Anticancer drug target screening method and system based on hypergraph-knowledge graph double channels

Abstract

The invention relates to the technical field of bioinformatics and computational biology, and provides an anti-cancer drug target screening method and system based on hypergraph-knowledge graph dual-channel, a hypergraph-knowledge graph dual-channel architecture is constructed through a feature extraction module, a self-adaptive multi-modal fusion module dynamically integrates multi-source heterogeneous data, and the anti-cancer drug target screening method and system based on the hypergraph-knowledge graph dual-channel architecture are obtained. The method comprises the following steps: adaptively fusing high-order interaction features and semantic association features of a gene, generating a high-quality negative sample by using a generative adversarial network to optimize a training process, finally optimizing feature representation through a discriminator, and calculating a synthetic lethal probability score between any gene pair. According to the application, firstly, the extracted multi-dimensional features are normalized through the feature extraction module, then the importance of different features is dynamically weighted based on the attention mechanism, finally, the synthesis lethal relationship prediction probability of the gene pair is output through the multi-layer sensor, and under the theoretical framework of the gene synthesis lethal effect, the synthesis lethal effect of the gene pair is predicted. And an innovative solution is provided for anti-cancer drug target screening.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioinformatics and computational biology, and in particular relates to a method and system for screening anticancer drug targets based on a hypergraph-knowledge graph dual channel. Background Art

[0002] In the fields of bioinformatics and computational biology, the prediction technology of synthetic lethality (SL) gene pairs has always 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 technology. These experimental methods use gene knockout or gene editing to inhibit the function of the target gene, and by observing the survival state of cells under different gene inactivation combinations, it is determined whether the two genes have a synthetic lethal relationship. Although the experimental technology has high reliability and authenticity, it has many inherent limitations in practical applications. First, the experimental cost is high, and the investment in equipment and reagents is huge; second, the experimental cycle is long, which is difficult to meet the needs of large-scale and systematic screening; in addition, due to the large differences in cell lines and experimental conditions, the experimental results may have batch effects and off-target effects, resulting in poor consistency and reproducibility of the results. 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 synthetic lethal interaction prediction methods have gradually become a research hotspot. These methods integrate multi-dimensional omics data such as gene expression profiles, mutation characteristics, and protein interaction networks to build machine learning models to predict potential synthetic lethal gene pairs, providing new ideas for the screening of anticancer drug targets.

[0004] The fusion of multimodal data provides a new perspective and means for SL gene pair prediction, which reduces the high cost and low throughput of experimental screening to a certain extent, and helps to mine potential SL interactions from massive data. However, it still has the following shortcomings in data fusion and negative sample construction:

[0005] In terms of data fusion, the current mainstream methods for multi-modal data integration mainly rely on simple linear concatenation, that is, directly connecting the feature vectors of different modalities to form a unified input feature. This method fails to fully reflect the differences and specific contributions of each modality data in the SL mechanism. For example, gene expression data can reflect the gene activity state of cells under specific conditions, while the protein interaction network mainly provides physical connections and signal transduction information at the molecular level. There are essential differences between the two in biological significance. The simple linear concatenation method does not assign dynamic weights to different modality data, making it difficult to effectively improve the quality of information integration and the comprehensive performance of the prediction model. As a result, key information may be diluted or ignored by noise data, thereby affecting the overall performance of the model.

[0006] In terms of negative sample construction, most existing methods usually directly use gene pairs that have not been experimentally verified as negative samples for model training. This processing method has potential risks because some gene pairs that have not been experimentally confirmed may actually have synthetic lethal interactions, but have not been detected by experiments. If these gene pairs with real SL relationships are wrongly classified into the negative class, it will not only disrupt the model training process, interfere with the determination of the boundary between positive and negative samples by the model, but also may lead to a high misjudgment rate and overfitting phenomenon during the model prediction process. The lack of rigor in the negative sample screening mechanism has thus become another key factor affecting the performance of the prediction model based on multi-modal data fusion.

[0007] The published literature with the patent publication number CN113299338B discloses a method, system, terminal and medium for predicting synthetic lethal gene pairs based on a knowledge graph. It constructs an association network between genes by using various biological information such as biological processes, diseases and pathways in the knowledge graph, and realizes the prediction of SL gene pairs through methods such as subgraph extraction and feature aggregation. In this method, by integrating the subgraph features containing gene neighborhood relationships extracted from the knowledge graph, and using deep learning technologies such as graph neural networks to update and fuse gene features, the complex biological associations and common mechanisms between genes are considered to a certain extent. However, this existing technology still faces certain limitations in multi-modal information integration: on the one hand, it lacks the assignment of dynamic adaptive weights to the specific biological significance contained in different data modalities, resulting in insufficient utilization of some key information during the concatenation process; on the other hand, the practice of using gene pairs that have not been fully 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 prediction results.

