Chemical carcinogenic dose prediction method based on multi-modal deep learning
Through multimodal deep learning methods, the classification and quantitative prediction model of chemical carcinogenicity of chemicals is constructed, which solves the insufficient prediction of existing models under multiple species, gender and exposure pathways, and achieves efficient and accurate prediction of carcinogenic doses, supporting chemical health risk management.
Patent Information
- Application Number
- CN202510520317.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-24
- Publication Date
- 2025-07-22
AI Technical Summary
The existing chemical carcinogenic dose prediction models have single modeling characteristics, no biological, exposed or tumor characteristics are integrated, and the algorithm capabilities are limited, making it difficult to achieve accurate predictions under multiple species, gender, exposure pathways and target organs.
Using multimodal deep learning method, a chemical carcinogenic classification and quantitative prediction model is constructed through relational graph convolution neural networks and graph attention pooling layers, a chemical carcinogenic classification and quantitative prediction model is constructed, and molecular graphs, biological, exposure and tumor characteristics are fused to generate embedded representations of oncogenic features to achieve carcinogenic dose prediction under multiple species, gender and exposure pathways.
It breaks through the limitations of traditional models, realizes efficient and accurate carcinogenic dose prediction for multiple tissues and organs, improves prediction accuracy and efficiency, and supports widespread application in chemical health risk assessment and management.
Smart Images
Figure CN120356553A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computational toxicology for artificial intelligence-assisted prediction of chemical toxicity, and relates to a method for predicting carcinogenic doses of chemicals based on multi-modal deep learning, which can predict the carcinogenic doses of chemicals for multiple tissues and organs in organisms.
[0002] Background
[0003] The carcinogenicity of chemicals is a core determinant in health effect assessment. The carcinogenic dose (TD 50 ) represents the dose at which 50% of the animals exposed to a certain chemical will develop tumors, and is a widely used quantitative standard for quantitatively measuring carcinogenicity. The carcinogenic process of chemicals is very complex and is interfered by multiple factors such as species, exposure route, and target organ. Carcinogenic chemicals may be exposed to organisms through different routes, and then interact in various tissues and organs to induce tumors. Therefore, it is necessary to quantitatively predict the doses that induce tumors for different exposure scenarios.
[0004] The relevant guidelines (OECD guideline 451) issued by the Organization for Economic Cooperation and Development (OECD) are carcinogenicity test guidelines based on rodents. This method is expensive, time-consuming, inefficient, and violates animal ethics, making it difficult to test the carcinogenic doses of numerous chemicals one by one. There is an urgent need to develop new technologies that are efficient and alternative to animal experiments. By constructing a quantitative structure-activity relationship (QSAR) model, the internal relationship between the molecular structure characteristics of chemicals and the induced carcinogenic effects can be revealed, and the rapid prediction of the carcinogenic doses of chemicals can be achieved, thus significantly saving time, reducing costs, and reducing the use of animal experiments.
[0005] However, traditional linear QSAR methods have obvious limitations in constructing the non-linear complex relationship between molecular structure and carcinogenic dose, and it is difficult to meet the prediction requirements, and the quantitative prediction accuracy of carcinogenic dose is poor. Currently, some studies have constructed quantitative prediction models for the carcinogenic doses of chemicals. "Molecular Diversity, 2010, 14(3): 581-594." used 805 chemicals in the Carcinogenic Potency Database (CPDB), adopted a backpropagation artificial neural network, and selected 27 physicochemical properties, topological indices, and connectivity indices as descriptors to construct a carcinogenic dose prediction model for rats. The determination coefficient R of the model on the training set 2is 0.532, the external validation coefficient is 0.212, and the generalization ability is poor. "Green Chemistry, 2022, 24(13): 5304-5319." used 33-55 kinds of fused / non-fused aromatic hydrocarbon chemicals, adopted Dragon and PaDEL descriptors, and constructed a quantitative prediction model for carcinogenic doses for male and female rats and mice respectively. The external validation coefficient is 0.626-0.926, and the model training data is limited. "Sensors (Basel), 2022, 22(21)." selected 561 chemicals from the CPDB database, used a hybrid neural network algorithm and 653 Mordred descriptors to build a model for predicting carcinogenic doses for rats. The coefficient of determination R 2 is 0.4.
