Machine Learning-Assisted QSAR Models on Contaminant Reactivity Toward Four Oxidants: Combining Small Data Sets and Knowledge Transfer

Integrated Transfer Learning and Multitask Learning Strategies to Construct Graph Neural Network Models for Predicting Bioaccumulation Parameters of Chemicals

Accurate prediction of parameters related to the environmental exposure of chemicals is crucial for the sound management of chemicals. However, the lack of large data sets for training models may result in poor prediction accuracy and robustness. Herein, integrated transfer learning (TL) and multitask learning (MTL) was proposed for constructing a graph neural network (GNN) model (abbreviated as TL-MTL-GNN model) using n-octanol/water partition coefficients as a source domain. The TL-MTL-GNN model was trained to predict three bioaccumulation parameters based on enlarged data sets that cover 2496 compounds with at least one bioaccumulation parameter. Results show that the TL-MTL-GNN model outperformed single-task GNN models with and without the TL, as well as conventional machine learning models trained with molecular descriptors or fingerprints. Applicability domains were characterized by a state-of-the-art structure-activity landscape-based (abbreviated as ADSAL) methodology. The TL-MTL-GNN model coupled with the optimal ADSAL was employed to predict bioaccumulation parameters for around 60,000 chemicals, with more than 13,000 compounds identified as bioaccumulative chemicals. The high predictive accuracy and robustness of the TL-MTL-GNN model demonstrate the feasibility of integrating the TL and MTL strategy in modeling small-sized data sets. The strategy holds significant potential for addressing small data challenges in modeling environmental chemicals.

Machine learning-driven QSAR models for predicting the cytotoxicity of five common microplastics

In the field of microplastics (MPs) toxicity prediction, machine learning (ML) computer simulation techniques are showing great potential. In this study, six ML algorithms were utilized to predict the toxicity of MPs on BEAS-2B cells based on quantitative structure-activity relationship (QSAR) models. Comparing the models of different algorithms, the extreme gradient boosting model showed the best fit and prediction performance (R2tra = 0.9876, R2test = 0.9286). Additionally, Williams plot analysis showed that the six models developed were able to predict stably within their applicability domain, with few outliers. Finally, the three feature importance methods-Embedded Feature Importance (EFI), Recursive Feature Elimination (RFE), and SHapley Additive exPlanations (SHAP)-consistently identified particle size as the most critical feature affecting toxicity prediction. The proposed QSAR model can be utilized for preliminary environmental exposure assessments of MPs and to better understand the associated health risks.

Quantitative structure-activity relationship predicting toxicity of pesticides towards Daphnia magna

Global pesticide usage reaching 2.7 million metric tons annually, brings a grave threat to non-target organisms, especially aquatic organisms, resulting in serious concerns. Predicting aquatic toxicity of pesticides towards Daphnia magna is significant. In this work, random forest (RF) algorithm, together with ten Dragon molecular descriptors, was successfully utilized to develop a quantitative structure-activity/toxicity relationship (QSAR/QSTR) model for the toxicity pEC50 of 745 pesticides towards Daphnia magna. The optimal QSTR model (RF Model I) based on the RF parameters of ntree = 50, mtry = 3 and nodesize = 5, yielded R2 = 0.877, MAE = 0.570, rms = 0.739 (training set of 596 pEC50), R2 = 0.807, MAE = 0.732, rms = 0.902 (test set of 149 pEC50), and R2 = 0.863, MAE = 0.602, rms = 0.774 (total set of 745 pEC50), which are accurate and satisfactory. The optimal RF model is comparable to other published QSTR models for Daphnia magna, although the optimal RF model possessed a small descriptor subset and dealt with a large dataset of pesticide toxicity pEC50. Thus, the investigation in this work provides a reliable, applicable QSTR model for predicting the toxicity pEC50 of pesticides towards Daphnia magna.

Meta-Analysis and Machine Learning Models for Anaerobic Biodegradation Rates of Organic Contaminants in Sediments and Sludge

Anaerobic biodegradation rates (half-lives) of organic chemicals are pivotal for environmental risk assessment and remediation. Traditional experimental evaluation, constrained by prolonged, oxygen-free conditions, struggles to keep pace with emerging contaminants. Data-driven machine learning (ML) models serve as promising complements. However, reported quantitative structure-biodegradation relationships or ML models on anaerobic biodegradation are mostly based on small data sets (<100 records) and neglect experimental conditions, usually achieving compromised predictions. This work aimed to develop ML models for predicting the biodegradation half-lives of organic pollutants in anaerobic environments (i.e., sediment/soil and sludge). Focusing on important features of both chemicals and experimental conditions, we first curated two data sets, one for sediment/soil (SED) and the other for sludge (SLD), covering 978 records for 206 chemicals from the literature, and then conducted a meta-analysis. Next, we built a binary classification (half-life of 30 days as the cutoff) model with an accuracy of 81% and a regression model with R2 of 0.56 for SED based on LightGBM (80% and 0.31 for SLD based on Extra tree, respectively). The model interpretations underscored the significance of experimental conditions (e.g., temperature and inoculum dosage), as evidenced by their high feature importance, and the models were found to correctly capture the effects of chemical substructures, for example, branched structures and aromatic rings prolonged half-lives while methyl group and ortho-substitution on rings shortened half-lives. The applicability domains of the models were also defined, resulting in reasonable prediction for the half-lives of 41% (SED) or 67% (SLD) of over 4000 persistent, bioaccumulative, and toxic chemicals. Overall, this study pioneers ML models for predicting the anaerobic degradation half-lives, offering valuable support for future evaluation and implementation of chemical anaerobic biodegradation.

