Multidimensional chromatographic fingerprint fusion with machine learning: Entropy-based feature evaluation for TCM...

Chromatography is a cornerstone methodology employed for quality evaluation of traditional Chinese medicine (TCM). However, the complexity of chromatographic data, where signals from multiple compounds overlap and interfere, often impedes the accurate identification of chemically significant features. This study proposed an integrated approach that combines multidimensional chromatographic fingerprinting with machine learning to trace the molecular origins of characteristic compounds in a representative TCM formula, Fufang E'jiao Jiang (FEJ). Following comprehensive chemical profiling, we constructed multi-dimensional datasets from chromatographic fingerprints, including TLC and LC-HRMS, with each dataset encompassing over 1700 features derived from retention time, m/z, and RGB values. Machine learning algorithms, such as random forest, were employed to select discriminative features, leading to the identification of 5 patterns in FEJ and 7 patterns in its intermediate products, primarily identified as ginsenosides. A simulation model further verified the significance of these features, showing that a single compound's chromatographic spot could effectively represent sample characteristics. We also introduced modified entropy values and obstacle factors to evaluate and weight the selected features. As a result, lobetyolin and ginsenoside Rf were recognized as key quality-related markers in FEJ and its intermediates, respectively. Experimental verification showed that this method can effectively deconvolute overlapping chromatographic signals and identify key quality-related features, providing an efficient and scalable computational framework for quality control in complex systems. In summary, this strategy is based on a general data structure and modular algorithm design, and hopefully to be applied to any sample system with complex chromatographic fingerprints (such as drug, environmental or food samples), without relying on specific domain knowledge.