[0006] In summary, although the above models all belong to the quantitative prediction models of chemical carcinogenic doses, they have the following common limitations: the modeling features are single, only molecular structure descriptors are used, and important influencing factors such as biological, exposure or tumor features are not integrated; the algorithm ability is limited, most models are based on simple neural networks and it is difficult to capture complex non-linear relationships; the experimental conditions are not fully considered, only rats are used, and the gender and exposure routes are not distinguished. Although different species and genders are involved, the data volume is small and the chemical structure diversity is limited, resulting in limited application scope of the models.
[0007] Based on the existing methods, it is still impossible to support the quantitative prediction of carcinogenic doses under multiple species, genders, exposure routes and target organs. Therefore, it is urgent to break through the modeling limitations of single features and simple algorithms, and construct a multi-modal deep learning model that couples molecular features, biological, exposure and tumor features to achieve efficient and accurate prediction of chemical carcinogenic doses. Summary of the Invention
[0008] The primary technical problem to be solved by the present invention is to provide a method for predicting multi-tissue specific carcinogenic doses of chemicals based on multi-modal deep learning.
[0009] The technical solution of the present invention:
[0010] A method for predicting carcinogenic doses of chemicals based on multi-modal deep learning, the steps are as follows:
[0011] Step S1, multi-modal data preprocessing module: Obtain the standard SMILES formula of the chemical to be tested, and encode biological features (species, gender), exposure features (exposure route), and tumor features (target organ, tumor type) into numbers;
[0012] Step S11. Obtain the standard SMILES formula of the chemical to be tested as the first modal data;
[0013] Step S12: According to the biological characteristics to be measured (species, gender) corresponding to the chemical to be measured, digitally encode the species and gender texts as the second modal data;
[0014] Step S13: According to the exposure characteristics to be measured (exposure route) corresponding to the chemical to be measured, digitally encode the exposure route text as the third modal data;
[0015] Step S14: According to the tumor characteristics to be measured (target organ, tumor) corresponding to the chemical to be measured, digitally encode the target organ and tumor texts as the fourth modal data.
[0016] Step S2, Multimodal data fusion module: Construct a multimodal coupling tensor that couples four modal features, and implement information transfer through the relational graph convolutional layer to update the state vectors of each node of the multimodal features. Further, through the attention pooling layer, fuse the cross-modal global features to obtain the carcinogenic feature embedding representation;
[0017] Step S21: Construct a molecular graph based on the standard SMILES formula of the chemical to be measured; use the Python DGL library to encode the molecular graph of the chemical to be measured to characterize the structural features of the molecule as the first modal data. It covers atomic type, structural information, bonding environment, physical and chemical properties, functional group category, conjugation, bond type, aromaticity;
[0018] Step S22: Add virtual nodes, and add biological features, exposure features, and tumor features as new nodes to the molecular graph to form a multimodal coupling tensor that couples the molecular graph, biological features, exposure features, and tumor features. The virtual nodes include: the second modal, add biological features as virtual nodes, and its feature vector is composed of the concatenation of species and gender; the third modal, add exposure features as virtual nodes, and its feature vector is composed of the concatenation of exposure route and exposure time; the fourth modal, add tumor features as virtual nodes, and its feature vector is composed of the concatenation of tumor type, target organ, and tumor quantity;
[0019] Step S23: Information transfer, input the multimodal coupling tensor into the relational graph convolutional layer, perform cross-modal node information transfer and feature update, and update the state vectors of each node in the multimodal features; finally, the feature vector of each node will be updated to the state vector that fuses the multimodal features The calculation formula is as follows:
[0020]
[0021] In the formula, is the feature vector of each target node v in the multi-modal features after l+1 iterations. The target node includes all features in the four modalities; r∈R is the set of edge types in the multi-modal features; the adjacent node u is the neighbor node connected to the target node v through an edge of relationship r; represents the set of neighbor nodes connected to the target node ν through an edge of relationship r; the σ function is a non-linear activation function, the ReLu function; is the learning weight matrix of the adjacent node u; is the learning weight matrix of the target node v;
[0022] Step S24, fuse global information, input the node state vector of the updated coupled multi-modal features into the attention pooling layer, and calculate the weight ω of each node in each multi-modal feature v , and then generate a carcinogenic feature embedding representation that fuses global information; the weight ω of each node v The calculation formula is as follows:
[0023] ω v = sigmoid(W·h v + bias)
[0024] In the formula, both W and bias are learnable parameter matrices;
[0025] The carcinogenic feature embedding representation of the chemical to be tested is calculated based on the calculated node state vectors and attention weights. The calculation formula is as follows:
[0026]
[0027] Step S3, carcinogenicity classification prediction module: Use a relational graph convolutional neural network to construct a carcinogenicity classification model, which can be used for chemical carcinogenicity classification;
[0028] Step S31, carcinogenicity classification model architecture, use the relational graph convolutional neural network algorithm and combine the graph attention pooling layer to construct the model architecture. The model input is the carcinogenic feature embedding representation after the fusion of four types of multi-modal features, and the output is the carcinogenic probability prediction value;
[0029] Step S32, carcinogenicity classification model construction and verification, sort out 16,595 carcinogenicity test data from the Carcinogenic Potential Database CPDB. Based on this dataset, train the model, use the binary cross-entropy loss function as the optimization target of the basic model, and update the hyperparameters through the backpropagation algorithm to construct a carcinogenicity classification model. After verification, the model accuracy reaches 88%.
