Fusion network representation and deep learning-based efficacy evaluation method for traditional chinese medicine in colorectal cancer
By integrating network representation and deep learning methods, a traditional Chinese medicine-compound heterogeneity network and efficacy evaluation framework were constructed, which solved the problem of the correlation between the synergistic effect of multiple components of traditional Chinese medicine and the mechanism of single-cell diseases, and realized the accurate efficacy evaluation and efficient screening of traditional Chinese medicine for colorectal cancer.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU NORMAL UNIVERSITY
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies lack an effective computational framework to integrate the synergistic effects of multiple components in traditional Chinese medicine (TCM) with the molecular mechanisms of diseases at the single-cell level, resulting in low efficiency of repurposing old TCM drugs and difficulty in accurately locating TCM drugs with potential therapeutic effects.
This study employs a method that integrates network representation and deep learning. By using graph isomorphic networks (GIN) and network representation learning algorithms to capture the chemical structure and network topology features of traditional Chinese medicine (TCM) components, and combining this with single-cell differential expression analysis, a TCM-compound heterogeneous network is constructed. Furthermore, a deep learning model is used to extract protein sequence features, thereby generating embedding vectors between TCM active ingredients and disease genes, and constructing an end-to-end efficacy evaluation framework.
This study achieved a cross-dimensional correlation between the multi-component characteristics of traditional Chinese medicine and single-cell pathological mechanisms, accurately screened out traditional Chinese medicines with potential therapeutic effects on colorectal cancer, and improved the conversion efficiency of old traditional Chinese medicines for new uses.
Smart Images
Figure CN122158186A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioinformatics technology, specifically computational pharmacology and artificial intelligence research in traditional Chinese medicine, and relates to a method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by integrating network representation and deep learning. Background Technology
[0002] Colorectal cancer (CRC) is one of the most common and deadliest malignant tumors worldwide, and its incidence is currently on a significant upward trend. Despite the availability of various treatment options, the high heterogeneity among individual patients and the drug resistance that develops during treatment remain key obstacles that limit clinical efficacy and affect patient prognosis.
[0003] Traditional drug discovery faces severe challenges due to high costs, lengthy cycles, and high failure rates, urgently requiring new strategies and sources to accelerate the translational medicine process. Drug repurposing, by utilizing known pharmacokinetic and toxicological data of compounds, effectively circumvents the most time-consuming preclinical stage of new drug development. Therefore, it is considered a key method to improve the success rate of drug clinical translation and rapidly obtain potential therapies. To address the challenge of the synergistic effects of multiple components, targets, and pathways in traditional Chinese medicine (TCM), network pharmacology has emerged and has become an indispensable key method for decoding the complex mechanisms of action of TCM.
[0004] The development of single-cell sequencing technology and deep learning has provided a new dimension for overcoming the heterogeneity challenge of colorectal cancer. Single-cell sequencing technology can reveal cellular diversity in the tumor microenvironment at single-cell resolution and accurately identify specific cancer-related cells. At the same time, deep learning, with its superior feature extraction and pattern recognition capabilities, can efficiently mine potential therapeutic patterns and predict drug efficacy from complex drug-target protein data.
[0005] While single-cell sequencing and deep learning provide powerful tools for precision medicine, the current field of AI drug discovery still lacks a computational framework that can effectively integrate the synergistic effects of multiple components in traditional Chinese medicine (TCM) with the molecular mechanisms of diseases at the single-cell level. This gap in the connection between the complexity of TCM components and the carcinogenic mechanisms at the single-cell level makes it difficult to accurately locate many TCMs with potential therapeutic effects, severely restricting the efficiency of repurposing existing TCM drugs. Therefore, developing a novel deep learning architecture that can cross-dimensionally link the characteristics of multiple components in TCM with single-cell pathological mechanisms has significant scientific and practical value for rapidly identifying potential therapies for colorectal cancer using existing drug libraries. Summary of the Invention
[0006] The purpose of this invention is to provide a method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by integrating network representation and deep learning. This method is used to screen traditional Chinese medicines with potential therapeutic effects. It uses graph isomorphic networks (GIN) and network representation learning algorithms to capture the chemical structure and network topology features of traditional Chinese medicine components to characterize their synergistic effects. At the same time, it combines single-cell differential expression analysis to accurately identify the core driver genes of colorectal cancer and uses protein sequence features to directly generate embedding vectors of drug targets and disease genes.
