A method and device for predicting developmental toxicity of chemicals based on multimodal fusion

Abstract

The present invention discloses a method and device for predicting the developmental toxicity of chemicals based on multimodal fusion. The method comprises the following steps: obtaining the data required for the model (SMILES of the compound); saving the SMILES of the compound into the device; the device automatically calls the preset data files, Python and related libraries in the environment to calculate the data of three modes of chemical descriptors, biological activity and batch key molecule docking, and generates thousands of features that are further called and input into the pre-trained developmental toxicity prediction model (multimodal fusion modeling); the output result of the developmental toxicity prediction model is the developmental toxicity risk prediction value and suspicious risk structure of the chemical source of the data. If the predicted value is greater than 0.5, the genes corresponding to the proteins with higher docking scores of the key molecule docking are simultaneously screened out for annotation and functional analysis, so as to achieve the purpose of segmented prediction, which is of great significance for the identification of emerging chemicals.

Description

Technical Field

[0001] The present invention relates to a method for predicting the developmental toxicity of chemicals based on multimodal fusion, and also relates to a corresponding device for predicting the developmental toxicity of chemicals based on multimodal fusion, belonging to the field of artificial intelligence-assisted prediction of chemical toxicity, and belonging to a multidisciplinary cross-study of computer science, toxicology, chemical informatics, and bioinformatics. Background Art

[0002] So far, human society has discovered and produced a large number of chemical substances. Taking Pubchem (https: / / pubchem.ncbi.nlm.nih.gov / ) as an example, the website has included 118 million compounds and 318 million substances as of 2024 / 05 / 15. However, the number of compounds with reported toxicity test results is less than 300,000, which is a drop in the bucket in the vast sea of ​​compounds. Synthetic chemicals have brought convenience to modern life and made great contributions to social development. Food additives, cosmetic raw materials, medicines, pesticides, organic intermediates, etc. are commonly used categories of chemicals in commerce and are indispensable chemicals in industry, agriculture and daily life. According to the United Nations Global Chemicals Outlook II, chemical sales are expected to continue to grow in the next few decades. However, exposure to chemical mixtures is considered one of the main threats to human health.

[0003] Most toxicity / toxicokinetic prediction models are based on the QSAR concept, which converts structural or characteristic information into a digital vector that can be recognized by a computer. This can be done using molecular descriptors and molecular fingerprints for further machine learning modeling. Databases contain experimental activity data and / or descriptor values ​​collected from academic publications, while computer programs usually generate values ​​from existing embedded models.

[0004] Based on artificial intelligence, the use of chemical databases, molecular descriptors, fingerprints and model algorithms are important factors in model development. With the development of information technology, multimodal methods have entered the field of scientific research. Multimodal deep learning has been proposed, and due to its high degree of nonlinearity, it has proven its advantages in representing multimodal data. If multimodal fusion provides a new breakthrough for computational toxicology prediction models, it can promote the development of intelligent systems in a more intelligent direction. Summary of the invention

[0005] The primary technical problem to be solved by the present invention is to provide a method for predicting the developmental toxicity of chemicals based on multimodal fusion.

[0006] Another technical problem to be solved by the present invention is to provide a device for predicting the developmental toxicity of chemicals based on multimodal fusion.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] According to a first aspect of an embodiment of the present invention, a method for predicting developmental toxicity of a chemical is provided, comprising the following steps:

[0009] Step S1, obtaining the SMILES formula of the chemical to be predicted;

[0010] Step S2, the device calls the internal file to calculate the data of three modes of the chemical, including molecular structure, biological activity and batch key molecule docking activity;

[0011] Step S3, inputting the calculated three modal data into a pre-trained developmental toxicity prediction model;

[0012] Step S4: The prediction result of the developmental toxicity prediction model is the probability value of the chemical having developmental toxicity;

[0013] Furthermore, the present invention can achieve detailed prediction of the developmental toxicity of chemicals through the following steps:

[0014] Step S5: If the predicted probability value is greater than 0.5, further call the self-written function to discretize the batch of key molecule docking data to obtain the gene corresponding to the variable with a value of 1;

[0015] Step S6: Perform enrichment analysis on the acquired gene sequence, and use the enrichment analysis result as the segmentation prediction result. Preferably, an effect diagram is output.

