Modelling and predicting liquid chromatography retention time for PFAS with no-code machine learning†

Abstract

Machine learning is increasingly popular and promising in environmental science due to its potential in solving various environmental problems. One such worldwide issue is the pollution caused by the persistent chemicals – per- and polyfluoroalkyl substances (PFAS), threatening the environment and human beings. Here, we introduce a no-code machine learning approach for modelling the quantitative structure–retention relationship (QSRR) of liquid chromatographic retention time (LC-RT) for PFAS. This approach aims to streamline the modelling process, particularly for environmental professionals who may find intensive coding cumbersome. The QSRR models were developed using the no-code machine learning tool, Orange, employing simple 2D molecular descriptors as input features. Through a systematic analysis, 12 descriptors were identified as pivotal properties essential for developing optimal models (including multiple linear regression – MLR and support vector machine – SVM). These selected models demonstrate great internal validation metrics (R² > 0.98, MAE < 6.5 s) and reasonable external robustness (R² > 0.80, MAE ∼ 40 s). Furthermore, a concise model interpretation was conducted to elucidate the molecular factors influencing LC-RT. It is anticipated that our models, capable of predicting the LC-RT for over 2000 PFAS within the Norman Network, will be instrumental in addressing this environmental challenge. This study not only contributes valuable insights into PFAS LC behaviour but also serves as a catalyst for future endeavours in the development and applications of no-code machine learning models.