Chemical Skin and Eye Toxicity Screening Method Based on Multi-Task Graph Neural Network Model

By constructing chemical molecular graphs using a multi-task graph neural network model and extracting molecular structure characterization using RGCN, simultaneous prediction of chemical skin corrosion and irritation, skin sensitization, and eye corrosion and irritation was achieved. This solves the problems of insufficient model accuracy and generalization ability in existing technologies, and improves the accuracy of prediction and screening efficiency.

CN123120102APending Publication Date: 2026-11-17DALIAN UNIV OF TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611378095.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to efficiently utilize information about the molecular structure of chemicals, resulting in limited accuracy and generalization ability of prediction models for chemical skin corrosion irritation, skin sensitization, and eye corrosion irritation. Furthermore, traditional models cannot fully leverage the correlation information between different toxicity endpoints.

Method used

The Multi-Task Graph Neural Network (MTL-GNN) model is adopted to construct chemical molecular graphs, extract molecular structure characteristics using Relational Graph Convolutional Network (RGCN), and combine task-specific feature generation and prediction modules to achieve simultaneous prediction of skin corrosion irritation, skin sensitization and eye corrosion irritation.

Benefits of technology

It improves the accuracy and efficiency of predicting skin and eye-related toxicities of chemicals, simplifies the high-throughput screening process, and provides an effective tool for chemical management and risk control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120102A_ABST
    Figure CN123120102A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of chemical toxicity prediction technology, and relates to a method for screening chemical skin and eye toxicity based on a multi-task graph neural network model. This invention constructs datasets for skin corrosion and irritation, skin sensitization, and eye corrosion and irritation by aggregating multi-source data. Based on this, this invention constructs an MTL-GNN model to achieve simultaneous prediction of the three endpoints within the same model framework. This allows the model to learn the differences in characteristics of each endpoint while utilizing the structural correlation information that may exist between different endpoints, thereby improving the reliability and screening efficiency of chemical skin and eye-related toxicity prediction. This method is simple and efficient, and can play a role in high-throughput screening of chemicals with skin corrosion and irritation, skin sensitization, and eye corrosion and irritation, providing a basic tool for improving chemical management and serving chemical risk control and the treatment of new pollutants.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of chemical toxicity prediction technology, and relates to a method for screening chemical skin and eye toxicity based on a multi-task graph neural network model. Background Technology

[0002] Chemicals are a vital material foundation for the operation of modern society and economic development. However, some chemicals, upon direct contact with human skin and eye tissues, may cause skin corrosion and irritation, skin sensitization, and eye corrosion and irritation. Skin corrosion and irritation refers to the ability of a chemical to cause irreversible damage / reversible inflammatory reactions after skin contact; skin sensitization refers to the ability of a chemical to induce an immune response after skin contact and trigger an allergic skin reaction upon re-exposure; and eye corrosion and irritation refers to the ability of a chemical to cause irreversible / reversible damage to the cornea, conjunctiva, or iris after contact with the ocular surface tissues. Published testing guidelines (OECD 404, OECD 429, and OECD 405) are used for testing the skin corrosion and irritation, skin sensitization, and eye corrosion and irritation toxicity of chemicals. However, these traditional animal experiments suffer from problems such as long testing cycles, high costs, low throughput, and ethical constraints, making it difficult to meet the needs of rapid screening for large numbers of chemicals.

[0003] Computational toxicology techniques integrating machine learning algorithms offer an effective approach to predicting chemical toxicity. Among these, quantitative structure-activity relationship (QSAR) models can achieve high-throughput screening without animal testing by establishing the correlation between chemical molecular structure and toxicity. Traditional QSAR models typically rely on predefined molecular fingerprints or descriptors, making it difficult to characterize atomic connections and local chemical environments within molecules. In contrast, graph neural network models use molecular graphs as input, representing atoms and chemical bonds as nodes and edges, respectively, directly learning molecular topological information and are better suited for characterizing the relationship between complex molecular structures and toxic effects. Furthermore, multi-task learning (MTL) allows sharing of underlying molecular representations within the same model and setting independent prediction outputs for different endpoints. This preserves endpoint differences while utilizing relevant task information, improving the accuracy of predicting chemical ocular irritation, skin sensitization, and ocular irritation / irritation.

