Improving protein function prediction using protein sequence and GO-term similarities.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Current protein functional annotation relies heavily on highly similar protein sequences.
• Gene Ontology (GO) terms, especially the Biological Process ontology, are complex and challenging for computational prediction.
• Less similar proteins can still provide valuable information for functional annotation.

Purpose of the Study:

• To develop and evaluate Label-Space Dimensionality Reduction (LSDR) techniques for improved protein functional annotation.
• To address the challenge of predicting complex GO terms by transforming them into a more manageable latent representation.
• To explore the utility of less similar proteins in functional prediction.

Main Methods:

• Utilized Sequence Similarity Profile (SSP) for protein comparison against annotated training sets.
• Introduced two novel LSDR methods: one leveraging GO structure and another based on semantic similarity of terms.
• Compared LSDR methods against existing techniques and baseline approaches using cross-validation.

Main Results:

• LSDR methods, including the novel GO-aware approaches, significantly improved the Critical Assessment of Functional Annotation (CAFA) performance.
• GO-aware LSDR demonstrated superiority over generic LSDR for Arabidopsis thaliana protein annotation.
• The combination of SSP representation and k-nearest neighbors (kNN) classifier outperformed state-of-the-art and baseline methods in cross-validated F-measure.

Conclusions:

• LSDR techniques effectively exploit GO term redundancy for more accurate protein function prediction.
• Incorporating less similar proteins and GO structure enhances functional annotation capabilities.
• The proposed SSP-LSDR approach offers a robust and high-performing solution for protein function prediction.