Development of quantitative structure property relationship models and tool for predicting the soil adsorption coefficient (logKOC)

Xianhai Yang; Yue Yang; Peter Watson; Huihui Liu

doi:10.1016/j.envpol.2025.125703

Development of quantitative structure property relationship models and tool for predicting the soil adsorption coefficient (logK_OC)

Environ Pollut. 2025 Jan 16:368:125703. doi: 10.1016/j.envpol.2025.125703. Online ahead of print.

Authors

Xianhai Yang¹, Yue Yang¹, Peter Watson², Huihui Liu³

Affiliations

¹ School of Environmental and Biological Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China.
² Los Alamos National Laboratory, New Mexico Los Alamos, 87545, USA.
³ School of Environmental and Biological Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China. Electronic address: hhliu@njust.edu.cn.

PMID: 39824331
DOI: 10.1016/j.envpol.2025.125703

Abstract

The soils/sediments organic carbon sorption coefficient (K_OC) of organic substances is one of the indispensable environmental behavioral parameters in chemicals management. Because the test procedure used to measure K_OC is normally expensive and time-consuming, predictive methods are considered vitally important technology to fill the data gap of K_OC. In this study, quantitative structure-property relationship (QSPR) models are developed using a data set with 1477 experimental logK_OC values and seven typical machine learning algorithms. We obtained three types of optimum models, i.e. one logarithm of n-octanol/water partition coefficient (logK_OW)-based univariate model, four logK_OW-based and three non logK_OW-based multi-variables machine learning models. The assessment results related to internal (goodness-of-fit and robustness) and external predictive ability indicate that all the optimum models exhibit good goodness-of-fit, robustness and predictive ability because the statistical parameters of all of those models met the accept criteria for goodness-of-fit (R²_Train > 0.700), robustness (Q²_LOO, Q²_LMO & Q²_BOOT > 0.600) and predictive ability (Q²_EXT > 0.700, CCC > 0.850, r²_m > 0.500, Δr²_m < 0.200). For convenient use, a software tool named "logKOC Predictor" was developed employing the aforementioned three types of optimum models. An external data set with 70 experimental logK_OC values was then used to test the predictive ability of the "logKOC Predictor". Results show that the tool may reliably estimate an unknown logK_OC value of a given target substance if the substance is in the applicability domain of developed models and its predicted data is marked as "High reliability".

Keywords: Machine learning; Organic substance; Predictive tool; Reliability evaluation; Soils/sediments organic carbon sorption coefficient; n-Octanol/water partition coefficient.