OffsampleAI

OffsampleAI identifies off-sample ion images in imaging mass spectrometry to remove matrix-induced artifacts such as the MALDI matrix 2,5-dihydroxybenzoic acid (DHB) and thereby improve metabolite identification and downstream analyses.

Key Features:

• Artificial Intelligence Integration: Employs machine learning and deep learning methods to detect off-sample ions in imaging MS data.
• Gold-standard Dataset: Developed using 23,238 expert-tagged ion images from 87 public datasets in the METASPACE knowledge base.
• Residual Deep Learning: Implements residual deep learning that achieves an F1-score of 0.97 in replicating expert judgments.
• Semi-Automated Spatio-Molecular Biclustering: Uses spatio-molecular biclustering to identify off-sample ions with an F1-score of 0.96 by analyzing spatial and molecular patterns.
• Molecular Co-Localization: Applies molecular co-localization analysis to detect co-localized off-sample ion clusters with an F1-score of 0.90.
• MALDI DHB Characterization: Validated on MALDI matrix DHB to recognize and characterize properties of matrix-induced ion clusters.

Scientific Applications:

• Spatial Metabolomics Data Cleaning: Removes off-sample ion artifacts from imaging MS datasets to produce cleaner spatial metabolomics data.
• Metabolite Identification: Reduces matrix-related confounding signals to improve accuracy of metabolite assignments.
• Statistical Analysis: Provides cleaner input for statistical analyses of imaging MS data to improve reliability of results.
• Downstream Data Processing: Enhances quality of datasets used for downstream computational analyses and interpretation.

Methodology:

Computational methods include integration of open-access METASPACE data (23,238 expert-tagged ion images from 87 public datasets) and application of residual deep learning, semi-automated spatio-molecular biclustering, and molecular co-localization analysis.