Yuxuan Zhang
Research Subjects
Yuxuan Zhang has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Jingwen Chen
Dalian University of Technology
7 shared publications - Yuwei Liu
Key Laboratory of Industrial Ecology and Environmental Engineering (Ministry of Education), Dalian Key Laboratory on Chemicals Risk Control and Pollution Prevention Technology, School of Environmental Science and Technology
4 shared publications - Guijia Bi
Dalian University of Technology
2 shared publications - Haobo Wang
Key Laboratory of Industrial Ecology and Environmental Engineering (Ministry of Education), Dalian Key Laboratory on Chemicals Risk Control and Pollution Prevention Technology, School of Environmental Science and Technology
2 shared publications - Huaijun Xie
Key Laboratory of Industrial Ecology and Environmental Engineering (Ministry of Education), Dalian Key Laboratory on Chemicals Risk Control and Pollution Prevention Technology, School of Environmental Science and Technology
2 shared publications - Wenjia Liu
Key Laboratory of Industrial Ecology and Environmental Engineering (Ministry of Education), Dalian Key Laboratory on Chemicals Risk Control and Pollution Prevention Technology, School of Environmental Science and Technology
2 shared publications
Latest Publications
- 13 August 2026
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 39 (9) , 1787-1800
Transfer Learning Empowers the Development of Better-Performing Models for In Vivo Cardiotoxicity by Learning from In Vitro hERG Binding Data
Evaluating chemical hazards, such as cardiotoxicity, is critical for sound chemical management but is challenging due to the labor-intensive and time-consuming nature of in vivo testing. In silico models can efficiently and cost-effectively screen potential cardiotoxic compounds from a large number of chemicals. However, the available small-scale in vivo toxicity data sets prevent models from achieving superior prediction accuracy. Herein, a transfer learning-based graph neural network (TL-GNN) model for predicting in vivo cardiotoxicity was developed by fine-tuning from a pretrained model on a data set regarding the human ether-a-go-go-related gene binding, a molecular initiating event that may ultimately cause cardiotoxicity. The TL-GNN model not only outperformed the GNN without TL and classic machine learning models, but also those fine-tuned from GNN models pretrained on a nonmechanistic octanol–water partition coefficient data set and Bidirectional Encoder Representations from Transformers models pretrained on a data set with over a million unlabeled compounds. A structure–activity landscape-based method was adopted to characterize the applicability domains of the TL-GNN model. The TL-GNN model with defined applicability domains was applied to screen approximately 600,000 chemicals across diverse categories, identifying over 25,000 as cardiotoxic. The robustness of the TL-GNN model suggests the potential of the modeling strategy for complex in vivo toxicity by learning from adverse outcome pathway (AOP)-related in vitro targets.
- 13 August 2026
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 39 (9) , 1787-1800
Transfer Learning Empowers the Development of Better-Performing Models for In Vivo Cardiotoxicity by Learning from In Vitro hERG Binding Data
Evaluating chemical hazards, such as cardiotoxicity, is critical for sound chemical management but is challenging due to the labor-intensive and time-consuming nature of in vivo testing. In silico models can efficiently and cost-effectively screen potential cardiotoxic compounds from a large number of chemicals. However, the available small-scale in vivo toxicity data sets prevent models from achieving superior prediction accuracy. Herein, a transfer learning-based graph neural network (TL-GNN) model for predicting in vivo cardiotoxicity was developed by fine-tuning from a pretrained model on a data set regarding the human ether-a-go-go-related gene binding, a molecular initiating event that may ultimately cause cardiotoxicity. The TL-GNN model not only outperformed the GNN without TL and classic machine learning models, but also those fine-tuned from GNN models pretrained on a nonmechanistic octanol–water partition coefficient data set and Bidirectional Encoder Representations from Transformers models pretrained on a data set with over a million unlabeled compounds. A structure–activity landscape-based method was adopted to characterize the applicability domains of the TL-GNN model. The TL-GNN model with defined applicability domains was applied to screen approximately 600,000 chemicals across diverse categories, identifying over 25,000 as cardiotoxic. The robustness of the TL-GNN model suggests the potential of the modeling strategy for complex in vivo toxicity by learning from adverse outcome pathway (AOP)-related in vitro targets.