[0016] Preferably, the following processing is performed when calculating the required data:

[0017] Step S21, retrieve key genes related to human developmental disease phenotypes through the HPO phenotype database, use Alpha Fold to predict a series of corresponding protein structure pdbqt files, remove ligands, add hydrogen, etc., and save them in the corresponding device;

[0018] Step S22, according to the SMILES formula of the chemical to be tested, using Python and other related libraries, calculate the 3D molecular structure of the chemical (chemical structure can be characterized by a set of numerical values, which are called molecular fingerprints or descriptors. They may characterize the properties of the molecule, such as log P, molecular weight, hydrogen bond donors, acceptors, rotatable bonds, etc., which can be linked to the experimental evidence of the molecule; for each level of molecular characterization, hundreds or thousands of structural features can be calculated. There are a variety of molecular descriptors and fingerprints that encode structural, topological, geometric, electrostatic, quantum chemical, thermodynamic, fragmentation features, etc.), and abstractly extract thousands of features, which is the first modal data;

[0019] Step S23, according to the SMILES formula of the chemical to be tested, using Python and other related libraries, calculate the biological activity data of the chemical (the biological activity data used in the present invention is calculated by the Chemical Checker tool. Chemical Checker extends the principle of small molecule similarity to all levels of biology. CC divides the data into five levels, from the chemical properties of the compound to the clinical results, and the complexity increases. The biological activity data is expressed in a universal vector format, and detailed information of the vectors of 25 CC spaces is obtained), and thousands of features are extracted, which is the second modal data;

[0020] Step S24: According to the SMILES formula of the chemical to be tested, Python and other related libraries are called to calculate the 3D molecular structure of the chemical, and the ligand removal and hydrogenation treatment are performed to generate a pbdqt file. The autodock vina package is called to perform batch molecular docking to generate a docking active sequence with a total of about 1k features, which is the third modal data.

[0021] Preferably, when calculating the first modal data:

[0022] Step 221, the modal feature is saved as a 50×50 matrix, all values ​​are filled toward the center, and the edges are filled with random noise.

[0023] Preferably, when calculating the second modal data:

[0024] Step 231: The modal features are saved as a 25×128 matrix.

[0025] Preferably, when calculating the third modal data:

[0026] Step 241, the present invention pre-trains an additional LDA topic model based on text data mining of gene cards database for gene classification;

[0027] Step 242: According to the gene classification model of step 241, the data of the third mode is split into several sequences for storage.

[0028] Preferably, we process the calculated three modal data as follows:

[0029] The original feature data calculated in this study were processed with outliers, recoded, standardized, and resampled in turn.

[0030] Step S3: Normalize the data to between 0 and 1 before input to prevent errors caused by dimensional differences between different variables.

[0031] Preferably, the developmental toxicity prediction model is obtained through the following steps:

[0032] Step S41, obtaining multiple groups of chemical SMILES numbers and developmental toxicity outcomes, calculating three modal data for each group of data according to the above method, and integrating these data into a corresponding format for modeling.

[0033] Step S42: obtaining multiple groups of data with or without developmental toxicity to train a pre-designed model structure, and further performing parameter optimization to obtain an optimal developmental toxicity prediction model.

[0034] The developmental toxicity prediction model adopts a multimodal fusion method and uses a Dropout layer for regularization. The model complexity is reduced by adjusting the number of nodes in the middle layer. The last Dense layer uses a sigmoid activation function to output a probability for a binary classification problem:

[0035] After the first modality data is passed in, the structure of the convolutional neural network (CNN) is used;

[0036] After the second modality data is passed in, the structure of the convolutional neural network (CNN) is used;

[0037] After the third modality data is passed in, the structure of the attention mechanism (Transformer) is used;

[0038] The last network layer contains 1 node and uses the sigmoid activation function for the output of the binary classification problem.

[0039] Preferably, after making a prediction, the developmental toxicity prediction model uses the Shapley method to calculate the SHAP value of each structural feature and visualize it.

[0040] Preferably, the developmental toxicity prediction model has the functions of subsequent processing and segmented prediction after making the prediction:

[0041] Step S51: When the model prediction value is greater than 0.5, the data of the third modality is automatically discretized according to a set threshold value, and the corresponding gene sequence is returned.