[0030] Step S33: Application of chemical carcinogenicity classification. The 4 types of multimodal features of the chemical to be predicted are fused through Steps S1 and S2, and the resulting carcinogenic feature embedding representation is input into the chemical carcinogenicity classification model to obtain the carcinogenic dose prediction value of the chemical to be predicted. Set the carcinogenic probability threshold to 0.5. If the carcinogenic probability of the chemical is higher than the threshold, then determine that the chemical is a chemical with a high carcinogenic probability. Otherwise, determine it as a chemical with a low carcinogenic probability.
[0031] Step S4: Quantitative prediction module for carcinogenic dose of chemicals with high carcinogenic probability. Use a relational graph convolutional neural network to construct a quantitative prediction model for carcinogenic dose, which can be used for quantitative prediction of the carcinogenic dose of chemicals with a carcinogenic probability higher than 0.5;
[0032] Step S41: Carcinogenic dose model architecture. Use the relational graph convolutional neural network algorithm and combine it with a graph attention pooling layer to construct the model architecture. The input of the model is the carcinogenic feature embedding representation after fusing 4 types of multimodal features, and the output is the carcinogenic dose prediction value;
[0033] Step S42: Construction and verification of the quantitative prediction model for the carcinogenic dose of chemicals with high carcinogenic probability. Select 11,109 carcinogenic dose data of chemicals with high carcinogenic probability from 165,959 pieces of data. Based on this dataset, train the model, use the mean squared error loss function as the optimization objective of the basic model, and update the hyperparameters through the backpropagation algorithm to construct a quantitative prediction model for the carcinogenic dose of chemicals with high carcinogenic probability. After verification, the external verification coefficient of the model reaches 0.85;
[0034] Step S43: Prediction of chemical carcinogenic dose. The 4 types of multimodal features of the chemical to be predicted are fused through Steps S1 and S2, and the resulting carcinogenic feature embedding representation is input into the quantitative prediction model for the carcinogenic dose of chemicals with high carcinogenic probability to obtain the carcinogenic dose prediction value of the chemical to be predicted;
[0035] Step S5: Quantitative prediction module for carcinogenic dose of chemicals with low carcinogenic probability. Use a relational graph convolutional neural network to construct a quantitative prediction model for carcinogenic dose, which can be used for quantitative prediction of the carcinogenic dose of chemicals with a carcinogenic probability lower than 0.5;
[0036] Step S51: Carcinogenic dose model architecture. Use the relational graph convolutional neural network algorithm and combine it with a graph attention pooling layer to construct the model architecture. The input of the model is the carcinogenic feature embedding representation after fusing 4 types of multimodal features, and the output is the carcinogenic dose prediction value;
[0037] Step S52: Construction and verification of a quantitative prediction model for the carcinogenic dose of chemicals with a low carcinogenic probability. Select 4551 carcinogenic dose data of chemicals with a low carcinogenic probability from 16595 pieces of data. Based on this dataset, train the model, use the mean square error loss function as the optimization objective of the basic model, and update the hyperparameters through the backpropagation algorithm to construct a quantitative prediction model for the carcinogenic dose of chemicals with a low carcinogenic probability. After verification, the external verification coefficient of the model reaches 0.91;
[0038] Step S53: Prediction of the carcinogenic dose of chemicals. The carcinogenic feature embedding representation obtained by fusing the four types of multimodal features of the chemical to be predicted through Steps S1 and S2 is input into the quantitative prediction model for the carcinogenic dose of chemicals with a low carcinogenic probability to obtain the predicted value of the carcinogenic dose of the chemical to be predicted.
