Deep learning method and device for evaluating compound carcinogenic risk based on multi-modal data fusion, and storage medium

Abstract

The invention discloses a deep learning method and device for evaluating the carcinogenic risk of a compound based on multi-modal data fusion and a storage medium, and belongs to the technical field of chemical health risk evaluation, and the method comprises the steps: (1) building a carcinogenicity prediction data set; (2) constructing a cross-modal knowledge graph, and learning the cross-modal knowledge graph by using a convolutional neural network to obtain cross-modal feature representation; (3) extracting a molecular fingerprint feature representation and a molecular map feature representation, and fusing the molecular fingerprint feature representation and the cross-modal feature representation to obtain a fused feature representation; (4) a fusion model is constructed, a first deep neural network and a second deep neural network in the fusion model are used for processing different feature representations, and a classifier is used for prediction; (5) carrying out supervised training on the fusion model to optimize parameters of the fusion model; and (6) performing carcinogenic risk assessment on a to-be-detected compound by using the optimized fusion model. The method can comprehensively and accurately evaluate the carcinogenic risk of the compound and explain the carcinogenic mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of chemical health risk assessment, and specifically relates to a deep learning method, device and storage medium for assessing the carcinogenic risk of compounds based on multimodal data fusion. Background Art

[0002] High exposure to carcinogenic chemicals is one of the important factors leading to the increase in cancer incidence. Screening out potential carcinogenic chemicals and controlling them is crucial to reducing cancer incidence and protecting human life and health. Traditional methods for evaluating the carcinogenic risk of chemicals mainly rely on animal experiments. However, this method is time-consuming, inefficient, and may have ethical issues. It is difficult to assess the carcinogenic risk of a large number of chemicals one by one, so it is necessary to develop high-throughput screening technology.

[0003] Alternative methods for carcinogenicity testing, such as in vitro experiments and computer simulation methods, face a series of challenges. In vitro methods, such as cell transformation and stem cell proliferation experiments, face bottlenecks in evaluating non-genotoxic carcinogens and interpreting their modes of action. The quantitative structure-activity relationship (QSAR) method based on the structure-activity principle can quickly evaluate the potential carcinogenicity of a variety of compounds, helping researchers to screen chemical carcinogens more efficiently. However, it has problems such as limited model generalization ability and limited evaluation of complex mechanisms.

[0004] A Chinese patent document with publication number CN114743614A discloses an integrated learning method for screening carcinogenic chemicals, including: (1) constructing a chemical carcinogenicity dataset; (2) calculating molecular PubChem fingerprints and performing preprocessing and feature screening; (3) training an integrated model; (4) evaluating model performance using indicators such as accuracy; and (5) characterizing the model application domain with reference to OECD guidelines.

[0005] A Chinese patent document with publication number CN115565624A discloses a method for constructing a carcinogenicity prediction model for predicting carcinogenicity. The invention first collects data of multiple compounds, screens out carcinogenicity data and non-carcinogenicity data, and constructs a carcinogenicity prediction data set; then, a carcinogenicity-graph convolutional neural network model is constructed through the molecular graph structure data in the carcinogenicity prediction data set; an autoencoder model is constructed through the mass spectrometry data in the carcinogenicity prediction data set; then, feature fusion is performed on the feature matrices finally output by the carcinogenicity-graph convolutional neural network model and the autoencoder model, the fused feature matrices are comprehensively trained, and finally a prediction result is output; by analyzing the prediction results, when the accuracy requirements are met, a comprehensive carcinogenicity prediction model is obtained; the carcinogenicity of the compound can be predicted by the obtained comprehensive carcinogenicity prediction model.

[0006] However, the performance of the above machine learning / deep learning models does not increase linearly with the complexity, number, algorithm complexity or architecture complexity of the input descriptors, and faces the limitations of low accuracy and high false negatives. Therefore, it is necessary to establish a carcinogenicity prediction model based on multimodal data fusion, improve the accuracy of data-driven carcinogenic risk assessment, and clarify the toxicity mechanism based on the harmful outcome pathway. Summary of the invention

[0007] The present invention provides a deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion. By constructing a cross-modal knowledge graph combined with rich molecular representations to obtain a cross-modal feature representation, a carcinogenicity prediction model is constructed by combining molecular fingerprint feature representation, cross-modal feature representation and molecular graph feature representation. This method can more comprehensively and accurately evaluate the carcinogenic risk of compounds, explain their mechanism of action at the molecular level, and improve the carcinogenic risk assessment of compounds.