[0042] Step S61: Perform enrichment analysis based on the gene sequence returned by S51, and make segmentation predictions on the pathways that may be affected.

[0043] According to a second aspect of an embodiment of the present invention, there is provided a device for predicting and subdividing developmental toxicity of chemicals based on multimodal fusion, comprising a processor and a memory, wherein the processor reads a computer program in the memory, and the result is displayed on a display, and is used to perform the following operations:

[0044] The data to be collected are obtained, and the data to be collected only include the SMILES formula of the chemical to be predicted.

[0045] Inputting the SMILES into the receiving interface of the whole device;

[0046] The output results of the developmental toxicity prediction model are the developmental toxicity risk probability of chemicals, as well as the risk structure feature ranking and enrichment analysis result segmentation prediction displayed on the display.

[0047] Beneficial effects of the present invention

[0048] The method and device for predicting the developmental toxicity of chemicals provided by the present invention provide broader and richer relevant features for predicting developmental toxicity prediction and detailed prediction only through the SMILES formula of the chemicals, greatly improving the prediction accuracy and efficiency. On the one hand, it can effectively assist in the identification and management of the developmental toxicity of new chemicals, and on the other hand, it can effectively save the time and energy of experimenters. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of a process for predicting and subdividing the developmental toxicity of a chemical provided in an embodiment of the present invention

[0050] Figure 2 Technical route and data flow for in vitro biological activity calculation provided by embodiments of the present invention

[0051] Figure 3 Technical route and data flow for molecular descriptor calculation provided by embodiments of the present invention

[0052] Figure 4 Technical route and data flow for batch molecular docking calculation provided by embodiments of the present invention

[0053] Figure 5 Module design and block function distribution provided for the embodiment of the present invention

[0054] Figure 6 t-SNE dimension reduction clustering visualization diagram of built-in training data provided by the embodiment of the present invention

[0055] Figure 7 Training curve and loss function curve of the built-in model provided in the embodiment of the present invention

[0056] Figure 8 The results of the external validation experiment of the built-in model provided by the embodiment of the present invention

[0057] Fig. 9 A schematic diagram of a device for predicting and subdividing the developmental toxicity of a chemical provided by an embodiment of the present invention DETAILED DESCRIPTION

[0058] The technical content of the present invention is described in detail below with reference to the accompanying drawings and specific examples.

[0059] For emerging chemicals, in order to provide toxicity prediction values ​​for identification and management as early as possible, such as Figure 1 As shown, the method for predicting and subdividing the developmental toxicity of chemicals provided in the embodiment of the present invention comprises the following steps:

[0060] Step S1, obtaining the SMILES formula of the chemical to be predicted;

[0061] Step S2: The device calls the internal file to calculate the data of the three modes of the chemical, including molecular structure, biological activity and batch key molecular docking activity (such as Figure 5 - Module 1a);

[0062] When calculating the three modal data required by the model, after the chemical SMILES is imported, the following steps are included:

[0063] Step S21, retrieve key genes related to human developmental disease phenotypes through the HPO phenotype database, use Alpha Fold to predict a series of corresponding protein structure pdbqt files, remove ligands, add hydrogen, etc., and save them in the corresponding device;

[0064] Step S22: According to the SMILES formula of the chemical to be tested, use Python and other related libraries to calculate the 3D molecular structure of the chemical, and abstractly extract thousands of features (preferably, refer to https: / / www.rdkit.org / docs / index.html for implementation). This is the first modal data (method as shown in Figure 2 shown);

[0065] Step 221, the modal feature is saved as a 50×50 matrix, all values ​​are filled toward the center, and the edges are filled with random noise.

[0066] Step S23: According to the SMILES formula of the chemical to be tested, use Python and other related libraries to calculate the biological activity data of the chemical and extract thousands of features, which is the second modal data (method as shown in Figure 3 As shown, the method and technology are derived from the literature Nat Biotechnol.2020;38(9):1087-1096);

[0067] Step 231: The modal features are saved as a 25×128 matrix.