- 29 June 2026
- Published by American Chemical Society (ACS) in Analytical Chemistry
- Vol. 98 (27) , 20340-20351
Machine Learning-Enhanced DGT Passive Sampling Coupled with Non-Targeted Analysis for High-Throughput Monitoring of Chemical Pollutants in Aquatic Systems
High-throughput monitoring of organic pollutants in aquatic environments remains a formidable challenge due to heterogeneous emission sources, irregular emission patterns, and the scarcity of analytical standards. The diffusive gradients in thin-films (DGT) technique provides time-integrated and representative sampling. High-resolution mass spectrometry-based nontargeted analysis (NTA) enables broad-spectrum chemical identification. However, the application of DGT-based NTA (DGT-NTA) is hindered by the lack of pollutant-specific parameters required for concentration calculations, including diffusion coefficients (D) and ionization efficiencies (IE). This study proposes a machine learning (ML)-assisted framework that enabled semiquantitative analysis in DGT-NTA workflows by predicting D and IE, thereby reducing the dependence on compound-specific standards in the routine quantification step. When applied to municipal wastewater effluents, the framework enabled semiquantitative estimation of 85 identified pollutants. Targeted validation demonstrated that the mean prediction error was 1.78-fold for compounds within the applicability domains of the constructed ML models for D and IE. An additional ecological risk assessment was conducted using predicted concentrations and aquatic toxicity data, thereby prioritizing 8 potentially high-risk pollutants. This framework provides a promising avenue for high-throughput monitoring of emerging pollutants in aquatic environments.
- 29 June 2026
- Published by American Chemical Society (ACS) in Analytical Chemistry
- Vol. 98 (27) , 20340-20351
Machine Learning-Enhanced DGT Passive Sampling Coupled with Non-Targeted Analysis for High-Throughput Monitoring of Chemical Pollutants in Aquatic Systems
High-throughput monitoring of organic pollutants in aquatic environments remains a formidable challenge due to heterogeneous emission sources, irregular emission patterns, and the scarcity of analytical standards. The diffusive gradients in thin-films (DGT) technique provides time-integrated and representative sampling. High-resolution mass spectrometry-based nontargeted analysis (NTA) enables broad-spectrum chemical identification. However, the application of DGT-based NTA (DGT-NTA) is hindered by the lack of pollutant-specific parameters required for concentration calculations, including diffusion coefficients (D) and ionization efficiencies (IE). This study proposes a machine learning (ML)-assisted framework that enabled semiquantitative analysis in DGT-NTA workflows by predicting D and IE, thereby reducing the dependence on compound-specific standards in the routine quantification step. When applied to municipal wastewater effluents, the framework enabled semiquantitative estimation of 85 identified pollutants. Targeted validation demonstrated that the mean prediction error was 1.78-fold for compounds within the applicability domains of the constructed ML models for D and IE. An additional ecological risk assessment was conducted using predicted concentrations and aquatic toxicity data, thereby prioritizing 8 potentially high-risk pollutants. This framework provides a promising avenue for high-throughput monitoring of emerging pollutants in aquatic environments.
- 12 June 2026
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 60 (25) , 17900-17913
Multimodal Machine Learning Models on Aqueous Direct Photodisappearance Rate Constants of Chemicals Exhibit Previously Unattainable Performance
Direct photodisappearance rate constant (kd) is a key parameter determining the persistence of chemical pollutants in surface waters. High-throughput in silico models are preferable to tedious and expensive experimental tests for obtaining kd values of multitudinous chemicals. Previous models mainly relied on molecular descriptors as predictor variables for kd prediction, neglecting environmental conditions that have great impacts on direct photodisappearance kinetics, resulting in models with limited applicability. This study collected and curated a comprehensive data set consisting of 1281 records of kd values for 304 chemicals. Multimodal machine learning models that simultaneously consider molecular structures and environmental conditions, such as light source, pH, temperature, dissolved oxygen level, and initial concentration of chemicals, were constructed accordingly. Results indicate that the multimodal models significantly improved the coefficient of determination on the validation set (from 0.369 to 0.756) over the unimodal models. An innovative applicability domain (AD) characterization based on feature–response landscape analysis was proposed to define the ADs of the models. The models, coupled with the AD characterization, serve as an effective tool for assessing the environmental photochemical persistence of chemicals, supporting sound chemical management.