[0007] The method of this invention is as follows:
[0008] Step (1) Data acquisition and preprocessing; The acquired data includes TCM-component-target data, TCM component SMILES symbol data, colorectal cancer single-cell sequencing data, protein sequence data, and protein-protein interaction network data; Active components with potential drug-like properties are screened from the TCM library, and the SMILES symbols of these components are converted into two-dimensional molecular graph topology structures;
[0009] Step (2) Construct a traditional Chinese medicine-compound heterogeneity network; extract feature vectors of active ingredients: use GIN as encoder, construct a molecular graph with atoms as nodes and chemical bonds as edges, and generate a 300-dimensional molecular embedding through nonlinear projection; use cosine similarity algorithm to quantify the feature vectors between active ingredients and construct a molecular structure similarity matrix between ingredients; construct a compound-compound similarity network based on the similarity matrix, and add the traditional Chinese medicine-component-target data to the compound-compound similarity network to form a traditional Chinese medicine-compound heterogeneity network;
[0010] Step (3) Disease gene localization; Disease gene analysis based on colorectal cancer single-cell sequencing data, including single-cell sequencing data quality control, differential analysis and disease differential gene screening;
[0011] Step (4) Extract heterogeneous topological representations of TCM nodes in the TCM-compound heterogeneous network, and extract essential semantic features of protein sequences from protein sequence data; through a multimodal heterogeneous feature extraction strategy, TCM data and biological protein sequence information of different dimensions are uniformly mapped to a high-dimensional vector space; through end-to-end representation learning, the data barrier between the chemical space of TCM and the biological space of disease is eliminated, and the microstructure and macroscopic topological features of TCM active ingredients are captured by a deep learning model, while extracting essential semantic features of protein sequences; through feature extraction and pooling, heterogeneous features are aligned into a unified 300-dimensional embedding vector, providing standardized multidimensional feature input for subsequent quantitative evaluation of anticancer drug efficacy;
[0012] Step (5) Construct a drug efficacy evaluation framework based on end-to-end deep learning; by constructing a heterogeneous interaction space between traditional Chinese medicine and target protein, supervised learning is carried out using known drug-target associations, the extracted traditional Chinese medicine topological embedding and protein sequence embedding are fused at the feature level, a joint feature vector is constructed, and it is input into a multilayer perceptron architecture for nonlinear mapping;
[0013] Step (6) involves splicing the heterogeneous topological representation of the Chinese medicine node to be tested with the essential semantic features of the protein sequence of the core driver gene of colorectal cancer, and inputting it into the trained fully connected neural network. The fully connected neural network outputs a continuously distributed prediction score, which serves as a quantitative basis for measuring the intervention efficacy of the Chinese medicine against colorectal cancer.
[0014] This invention precisely locates core driver genes in colorectal cancer through single-cell transcriptome differential analysis and constructs a traditional Chinese medicine (TCM)-compound heterogeneity network integrating the molecular structure and attribution relationships of active ingredients. The Node2Vec algorithm is used to extract 300-dimensional heterogeneous topological representations of TCM nodes, simultaneously combined with 300-dimensional essential semantic features of protein sequences extracted based on a one-dimensional convolutional neural network. A 600-dimensional joint vector is constructed through feature concatenation and input into a trained fully connected neural network, ultimately achieving quantitative evaluation and ranking screening of the efficacy of candidate TCM interventions against oncogenes. This invention establishes a direct and quantitative link between the TCM intervention space and single-cell pathological mechanisms, realizing end-to-end evaluation of the efficacy of TCM in treating colorectal cancer. Attached Figure Description
[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0016] The specific implementation of the present invention will be described in detail below with reference to the technical solution and accompanying drawings.
