A method and system for predicting drug IC50 based on molecular structure and gene expression

Abstract

This invention discloses a method and system for predicting drug IC50 based on molecular structure and gene expression. By comprehensively characterizing the drug's molecular structure, chemical structure and property information are extracted using molecular fingerprinting and molecular mapping, and then fused with the cell's gene expression matrix to establish a unified feature expression system. A gene regulatory network reflecting potential regulatory relationships between genes is inferred based on a variational autoencoder (VAE). Furthermore, a multi-layer feature extraction mechanism combining global and local information is constructed. Global information is used to learn the overall regulatory structure between genes in the entire regulatory network through a graph neural network, while local information is captured by dividing and analyzing subgraphs to identify local interactions between drugs and key regulatory factors. This invention can more accurately characterize the mechanism of drug influence on cellular systems, significantly improve the prediction accuracy of IC50 values ​​and the interpretability of the model, and has good adaptability and broad application prospects.

Description

Technical Field

[0001] This invention belongs to the field of drug response prediction technology, and in particular relates to a method and system for predicting drug IC50 based on molecular structure and gene expression. Background Technology

[0002] In research fields such as anti-tumor drugs, high-throughput drug screening, and precision medicine, predicting the sensitivity of drugs to cells or individuals is of great significance. Among these, the half-maximal inhibitory concentration (IC50), as an important indicator of a drug's ability to inhibit cell proliferation or survival, is widely used in various stages of drug development. With the continuous development of molecular biology and computational biology, researchers have proposed various methods for predicting IC50 values, mainly including experimental determination and computational modeling based on bioinformatics.

[0003] Traditional research commonly employs experimental perturbation combined with high-throughput sequencing technology. This involves applying drugs to cells and measuring subsequent changes in gene expression to indirectly obtain the drug's effects, then using the expression matrix to construct an IC50 prediction model. However, this approach is not only time-consuming and costly, but also requires strict experimental design and operating conditions, resulting in poor reproducibility and large batch-to-batch variability. Especially in large-scale drug screening and cross-tissue prediction, its efficiency and stability are insufficient to meet practical needs.

[0004] In recent years, with the advancement of gene regulatory network construction technology, some studies have attempted to predict IC50 values ​​based on transcriptional regulatory network information. These methods construct regulatory relationship maps between genes to uncover the potential impact of drugs on the regulatory network, thereby aiding in the assessment of cellular drug sensitivity. However, in practical applications, most existing regulatory network models focus on local pathways or specific functional modules, failing to adequately consider the interactions between transcription factors and the cross-influence between different regulatory sub-maps, thus unable to comprehensively capture the complex regulatory mechanisms induced by drugs. This limitation restricts the model's expressive power, often making it difficult to obtain prediction results with high accuracy and high generalization performance.

[0005] Furthermore, although existing research has recognized the importance of drug structure to its biological activity and has attempted to introduce molecular structural features for drug response modeling, most current methods still treat drug information and intracellular gene expression information separately, lacking a unified fusion mechanism and failing to fully leverage the complementarity of these two types of information in reflecting drug action mechanisms. Traditional structural information modeling methods often rely on simple characterization techniques such as molecular fingerprints, physicochemical properties, or topological structures, making it difficult to comprehensively uncover key features in drug structures; at the same time, they ignore the dynamic expression responses triggered by different drugs in specific cellular contexts, resulting in insufficient model fitting ability to actual IC50 values.

[0006] In summary, current IC50 prediction research still faces numerous challenges. While experimental methods can provide effective data support to some extent, they are resource-intensive and limited in application scenarios. Modeling methods based on gene regulatory networks neglect the interactions between transcription factors, leading to insufficient characterization of expression relationships. The disconnect between structural information modeling and expression data also limits the improvement of the model's comprehensive representational capabilities. Therefore, a new strategy is urgently needed to fully integrate the structural information and gene expression characteristics of drugs, constructing a more biologically representative and predictive model framework to achieve more accurate and efficient prediction of drug IC50 values. This will have significant implications for improving drug screening efficiency, reducing R&D costs, and advancing the development of personalized medicine. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method and system for predicting drug IC50 based on molecular structure and gene expression. The specific technical solution provided by this invention is as follows:

