of high-quality negative samples, or "false features" of genes, to expand the training dataset and avoid data uniformity. These generated negative samples are then input into the trained adaptive modality fusion module for subsequent adversarial training of the model. During adversarial training, the model continuously learns and optimizes, thereby improving its performance and enabling it to more accurately predict synthetic lethal gene pairs.

[0088] As a preferred example of this application, in step S4, through the adversarial training mechanism between the generator and the discriminator, the high-quality synthetic features h′ generated by the generator and the real sample features h are jointly input into the discriminator for adversarial training. The discriminator learns the decision boundary for distinguishing between real features and generated features through a multi-layer neural network, and its objective function is expressed as:

[0089]

[0090] Where G represents the generator, D represents the discriminator, D(x) represents the discriminator's judgment result on the real sample x, G(z) represents the fake sample generated by the generator, z represents the random noise vector, and E x～p(x) Let E represent the expectation of the true sample x, where x is sampled from the true data distribution p(x). z～q(z) Let q(z) represent the expectation of random noise z, where z is sampled from the noise distribution q(z).

[0091] The anticancer drug target screening method based on a hypergraph-knowledge graph dual-channel approach described in this application employs a feature discrimination optimization method based on adversarial learning. A generator produces high-quality synthetic features, which are then input into a discriminator along with real sample features for adversarial training. This adversarial training mechanism continuously optimizes the generator and discriminator, enabling the model to more accurately predict synthetic lethal gene pairs. By adopting this method, the model's ability to discriminate features is enhanced, allowing it to better capture data features, improve prediction accuracy, and maintain high generalization performance even when faced with new data, providing more reliable results for predicting synthetic lethal gene pairs.

[0092] As a preferred example of this application, in step S4, based on adversarial learning-based feature discrimination optimization, a high-quality synthetic feature is generated by a generator and input into the discriminator along with the real sample feature for adversarial training. The discriminator learns the decision boundary that distinguishes between real features and generated features through a multi-layer neural network, and combines the features h of gene p and gene q. p ,h q The input is concatenated in the discriminator. The concatenated vector is then processed by a fully connected layer, a ReLU activation function, and a Sigmoid activation function to obtain the probability that gene p and gene q are synthetic lethal partners.

[0093] This scheme constructs an adversarial discriminant process by comparing real sample features with synthetic features generated by the generator. The real sample features extracted by the model and the negative sample features generated by the generator are input into the discriminator for adversarial training; the features h of genes p and q... p ,h q The inputs are fed into the discriminator for splicing and similarity scoring. The similarity score, provided by the discriminator, represents the probability that a pair of genes are synthetic lethal partners.

[0094] In the example of this application, the concatenation operation is a general linear concatenation; the discriminator consists of two fully connected layers, a ReLU activation function, and a Sigmoid activation function, and the linearly concatenated vector can be represented as:

[0095] h con =h p +h q (11)

[0096] The features of genes p and q are linearly concatenated to form a joint vector. This vector contains mixed information from real and fabricated data. The concatenated vector is then input into a discriminator, which performs a fully connected operation on it and uses the ReLU activation function to reduce the feature dimension and perform a non-linear transformation. The feature dimension is changed from twice that of the real sample to the same as the real sample's feature dimension. Then, a fully connected layer maps the feature dimension to 1 dimension, and the Sigmoid activation function maps the output to the probability interval [0, 1]. This yields the similarity score between genes p and q, representing the probability that genes p and q are synthetic lethal partners. This value characterizes the confidence that gene pair (p, q) constitutes a synthetic lethal partner relationship and is used for subsequent anticancer drug target screening. It can be expressed as:

[0097]

[0098] Where φ represents the ReLU activation function. This indicates a splicing operation.

[0099] Based on this adversarial learning framework, the model can more accurately identify whether there is a synthetic lethal relationship between gene p and gene q through the discriminator. This not only significantly improves prediction accuracy but also makes it more robust in dealing with potential noise and data imbalance. The adversarial training process allows the discriminator to continuously compete with high-quality negative samples generated by the generator, prompting the discriminator to learn to capture key feature differences and make rigorous judgment decisions accordingly. At the same time, the generator continuously corrects its method of generating synthetic features during the trial and error process, making it closer and closer to the real data. This gradually improves the model's understanding of various patterns and its characterization of complex interaction information, thereby accurately determining whether a gene is a synthetic lethal partner, improving the model's prediction accuracy and generalization performance, and providing a more reliable basis for the screening of anticancer drug targets.