Development of Machine Learning Models for Ion-Selective Electrode Cation Sensor Design

Polyvinyl chloride (PVC) membrane-based ion-selective electrode (ISE) sensors are common tools for water assessments, but their development relies on time-consuming and costly experimental investigations. To address this challenge, this study combines machine learning (ML), Morgan fingerprint, and Bayesian optimization technologies with experimental results to develop high-performance PVC-based ISE cation sensors. By using 1745 data sets collected from 20 years of literature, appropriate ML models are trained to enable accurate prediction and a deep understanding of the relationship between ISE components and sensor performance (R 2 = 0.75). Rapid ionophore screening is achieved using the Morgan fingerprint based on atomic groups derived from ML model interpretation. Bayesian optimization is then applied to identify optimal combinations of ISE materials with the potential to deliver desirable ISE sensor performance. Na+, Mg2+, and Al3+ sensors fabricated from Bayesian optimization results exhibit excellent Nernst slopes with less than 8.2% deviation from the ideal value and superb detection limits at 10-7 M level based on experimental validation results. This approach can potentially transform sensor development into a more time-efficient, cost-effective, and rational design process, guided by ML-based techniques.

Machine learning for predicting halogen radical reactivity toward aqueous organic chemicals

Rapid advances in machine learning (ML) provide fast, accurate, and widely applicable methods for predicting free radical-mediated organic pollutant reactivity. In this study, the rate constants (logk) of four halogen radicals were predicted using Morgan fingerprint (MF) and Mordred descriptor (MD) in combination with a series of ML models. The findings highlighted that making accurate predictions for various datasets depended on an effective combination of descriptors and algorithms. To further alleviate the challenge of limited sample size, we introduced a data combination strategy that improved prediction accuracy and mitigated overfitting by combining different datasets. The Light Gradient Boosting Machine (LightGBM) with MF and Random Forest (RF) with MD models based on the unified dataset were finally selected as the optimal models. The SHapley Additive exPlanations revealed insights: the MF-LightGBM model successfully captured the influence of electron-withdrawing/donating groups, while autocorrelation, walk count and information content descriptors in the MD-RF model were identified as key features. Furthermore, the important contribution of pH was emphasized. The results of the applicability domain analysis further supported that the developed model can make reliable predictions for query compounds across a broader range. Finally, a practical web application for logk calculations was built.

Precisely Identifying the Sources of Magnetic Particles by Hierarchical Classification-Aided Isotopic Fingerprinting

Magnetic particles (MPs), with magnetite (Fe3O4) and maghemite (γ-Fe2O3) as the most abundant species, are ubiquitously present in the natural environment. MPs are among the most applied engineered particles and can be produced incidentally by various human activities. Identification of the sources of MPs is crucial for their risk assessment and regulation, which, however, is still an unsolved problem. Here, we report a novel approach, hierarchical classification-aided stable isotopic fingerprinting, to address this problem. We found that naturally occurring, incidental, and engineered MPs have distinct Fe and O isotopic fingerprints due to significant Fe/O isotope fractionation during their generation processes, which enables the establishment of an Fe-O isotopic library covering complex sources. Furthermore, we developed a three-level machine learning model that not only can distinguish the sources of MPs with a high precision (94.3%) but also can identify the multiple species (Fe3O4 or γ-Fe2O3) and synthetic routes of engineered MPs with a precision of 81.6%. This work represents the first reliable strategy for the precise source tracing of particles with multiple species and complex sources.

Machine learning framework for predicting cytotoxicity and identifying toxicity drivers of disinfection byproducts

Drinking water disinfection can result in the formation disinfection byproducts (DBPs, > 700 have been identified to date), many of them are reportedly cytotoxic, genotoxic, or developmentally toxic. Analyzing the toxicity levels of these contaminants experimentally is challenging, however, a predictive model could rapidly and effectively assess their toxicity. In this study, machine learning models were developed to predict DBP cytotoxicity based on their chemical information and exposure experiments. The Random Forest model achieved the best performance (coefficient of determination of 0.62 and root mean square error of 0.63) among all the algorithms screened. Also, the results of a probabilistic model demonstrated reliable model predictions. According to the model interpretation, halogen atoms are the most prominent features for DBP cytotoxicity compared to other chemical substructures. The presence of iodine and bromine is associated with increased cytotoxicity levels, while the presence of chlorine is linked to a reduction in cytotoxicity levels. Other factors including chemical substructures (CC, N, CN, and 6-member ring), cell line, and exposure duration can significantly affect the cytotoxicity of DBPs. The similarity calculation indicated that the model has a large applicability domain and can provide reliable predictions for DBPs with unknown cytotoxicity. Finally, this study showed the effectiveness of data augmentation in the scenario of data scarcity.