[0008] Although the existing technologies have made some progress in predicting synthetic lethal (SL) gene pairs by 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 technologies in this field and the application efficiency in anti-cancer drug target screening. Summary of the Invention

[0009] In view of this, the present invention aims to solve the problems such as high cost of experimental methods, low throughput, simplicity of data fusion methods, and lack of rigor in negative sample construction strategies in predicting synthetic lethal gene pairs in anti-cancer drug target screening. This application discloses an anti-cancer drug target screening method and system based on a hypergraph-knowledge graph dual-channel, which adaptively fuses gene high-order interaction features and semantic association features by constructing a hypergraph-knowledge graph dual-channel architecture, and introduces a generative adversarial network to generate high-quality negative samples to optimize the training process, significantly improving the prediction accuracy of synthetic lethal relationships and the efficiency of target screening.

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

[0011] The first object of this application is to disclose an anti-cancer drug target screening method based on a hypergraph-knowledge graph dual-channel, including four processes: gene semantic feature extraction and hypergraph structure learning based on a knowledge graph, feature optimization based on adaptive multi-modal fusion, negative sample enhancement based on adversarial learning, and feature discrimination optimization based on adversarial learning, where:

[0012] Step S1: Gene semantic feature extraction and hypergraph structure learning based on a knowledge graph 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 perform embedding representation on the gene-related biomedical knowledge graph SynLethKG to extract the deep semantic feature vectors of genes. At the same time, based on the known gene synthetic lethal relationship 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 is composed of a group (two or more) of gene nodes with synthetic lethal relationships, which is used to model the high-order interaction relationships between genes. The hypergraph data is input into the feature extraction module based on a hypergraph neural network, and the neighbor node information is aggregated through a multi-layer message passing mechanism, and finally a comprehensive feature representation containing gene high-order interaction features and low-order structure information is obtained.

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

[0015] Input the gene structure feature representation and semantic feature representation obtained in step S1 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 on the features based on the attention scores. In this way, the model can adaptively adjust the contribution degrees of different modal features according to the requirements of downstream tasks, thereby enhancing the discriminative ability of the feature representation.

[0016] Step S3: The negative sample enhancement method based on adversarial learning uses the adversarial training mechanism of the generator and the discriminator to generate high-quality negative samples similar to the real sample distribution, thereby expanding the training data set and enhancing the generalization ability of the model.

[0017] Input the jointly fused features after multimodal fusion into the gene feature generator (Generator), and inject a random noise vector z ∼ N(0, I) at the same time to enhance the generation diversity. The generator learns the distribution characteristics of real samples through a multi-layer neural network and generates high-quality synthetic negative sample features. These synthetic negative samples have the following advantages: (1) Approaching the real sample distribution boundary through the adversarial training mechanism, increasing the discrimination difficulty of negative samples; (2) Avoiding the noise that may be introduced by using unvalidated gene pairs as negative samples; (3) Generating discriminative feature representations similar to the positive sample distribution.

[0018] Step S4: The feature discrimination optimization based on adversarial learning dynamically enhances the discriminative ability of the feature representation through the adversarial training mechanism of the generator and the discriminator, thereby enhancing the generalization performance and prediction accuracy of the model.

[0019] Input the high-quality synthetic features h′ generated by the generator and the real sample features h into the discriminator D for adversarial training. The discriminator learns the decision boundary for distinguishing real features from 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 discrimination result of the discriminator 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) represents the expectation of the real sample x, where x is sampled from the real data distribution p(x), and E z～q(z) represents the expectation of the 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 objectives:

[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 generalization ability of the model on unseen data; (5)Promote the collaborative optimization of the generator and the discriminator.

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

[0024] Another object of the present invention is to propose an anti-cancer drug target screening system based on a hypergraph-knowledge graph dual channel, including:

[0025] A knowledge graph feature extraction module, which is used to embed the gene-related biomedical knowledge graph by using a knowledge representation learning model, capture the semantic relationships between genes, and extract biologically meaningful gene semantic feature representations;

[0026] A hypergraph structure learning module, which is used to integrate known gene synthetic lethal data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethal relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on a hypergraph neural network, so as to obtain a comprehensive feature representation containing gene high-order interaction features and low-order structure information;

[0027] An adaptive multi-modal fusion module, which 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 adopt a dynamic weighting strategy based on the attention mechanism. Through unbiased linear transformation and the SoftMax function, the weights of each modal information are learned, and each modal information is fused to obtain a joint embedding;

[0028] A generative adversarial network training module, including a generator and a discriminator. The generator adopts a feed-forward neural network structure, including multiple linear layers, batch normalization layers, and activation functions, and combines random noise with the same dimension as the input features to generate high-quality negative samples similar to the real sample distribution. The discriminator is used to perform adversarial training on real samples and generated samples to optimize the data set;

[0029] A synthetic lethal relationship prediction module, which is used to calculate the synthetic lethal probability score between any gene pair by using the joint embedding output by the adaptive multi-modal fusion module after the discriminator is trained, so as to perform probability prediction on potential synthetic lethal partner relationships.