[0017] like Figure 1 A method for evaluating the efficacy of traditional Chinese medicine in colorectal cancer, integrating network representation and deep learning, is described below:
[0018] Step (1) Data acquisition and preprocessing;
[0019] Data acquisition included: screening TCM-component-target data and TCM component SMILES symbol data through the TCMSP database; obtaining colorectal cancer single-cell sequencing data through the Gene Expression Organisation (GEO) database, including single-cell RNA sequencing of 63,689 cells from 23 CRC patients, including 23 primary colorectal cancer samples and 10 matched normal mucosal samples; obtaining protein sequence data through the UniProt database; and downloading protein-protein interaction (PPI) network data from the STRING database, screening edges with confidence scores higher than 700, and obtaining a PPI network containing 16,201 nodes and 236,930 edges.
[0020] Data preprocessing: according to ADME standards ( , ), Indicates oral bioavailability, This indicates drug similarity. Active ingredients with potential drug-like properties were screened from a traditional Chinese medicine (TCM) database, resulting in 3942 TCM-active ingredient pairs and 88423 TCM-potential target associations. The RDKit toolkit was used to transform the SMILES symbols of these ingredients into two-dimensional molecular graph topologies, providing structured input for subsequent GIN-based feature embedding.
[0021] Step (2) Constructing a network of isomers of traditional Chinese medicine and its compounds;
[0022] (2-1) Extraction of active ingredient feature vectors: GIN is used as the encoder, and a molecular graph is constructed with atoms as nodes and chemical bonds as edges. Local chemical environment information is iteratively aggregated and dynamically weighted using an attention mechanism, resulting in a 300-dimensional molecular embedding through nonlinear projection. Layer nodes The update rules are as follows: ;in, Represents a node In the Feature vectors after layer iteration Representing the A multilayer perceptron is used to perform non-linear mapping on the aggregated features. For nodes The set of neighboring atoms, It is for nodes The features of all direct neighbor nodes are summed and aggregated. It is a learnable parameter.
[0023] (2-2) The cosine similarity algorithm is used to quantify and characterize the feature vectors between active ingredients, and a molecular structure similarity matrix between the ingredients is constructed; active ingredients and cosine similarity , and These are respectively represented as active ingredients. and The feature vectors are used. A compound-compound similarity network is constructed based on the similarity matrix. If the similarity value between two compounds exceeds a set threshold of 95%, an edge is added between the two compounds, thus forming a compound-compound similarity network. Based on the collected traditional Chinese medicine-component-target data, the attribution relationships between traditional Chinese medicine and components are added to the compound-compound similarity network, forming a traditional Chinese medicine-compound heterogeneity network.
[0024] Step (3) Disease gene localization; Disease gene analysis based on colorectal cancer single-cell sequencing data, including single-cell sequencing data quality control, differential analysis and disease differential gene screening.
[0025] (3-1) Single-cell sequencing data quality control: Cells with fewer than 300 gene detections and fewer than 800 transcripts were removed to exclude empty droplets and low-quality cells. Cells with abnormal mitochondrial gene expression ratios greater than 10% were filtered to reduce interference from dead or stressed cells on the analysis results. Potential red blood cells and blood-derived cells in the sample were identified and removed by assessing the expression levels of hemoglobin-related genes and setting a 10% rejection threshold.
[0026] (3-2) Differential Analysis: Epithelial cells were extracted from the quality-controlled data. Based on the sample source and tissue type of each cell, the epithelial cells were grouped according to sample-tissue state combinations. The original UMI counts of each group were summarized to construct a pseudo-batch expression matrix. The pseudo-batch expression matrix contained genes... , Indicates gene In a single cell In state The original expression value in Indicates belonging to the sample And in a state The collection of all individual cells was analyzed using the edgeR package to obtain the logarithmic fold change value and significance level after multiple test correction for each gene, which was used to screen differentially expressed genes associated with colorectal cancer.
[0027] (3-3) Disease-differentiated gene screening: Screening for genes that meet the significance level in tumor tissue. Average difference multiple And the detection rate in tumor cell populations Genes that are upregulated are selected. Edges between nodes where both ends are upregulated genes are preserved in the protein-protein interaction network data. A differential gene DEG subnetwork is constructed, isolated nodes are removed, and the degree centrality and betweenness centrality of the network are calculated. Core genes are selected based on the overall network score.
