Machine Learning for Group-Targeted Elution Order Prediction: Substituted Flavones as a Case Study

Prediction of the elution order of close structural analogs and isomers is a critical step in plant metabolites dereplication. The application of machine learning (ML) is an efficient approach to automate peak annotation by implementing structure-retention relationships. In this study, four ML models were trained to predict the elution order flavonoid derivative pairs containing a flavone aglycone bearing hydroxy and methoxy substituents. Additionally, retention times were indirectly estimated from model scores using linear regression and linear interpolation. Two data sets, comprising 51 compounds (1275 pairs) and 48 compounds (356 pairs), were constructed from available literature data. These data sets were used to explore alternative model training strategies and to conduct internal and external validation. A specially designed molecular fingerprint was employed to encode structural features of the flavone scaffold and its substituents, optimizing the representation for this widespread class of phytochemicals, taken as an example. Ranking neural network-based (NN) models with binary cross-entropy (BCE) and margin ranking (MR) loss functions employed a simplified version of the fingerprint, whereas logistic regression models were tested using both a condensed (20-bit) and an extended (92-bit) fingerprint that incorporated interaction between substituents. The pairwise error rates for elution order prediction were predominantly below 10%, demonstrating reliable performance under reversed-phase LC conditions with an acetonitrile gradient. Linear regression slightly outperformed the other models with the statistical significance indicated by Friedman and Wilcoxon tests. Although overall performance metrics were comparable, the use of a large uniform data set was found to be preferable over fragmented literature-derived data. The positive and negative effects of hydroxy and methoxy groups in different positions, along with their interactions on chromatographic retention, were analyzed following the visualization of model weights.