[0008] The specific technical solutions adopted are as follows:

[0009] A deep learning method for assessing the carcinogenic risk of compounds based on multimodal data fusion, including:

[0010] (1) Establish a carcinogenicity prediction dataset, which includes multiple compounds and their real molecular labels of whether they are carcinogenic or not;

[0011] (2) Using the compound-protein association data, protein-protein association data, protein-biological process data, protein-cell component data, and protein-molecular function data obtained from the database, a cross-modal knowledge graph is constructed; a convolutional neural network is used to learn the cross-modal knowledge graph to obtain a cross-modal feature representation;

[0012] (3) Extracting the molecular fingerprint feature representation of the compounds in the carcinogenicity prediction dataset, fusing the molecular fingerprint feature representation with the cross-modal feature representation to obtain a fused feature representation; extracting the molecular graph feature representation of the compounds in the carcinogenicity prediction dataset;

[0013] (4) constructing a fusion model combining a first deep neural network, a second deep neural network and a classifier, wherein the first deep neural network is used to extract a fusion feature vector based on the fusion feature representation, the second deep neural network is used to extract a graph feature vector based on the molecular graph feature representation of the compound, and the classifier is used to perform prediction and classification based on the fusion feature vector and the graph feature vector to obtain a molecular label prediction result;

[0014] (5) Using the carcinogenicity prediction dataset, the fusion model was supervised and trained based on the molecular label prediction results and the real molecular labels to optimize the fusion model parameters;

[0015] (6) The optimized fusion model is used to evaluate the test compound and predict the carcinogenic risk of the compound. If the prediction result is positive, the potential carcinogenic mechanism of the compound is explained.

[0016] Toxicological information of chemicals is usually scattered in various databases, including active proteins, affected pathways, animal models, and potential carcinogenic effects. Chemical decisions made based on existing computational methods are opaque and require experts in the field of toxicology to conduct labor-intensive experimental screening, which takes a lot of time to determine the toxicological association between chemicals and carcinogenic toxicity endpoints. Determining the mechanism of action is crucial for risk assessment. In order to solve problems such as data dispersion and inconsistent identifiers, a cross-modal knowledge graph is used to interconnect multimodal data, and further combined with multiple feature representations of compounds to construct a carcinogenicity prediction fusion model to predict compounds that may have carcinogenic effects. By integrating compounds from multiple databases including CPDB (Consensus Pathway Database), carcinogenicity data and related toxicology data of multiple compounds were obtained.

[0017] Specifically, in step (1), the compounds collected from the database are cleaned and used to establish a carcinogenicity prediction data set. The cleaning step includes: removing inorganic substances, organometallic compounds, salt compounds, mixtures and repeated substances. The obtained carcinogenicity prediction data set includes a total of 3762 sample compounds, including 2306 negative compounds and 1456 positive compounds.

[0018] In step (2), the cross-modal knowledge graph includes compound nodes, protein nodes, biological process (BP) nodes, cellular component (CC) nodes, and molecular function (MF) nodes;

[0019] The compound node includes the compound SMILES string;

[0020] Protein nodes include protein sequences;

[0021] Biological process nodes, cellular component nodes, and molecular function nodes include text descriptions;

[0022] Different types of nodes are connected by edges. The edges between compound nodes and protein nodes represent the active relationship between compounds and proteins. The edges between protein nodes represent the interaction relationship between proteins. The edges between protein nodes and biological process nodes represent the participation or regulatory relationship of proteins in biological processes. The edges between protein nodes and cell component nodes represent the localization or action relationship of proteins in specific cell components. The edges between protein nodes and molecular function nodes represent the relationship between proteins performing specific molecular functions.

[0023] Through the graph structure, different types of nodes are connected through edges to form a complete multimodal knowledge graph. Deep learning methods are then used to explore hidden relationships, predict possible interactions between chemicals and proteins, and reveal the potential associations between certain proteins and specific biological processes, cellular components or molecular functions.

