A drug synergistic effect prediction method and system based on multi-level information fusion

Abstract

This invention relates to the field of bioinformatics, specifically to a method and system for predicting drug synergistic effects based on multi-level information fusion. The method includes acquiring the SMILES structure of a drug, gene expression profiles of cancer cell lines, and drug combination response data; converting the SMILES structure into a molecular graph, standardizing the gene expression profile, and labeling samples; extracting system-level features using a graph converter layer and a hypergraph neural network through a system-level feature extraction module; extracting molecular-level features using a graph attention network and an external substructure perception layer through a molecular-level feature extraction module; and adaptively weighting and fusing the system-level and molecular-level features through a multi-level information fusion prediction module to output the drug synergistic effect prediction result. This invention overcomes the limitations of single-level representation in existing technologies by simultaneously capturing multi-level information at both the system and molecular levels, significantly improving the accuracy and generalization ability of drug synergistic effect prediction.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, specifically to a method and system for predicting drug synergistic effects based on multi-level information fusion. Background Technology

[0002] As biomedical data becomes increasingly complex, computational methods, due to their scalability and cost-effectiveness, have become a key tool for predicting drug synergies. Traditional machine learning models (such as random forests and support vector machines) have proven effective in reducing experimental burden and improving screening efficiency. Compared to these methods, deep learning models extract hierarchical features through multi-layer nonlinear transformations, making them particularly suitable for large-scale heterogeneous datasets. Among them, graph neural networks (GNNs) have shown unique potential in the field of drug synergy prediction by modeling complex relational structures. GNNs can effectively capture both the biological and topological interactions between drugs and cell lines, while simultaneously learning information-valuable molecular representations.

[0003] Currently, research on drug synergistic prediction based on Generative Neural Networks (GNNs) mainly falls into two categories: relation-based methods and structure-based methods. Relationship-based methods aim to model the complex interactions between drug combinations and cell lines by constructing heterogeneous networks or knowledge graphs that integrate biological and pharmacological relationships. For example, Liu et al. constructed a drug-drug-cell line hypergraph and used a hypergraph neural network for synergistic prediction; Zhang et al. constructed a knowledge graph containing drugs, targets, enzymes, and transporters, extracting entity semantic information through embedding learning. However, these methods often focus on high-level interactions between drugs and cell lines, potentially neglecting key molecular-level information. Furthermore, the dependence of knowledge graphs on predefined entities and relationships limits their generalization ability to novel drugs or cell lines.

[0004] Structure-based methods model drug molecules as graph structures and apply graph neural networks (GNNs) to extract structural features. For example, Chen et al. used graph transformers to simulate higher-order dependencies between drugs and cell lines; Wang et al. converted SMILES strings into drug molecule graphs and embedded drug features using graph attention networks and graph convolutional networks. However, these methods mainly focus on local molecular features and often ignore the complex multidirectional interactions between drug combinations and cell lines. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for predicting drug synergy based on multi-level information fusion. By synergistically integrating system-level interaction information and molecular-level substructure information of drugs, the accuracy and generalization ability of drug synergy combination prediction are significantly improved.

[0006] According to a first aspect of the embodiments of this disclosure, a method for predicting drug synergy based on multi-level information fusion is provided, comprising the following steps: Obtain the SMILES structural formulas of drugs, gene expression profiles of cancer cell lines, and drug combination response data; The SMILES structural formula was converted into a drug molecule map, the gene expression profile was standardized, and the samples were labeled according to the drug synergy score. An initial drug feature is encoded using a graph converter layer, and cell line features are encoded using a multilayer perceptron. A hypergraph containing drug nodes and cell line nodes is constructed. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; A graph attention network is used to perform multi-layer convolution on the drug molecule graph. The pre-activation outputs of each layer are extracted and aggregated by self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. The similarity features of the drug interstitial structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; For the system-level features With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

[0007] In one embodiment, the hypergraph containing drug nodes and cell line nodes is constructed as follows: in Includes drug nodes With cell line nodes hyper-edge set Contains triplets corresponding to synergistic drug-drug-cell line interactions, diagonal matrix Assign a weight to each superedge.

