Rapid Prediction of Chemical Ecotoxicity Through Genetic Algorithm Optimized Neural Network Models

Abstract

Evaluating potentially hazardous effects of chemicals on ecosystems has always been an important research topic traditionally studied using laboratory or field experiments. Experiment-based ecotoxicity test results are only available for a limited number of chemicals due to the extensive experimental effort and cost. Given the ever-increasing number of chemicals involved in the modern production process and products, rapidly characterizing chemical ecotoxicity at lower costs has become critical for guiding technology and policy development for chemical risk management. In this study, artificial neural network models are developed to predict chemical ecotoxicity (HC 50 ) based on experimental data to fill data gaps in a widely used database (USEtox). To reduce the manual tuning effort on optimal network architecture, a genetic algorithm is investigated to automatically search and configure the network architecture. The resulting neural network model reached an average test R 2 of 0.632 and had a trivial difference with the global optimal regarding validation MSE. The findings of this study can rapidly predict the ecotoxicity of chemicals and further help to understand the potential risk of chemicals and develop strategies for risk management.

Advances, Norms, and Perspectives in Product Chemical Footprint Research

“…It assesses the chemical composition of a product in a simplified way based on information available in Safety Data Sheets (SDS) [ 39 ]. Ping Hou et al developed a neural network model with an architecture optimized by a genetic algorithm to efficiently predict the ecotoxicity of chemicals (HC50 values in USEtox) [ 41 ].…”

“…The scoring system and strategy tool are qualitative calculation methods with some subjective limitations. Neural network models can quickly predict the ecotoxicity of chemicals and fill in data gaps (HC50), but the disadvantage is that extrapolated HC50 has uncertainty, and model results may not be directly applicable to risk assessment [ 41 ].…”

Machine learning models based on molecular descriptors to predict human and environmental toxicological factors in continental freshwater

“…Such models already exist, like QSAR models that are mostly linear models based on the chemical structure of compounds (Danish QSAR database (DTU, 2015), ECOSAR (Mayo-Bean et al, 2011), VEGA (Benfenati et al, 2013)) and are used to predict ecotoxicological data (LC50) needed for REACH for example. Recently, machine learning algorithms have been used to predict hazardous concentration 50% (HC50) based on 14 physico-chemical characteristics (Hou et al, 2020a) or on 691 more various variables (Hou et al, 2020b). In the case of USEtox®, despite its wide use in LCA, it only offers characterization factors for approximately 3000 chemicals and even for this limited number of compounds, 19% of ecotoxicity CFs and 67% of human toxicity CFs are missing.…”

“…Recently, machine learning algorithms have been used to predict some midpoints based on molecular descriptors and environmental parameters (Marvuglia et al, 2014 and 2015; Song et al, 2017; Lysenko et al 2018) and a first review on this subject could be found in Wu and Wang (2018). After these first works, predictions of hazardous concentration 50% (HC50) based on 14 physicochemical characteristics (Hou et al, 2020a) or on 691 more various variables (Hou et al, 2020b) were carried out. Nevertheless, their input variables need some experiments and could be difficult to collect.…”

Direct prediction of bioaccumulation of organic contaminants in plant roots from soils with machine learning models based on molecular structures

“…Machine learning and deep learning algorithms have been widely used in image recognition, natural language processing, and with chemistry applications including reaction prediction and molecular property prediction 11,12 . Recently, as big data-based assessment and decision-making tools, machine learning models were successfully applied to predict some characterization parameters of LCIA, such as chemical USEtox HC50 values [13][14][15] . However, applying these algorithms to other LCIA characterization parameters, such as in the plant-soil system, remain challenging.…”