[0004] Currently, existing studies have constructed predictive models for skin corrosion irritation, skin sensitization, or eye corrosion irritation. For example, the paper "Toxicol. Res., 2023, 39, 295-305" constructed a binary classification model for skin corrosion irritation based on 545 liquid chemicals, using 22 physicochemical descriptors and an extreme gradient boosting algorithm; the model accuracy ( The recall rate was 0.734. The specificity was 0.821. The area under the receiver operating characteristic curve (AUC) of the optimal model on the test set was 0.516; however, its modeling data was limited to liquid chemicals, and the model's applicability was limited. The literature "Toxicol. Lett., 2025, 408, 1-12" constructed a binary classification model based on 3316 experimental data points of eye corrosion irritation using the Molecular Access System Fingerprint (MACCS) and Random Forest (RF) algorithms; its optimal model's AUC on the test set was 0.516. The value is 0.804. The value was 0.730. The literature "Environ. Health Perspect., 2022, 130(2): 027012" collected skin sensitization data for 1000 chemicals, and used the RF algorithm, combined with Morgan fingerprint, MACCS fingerprint, and Mordred descriptor to construct a binary classification prediction model for skin sensitization; after five-fold cross-validation, the model... It is 0.700. The value of 0.660 indicates that the generalization ability of this type of single-task model based on traditional molecular fingerprints and descriptors is limited, and there is still room for improvement in prediction performance.

[0005] Existing technologies typically construct separate single-task prediction models for skin corrosion and irritation, skin sensitization, and eye corrosion and irritation, making it difficult to fully utilize the potential correlation information between different toxicity endpoints. Furthermore, existing models largely rely on predefined molecular characterizations such as physicochemical descriptors, structural fragments, or molecular fingerprints, failing to adequately utilize atomic connections and local chemical environments within molecules. Therefore, there is still room for improvement in the predictive performance and generalization ability of these models. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a chemical skin and eye toxicity screening method based on a multi-task graph neural network model. Using chemical molecular graphs as input, the method learns the chemical molecular structure characterization through a shared molecular graph encoding process and combines independent output layers for each endpoint to complete predictions. This approach improves the accuracy of predicting chemical skin corrosiveness, skin sensitization, and eye corrosiveness by utilizing the correlations between endpoints while preserving the differences in judgment between different endpoints.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A chemical skin and eye toxicity screening method based on a multi-task graph neural network model, the specific steps of which are as follows: Step 1: Construct a multi-task modeling dataset; To address three endpoints—skin corrosion and irritation, skin sensitization, and eye corrosion and irritation—data on the skin corrosion and irritation, skin sensitization, and eye corrosion and irritation of chemicals were collected and organized. The data was then standardized in structure, duplicate records were merged, labels were unified, and multi-task data were integrated to construct a multi-task modeling dataset. The specific steps are as follows: Step 1.1: Collect and organize data on the skin corrosiveness, skin sensitization, and eye corrosiveness of chemicals from publicly available data to obtain the raw dataset; Step 1.2: Organize the chemical structure information and original toxicity labels for each endpoint in the original dataset; organize the original experimental labels for skin corrosion and irritation, skin sensitization, and eye corrosion and irritation according to the toxicity endpoints; among them, the experimental labels for the skin corrosion and irritation endpoint include skin corrosion, skin irritation, and no skin corrosion and irritation; the experimental labels for the skin sensitization endpoint include no sensitization, weak sensitization, moderate sensitization, strong sensitization, and extremely strong sensitization; the experimental labels for the eye corrosion and irritation endpoint include eye corrosion, eye irritation, and no eye corrosion and irritation; retain records with clear simplified linear input specification (SMILES) structures and corresponding toxicity endpoint experimental labels; Step 1.3: Standardize the chemical SMILES structure, including removing salts, solvents, mixtures and inorganic substances; for structures containing multiple organic molecular fragments, retain the organic molecular fragment with the most heavy atoms; and perform charge neutralization, isotope information removal, explicit hydrogen removal, structural validity check and free radical molecule elimination, and output the standardized SMILES structure. Step 1.4: Using the standardized SMILES structure as a unified molecular identifier, merge duplicate records for the same chemical. For duplicate records of the same chemical with identical original experimental labels at the same toxicity endpoint, retain one record; for duplicate records of the same chemical with inconsistent original experimental labels at the same toxicity endpoint, prioritize retaining the experimental label with the higher degree of hazard, and use it as the merged toxicity endpoint label for that chemical at the corresponding toxicity endpoint. Step 1.5: Based on the hazard category definitions of skin corrosion and irritation, skin sensitization, and eye corrosion and irritation, convert the merged toxicity endpoint labels into binary toxicity labels respectively; for skin corrosion and irritation endpoints, classify skin corrosion and skin irritation labels as positive labels, and non-skin corrosion and irritation labels as negative labels; for skin sensitization endpoints, classify weak sensitization, moderate sensitization, strong sensitization, and extremely strong sensitization labels as positive labels, and non-sensitization labels as negative labels; for eye corrosion and irritation endpoints, classify eye corrosion and eye irritation labels as positive labels, and non-eye corrosion and irritation labels as negative labels; Step 1.6: Using the standardized SMILES structure as a unified molecular identifier, integrate the three binary toxicity endpoint datasets obtained in Step 1.5; merge the binary toxicity labels of the same chemical under the endpoints of skin corrosion and irritation, skin sensitization, and eye corrosion and irritation into the same record; for chemicals with only partial toxicity labels, retain their existing labels and record the missing endpoint labels as null values, and finally construct the multi-task modeling dataset.