- 12 June 2026
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 60 (25) , 17900-17913
Multimodal Machine Learning Models on Aqueous Direct Photodisappearance Rate Constants of Chemicals Exhibit Previously Unattainable Performance
Direct photodisappearance rate constant (kd) is a key parameter determining the persistence of chemical pollutants in surface waters. High-throughput in silico models are preferable to tedious and expensive experimental tests for obtaining kd values of multitudinous chemicals. Previous models mainly relied on molecular descriptors as predictor variables for kd prediction, neglecting environmental conditions that have great impacts on direct photodisappearance kinetics, resulting in models with limited applicability. This study collected and curated a comprehensive data set consisting of 1281 records of kd values for 304 chemicals. Multimodal machine learning models that simultaneously consider molecular structures and environmental conditions, such as light source, pH, temperature, dissolved oxygen level, and initial concentration of chemicals, were constructed accordingly. Results indicate that the multimodal models significantly improved the coefficient of determination on the validation set (from 0.369 to 0.756) over the unimodal models. An innovative applicability domain (AD) characterization based on feature–response landscape analysis was proposed to define the ADs of the models. The models, coupled with the AD characterization, serve as an effective tool for assessing the environmental photochemical persistence of chemicals, supporting sound chemical management.
- 9 March 2026
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 60 (15) , 11542-11552
Counted Fingerprint-Enhanced Graph Neural Network Models Enable Accurate Screening of hERG Blockers from Diverse Categories of Chemicals
Chemical-induced blockade of the human ether-a-go-go-related gene (hERG) K+ channels may lead to fatal cardiac arrhythmia. Given the ever-increasing number of chemicals, developing in silico models is preferable to time-consuming experimental tests for efficiently screening potential hERG blockers. Nonetheless, the existing models were primarily constructed with structural nondiversity data sets and adopted one type of molecular representation, limiting their prediction accuracy and applicability domain (AD) coverage. Herein, dual feature-based neural network (DFNN) models with fused molecular fingerprint (MF) and molecular graph (MG) features were constructed based on an enlarged hERG blockade data set for high-throughput screening of hERG blockers. The optimal DFNN model achieved average area under the receiver operating characteristic curve, sensitivity, and specificity values of 0.950, 0.844, and 0.909, respectively, outperforming single MF- or MG-based neural network models, MF-based machine learning models, and previous models. ADs of the optimal model were characterized by an advanced structure-activity landscape analysis method. The model with defined ADs was applied to screen over 500,000 industrial chemicals, drugs, and natural products, yielding over 26,000 hERG blocker identifications. The state-of-the-art performance and robustness of the DFNN model underscore the effectiveness of the feature fusion strategy in modeling processes, holding significance for other end points.
- 9 March 2026
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 60 (15) , 11542-11552
Counted Fingerprint-Enhanced Graph Neural Network Models Enable Accurate Screening of hERG Blockers from Diverse Categories of Chemicals
Chemical-induced blockade of the human ether-a-go-go-related gene (hERG) K+ channels may lead to fatal cardiac arrhythmia. Given the ever-increasing number of chemicals, developing in silico models is preferable to time-consuming experimental tests for efficiently screening potential hERG blockers. Nonetheless, the existing models were primarily constructed with structural nondiversity data sets and adopted one type of molecular representation, limiting their prediction accuracy and applicability domain (AD) coverage. Herein, dual feature-based neural network (DFNN) models with fused molecular fingerprint (MF) and molecular graph (MG) features were constructed based on an enlarged hERG blockade data set for high-throughput screening of hERG blockers. The optimal DFNN model achieved average area under the receiver operating characteristic curve, sensitivity, and specificity values of 0.950, 0.844, and 0.909, respectively, outperforming single MF- or MG-based neural network models, MF-based machine learning models, and previous models. ADs of the optimal model were characterized by an advanced structure-activity landscape analysis method. The model with defined ADs was applied to screen over 500,000 industrial chemicals, drugs, and natural products, yielding over 26,000 hERG blocker identifications. The state-of-the-art performance and robustness of the DFNN model underscore the effectiveness of the feature fusion strategy in modeling processes, holding significance for other end points.
