Machine learning for discovering missing or wrong protein function annotations : A comparison using updated benchmark...

Area of Science:

• Bioinformatics and Computational Biology
• Machine Learning in Proteomics
• Functional Genomics

Background:

• Massive proteomic data generation necessitates automated protein function annotation.
• Existing machine learning studies often use outdated datasets (Functional Catalogue [FunCat], Gene Ontology [GO]), limiting predictive accuracy.
• This study addresses the need for updated benchmark datasets for hierarchical multi-label classification (HMC) in protein function prediction.

Purpose of the Study:

• To create and provide updated FunCat and GO yeast annotation datasets.
• To establish baseline performance results for four HMC methods on these new datasets.
• To evaluate the ability of predictive models to discover novel or incorrect annotations using updated versus old data.

Main Methods:

• Generated 24 new datasets by querying recent versions of FunCat and GO yeast annotations.
• Compared the performance of four HMC methods, including Clus-Ensemble (predictive clustering trees) and HMC-GA (genetic algorithms).
• Trained models on old data and evaluated against recent annotations to assess discovery capabilities.

Main Results:

• Clus-Ensemble outperformed more recent methods on standard evaluation tasks using the updated datasets.
• Clus-Ensemble excelled at discovering new FunCat annotations, while HMC-GA was better at detecting removed annotations.
• Similar trends were observed for GO datasets, though differences between methods were less pronounced for detecting removed annotations.

Conclusions:

• Protein function prediction remains a challenging area requiring further investigation.
• The provided updated datasets and baseline results serve as essential guidelines for future HMC research.
• Old datasets retain value for other machine learning tasks, emphasizing the importance of data versioning.