DDA/DIA mass spectrometry coupled with molecular networking and machine learning: high-coverage profiling of...

Pyrrolizidine alkaloids (PAs) are naturally occurring toxic secondary metabolites found in plants. However, the identification of PAs often relies on commercially available standards, which are expensive and limited in availability. To overcome this limitation, we developed molecular networks (MNs) from both data-dependent acquisition (DDA) and data-independent acquisition (DIA) data, utilizing specialized platforms tailored for each data type. Three approaches were employed for compound annotation: comparison with reference standards, library-based screening, and the deduction of candidate compounds based on precursor and characteristic fragment ions. Through the analysis of ten distinct herb species, DDA and DIA data led to the detection of 152 and 78 PA-like compounds, respectively. Meanwhile, several previously unreported compounds were identified, including 63 compounds in DDA data and 28 compounds in DIA data. Furthermore, a retention time (RT) prediction model was constructed using machine learning (ML) to enhance the accuracy of compound annotation. The prediction model, trained on 27 PA standards, exhibited a maximum error of 0.76 min in the test set, with an overall error within 1.5 min. The results show that there are 52.0% and 62.8% compound errors within 1.5 min for DDA and DIA data in the experiment. These combined methods significantly expand the number of detected compounds and improve the accuracy of PA annotation.