[0100] Finally, by linearly splicing the modal fusion features h_joint_x and h_joint_y of genes x and y and passing them through a discriminator, the probability P that genes x and y are synthetic lethal partners can be obtained.

[0101] In the examples used in this application, the dataset used is SynLethDB 2.0. SynLethDB 2.0 is the first comprehensive database containing a large number of synthetic lethal (SL) gene pairs, collected from various sources including biochemical analysis, other relevant databases, computation, prediction, and text mining. SynLethDB 2.0 covers humans and four model species: mice, fruit flies, worms, and yeast. The current version contains 50,868 synthetic lethal pairs. The current version of SynLethDB 2.0 also includes a knowledge graph containing 11 entities and 27 relationships.

[0102] This application discloses a method for screening anticancer drug targets based on a hypergraph-knowledge graph dual-channel approach. It employs a hypergraph-knowledge graph dual-channel structure, an adaptive multimodal fusion mechanism, and an adversarial learning-based feature discrimination optimization method. First, it uses the TransE model to convert complex semantic relationships in a gene-related biomedical knowledge graph into low-dimensional deep feature vectors. After linear layer mapping, refined gene semantic information is extracted. Simultaneously, it integrates gene synthesis lethality data from SynLethDB-v2.0 to construct a hypergraph network. Within this network, a hypergraph neural network dynamically aggregates high-order interactions and local structural information between genes through multi-layer message passing and attention mechanisms. The method then employs an unbiased linear layer combined with the SoftMax function to adaptively learn the importance of structural and semantic information and generate a unified joint embedding. Next, a generator injects random noise with the same dimension as the original features into the input real sample features to generate high-quality negative samples that closely resemble the real distribution. These negative samples are then input into the discriminator along with the real samples. The discriminator uses linear concatenation, fully connected layers, ReLU, and sigmoid activation functions for nonlinear feature dimensionality reduction and mapping, constructing a decision boundary that accurately distinguishes between positive and negative samples. This allows for continuous optimization of the generator and discriminator performance during adversarial training, enabling the model to more accurately capture potential synthetic lethal relationships between genes. This method not only fully integrates multi-source heterogeneous data, overcoming the information dilution and noise introduction problems easily caused by fixed-weight concatenation in traditional data fusion, but also improves the robustness and generalization ability of the model in negative sample construction through adversarial training. This significantly improves the prediction accuracy, stability, and execution efficiency of anticancer drug target screening, providing a computational tool for anticancer drug development.

[0103] This application also discloses an anticancer drug target screening system based on a hypergraph-knowledge graph dual-channel approach, including:

[0104] The knowledge graph feature extraction module is used to embed gene-related biomedical knowledge graphs using a knowledge representation learning model, capture semantic relationships between genes, and extract gene semantic feature representations with biological significance.

[0105] The hypergraph structure learning module is used to integrate known gene synthetic lethality data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethality relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on the hypergraph neural network, thereby obtaining a comprehensive feature representation that includes high-order gene interaction features and low-order structural information.

[0106] An adaptive multimodal fusion module is used to input the gene semantic feature representation obtained by the knowledge graph feature extraction module and the comprehensive feature representation obtained by the hypergraph structure learning module, and adopts a dynamic weighting strategy based on attention mechanism to learn the weights of each modality information through unbiased linear transformation and SoftMax function, and fuses the information of each modality to obtain joint embedding;

[0107] The generative adversarial network training module includes a generator and a discriminator. The generator adopts a feedforward neural network structure, which includes multiple linear layers, batch normalization layers and activation functions. It combines random noise of the same dimension as the input features to generate high-quality negative samples that are similar to the distribution of real samples. The discriminator is used to perform adversarial training on real samples and generated samples to optimize the dataset.

[0108] The synthetic lethal relationship prediction module is used to calculate the synthetic lethal probability score between any gene pairs by using the joint embedding output by the adaptive multimodal fusion module after the discriminator training is completed, thereby predicting the probability of potential synthetic lethal partner relationships.