[0039] Advantageous effects of the present invention:
[0040] (1) The quantitative prediction method for the carcinogenic dose of chemicals provided by the present invention covers a total of 6 species including rats, mice, hamsters, monkeys, prosimians, and dogs, four exposure routes of intraperitoneal, intravenous, oral, and inhalation, three population gender categories of female, male, and non-gender-specific, and 39 tissue organs including adrenal gland, aorta, brain (including cerebrum, cerebellum, medulla oblongata / pons, and pituitary gland), cecum, cervix, coagulating gland, epididymis, eye, humerus (femorotibial joint), gonads (testis, ovary), heart, kidney, lacrimal gland, large intestine, liver, lymph node, esophagus, ovary, pancreas, parathyroid gland, prostate gland, stomach, rectum, testis, salivary gland, thymus, bone marrow section / fresh bone marrow aspiration, thyroid gland, seminal vesicle, tongue, skeletal muscle, trachea and lungs (soaked in formalin after inflation and fixation), skin with mammary gland area, bladder, small intestine (including intestinal-associated lymphoid tissue), uterus, spinal cord (cervical, mid-thoracic, and lumbar regions), vagina, and spleen. It breaks through the modeling limitations of traditional carcinogenicity prediction methods that only consider molecular structure features and ignore biological, exposure, and tumor features, and the application domain of the model is wider.
[0041] (2) The quantitative prediction method for the carcinogenic dose of chemicals provided by the present invention introduces multimodal feature fusion technology and graph neural network modeling algorithm, which can not only accurately classify the carcinogenicity of chemicals, but also quantitatively predict the carcinogenic dose under specific species and target organs, greatly improving the prediction accuracy and efficiency, efficiently filling the carcinogenicity classification and carcinogenic dose quantitative data of chemicals, providing a basic tool for chemical health risk assessment and management, and serving the national major needs of chemical risk control and new pollutant treatment. Description of the Drawings
[0042] Figure 1 It is a schematic flowchart of a method for predicting the carcinogenic dose of chemicals based on multimodal deep learning provided by an embodiment of the present invention. Specific Embodiment
[0043] The following further illustrates the specific embodiments of the present invention in conjunction with the accompanying drawings and technical solutions.
[0044] Example 1: Given a chemical to be predicted, β-propiolactone (CAS: 57-57-8, CID: 2365, Molecular Weight: 72.06 g / mol, Name: Propiolactone), its carcinogenicity and carcinogenic dose need to be predicted. The following steps are required:
[0045] Step S1: First, based on obtaining the standard SMILES formula C1COC1=O as the first modality, setting the organism as female mice as the second modality data, encoding the mouse in the biological characteristic species as a number, and encoding female as a number; setting the exposure as gavage as the third modality data, encoding gavage in the exposure characteristic as a number, and setting the exposure time as 104 days; setting the tissue organ and tumor type as forestomach and squamous cell carcinoma as the fourth modality data. Encoding forestomach in the tumor characteristic as a number, encoding squamous cell carcinoma as a number, and the number of tumors as 1;
[0046] Step S2: Invoke the multi-modal data preprocessing module, encode the molecular graph of this chemical using the Python DGL library, and add biological (mice and female), exposure (gavage and 104 days), and tumor characteristics (forestomach and squamous cell carcinoma) as virtual nodes to generate a multi-modal coupling tensor of the substance to be predicted. Input the multi-modal coupling tensor into the relational graph convolutional layer to update the state vectors of each node in the multi-modal features. Input the node state vectors of the updated coupled multi-modal features into the attention pooling layer, calculate the weights of each node in each multi-modal feature, and then generate a multi-modal carcinogenic feature embedding representation of β-propiolactone that fuses global information;
[0047] Step S3: Invoke the carcinogenicity classification prediction module. Calculate the carcinogenic probability of this substance to be 0.99, and the carcinogenic probability is higher than the threshold of 0.5;
[0048] Step S4: Invoke the module for quantitatively predicting the carcinogenic dose of high-carcinogenic-probability chemicals, and calculate the carcinogenic dose of this substance to be 1.42 mg / kg / d.
[0049] Currently, the International Agency for Research on Cancer (IARC) classifies it as a Group 2B carcinogen, and the CPDB database reports that the experimental value of the carcinogenic dose of this substance in this exposure scenario is 1.33 mg / kg / d. In summary, the classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 0.09 mg / kg / d.
