Semi-supervised Bayesian integration of multiple spatial proteomics datasets.

Area of Science:

• Proteomics
• Systems Biology
• Bioinformatics

Background:

• Protein subcellular localization is crucial for function.
• Spatial proteomics and other omics data offer insights into protein localization.
• Existing integration methods are limited in data types and uncertainty quantification.

Purpose of the Study:

• To develop a semi-supervised Bayesian approach for integrating spatial proteomics with diverse data sources.
• To improve the inference of protein subcellular localization by quantifying prediction uncertainty.
• To offer a flexible method for integrating categorical, continuous, and temporal data.

Main Methods:

• Developed a semi-supervised Bayesian model for integrating spatial proteomics with other data.
• Inferred model parameters from labeled marker proteins and unlabeled data.
• Quantified prediction uncertainty in protein localization inference.
• Applied the method to Toxoplasma gondii spatial proteomics and cell cycle gene expression data.

Main Results:

• The proposed Bayesian approach outperforms existing transfer learning methods.
• Demonstrated flexibility in modeling various data types including annotations, abundance, and time-series expression.
• Identified protein expression programs peaking at the end of the first cell cycle in T. gondii.
• Revealed heterogeneous populations within dense granule proteins, suggesting diverse functions.

Conclusions:

• The novel Bayesian method significantly improves protein subcellular localization inference.
• The approach provides a flexible and robust framework for integrative omics analysis.
• Findings offer new insights into the functional roles and localization of T. gondii proteins.
• The method is available as the mdir R package for broader scientific use.