Predicting gene function using hierarchical multi-label decision tree ensembles.

Area of Science:

• Computational Biology
• Bioinformatics

Background:

• Automated gene function prediction for Open Reading Frames (ORFs) remains challenging in well-studied model organisms like S. cerevisiae, A. thaliana, and M. musculus.
• Existing machine learning methods vary in predictive performance, efficiency, and usability, necessitating further investigation.

Purpose of the Study:

• To evaluate decision tree-based models for predicting multiple ORF functions.
• To develop and assess an algorithm for learning hierarchical multi-label decision trees.

Main Methods:

• Developed a novel algorithm for learning hierarchical multi-label decision trees capable of predicting all ORF functions simultaneously while respecting functional hierarchies (e.g., FunCat, Gene Ontology).
• Evaluated the predictive performance of individual decision trees and ensembles against existing methods.

Main Results:

• The developed algorithm produced decision trees with superior predictive performance compared to previously described methods.
• Ensembles of decision trees demonstrated competitive accuracy with state-of-the-art statistical learning and functional linkage methods.
• The ensemble approach proved computationally efficient and user-friendly.

Conclusions:

• Decision tree-based methods, particularly ensembles, represent a state-of-the-art approach for ORF function prediction.
• These methods offer a balance of high predictive performance, computational efficiency, and ease of use.