[0030] In this application, the feature extraction module first normalizes the extracted multi-dimensional features, then dynamically weights the importance of different features based on the attention mechanism, and finally outputs the predicted probability of the synthetic lethal relationship of gene pairs through a multi-layer perceptron. Under the theoretical framework of gene synthetic lethality, an innovative solution is provided for the screening of anti-cancer drug targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0032] Figure 1 is a flowchart of the method for screening anti-cancer drug targets based on the hypergraph-knowledge graph dual-channel according to the embodiment of the present invention;

[0033] Figure 2 is an example graph of the hypergraph constructed from the gene synthetic lethality data according to the embodiment of the present invention;

[0034] Figure 3 is a structure diagram of the generator according to the embodiment of the present invention;

[0035] Figure 4 is a structure diagram of the discriminator according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] In order to make the technical means, objectives and effects of the present invention easy to understand, the embodiments of the present invention will be described in detail below with reference to specific drawings.

[0037] It should be noted that all terms indicating directions and positions in the present invention, such as: "upper", "lower", "left", "right", "front", "rear", "vertical", "horizontal", "inner", "outer", "top", "bottom", "lateral", "longitudinal", "center", etc., are only used to explain the relative positional relationship and connection situation between components in a certain specific state (as shown in the drawings), and are only for the convenience of describing the present invention, rather than requiring the present invention to be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation to the present invention. In addition, the descriptions involving "first", "second", etc. in the present invention are only for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features.

[0038] In the description of the present invention, unless otherwise clearly defined and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0039] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0040] As Figure 1 shown, this application discloses an anti-cancer drug target screening method based on a hypergraph-knowledge graph dual channel, including the following steps:

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

[0042] Through the knowledge representation learning model, capture the semantic relationships between genes to obtain a gene semantic feature representation with biological significance; at the same time, integrate the known gene synthetic lethality data to construct a gene hypergraph network G=(V, E), where V represents the gene node set, and E represents the hyperedge set, and each hyperedge is composed of a group of gene nodes with synthetic lethality relationships, model the high-order interaction relationships between genes, input the hypergraph data into the feature extraction module based on the hypergraph neural network, and aggregate the neighbor node information through a multi-layer message passing mechanism to obtain a comprehensive feature representation including gene high-order interaction features and low-order structure information;

[0043] S2: Adaptive multimodal fusion

[0044] Input the gene semantic feature representation and the 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 weights of each modal information through the learned parameters;

[0045] S3: Generative adversarial network training

[0046] Construct a generator that generates high-quality negative samples similar to the distribution of real samples by introducing random noise vectors. At the same time, construct a discriminator to distinguish real samples from samples synthesized by the generator. Through the adversarial learning mechanism of the generator and the discriminator, expand the training dataset and learn the data distribution characteristics of real samples.

[0047] S4: Synthetic lethality relationship prediction

[0048] After the discriminator training is completed, use the adaptive fusion mechanism to integrate the structural feature representation and semantic feature representation of genes, and calculate the synthetic lethality probability score between any gene pair.

[0049] The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel disclosed in this application comprehensively captures the complex relationships between genes by combining the knowledge graph and hypergraph neural network. First, use the knowledge representation learning model to perform embedding operations on the biomedical knowledge graph related to genes, and extract gene semantic features with biological significance. These features can reflect the association information of genes in multiple aspects such as biological processes, diseases, and signal pathways. At the same time, combine the known synthetic lethality data to construct a gene hypergraph network, where each hyperedge in the network consists of a group of gene nodes with synthetic lethality relationships to model the high-order interaction relationships between genes. Then, input the hypergraph data into the feature extraction module based on the hypergraph neural network, and aggregate the information of neighboring nodes through a multi-layer message passing mechanism to obtain a comprehensive feature representation that reflects both local structure and high-order interaction features. Subsequently, send the semantic feature representation extracted from the knowledge graph and the comprehensive feature representation generated by the hypergraph neural network into the adaptive multi-modal fusion module. This module adopts a dynamic weighting strategy based on the attention mechanism to learn the importance of different modal information and achieve effective integration of each modal feature. Then, use the generative adversarial network to construct a generator and a discriminator. By introducing a random noise vector z ∼ N(0, I), generate high-quality negative samples similar to the distribution of real samples, enhance the generation diversity, and use the discriminator to perform adversarial learning on real samples and generated samples to optimize the training dataset, improve the discrimination difficulty of negative samples, avoid the noise that may be introduced by using unvalidated gene pairs as negative samples, and generate discriminative feature representations similar to the distribution of positive samples. Finally, after the discriminator training is completed, use the adaptively fused feature representation to calculate the synthetic lethality probability score between any gene pair, so as to realize the prediction of potential synthetic lethality relationships and provide accurate data support for anti-cancer drug target screening.

[0050] The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel disclosed in this application realizes the efficient fusion of gene semantic information and structural information by fully integrating two data sources, namely the knowledge graph and the hypergraph neural network, effectively breaking through the limitations of traditional simple linear splicing in the process of multi-modal data fusion. It uses an adaptive multi-modal fusion module to dynamically allocate the weights of each modal information, thereby ensuring that key information is not diluted and improving the accuracy of feature expression. On this basis, a high-quality negative sample is constructed through a generative adversarial network, effectively reducing the noise problem introduced by the traditional negative sample direct construction method, enhancing the training stability and generalization ability of the model, not only improving the accuracy and discrimination ability of synthetic lethal relationship prediction, but also shortening the cycle of anti-cancer drug target screening and reducing the screening cost, with high practicality and promotion value.

