Resolution-Adaptive Binning Enhances Machine Learning Modeling by Interbatch and Multiplatform Orbitrap-Based Shotgun Mass Spectrometry Data Integration

Machine learning (ML) modeling on mass spectrometry (MS)-based shotgun data facilitates feature selection and disease modeling. However, batch-specific models often struggle with limited transferability and generalizability, necessitating data integration from multiple batches and platforms. Traditional binning methods can either disintegrate or aggregate <i>m</i>/<i>z</i> features, making data combination unreliable. In this study, we introduce a mass resolution-adaptive binning and integration strategy to overcome these challenges. This approach recovers 88-99% of ground truth features in a low mass region (70-434 <i>m</i>/<i>z</i>) from 49 mixed standard solutions at 250, 500, and 1000 ppb. Compared to conventional methods, it demonstrates stable binning and integration across low (100-450 <i>m</i>/<i>z</i>), mid (450-900 <i>m</i>/<i>z</i>), and high (900-1500 <i>m</i>/<i>z</i>) mass regions, resulting in superior predictive models. Using a mouse model of hepatocellular carcinoma as a proof-of-concept study, we identify 10 generic metabolites that showcase advancements in using ambient MS imaging (MSI) data for modeling and deploy the attained model to shotgun data. This facilitates disease detection via various sample introduction methods, including MSI on liver cryosections (<i>F</i>1 score = 0.87) and glass smears (<i>F</i>1 score = 0.80), as well as rapid direct infusion analysis (recall = 0.89 and precision = 0.63). This novel mass resolution-adaptive binning and integration strategy offers a promising approach for integrating different data sets, potentially improving disease detection accuracy in MS applications.