[0008] A method for predicting the IC50 of drugs based on molecular structure and gene expression, comprising:

[0009] Step S1: Generate molecular fingerprints and molecular maps for characterizing molecular structures based on drug SMILES;

[0010] Step S2: Obtain the gene expression matrix of the tumor tissue before drug addition and reduce its dimensionality;

[0011] Step S3: Load the drug molecule information into the dimensionality-reduced gene expression matrix;

[0012] Step S4: Infer the gene regulatory network diagram based on the gene expression matrix containing drug molecule information;

[0013] Step S5: Predict the IC50 value of the drug based on the gene regulatory network diagram:

[0014] The GAT network is used to extract the full-map attention features of the gene regulation network graph. At the same time, the gene regulation network graph is divided into multiple subgraphs according to transcription factors, and the GAT network is used to extract the corresponding subgraph attention features of each subgraph. The attention features of each subgraph are fused through the attention mechanism and input together with the full-map attention features into a fully connected neural network to output the IC50 prediction value.

[0015] Further: In step S1, drug SMILES is converted into molecular objects by calling RDKit library functions, and then a molecular fingerprint with specified parameters is generated; atomic and chemical bond information is obtained from drug SMILES by calling RDKit library functions, and a molecular graph is constructed with atoms as nodes and chemical bonds as edges, with atomic properties as atomic features and chemical bond properties as bond features.

[0016] Further: In step S2, dimensionality reduction of the gene expression matrix is ​​achieved by constructing an interpretable fully connected neural network. This interpretable neural network includes an input layer for inputting the gene expression matrix data, an output layer for outputting the dimensionality-reduced gene expression matrix data, and several hidden layers located between the input and output layers. The neurons in each layer are connected to all neurons in adjacent layers. Interpretability of the neural network is achieved by introducing a mask matrix, which is used to set the validity of neuron connections between adjacent layers. If the mask matrix... M The i Line 1 j Column elements M i,j = 1, then it means the first one in the previous layer. i The first neuron and the second neuron in the subsequent layer j If the connections between neurons are effective, then M i,j = 0 indicates that the connection is disabled.

[0017] Further, step S3 includes:

[0018] Molecular fingerprints are input into a GRU network to extract fingerprint features;

[0019] Molecular graphs are input into a GCN network to extract molecular graph features;

[0020] The fingerprint features and molecular map features are spliced ​​together to form the drug molecule fusion features;

[0021] The drug molecule fusion features are multiplied element-wise with the dimension-reduced gene expression matrix to obtain a gene expression matrix containing drug molecule information. If the dimension of the dimension-reduced gene expression matrix is ​​inconsistent with that of the drug molecule fusion features during element-wise multiplication, the gene expression matrix or the drug molecule fusion features are expanded and completed through a broadcast mechanism.

[0022] Further: In step S4, the gene expression matrix containing drug molecule information is input into the DeepSEM network to obtain the corresponding gene regulation network graph; the DeepSEM network is a variational autoencoder (VAE), which includes an encoder, a GRN layer, an inverse GRN layer, and a decoder connected in sequence; the encoder and decoder are both multilayer perceptrons, and their weights are shared among different genes; the GRN layer and the inverse GRN layer are both gene interaction matrices used to construct the gene regulation network graph.

[0023] Furthermore, in step S5, it is assumed that the gene regulatory network diagram is represented as follows: , V For a set of nodes, E As an edge set, the gene regulatory network graph is divided into multiple subgraphs according to transcription factors, including: selecting seed nodes. According to the seed node Get the set of one-hop neighbor nodes Then the subgraph The set of nodes is subgraph The set of edges is , For a set of nodes Different nodes in Represents a node The edges between; the seed node Transcription factors are obtained from the gene expression matrix based on an interpretable neural network.