[0050] Example 2. Given a chemical to be predicted, isomaltose (CAS: 68424-04-4, CID: 71306906, Name: Isomaltose, Molecular Weight: 342.3 g / mol), its carcinogenicity and carcinogenic dose need to be predicted. The following steps are required:
[0051] Step S1. First, based on the obtained standard SMILES formula C([C@@H]1[C@H]([C@@H]([C@H](C(O1)OC[C@@H]2[C@H]([C@@H]([C@H]([C@H](O2)O)O)O)O)O)O)O)O)O as the first modality, set the organism as female mice as the second modality data, encode the mouse in the biological characteristic species as a number, and encode female as a number; set the exposure as oral exposure via diet as the third modality data, encode oral exposure via diet in the exposure characteristics as a number, and set the exposure time as 104 days; set the tissue organ and tumor type as liver and hepatocellular carcinoma as the fourth modality data. Encode the liver as a number, encode hepatocellular carcinoma as a number, and the number of tumors is 1;
[0052] Step S2. Call the multi-modal data preprocessing module, use the Python DGL library to encode the molecular graph of this chemical, and add biological (mice and female), exposure (diet and 104 days), and tumor characteristics (liver and hepatocellular carcinoma) as virtual nodes to generate the multi-modal coupling tensor of the substance to be predicted. Input the multi-modal coupling tensor into the Relational Graph Convolution (RGCN) layer to update the state vectors of each node in the multi-modal features. Input the node state vectors of the updated coupled multi-modal features into the attention pooling layer to calculate the weight ω of each node in each multi-modal feature v , and then generate the multi-modal carcinogenic feature embedding representation of isomaltose that fuses global information;
[0053] Step S3. Call the carcinogenicity classification prediction module. Calculate the carcinogenic probability of this substance to be 0.32, and the carcinogenic probability is lower than the threshold of 0.5;
[0054] Step S5. Call the carcinogenic dose quantitative prediction module for low carcinogenic probability chemicals, and calculate the carcinogenic dose of this substance to be 1×10 31 mg / kg / d.
[0055] Currently, the CPDB database reports that the carcinogenic probability of this substance in this exposure scenario is 0, and the experimental value of the carcinogenic dose is 1×10 31 mg / kg / d. In summary, the classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 0.09 mg / kg / d.
[0056] Example 3: Given a chemical to be predicted, 7,12-dimethylbenz[a]anthracene (CAS: 57-97-6, CID: 6001, Molecular Weight: 256.3 g / mol, Name: 7,12-Dimethylbenz[a]anthracene, DMBA), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CC1=C2C=CC3=CC=CC=C3C2=C(C4=CC=CC=C14)C as the first modality, set the organism as male mice as the second modality data, set the exposure as oral exposure via diet and the exposure time as 104 days as the third modality data, and set the tissue organ and tumor type as forestomach and malignant hemangioendothelioma as the fourth modality data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate the carcinogenic feature embedding representation (Embedding), call the carcinogenicity classification prediction module, calculate the carcinogenic probability of this substance to be 0.90, call the carcinogenic dose quantitative prediction module for high carcinogenic probability chemicals, and calculate the carcinogenic dose of this substance to be 0.19 mg / kg / d. Currently, the International Agency for Research on Cancer (IARC) classifies it as a Group 1 carcinogen, and the CPDB database reports the experimental value of the carcinogenic dose of this substance in this exposure scenario to be 0.08 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 0.11 mg / kg / d.
[0057] Example 4: Given a chemical to be predicted, N-nitrosomethylundecylamine (CAS: 68107-26-6, CID: 129276, Molecular Weight: 214.35 g / mol, Name: N-Nitrosomethylundecylamine), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CCCCCCCCCC CCN(C)N=O as the first modality, set the organism as male rats as the second modality data, set the exposure as gavage exposure and the exposure time as 104 days as the third modality data, and set the tissue organ and tumor type as lung and alveolar cell tumor as the fourth modality data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate the carcinogenic feature embedding representation (Embedding), call the carcinogenicity classification prediction module, calculate the carcinogenic probability of this substance to be 0.83, call the carcinogenic dose quantitative prediction module for high carcinogenic probability chemicals, and calculate the carcinogenic dose of this substance to be 8.37 mg / kg / d. It is listed as a carcinogen in the California Safe Cosmetics Program's Reportable Ingredients List, and the CPDB database reports the experimental value of the carcinogenic dose of this substance in this exposure scenario to be 2.35 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 5.02 mg / kg / d.
