Abstract:In the field of environmental science and engineering, quantitative structure-activity relationship (QSAR) means the quantitative relationship between the structure of molecules (or their aggregates e.g.. nanoparticles) and certain endpoints. Herein, endpoints generally refer to physicochemical properties. biological effects or environmental behavior parameters, etc. that can be measured or modeled. Based on data sets of chemical structures and their known endpoint values (i.e., training set), QSAR models could, by means of specific algorithms, establish the mathematical relationships between the digital features that characterize the molecular structure (i.e.. descriptors) and the endpoint values. Then, the established mathematical relationships can be employed to predict the endpoint values for given chemical structures. QSAR models are important tools for filling the data gap in environmental safety of chemicals and addressing the issues from so-called "emerging pollutants" that are closely related to the improper management of chemicals. Notably. QSAR models are intrinsically data-driven models. The relationships presented in the training set are not necessarily applicable to arbitrary chemical structures. The reliability of QSAR models is always limited to certain applicability domains. Therefore, acceptance of QSAR models in sound management of chemicals requires clearly defined applicability domains. This study reviewed three concepts of the applicability domain: Descriptor domain, structural domain and mechanism domain. For characterizing descriptor domain, methods based on hyper-rectangle, convex hull, joint probability density estimation and various types of distances were described. Notably, when Boolean fingerprints are used as descriptors, these methods become meaningless. Thus. implementation_ characters and advantages of the structural domain based on fingerprints and similarity, were specially introduced. Moreover, structure-activity landscapes (SALs), as exemplified by a network-like similarity graph (NSG) and a 3D topography of the endpoint and the descriptor values, were also demonstrated. "Activity cliffs" manifested in the SALs were highlighted. Based on the NSG, local discontinuity score (LDS) was calculated to quantitatively describe the "smoothness" of the SAL. LDS was also demonstrated to be linearly correlated to the cross-validation log loss of random forest classifiers for 59 Tox21 bioassay endpoints, further implying that smoothness of the training SAL could be an indicator of the performance of the established QSAR models. The causes of the "activity cliff" were further discussed and assumed to be dependent on the complexity and spatial heterogeneity of the investigated endpoint systems. Particularly, the chaos and emergence behavior of complex systems could be sensitive to tiny but specific structural changes in the small molecules and thus cannot be satisfactorily predicted based purely on the molecular descriptors. Hence. the understanding of the mechanisms underlying the endpoints was emphasized, which corresponds to the concept of the mechanism domain. In conclusion, in order to understand the feasibility of the molecular descriptors, explain the mechanisms of the QSARs. and reasonably select defining methods for the applicability domain, it is essential to address three questions: (1) What system does the endpoint intrinsically describe? (2) How complex and spatially heterogeneous is the system? (3) Whether the endpoint considers the emergence of systems behavior? With the enhancement of chemicals management and the emphasis on the treatment of novel pollutants, the development and use of QSAR models are anticipated to be increased. Only with basic knowledge of the mechanisms underlying the modeled endpoints, with sufficient recognition of the characters of the established QSAR models, and with properly selected applicability domain characterization methods. could the QSAR models yield relatively reliable predictions and thus benefit the environmental management of chemicals in China.

Assessment of performance of the profilers provided in the OECD QSAR toolbox for category formation of chemicals

ToxProfiler: A novel human-based reporter assay for in vitro chemical safety assessment

First report on pesticide sub-chronic and chronic toxicities against dogs using QSAR and chemical read-across

Nano(Q)SAR: Challenges, Pitfalls and Perspectives

Quantitative Structure-Activity Relationship Prediction of Carcinogenicity of N-nitroso Compounds Based on Category Approach and Read-Across

In Silico Phototoxicity Prediction of Drugs and Chemicals by using Derek Nexus and QSAR Toolbox

High Throughput Read-Across for Screening a Large Inventory of Related Structures by Balancing Artificial Intelligence/Machine Learning and Human Knowledge

QSAR assessment of aquatic toxicity potential of diverse agrochemicals

Bio-QSARs 2.0: Unlocking a new level of predictive power for machine learning-based ecotoxicity predictions by exploiting chemical and biological information

Metabolism-based category formation for the prioritisation of genotoxicity hazard assessment for plant protection product residues (Part 4): α-Chloroacetamides

Molecular Similarity in Predictive Toxicology with a Focus on the q-RASAR Technique

Systematically Evaluating Read-Across Prediction and Performance Using a Local Validity Approach Characterized by Chemical Structure and Bioactivity Information

An improved method for toxicological profiling of chemical substances

Characterization of Applicability Domains for QSAR Models

Utility of in vivo metabolomics to support read-across for UVCB substances under REACH

Comparison of seven in silico tools for evaluating of daphnia and fish acute toxicity: case study on Chinese Priority Controlled Chemicals and new chemicals

Predicting oxidative stress induced by organic chemicals by using quantitative Structure–Activity relationship methods

How Precise Are Our Quantitative Structure-Activity Relationship Derived Predictions for New Query Chemicals?

Molecular similarity in chemical informatics and predictive toxicity modeling: from quantitative read-across (q-RA) to quantitative read-across structure–activity relationship (q-RASAR) with the application of machine learning

Ecotoxicological QSAR Modelling of the Acute Toxicity of Fused and Non-Fused Polycyclic Aromatic Hydrocarbons (fnfpahs) Against Two Aquatic Organisms: Consensus Modelling and Comparison with ECOSAR

ED Profiler: Machine Learning Tool for Screening Potential Endocrine-Disrupting Chemicals