[0008] Step 2: Construct chemical molecular diagrams; Based on the standardized SMILES structure, chemical molecules are converted into chemical molecule graphs, where atoms are nodes and chemical bonds are edges. Node features include atom type, atom adjacency, formal charge, number of free radical electrons, hybridization type, aromaticity, total number of hydrogen atoms, and chirality information. Edge features are jointly encoded by bond type, conjugation, cyclicity, bond stereochemistry, and atom pair type, ultimately obtaining chemical molecule graphs for input to the MTL-GNN model.

[0009] Step 3: MTL-GNN model construction; An MTL-GNN model was constructed, with chemical molecular diagrams as input and skin corrosion irritation, skin sensitization, and eye corrosion irritation as prediction endpoints. The MTL-GNN model includes a molecular graph encoding module, a task-specific feature generation module, and a task prediction module.

[0010] The molecular graph encoding module is used to extract molecular structural representations that can be used for all three toxicity endpoints. This module takes a chemical molecular graph as input, with atoms as nodes and chemical bonds as edges, and uses node and edge features as initial features. A relational graph convolutional network (RGCN) is used to encode the molecular graph, and the structural information of adjacent atoms is aggregated based on edge features to update the node representations, thus obtaining shared atom representations. These shared atom representations serve as the underlying molecular structural information used for all three toxicity endpoints, enabling the model to utilize potential structural relationships between different toxicity endpoints, and also serving as input to the subsequent task-specific feature generation module. The task-specific feature generation module generates features for each of the three toxicity endpoints. Using shared atom representations as input, it calculates the weights of different atoms on the corresponding toxicity endpoints through linear transformation and the Sigmoid function, and then performs a weighted summation of the shared atom representations to obtain the task-related molecular representations for the corresponding endpoints. The task prediction module sets up independent multilayer perceptron classification heads for each of the three toxicity endpoints, takes the task-related molecular characterization of each toxicity endpoint as input, and outputs the prediction results of the corresponding toxicity endpoints.

[0011] Step 4: MTL-GNN model training; The MTL-GNN model constructed in step 3 was trained using the multi-task modeling dataset from step 1 and the chemical molecular graphs obtained in step 2. The trained MTL-GNN model was then used to predict three toxicity endpoints of chemicals: skin corrosion irritation, skin sensitization, and eye corrosion irritation, thus enabling chemical skin and eye toxicity screening.

