GOLabeler: improving sequence-based large-scale protein function prediction by learning to rank.

Area of Science:

• Bioinformatics
• Computational Biology
• Proteomics

Background:

• Gene Ontology (GO) annotations are essential for understanding protein functions but are experimentally determined for less than 1% of known proteins.
• Automated protein function prediction (AFP) is critical, particularly sequence-based AFP (SAFP), as protein sequences are often the sole available data.
• Existing homology-based SAFP methods struggle with 'difficult' proteins that have low sequence identity to annotated proteins.

Purpose of the Study:

• To develop an effective and efficient method for sequence-based automated protein function prediction (SAFP).
• To specifically address the challenge of predicting functions for 'difficult' proteins with limited sequence homology to known proteins.

Main Methods:

• Propose GOLabeler, a novel approach integrating five component classifiers trained on diverse sequence features.
• Features include GO term frequency, sequence alignment, amino acid trigrams, domain/motif information, and biophysical properties.
• Employ learning to rank (LTR), a machine learning paradigm effective for complex multilabel classification problems.

Main Results:

• GOLabeler demonstrates significant performance advantages over existing state-of-the-art AFP methods.
• Extensive evaluation on large-scale datasets confirms the method's efficacy and robustness.
• The integrated approach successfully captures diverse, deep-rooted information from protein sequences.

Conclusions:

• GOLabeler offers a powerful new tool for sequence-based automated protein function prediction.
• The method shows particular promise for annotating the functions of difficult-to-predict proteins.
• This work advances the field of computational proteomics by improving the accuracy and scope of protein function prediction.