Interpretable Machine Learning-based Automated HPLC/MS2 Platform

Identification of unknown compounds in complex mixtures is a difficult, time consuming and challenging problem in several areas of chemistry. High-performance liquid chromatography (HPLC) coupled to tandem mass spectrometry (MS2) based on collision-activated dissociation (CAD) is a standard approach used to identify unknown compounds in complex mixtures. However, CAD often produces similar fragmentation patterns for isomeric or related ionized analytes, which makes it difficult to identify specific functional groups. MS/MS methods based on diagnostic gas-phase ion-molecule reactions provides a powerful, predictable, and reliable alternative for the identification of specific functional groups in unknown ions and differentiation of isomeric ions. However, the interpretation of the experimental results, the selection of the appropriate neutral reagents for new analytes, and the optimization of the conditions for their introduction is a manual, time-consuming and challenging process. We have developed a chemical graph-based interpretable machine learning approach that enables automated identification of functionalities in previously unknown protonated analytes. Furthermore, the approach also facilitates the automatic selection of neutral reagents for previously unstudied analytes and automatic optimization of the pulsing-in and pumping-out times for the neutral reagents introduced into the mass spectrometer. We believe that this work lays the foundation of developing automated HPLC/MS2 platforms with applications to identify unknown compounds in complex mixtures in all areas of chemical science.