[0024] Preferably, in step S5, the attention mechanism used to fuse the attention features of each subgraph employs a Transformer encoder. The Transformer encoder consists of multiple encoder blocks, each of which includes input embedding, position embedding, a multi-head self-attention mechanism, and a feedforward neural network. The attention features of each subgraph are processed as input sequences through position encoding to preserve the order fine-grained system, then mapped into query vectors, key vectors, and value vectors. Similarity is calculated through dot product, and normalization is performed using the softmax function to generate attention weights. Next, the attention weights are multiplied by the value vectors and weighted summed to obtain the output of the multi-head self-attention mechanism. Finally, the feedforward neural network performs a nonlinear transformation on the output of the multi-head self-attention mechanism to obtain the subgraph attention fusion features.

[0025] Further, in step S5, the fully connected neural network includes an input layer for inputting subgraph attention fusion features and full-graph attention features, an output layer for outputting IC50 prediction values, and several fully connected layers located between the input layer and the output layer, wherein the neurons in each layer are connected to all neurons in the adjacent layer.

[0026] A drug IC50 prediction system based on the above method includes a fingerprint feature generation module, a molecular graph feature generation module, a gene expression matrix dimensionality reduction module, a drug molecule information loading module, a gene regulatory network inference module, and an IC50 prediction module.

[0027] The fingerprint feature generation module generates molecular fingerprints based on drug smiles and extracts fingerprint features from the molecular fingerprints using a GRU network. The molecular graph feature generation module generates molecular graphs based on drug smiles and extracts molecular graph features from the molecular graphs using a GCN network. The gene expression matrix dimensionality reduction module reduces the dimensionality of the gene expression matrix. The drug molecule information loading module concatenates the fingerprint features and molecular graph features into a drug molecule fusion feature and multiplies the drug molecule fusion feature element-wise with the dimensionality-reduced gene expression matrix to obtain a gene expression matrix containing drug molecule information. The gene regulation network inference module inputs the gene expression matrix containing drug molecule information into a DeepSEM network to obtain the corresponding gene regulation network graph. The IC50 prediction module outputs the predicted IC50 value of the drug based on the gene regulation network graph.

[0028] Furthermore, the IC50 prediction module includes a full-graph attention feature extraction module, a subgraph partitioning module, a subgraph attention feature extraction and fusion module, and a fully connected neural network. The full-graph attention feature extraction module is used to extract full-graph attention features of the gene regulation network graph using a GAT network. The subgraph partitioning module is used to divide the gene regulation network graph into multiple subgraphs according to transcription factors. The subgraph attention feature extraction and fusion module is used to extract the corresponding subgraph attention features of each subgraph using a GAT network and fuse the attention features of each subgraph through an attention mechanism. The fully connected neural network is used to output the IC50 prediction value based on the input full-graph attention features and the fused subgraph attention features.

[0029] Compared to existing technologies that rely solely on gene expression matrices or drug perturbation experimental data for IC50 prediction, this invention achieves synergistic modeling of drug structural features and intracellular response mechanisms by comprehensively characterizing drug information and deeply integrating it with gene expression data. This fundamentally enhances the predictive model's ability to understand drug bioactivity. The characterization of drug information not only encompasses static features such as molecular fingerprints and chemical structures but also, through its combination with gene expression matrices, dynamically reflects the regulatory effects of drugs in specific biological environments, thus addressing the problem of insufficient description of drug action mechanisms in existing methods.

[0030] Furthermore, this invention introduces an interpretable neural network to effectively reduce the dimensionality of high-dimensional gene expression data, preserving biologically significant features. Based on this, a variational autoencoder (VAE) is used to infer gene regulatory networks, establishing a global network structure that reflects the potential regulatory relationships between genes. Through the deep generative capabilities of VAEs, not only is the modeling ability for complex nonlinear relationships between genes improved, but the generalization and reconstruction accuracy of gene regulatory networks are also enhanced, allowing the model to learn the potential distribution of the data more fully.