Understanding Rejection Mechanisms of Trace Organic Contaminants by Polyamide Membranes via Data-Knowledge Codriven Machine Learning

Data-driven machine learning (ML) provides a promising approach to understanding and predicting the rejection of trace organic contaminants (TrOCs) by polyamide (PA). However, various confounding variables, coupled with data scarcity, restrict the direct application of data-driven ML. In this study, we developed a data-knowledge codriven ML model via domain-knowledge embedding and explored its application in comprehending TrOC rejection by PA membranes. Domain-knowledge embedding enhanced both the predictive performance and the interpretability of the ML model. The contribution of key mechanisms, including size exclusion, charge effect, hydrophobic interaction, etc., that dominate the rejections of the three TrOC categories (neutral hydrophilic, neutral hydrophobic, and charged TrOCs) was quantified. Log D and molecular charge emerge as key factors contributing to the discernible variations in the rejection among the three TrOC categories. Furthermore, we quantitatively compared the TrOC rejection mechanisms between nanofiltration (NF) and reverse osmosis (RO) PA membranes. The charge effect and hydrophobic interactions possessed higher weights for NF to reject TrOCs, while the size exclusion in RO played a more important role. This study demonstrated the effectiveness of the data-knowledge codriven ML method in understanding TrOC rejection by PA membranes, providing a methodology to formulate a strategy for targeted TrOC removal.

Reassessing the Quantum Yield and Reactivity of Triplet-State Dissolved Organic Matter via Global Kinetic Modeling

Measuring the quantum yield and reactivity of triplet-state dissolved organic matter (3DOM*) is essential for assessing the impact of DOM on aquatic photochemical processes. However, current 3DOM* quantification methods require multiple fitting steps and rely on steady-state approximations under stringent application criteria, which may introduce certain inaccuracies in the estimation of DOM photoreactivity parameters. Here, we developed a global kinetic model to simulate the reaction kinetics of the hv/DOM system using four DOM types and 2,4,6-trimethylphenol as the probe for 3DOM*. Analyses of residuals and the root-mean-square error validated the exceptional precision of the new model compared to conventional methods. 3DOM* in the global kinetic model consistently displayed a lower quantum yield and higher reactivity than those in local regression models, indicating that the generation and reactivity of 3DOM* have often been overestimated and underestimated, respectively. The global kinetic model derives parameters by simultaneously fitting probe degradation kinetics under different conditions and considers the temporally increasing concentrations of the involved reactive species. It minimizes error propagation and offers insights into the interactions of different species, thereby providing advantages in accuracy, robustness, and interpretability. This study significantly advances the understanding of 3DOM* behavior and provides a valuable kinetic model for aquatic photochemistry research.

Quantitative identification of the co-exposure effects of e-waste pollutants on human oxidative stress by explainable machine learning

Global electronic waste (e-waste) generation continues to grow. The various pollutants released during precarious e-waste disposal activities can contribute to human oxidative stress. This study encompassed 129 individuals residing near e-waste dismantling sites in China, with elevated urinary concentrations of e-waste-related pollutants including heavy metals, polycyclic aromatic hydrocarbons (PAHs), organophosphorus flame retardants (OPFRs), bisphenols (BPs), and phthalate esters (PAEs). Utilizing an explainable machine learning framework, the study quantified the co-exposure effects of these pollutants, finding that approximately 23% and 18% of the variance in oxidative DNA damage and lipid peroxidation, respectively, was attributable to these substances. Heavy metals emerged as the most critical factor in inducing oxidative stress, followed by PAHs and PAEs for oxidative DNA damage, and BPs, OPFRs, and PAEs for lipid peroxidation. The interactions between different pollutant classes were found to be weak, attributable to their disparate biological pathways. In contrast, the interactions among congeneric pollutants were strong, stemming from their shared pathways and resultant synergistic or additive effects on oxidative stress. An intelligent analysis system for e-waste pollutants was also developed, which enables more efficient processing of large-scale and dynamic datasets in evolving environments. This study offered an enticing peek into the intricacies of co-exposure effect of e-waste pollutants.

AI-enhanced chemical paradigm: From molecular graphs to accurate prediction and mechanism

The development of accurate and interpretable models for predicting reaction constants of organic compounds with hydroxyl radicals is vital for advancing quantitative structure-activity relationships (QSAR) in pollutant degradation. Methods like molecular descriptors, molecular fingerprinting, and group contribution methods have limitations, as traditional machine learning struggles to capture all intramolecular information simultaneously. To address this, we established an integrated graph neural network (GNN) with approximately 12 million learnable parameters. GNN represents atoms as nodes and chemical bonds as edges, thus transforming molecules into a graph structures, effectively capturing microscopic properties while depicting atom connectivity in non-Euclidean space. Our datasets comprise 1401 pollutants to develop an integrated GNN model with Bayesian optimization, the model achieves root mean square errors of 0.165, 0.172, and 0.189 on the training, validation, and test datasets, respectively. Furthermore, we assess molecular structure similarity using molecular fingerprint to enhance the model's applicability. Afterwards, we propose a gradient weight mapping method for model explainability, uncovering the key functional groups in chemical reactions in artificial intelligence perspective, which would boost chemistry through artificial intelligence extreme arithmetic power.

