Predicting gene function using few positive examples and unlabeled ones.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Functional genomic data aids computational gene function prediction.
• Gene function prediction is typically a binary classification task.
• Limited positive examples hinder classifier training for many gene functions.

Purpose of the Study:

• To develop a novel approach (SPE_RNE) for training gene function classifiers.
• To address the challenge of imbalanced datasets in gene function prediction.

Main Methods:

• Enlarging positive examples by creating synthetic data.
• Iteratively training Support Vector Machines (SVM) to select representative negative examples.
• Training optimal SVM classifiers using grid search on combined data (Yeast protein sequence, microarray expression, protein-protein interaction, GO annotation).

Main Results:

• SPE_RNE demonstrates superior performance compared to three other methods (twoclass, twoclassbal, PSoL).
• The approach shows better generalized performance and practical prediction capacity on test and unknown gene datasets.
• Evaluated using recall (R), precision (P), and F-measure (F).

Conclusions:

• The proposed SPE_RNE method offers improved gene function prediction.
• The approach exhibits robust performance across different datasets.
• The method is adaptable for predicting gene function in other organisms, including humans.