[0031] At the information extraction level, this invention constructs a gene regulatory network feature expression strategy that combines global and local approaches. On the one hand, it utilizes graph neural networks to process the entire gene regulatory network, extracting global regulatory structural information and revealing the overall regulatory pattern among genes. On the other hand, it focuses on sub-graphs of regulatory factors closely related to drug action, extracting the expression features of key genes in local sub-networks, thereby enhancing the model's ability to identify the interaction between drugs and target genes. This strategy of fusing global and local information not only improves the model's understanding of drug action mechanisms but also enhances the biological interpretability and credibility of the IC50 value prediction results.

[0032] In summary, this invention significantly improves the accuracy, robustness, and interpretability of the IC50 prediction model by fusing drug molecular structure information with gene expression data, constructing a gene regulatory network based on VAE, and employing a global-local information joint extraction strategy. It has stronger generalization ability and clinical application potential, providing important technical support for efficient drug screening and personalized medicine. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0034] Figure 1 This is a schematic diagram of the technical solution framework provided by an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0036] Example 1

[0037] This embodiment provides a method for predicting the IC50 of drugs based on molecular structure and gene expression, including the following steps:

[0038] 1. Generating molecular fingerprints and molecular maps for characterizing drug molecular structures based on SMILES

[0039] SMILES (Simplified Molecular-Input Line-Entry System) is a chemical language for describing molecular structures. It represents the atoms, bonds, and ring structures of molecules as linear strings, enabling computers to efficiently process and analyze molecular information. ECFP (Extended Connectivity Fingerprints) is a topological molecular fingerprinting method that effectively captures the ring structures, branching structures, and interatomic connections in drug molecules. By recursively considering the structural information of each atom and its neighborhood, ECFP molecular fingerprints can describe the molecular topology as a fixed-length binary vector. Molecular graphs can represent drug molecules as graph structures, with atoms as nodes and chemical bonds as edges, providing a more intuitive representation of the intermolecular connections.

[0040] The SMILES string is converted into a molecular object by calling the MolFromSmiles function of the RDKit library, and then the GetMorganFingerprintAsBitVect function is called to generate an ECFP fingerprint with specified parameters (radius, length). In this embodiment, a circular topological fingerprint with a radius of 2 and a length of 1024 is used. Similarly, after converting the SMILES string into a molecular object, atomic and bond information is obtained by calling the corresponding functions in the RDKit library, and then this information is used to construct a molecular graph. For example, the NetworkX library is used to create and visualize the molecular graph. Nodes in the molecular graph represent atoms, edges represent chemical bonds, atomic properties are called atomic features, and bond properties are called bond features. Examples of atomic and bond features are shown in Table 1.

[0041] Table 1 Examples of atomic and bond characteristics

[0042]

[0043] 2. Obtain and dimensionality-reduced gene expression matrix of tumor tissue before drug addition.

[0044] Gene expression data from tumor tissues before drug administration were extracted to construct a gene expression matrix (vector), which was then preprocessed for dimensionality reduction. Gene expression matrices are typically high-dimensional, containing a large number of genes and their corresponding expression values. Common dimensionality reduction methods for more efficient processing of this data include Principal Component Analysis (PCA), Singular Value Decomposition (SVD), and Non-negative Matrix Factorization (NMF). Interpretable neural networks (such as neural network visualization and neuron selection) can also be used for dimensionality reduction. This method selects a subset of neurons to extract the most relevant features, reducing redundant information and retaining the most biologically meaningful parts of the data. Dimensionality reduction not only makes the model more efficient but also improves its interpretability, allowing researchers to clearly see which genes and transcription factors play key roles in drug response, gene regulation, and other tasks. During dimensionality reduction, the features of genes and transcription factors are typically reflected in different layers of the neural network. By selecting and weighting a subset of neurons in the neural network, fine-grained feature learning of genes and transcription factors can be performed at each layer of the network. Selecting a subset of neurons allows the network to focus on the most relevant genes or transcription factors, thereby improving learning efficiency and accuracy.