[0008] In one embodiment, system-level features are extracted. The method is as follows: in Let the node degree matrix be... Hypermarginality matrix For trainable parameters, For the first m-layer node feature matrix For the transpose of the hypergraph incidence matrix, For decoders.

[0009] In one embodiment, the graph attention network has nodes at layer l+1. i Embedded Implement it in the following way: in, This indicates concatenating the outputs of K attention heads, where σ represents the activation function. For trainable weight matrix, Let i represent the neighborhood of node i; Let j be the feature vector of node j in the l-th layer; This represents the attention coefficient of node j to node i in the k-th attention head: in, This represents the weight vector of the k-th attention head.

[0010] In one embodiment, the drug substructure features for: in, For the first l Layer nodes i The pre-activated output, SAGPooling is a self-attention graph pooling operation.

[0011] In one embodiment, the external substructure sensing layer processes the data as follows: in, This represents the concatenated output of K attention heads. , Represents trainable weights. , Represents the bias parameter. This indicates a summation operation over the second dimension.

[0012] In one embodiment, the fusion feature The method of obtaining it is: in, As a system-level feature, Characterized at the molecular level. , They are respectively , The projection, It is a multi-level interaction factor.

[0013] According to a second aspect of the present disclosure, a drug synergy prediction system based on multi-level information fusion is provided, comprising: The data acquisition module acquires the SMILES structural formulas of drugs, gene expression profiles of cancer cell lines, and drug combination response data. The data processing module converts the SMILES structural formula into a drug molecule map, standardizes the gene expression profile, and labels the samples according to the drug synergy score. The system-level feature extraction module uses a graph converter layer to encode initial drug features and a multilayer perceptron to encode cell line features, constructing a hypergraph containing drug nodes and cell line nodes. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; The molecular-level feature extraction module employs a graph attention network to perform multi-layer convolution on the drug molecule graph, extracts the pre-activation outputs of each layer, and aggregates them using self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. The similarity features of the drug interstitial structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; The multi-level information fusion prediction module analyzes the system-level features. With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned method for predicting drug synergy based on multi-level information fusion.

[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for predicting drug synergy based on multi-level information fusion.

[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention models high-order topological interactions between drug pairs and cell lines at the system level, explicitly characterizing the overall action pattern of drug combinations in different cellular environments. This module utilizes a hypergraph neural network to jointly model the drug-drug-cell line triple, enabling the model to fully understand the complex dependencies between drug combinations and the cellular environment, thereby enhancing the ability to recognize synergistic effects.

[0017] 2. This invention employs deep convolution on drug molecule graphs to capture local structural relationships at the atomic level and between chemical bonds. The multi-head mechanism of the graph attention network allows each attention head to focus on different chemical properties, such as bond type, functional groups, or spatial arrangement, thereby generating a comprehensive and detailed molecular characterization. Furthermore, the introduction of an external substructure sensing layer enables the model to capture the pharmacological similarities implied by shared substructures among drugs, overcoming the limitations of existing methods that only focus on the internal features of a single drug.

[0018] 3. This invention effectively integrates system-level and molecular-level features. Employing an attention-based adaptive fusion strategy, it dynamically adjusts weights based on the contribution of different features to the prediction results, enabling the model to accurately capture high-order interaction patterns of drug combinations. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0020] Figure 1 This is an architecture diagram of a drug synergy prediction method based on multi-level information fusion. Figure 2 Experimental strategy diagram for partitioning training and testing datasets; Figure 3 This is a performance comparison chart of drug synergistic prediction models under cold start conditions. Detailed Implementation

[0021] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0025] Example 1: like Figure 1 As shown, this embodiment provides a method for predicting drug synergy based on multi-level information fusion, including the following steps: S1. Obtain the SMILES structural formula of the drug, the gene expression profile of the cancer cell line, and the drug combination response data; This embodiment uses Merck and DrugComb as benchmark evaluation datasets. The SMILES structure data for the drugs were obtained from the DrugBank database, and the gene expression profile data of the cancer cell lines were obtained from the CCLE database. S2. Convert the SMILES structural formula into a drug molecule map, standardize the gene expression profile, and label the samples according to the drug synergy score; Specifically, gene expression values ​​were extracted from the whole-genome read count matrix and standardized using the Transcripts Per Million (TPM) method to ensure comparability across samples. For molecular representation, the RDKit tool was used to convert the SMILES string into a molecular graph, where atoms and bonds correspond to nodes and edges, respectively. Each atom is encoded using a binary feature vector containing five attributes: atom symbol, number of neighboring atoms, number of bonded hydrogen atoms, implicit valence, and atomicity.