[0024] Specifically, in step (2), when a convolutional neural network is used to learn a cross-modal knowledge graph, the cross-modal knowledge graph can be learned by constructing a cross-modal conversion model based on a convolutional neural network. During the learning process, the embedding vectors of entities and relationships in the cross-modal knowledge graph are initialized, and a projection layer is designed for each node and relationship. The embedding vectors of different feature dimensions are mapped to a unified latent space through the projection layer. The convolutional layer, batch normalization and activation function in the convolutional neural network are used to extract the deep features of the head entity in the latent space, and the feature mapping of the head entity in the tail entity space is represented to obtain a cross-modal feature representation. The paired and unpaired node relationships are optimized through the InfoNCE loss function.

[0025] Preferably, in step (3), the molecular fingerprint feature representation of the compound includes MACCS molecular fingerprint, Morgan molecular fingerprint and mol2vec descriptor.

[0026] Specifically, according to the SMILES expression of the sample compound, the RDKit software package was used to calculate the MACCS molecular fingerprint and Morgan fingerprint, and the mol2vec pre-trained model was used to calculate the mol2vec descriptor. Then, feature fusion was performed to remove missing values ​​and features with a variance of 0 and a Pearson correlation coefficient greater than 0.95 to obtain the molecular fingerprint feature representation of the compound.

[0027] In step (3), the Deep Graph Library (DGL) software package is used to convert the SMILES expression of the sample compound into a molecular graph to characterize atoms, bonds and local environment. Specifically, the molecular graph feature representation of the compound includes node features for describing the atomic information of the compound components and edge features for describing the bonding relationship between atoms; node features include atom type, hybridization type, formal charge, number of free radical electrons, atomic degree, atomic implicit valence or chirality; bond features include bond type, whether it is a conjugated bond, whether it is a ring or whether it is an aromatic bond.

[0028] Preferably, in step (4), the first deep neural network is constructed based on the KAN framework, and the second deep neural network is constructed based on the GCN framework.

[0029] Preferably, in step (5), accuracy, precision, recall, F1 score, false negative rate, area under the receiver operating characteristic curve (AUROC), or balanced accuracy (balanced_accuracy) are used as indicators to optimize the fusion model parameters.

[0030] The present invention also provides a device, comprising:

[0031] Dataset construction module: construct a carcinogenicity prediction dataset based on the collected compounds and their real molecular labels of whether they are carcinogenic or not;

[0032] Knowledge graph construction module: for the compounds in the carcinogenicity prediction dataset, a cross-modal knowledge graph containing compound-protein association data, protein-protein association data, protein-biological process data, protein-cellular component data, and protein-molecular function data is constructed;

[0033] Knowledge representation learning module: used to learn entities and relationships in cross-modal knowledge graphs using convolutional neural networks to obtain cross-modal feature representations;

[0034] Feature extraction and feature fusion module: extract molecular graph feature representation and molecular fingerprint feature representation for the compounds in the carcinogenicity prediction dataset, and fuse the molecular fingerprint feature representation with the cross-modal feature representation to obtain a fused feature representation;

[0035] Fusion model construction module: construct a fusion model combining the first deep neural network, the second deep neural network and the classifier according to the fusion feature representation and the molecular graph feature representation;

[0036] Model training module, used to train and optimize the fusion model;

[0037] The evaluation and identification module is used to evaluate the test compound using the optimized fusion model and predict the carcinogenic risk of the compound. If the prediction result is positive, the potential carcinogenic mechanism of the compound is explained.

[0038] The present invention also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the deep learning method for evaluating the carcinogenic risk of a compound based on multimodal data fusion.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention adopts a cross-modal conversion model based on a convolutional neural network to learn the cross-modal knowledge graph to capture key carcinogenic feature information. The cross-modal conversion model has learned the carcinogenic mechanism of 1548 compounds (with complete relevant information) and can obtain the cross-modal feature representation of 3672 compounds. The accuracy of the cross-modal conversion model based on a convolutional neural network on the external validation set reached 0.8579, and the PRC reached 0.8908, which is the best overall. This shows that the cross-modal conversion model has a strong generalization ability, can accurately capture the carcinogenic feature information of the compound, and performs well on unbalanced data sets.