[0051] As a preferred example of this application, in step S1, the TransE knowledge representation learning model is used to perform an embedding representation on the gene-related biomedical knowledge graph SynLethKG by optimizing the energy function:

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

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

[0054] Where h represents the head entity, r represents the relationship, 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 regards the relationship as a translation vector from the head entity to the tail entity. Given a triple (head entity h, relationship r, tail entity t), the goal of TransE is to make h+r≈t, and the learning process is completed by minimizing the distance between h+r and t. Finally, a deep semantic feature vector that can accurately reflect the gene semantic information is obtained. The gene deep semantic feature vector is used as the input of the subsequent hypergraph structure learning and adaptive multi-modal fusion module, so that the hypergraph network and knowledge graph data composed of known gene synthetic lethal data can be effectively fused at a higher level, and provide reliable basic data support for the training of the generative adversarial network.

[0056] This solution makes full use of the advantages of the TransE model, transforms the complex semantic relationships in the gene-related biomedical knowledge graph SynLethKG into a vector form that can be processed by a computer, and the extracted gene deep semantic feature vectors can more accurately capture the semantic associations between genes. Compared with traditional methods, it can more comprehensively reflect the functions and action mechanisms of genes in the biomedical environment. By continuously iteratively optimizing, the head entity vector plus the relationship 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, this model extracts deep semantic feature vectors that can accurately reflect gene semantic information.

[0057] As a preferred example of this application, in step S1, the comprehensive feature representation of gene high-order interaction features and low-order structural information is:

[0058]

[0059] where β represents the attention weight, H ij represents the hypergraph incidence matrix, and h i (0) represents the node feature matrix.

[0060] In the example of this application, gene synthetic lethality data from multiple data sources is integrated, and a synthetic lethality hypergraph is constructed according to the synthetic lethality data. Figure 2 For the constructed hypergraph example, it is then 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 incidence matrix of the hypergraph, W represents the weight matrix of the hyperedges, and D v represents the degree matrix of the nodes.

[0063] Hyperedges can connect multiple nodes. By aggregating the local information of the nodes into the hyperedges through the hyperedges, and then passing the global information back to the nodes through the hyperedges, the relationships between multiple nodes are directly modeled to extract high-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 the low-order structural information of the hypergraph, the attention mechanism is applied to aggregate the features of neighbor nodes, that is:

[0065]

[0066] where N i represents the set of neighbor nodes of node i, w1 is a learnable parameter matrix, ReLU() represents the activation function, and h i (0) represents the aggregated node feature, and a ij represents the attention weight between node i and node j.

[0067] It can be calculated by 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 relationship between node features; A represents the hypergraph adjacency matrix; softmax() represents the activation function.

[0070] To learn the high-order structure information of the hypergraph, it is necessary to consider the importance of different hyperedges to node features; by combining the feature representation matrices of nodes, a comprehensive representation containing both high-order and low-order structure information of the hypergraph can be obtained, that is:

[0071]

[0072] where ε represents the set of hyperedges containing node j, is the attention score of the hyperedge and can be calculated by the following formula:

[0073]

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

[0075] In the process of using the hypergraph neural network to extract structural information in this application, an attention mechanism based on a single-layer feedforward neural network is further applied to dynamically allocate weights to node neighbor features. Among them, the aggregation from multiple nodes within a hyperedge to the hyperedge is realized, and the global information is passed back from the hyperedge to each node. The relative importance of different hyperedges in node feature learning is modeled by combining the hypergraph adjacency matrix, incidence matrix, weight matrix, and node degree matrix. Finally, a comprehensive representation containing high-order and low-order structural information is formed, and the prediction of synthetic lethal gene pairs is carried out based on this comprehensive representation. Based on the improved scheme of the above hypergraph construction and attention aggregation mechanism, the multi-node interaction information contained in the gene synthetic lethal data can be fully mined, avoiding information loss or aggregation bias that may be caused by relying only on simple adjacency methods. By introducing the attention mechanism of a single-layer feedforward neural network in the hypergraph neural network to calculate the relative importance between nodes and hyperedges, the gain effect brought by high-order associations can be captured while retaining local structural features, which can better reflect the potential interactions between genes and significantly enhance the attention to key hyperedges in complex biological networks. Therefore, when the model processes the synthetic lethal prediction task, it can take into account both the overall impact brought by the combined inactivation of multiple genes and the differential information within each hyperedge, improving the judgment accuracy of the potential synthetic lethal relationship 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 composed of linear layers with both input and output feature dimensions of 32 to obtain the semantic information of genes. At the same time, the gene synthetic lethal relationship data is input into the hypergraph neural network, and the gene synthetic lethal relationship data comes from SynLethDB-v2.0.