[0026] Following the method of Wang et al., samples were partitioned based on the Loewe additive score, and a strict threshold was applied to minimize noise, thus constructing a benchmark dataset. Drug combinations with scores higher than 10 were marked as positive samples, and those lower than 0 were marked as negative samples, thereby obtaining a relatively balanced dataset.

[0027] To address the dimensionality imbalance between drug characteristics and cell line characteristics, this embodiment selects key genes identified by the Comprehensive Network Cell Feature Library (LINCS L1000) project. This project defines approximately 1000 landmark genes, which can explain about 80% of the variability in connectivity map data. By performing intersection analysis of these landmark genes with the CCLE lineage feature set, feature screening of gene expression profiles is completed. Further integration of CCLE and GENCODE annotation data removes redundant entries and non-coding RNA transcripts.

[0028] To ensure data quality, the following refinements were performed on the two benchmark datasets: (1) entries unrelated to drug combinations were excluded; (2) samples lacking drug or cell line characteristics were excluded; and (3) data affected by noise or classified as additive effects according to the Loewe criteria were excluded. In addition, to address the class imbalance problem in the DrugComb dataset, stratified sampling was performed on negative class samples.

[0029] S3. A graph converter layer is used to encode the initial drug features, and a multilayer perceptron is used to encode the cell line features, constructing a hypergraph containing drug nodes and cell line nodes. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; Specifically, the SRAM module is used to capture system-level features of drug pairs and introduces a similarity-based auxiliary task to enhance the generalization ability of drug-cell line interaction characterization. Considering the excellent performance of the Transformer model in tasks such as molecular graph processing, this embodiment uses a Graph Transformer (GT) layer for drug feature extraction. The GT layer uses a self-attention mechanism to dynamically adjust the feature weights of neighboring nodes, which not only flexibly models the molecular graph structure but also simultaneously captures local substructure information and global context information.

[0030] Initial drug characteristics Obtained through GT processing Cell line characteristics are generated via MLP encoding. Then a hypergraph is constructed. ,in Includes drug nodes With cell line nodes The node feature matrix is ​​denoted as Hyper-edgeset It must contain triples corresponding to synergistic drug-drug-cell line interactions; otherwise, no edges are formed. Diagonal matrix. Assign a weight to each superedge.

[0031] To capture latent node representations, a hypergraph neural network framework is employed. This framework utilizes hypergraphs to model higher-order relationships and achieves representation learning through hyperedge convolution. As defined in the following equation, This process integrates the node degree matrix. Hypermarginality Matrix and trainable parameters The decoder is implemented using a fully connected layer: S4. A graph attention network is used to perform multi-layer convolution on the drug molecule graph. The pre-activation outputs of each layer are extracted and aggregated using self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. The similarity features of the drug interstitial structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; Substructure-level analysis has become a crucial step in predicting drug synergistic effects, with numerous studies demonstrating a close correlation between specific molecular substructures and drug synergistic effects. This embodiment treats substructures as discrete entities, capturing fine-grained molecular-level structural information by embedding and encoding key drug sub-components. Given that structurally similar substructures often have similar therapeutic effects, the MSLM module employs a graph neural network to identify key substructures in synergistic drugs. Furthermore, this embodiment introduces an External Substructure Perception Layer (ESPL) to capture substructure similarities between drugs, thus overcoming the shortcomings of existing methods that primarily focus on internal drug substructure features while neglecting inter-drug substructure relationships.