Solar photo-Fenton and persulphate-based processes for landfill leachate treatment: A critical review

Landfilling is the most usual solid waste management strategy for solid residues disposal. However, it entails several drawbacks such as the generation of landfill leachate that seriously threaten human life and the environment due to their toxicity and carcinogenic character. Among various technologies, solar photo-Fenton and sulphate-based processes have proven to be suitable for the treatment of these polluted streams. This review critically summarises the last three decades of studies in this field. It is found that the solar homogeneous photo-Fenton process should be preferably used as a pre- and post-treatment of biological technologies and as a standalone treatment for young, medium, and mature leachates, respectively. Studies on heterogeneous solar photo-Fenton process are lacking so that this technology may be scaled-up for industrial applications. Sulphate radicals are attractive for removing both COD and ammonia. However, no study has been reported on solar sulphate activation for landfill leachate treatment. This review discusses the main advances and challenges on treating landfill leachate through solar AOPs, it compares solar photo-Fenton and solar persulphate-based treatments, indicates the future research directions and contributes for a better understanding of these technologies towards sustainable treatment of landfill leachate in sunny and not-so-sunny regions.

Predicting reactivity dynamics of halogen species and trace organic contaminants using machine learning models

Reactions of reactive halogen species (Cl•, Br•, and Cl2•-) with trace organic contaminants (TrOCs) have received much attention in recent years, and their k values are fundamental parameters for understanding their reaction mechanisms. However, k values are usually unknown. In this study, we developed machine learning (ML)-based quantitative structure-activity relationship (QSAR) models to predict k values. We tested five algorithms, namely, random forest, neural network, XGBoost, support vector machine (SVM), and multilinear regression, using molecular descriptors (MDs) and molecular fingerprints (MFs) as inputs. The optimal algorithms were MD-XGBoost for Cl• and Br•, and MF-SVM for Cl2•-, respectively, with R2test values of 0.876, 0.743, and 0.853. We found that electron-withdrawing/donating groups tended to interfere with the reactivity of Cl2•- more than Cl• and Br•. This explains why MFs are better inputs for predictive models of Cl2•-, whereas MDs are more suitable for Cl• and Br•. Furthermore, we interpreted the models using SHAP analysis, and the results indicated that our models accurately predicted k values both statistically and mechanistically. Our models provide useful tools for obtaining unknown k values and help researchers understand the inherent relationships between the models.

Prediction of free radical reactions toward organic pollutants with easily accessible molecular descriptors

Machine learning (ML) is becoming an efficient tool for predicting the fate of aquatic contaminants owing to the preponderance of big data. However, whether ML can "learn" the differences in reactivity among different free radicals has not yet been tested. In this work, the effectiveness of combining ML algorithms with molecular fingerprints to predict the reactivity of three free radicals was evaluated. First, a dataset containing 211 organic pollutants and their respective rate constants with the carbonate radical (CO3•-) was used to develop predictive models using both linear regression and ML methods. The use of topological atomic alignment information, in the form of the molecular access system (MACCS) and Morgan Fingerprint, and the electronic structure features (energy levels of the lowest unoccupied and highest occupied molecular orbitals, ELUMO and EHOMO, and the energy gap between ELUMO and EHOMO) gave satisfactory predictive performances (ML model with Random Forest algorithm with MACCS: RMSEtest = 0.787; linear regression model with energy levels: RMSEtest = 0.641). Additionally, the model interpretation correctly described that the key reactivity features for CO3•- were relatively close to those for SO4•- rather than those for •OH. These results suggest that combination of ML algorithms with easily accessible molecular fingerprints would be a powerful tool to accurately predict the radical reactions towards organic compounds.

Improving predictions and understanding of primary and ultimate biodegradation rates with machine learning models

This study aimed to develop machine learning based quantitative structure biodegradability relationship (QSBR) models for predicting primary and ultimate biodegradation rates of organic chemicals, which are essential parameters for environmental risk assessment. For this purpose, experimental primary and ultimate biodegradation rates of high consistency were compiled for 173 organic compounds. A significant number of descriptors were calculated with a collection of quantum/computational chemistry software and tools to achieve comprehensive representation and interpretability. Following a pre-screening process, multiple QSBR models were developed for both primary and ultimate endpoints using three algorithms: extreme gradient boosting (XGBoost), support vector machine (SVM), and multiple linear regression (MLR). Furthermore, a unified QSBR model was constructed using the knowledge transfer technique and XGBoost. Results demonstrated that all QSBR models developed in this study had good performance. Particularly, SVM models exhibited high level of goodness of fit (coefficient of determination on the training set of 0.973 for primary and 0.980 for ultimate), robustness (leave-one-out cross-validated coefficient of 0.953 for primary and 0.967 for ultimate), and external predictive ability (external explained variance of 0.947 for primary and 0.958 for ultimate). The knowledge transfer technique enhanced model performance by learning from properties of two biodegradation endpoints. Williams plots were used to visualize the application domains of the models. Through SHapley Additive exPlanations (SHAP) analysis, this study identified key features affecting biodegradation rates. Notably, MDEO-12, APC2D1_C_O, and other features contributed to primary biodegradation, while AATS0v, AATS2v, and others inhibited it. For ultimate biodegradation, features like No. of Rotatable Bonds, APC2D1_C_O, and minHBa were contributors, while C1SP3, Halogen Ratio, GGI4, and others hindered the process. Also, the study quantified the contributions of each feature in predictions for individual chemicals. This research provides valuable tools for predicting both primary and ultimate biodegradation rates while offering insights into the mechanisms.