[0012] The beneficial effects of this invention are as follows: This invention constructs datasets on skin corrosiveness and irritation, skin sensitization, and eye corrosiveness and irritation by aggregating multi-source data. Based on this, an MTL-GNN model is built to simultaneously predict the three endpoints within the same model framework. This allows the model to learn the differential characteristics of each endpoint while utilizing potential structural correlations between them, thereby improving the reliability and screening efficiency of chemical skin and eye-related toxicity prediction. This method is simple and efficient, and can play a role in high-throughput screening of chemicals with skin corrosiveness and irritation, skin sensitization, and eye corrosiveness and irritation. It provides a fundamental tool for improving chemical management and serves chemical risk control and the treatment of new pollutants. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0014] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0015] The chemical skin and eye toxicity screening method based on a multi-task graph neural network model of the present invention has the following basic process: Figure 1 As shown.

[0016] After completing the data collection and preprocessing as described in step 1 above, a multi-task modeling dataset is constructed. The skin corrosion and irritation dataset contains 6105 chemicals, the skin sensitization dataset contains 2286 chemicals, and the eye corrosion and irritation dataset contains 5779 chemicals.

[0017] After completing steps 2 and 3 above to construct the chemical molecular graph and the MTL-GNN model, the model is trained. The molecular graph node feature dimension is 40, the RGCN hidden layer dimension is set to 128, and the classification head hidden layer dimension is set to 128. The RGCN layer dropout is set to 0.2, and the classification head dropout is set to 0.3.

[0018] The multi-task modeling dataset was divided into training, test, and validation sets in an 8:1:1 ratio. The training set was used for model parameter updates, the test set for model selection and early stopping detection, and the validation set for final performance evaluation. During training, all three tasks used the binary classification cross-entropy loss function. To reduce the impact of class imbalance on model training, positive sample weights were set based on the positive and negative sample distribution of each task in the training set, where the positive sample weight was the ratio of the number of negative samples to the number of positive samples in the training set for that task. The model used the Adam optimizer with a learning rate of 0.001 and a weight decay of 1 × 10⁻⁶. -5The batch size is 32, the maximum number of training rounds is 300, and the early stop patience value is 50. After each training round, the average of the three tasks is calculated on the test set. A ROC The current model is saved when the average AROC improves; if there is no improvement for 50 consecutive rounds, training is stopped and the best-performing model on the test set is loaded. To examine the stability of the model training results, the MTL-GNN model is repeatedly trained using 10 random seeds from 2020 to 2029.

[0019] Model evaluation metrics include accuracy ( Balance accuracy () ), recall rate ( ), specificity ( ), accuracy ( ), Score and area under the receiver operating characteristic curve (ROC) ).

[0020] The results obtained from 10 repeated training iterations were statistically analyzed, and the standard deviation of each evaluation index was used as the criterion. Characterize the robustness of the model.

[0021] On the training set, the performance of the skin corrosion and irritation prediction model is: , , , , , , , , , , , , , On the validation set, the performance of the skin corrosion and irritation prediction model is: , , , , , , ; On the training set, the performance of the skin sensitization prediction model is: , , , , , , , , , , , , , On the validation set, the performance of the skin sensitization prediction model is: , , , , , , ; On the training set, the predictive performance for eye corrosion and irritation was [performance value missing]. , , , , , , , , , , , , , On the validation set, the performance of the eye corrosion irritation prediction model is: , , , , , , .

[0022] Employing structure-property relationship-based morphology ( Application areas of analysis () Characterization methods, i.e. .in, Represents the weighted molecular similarity density. Indicates the weighted response difference; subscript This indicates the chemical to be predicted; by adjusting the weighted molecular similarity density threshold ( ) and weighted response difference threshold ( To characterize the model If the chemical to be predicted Not less than ,and Not higher than Then the chemical is determined to be in Inside. According to Characterization results, skin corrosion and irritation task recommendations The threshold is: .through After characterization, , Skin sensitization task recommendation The threshold is: .through After characterization, , Recommended tasks for eye corrosion and irritation The threshold is: .through After characterization, = 0.780, = 0.735.