[0077] This scheme uses the TransE model to process the knowledge graph and extracts the semantic features contained therein. These features contain rich information about entities and relationships in the knowledge graph. Then, these semantic features are input into a semantic mapper composed of linear layers, and the linear layers perform linear transformations on the features, mapping the high-dimensional semantic features to a specific dimensional space, preferably a 32-dimensional specific space, so as to obtain more representative gene semantic information. At the same time, the gene synthetic lethal relationship data obtained from SynLethDB-v2.0 is fed into the hypergraph neural network, which uses a multi-layer message passing mechanism to model and aggregate the complex high-order interaction relationships between genes, realizing the comprehensive extraction of gene local structure and global interaction features. By integrating gene semantic information and hypergraph structure information, it provides a more comprehensive and accurate basis for subsequent drug target screening.

[0078] As a preferred example of this application, in step S2, a trainable parameter matrix W is used to evaluate the importance of the semantic feature vector and the structural feature vector, and the features are weighted and fused based on the attention scores to calculate the modality attention score a m :

[0079]

[0080] where a m is the weight of the modality information, is the dot product operation, and W m , W n are the learnable parameters corresponding to modalities m and n respectively, and M = {structure, semantics}, representing the structural information and semantic information.

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

[0082]

[0083] h merge represents the node embedding after modality fusion, which is used to accurately predict the potential synthetic lethal partner gene of a given target gene g p , providing key decision-making support for anti-cancer drug target screening.

[0084] Specifically, in the design process of the adaptive multi-modal fusion module, the importance of features is evaluated through the trainable parameter matrix W, and the features are weighted and fused based on the attention scores. In this way, the model can adaptively adjust the contribution degrees of different modal features according to the requirements of downstream tasks. When calculating the modality attention score, corresponding formulas are used for the semantic feature vector and the structural feature vector to determine the weights of different modal features in the fusion process, making the fused feature representation more robust and discriminative, thereby improving the prediction performance. By adopting the above scheme, it is possible to dynamically integrate the multi-source heterogeneous data of genes, effectively improve the robustness of feature representation and prediction performance, dynamically learn the weight distribution and weighted fuse features through the attention mechanism, make the model more flexible to adapt to the requirements of downstream tasks, and further enhance the prediction accuracy of synthetic lethal gene pairs.

[0085] As a preferred example of the present application, in step S2, the gene structure information h_structure and semantic information h_semantics are input into the adaptive multi-modal 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, and then the jointly adapted joint embedding h_joint of different modal information is obtained for subsequent discriminator prediction of the probability of synthetic lethality between a pair of genes and model inference. In the example of the present application, when performing adaptive multi-modal fusion in step S2, the input data containing gene structure information and semantic information is fed into the adaptive multi-modal 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 the key information features in both types of information and reflect the importance of each modality through the obtained intermediate output results. 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 weighted fusion on the original structure information and semantic information according to the obtained weights, thereby obtaining a unified joint embedding h_joint, which contains both the physical structure features and semantic features of the gene, comprehensively reflecting the multi-dimensional information of the gene under different data sources, and providing accurate and representative feature inputs for subsequent discriminator prediction of the probability of synthetic lethality relationship between gene pairs.

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

[0087] As a preferred example of the present application, in step S3, the gene structure information and semantic information extracted by the model are respectively input into the 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 the present application, the generator adopts a feed-forward neural network structure composed of four linear layers, with a batch normalization layer and a GELU activation function set between each layer, and the last layer uses a Sigmoid activation function to limit the output to the interval [0, 1], thereby generating high-quality negative samples that are similar to the real sample distribution and have diversity. Specifically, Figure 3It is the structural diagram of the generator. The generator consists of a sequential model, including four fully connected layers (Linear), three batch normalization layers (BatchNorm1d), GELU activation function, and Sigmoid activation function. The structure, semantic information of genes, 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, so as to enhance the diversity of the generated samples. Through such a generator structure, high-quality negative samples, that is, "false features" of genes, can be generated to expand the training data set and avoid data singularity. Then, these generated negative samples are input into the trained adaptive modality fusion module for subsequent adversarial training of the model. During the adversarial training process, the model continuously learns and optimizes, thereby improving the model 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 of 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 real features from generated features through a multi-layer neural network, and its objective function is expressed as:

[0089]

[0090] Among them, G represents the generator, D represents the discriminator, D(x) represents the discrimination result of the discriminator 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) represents the expectation of the real sample x, where x is sampled from the real data distribution p(x), and E z～q(z) represents the expectation of the random noise z, where z is sampled from the noise distribution q(z).

[0091] The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel described in this application adopts a feature discrimination optimization method based on adversarial learning. The generator generates high-quality synthetic features, and these synthetic features and real sample features are input into the discriminator for adversarial training. Through this adversarial training mechanism, the generator and discriminator are continuously optimized, enabling the model to more accurately predict synthetic lethal gene pairs. By adopting the above method, the discrimination ability of the model for features can be enhanced, enabling the model to better capture data features, improving the accuracy of prediction, and maintaining high generalization performance when facing new data, providing more reliable results for the prediction of synthetic lethal gene pairs.

