Beyond trial-and-error: Fast in silico optimization of electromembrane extraction methods based on machine learning...

Electromembrane extraction (EME) is a rapid and solvent-efficient sample preparation technique, yet method development remains highly empirical and time-consuming, as extraction recovery depends on complex relationships between analyte properties and operational parameters. Generic EME methods provide useful starting points but cannot account for the chemical diversity encountered in real applications, and no tools currently exist to predict extraction performance or guide optimization. Here, we present the first data-driven framework for quantitative prediction and in silico optimization of EME.

Using commercial instrumentation and central composite designs, the effects of sample pH, voltage, and time were systematically explored across five generic supported liquid membranes, for 182 chemically diverse analytes in human plasma. The resulting dataset was used to develop neural network models that combine molecular descriptors with experimental conditions to estimate extraction recoveries. The models showed strong predictive performance (R²_test = 0.86-0.97; mean absolute error = 6.3-9.8% for recoveries >10%). To enable practical application, we developed EME Predictor (available at https://github.com/fredehan/EME-Predictor), a freely available desktop tool for predicting recoveries of new analytes and optimizing extraction parameters. External validation using a set of 47 compounds from forensic toxicology confirmed good model generalization, including for previously unseen compounds.

This work demonstrates how data-driven modelling can transform EME method development from trial-and-error to rational optimization. By combining machine learning with a practical desktop application, the proposed framework improves the efficiency, accessibility, and broader adoption of EME for analytical method development.