An Interpretable Deep Learning Approach for Biomarker Detection in LC-MS Proteomics Data.

Area of Science:

• Proteomics
• Computational Biology
• Biomarker Discovery

Background:

• Deep learning (DL) in proteomics faces challenges: high dimensionality, low sample size, noise, and lack of interpretability for clinical use.
• Liquid chromatography-mass spectrometry (LC-MS) is crucial for quantifying protein mixtures but requires robust analysis methods.

Purpose of the Study:

• To present DLearnMS, a DL framework for biomarker detection in LC-MS proteomics data.
• To address challenges of DL in proteomics, including interpretability and data complexity.
• To improve biomarker identification accuracy and robustness.

Main Methods:

• Developed DLearnMS, a DL framework utilizing convolutional neural networks for LC-MS data analysis.
• Employed layer-wise relevance propagation for biomarker identification and network interpretability.
• Evaluated DLearnMS against conventional LC-MS biomarker detection approaches.

Main Results:

• DLearnMS effectively learns clinical states from LC-MS data instances.
• Biomarkers are identified through interpretable DL models, enabling detection of discriminating data regions.
• The framework achieves superior performance, reducing false positive peaks while maintaining true positive rates.

Conclusions:

• DLearnMS provides an interpretable and robust DL approach for biomarker detection in LC-MS proteomics.
• The framework simplifies analysis by eliminating the need for explicit preprocessing steps.
• DLearnMS demonstrates significant advantages over conventional methods for clinical applications.