Learning from Heterogeneous Data Sources: An Application in Spatial Proteomics.

Area of Science:

• Cell Biology
• Proteomics
• Bioinformatics

Background:

• Sub-cellular protein localization is crucial for cellular function and regulated by post-translational modifications.
• High-throughput mass spectrometry (MS) enables spatial proteomics, mapping protein distribution.
• Complementary data sources like microscopy and sequence data offer valuable insights.

Purpose of the Study:

• To develop a novel transfer learning classification framework for improved sub-cellular protein assignment.
• To integrate heterogeneous data sources for enhanced spatial proteomics analysis.
• To increase the quantity and quality of protein localization predictions.

Main Methods:

• Utilized a nearest-neighbor or support vector machine system for classification.
• Implemented a transfer learning approach to integrate diverse datasets.
• Evaluated algorithms on five experimental datasets across four species.

Main Results:

• Achieved high generalization accuracy in classifying proteins to tens of sub-cellular compartments.
• Successfully applied the method to identify and localize previously uncharacterized proteins in mouse embryonic stem cells.
• Validated findings against a high-resolution mouse stem cell proteome map.

Conclusions:

• The developed framework significantly improves sub-cellular protein localization prediction by integrating heterogeneous data.
• The methodology offers a robust tool for spatial proteomics research, enhancing data analysis and discovery.
• The framework is available as part of the open-source Bioconductor pRoloc suite.