[0028] node , The total number of nodes in the network, if the number of nodes With nodes If there are interactive edges, then the elements in the adjacency matrix... ;otherwise .
[0029] node normalized betweenness centrality , For nodes arrive The shortest path, This indicates the nodes passed through in these shortest paths. The number of entries;
[0030] node Overall online rating , and Let these represent the mean and standard deviation of the degree centrality of all nodes in the network, respectively. and Let represent the mean and standard deviation of the betweenness centrality of all nodes in the network, respectively.
[0031] Nodes with a network score greater than 2 are selected as core genes, i.e., the disease genes to be located.
[0032] Step (4) extracts the heterogeneous topological representation of TCM nodes in the TCM-compound heterogeneous network and extracts the essential semantic features of protein sequences from the protein sequence data. Through a multimodal heterogeneous feature extraction strategy, TCM data of different dimensions and biological protein sequence information are uniformly mapped to a high-dimensional vector space. End-to-end representation learning eliminates the data barrier between the chemical space of TCM and the biological space of the disease, and a deep learning model is used to capture the microstructure and macroscopic topological features of the active ingredients of TCM, while simultaneously extracting the essential semantic features of the protein sequences. Through feature extraction and pooling, the heterogeneous features are aligned into a unified 300-dimensional embedding vector, thereby providing standardized multidimensional feature input for subsequent quantitative evaluation of anticancer drug efficacy.
[0033] (4-1) Heterogeneous Topological Representation of Traditional Chinese Medicine Nodes: On the heterogeneous network of traditional Chinese medicine and compounds, topological feature extraction techniques based on graph representation learning are used to capture the global evolutionary patterns and local clustering characteristics of traditional Chinese medicine in the chemical composition space. Specifically:
[0034] Using a biased random walk algorithm, the return parameters are adjusted. and forward parameters Node sequences are generated to achieve a balance between exploring network homogeneity and structure during the walking process, thereby comprehensively capturing local and global topological information in the heterogeneous network of traditional Chinese medicine and compounds. Nodes To the next node transition probability , Representative node and nodes Edge weights between them This is the weighting adjustment factor. , This represents the previous node to be visited. Indicates the target node to be traversed. Represents a node and The shortest distance between them.
[0035] A traditional Chinese medicine node in the traditional Chinese medicine-compound heterogeneity network is randomly selected as the starting node. Starting from the starting node, the next node is randomly selected according to the transition probability distribution of the current node. This process is repeated until the termination condition is met. The resulting nodes are arranged in sequence as the node sequence of the random walk.
[0036] The node sequence of random walk is input into the Skip-gram model. With the optimization objective of maximizing the occurrence probability of neighboring nodes, the model is iteratively updated through the gradient descent algorithm to learn the heterogeneous topological representation of the Chinese medicine nodes.
[0037] At a given central node In the case of a neighborhood node set obtained through random walk, The probability of a node appearing in , As the central node The mapping function, Let represent the set of all nodes in the network. By maximizing the co-occurrence likelihood probability of the central node and its neighboring nodes in the feature space, the mapping function is continuously optimized using the stochastic gradient descent algorithm. The parameters are used to transform the topological proximity between nodes into a distance metric between feature vectors, thereby extracting a 300-dimensional embedding vector containing topological information of heterogeneous networks.
[0038] By executing the Node2Vec algorithm on a heterogeneous network of traditional Chinese medicine (TCM) compounds, a biased random walk strategy is used to map the complex topology into a sequence of nodes. Subsequently, deep representation learning is performed using a Skip-gram model, ultimately transforming the TCM nodes into heterogeneous topological representations of TCM nodes capable of capturing globally correlated features. .