Developing a hybrid model for predicting the reaction kinetics between chlorine and micropollutants in water

Understanding the reactivities of chlorine towards micropollutants is crucial for assessing the fate of micropollutants in water chlorination. In this study, we integrated machine learning with kinetic modeling to predict the reaction kinetics between micropollutants and chlorine in deionized water and real surface water. We first established a framework to predict the apparent second-order rate constants for micropollutants with chlorine by combining Morgan molecular fingerprints with machine learning algorithms. The framework was tuned using Bayesian optimization and showed high prediction accuracy. It was validated through experiments and used to predict the unreported apparent second-order rate constants for 103 emerging micropollutants with chlorine. The framework also improved the understanding of the structure-dependence of micropollutants' reactivity with chlorine. We incorporated the predicted apparent second-order rate constants into the Kintecus software to establish a hybrid model to profile the time-dependent changes of micropollutant concentrations by chlorination. The hybrid model was validated by experiments conducted in real surface water in the presence of natural organic matter. The hybrid model could predict how much micropollutants were degraded by chlorination with varied chlorine contact times and/or initial chlorine dosages. This study advances fundamental understanding of the reaction kinetics between chlorine and emerging micropollutants, and also offers a valuable tool to assess the fate of micropollutants during chlorination of drinking water.

Predicting Transformation Products during Aqueous Oxidation Processes: Current State and Outlook

Water quality and its impacts on human and ecosystem health presents tremendous global challenges. While oxidative water treatment can solve many of these problems related to hygiene and micropollutants, identifying and predicting transformation products from a large variety of micropollutants induced by dosed chemical oxidants and in situ formed radicals is still a major challenge. To this end, a better understanding of the formed transformation products and their potential toxicity is needed. Currently, no theoretical tools alone can predict oxidatively induced transformation products in aqueous systems. Coupling experimental and theoretical studies has advanced the understanding of reaction kinetics and mechanisms significantly. This perspective article highlights the key progress made concerning experimental and computational approaches to predict transformation products. Knowledge gaps are identified, and the research required to advance the predictive capability is discussed.

Machine Learning Models for Inverse Design of the Electrochemical Oxidation Process for Water Purification

In this study, a machine learning (ML) framework is developed toward target-oriented inverse design of the electrochemical oxidation (EO) process for water purification. The XGBoost model exhibited the best performances for prediction of reaction rate (k) based on training the data set relevant to pollutant characteristics and reaction conditions, indicated by Rext2 of 0.84 and RMSEext of 0.79. Based on 315 data points collected from the literature, the current density, pollutant concentration, and gap energy (Egap) were identified to be the most impactful parameters available for the inverse design of the EO process. In particular, adding reaction conditions as model input features allowed provision of more available information and an increase in the sample size of the data set to improve the model accuracy. The feature importance analysis was performed for revealing the data pattern and feature interpretation by using Shapley additive explanations (SHAP). The ML-based inverse design for the EO process was generalized to a random case for tailoring the optimum conditions with phenol and 2,4-dichlorophenol (2,4-DCP) serving as model pollutants. The resulting predicted k values were close to the experimental k values by experimental verification, accounting for the relative error lower than 5%. This study provides a paradigm shift from conventional trial-and-error mode to data-driven mode for advancing research and development of the EO process by a time-saving, labor-effective, and environmentally friendly target-oriented strategy, which makes electrochemical water purification more efficient, more economic, and more sustainable in the context of global carbon peaking and carbon neutrality.

Count-Based Morgan Fingerprint: A More Efficient and Interpretable Molecular Representation in Developing Machine Learning-Based Predictive Regression Models for Water Contaminants' Activities and Properties

In this study, we introduce the count-based Morgan fingerprint (C-MF) to represent chemical structures of contaminants and develop machine learning (ML)-based predictive models for their activities and properties. Compared with the binary Morgan fingerprint (B-MF), C-MF not only qualifies the presence or absence of an atom group but also quantifies its counts in a molecule. We employ six different ML algorithms (ridge regression, SVM, KNN, RF, XGBoost, and CatBoost) to develop models on 10 contaminant-related data sets based on C-MF and B-MF to compare them in terms of the model's predictive performance, interpretation, and applicability domain (AD). Our results show that C-MF outperforms B-MF in nine of 10 data sets in terms of model predictive performance. The advantage of C-MF over B-MF is dependent on the ML algorithm, and the performance enhancements are proportional to the difference in the chemical diversity of data sets calculated by B-MF and C-MF. Model interpretation results show that the C-MF-based model can elucidate the effect of atom group counts on the target and have a wider range of SHAP values. AD analysis shows that C-MF-based models have an AD similar to that of B-MF-based ones. Finally, we developed a "ContaminaNET" platform to deploy these C-MF-based models for free use.