- 12 November 2025
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 59 (46) , 24976-24986
Generative AI-Empowered Screening Strategy for Chemical Pollutants: A Case on Per- and Polyfluoroalkyl Substances
Identifying unknown chemical pollutants is essential for effective risk management. However, current analytical techniques are restricted to structures covered by existing databases, capturing only the tip of the iceberg in the vast pollutant chemical space. Generative artificial intelligence (AI) holds promise for exploring this space and revealing the structures of unknown pollutants. This study proposes a chemical language model (CLM)-assisted screening strategy and applies it to identify unknown perfluoroalkyl and polyfluoroalkyl substances (PFASs). Using CLMs, over 1.4 million PFAS structures were generated, expanding the PFAS chemical space by 21.6%. The generated structures were compiled into a suspect list for screening. This strategy achieved a Top-1 annotation accuracy of 87% on spiked samples. When applied to a fluorochemical wastewater influent sample, it successfully enabled the annotation of 88 previously overlooked PFAS features. To further demonstrate its utility, the generated suspect list was integrated with an existing computational tool for analyzing the effluent sample, leading to the annotation of 100 new PFAS features. These findings underscore the substantial potential of generative AI for elucidating the structures of unknown pollutants, providing a new approach for high-throughput pollutant annotation.
- 12 November 2025
- Published by American Chemical Society (ACS) in Environmental Science & Technology
- Vol. 59 (46) , 24976-24986
Generative AI-Empowered Screening Strategy for Chemical Pollutants: A Case on Per- and Polyfluoroalkyl Substances
Identifying unknown chemical pollutants is essential for effective risk management. However, current analytical techniques are restricted to structures covered by existing databases, capturing only the tip of the iceberg in the vast pollutant chemical space. Generative artificial intelligence (AI) holds promise for exploring this space and revealing the structures of unknown pollutants. This study proposes a chemical language model (CLM)-assisted screening strategy and applies it to identify unknown perfluoroalkyl and polyfluoroalkyl substances (PFASs). Using CLMs, over 1.4 million PFAS structures were generated, expanding the PFAS chemical space by 21.6%. The generated structures were compiled into a suspect list for screening. This strategy achieved a Top-1 annotation accuracy of 87% on spiked samples. When applied to a fluorochemical wastewater influent sample, it successfully enabled the annotation of 88 previously overlooked PFAS features. To further demonstrate its utility, the generated suspect list was integrated with an existing computational tool for analyzing the effluent sample, leading to the annotation of 100 new PFAS features. These findings underscore the substantial potential of generative AI for elucidating the structures of unknown pollutants, providing a new approach for high-throughput pollutant annotation.
- 15 August 2025
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 38 (9) , 1460-1471
Enlarged Data Sets and Innovative Applicability Domain Characterization Empower ML Models to Reliably Bridge hERG Binding Data Gaps in Diverse Chemicals
Chemicals may cause cardiotoxicity by binding to the K+ channel encoded by the human ether-à-go-go-related gene (hERG). Given the ever-increasing number of chemicals, developing in silico models to efficiently fill the hERG binding affinity data gap is more desirable than conducting time-consuming experimental tests. However, previous data sets with limited chemical space hindered the development of models with high prediction accuracy and broad applicability domains (ADs). Herein, an expanded hERG binding affinity data set containing diverse categories of chemicals was constructed and subsequently employed to develop machine learning models. ADs of the constructed models were defined by an innovative structure-activity landscape (SAL)-based AD characterization (ADSAL), which considers activity cliffs within SALs formed by molecules with similar structures but inconsistent bioactivities. The optimal model constrained by the ADSAL achieved a coefficient of determination up to 0.89 on the external-validation set, which significantly outperformed previous models. The model coupled with the ADSAL constraint was applied to predict hERG binding affinities for more than 100,000 chemicals from multiple inventories, identifying over 5,000 potential hERG blockers. The model with ADSAL can serve as an efficient and reliable tool for bridging the hERG-mediated cardiotoxicity data vacancy to support sound chemical management.