[0039] (4-2) Essential Semantic Features of Protein Sequences: Initial mapping was achieved using a biological language vocabulary, followed by the use of a one-dimensional convolutional kernel (1D-CNN) to capture local biochemical primitive patterns reflecting active sites and physicochemical essence. Subsequently, global max pooling was used to extract signals from each feature channel, completing feature reduction while eliminating sequence length differences. Traditional Chinese medicine target proteins and colorectal cancer core driver genes were processed separately, ultimately obtaining fixed-length 300-dimensional feature vectors, denoted as the essential semantic features of the protein sequences of traditional Chinese medicine target proteins. Essential semantic features of protein sequences of core driver genes in colorectal cancer .
[0040] Step (5) constructs a drug efficacy evaluation framework based on end-to-end deep learning; by constructing a heterogeneous interaction space between traditional Chinese medicine (TCM) and target proteins, supervised learning is performed using known drug-target associations, and the extracted TCM topological embeddings and protein sequence embeddings are fused at the feature level to construct a joint feature vector, which is then input into a multilayer perceptron architecture for nonlinear mapping. This module can not only learn existing drug intervention patterns, but also perform efficacy propensity scoring on unknown "TCM-colorectal cancer gene" pairs, thereby achieving high-quality screening of colorectal cancer intervention drugs.
[0041] (5-1) Construct a training set, which includes positive and negative samples;
[0042] Constructing positive samples: Based on known protein-target interaction pairs of traditional Chinese medicine, the corresponding traditional Chinese medicine nodes are heterogeneously topologically characterized. Essential semantic features of protein sequences of target proteins in traditional Chinese medicine The features are concatenated to form a 600-dimensional joint feature vector, and the joint feature vector of traditional Chinese medicine and target protein is labeled. tag value There are a total of 88,423 positive sample data.
[0043] Constructing negative samples: A random negative sampling method is used to randomly pair up Chinese herbal medicines and target proteins in a set of unrelated traditional Chinese medicines to generate pseudo-association pairs, and to represent the heterogeneous topology of the corresponding Chinese herbal medicine nodes. Essential semantic features of protein sequences of target proteins in traditional Chinese medicine The features are concatenated to form a 600-dimensional joint feature vector, and the joint feature vector of traditional Chinese medicine and target protein is labeled. tag value There are a total of 49,799 negative samples. Since the positive and negative samples are unevenly distributed, the SMOTE algorithm is further introduced to synthesize and enhance the minority class samples, thereby achieving complete balance in the dataset and ensuring that the model possesses rigorous discriminative ability during training.
[0044] The concatenation method uses the feature concatenation operator to construct a joint feature vector of traditional Chinese medicine and target protein. ,
[0045] , This refers to feature splicing. Feature splicing effectively preserves the topological information of the drug and the biological sequence essence of the protein, avoiding information loss caused by premature linear weighting.
[0046] (5-2) Combine the feature vectors of traditional Chinese medicine-target proteins from positive and negative samples. A fully connected neural network (FCNN) was input, and its performance under different hyperparameters was verified using grid search and exploratory experiments. After multiple rounds of iteration, the optimal parameter configuration for the fully connected neural network was determined as follows: the hidden layers adopted a three-layer decreasing neuron structure with dimensions [512, 128, 32], and the learning rate was set to 0.001 to balance convergence speed and training stability; the weight decay coefficient was set to 0.001 to effectively suppress overfitting of the fully connected neural network through L2 regularization and enhance its generalization ability on the independent validation set. The binary cross-entropy loss function (BCE Loss) was used as the optimization objective during training.
[0047] Loss value , The total number of samples in the training set. For real labels, The model predicts probability values, and the final output scalar, after being mapped by the Sigmoid function, yields the predicted probability of interaction between the traditional Chinese medicine and the target protein pair. , The larger the value, the stronger the effect of the traditional Chinese medicine on the target protein.
[0048] Step (6) Characterize the heterogeneous topology of the nodes of the Chinese herbal medicine to be tested. Essential semantic features of protein sequences related to core driver genes in colorectal cancer By splicing the data, a 600-dimensional joint feature vector of traditional Chinese medicine and disease genes is obtained. ,Will The input is fed into a fully connected neural network that has been trained. The fully connected neural network outputs a continuously distributed prediction score, which serves as a quantitative basis for measuring the efficacy of this traditional Chinese medicine in intervening in colorectal cancer. Output value Defined as the efficacy tendency score of traditional Chinese medicine (TCM) against disease genes, the closer the score is to 1, the higher the degree of matching between the TCM and the oncogene in terms of biological semantics and topological association, i.e., the greater the intervention potential. By ranking all candidate TCM-disease gene pairs in descending order of their scores, core drugs with significant intervention effects against colorectal cancer can be accurately identified from the TCM database.
