QuanFormer: A Transformer-Based Precise Peak Detection and Quantification Tool in LC-MS-Based Metabolomics.

Area of Science:

• Metabolomics
• Analytical Chemistry
• Bioinformatics

Background:

• Metabolomic analysis using liquid chromatography-mass spectrometry (LC-MS) faces challenges in accurately detecting and quantifying complex metabolite signals.
• Existing peak picking methods often yield high false positive rates, compromising the reliability and reproducibility of metabolomic results.
• Accurate quantification of metabolite peaks is crucial for identifying biomarkers and understanding biological pathways.

Purpose of the Study:

• To develop a deep learning-based method, QuanFormer, for accurate quantification of peak signals in LC-MS metabolomic data.
• To address the limitations of current peak picking algorithms by enhancing accuracy and reducing false positives.
• To evaluate QuanFormer's performance against existing methods and demonstrate its utility in clinical metabolomic studies.

Main Methods:

• Developed QuanFormer, a deep learning model integrating Convolutional Neural Networks (CNNs) for feature extraction and Transformer architecture for global computation.
• Trained QuanFormer on a dataset of nearly 20,000 annotated regions-of-interest (ROIs) using bipartite matching for unique prediction.
• Validated QuanFormer's performance on a test set, assessing accuracy in true/false peak identification and retention time shift correction.

Main Results:

• QuanFormer achieved 96.5% average precision on the test set and maintained over 90% accuracy in distinguishing true from false peaks without retraining.
• The method demonstrated superior performance compared to MZmine 3 and PeakDetective, yielding more detected peaks and higher quantification accuracy.
• QuanFormer successfully corrected retention time shifts for peak alignment and was applied to identify potential biomarkers in a breast cancer patient cohort.

Conclusions:

• QuanFormer represents a significant advancement in LC-MS-based metabolomic data analysis, offering robust and accurate peak quantification.
• The deep learning approach enhances the reliability and reproducibility of metabolomic studies, facilitating biomarker discovery.
• QuanFormer is an open-source tool, promoting wider adoption and further development in the field of metabolomics.