[0068] Step 231, the present invention pre-trains an additional LDA topic model based on text data mining of gene cards database for gene classification;

[0069] Step S24: According to the SMILES formula of the chemical to be tested, call Python and other related libraries to calculate the 3D molecular structure of the chemical, and perform ligand removal and hydrogenation treatment to generate a pbdqt file. Call the autodock vina package to perform batch molecular docking and generate a docking active sequence with a total of about 1k features. This is the third modal data (method as shown in Figure 4 shown);

[0070] Step 241, the present invention pre-trains an LDA topic model based on text data mining of gene cards database for gene classification (wherein the LDA topic model works based on Bayesian statistical theory by constructing a generative model with a three-layer structure of document-topic-vocabulary; its principle includes two parts: document generation process (first, a series of topics are randomly selected for each document; then, for each word in the document, a topic is randomly selected according to the topic distribution of the current position; finally, a word is randomly generated according to the vocabulary distribution under the selected topic) and model training (the LDA model maximizes the probability of generating a document by adjusting the parameters of the topic-vocabulary distribution and the document-topic distribution. This is usually implemented using algorithms such as Gibbs Sampling or Variational Inference));

[0071] Step 242: According to the gene classification model of step 241, the data of the third mode is split into several sequences for storage.

[0072] The method of using the three modal data of chemicals calculated in step S2 to obtain multiple groups of the above three modal data of chemicals with or without developmental toxicity is used to train the developmental toxicity prediction model.

[0073] Step S3, inputting the calculated three modal data into a pre-trained developmental toxicity prediction model;

[0074] like Figure 5 - As shown in module 1b, the pre-designed developmental toxicity prediction model adopts a multimodal fusion method and uses a Dropout layer for regularization. The model complexity is reduced by adjusting the number of nodes in the middle layer. The last Dense layer uses a sigmoid activation function to output a probability for a binary classification problem, where: after the first modality data is passed in, the convolutional neural network (CNN) structure is used; after the second modality data is passed in, the convolutional neural network (CNN) structure is used; after the third modality data is passed in, the attention mechanism (Transformer) structure is used; the training data of the entire model is visualized after t-SNE dimension reduction as shown below: Figure 6As shown, CAESAR database is the training set and PG database is the external validation set; Figure 7 The accuracy and loss function of the entire project during training are shown as they change with the training batches, revealing the good convergence of this study.

[0075] Step S4: Figure 5 - As shown in module 1c, the result predicted by the developmental toxicity prediction model is the probability value of the chemical having developmental toxicity; Figure 8 The confusion matrix of the actual values ​​of the entire validation set and the predicted set is displayed, indicating that the model with good phenotype in the training set still has a good effect in the external validation set, and the overall prediction accuracy is around 0.8.

[0076] Step S5: If the predicted probability value is greater than 0.5, further call the self-written function to discretize the batch of key molecule docking data to obtain the gene corresponding to the variable with a value of 1;

[0077] Step S6: perform GO or KEGG annotation on the obtained related genes and perform enrichment analysis. The more genes enriched in a certain pathway, the more susceptible this pathway is. Sort the suspicious pathways as segmented predictions and output effect diagrams.

[0078] The collected chemical SMILES formula obtained in step S1 is input into the receiving interface of the entire device; if the output probability of this model is greater than 0.5, it means that the chemical to be tested is likely to have developmental toxicity. This model can output the possible risk structure and suspicious subdivision prediction pathways of the chemical to be tested.

[0079] In order to realize the method for predicting and subdividing the developmental toxicity of chemicals based on multimodal fusion provided by the present invention, the present invention also provides a device for predicting and subdividing the developmental toxicity of chemicals based on multimodal fusion. Fig. 9 As shown, the device for predicting and subdividing the developmental toxicity of chemicals based on multimodal fusion includes a memory 31, a processor 32 and a display 33. The processor 32 reads the computer program in the memory 31 so that the prediction device executes the prediction model, and the results are displayed on the display 33.