Prediction of Photochemical Properties of Dissolved Organic Matter Using Machine Learning

Apparent quantum yields (Φ) of photochemically produced reactive intermediates (PPRIs) formed by dissolved organic matter (DOM) are vital to element cycles and contaminant fates in surface water. Simultaneous determination of ΦPPRI values from numerous water samples through existing experimental methods is time consuming and ineffective. Herein, machine learning models were developed with a systematic data set including 1329 data points to predict the values of three ΦPPRIs (Φ3DOM*, Φ1O2, and Φ·OH) based on DOM spectral parameters, experimental conditions, and calculation parameters. The best predictive performances for Φ3DOM*, Φ1O2, and Φ·OH were achieved using the CatBoost model, which outperformed the traditional linear regression models. The significances of the wavelength range and spectral parameters on the three ΦPPRI predictions were revealed, suggesting that DOM with lower molecular weight, lower aromatic content, and a more autochthonous portion possessed higher ΦPPRIs. Chain models were constructed by adding the predicted Φ3DOM* as a new feature into the Φ1O2 and Φ·OH models, which consequently improved the predictive performance of Φ1O2 but worsened the Φ·OH prediction likely due to the complex formation pathways of ·OH. Overall, this study offered robust ΦPPRI prediction across interlaboratory differences and provided new insights into the relationship between PPRIs formation and DOM properties.

Abiotic Reduction of Organic and Inorganic Compounds by Fe(II)-Associated Reductants: Comprehensive Data Sets and Machine Learning Modeling

Iron-associated reductants play a crucial role in providing electrons for various reductive transformations. However, developing reliable predictive tools for estimating abiotic reduction rate constants (logk) in such systems has been impeded by the intricate nature of these systems. Our recent study developed a machine learning (ML) model based on 60 organic compounds toward one soluble Fe(II)-reductant. In this study, we built a comprehensive kinetic data set covering the reactivity of 117 organic and 10 inorganic compounds toward four major types of Fe(II)-associated reductants. Separate ML models were developed for organic and inorganic compounds, and the feature importance analysis demonstrated the significance of resonance structures, reducible functional groups, reductant descriptors, and pH in logk prediction. Mechanistic interpretation validated that the models accurately learned the impact of various factors such as aromatic substituents, complexation, bond dissociation energy, reduction potential, LUMO energy, and dominant reductant species. Finally, we found that 38% of the 850,000 compounds in the Distributed Structure-Searchable Toxicity (DSSTox) database contain at least one reducible functional group, and the logk of 285,184 compounds could be reasonably predicted using our model. Overall, the study is a significant step toward reliable predictive tools for anticipating abiotic reduction rate constants in iron-associated reductant systems.

Investigation of the Binding Fraction of PFAS in Human Plasma and Underlying Mechanisms Based on Machine Learning and Molecular Dynamics Simulation

More than 7000 per- and polyfluorinated alkyl substances (PFAS) have been documented in the U.S. Environmental Protection Agency's CompTox Chemicals database. These PFAS can be used in a broad range of industrial and consumer applications but may pose potential environmental issues and health risks. However, little is known about emerging PFAS bioaccumulation to assess their chemical safety. This study focuses specifically on the large and high-quality data set of fluorochemicals from the related environmental and pharmaceutical chemicals databases, and machine learning (ML) models were developed for the classification prediction of the unbound fraction of compounds in plasma. A comprehensive evaluation of the ML models shows that the best blending model yields an accuracy of 0.901 for the test set. The predictions suggest that most PFAS (∼92%) have a high binding fraction in plasma. Introduction of alkaline amino groups is likely to reduce the binding affinities of PFAS with plasma proteins. Molecular dynamics simulations indicate a clear distinction between the high and low binding fractions of PFAS. These computational workflows can be used to predict the bioaccumulation of emerging PFAS and are also helpful for the molecular design of PFAS to prevent the release of high-bioaccumulation compounds into the environment.

Machine learning-driven QSAR models for predicting the mixture toxicity of nanoparticles

Research on theoretical prediction methods for the mixture toxicity of engineered nanoparticles (ENPs) faces significant challenges. The application of in silico methods based on machine learning is emerging as an effective strategy to address the toxicity prediction of chemical mixtures. Herein, we combined toxicity data generated in our lab with experimental data reported in the literature to predict the combined toxicity of seven metallic ENPs for Escherichia coli at different mixing ratios (22 binary combinations). We thereafter applied two machine learning (ML) techniques, support vector machine (SVM) and neural network (NN), and compared the differences in the ability to predict the combined toxicity by means of the ML-based methods and two component-based mixture models: independent action and concentration addition. Among 72 developed quantitative structure-activity relationship (QSAR) models by the ML methods, two SVM-QSAR models and two NN-QSAR models showed good performance. Moreover, an NN-based QSAR model combined with two molecular descriptors, namely enthalpy of formation of a gaseous cation and metal oxide standard molar enthalpy of formation, showed the best predictive power for the internal dataset (R²_test = 0.911, adjusted R²_test = 0.733, RMSE_test = 0.091, and MAE_test = 0.067) and for the combination of internal and external datasets (R²_test = 0.908, adjusted R²_test = 0.871, RMSE_test = 0.255, and MAE_test = 0.181). In addition, the developed QSAR models performed better than the component-based models. The estimation of the applicability domain of the selected QSAR models showed that all the binary mixtures in training and test sets were in the applicability domain. This study approach could provide a methodological and theoretical basis for the ecological risk assessment of mixtures of ENPs.