[0109] This invention screens anticancer drug targets based on synthetic lethal mechanisms, achieving drug target prediction by integrating multimodal synthetic lethal data. The invention includes a feature extraction module (knowledge graph feature extraction module, hypergraph structure learning module), an adaptive modality fusion module, a generative adversarial network module, and a synthetic lethal relationship prediction module. First, gene structural features containing high-order and low-order relationships are extracted based on a hypergraph neural network, while semantic features of genes are extracted based on a biological knowledge graph. Then, the adaptive modality fusion module dynamically adjusts the weights of each modality based on learned parameters, achieving rational utilization of each modality's information. Next, the generative adversarial network module achieves collaborative optimization through an adversarial learning mechanism between the generator and discriminator; the generator learns the data distribution characteristics of real samples by introducing random noise vectors, generating high-quality negative samples with a distribution similar to real samples; simultaneously, the discriminator distinguishes between real samples and samples synthesized by the generator, improving the accuracy of sample recognition. During training, the generator and discriminator continuously optimize through adversarial training, ultimately reaching a dynamic equilibrium and completing iterative updates of the model parameters. Finally, after the discriminator is trained, the synthetic lethality prediction module integrates the structural and semantic feature representations of genes through an adaptive fusion mechanism to calculate the synthetic lethality probability score between any gene pair.

[0110] Specifically, the anticancer drug target screening system based on a hypergraph-knowledge graph dual-channel approach described in this application first normalizes the extracted multi-dimensional features through a feature extraction module, then dynamically weights the importance of different features based on an attention mechanism, and finally outputs the predicted probability of synthetic lethality relationships of gene pairs through a multilayer perceptron. This invention provides an innovative solution for anticancer drug target screening within the theoretical framework of gene synthetic lethality.

[0111] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using hardware related to computer programs. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, FLASH, floppy disk, magnetic disk or optical disk, hard disk, and cloud computing. The computer processor can be a central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP), or field-programmable gate array (FPGA), meaning that the steps of the anticancer drug target screening method based on a hypergraph-knowledge graph dual-channel approach can be implemented using the aforementioned processor.

[0112] In summary, this invention presents a method and system for screening anticancer drug targets based on a hypergraph-knowledge graph dual-channel approach. This approach comprehensively captures complex interactions between genes using a hypergraph-knowledge graph dual-channel method, and effectively integrates multi-source information using an adaptive multimodal fusion module. Simultaneously, a generative adversarial network training module is used to construct high-quality negative samples to optimize the training dataset. With the support of a discriminator, the synthetic lethal probability between gene pairs is accurately calculated. This not only overcomes the high cost, long cycle, and inconsistency problems of traditional high-throughput experimental methods, but also effectively solves the problems of insufficient feature expression and training instability caused by fixed-weight splicing and the introduction of noisy data in existing multimodal data fusion methods during information fusion and negative sample construction. This significantly improves the efficiency and accuracy of anticancer drug target screening, providing a scientific, precise, and reliable computational tool for anticancer drug development.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for screening anticancer drug targets based on a dual-channel approach of hypergraph and knowledge graph, characterized in that, Includes the following steps: S1: Through a knowledge representation learning model, capture the semantic relationships between genes to obtain gene semantic feature representations with biological significance; at the same time, integrate known gene synthesis lethal data to construct a gene hypergraph network, input the hypergraph data into the feature extraction module based on the hypergraph neural network, and aggregate neighbor node information through a multi-layer message passing mechanism to obtain a comprehensive feature representation containing high-order gene interaction features and low-order structural information. S2: Input the gene semantic feature representation and comprehensive feature representation obtained in step S1 into the adaptive multimodal fusion module. The adaptive multimodal fusion module adopts a dynamic weighting strategy based on the attention mechanism and dynamically adjusts the weight of each modality information through the learned parameters. In step S2, a trainable parameter matrix is ​​used. Importance is evaluated for semantic and structural feature vectors, and features are weighted and fused based on attention scores to calculate modality. Attention weights : (8) Modal set , representing structural information and semantic information respectively. It's a dot product operation. For the mode whose weights are currently to be calculated, To sum the modal indices for traversal, , These are the corresponding modes ( The learnable parameters of ) , They represent the corresponding modes ( ) eigenvectors, Represents the hyperbolic tangent function; The gene structure information h_structure and semantic information h_semantics are input into the adaptive multimodal fusion module. The importance of different modal information is learned through an unbiased linear layer Linear, and then its weights are obtained through the SoftMax function. This results in the joint embedding h_joint that automatically adapts to different modal information, which is used by the discriminator to predict the probability of synthetic lethality between a pair of genes and for model inference. S3: Construct generators and discriminators, introduce random noise vectors to generate high-quality negative samples that are similar to the distribution of real samples, and expand the training dataset and learn the data distribution characteristics of real samples through the adversarial learning mechanism of generators and discriminators. S4: After the discriminator is trained, the structural and semantic feature representations of genes are integrated using an adaptive fusion mechanism to calculate the synthetic lethality probability score between any pair of genes.

2. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 1, characterized in that, In step S1, the TransE knowledge representation learning model is used to embed the gene-related biomedical knowledge graph SynLethKG, and the energy function is optimized: (1) Map entities and relationships in the knowledge graph to a low-dimensional vector space to extract deep semantic feature vectors of genes; in Indicates the head entity. Indicates a relationship. Indicates the tail entity. This represents the Euclidean distance between vectors.

3. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 2, characterized in that, In step S1, the combined feature of high-order gene interaction features and low-order structural information is represented as follows: (2) in, Indicates attention weights, Represents the hypergraph incidence matrix. This represents the node feature matrix.

4. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 3, characterized in that, In step S1, the knowledge graph is input into the TransE model to extract semantic features, and then the semantic features are input into a semantic mapper consisting of linear layers with input and output feature dimensions of 32 to obtain the semantic information of the gene. At the same time, the gene synthesis lethality relationship data is input into the hypergraph neural network, and the gene synthesis lethality relationship data comes from SynLethDB-v2.

0.

5. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 1, characterized in that, In step S3, the gene structure information and semantic information extracted by the model are input into the gene-based generator, and random noise with the same dimension as the information is injected into the generator. The generator learns the distribution characteristics of real samples through a multi-layer neural network.

6. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 1, characterized in that, In step S4, the high-quality synthetic features generated by the generator are trained through an adversarial training mechanism between the generator and the discriminator. Features of real samples Adversarial training is performed in a common input discriminator. The discriminator learns the decision boundary to distinguish between real and generated features through a multi-layer neural network. Its objective function is expressed as: (10) in, Represents a generator. Indicates the discriminator, This indicates that the discriminator is effective against real samples. The judgment result, This represents fake samples generated by the generator. Represents a random noise vector. Indicates the real sample The expectations, among which From the distribution of real data Sampling from the middle, Indicates random noise The expectations, among which From noise distribution Sampling.

7. The method for screening anticancer drug targets based on a dual-channel hypergraph-knowledge graph according to claim 6, characterized in that, In step S4, feature discrimination optimization based on adversarial learning is performed. A generator produces high-quality synthetic features, which are input into the discriminator along with real sample features for adversarial training. The discriminator learns the decision boundary to distinguish between real and generated features through a multi-layer neural network, thus integrating the features of gene p and gene q. The input is concatenated in the discriminator. The concatenated vector is processed by a fully connected layer, a ReLU activation function, and a Sigmoid activation function to obtain the probability that gene p and gene q are synthetic lethal partners. After the discriminator is trained, the probability P that genes x and y are synthetic lethal partners is obtained by linearly splicing the modality fusion features h_joint_x and h_joint_y of genes x and y and passing them through the discriminator.

8. A dual-channel anticancer drug target screening system based on hypergraph-knowledge graph, characterized in that, The method for performing the anticancer drug target screening method based on a hypergraph-knowledge graph dual-channel approach as described in any one of claims 1 to 7 includes: The knowledge graph feature extraction module is used to embed gene-related biomedical knowledge graphs using a knowledge representation learning model, capture semantic relationships between genes, and extract gene semantic feature representations with biological significance. The hypergraph structure learning module is used to integrate known gene synthetic lethality data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethality relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on the hypergraph neural network, thereby obtaining a comprehensive feature representation that includes high-order gene interaction features and low-order structural information. An adaptive multimodal fusion module is used to input the gene semantic feature representation obtained by the knowledge graph feature extraction module and the comprehensive feature representation obtained by the hypergraph structure learning module, and adopts a dynamic weighting strategy based on attention mechanism to learn the weights of each modality information through unbiased linear transformation and SoftMax function, and fuses the information of each modality to obtain joint embedding; The generative adversarial network training module includes a generator and a discriminator. The generator adopts a feedforward neural network structure, which includes multiple linear layers, batch normalization layers and activation functions. It combines random noise of the same dimension as the input features to generate high-quality negative samples that are similar to the distribution of real samples. The discriminator is used to perform adversarial training on real samples and generated samples to optimize the dataset. The synthetic lethal relationship prediction module is used to calculate the synthetic lethal probability score between any gene pairs by using the joint embedding output by the adaptive multimodal fusion module after the discriminator training is completed, thereby predicting the probability of potential synthetic lethal partner relationships.

Images