Toward Machine Learning Electrospray Ionization Sensitivity Prediction for Semiquantitative Lipidomics in Stem Cells.

Area of Science:

• Analytical Chemistry
• Biochemistry
• Data Science

Background:

• Mass spectrometry (MS) is crucial for metabolomics and lipidomics but faces challenges in multiyear, multi-batch studies due to electrospray ionization (ESI) response variability.
• This variability hinders accurate batch-to-batch comparisons, creating a divide between broad discovery work and scope-limited targeted quantitation.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting ESI sensitivity of lipid classes relevant to stem cell potency.
• To enable semi-quantitation of lipids across batches in long-term stem cell studies, bridging the gap between discovery and targeted approaches.

Main Methods:

• Utilized molecular descriptors from lipid chemical structures as input for an ML model to predict ESI response.
• Validated the model using cultured stem cell samples from diverse donors, assessing performance in both positive and negative ion modes.

Main Results:

• The ML model achieved moderate accuracy (global percent errors of 40% for positive and 20% for negative modes), sufficient for semi-quantitation.
• Demonstrated high precision (16.9% positive, 7.5% negative modes), indicating potential for data harmonization across batches.
• The chosen molecular descriptors improved accuracy compared to previous literature.

Conclusions:

• The developed ML model effectively predicts ESI sensitivity, enabling semi-quantitative lipidomics in challenging multi-batch stem cell research.
• This approach offers a viable solution for lipid marker concentration estimation across batches without reliance on unavailable isotopic standards.
• The findings show significant promise for advancing data harmonization and quantitative capabilities in lipidomics.