[0045] For example, a gene expression matrix can be input into an interpretable neural network. Interpretability is achieved by introducing a mask matrix, where each neuron is connected to a specific subset of the input data, and the remaining weights are set to zero. Assume the mask matrix is ​​M, and the mask elements are M... i,j = 1 indicates that there is a connection between input feature i and neuron j, M i,j = 0 indicates that the connection is disabled. The forward propagation of a fully connected neural network can be represented as: Where X is the network input and W is the network weight matrix. This indicates element-wise multiplication of matrices, with Z representing the network output. During backpropagation, the mask matrix M determines that some elements of matrix W in the region have zero gradients; that is, the weights at these locations will not be updated. Where L represents the loss function, This represents the gradient of the loss with respect to the output.

[0046] 3. Load drug molecule information into the gene expression matrix

[0047] First, the molecular fingerprint is input into a gated neural network (GRU) to extract fingerprint features, and the molecular graph is input into a graph convolutional network (GCN) to extract molecular graph features. Then, the fingerprint features and molecular graph features are fused together to form the drug molecule fusion feature, i.e.:

[0048] ,

[0049] Among them, Mf and M g These represent fingerprint features and molecular map features, respectively. M represents vector concatenation. d This refers to the fusion characteristics of the spliced ​​drug molecules.

[0050] Finally, the drug molecule fusion features are multiplied by the dimension-reduced gene expression matrix to obtain a gene expression matrix containing drug molecule information. Specifically, the dimension-reduced gene expression matrix is ​​expanded and completed using a broadcast mechanism to make its dimension consistent with the drug molecule fusion features. Then, the gene expression matrix and the drug molecule fusion features are multiplied element-wise (corresponding element-wise) to obtain the final gene expression matrix containing drug molecule information.

[0051] In this way, the characteristic information of drug molecule structure is introduced into gene expression data, thereby providing more sufficient features for subsequent IC50 value prediction.

[0052] 4. Inferring gene regulatory network diagrams based on gene expression matrices containing drug molecule information.

[0053] A gene regulatory network can be defined as a graph, which represents genes and their interactions in the form of nodes and edges, where nodes represent genes and edges represent the regulatory relationships (activation or inhibition) between genes.

[0054] A gene expression matrix containing drug molecule information is input into a DeepSEM network to obtain the corresponding gene regulatory network diagram. For example... Figure 1 As shown, the DeepSEM network (Modeling gene regulatory networks using neural network architectures[J]. Nature Computational Science, 2021, 1(7): 491-501) is a variational autoencoder (VAE) consisting of four modules: encoder, GRN layer, inverse GRN layer, and decoder. Both the encoder and decoder are multilayer perceptrons that take a single gene as input, and their weights are shared across different genes. The GRN and inverse GRN layers are gene interaction matrices that explicitly model the gene regulatory network graph and guide the information flow of the neural network. The VAE encoder can learn low-dimensional feature representations in the latent space of the input data, thereby capturing the latent structure in the data. In this way, the information in the gene expression matrix and the effects of drugs are compressed into the latent space, providing important features for subsequent gene regulatory network inference in the GRN layer.

[0055] The encoding principle of VAE is to map input data to a structured latent space through probabilistic modeling. Unlike the deterministic encoding of traditional autoencoders, VAE encoders map input data... x The parameter mapped to the probability distribution in the latent space: mean (Indicating the central location of the latent variables) and variance (Describes the degree of dispersion of the data).

[0056] Right now

[0057]

[0058] This design assumes latent variables It follows a Gaussian distribution, thus giving the model the ability to generate diverse results.

[0059] Due to direct sampling This can lead to gradient non-differentiability (sampling is a random process). VAE introduces the reparameterization technique to separate randomness and preserve differentiability, i.e.

[0060]

[0061]

[0062] For GRN layers and inverse GRN layers, we have:

[0063]

[0064]

[0065] For the decoder, we have:

[0066]

[0067] This represents the input data after decoding and reconstruction.

[0068] 5. Predicted IC50 values ​​of drugs based on gene regulatory network diagrams

[0069] like Figure 1 As shown, the gene regulation network graph is input into a graph attention network (GAT network) to extract the full graph attention features; at the same time, the gene regulation network graph is divided into multiple subgraphs according to transcription factors, and the GAT is used to extract the corresponding subgraph attention features of each subgraph; then the attention features of each subgraph are fused through an attention mechanism, and the fused features and the full graph attention features are input into a fully connected neural network to output the IC50 prediction value.

