Gene function prediction using labeled and unlabeled data

Abstract

Background: In general, gene function prediction can be formalized as a classification problem based on machine learning technique. Usually, both labeled positive and negative samples are needed to train the classifier. For the problem of gene function prediction, however, the available information is only about positive samples. In other words, we know which genes have the function of interested, while it is generally unclear which genes do not have the function, i.e. the negative samples. If all the genes outside of the target functional family are seen as negative samples, the imbalanced problem will arise because there are only a relatively small number of genes annotated in each family. Furthermore, the classifier may be degraded by the false negatives in the heuristically generated negative samples.

Results: In this paper, we present a new technique, namely Annotating Genes with Positive Samples (AGPS), for defining negative samples in gene function prediction. With the defined negative samples, it is straightforward to predict the functions of unknown genes. In addition, the AGPS algorithm is able to integrate various kinds of data sources to predict gene functions in a reliable and accurate manner. With the one-class and two-class Support Vector Machines as the core learning algorithm, the AGPS algorithm shows good performances for function prediction on yeast genes.

Conclusion: We proposed a new method for defining negative samples in gene function prediction. Experimental results on yeast genes show that AGPS yields good performances on both training and test sets. In addition, the overlapping between prediction results and GO annotations on unknown genes also demonstrates the effectiveness of the proposed method.

Figure 1

The number of genes predicted correctly for the 13 functional classes. The prediction results obtained by the five methods: AGPS, PSoL, two-class SVMs, one-class SVMs and kernel integration methods, where two-class SVMs_balanced means the results by two-class SVMs trained on balanced data and the same for kernel integration. The height of the bar in the figure means the number of genes that the five methods can recover correctly from unlabeled genes for each functional class, respectively.

Figure 2

Comparison of the five methods class by class. Comparison of the performance among the five methods, where two-class SVMs_balanced means the results by two-class SVMs trained on balanced data and the same for kernel integration. The number of classes versus one ROC score threshold is countered, and a higher curve means a better result.

Figure 3

Comparison of AGPS and PSoL class by class. Comparison of the performance between the two single-class methods, i.e. AGPS and PSoL, class by class. The ROC scores obtained by the two methods for each functional class are compared.

Figure 4

Schematic flow chart of the proposed method. Schematic flow chart of the proposed method. First, the protein interaction data, gene expression profiles and protein complex data for yeast genes are integrated into one functional linkage graph; Then, the SVD technique is utilized to project the gene vectors into low-dimensional feature space by uncovering the dominant structure of the functional linkage graph; Finally, the AGPS algorithm is utilized to predict the functions of genes.