[0092] As a preferred example of this application, in step S4, for the feature discrimination optimization based on adversarial learning, high-quality synthetic features are generated by the generator and jointly input into the discriminator for adversarial training. The discriminator learns the decision boundary for distinguishing real features from generated features through a multi-layer neural network, and inputs the features h p , h q of gene p and gene q into the discriminator for splicing. The spliced vector is processed through 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 discrimination 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 p , h q of gene p and gene q are input into the discriminator for splicing and similarity scoring. The similarity score is given by the discriminator and represents the probability that a pair of genes are synthetic lethal partners.

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

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

[0096] The features of gene p and gene q are linearly concatenated to form a joint vector, which contains mixed information from real data and forged data. The concatenated vector is then input into the discriminator, which performs a full connection operation on the concatenated vector and cooperates with the ReLU activation function to achieve dimensionality reduction and nonlinear transformation of the feature dimension. The feature dimension is changed from twice the feature dimension of the real sample to the same as the feature dimension of the real sample. Then, a fully connected layer is used to map the feature dimension to 1 dimension, and the output is mapped to the [0, 1] probability interval through the Sigmoid activation function. The similarity score of gene p and gene q is obtained, that is, the probability that genes p and q are synthetic lethal partners. This value represents the confidence that the gene pair (p, q) constitutes a synthetic lethal partner relationship, which is used for subsequent anticancer drug target screening and can be expressed as:

[0097]

[0098] Where φ represents the ReLU activation function, Represents a concatenation operation.

[0099] Based on this adversarial learning framework, the model can use the discriminator to more accurately identify whether there is a synthetic lethal relationship between gene p and gene q, which not only significantly improves the prediction accuracy, but also is more robust in dealing with potential noise and data imbalance problems. The adversarial training process enables the discriminator to continuously compete with the high-quality negative samples generated by the generator, prompting the discriminator to learn to capture key feature differences and make rigorous judgments based on them. At the same time, the generator continuously corrects the way of generating synthetic features in the process of continuous trial and error, making it closer and closer to the real data, and gradually improving the model's internal understanding of various patterns and the characterization of complex interactive information, so as to accurately judge whether the gene is a synthetic lethal partner, improve the model's prediction accuracy and generalization performance, and provide a more reliable basis for the screening of anti-cancer drug targets.

[0100] Finally, by linearly concatenating the modal fusion features h_joint_x and h_joint_y of gene x and gene y and passing them through the discriminator, we can obtain the probability P that gene x and gene y are synthetic lethal partners.

[0101] In the examples of this application, the dataset used in the present invention 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 related databases, computation, prediction, and text mining. SynLethDB 2.0 covers humans and 4 model species, namely mice, Drosophila, worms, and yeast. The current version contains 50,868 synthetic lethal pairs. The current version of SynLethDB 2.0 also contains a knowledge graph, which contains 11 entities and 27 relationships.

[0102] The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel disclosed in this application adopts a hypergraph-knowledge graph dual-channel structure, an adaptive multi-modal fusion mechanism, and a feature discrimination optimization method based on adversarial learning. First, the TransE model is used to convert the complex semantic relationships in the gene-related biomedical knowledge graph into low-dimensional deep feature vectors, and after mapping through a linear layer, refined gene semantic information is extracted. At the same time, gene synthetic lethal data from SynLethDB-v2.0 is integrated to construct a hypergraph network. In this network, the hypergraph neural network dynamically aggregates the high-order interactions and local structural information between genes through a multi-layer message passing mechanism and an attention mechanism. Then, an unbiased linear layer combined with the SoftMax function is used to adaptively learn the importance of the structural information and semantic information and generate a unified joint embedding. Furthermore, the generator injects random noise with the same dimension as the original feature into the input real sample features to generate high-quality negative samples close to the real distribution, and inputs them together with the real samples into the discriminator. The discriminator uses linear splicing, fully connected layers, ReLU, and Sigmoid activation functions for non-linear feature dimensionality reduction and mapping to construct a decision boundary that can accurately distinguish positive and negative samples. Thus, the performance of the generator and discriminator is continuously optimized during the adversarial training process, enabling the model to more accurately capture the potential synthetic lethal relationships between genes. This method not only fully integrates multi-source heterogeneous data, overcomes the problems of information dilution and noise introduction easily caused by fixed-weight splicing in traditional data fusion, but also uses the adversarial training mechanism to improve the robustness and generalization ability of the model in negative sample construction, thereby significantly improving the prediction accuracy, stability, and execution efficiency of anti-cancer drug target screening, providing a computational tool for anti-cancer drug research and development.

