Negative example selection for protein function prediction: the NoGO database.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Negative examples, crucial for machine learning in protein function prediction, are underrepresented in genomic and proteomic databases.
• Current methods for selecting negative examples often rely on heuristics, lacking rigorous validation due to annotation limitations.

Purpose of the Study:

• To rigorously compare existing heuristics for negative example selection in protein function prediction.
• To introduce and evaluate novel algorithms for generating accurate negative examples, particularly in sparse annotation contexts.

Main Methods:

• Comparative analysis of heuristic-based negative example selection strategies.
• Adaptation of Positive-Unlabeled learning algorithms from text classification.
• Development of two novel algorithms: one based on empirical conditional probability and another using topic modeling.
• Evaluation using a temporal holdout and a novel validation strategy across multiple organisms.

Main Results:

• The proposed novel algorithms significantly outperform existing state-of-the-art methods in predicting negative examples.
• The new methods demonstrate a lower rate of incorrect negative example predictions.
• The study provides a benchmark for evaluating negative example prediction strategies.

Conclusions:

• The developed algorithms offer a more accurate approach to generating negative examples for protein function prediction.
• These methods can be broadly applied to enhance various protein function prediction tools.
• The NoGO database is introduced to provide a general resource of negative examples for multiple organisms.