Claims

1. A method for predicting developmental toxicity of chemicals based on multimodal fusion, characterized in that The steps include: Step S1, obtaining the SMILES formula of the chemical to be predicted; Step S2, obtaining data of three modes of the chemical, including molecular structure, biological activity and key molecule docking activity; comprising the following steps: Step S21, retrieve key genes related to human developmental disease phenotypes through the HPO phenotype database, use AlphaFold to predict a series of corresponding protein structure pdbqt files, remove ligands, hydrogenate them, and save them for subsequent calls; Step S22: Calculate the 3D molecular structure of the chemical according to the SMILES formula of the chemical to be tested, and abstractly extract the features, which is the first modal data. When the first modal data is obtained: Step 221, the modal feature is saved as a 50×50 matrix, all values ​​are filled toward the center, and the edges are filled with random noise; Step S23: Calculate the bioactivity data of the chemical according to the SMILES formula of the chemical to be tested, and extract the features, which is the second modal data. When the second modal data is obtained: Step 231, the modal feature is saved as a 25×128 matrix; Step S24, according to the SMILES formula of the chemical to be tested, the 3D molecular structure of the chemical is calculated, and the ligand removal and hydrogenation treatment is performed to generate a pbdqt file; Call the pdbqt file of the protein macromolecule in step S21 to perform batch molecular docking together to generate a docking active sequence, which is the third modal data; Step S3, inputting the three modal data into a pre-trained developmental toxicity prediction model; Step S4: The prediction result of the developmental toxicity prediction model is the probability value of the chemical having developmental toxicity; when the third modal data is obtained: Step 241, performing gene classification based on the LDA topic model mined from the gene cards database text data; the LDA topic model is obtained by the following steps: Ⅰ Randomly select a series of topics for each gene description document; Ⅱ For each word in the document, randomly select a topic based on the topic distribution of the current position; Ⅲ Randomly generate a vocabulary according to the vocabulary distribution under the selected topic; IV. Using Gibbs sampling or variational inference algorithm, the probability of generating documents is maximized by adjusting the parameters of topic-vocabulary distribution and document-topic distribution, thereby obtaining the LDA topic model. Step 242: According to the LDA topic model of step 241, the data of the third mode is split into several sequences for storage.

2. The method according to claim 1, characterized in that: The original feature data obtained in steps S22, S23, and S24 are sequentially processed for outliers, recoded, standardized, and resampled before being input into the model.

3. The method according to claim 1, characterized in that Step S3 normalizes the data to between 0 and 1 before input to prevent errors caused by dimensional differences between different variables.

4. The method according to claim 1, characterized in that: The developmental toxicity prediction model is obtained through the following steps: Step S31, obtaining multiple groups of SMILES formulas and developmental toxicity outcomes of chemicals, and each group of data obtains data of three modes according to step S2; Step S32, dividing the multiple groups of data obtained in S31 into training sets and test sets, taking the data of the three modalities as input and the corresponding developmental toxicity outcomes as output to train the pre-designed model structure, further performing parameter optimization, and obtaining the optimal developmental toxicity prediction model.

5. The method according to claim 1 or 4, characterized in that: The model structure pre-designed for the developmental toxicity prediction model adopts a multimodal fusion structure and uses a Dropout layer for regularization. The model complexity is reduced by adjusting the number of nodes in the middle layer. The last Dense layer uses a sigmoid activation function to output a probability for a binary classification problem: After the first modality data is passed in, the structure of the convolutional neural network CNN is used; After the second modality data is passed in, the structure of the convolutional neural network CNN is used; After the third modality data is passed in, the structure of the attention mechanism Transformer is used; The last network layer contains 1 node and uses the sigmoid activation function for the output of the binary classification problem.

6. The method according to claim 1, characterized in that It also includes: Step S5: If the predicted probability value is greater than 0.5, discretize the key molecule docking data of the chemical batch to obtain the gene sequence corresponding to the variable with a value of 1; Step S6: Perform enrichment analysis on the obtained gene sequences, and use the enrichment analysis results as segmentation prediction results.

7. The method according to claim 1, characterized in that: Step S61: According to the gene sequence returned by S5, GO or KEGG annotations are performed on the obtained related genes, and enrichment analysis is performed. The more genes enriched in a certain pathway, the more susceptible this pathway is, and the pathways that may be affected are subdivided and predicted.

8. A device for predicting developmental toxicity of chemicals based on multimodal fusion, characterized in that The device comprises a processor, a memory and a display, wherein the processor reads a computer program in the memory so that the prediction device executes the prediction model according to any one of claims 1 to 7, and the result is displayed on the display.

Images