[0041] (2) The method of the present invention fuses multimodal data, constructs a cross-modal knowledge graph, and uses a deep learning method to evaluate the carcinogenic risk of compounds. It can not only evaluate the potential carcinogenicity of compounds, but also explain the mechanism of action of compounds with positive predicted carcinogenic risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flowchart of the deep learning method for assessing the carcinogenic risk of compounds based on multimodal data fusion.

[0043] Figure 2 This is a diagram showing the effect of using convolutional neural networks to learn cross-modal knowledge graphs.

[0044] Figure 3 This is a diagram showing the effect of using the optimized fusion model to predict the carcinogenicity of compounds. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the examples. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0046] Example 1

[0047] Specifically, the flowchart of the deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion is as follows: Figure 1 As shown;

[0048] (1) Establish a carcinogenicity prediction dataset;

[0049] Based on the lists of the International Agency for Research on Cancer (IARC) and others, we obtained real molecular labels of various compounds and their carcinogenicity. We cleaned the collected compounds to remove inorganic substances, organometallic compounds, salt compounds, mixtures and duplicate substances. After data cleaning and unified classification standards, we obtained a total of 3762 sample compounds, including 2306 negative compounds and 1456 positive compounds, and established a carcinogenicity prediction data set.

[0050] (2) Construct a cross-modal knowledge graph and use a convolutional neural network to learn the cross-modal knowledge graph to obtain a cross-modal feature representation;

[0051] The crawler obtains the compound BioAssay Results information on Pubchem, filters the data with taxid 9606 and activity Active, removes duplicate records, eliminates abnormal data, corrects erroneous data, and obtains compound-protein association data (compound and active protein pair data set);

[0052] We selected the node types of protein, BP, CC and MF from PrimeKG, and recorded the features of different nodes to obtain text descriptions of protein sequences and biological processes, cellular components and molecular functions. We obtained protein-protein association data (protein-protein paired datasets), protein-biological process data (protein-biological process paired datasets), protein-cellular component data (protein-cellular component paired datasets), and protein-molecular function data (protein-molecular function paired datasets) for building cross-modal knowledge graphs.

[0053] Specifically, the cross-modal knowledge graph includes compound nodes, protein nodes, biological process nodes, cell component nodes and molecular function nodes; compound nodes include compound SMILES strings; protein nodes include protein sequences; biological process nodes, cell component nodes and molecular function nodes include text descriptions; different types of nodes are connected by edges, the edges between compound nodes and protein nodes represent the active relationship between compounds and proteins, the edges between protein nodes and protein nodes represent the interaction relationship between proteins, the edges between protein nodes and biological process nodes represent the participation or regulation relationship of proteins in biological processes, the edges between protein nodes and cell component nodes represent the localization or action relationship of proteins in specific cell components, and the edges between protein nodes and molecular function nodes represent the relationship between proteins performing specific molecular functions; as shown in Table 1;

[0054] Table 1 Entities, relations and quantity statistics of cross-modal knowledge graph

[0055]

[0056] The cross-modal knowledge graph is learned by constructing a cross-modal conversion model based on convolutional neural networks. During the learning process, the embedding vectors of entities and relationships in the cross-modal knowledge graph are initialized, and a projection layer is designed for each node and relationship. The embedding vectors of different feature dimensions are mapped to a unified latent space through the projection layer. The convolutional layer, batch normalization and activation function in the convolutional neural network are used to extract the deep features of the head entity in the latent space, and the feature mapping of the head entity in the tail entity space is represented to obtain a cross-modal feature representation. The paired and unpaired node relationships are optimized through the InfoNCE loss function.

[0057] Specifically, the projected head node embedding, head node type, tail node type and relationship type embedding are stacked, and the input embedding is re-arranged to adapt to the 1D convolution operation. Feature extraction is performed through 3 convolutional layers and the subsequent batch normalization layer and ReLU activation function. The convolutional layer extracts local features of different ranges with different kernel sizes (kernel_size=7, kernel_size=5 and kernel_size=3) and padding parameters. After that, the result dimension is rearranged and the first position feature is taken, which is added to the head node embedding input as the output to retain the original information and ensure convergence and stability. It also represents the features of the head node converted to the target modal space. The InfoNCE loss function is used to calculate the loss based on the tail node and negative tail node embedding. The convolutional neural network extracts and transforms node embedding features, reorganizes and maps different modal information in the feature space, and adjusts parameters in combination with InfoNCE to provide more discriminative and representative feature representations for cross-modal tasks and improve model performance.