[0049] Results evaluation: Accuracy, recall, precision, F1 score, and area under the ROC curve (AUC) were used as evaluation metrics to verify the predictive performance of the model.
[0050] Accuracy is a core metric that measures the consistency between the model's predictions of the association between "traditional Chinese medicine and target protein" and the actual labels. It reflects the overall accuracy of the model's judgments on positive and negative samples. It represents the proportion of samples correctly predicted by the model (including both positive and negative samples) out of the total number of samples. .
[0051] Recall represents the proportion of "traditional Chinese medicine-target protein" association pairs that the model successfully predicted. Precision represents the percentage of "effective drugs" predicted by the model that are actually effective. .
[0052] F1 score is a harmonic mean that comprehensively measures the model's recall and precision in drug screening. .
[0053] AUC is defined as the area under the ROC curve, where the horizontal axis of the ROC curve represents the false positive rate (FPR) and the vertical axis represents the true positive rate (TPR). , , ,in, The number of positive samples The number of negative samples The rank of the samples after sorting by predicted probability.
[0054] in, This indicates that the model predicted the effectiveness of the traditional Chinese medicine-target protein pair, and it was indeed effective in practice. This indicates that the model predicted the Chinese medicine-target protein pair to be effective, but in reality, it was a randomly generated invalid pair. This indicates a drug-target pair that is truly ineffective and is also predicted to be ineffective by the model; This represents a real, effective drug target pair that the model incorrectly predicted as ineffective. Five metrics are calculated based on the prediction results from the algorithm design module; higher values for all metrics indicate better predictive performance of the model.
Claims
1. A method for evaluating the efficacy of traditional Chinese medicine in colorectal cancer by integrating network representation and deep learning, characterized by: Step (1) Data acquisition and preprocessing; The acquired data includes TCM-component-target data, TCM component SMILES symbol data, colorectal cancer single-cell sequencing data, protein sequence data, protein-protein interaction data, and network data; Active components with potential drug-like properties are screened from the TCM library, and the SMILES symbols of these components are converted into two-dimensional molecular graph topology structures; Step (2) Construct a traditional Chinese medicine-compound heterogeneity network; extract feature vectors of active ingredients: use GIN as encoder, construct a molecular graph with atoms as nodes and chemical bonds as edges, and generate a 300-dimensional molecular embedding through nonlinear projection; use cosine similarity algorithm to quantify the feature vectors between active ingredients and construct a molecular structure similarity matrix between ingredients; construct a compound-compound similarity network based on the similarity matrix, and add the traditional Chinese medicine-component-target data to the compound-compound similarity network to form a traditional Chinese medicine-compound heterogeneity network; Step (3) Disease gene localization; Disease gene analysis based on colorectal cancer single-cell sequencing data, including single-cell sequencing data quality control, differential analysis and disease differential gene screening; Step (4) Extract heterogeneous topological representations of Chinese medicine nodes in the Chinese medicine-compound heterogeneous network, and extract essential semantic features of protein sequences from protein sequence data; through a multimodal heterogeneous feature extraction strategy, unify the mapping of Chinese medicine data and biological protein sequence information of different dimensions to a high-dimensional vector space; through end-to-end representation learning, eliminate the data barrier between the chemical space of Chinese medicine and the biological space of disease, use deep learning models to capture the microstructure and macroscopic topological features of active ingredients of Chinese medicine, and extract essential semantic features of protein sequences at the same time; By extracting and pooling features, heterogeneous features are aligned into a unified 300-dimensional embedding vector, providing standardized multidimensional feature inputs for subsequent quantitative evaluation of anticancer drug efficacy. Step (5) Construct a drug efficacy evaluation framework based on end-to-end deep learning; by constructing a heterogeneous interaction space between traditional Chinese medicine and target protein, supervised learning is carried out using known drug-target associations, the extracted traditional Chinese medicine topological embedding and protein sequence embedding are fused at the feature level, a joint feature vector is constructed, and it is input into a multilayer perceptron architecture for nonlinear mapping; Step (6) involves splicing the heterogeneous topological representation of the Chinese medicine node to be tested with the essential semantic features of the protein sequence of the core driver gene of colorectal cancer, and inputting it into the trained fully connected neural network. The fully connected neural network outputs a continuously distributed prediction score, which serves as a quantitative basis for measuring the intervention efficacy of the Chinese medicine against colorectal cancer.
2. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by fusing network representation and deep learning as described in claim 1, characterized in that: In step (2), the active ingredient and cosine similarity , and These are respectively represented as active ingredients. and If the similarity between two compounds exceeds a set threshold, an edge is added between the two compounds to form a compound-compound similarity network.
3. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by integrating network representation and deep learning as described in claim 1, characterized in that, Step (3) is as follows: (3-1) Single-cell sequencing data quality control: remove cells with gene detection count and transcript count less than the threshold, filter cells with abnormal mitochondrial gene expression ratio greater than the threshold, and identify and remove potential red blood cells and blood-derived cells in the sample by assessing the expression level of hemoglobin-related genes. (3-2) Differential Analysis: Epithelial cells were extracted from the quality-controlled data. Based on the sample source and tissue type of each cell, the epithelial cells were grouped according to sample-tissue state combinations. The original UMI counts of each group were summarized to construct a pseudo-batch expression matrix. The pseudo-batch expression matrix contained genes... , Indicates gene In a single cell In state The original expression value in Indicates belonging to the sample And in a state The collection of all individual cells was analyzed using the edgeR package to obtain the logarithmic fold change value of each gene and the significance level after multiple test correction. (3-3) Disease differential gene screening: Genes with a significance level greater than the threshold are selected as upregulated genes in tumor tissues. The edges between nodes of protein-protein interaction network data that are both upregulated genes are retained. A differential gene DEG sub-network is constructed, isolated nodes are removed, the degree centrality and betweenness centrality of the network are calculated, and core genes are screened based on the comprehensive network score.
4. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by integrating network representation and deep learning as described in claim 3, characterized in that, Step (3-3) is as follows: node , The total number of nodes in the network, if the number of nodes With nodes If there are interactive edges, then the elements in the adjacency matrix... ;otherwise ; node normalized betweenness centrality , For nodes arrive The shortest path, This indicates the nodes passed through in these shortest paths. The number of entries; node Overall online rating , and Let these represent the mean and standard deviation of the degree centrality of all nodes in the network, respectively. and Let represent the mean and standard deviation of the betweenness centrality of all nodes in the network, respectively. Nodes with network scores greater than a threshold are selected as core genes, i.e., the disease genes to be located.
5. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by fusing network representation and deep learning as described in claim 1, characterized in that, Step (4) is as follows: (4-1) Heterogeneous topological representation of Chinese medicine nodes: On the heterogeneous network of Chinese medicine and compounds, the topological feature extraction technology based on graph representation learning is used to capture the global evolution law and local clustering characteristics of Chinese medicine in the chemical composition space. (4-2) Essential semantic features of protein sequences: Initial mapping is achieved through a biological language vocabulary, followed by the capture of local biochemical primitive patterns reflecting active sites and physicochemical essence using one-dimensional convolutional kernels; global max pooling technology is used to extract signals from each feature channel, completing feature reduction while eliminating sequence length differences; traditional Chinese medicine target proteins and colorectal cancer core driver genes are processed separately, finally obtaining fixed-length 300-dimensional feature vectors, which are denoted as the essential semantic features of protein sequences of traditional Chinese medicine target proteins. Essential semantic features of protein sequences of core driver genes in colorectal cancer .
6. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by fusing network representation and deep learning as described in claim 5, characterized in that, Step (4-1) specifically involves: Using a biased random walk algorithm, the return parameters are adjusted. and forward parameters Generate node sequences to capture local and global topological information in the heterogeneous network of traditional Chinese medicine and compounds; nodes To the next node transition probability , Representative node and nodes Edge weights between them This is the weighting adjustment factor. , This represents the previous node to be visited. Indicates the target node to be traversed. Represents a node and The shortest distance between them; A traditional Chinese medicine node in the traditional Chinese medicine-compound heterogeneity network is randomly selected as the starting node. Starting from the starting node, the next node is randomly selected according to the transition probability distribution of the current node. This process is repeated until the termination condition is met. The resulting nodes are arranged in sequence as the node sequence of the random walk. The node sequence of random walk is input into the Skip-gram model. With the optimization objective of maximizing the occurrence probability of neighboring nodes, the model is iteratively updated through the gradient descent algorithm to learn the heterogeneous topological representation of the Chinese medicine nodes. At a given central node In the case of a neighborhood node set obtained through random walk, The probability of a node appearing in , As the central node The mapping function, Let be the set of all nodes in the network; optimize the mapping function by maximizing the co-occurrence likelihood probability of the center node and its neighboring nodes in the feature space using the stochastic gradient descent algorithm. The parameters are used to transform the topological proximity between nodes into a distance metric between feature vectors, and to extract a 300-dimensional embedding vector containing heterogeneous network topology information. By executing the Node2Vec algorithm on a heterogeneous network of traditional Chinese medicine (TCM) compounds, a biased random walk strategy is used to map the complex topology into a sequence of nodes. Subsequently, deep representation learning is performed using a Skip-gram model, ultimately transforming the TCM nodes into heterogeneous topological representations of TCM nodes capable of capturing globally correlated features. .
7. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by integrating network representation and deep learning as described in claim 5, characterized in that, Step (5) specifically involves: (5-1) Construct a training set, which includes positive and negative samples; Constructing positive samples: Based on known protein-target interaction pairs of traditional Chinese medicine, the corresponding traditional Chinese medicine nodes are heterogeneously topologically characterized. Essential semantic features of protein sequences of target proteins in traditional Chinese medicine The features are concatenated to form a 600-dimensional joint feature vector, and the joint feature vector of traditional Chinese medicine and target protein is labeled. tag value ; Constructing negative samples: A random negative sampling method is used to randomly pair up Chinese herbal medicines and target proteins in a set of unrelated traditional Chinese medicines to generate pseudo-association pairs, and to represent the heterogeneous topology of the corresponding Chinese herbal medicine nodes. Essential semantic features of protein sequences of target proteins in traditional Chinese medicine The features are concatenated to form a 600-dimensional joint feature vector, and the joint feature vector of traditional Chinese medicine and target protein is labeled. tag value ; The splicing method uses feature splicing operators to construct a joint feature vector of traditional Chinese medicine and target proteins. , Indicates feature splicing; (5-2) Combine the feature vectors of traditional Chinese medicine-target proteins from positive and negative samples. A fully connected neural network was input, and its performance under different hyperparameters was verified using grid search and trial-and-error experiments. During training, a binary cross-entropy loss function was used as the optimization objective; the loss value... , The total number of samples in the training set. For real labels, The model predicts probability values, and the final output scalar, after being mapped by the Sigmoid function, yields the predicted probability of interaction between the traditional Chinese medicine and the target protein pair. , The larger the value, the stronger the effect of the traditional Chinese medicine on the target protein.
8. The method for evaluating the efficacy of traditional Chinese medicine for colorectal cancer by fusing network representation and deep learning as described in claim 1, characterized in that, Step (6) specifically involves: Heterogeneous topological characterization of the nodes of the Chinese herbal medicine under test Essential semantic features of protein sequences related to core driver genes in colorectal cancer By splicing the data, a 600-dimensional joint feature vector of traditional Chinese medicine and disease genes is obtained. , Indicates feature concatenation, The input is fed into the trained fully connected neural network, and the output value is... The efficacy tendency score of traditional Chinese medicine (TCM) on disease genes is calculated. The closer the score is to 1, the higher the degree of matching between the TCM and the oncogene in terms of biological semantics and topological association, that is, the greater the intervention potential.