[0058] Example 5: Given a chemical to be predicted, N-Nitrosomethylundecylamine (CAS: 68107-26-6, CID: 129276, Molecular Weight: 214.35 g / mol, Name: N-Nitrosomethylundecylamine), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CCCCCCCCCCCCN(C)N=O as the first modality, set the organism as male rats as the second modality data, set the exposure as gavage exposure and the exposure time as 104 days as the third modality data, and set the tissue organ and tumor type as liver and hepatocellular adenoma as the fourth modality data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate a carcinogenic feature embedding representation (Embedding). Call the carcinogenicity classification prediction module to calculate the carcinogenic probability of this substance as 0.97. Call the carcinogenic dose quantitative prediction module for highly carcinogenic probability chemicals to calculate the carcinogenic dose of this substance as 3.29 mg / kg / d. It is listed as a carcinogen in the California Safe Cosmetics Program's Reportable Ingredients Report. The CP DB database reports that the experimental value of the carcinogenic dose of this substance under this exposure scenario is 2.40 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 0.89 mg / kg / d.
[0059] Example 6: Given a chemical to be predicted, N-Nitrosomethylundecylamine (CAS: 68107-26-6, CID: 129276, Molecular Weight: 214.35 g / mol, Name: N-Nitrosomethylundecylamine), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CCCCCCCCCCCCN(C)N=O as the first modality, set the organism as male rats as the second modality data, set the exposure as gavage exposure and the exposure time as 104 days as the third modality data, and set the tissue organ and tumor type as esophagus and squamous cell carcinoma as the fourth modality data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate a carcinogenic feature embedding representation (Embedding). Call the carcinogenicity classification prediction module to calculate the carcinogenic probability of this substance as 0.98. Call the carcinogenic dose quantitative prediction module for highly carcinogenic probability chemicals to calculate the carcinogenic dose of this substance as 21.85 mg / kg / d. It is listed as a carcinogen in the California Safe Cosmetics Program's Reportable Ingredients Report. The CP DB database reports that the experimental value of the carcinogenic dose of this substance under this exposure scenario is 15.40 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 6.46 mg / kg / d.
[0060] Example 7: Given a chemical to be predicted, N-Nitrosomethylundecylamine (CAS: 68107-26-6, CID: 129276, Molecular Weight: 214.35 g / mol, Name: N-Nitrosomethylundecylamine), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CCCCCCCCCCCCN(C)N=O as the first modal data, set the organism as male rats as the second modal data, set the exposure as gavage exposure and the exposure time as 104 days as the third modal data, and set the tissue organ and tumor type as forestomach and papilloma as the fourth modal data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate the carcinogenic feature embedding representation (Embedding). Call the carcinogenicity classification prediction module, and calculate the carcinogenic probability of this substance to be 0.99. Call the carcinogenic dose quantitative prediction module for high-carcinogenic-probability chemicals, and calculate the carcinogenic dose of this substance to be 15.06 mg / kg / d. It is listed as a carcinogen in the California Safe Cosmetics Program's Reportable Ingredients List. The CP DB database reports that the experimental value of the carcinogenic dose of this substance in this exposure scenario is 15.40 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 0.34 mg / kg / d.
[0061] Example 8: Given a chemical to be predicted, 7,12-Dimethylbenz[a]anthracene (CAS: 57-97-6, CID: 6001, Molecular Weight: 256.3 g / mol, Name: 7,12-Dimethylbenz[a]anthracene, DMBA), its carcinogenicity and carcinogenic dose are to be predicted. First, set the SMILES formula CC1=C2C=CC3=CC=CC=C3C2=C(C4=CC=CC=C14)C as the first modal data, set the organism as male mice as the second modal data, set the exposure as oral exposure via diet and the exposure time as 104 days as the third modal data, and set the tissue organ and tumor type as forestomach and malignant hemangioendothelioma as the fourth modal data. Call the multi-modal data preprocessing module and the multi-modal data fusion module to generate the carcinogenic feature embedding representation (Embedding). Call the carcinogenicity classification prediction module, and calculate the carcinogenic probability of this substance to be 0.98. Call the carcinogenic dose quantitative prediction module for high-carcinogenic-probability chemicals, and calculate the carcinogenic dose of this substance to be 1.64 mg / kg / d. Currently, the International Agency for Research on Cancer (IARC) has listed it as a Group 1 carcinogen. The CPDB database reports that the experimental value of the carcinogenic dose of this substance in this exposure scenario is 0.29 mg / kg / d. The classification screening result is consistent with the experimental value, and the quantitative prediction result is relatively consistent with the experimental value, with a difference of 1.4 mg / kg / d.