[0058] After learning the cross-modal knowledge graph by building a cross-modal conversion model based on a convolutional neural network, cross-modal retrieval can be performed by matching the converted embedding with the candidate samples in the target modality embedding space. Cross-modal matching of chemicals and proteins, proteins and BP, MF and CC is performed on the validation set, and the recall rate is used to evaluate the matching effect (that is, the effect of learning the cross-modal knowledge graph using a convolutional neural network). Figure 2 As shown in the figure, the established cross-modal conversion model shows good results in compound target prediction and protein biological function description, with reall@5 reaching 0.503, 0.46, 0.5 and 0.78 respectively. This shows that the mechanism of action of the compound can be explained through pairwise cross-modal search conversion.

[0059] Cross-modal knowledge graphs connect chemicals and proteins, enriching the characteristic representation of chemicals. Protein and biological properties enhance protein representation learning. Convolutional neural networks are used to learn cross-modal knowledge graphs to convert protein embeddings into biomolecular functional embeddings, thereby better aligning protein sequences with functional semantics.

[0060] (3) Extracting molecular fingerprint feature representation and molecular graph feature representation, and fusing the molecular fingerprint feature representation with the cross-modal feature representation to obtain a fused feature representation;

[0061] The SMILES expression of the sample compound was exported, and the 166 substructure MACCS molecular fingerprints of the chemical based on SMART and the 2048-bit Morgan molecular fingerprint with a diameter of 2 were calculated using the RDKit software package. A 300-dimensional vector representation was constructed for the sample compound using the mol2vec pre-trained model to obtain the mol2vec descriptor, which was then fused with the RDKit fingerprint to further remove missing values ​​and features with a variance of 0. The Pearson correlation coefficient between each feature was calculated, and features greater than 0.95 were removed to prevent overfitting. Finally, 3189 columns of molecular descriptors, i.e., a feature matrix of 3762 rows and 3189 columns, were obtained to obtain the molecular fingerprint feature representation of the compound. The molecular fingerprint feature representation was fused with the cross-modal feature representation to obtain a fused feature representation.

[0062] The DGL software package with PyTorch as the port is used to encode 9 atomic features and 5 bond features, and convert the molecules into molecular graph feature representation to characterize atoms, bonds and local environment. The specific encoding rules of atoms and bonds are shown in Table 2.

[0063] Table 2 Characteristic coding of atoms and bonds

[0064]

[0065] (4) constructing a fusion model, in which the first deep neural network and the second deep neural network in the fusion model are used to process different feature representations respectively, and the classifier in the fusion model is used for prediction;

[0066] The first deep neural network is built based on the KAN framework, and the second deep neural network is built based on the GCN framework.

[0067] KAN is implemented through the KANLinear class, which combines B-splines functions and linear transformations with SiLU, an activation function that can be learned on weights, to map input features to a grid space and perform nonlinear combinations through learnable spline coefficients. The goal is to enhance the model's ability to express non-graph structured data (fingerprint data and bioactive features) through complex nonlinear mappings. The KAN class forms a deep network by stacking multiple KANLinear layers, each of which performs further nonlinear transformations on the input and dynamically updates the grid during training. The activation and entropy of the model are controlled through a regularized loss function, thereby enhancing the model's stability and generalization ability.

[0068] GCN processes the node features in the graph layer by layer through the GraphConv layer in the DGL library, and introduces nonlinearity using the ReLU activation function to gradually extract high-order features in the graph structure. To prevent overfitting, a dropout layer with a dropout rate of 0.41 is added after each convolution layer, and some node features are randomly discarded during training to enhance the generalization ability of the model. After the convolution operation, the mean_nodes function of DGL is used for global pooling operation to average all node features in the graph and generate a fixed-length feature vector to represent the features of the entire graph. The pooled features are mapped to the output space through a linear layer for predicting the category of the graph.