[0103] This application also discloses an anti-cancer drug target screening system based on the hypergraph-knowledge graph dual-channel, including:

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

[0105] A hypergraph structure learning module for integrating known synthetic lethal gene data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethal relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on a hypergraph neural network to obtain a comprehensive feature representation containing high-order interaction features and low-order structural information of genes;

[0106] An adaptive multi-modal fusion module for inputting 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 adopting a dynamic weighting strategy based on an attention mechanism to learn the weights of each modal information through unbiased linear transformation and the SoftMax function, and fusing each modal information to obtain a joint embedding;

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

[0108] A synthetic lethal relationship prediction module for calculating the synthetic lethal probability score between any gene pair using the joint embedding output by the adaptive multi-modal fusion module after the discriminator training is completed, so as to perform probability prediction on potential synthetic lethal partner relationships.

[0109] The present invention screens anti-cancer drug targets based on the synthetic lethal mechanism and realizes drug target prediction by integrating multi-modal synthetic lethal data. The present invention includes a feature extraction module (knowledge graph feature extraction module, hypergraph structure learning module), an adaptive modal fusion module, a generative adversarial network module and a synthetic lethal relationship prediction module. First, gene structure features containing high-order and low-order relationships are extracted based on a hypergraph neural network, and at the same time, gene semantic features are extracted based on a biological knowledge graph. Then, the adaptive modal fusion module dynamically adjusts the weights of each modal information according to the learned parameters to realize the reasonable utilization of each modal information. Next, the generative adversarial network module realizes collaborative optimization through the adversarial learning mechanism of the generator and the discriminator; the generator learns the data distribution characteristics of real samples by introducing a random noise vector and generates high-quality negative samples with a distribution similar to that of real samples; at the same time, the discriminator distinguishes real samples from the samples synthesized by the generator to improve the accuracy of sample recognition. During the training process, the generator and the discriminator are continuously optimized through adversarial training, and finally reach a dynamic balance to complete the iterative update of the model parameters. Finally, after the discriminator training is completed, the synthetic lethal relationship prediction module integrates the structural feature representation and semantic feature representation of genes through an adaptive fusion mechanism and calculates the synthetic lethal probability score between any gene pair.

[0110] Specifically, for the anti-cancer drug target screening system based on the hypergraph-knowledge graph dual-channel described in this application, first, the feature extraction module normalizes the extracted multi-dimensional features, then dynamically weights the importance of different features based on the attention mechanism, and finally outputs the prediction probability of the synthetic lethal relationship of gene pairs through a multi-layer perceptron. Under the theoretical framework of gene synthetic lethality, the present invention provides an innovative solution for anti-cancer drug target screening.

[0111] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to a computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the aforementioned storage medium includes: ROM, RAM, FLASH, floppy disk, magnetic disk, optical disk, mechanical hard disk, cloud, and other various media that can store program codes. The computer processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), that is, the steps of the anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel can be implemented through the above processors.

[0112] In summary, for the anti-cancer drug target screening method and system based on the hypergraph-knowledge graph dual-channel of the present invention, the hypergraph-knowledge graph dual-channel method is used to comprehensively capture the complex interactions between genes, and an adaptive multi-modal fusion module is used to effectively integrate multi-source information. At the same time, a generative adversarial network training module is used to construct high-quality negative samples to optimize the training data set, so as to accurately calculate the synthetic lethal probability between gene pairs with the support of the discriminator. It not only overcomes the problems of high cost, long cycle, and inconsistency in traditional high-throughput experimental methods, but also effectively solves the problems of insufficient feature expression and unstable training caused by fixed-weight splicing and noise data introduction in existing multi-modal data fusion methods for information fusion and negative sample construction. It greatly improves the efficiency and accuracy of anti-cancer drug target screening, and provides a scientific, accurate, and reliable computing tool for anti-cancer drug research and development.

[0113] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for screening anti-cancer drug targets based on a dual-channel of hypergraph and knowledge graph, characterized in that, It includes the following steps: S1: Through a knowledge representation learning model, capture the semantic relationships between genes to obtain a gene semantic feature representation with biological significance; meanwhile, integrate the known gene synthetic lethality data, construct a gene hypergraph network, input the hypergraph data into a feature extraction module based on a hypergraph neural network, and aggregate neighbor node information through a multi-layer message passing mechanism to obtain a comprehensive feature representation containing gene high-order interaction features and low-order structure information; S2: Input the gene semantic feature representation and the comprehensive feature representation obtained in step S1 into an adaptive multi-modal fusion module. The adaptive multi-modal fusion module adopts a dynamic weighting strategy based on an attention mechanism to dynamically adjust the weights of each modal information through learned parameters; S3: Construct a generator and a discriminator, introduce a random noise vector, generate high-quality negative samples similar to the distribution of real samples, and expand the training data set and learn the data distribution characteristics of real samples through the adversarial learning mechanism between the generator and the discriminator; S4: After the discriminator training is completed, use an adaptive fusion mechanism to integrate the structural feature representation and semantic feature representation of genes, and calculate the synthetic lethality probability score between any gene pair.

2. The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual channel according to claim 1, wherein In step S1, the TransE knowledge representation learning model is used to perform embedding representation on the gene-related biomedical knowledge graph SynLethKG. By optimizing the energy function: E(h,r,t)=||h+r-t|| (1) Map the entities and relationships in the knowledge graph to a low-dimensional vector space to extract the deep semantic feature vector of genes; where h represents the head entity, r represents the relationship, t represents the tail entity, and || || represents the Euclidean distance between vectors.