[0023] Example 1: Given a chemical, vinyl acetate, with Chemical Abstracts Service (CAS) registration number 108-05-4 and its normalized SMILES code C=COC(C)=O, it is used to predict its skin irritation, skin sensitization, and eye irritation.

[0024] Calculate its relationship with the skin corrosive and irritating chemical molecules in the training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the skin corrosion and irritation model. Internally, the MTL-GNN constructed in this invention was used for prediction, and the skin corrosion and irritation prediction result was "positive", which is consistent with the experimental results.

[0025] Calculate its relationship with the molecules of chemicals in the skin sensitization training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the skin sensitization model. Internally, the skin sensitization prediction result obtained by using the MTL-GNN constructed in this invention was "negative," consistent with the experimental results.

[0026] Calculate its relationship with eye-corrosion and irritation-causing chemical molecules in the training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the eye corrosion and irritation model. Internally, the MTL-GNN constructed in this invention was used for prediction, and the prediction result for eye corrosion and irritation was "negative," consistent with the experimental results.

[0027] Example 2: Given a chemical 4-amino-3,6-dichloropyridine-2-carboxylic acid (CAS No.: 150114-71-9), whose normalized SMILES code is Nc1cc(Cl)nc(C(=O)O)c1Cl, its skin corrosiveness, skin sensitization, and eye corrosiveness are predicted.

[0028] Calculate its relationship with the skin corrosive and irritating chemical molecules in the training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the skin corrosion and irritation model. Internally, the MTL-GNN constructed in this invention was used for prediction, and the skin corrosion and irritation prediction result was "negative," consistent with the experimental results.

[0029] Calculate its relationship with the molecules of chemicals in the skin sensitization training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the skin sensitization model. Internally, the skin sensitization prediction result obtained by using the MTL-GNN constructed in this invention was "negative," consistent with the experimental results.

[0030] Calculate its relationship with eye-corrosion and irritation-causing chemical molecules in the training set. and Calculations yielded , ,satisfy condition( and Therefore, this chemical is located in the eye corrosion and irritation model. Internally, the MTL-GNN constructed in this invention was used for prediction, and the prediction result for eye corrosion and irritation was "positive", consistent with the experimental results.

Claims

1. A method for screening the skin and eye toxicity of chemicals based on a multi-task graph neural network model, characterized in that, The specific steps are as follows: Step 1: Construct a multi-task modeling dataset; For the three endpoints of skin corrosion and irritation, skin sensitization and eye corrosion and irritation, data on the skin corrosion and irritation, skin sensitization and eye corrosion and irritation of chemicals were collected and organized. Then, the data were standardized in structure, duplicate records were merged, labels were unified and multi-task data were integrated to construct a multi-task modeling dataset. Step 2: Construct chemical molecular maps; Based on the standardized SMILES structure, chemical molecules are converted into chemical molecule graphs, where atoms are nodes and chemical bonds are edges. Node features include atom type, atom adjacency, formal charge, number of free radical electrons, hybridization type, aromaticity, total number of hydrogen atoms, and chirality information. Edge features are jointly encoded by bond type, conjugation, cyclicity, bond stereochemistry, and atom pair type, ultimately obtaining chemical molecule graphs for input to the MTL-GNN model. Step 3: MTL-GNN model construction; An MTL-GNN model was constructed, with chemical molecular diagrams as input and skin corrosion irritation, skin sensitization, and eye corrosion irritation as prediction endpoints. The MTL-GNN model includes a molecular graph encoding module, a task-specific feature generation module, and a task prediction module. Step 4: MTL-GNN model training; The MTL-GNN model constructed in step 3 was trained using the multi-task modeling dataset in step 1 and the chemical molecular graph obtained in step 2. The trained MTL-GNN model was used to predict the three toxic endpoints of chemical skin corrosion irritation, skin sensitization and eye corrosion irritation, so as to achieve chemical skin and eye toxicity screening.