[0070] The Genetic Artificial Intelligence (GAI) network is a neural network architecture that uses information about the relationships between nodes to calculate node weights. In a gene regulation network graph, nodes represent genes, and edges represent the regulatory relationships between genes. Through GAI, the model can automatically learn the importance of each node (gene) in the graph structure and weight neighboring nodes through an attention mechanism, thereby extracting complex dependencies between genes. This allows the model to not only identify direct effects between genes but also capture long-distance dependencies.

[0071] In gene regulatory network diagrams, different genes may exhibit different interaction patterns under different transcription factors. Therefore, subdividing the gene regulatory network into subgraphs can help to analyze local gene networks under specific transcription factors in depth. Based on the different transcription factors, by extracting corresponding subgraphs, the model can treat each transcription factor as a separate local graph and perform local graph attention calculations. This allows the model to focus on the influence of different transcription factors on gene expression and calculate attention weights separately for gene subsets of different transcription factors.

[0072] Assuming the gene regulatory network diagram is represented as follows: ,in, V Represents a set of nodes. E Represents the edge set. Different subgraphs are obtained based on specific transcription factors, including: selecting seed nodes. ,in These are transcription factors obtained from the gene expression matrix using an interpretable neural network, based on seed nodes. Get the set of one-hop neighbor nodes Then, subgraph Node set subgraph edge set The features of the nodes are the latent variables z obtained after passing through the VAE encoder.

[0073] After extracting multiple local subgraphs, an attention mechanism is used to calculate the attention relationships between these subgraphs. This process allows the model to assess the associations between different transcription factors and their impact on overall gene expression. By calculating the attention weights for each subgraph, the model can identify which transcription factors are most critical to gene expression and perform further information fusion accordingly. Specifically, the attention mechanism employs a Transformer encoder, which consists of multiple repeating encoder blocks. Each block contains input embeddings, position embeddings, a multi-head self-attention mechanism, and a feedforward neural network. The multi-head self-attention mechanism generates new representations by calculating the relationships between different parts of the input sequence. Specifically, the input sequence is processed through positional encoding to preserve the order coherence, and then mapped into query vectors, key vectors, and value vectors. These vectors are similarized using dot products and normalized using a softmax function to generate attention weights. Finally, the attention weights are multiplied by the value vectors and weighted summed to obtain the final output. The feedforward neural network performs a non-linear transformation on the output of the self-attention mechanism to further extract features.

[0074] In addition to calculating attention for local subgraphs, the model also computes global graph attention for the entire gene regulatory network. Global graph attention provides a holistic view of gene interactions, capturing global dependencies between all genes. By fusing global graph attention with the attention of each subgraph, the model can effectively combine local and global information to comprehensively understand the role of genes in drug response.

[0075] Fully connected neural networks, with their multiple fully connected layers, combine multi-level graph attention features to generate a comprehensive feature vector suitable for predicting drug resistance (IC50 value). The final IC50 prediction is then obtained from the output layer of the fully connected network. In this way, the model can not only identify the role of a single gene but also understand the complex interactions between different genes and transcription factors, thereby accurately predicting drug resistance.

[0076] In this embodiment, the fully connected neural network used to output the IC50 prediction value was fully trained and validated on a pre-constructed dataset, which can guarantee the accuracy of the prediction results to a certain extent. Regarding the construction of the dataset, this embodiment conducted a large number of drug experiments by establishing a PDX (patient-derived xenograft) model using tumor tissue to obtain the actual IC50 values ​​of the corresponding drugs. Before the drug experiments, gene expression data of tumor tissue before drug addition and SMILES of each drug molecule structure were extracted.

[0077] Example 2

[0078] Based on the above method, this embodiment provides a drug IC50 prediction system based on molecular structure and gene expression. The system mainly includes the following modules:

[0079] The fingerprint feature generation module is used to generate molecular fingerprints based on drug SMILES and extract fingerprint features from the molecular fingerprints using a GRU network.

