AUGMENTED DOUBLY ROBUST POST-IMPUTATION INFERENCE FOR PROTEOMIC DATA.

Area of Science:

• Proteomics
• Bioinformatics
• Statistical Inference

Background:

• Mass spectrometry proteomics generates quantitative data crucial for understanding molecular mechanisms.
• Large proportions of missing values in proteomics data present significant analytical challenges.
• Existing imputation strategies can introduce bias, compromising downstream analysis validity.

Purpose of the Study:

• To develop a statistical framework for valid and efficient inference in proteomics data with missing values.
• To combine machine learning for improved imputation with debiasing techniques for accurate downstream analysis.
• To provide a robust method that maintains control over false positives while enhancing discovery.

Main Methods:

• A doubly robust estimation-inspired statistical framework.
• Integration of variational autoencoders for high-dimensional peptide data imputation.
• A parametric model for propensity score estimation to debias imputed outcomes.
• Compatibility with the double machine learning framework.

Main Results:

• The proposed framework demonstrates empirical superiority over existing methods in simulation studies.
• Application to single-cell and bulk Alzheimer's disease proteomics data yielded meaningful discoveries.
• The method effectively utilizes imputed data while maintaining good control of false positives.

Conclusions:

• The developed statistical framework offers a robust solution for analyzing proteomics data with missing values.
• This approach enhances the reliability of inferences and facilitates new discoveries in biological research.
• The method proves effective in both simulated and real-world biological datasets, including single-cell and disease-specific studies.