[0032] In describing drug characteristics After being embedded in a graph structure, the model can efficiently extract key substructures. application The GAT (Generative Attention Array) layer aggregates atomic-level information through attention weights to form unique substructures at each layer. GAT adaptively adjusts the influence of neighboring nodes, prioritizing more relevant interactions during feature learning. To enhance representation capabilities, a multi-head attention mechanism is employed, allowing each head to focus on different chemical properties (such as bond type, functional groups, or spatial arrangement). By integrating these diverse perspectives, a comprehensive node representation is generated. Node i is embedded in layer (l+1). As shown below: in This indicates that the outputs of K attention heads are concatenated, and σ represents the activation function; ELU is used as the activation function because it helps prevent the "dead neuron" problem of ReLU and accelerates network convergence.

[0033] Attention coefficient This represents the importance of node j to node i, and this coefficient appears in the k-th attention head. Where W represents the trainable weight matrix. This indicates feature concatenation, and T indicates transpose. This represents the neighborhood of node i. In this embodiment, pre-activation outputs are extracted from each layer l=1,2,3,...,N. (recorded as) SAGPooling further optimizes this representation using a self-attention mechanism: it evaluates node importance based on both feature content and graph topology, retaining the most informative nodes to construct a global representation.

[0034] Shared substructures often imply similar pharmacological properties. Therefore, this embodiment introduces an External Substructure Perception Layer (ESPL) module to implicitly capture substructure similarities between drugs. Simultaneously, to extract chemically meaningful functional group information from different feature subspaces, this embodiment employs a multi-head attention mechanism. The aforementioned deep feature extraction process is as follows: Where K represents the number of attention heads. The Original Substructure Extraction (OSE) module captures the correlations and similarities between substructures in the original feature space; while the High-dimensional Substructure Extraction (HSE) module performs a deeper transformation on the normalized features to identify more abstract intermolecular similarities. Both OSE and HSE operate independently of the molecular graph structure. As dedicated substructure similarity extraction components, their specific operations are as follows: in This represents the concatenated output of K attention heads, where W and b represent the trainable weights and bias parameters, respectively. This indicates a summation operation over the second dimension.

[0035] A comprehensive interaction model of drug pairs is performed based on the external substructure features and their relationship with cell lines. The drug pair interaction is represented as follows: ,in Calculate interaction ratings, Identify key substructures. Parameters , a and b are both learnable parameters. Drug-cell line interaction is represented as DC= ,in This represents the normalized cell line characteristics converted by a multilayer sensor, while This represents the interaction score between the drug and cell line derived through the attention mechanism. By applying attention weights to the basic interaction features, the final characterization can be obtained, effectively capturing the strength of the association between the drug and cell line. Furthermore, the drug characterization output by the ESPL module is concatenated with the cell line features to form the local information required for cross-level information fusion, denoted as... .

[0036] S5. Regarding the system-level features With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

[0037] The Multi-level Information Fusion Module (MIFM) integrates system-level features output from the System-level Feature Extraction Module (SRAM) and molecular-level features output from the Molecular-level Feature Extraction Module (MSLM) through an adaptive weighting mechanism. This module first uses an attention mechanism to identify key elements in the features at each level, then adjusts the original features through interaction factors, and finally outputs a weighted fused comprehensive representation.

[0038] in and They come from SRAM and MSLM modules respectively, and and yes and The projection of I represents the multi-level interaction factor. This process achieves a comprehensive fusion of system-level and molecular-level features, significantly improving prediction accuracy.

[0039] This invention proposes a multi-level information capture model, MMSSyn, to extract and integrate heterogeneous drug and cell line information at the molecular and systemic levels. Through an adaptive fusion mechanism across representational features, this model significantly improves the prediction accuracy of drug synergistic effects. In warm-start scenarios on the Merck and DrugComb benchmark datasets, MMSSyn achieves state-of-the-art performance. Furthermore, in cold-start scenarios with tissue absence, drug absence, and independent dataset evaluation, the model maintains robust generalization ability, as shown in Table 1.

