LipoCLEAN: A Machine Learning Filter to Improve Untargeted Lipid Identification Confidence.

Area of Science:

• Lipidomics
• Mass Spectrometry
• Bioinformatics

Background:

• Automated lipid identification software in untargeted lipidomics generates many putative identifications requiring manual validation.
• Current software tools do not fully leverage all available data for assessing lipid identification quality.

Purpose of the Study:

• To develop a machine learning model for holistic quality scoring of lipid identifications.
• To improve the accuracy and confidence of lipid identifications in untargeted lipidomics.

Main Methods:

• Implemented a machine learning model integrating underutilized metrics like isotope ratios and chromatographic behavior.
• Developed a multidimensional rescoring approach for lipid identification quality assessment.
• Tested the model's generalizability across different chromatography methods and mass spectrometry instruments.

Main Results:

• Approximately 50% of automated lipid identifications using tandem mass spectrometry were found to be incorrect.
• The multidimensional rescoring method reduced false discoveries to 7% while retaining 80% of true positives.
• The developed method demonstrated broad applicability across various chromatography and MS instrument families.

Conclusions:

• Machine learning-based quality scoring significantly enhances the reliability of lipid identifications.
• LipoCLEAN offers a robust solution for improving data quality in lipidomics research.
• The tool is publicly available, facilitating wider adoption and reproducibility.