[0080] The molecular graph feature generation module is used to generate molecular graphs based on drug SMILES and extract molecular graph features from the molecular graphs using a GCN network.

[0081] The gene expression matrix dimensionality reduction module is used to reduce the dimensionality of the gene expression matrix.

[0082] The drug molecule information loading module is used to concatenate fingerprint features and molecular graph features into drug molecule fusion features, and then multiply the drug molecule fusion features element-wise with the dimension-reduced gene expression matrix to obtain a gene expression matrix containing drug molecule information.

[0083] The gene regulation network inference module is used to input the gene expression matrix containing drug molecule information into the DeepSEM network to obtain the corresponding gene regulation network diagram.

[0084] The IC50 prediction module specifically includes a full-graph attention feature extraction module, a subgraph partitioning module, a subgraph attention feature extraction and fusion module, and a fully connected neural network. The full-graph attention feature extraction module uses a GAT network to extract full-graph attention features from the gene regulation network graph. The subgraph partitioning module divides the gene regulation network graph into multiple subgraphs based on transcription factors. The subgraph attention feature extraction and fusion module uses a GAT network to extract the corresponding subgraph attention features for each subgraph separately and then fuses these subgraph attention features using an attention mechanism. The fully connected neural network outputs the predicted IC50 value based on the input full-graph attention features and the fused subgraph attention features.

[0085] The above system can execute the drug IC50 prediction method based on molecular structure and gene expression as described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to the drug IC50 prediction method based on molecular structure and gene expression provided in Embodiment 1 of this invention.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general-purpose hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to enable a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting drug IC50 based on molecular structure and gene expression, characterized in that, The method comprises the following steps: Step S1, generating a molecular fingerprint and a molecular graph for characterizing the molecular structure based on the drug SMILES; Step S2, obtaining a gene expression matrix of tumor tissue before drug addition and reducing dimension; Step S3, loading drug molecular information into the reduced gene expression matrix; Step S4, inferring a gene regulatory network graph based on the gene expression matrix containing drug molecular information; Step S5, predicting the IC50 value of the drug based on the gene regulatory network graph: The GAT network is used to extract the full-graph attention features of the gene regulatory network graph; meanwhile, the gene regulatory network graph is divided into multiple subgraphs according to transcription factors, and the GAT network is used to extract the corresponding subgraph attention features of each subgraph; The subgraph attention features are fused through an attention mechanism and input into a fully connected neural network together with the full-graph attention features to output the IC50 prediction value.

2. The drug IC50 prediction method of claim 1, wherein: In step S1, the drug SMILES is converted into a molecular object by calling the RDKit library function, and then a molecular fingerprint with specified parameters is generated; the atom and chemical bond information is obtained from the drug SMILES by calling the RDKit library function, and the molecular graph is constructed with atoms as nodes and chemical bonds as edges, the atom attributes as atom features, and the chemical bond attributes as bond features.

3. The drug IC50 prediction method of claim 1, wherein: In step S2, the dimensionality reduction of the gene expression matrix is realized by constructing an interpretable fully connected neural network, which comprises an input layer for inputting gene expression matrix data, an output layer for outputting reduced gene expression matrix data, and a plurality of hidden layers between the input layer and the output layer, each layer containing neurons connected with all neurons in the adjacent layer; The explainability of the neural network is achieved by introducing a mask matrix for setting the effectiveness of the neuron connections between adjacent layers, if the mask matrix M of the first i row and the first j column element M i,j = 1, then the connection between the first i neuron of the previous layer and the first j neuron of the subsequent layer is effective, if M i,j = 0 then the connection is disabled.

4. The drug IC50 prediction method of claim 1, wherein Step S3 comprises: Inputting the molecular fingerprint into the GRU network to extract fingerprint features; Inputting the molecular graph into the GCN network to extract molecular graph features; Concatenating the fingerprint features and the molecular graph features into drug molecular fusion features; Element-wise multiplying the drug molecular fusion features and the reduced gene expression matrix to obtain a gene expression matrix containing drug molecular information; when the dimensions of the reduced gene expression matrix and the drug molecular fusion features are inconsistent, the gene expression matrix or the drug molecular fusion features is expanded and completed through the broadcast mechanism.