[0040] Table 1 Performance comparison of MMSSyn and its variants To comprehensively evaluate the performance of MMSSyn, this embodiment conducts tests under both warm-start and cold-start scenarios. In the warm-start scenario, five-fold cross-validation is performed on the complete dataset, with each fold of data serving as the test set sequentially, and the remaining four folds used for training. The mean and standard deviation of the performance metrics for each fold are reported. The cold-start scenario simulates real-world applications and evaluates the model's generalization ability to unseen data, specifically employing three strategies: drug removal, tissue removal, and independent dataset evaluation. Figure 2The details of the drug exclusion method and tissue exclusion method are shown. Each data point represents a drug-cell line triplet, and the green area highlights the test set partitioning under different partitioning schemes. Figure 2 As shown, the drug exclusion method removes all samples involving a specific drug (such as Drug3) from the training set and assigns them all to the test set to evaluate the model's ability to predict the synergistic effects of new drugs. Figure 2 The tissue exclusion rule shown assigns all samples associated with a specific tissue type (e.g., tissue 1) to the test set, with the remaining samples used for training. In independent validation, the model was trained on the Merck dataset and tested on an external AstraZeneca dataset containing 668 drug-cell line pairs consisting of 57 drugs and 24 cancer cell lines.

[0041] Table 2 Performance comparison of MMSSYN and baseline methods on the MERCK dataset (mean ± standard deviation) Table 3 Performance comparison of MMSSYN and baseline methods on the DrugComb dataset (mean ± standard deviation) To evaluate model performance, this embodiment compares MMSSyn with existing state-of-the-art deep learning models on the Merck and DrugComb datasets. Furthermore, to enhance data diversity and verify the general assumption that "drug administration order does not affect synergistic effects," symmetrical combinations of each drug pair were specifically included in the dataset.

[0042] Tables 2 and 3 present the hot-start performance comparison results of MMSSyn on the Merck and DrugComb datasets, respectively. It can be seen that MMSSyn consistently outperforms other comparative models across all evaluation metrics, achieving an AUC of 97.9% and an AUPR of 97.8% on the Merck dataset, fully demonstrating its powerful classification ability for drug synergies. Compared to traditional machine learning models such as Support Vector Machines and Random Forests, MMSSyn's performance improvement is more significant, highlighting the advantages of deep learning in drug synergy prediction. Although the performance gap with deep learning models such as DFFNDDS and CCSynergy has narrowed, MMSSyn still maintains its lead, further validating the superior ability of graph neural networks to capture intermolecular structural relationships. Furthermore, MMSSyn surpasses graph-based models such as SDDSynergy and HypergraphSynergy, demonstrating the importance of multi-level feature representation and integration in simulating complex drug-cell line interactions.

[0043] Figure 3The superior performance of MMSSyn in cold-start experiments is demonstrated. It can be seen that MMSSyn outperforms the benchmark model in various test scenarios, exhibiting strong adaptability and generalization ability. Its robust performance in complex scenarios such as distribution shifts and novel drug combinations indicates that the model has overcome the limitations of conventional training conditions. Especially in the independent validation task, the model maintains high prediction accuracy even when it has not encountered any test drug combinations during the training phase. These results fully demonstrate that the model can not only effectively generalize from existing patterns but also transfer knowledge to unfamiliar scenarios.

[0044] Example 2: This embodiment provides a drug synergy prediction system based on multi-level information fusion, including: The data acquisition module acquires the SMILES structural formulas of drugs, gene expression profiles of cancer cell lines, and drug combination response data. The data processing module converts the SMILES structural formula into a drug molecule map, standardizes the gene expression profile, and labels the samples according to the drug synergy score. The system-level feature extraction module uses a graph converter layer to encode initial drug features and a multilayer perceptron to encode cell line features, constructing a hypergraph containing drug nodes and cell line nodes. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; The molecular-level feature extraction module employs a graph attention network to perform multi-layer convolution on the drug molecule graph, extracts the pre-activation outputs of each layer, and aggregates them using self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. The similarity features of the drug interstitial structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; The multi-level information fusion prediction module analyzes the system-level features. With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