- 15 August 2025
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 38 (9) , 1460-1471
Enlarged Data Sets and Innovative Applicability Domain Characterization Empower ML Models to Reliably Bridge hERG Binding Data Gaps in Diverse Chemicals
Chemicals may cause cardiotoxicity by binding to the K+ channel encoded by the human ether-à-go-go-related gene (hERG). Given the ever-increasing number of chemicals, developing in silico models to efficiently fill the hERG binding affinity data gap is more desirable than conducting time-consuming experimental tests. However, previous data sets with limited chemical space hindered the development of models with high prediction accuracy and broad applicability domains (ADs). Herein, an expanded hERG binding affinity data set containing diverse categories of chemicals was constructed and subsequently employed to develop machine learning models. ADs of the constructed models were defined by an innovative structure-activity landscape (SAL)-based AD characterization (ADSAL), which considers activity cliffs within SALs formed by molecules with similar structures but inconsistent bioactivities. The optimal model constrained by the ADSAL achieved a coefficient of determination up to 0.89 on the external-validation set, which significantly outperformed previous models. The model coupled with the ADSAL constraint was applied to predict hERG binding affinities for more than 100,000 chemicals from multiple inventories, identifying over 5,000 potential hERG blockers. The model with ADSAL can serve as an efficient and reliable tool for bridging the hERG-mediated cardiotoxicity data vacancy to support sound chemical management.
- 27 June 2025
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 38 (7) , 1192-1202
Machine Learning Models Based on Enlarged Chemical Spaces for Screening Carcinogenic Chemicals
Machine learning (ML) models for screening carcinogenic chemicals are critical for the sound management of chemicals. Previous models were built on small-scale datasets and lacked applicability domain (AD) characterization that is necessary for regulatory applications of the models. In the current study, an enlarged dataset containing 1697 compounds (940 carcinogens and 757 non-carcinogens) was curated and employed to construct screening models based on 12 types of molecular fingerprints, four ML algorithms, and two graph neural networks. The AD of the optimal model was defined by a state-of-the-art characterization methodology (ADSAL) based on the analysis of structure-activity landscapes (SALs). Results showed that an optimal model based on the random forest algorithm with the PubChem fingerprints outperformed previous ones, with an area under the receiver operating characteristic curve of 86.2% on the validation set imposed with the ADSAL. The optimal model, coupled with the ADSAL, was employed to screen carcinogenic chemicals in the Inventory of Existing Chemical Substances of China (IECSC) and plastic additives datasets, identifying 1282 chemicals from the IECSC and 841 plastic additives as carcinogenic chemicals. The screening model coupled with ADSAL may serve as a promising tool for prioritizing chemicals of carcinogenic concern, facilitating the sound management of chemicals.
- 27 June 2025
- Published by American Chemical Society (ACS) in Chemical Research in Toxicology
- Vol. 38 (7) , 1192-1202
Machine Learning Models Based on Enlarged Chemical Spaces for Screening Carcinogenic Chemicals
Machine learning (ML) models for screening carcinogenic chemicals are critical for the sound management of chemicals. Previous models were built on small-scale datasets and lacked applicability domain (AD) characterization that is necessary for regulatory applications of the models. In the current study, an enlarged dataset containing 1697 compounds (940 carcinogens and 757 non-carcinogens) was curated and employed to construct screening models based on 12 types of molecular fingerprints, four ML algorithms, and two graph neural networks. The AD of the optimal model was defined by a state-of-the-art characterization methodology (ADSAL) based on the analysis of structure-activity landscapes (SALs). Results showed that an optimal model based on the random forest algorithm with the PubChem fingerprints outperformed previous ones, with an area under the receiver operating characteristic curve of 86.2% on the validation set imposed with the ADSAL. The optimal model, coupled with the ADSAL, was employed to screen carcinogenic chemicals in the Inventory of Existing Chemical Substances of China (IECSC) and plastic additives datasets, identifying 1282 chemicals from the IECSC and 841 plastic additives as carcinogenic chemicals. The screening model coupled with ADSAL may serve as a promising tool for prioritizing chemicals of carcinogenic concern, facilitating the sound management of chemicals.