Machine learning-assisted retention time predictions on a cellulose Tris(3,5)-dimethylphenylcarbamate column in polar organic mode

This is the authors' abstract. We don't add key points for this paper.

Background Enantioseparation in HPLC is a considerable challenge in analytical chemistry, frequently requiring numerous trials with varying experimental conditions to achieve baseline separations. To address this issue, we propose a solution that utilizes consensus modelling based on partial least squares (PLS) regression method together with neural network (NN) algorithms and a graph neural network (GNN) method to predict the retention times of compounds on Lux Cellulose-1 chiral stationary phase under various polar organic mode mobile phases. Results A homogeneous dataset was collected for the developed machine learning methods, consisting of 535 unique molecules and 1,414 retention time measurements under four polar organic mode conditions (acidic and basic methanol, acidic and basic acetonitrile). The PLS+NN consensus model showed outstanding results in condition-specific predictions, achieving R 2 values over 0.70 and RMSE values below 0.40 in most cases. Conversely, the GNN model excelled in combined predictions under all conditions, achieving a R 2 of 0.58 and RMSE CV of 0.49 during cross-validation, as well as a R 2 of 0.85 and RMSE Test of 0.25 on the test set. Significance Our research presents a novel approach for predicting chiral separations, offering an easy-to-use, open-access web tool to the scientific community. The robust GNN model was used to create a web server (https://chiralscreen.com) that enables the prediction of retention times, separation capabilities, and elution orders for various compounds. Furthermore, the software helps determine optimal initial chromatographic conditions for separations.