[0045] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0046] Example 3: An electronic device is provided for running the aforementioned "a method for predicting drug synergistic effects based on multi-level information fusion". The electronic device includes: a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain the SMILES structural formula of the drug, the gene expression profile of the cancer cell line, and the drug combination response data; S2. Convert the SMILES structural formula into a drug molecule map, standardize the gene expression profile, and label the samples according to the drug synergy score; S3. A graph converter layer is used to encode the initial drug features, and a multilayer perceptron is used to encode the cell line features, constructing a hypergraph containing drug nodes and cell line nodes. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; S4. A graph attention network is used to perform multi-layer convolution on the drug molecule graph. The pre-activation outputs of each layer are extracted and aggregated using self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. The similarity features of the drug interstitial structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; S5. Regarding the system-level features With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

[0047] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0048] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0049] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0050] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0051] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for predicting drug synergistic effects based on multi-level information fusion, characterized in that, Includes the following steps: Obtain the SMILES structural formulas of drugs, gene expression profiles of cancer cell lines, and drug combination response data; The SMILES structural formula was converted into a drug molecule map, the gene expression profile was standardized, and the samples were labeled according to the drug synergy score. An initial drug feature is encoded using a graph converter layer, and cell line features are encoded using a multilayer perceptron. A hypergraph containing drug nodes and cell line nodes is constructed. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; A graph attention network is used to perform multi-layer convolution on the drug molecule graph. The pre-activation outputs of each layer are extracted and aggregated by self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. ; The similarity features of the drug mesenchymal structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; For the system-level features With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

2. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, The hypergraph containing drug nodes and cell line nodes is constructed as follows: in Includes drug nodes With cell line nodes hyper-edge set Contains triplets corresponding to synergistic drug-drug-cell line interactions, diagonal matrix Assign a weight to each superedge.

3. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, Extracting system-level features The method is as follows: in Let the node degree matrix be... Hypermarginality matrix For trainable parameters, For the first m-layer node feature matrix This is the transpose of the hypergraph incidence matrix. For decoders.

4. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, The nodes of the (l+1)th layer of the graph attention network i Embedded Implement it in the following way: in, This indicates concatenating the outputs of K attention heads, where σ represents the activation function. For trainable weight matrix, Let i represent the neighborhood of node i; Let j be the feature vector of node j in the l-th layer; This represents the attention coefficient of node j to node i in the k-th attention head: in, This represents the weight vector of the k-th attention head.

5. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, The drug substructure features for: in, For the first l Layer nodes i The pre-activated output, SAGPooling is a self-attention graph pooling operation.

6. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, The processing method of the external substructure sensing layer is as follows: in, This represents the concatenated output of K attention heads. , Represents trainable weights. , Represents the bias parameter. This indicates a summation operation over the second dimension.

7. The method for predicting drug synergy based on multi-level information fusion according to claim 1, characterized in that, The fusion feature The method of obtaining it is: in, As a system-level feature, Characterized at the molecular level. , They are respectively , The projection, It is a multi-level interaction factor.

8. A drug synergy prediction system based on multi-level information fusion, characterized in that, include: The data acquisition module acquires the SMILES structural formulas of drugs, gene expression profiles of cancer cell lines, and drug combination response data. The data processing module converts the SMILES structural formula into a drug molecule map, standardizes the gene expression profile, and labels the samples according to the drug synergy score. The system-level feature extraction module uses a graph converter layer to encode initial drug features and a multilayer perceptron to encode cell line features, constructing a hypergraph containing drug nodes and cell line nodes. The hypergraph is then convolved using a hypergraph neural network to extract system-level features. ; The molecular-level feature extraction module employs a graph attention network to perform multi-layer convolution on the drug molecule graph, extracts the pre-activation outputs of each layer, and aggregates them using self-attention graph pooling to obtain drug substructure features. The drug substructure features are detected through an external substructure sensing layer. Perform multi-head attention transformation to extract drug interstitial structural similarity features. ; The similarity features of the drug mesenchymal structure Cell line characteristics converted by multilayer sensor By splicing the data, molecular-level features can be obtained. ; The multi-level information fusion prediction module analyzes the system-level features. With the aforementioned molecular-level features Projection is performed, interaction weights are calculated using an attention mechanism, and these interaction weights are multiplied by the basic interaction features to obtain the fused features. Based on the fusion features Output drug synergistic prediction results.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the drug synergy prediction method based on multi-level information fusion as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a drug synergy prediction method based on multi-level information fusion as described in any one of claims 1-7.

Images