2. The chemical skin and eye toxicity screening method based on a multi-task graph neural network model according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: Collect and organize data on the skin corrosiveness, skin sensitization, and eye corrosiveness of chemicals from publicly available data to obtain the raw dataset; Step 1.2: Organize the chemical structure information and original toxicity labels for each endpoint in the original dataset; organize the original experimental labels for skin corrosion and irritation, skin sensitization, and eye corrosion and irritation according to the toxicity endpoints; among them, the experimental labels for the skin corrosion and irritation endpoint include skin corrosion, skin irritation, and no skin corrosion and irritation; the experimental labels for the skin sensitization endpoint include no sensitization, weak sensitization, moderate sensitization, strong sensitization, and extremely strong sensitization; the experimental labels for the eye corrosion and irritation endpoint include eye corrosion, eye irritation, and no eye corrosion and irritation; retain records with clear simplified molecular linear input canonical (SMILES) structures and corresponding toxicity endpoint experimental labels; Step 1.3: Standardize the chemical SMILES structure, including removing salts, solvents, mixtures and inorganic substances; for structures containing multiple organic molecular fragments, retain the organic molecular fragment with the most heavy atoms; and perform charge neutralization, isotope information removal, explicit hydrogen removal, structural validity check and free radical molecule elimination, and output the standardized SMILES structure. Step 1.4: Using the standardized SMILES structure as a unified molecular identifier, merge duplicate records of the same chemical. For duplicate records of the same chemical with consistent original experimental labels at the same toxicity endpoint, retain one of the records. For duplicate records of the same chemical with inconsistent original experimental labels at the same toxicity endpoint, prioritize retaining the experimental label with higher hazard level and use it as the merged toxicity endpoint label for the corresponding toxicity endpoint. Step 1.5: Based on the hazard category definitions of skin corrosion and irritation, skin sensitization, and eye corrosion and irritation, convert the merged toxicity endpoint labels into binary toxicity labels respectively; for skin corrosion and irritation endpoints, classify skin corrosion and skin irritation labels as positive labels, and non-skin corrosion and irritation labels as negative labels; for skin sensitization endpoints, classify weak sensitization, moderate sensitization, strong sensitization, and extremely strong sensitization labels as positive labels, and non-sensitization labels as negative labels; for eye corrosion and irritation endpoints, classify eye corrosion and eye irritation labels as positive labels, and non-eye corrosion and irritation labels as negative labels; Step 1.6: Using the standardized SMILES structure as a unified molecular identifier, integrate the three binary toxicity endpoint datasets obtained in Step 1.5; merge the binary toxicity labels of the same chemical under the endpoints of skin corrosion and irritation, skin sensitization, and eye corrosion and irritation into the same record; for chemicals with only partial toxicity labels, retain their existing labels and record the missing endpoint labels as null values, and finally construct the multi-task modeling dataset.

3. The chemical skin and eye toxicity screening method based on a multi-task graph neural network model according to claim 1, characterized in that, In step 3, The molecular graph encoding module is used to extract molecular structural representations that can be used for all three toxicity endpoints. This module takes a chemical molecular graph as input, with atoms as nodes and chemical bonds as edges, and uses node and edge features as initial features. A relational graph convolutional network (RGCN) is used to encode the molecular graph, and the structural information of adjacent atoms is aggregated based on edge features to update the node representations, thus obtaining shared atom representations. These shared atom representations serve as the underlying molecular structural information used for all three toxicity endpoints, enabling the model to utilize potential structural relationships between different toxicity endpoints, and also serving as input to the subsequent task-specific feature generation module. The task-specific feature generation module generates features for each of the three toxicity endpoints. Using shared atom representations as input, it calculates the weights of different atoms on the corresponding toxicity endpoints through linear transformation and the Sigmoid function, and then performs a weighted summation of the shared atom representations to obtain the task-related molecular representations for the corresponding endpoints. The task prediction module sets up independent multilayer perceptron classification heads for each of the three toxicity endpoints, takes the task-related molecular characterization of each toxicity endpoint as input, and outputs the prediction results of the corresponding toxicity endpoints.