[0069] The outputs from GNN and KAN are added to generate a combined output for the binary classification task.

[0070] (5) Conduct supervised training on the fusion model and optimize the fusion model parameters;

[0071] The data is processed in batches during the training process and loaded in batches using DataLoader, which improves the training efficiency and effectively utilizes GPU resources. The cross entropy loss function is used in the training process to minimize the gap between the model prediction value and the actual label. The weights are continuously updated iteratively using the adam optimization algorithm to learn and train the model to obtain the optimal weights. The ReduceLROnPlateau learning rate scheduler is used. When the performance of the model on the validation set does not improve, the scheduler will automatically reduce the learning rate to better perform gradient descent. When the performance on the validation set reaches the best, the model will be saved. The early stopping mechanism is used to monitor the performance of the validation set. If the performance of the validation set does not improve within 10 epochs, the training is terminated early to prevent overfitting and make the model more stable when converging. The generalization ability of the model is evaluated using the test set. The performance of the model is evaluated using multiple indicators, including accuracy, precision, recall, f1_score or ROC-AUC.

[0072] The hyperparameters of the model adjustment are as follows: In GCN, the number of hidden layers of the graph convolution layer is 3, the number of hidden units is 128, 64 and 32 respectively, and the dropout of the graph convolution layer is 0.41; in KAN, the number of hidden layers is 2, and the number of hidden units is 64 and 32 respectively. The optimizer is Adam, the loss function is the cross entropy loss CrossEntropyLoss with class weights, the loss function weight class_weights of the imbalanced class is torch.tensor([1.0,6.9]), and the loss weight of the minority class 1 is higher. The learning rate lr is 0.001, the weight decay coefficient weight_decay is 0.0001, the batch_size is 64, and the total number of epochs of training iterations is 50. The patience parameter of the Early Stopping method is 10, and the training is stopped after 10 rounds without improvement.

[0073] The data set was randomly split into training set, validation set and test set in a ratio of 7:2:1. The method of the present invention was evaluated based on the test set. The test set contained 753 substances. After the SMILES expression of the test set compounds was derived, the fusion feature representation and the molecular graph feature representation were extracted and input into the fusion model to obtain the carcinogenicity test results. The prediction results of 635 molecular labels in the test results were consistent with the true molecular labels. The method accuracy was 0.85 and the balance accuracy was 0.83. The balance accuracy of the fusion model and other models is compared as follows Figure 3 shown.

[0074] (6) Use the optimized fusion model to evaluate the carcinogenic risk of the test compound;

[0075] Given a chemical Butyl(3-hydroxypropyl)nitrosamine(BBN) (CAS No.: 51938-13-7), the above method and the optimized fusion model were used to evaluate the test compound and predict the carcinogenic risk of the compound. The result showed that its activity was 1 (the predicted result was positive), which was consistent with the conclusion of the paper. The mechanism of action of the compound was explained as follows: the CYP family (CYP2C9, CYP2C19, CYP2D6, CYP1A2) is directly involved in the phase I oxidative metabolism of BBN, and NR1I2 may indirectly affect the metabolism of BBN by regulating the expression of CYP family enzymes to generate active intermediates.

[0076] The embodiments described above provide a detailed description of the technical solutions of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements or similar substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion, characterized in that: include: (1) Establish a carcinogenicity prediction dataset, which includes multiple compounds and their real molecular labels of whether they are carcinogenic or not; (2) Using the compound-protein association data, protein-protein association data, protein-biological process data, protein-cell component data, and protein-molecular function data obtained from the database, a cross-modal knowledge graph is constructed; a convolutional neural network is used to learn the cross-modal knowledge graph to obtain a cross-modal feature representation; (3) Extracting the molecular fingerprint feature representation of the compounds in the carcinogenicity prediction dataset, fusing the molecular fingerprint feature representation with the cross-modal feature representation to obtain a fused feature representation; extracting the molecular graph feature representation of the compounds in the carcinogenicity prediction dataset; (4) constructing a fusion model combining a first deep neural network, a second deep neural network and a classifier, wherein the first deep neural network is used to extract a fusion feature vector based on the fusion feature representation, the second deep neural network is used to extract a graph feature vector based on the molecular graph feature representation of the compound, and the classifier is used to perform prediction and classification based on the fusion feature vector and the graph feature vector to obtain a molecular label prediction result; (5) Using the carcinogenicity prediction dataset, the fusion model was supervised and trained based on the molecular label prediction results and the true molecular labels to optimize the fusion model parameters; (6) The optimized fusion model is used to evaluate the test compound and predict the carcinogenic risk of the compound. If the prediction result is positive, the potential carcinogenic mechanism of the compound is explained.

2. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (2), the cross-modal knowledge graph includes compound nodes, protein nodes, biological process nodes, cell component nodes, and molecular function nodes; The compound node includes the compound SMILES string; Protein nodes include protein sequences; Biological process nodes, cellular component nodes, and molecular function nodes include text descriptions; Different types of nodes are connected by edges. The edges between compound nodes and protein nodes represent the active relationship between compounds and proteins. The edges between protein nodes represent the interaction relationship between proteins. The edges between protein nodes and biological process nodes represent the participation or regulatory relationship of proteins in biological processes. The edges between protein nodes and cell component nodes represent the localization or action relationship of proteins in specific cell components. The edges between protein nodes and molecular function nodes represent the relationship between proteins performing specific molecular functions.

3. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (2), the embedding vectors of entities and relationships in the cross-modal knowledge graph are initialized, and a projection layer is designed for each node and relationship. The embedding vectors of different feature dimensions are mapped to a unified latent space through the projection layer. The convolutional layer, batch normalization and activation function in the convolutional neural network are used to extract the deep features of the head entity in the latent space, and the feature mapping of the head entity in the tail entity space is represented to obtain a cross-modal feature representation. The paired and unpaired node relationships are optimized using the InfoNCE loss function.

4. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (3), the molecular fingerprint feature representation of the compound includes MACCS molecular fingerprint, Morgan molecular fingerprint and mol2vec descriptor.

5. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (3), the molecular graph feature representation of the compound includes node features for describing the atomic information of the compound components and edge features for describing the bonding relationship between atoms; Node features include atom type, hybridization type, formal charge, radical electron number, atomic degree, atomic implicit valence or chirality; Bond characteristics include the type of bond, whether it is a conjugated bond, whether it forms a ring, or whether it is an aromatic bond.

6. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (4), the first deep neural network is constructed based on the KAN framework, and the second deep neural network is constructed based on the GCN framework.

7. The deep learning method for evaluating the carcinogenic risk of compounds based on multimodal data fusion according to claim 1, characterized in that: In step (5), the accuracy, precision, recall, F1 score, false negative rate, area under the receiver operating characteristic curve (AUROC), or balanced accuracy (balanced_accuracy) are used as indicators to optimize the fusion model parameters.

8. A device, characterized in that: The device comprises: Dataset construction module: construct a carcinogenicity prediction dataset based on the collected compounds and their real molecular labels of whether they are carcinogenic or not; Knowledge graph construction module: for the compounds in the carcinogenicity prediction dataset, a cross-modal knowledge graph containing compound-protein association data, protein-protein association data, protein-biological process data, protein-cellular component data, and protein-molecular function data is constructed; Knowledge representation learning module: used to learn entities and relationships in cross-modal knowledge graphs using convolutional neural networks to obtain cross-modal feature representations; Feature extraction and feature fusion module: extract molecular graph feature representation and molecular fingerprint feature representation for the compounds in the carcinogenicity prediction dataset, and fuse the molecular fingerprint feature representation with the cross-modal feature representation to obtain a fused feature representation; Fusion model construction module: construct a fusion model combining the first deep neural network, the second deep neural network and the classifier according to the fusion feature representation and the molecular graph feature representation; Model training module, used to train and optimize the fusion model; The evaluation and identification module is used to evaluate the test compound using the optimized fusion model and predict the carcinogenic risk of the compound. If the prediction result is positive, the potential carcinogenic mechanism of the compound is explained.

9. A computer-readable storage medium storing computer instructions, characterized in that: When the instruction is executed by the processor, the deep learning method for evaluating the carcinogenic risk of a compound based on multimodal data fusion as described in any one of claims 1-7 is implemented.

Images