Machine-Learning-Based Prediction of Plant Cuticle-Air Partition Coefficients for Organic Pollutants: Revealing Mechanisms from a Molecular Structure Perspective

Abstract

Accurately predicting plant cuticle-air partition coefficients (K_ca) is essential for assessing the ecological risk of organic pollutants and elucidating their partitioning mechanisms. The current work collected 255 measured K_ca values from 25 plant species and 106 compounds (dataset (I)) and averaged them to establish a dataset (dataset (II)) containing K_ca values for 106 compounds. Machine-learning algorithms (multiple linear regression (MLR), multi-layer perceptron (MLP), k-nearest neighbors (KNN), and gradient-boosting decision tree (GBDT)) were applied to develop eight QSPR models for predicting K_ca. The results showed that the developed models had a high goodness of fit, as well as good robustness and predictive performance. The GBDT-2 model (Radj2 = 0.925, QLOO2 = 0.756, QBOOT2 = 0.864, Rext2 = 0.837, Qext2 = 0.811, and CCC = 0.891) is recommended as the best model for predicting K_ca due to its superior performance. Moreover, interpreting the GBDT-1 and GBDT-2 models based on the Shapley additive explanations (SHAP) method elucidated how molecular properties, such as molecular size, polarizability, and molecular complexity, affected the capacity of plant cuticles to adsorb organic pollutants in the air. The satisfactory performance of the developed models suggests that they have the potential for extensive applications in guiding the environmental fate of organic pollutants and promoting the progress of eco-friendly and sustainable chemical engineering.

Keywords: QSPR; machine learning; organic pollutants; plant cuticle–air partition coefficient.

Figure 1

Descriptor correlation graph of the QSPR models. (a) Datasets (I); (b) Datasets (II).

Figure 2

Plots of the observed versus predicted for log K_ca based on dataset (I). (a) Training set; (b) Test set; Plots of the observed versus predicted for log K_ca based on dataset (II). (c) Training set; (d) Test set.

Figure 3

Application domain characterized by Williams plots: the MLR-1 (a) and GBDT-2 (b) models for log K_ca.

Figure 4

Relationship between SHAP value and the values of different descriptors for dataset (I) (a) and dataset (II) (b).

Figure 5

Distribution of experimental data of log K_ca values. (a) Distribution of 255 experimental log K_ca values for 106 compounds; (b) Distribution of average experimental log K_ca values for 106 compounds.