Machine-learning-aided application of high-gravity technology to enhance ammonia recovery of fresh waste leachate

Stripping is widely applied for the removal of ammonia from fresh waste leachate. However, the development of air stripping technology is restricted by the requirements for large-scale equipment and long operation periods. This paper describes a high-gravity technology that improves ammonia stripping from actual fresh waste leachate and a machine learning approach that predicts the stripping performance under different operational parameters. The high-gravity field is implemented in a co-current-flow rotating packed bed in multi-stage cycle series mode. The eXtreme Gradient Boosting algorithm is applied to the experimental data to predict the liquid volumetric mass transfer coefficient (K_La) and removal efficiency (η) for various rotation speeds, numbers of stripping stages, gas flow rates, and liquid flow rates. Ammonia stripping under a high-gravity field achieves η = 82.73% and K_La = 5.551 × 10^-4 s^-1 at a pH value of 10 and ambient temperature. The results suggest that the eXtreme Gradient Boosting model provides good accuracy and predictive performance, with R² values of 0.9923 and 0.9783 for K_La and η, respectively. The machine learning models developed in this study are combined with experimental results to provide more comprehensive information on rotating packed bed operations and more accurate predictions of K_La and η. The information mining behind the model is an important reference for the rational design of high-gravity-field-coupled ammonia stripping projects.

Short-term Lake Erie algal bloom prediction by classification and regression models

The recent outbreaks of harmful algal blooms in the western Lake Erie Basin (WLEB) have drawn tremendous attention to bloom prediction for better control and management. Many weekly to annual bloom prediction models have been reported, but they only employ small datasets, have limited types of input features, build linear regression or probabilistic models, or require complex process-based computations. To address these limitations, we conducted a comprehensive literature review, complied a large dataset containing chlorophyll-a index (from 2002 to 2019) as the output and a novel combination of riverine (the Maumee & Detroit Rivers) and meteorological (WLEB) features as the input, and built machine learning-based classification and regression models for 10-d scale bloom predictions. By analyzing the feature importance, we identified 8 most important features for the HAB control, including nitrogen loads, time, water levels, soluble reactive phosphorus load, and solar irradiance. Here, both long- and short-term nitrogen loads were for the first time considered in HAB models for Lake Erie. Based on these features, the 2-, 3-, and 4-level random forest classification models achieved an accuracy of 89.6%, 77.0%, and 66.7%, respectively, and the regression model achieved an R² value of 0.69. In addition, long-short term memory (LSTM) was implemented to predict temporal trends of four short-term features (N, solar irradiance, and two water levels) and achieved the Nash-Sutcliffe efficiency of 0.12-0.97. Feeding the LSTM model predictions for these features into the 2-level classification model reached an accuracy of 86.0% for predicting the HABs in 2017-2018, suggesting that we can provide short-term HAB forecasts even when the feature values are not available.

Machine learning approaches to predict the apparent rate constants for aqueous organic compounds by ferrate

The apparent second-order rate constant with hexavalent ferrate (Fe(VI)) (k_Fe(VI)) is a key indicator to evaluate the removal efficiency of a molecule by Fe(VI) oxidation. k_Fe(VI) is often determined by experiment, but such measurements can hardly catch up with the rapid growth of organic compounds (OCs). To address this issue, in this study, a total of 437 experimental second-order k_Fe(VI) rate constants at a range of conditions (pH and temperature) were used to train four machine learning (ML) algorithms (lasso regression (LR), ridge regression (RR), extreme gradient boosting (XGBoost), and the light gradient boosting machine (LightGBM)). Using the Morgan fingerprint (MF)) of a range of organic compounds (OCs) as the input, the performance of the four algorithms was comprehensively compared with respect to the coefficient of determination (R²) and root-mean-square error (RMSE). It is shown that the RR, XGBoost, and LightGBM models displayed generally acceptable performance k_Fe(VI) (R²_test > 0.7). In addition, the shapely additive explanation (SHAP) and feature importance methods were employed to interpret the XGBoost/LightGBM and RR models, respectively. The results showed that the XGBoost/LightGBM and RR models suggestd pH as the most important predictor and the tree-based models elucidate how electron-donating and electron-withdrawing groups influence the reactivity of the Fe(VI) species. In addition, the RR model share eight common features, including pH, with the two tree-based models. This work provides a fast and acceptable method for predicting k_Fe(VI) values and can help researchers better understand the degradation behavior of OCs by Fe(VI) oxidation from the perspective of molecular structure.

Classification and Regression Machine Learning Models for Predicting Aerobic Ready and Inherent Biodegradation of Organic Chemicals in Water

Machine learning (ML) is viewed as a promising tool for the prediction of aerobic biodegradation, one of the most important elimination pathways of organic chemicals from the environment. However, available models only have small datasets (<3200 records), make binary classification predictions, evaluate ready biodegradability, and do not incorporate experimental conditions (e.g., system setup and reaction time). This study addressed all these limitations by first compiling a large database of 12,750 records, considering both ready and inherent biodegradation under different conditions, and then developing regression and classification models using different chemical representations and ML algorithms. The best regression model (R² = 0.54 and root mean square error of 0.25) and classification model (the prediction accuracy from 85.1%) achieved very good performance. The model interpretation indicated that the models correctly captured the effects of chemical substructures, following the order of C═O > O═C-O > OH > CH₃ > halogen > branching > N > 6-member ring. The consideration of chemical speciation based on pK_a and α notations did not affect the regression model performance but significantly improved the classification model performance (the accuracy increased to 87.6%). The models also showed large applicability domains and provided reasonable predictions for more than 98% of over 850,000 environmentally relevant chemicals in the Distributed Structure-Searchable Toxicity database. These robust, trustable models were finally made widely accessible through two free online predictors with graphical user interface.

