mzLearn as a data-driven LC/MS signal detection algorithm that enables pre-trained generative models for untargeted...

Area of Science:

• Metabolomics
• Bioinformatics
• Computational Biology

Background:

• Metabolite alterations are crucial disease biomarkers.
• Large-scale untargeted metabolomics faces signal detection and data integration challenges.
• Developing pre-trained generative models requires high-quality, integrated datasets.

Purpose of the Study:

• Introduce mzLearn, a novel data-driven method for MS¹ signal detection and alignment.
• Overcome limitations in current metabolomics data processing.
• Enable the development of pre-trained foundation models for untargeted metabolomics.

Main Methods:

• mzLearn processes mzML files without user-set parameters.
• It employs a data-driven approach for signal detection and alignment.
• The method was validated across 15 public datasets and 22 studies involving 20,548 blood samples.

Main Results:

• mzLearn detected significantly more signals (11,442 avg.) compared to XCMS (7,100) and ASARI (4,655).
• Achieved higher true positive (89.0%) and lower false positive (12.5%) rates.
• Enabled development of a pre-trained variational autoencoder, improving renal cell carcinoma risk stratification and survival prediction.

Conclusions:

• mzLearn provides a robust, scalable solution for MS¹ feature detection and alignment in untargeted metabolomics.
• Facilitates the creation of high-quality feature matrices essential for foundation models.
• Advances the application of metabolomics in disease research and biomarker discovery.