Sparse regressions for predicting and interpreting subcellular localization of multi-label proteins

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Protein subcellular localization is crucial for understanding protein function.
• Predicting both single- and multi-location proteins is an active research area.
• Current predictors often rely on Gene Ontology (GO) terms but lack interpretability.

Purpose of the Study:

• To develop interpretable methods for predicting protein subcellular localization.
• To leverage sparse regression techniques for feature selection from GO terms.
• To differentiate between single- and multi-location proteins.

Main Methods:

• Comparison of multi-label sparse regression algorithms: multi-label LASSO (mLASSO) and multi-label elastic net (mEN).
• Application of a one-vs-rest strategy for large-scale GO term analysis.
• Utilizing GO terms to construct feature vectors for classification.

Main Results:

• Both mLASSO and mEN provide sparse and interpretable solutions for protein localization prediction.
• mLASSO identified 87 and mEN identified 429 essential GO terms out of over 8,000.
• mEN demonstrated superior performance and selected more features, particularly from biological process and molecular function categories.

Conclusions:

• Sparse regression methods (mEN and mLASSO) offer interpretable and superior performance in predicting protein subcellular localization.
• mEN outperforms mLASSO on a human benchmark dataset, highlighting its effectiveness.
• An online server, SpaPredictor, is available for utilizing these mLASSO and mEN prediction methods.