3. The anti-cancer drug target screening method based on a hypergraph-knowledge graph dual channel according to claim 2, characterized in that, In step S1, the comprehensive feature representation of gene high-order interaction features and low-order structure information is: Among them, β represents the attention weight, and H ij represents the hypergraph incidence matrix, and h i (0) represents the node feature matrix.

4. The anti-cancer drug target screening method based on a hypergraph-knowledge graph dual channel according to claim 3, wherein In step S1, input the knowledge graph into the TransE model to extract semantic features, and then input the semantic features into a semantic mapper composed of linear layers with both input and output feature dimensions of 32 to obtain the semantic information of genes. At the same time, input the gene synthetic lethality relationship data into the hypergraph neural network. The gene synthetic lethality relationship data comes from SynLethDB-v2.

0.

5. The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel according to any one of claims 1 to 4, characterized in that In step S2, the importance of the semantic feature vector and the structural feature vector is evaluated using the trainable parameter matrix W, and the features are weighted and fused based on the attention scores to calculate the modality attention score a m : where a m is the weight of the modal information, is the dot product operation, W m , W n are respectively the learnable parameters corresponding to the modalities m and n, M = {structure, semantics}, representing the structural information and semantic information.

6. The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual channel according to claim 5, characterized in that, In step S2, input the gene structure information h_structure and semantic information h_semantics into the adaptive multi-modal fusion module, learn the importance of different modal information through an unbiased linear layer Linear, and then obtain its weight through the SoftMax function, and then obtain the joint embedding h_joint that automatically adapts to different modal information for subsequent discriminator prediction of the probability of synthetic lethality between a pair of genes and model inference.

7. The anti-cancer drug target screening method based on a hypergraph-knowledge graph dual channel according to claim 1, characterized in that, In step S3, input the gene structure information and semantic information extracted by the model into a gene-based generator respectively, and at the same time inject random noise with the same dimension as the information into the generator. The generator learns the distribution characteristics of real samples through a multi-layer neural network.

8. The anti-cancer drug target screening method based on a hypergraph-knowledge graph dual channel according to claim 1, wherein, In step S4, through the adversarial training mechanism of the generator and the discriminator, the high-quality synthetic feature h' generated by the generator and the real sample feature h are jointly input into the discriminator for adversarial training. The discriminator learns the decision boundary for distinguishing real features from generated features through a multi-layer neural network, and its objective function is expressed as: Among them, G represents the generator, D represents the discriminator, D(x) represents the discrimination result of the discriminator 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) represents the expectation of the real sample x, where x is sampled from the real data distribution p(x), and E z～q(z) represents the expectation of the random noise z, where z is sampled from the noise distribution q(z).

9. The anti-cancer drug target screening method based on the hypergraph-knowledge graph dual-channel according to claim 8, wherein, In step S4, for the feature discrimination optimization based on adversarial learning, a high-quality synthetic feature is generated by a generator and jointly input into a discriminator with the real sample features for adversarial training. The discriminator learns the decision boundary for distinguishing real features from generated features through a multi-layer neural network, and the features h p , h q of gene p and gene q are input into the discriminator for splicing. The spliced vector is processed through 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 training is completed, by linearly concatenating the modality fusion features h_joint_x and h_joint_y of gene x and gene y and passing them through the discriminator, the probability P that gene x and gene y are synthetic lethal partners is obtained.

10. An anti-cancer drug target screening system based on a hypergraph-knowledge graph dual-channel, characterized in that, Including: A knowledge graph feature extraction module for embedding the gene-related biomedical knowledge graph using a knowledge representation learning model, capturing the semantic relationships between genes, and extracting gene semantic feature representations with biological significance; A hypergraph structure learning module for integrating known gene synthetic lethality data to construct a gene hypergraph network. The hypergraph network includes gene nodes and hyperedges composed of genes with synthetic lethal relationships. This module aggregates neighbor node information through a multi-layer message passing mechanism based on a hypergraph neural network to obtain a comprehensive feature representation containing gene high-order interaction features and low-order structure information; An adaptive multi-modal fusion module for inputting 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 adopting a dynamic weighting strategy based on an attention mechanism to learn the weights of each modality information through unbiased linear transformation and the SoftMax function, and fusing each modality information to obtain a joint embedding; A generative adversarial network training module, including a generator and a discriminator. The generator adopts a feed-forward neural network structure, including multiple linear layers, batch normalization layers, and activation functions, and combines random noise with the same dimension as the input features to generate high-quality negative samples similar to the real sample distribution. The discriminator is used to perform adversarial training on real samples and generated samples to optimize the data set; A synthetic lethality relationship prediction module for calculating the synthetic lethality probability score between any gene pair using the joint embedding output by the adaptive multi-modal fusion module after the discriminator training is completed, so as to perform probability prediction on potential synthetic lethal partner relationships.

Images