A deep learning framework for hepatocellular carcinoma diagnosis using MS1 data.

Area of Science:

• Proteomics
• Computational Biology
• Machine Learning

Background:

• Clinical proteomics is crucial for understanding disease mechanisms and identifying biomarkers.
• Accurate computational disease prediction enhances patient diagnosis and prognosis.
• Errors in peptide and protein identification hinder pathological diagnosis in proteomics.

Purpose of the Study:

• To develop an advanced deep learning model for classifying hepatocellular carcinoma (HCC) and adjacent non-tumor tissues.
• To enable direct classification from raw MS1 spectra, eliminating the need for peptide precursor identification.

Main Methods:

• Developed an end-to-end deep learning model named MS1Former.
• Utilized raw MS1 spectra for classification, focusing on subtle m/z differences.
• Validated the model on diverse datasets, including data-dependent acquisition, data-independent acquisition, and full-scan data.

Main Results:

• MS1Former accurately discriminates between HCC tumors and adjacent non-tumor tissues.
• The model demonstrates robust performance across various data acquisition methods.
• Achieved superior performance on multiple external validation datasets.
• Explored the interpretability of the deep learning model.

Conclusions:

• MS1Former offers a powerful and accurate method for HCC classification directly from MS1 spectra.
• The end-to-end framework simplifies the analysis pipeline by removing peptide identification steps.
• This approach holds promise for the classification of other tumor types in clinical proteomics.