5. The drug IC50 prediction method of claim 1, wherein: In step S4, the gene expression matrix containing drug molecular information is input into the DeepSEM network to obtain the corresponding gene regulatory network graph; the DeepSEM network is a variational autoencoder VAE, comprising an encoder, a GRN layer, an inverse GRN layer and a decoder connected in sequence; the encoder and the decoder are both multilayer perceptrons, and their weights are shared among different genes; the GRN layer and the inverse GRN layer are both gene interaction matrices, which are used to construct the gene regulatory network graph.

6. The drug IC50 prediction method of claim 3, wherein, In step S5, assuming that the gene regulatory network graph is represented as , V is a node set, E is an edge set, the gene regulatory network graph is divided into multiple sub-graphs according to transcription factors, including: selecting a seed node , obtaining a one-hop neighbor node set according to the seed node , then the node set of the sub-graph is , and the edge set of the sub-graph is , u , v is a different node in the node set , represents an edge between the nodes u , v ; the seed node is a transcription factor obtained from a gene expression matrix according to an interpretable fully connected neural network.

7. The drug IC50 prediction method of claim 1, wherein In step S5, the attention mechanism for fusing the subgraph attention features adopts a Transformer encoder composed of multiple encoder blocks, each of which contains input embedding, position embedding, multi-head self-attention mechanism and feedforward neural network; each subgraph attention feature is processed by position coding to preserve the sequential fine-grained system, then mapped into query vector, key vector and value vector and the similarity is calculated by dot product, and normalized using the softmax function to generate attention weights; then the attention weights are multiplied with the value vector and weighted summation is performed to obtain the output of the multi-head self-attention mechanism; finally, the output of the multi-head self-attention mechanism is nonlinearly transformed by the feedforward neural network to obtain the subgraph attention fusion feature.

8. The drug IC50 prediction method of claim 7, wherein, In step S5, the fully connected neural network includes an input layer for inputting the subgraph attention fusion feature and the whole graph attention feature, and an output layer for outputting the IC50 prediction value, and a plurality of fully connected layers between the input layer and the output layer, each layer containing neurons connected to all neurons in the adjacent layer.

9. A drug IC50 prediction system based on the method of any one of claims 1 to 8, characterized by, The fingerprint feature generation module, the molecular graph feature generation module, the gene expression matrix dimension reduction module, the drug molecule information loading module, the gene regulatory network inference module and the IC50 prediction module are included. The fingerprint feature generation module is configured to generate a molecular fingerprint according to the drug SMILES, and extract fingerprint features from the molecular fingerprint using a GRU network; the molecular graph feature generation module is configured to generate a molecular graph according to the drug SMILES, and extract molecular graph features from the molecular graph using a GCN network; the gene expression matrix dimension reduction module is configured to reduce the dimension of the gene expression matrix; the drug molecule information loading module is configured to concatenate the fingerprint features and the molecular graph features into drug molecule fusion features, and multiply the drug molecule fusion features with the reduced gene expression matrix element by element to obtain a gene expression matrix containing drug molecule information; the gene regulatory network inference module is configured to input the gene expression matrix containing drug molecule information into a DeepSEM network to obtain a corresponding gene regulatory network graph; and the IC50 prediction module is configured to output the IC50 prediction value of the drug according to the gene regulatory network graph.

10. The drug IC50 prediction system of claim 9, wherein, The IC50 prediction module includes a whole graph attention feature extraction module, a subgraph division module, a subgraph attention feature extraction and fusion module, and a fully connected neural network; the whole graph attention feature extraction module is configured to extract whole graph attention features of the gene regulatory network graph using a GAT network, the subgraph division module is configured to divide the gene regulatory network graph into multiple subgraphs according to transcription factors; the subgraph attention feature extraction and fusion module is configured to extract respective subgraph attention features of each subgraph using a GAT network, and fuse the subgraph attention features through an attention mechanism; and the fully connected neural network is configured to output the IC50 prediction value according to the whole graph attention features and the fused subgraph attention features input together.

Images