LCMS-Net: Deep Learning for Raw High Resolution Mass Spectrometry Data Applied to Forensic Cause-of-Death Screening

Area of Science:

• Biochemistry
• Computational Biology
• Analytical Chemistry

Background:

• Untargeted metabolomics using liquid chromatography-high resolution mass spectrometry (LC-HRMS) involves complex, time-consuming preprocessing.
• Existing workflows often lack reproducibility and may miss crucial metabolite data.
• Significant domain expertise is typically required for data analysis.

Purpose of the Study:

• To introduce LCMS-Net, an end-to-end deep learning model for automated LC-HRMS data analysis.
• To address challenges of time, reproducibility, and metabolite detection in current preprocessing methods.
• To provide a more efficient and robust computational tool for metabolomics research.

Main Methods:

• Developed LCMS-Net, a deep learning model operating directly on raw LC-HRMS data.
• Explicitly modeled spatial properties of the LC-HRMS data within the model architecture.
• Validated the model on cause-of-death screening and colon cancer detection case studies.

Main Results:

• LCMS-Net achieved a 9% F1-score improvement for cause-of-death screening over OPLS-DA.
• LCMS-Net demonstrated a 1.8% F1-score improvement for colon cancer detection over DeepMSProfiler.
• The model significantly reduced batch effects, with performance differing by only 3% across different instruments.
• LCMS-Net is computationally efficient, faster, and simpler than other end-to-end deep learning methods, without requiring pretraining.

Conclusions:

• LCMS-Net offers a fully automated, reproducible, and efficient workflow for LC-HRMS data analysis.
• The model enhances metabolite detection and minimizes batch effects, improving the reliability of metabolomics studies.
• LCMS-Net represents a significant advancement in computational metabolomics, applicable to various biological and clinical research areas.