Claims
1. A method for predicting the carcinogenic dose of chemicals based on multimodal deep learning, characterized in that The steps are as follows: Step S1, Multimodal Data Preprocessing Module: Obtain the standard SMILES formula of the chemical to be tested, and encode the biological characteristics, exposure characteristics, and tumor characteristics into numbers; Step S2, Multimodal Data Fusion Module: Construct a multimodal coupling tensor that couples the features of 4 modalities, and implement information transfer through the relational graph convolutional layer to update the state vectors of each node of the multimodal features. Further, fuse the cross-modal global features through the attention pooling layer to obtain the carcinogenic feature embedding representation; Step S3, Carcinogenicity Classification Prediction Module: Use a relational graph convolutional neural network to construct a carcinogenicity classification model, which can be used for the carcinogenicity classification of chemicals; Step S4, Quantitative Prediction Module for Carcinogenic Dose of Chemicals with High Carcinogenic Probability: Use a relational graph convolutional neural network to construct a quantitative prediction model for carcinogenic dose, which can be used for the quantitative prediction of the carcinogenic dose of chemicals with a carcinogenic probability higher than 0.5; Step S5, Quantitative Prediction Module for Carcinogenic Dose of Chemicals with Low Carcinogenic Probability: Use a relational graph convolutional neural network to construct a quantitative prediction model for carcinogenic dose, which can be used for the quantitative prediction of the carcinogenic dose of chemicals with a carcinogenic probability lower than 0.
5.
2. The method for predicting the carcinogenic dose of chemicals based on multimodal deep learning according to claim 1, wherein The specific implementation process of Step S1 is as follows: Step S11, Obtain the standard SMILES formula of the chemical to be tested as the first modal data; Step S12, According to the biological characteristics to be tested corresponding to the chemical to be tested, perform digital encoding on the species and gender texts as the second modal data; Step S13, According to the exposure characteristics to be tested corresponding to the chemical to be tested, perform digital encoding on the exposure route text as the third modal data; Step S14, According to the tumor characteristics to be tested corresponding to the chemical to be tested, perform digital encoding on the target organ and tumor texts as the fourth modal data.
3. The method for predicting the carcinogenic dose of chemicals based on multimodal deep learning according to claim 1, characterized in that The specific implementation process of Step S2 is as follows: Step S21, Construct a molecular graph based on the standard SMILES formula of the chemical to be tested; use the Python DGL library to encode the molecular graph of the chemical to be tested, and represent the structural characteristics of the molecule with edges and nodes as the first modal data; covering atomic type, structural information, bonding environment, physicochemical properties, functional group category, conjugation, bond type, aromaticity; Step S22, Add virtual nodes, and add biological characteristics, exposure characteristics, and tumor characteristics as new nodes to the molecular graph to form a multimodal coupling tensor that couples the molecular graph, biological characteristics, exposure characteristics, and tumor characteristics; The virtual nodes include: the second modality, add biological characteristics as virtual nodes, and its feature vector is composed of the concatenation of species and gender; the third modality, add exposure characteristics as virtual nodes, and its feature vector is composed of the concatenation of exposure route and exposure time; the fourth modality, add tumor characteristics as virtual nodes, and its feature vector is composed of the concatenation of tumor type, target organ, and tumor number; Step S23, Information transmission: Input the multi-modal coupling tensor into the relational graph convolutional layer, perform cross-modal node information transmission and feature update, and update the state vectors of each node in the multi-modal features; finally, the feature vector of each node will be updated to the state vector that fuses the multi-modal features The calculation formula is as follows: Wherein, is the feature vector of each target node v in the multi-modal features after l+1 iterations. The target node includes all features in the four modalities; r∈R is the set of edge types in the multi-modal features; the adjacent node u is the neighbor node connected to the target node v through the edge of the relationship r; represents the set of neighbor nodes connected to the target node ν through the edge of the relationship r; the σ function is a non-linear activation function, the ReLu function; is the learning weight matrix of the adjacent node u; is the learning weight matrix of the target node v; Step S24: Integrate global information. Input the node state vectors of the updated coupled multi-modal features into the attention pooling layer to calculate the weights ω of each node in each multi-modal feature v , and then generate a carcinogenic feature embedding representation that integrates global information; the weights ω of each node v are calculated as follows: ω v = sigmoid(W·h v + bias) In the formula, both W and bias are learnable parameter matrices; The carcinogenic feature embedding representation of the chemical to be tested is calculated based on the calculated state vectors and attention weights of each node, and the calculation formula is as follows:
4. The method for predicting the carcinogenic dose of chemicals based on multimodal deep learning according to claim 1, wherein The specific implementation process of Step S3 is as follows: Step S31: Carcinogenicity classification model architecture. A relational graph convolutional neural network algorithm is adopted and combined with a graph attention pooling layer to construct the model architecture. The input of the model is the carcinogenic feature embedding representation after the fusion of 4 types of multimodal features, and the output is the carcinogenic probability prediction value. Step S32: Construction and verification of the carcinogenicity classification model. 16,595 carcinogenicity test data are sorted out from the Carcinogenic Potential Database (CPDB). Based on this dataset, the model is trained. The binary cross-entropy loss function is used as the optimization objective of the basic model, and the hyperparameters are updated through the backpropagation algorithm to construct the carcinogenicity classification model. After verification, the accuracy of the model reaches 88%. Step S33: Application of chemical carcinogenicity classification. The carcinogenic feature embedding representation obtained by fusing the 4 types of multimodal features of the chemical to be predicted through Steps S1 and S2 is input into the chemical carcinogenicity classification model to obtain the carcinogenic dose prediction value of the chemical to be predicted. The carcinogenic probability threshold is set to 0.
5. If the carcinogenic probability of the chemical is higher than the threshold, the chemical is determined to be a chemical with a high carcinogenic probability; otherwise, it is determined to be a chemical with a low carcinogenic probability.
5. The method for predicting carcinogenic dose of chemicals based on multimodal deep learning according to claim 1, characterized in that The specific implementation process of Step S4 is as follows: Step S41: Carcinogenic dose model architecture. A relational graph convolutional neural network algorithm is adopted and combined with a graph attention pooling layer to construct the model architecture. The input of the model is the carcinogenic feature embedding representation after the fusion of 4 types of multimodal features, and the output is the carcinogenic dose prediction value. Step S42: Construction and verification of the quantitative prediction model for the carcinogenic dose of high-carcinogenic-probability chemicals. 11,109 carcinogenic dose data of high-carcinogenic-probability chemicals are selected from 165,959 pieces of data. Based on this dataset, the model is trained. The mean squared error loss function is used as the optimization objective of the basic model, and the hyperparameters are updated through the backpropagation algorithm to construct the quantitative prediction model for the carcinogenic dose of high-carcinogenic-probability chemicals. After verification, the external validation coefficient of the model reaches 0.
85. Step S43: Prediction of the carcinogenic dose of chemicals. The carcinogenic feature embedding representation obtained by fusing the 4 types of multimodal features of the chemical to be predicted through Steps S1 and S2 is input into the quantitative prediction model for the carcinogenic dose of high-carcinogenic-probability chemicals to obtain the carcinogenic dose prediction value of the chemical to be predicted.
6. The method for predicting the carcinogenic dose of chemicals based on multimodal deep learning according to claim 1, wherein The specific implementation process of Step S5 is as follows: Step S51: Carcinogenic dose model architecture. A relational graph convolutional neural network algorithm is adopted and combined with a graph attention pooling layer to construct the model architecture. The input of the model is the carcinogenic feature embedding representation after the fusion of 4 types of multimodal features, and the output is the carcinogenic dose prediction value. Step S52: Construction and verification of the quantitative prediction model for the carcinogenic dose of low-carcinogenic-probability chemicals. 4,551 carcinogenic dose data of low-carcinogenic-probability chemicals are selected from 16,595 pieces of data. Based on this dataset, the model is trained. The mean squared error loss function is used as the optimization objective of the basic model, and the hyperparameters are updated through the backpropagation algorithm to construct the quantitative prediction model for the carcinogenic dose of low-carcinogenic-probability chemicals. After verification, the external validation coefficient of the model reaches 0.
91. Step S53: Prediction of the carcinogenic dose of chemicals. The carcinogenic feature embedding representation obtained by fusing the four types of multimodal features of the chemical to be predicted through Steps S1 and S2 is input into the quantitative prediction model of the carcinogenic dose of chemicals with a low carcinogenic probability to obtain the predicted value of the carcinogenic dose of the chemical to be predicted.