Unsupervised Machine Learning for Differential Analysis in Proteomics.

Area of Science:

• Proteomics
• Biomarker Discovery
• Machine Learning

Background:

• Traditional statistical methods for differential proteomics analysis have limitations, including distributional assumptions and reliance on fold-change thresholds.
• These limitations can impact the accuracy and reliability of biomarker discovery and disease mechanism elucidation.

Purpose of the Study:

• To systematically evaluate unsupervised anomaly detection machine learning (ML) algorithms against established statistical methods for differential protein detection.
• To assess the performance of ML algorithms in terms of recall, precision, accuracy, and robustness to intersample heterogeneity.

Main Methods:

• Evaluation of 18 unsupervised anomaly detection ML algorithms using *in silico* simulated proteomic datasets.
• Probability-based transformation for cross-algorithm comparability.
• Validation using real-world proteomic data.

Main Results:

• Unsupervised ML methods, particularly the Minimum Covariance Determinant (MCD), demonstrated superior performance over statistical tests in recall, precision, and accuracy.
• MCD showed enhanced robustness to intersample heterogeneity in proteomic datasets.
• MCD-identified proteins covered canonical pathways and revealed novel tumor-associated biomolecules in real-world data.

Conclusions:

• Unsupervised ML methods, especially MCD, provide a robust and reliable alternative to traditional statistical approaches for differential proteomics analysis.
• These ML methods enhance the reliability of biomarker discovery and disease mechanism studies.
• The findings support the application of unsupervised ML in precision medicine research.