Predicting reaction rate constants of ozone with ionic/non-ionic compounds in water

Ozonation is a significant technology for the mitigation of pollutants in water. The second-order reaction rate constant (k_O3) of ozone (O₃) with compounds is essential for measuring their reactivity toward O₃ and understanding their fate during ozonation. However, there is a huge gap between the number of existing chemicals and the available experimental k_O3 values. Moreover, the reactivity of ionizable compounds with different ionization forms toward O₃ may differ greatly. In this study, two quantitative structure activity relationship (QSAR) models for non-ionic and ionic species, are respectively established with partial least squares (PLS) and support vector machine (SVM) methods based on the large datasets (324 non-ionic states and 188 ionic states). These models exhibit good fitting ability (non-ionic model: R²_tr > 0.760; ionic model: R²_tr > 0.780), robustness (Q²_CUM > 0.700), predictive performance (non-ionic model: R²_ext > 0.760; ionic model: R²_ext > 0.810) and wide applicability domain. The molecular parameters in two models are revealed to be significantly different, which may be attributed to the significant difference in molecular structures in two datasets and different reactivities of uncharged and charged states toward O₃. Additionally, the overall k_O3 for compounds at certain pH can be estimated by combining the two single QSAR models. These models and methods can become the effective tools for predicting the conversion rate of pollutants by O₃ in the urban sewage and drinking water treatment.

Multiple Roles of Dissolved Organic Matter in Advanced Oxidation Processes

Advanced oxidation processes (AOPs) can degrade a wide range of trace organic contaminants (TrOCs) to improve the quality of potable water or discharged wastewater effluents. Their effectiveness is impacted, however, by the dissolved organic matter (DOM) that is ubiquitous in all water sources. During the application of an AOP, DOM can scavenge radicals and/or block light penetration, therefore impacting their effectiveness toward contaminant transformation. The multiple ways in which different types or sources of DOM can impact oxidative water purification processes are critically reviewed. DOM can inhibit the degradation of TrOCs, but it can also enhance the formation and reactivity of useful radicals for contaminants elimination and alter the transformation pathways of contaminants. An in-depth analysis highlights the inhibitory effect of DOM on the degradation efficiency of TrOCs based on DOM's structure and optical properties and its reactivity toward oxidants as well as the synergistic contribution of DOM to the transformation of TrOCs from the analysis of DOM's redox properties and DOM's transient intermediates. AOPs can alter DOM structure properties as well as and influence types, mechanisms, and extent of oxidation byproducts formation. Research needs are proposed to advance practical understanding of how DOM can be exploited to improve oxidative water purification.

Transfer Learning Approach to Multitarget Temperature-Dependent Reaction Rate Prediction

Accurate prediction of temperature-dependent reaction rate constants of organic compounds is of great importance to both atmospheric chemistry and combustion science. Extensive work has been done on developing automated mechanism generation systems but the lack of quality reaction rate data remains a huge bottleneck in the application of highly detailed mechanisms. Machine learning prediction models have been recently adopted to alleviate the data gap in thermochemistry and have great potential to do the same for kinetic data with the recent release of quality reaction rate data compilations. The ultimate goal is to formulate easily accessible, general-purpose, temperature-dependent, and multitarget models for the prediction of reaction rates. To that end, we propose a model that depends on the well-known Morgan fingerprints as well as learned representations transferred from the QM9 data set. We propose the use of an Arrhenius-based loss where predictions of the three modified-Arrhenius parameters (A, n, and B = E_a/R) are given instead of the direct prediction of reaction rate constants. Our model is >35% more accurate compared to a baseline model of feed forward network (FFN) on Morgan fingerprints.

Insights into the Mechanism of Ozone Activation and Singlet Oxygen Generation on N-Doped Defective Nanocarbons: A DFT and Machine Learning Study

N-doped defective nanocarbon (N-DNC) catalysts have been widely studied due to their exceptional catalytic activity in many applications, but the O₃ activation mechanism in catalytic ozonation of N-DNCs has yet to be established. In this study, we systematically mapped out the detailed reaction pathways of O₃ activation on 10 potential active sites of 8 representative configurations of N-DNCs, including the pyridinic N, pyrrolic N, N on edge, and porphyrinic N, based on the results of density functional theory (DFT) calculations. The DFT results indicate that O₃ decomposes into an adsorbed atomic oxygen species (O_ads) and an ³O₂ on the active sites. The atomic charge and spin population on the O_ads species indicate that it may not only act as an initiator for generating reactive oxygen species (ROS) but also directly attack the organics on the pyrrolic N. On the N site and C site of the N₄V₂ system (quadri-pyridinic N with two vacancies) and the pyridinic N site at edge, O₃ could be activated into ¹O₂ in addition to ³O₂. The N₄V₂ system was predicted to have the best activity among the N-DNCs studied. Based on the DFT results, machine learning models were utilized to correlate the O₃ activation activity with the local and global properties of the catalyst surfaces. Among the models, XGBoost performed the best, with the condensed dual descriptor being the most important feature.