Proteomics and machine learning: Leveraging domain knowledge for feature selection in a skeletal muscle tissue...

Area of Science:

• Proteomics
• Bioinformatics
• Machine Learning

Background:

• Omics techniques like proteomics offer vital biological insights but are underutilized due to high dimensionality.
• Limited adoption of machine learning in proteomics hinders cross-study comparisons and data integration.

Purpose of the Study:

• To address the underutilization of proteomics data by developing a machine learning approach for high-dimensional datasets.
• To improve cross-study comparisons and facilitate the discovery of biological patterns in skeletal muscle proteomics.

Main Methods:

• Combined skeletal muscle proteomics datasets from five studies (in vitro, in vivo, adjacent tissues).
• Preprocessed data using MaxQuant and enriched with UniProt/Ensembl information.
• Utilized cellular composition to categorize data, training separate Random Forest models for each category.

Main Results:

• Integrating biological context improved model performance by reducing dimensionality and increasing signal-to-noise ratio.
• Preserved biologically relevant features, enabling discovery of patterns missed by traditional methods.
• Demonstrated suitability for diverse analyses, including biomarker discovery and classification.

Conclusions:

• Domain knowledge integration in machine learning enhances proteomics data analysis and biological pattern discovery.
• This approach facilitates tailored analyses and retains critical biological details.
• Further dataset incorporation is needed to expand clinical applications in proteomics.