Tailored SONAR-MSI: Converting SONAR-MS Data into Pseudoimages for Deep-Learning-Based Natural Products Analysis

LC-MS has become an essential tool for the analysis of complex samples. However, conventional MS data processing often involves cumbersome workflows and is prone to loss of information, particularly in the context of chemically diverse natural products (NPs). In this study, a novel workflow termed SONAR-MSI was established by integrating synchronized selected ion acquisition (SONAR) with pseudo-mass spectrometry imaging (MSI) and deep learning (DL) for NP quality analysis. Specifically, to enable direct application of convolutional neural networks (CNNs), a dedicated conversion protocol was established to transform SONAR-MS data into structured pseudoimages, while retaining comprehensive retention time, mass-to-charge ratio (m/z), and intensity information. Comparative evaluation revealed that SONAR significantly reduces spectral redundancy and enhances MS2 quality while minimizing data storage demands relative to conventional MS^E acquisition. As a case study, five closely related Ganoderma species were accurately classified using a SONAR-MSI-based CNN model, which achieved 100% accuracy, surpassing the performance of feature-table-based models (91.4%). Furthermore, the pixel-wise structure of SONAR-MSI allows interpretable mapping of metabolites to image coordinates, supporting both visualization and annotation. These findings establish SONAR-MSI as a robust and scalable approach for DL-assisted metabolomics, enabling efficient and information-rich NP analysis.