The gene similarity score distributions for the zebrafish unfiltered anatomy-based gene networks. The networks were constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method

The gene similarity score distributions for the mouse unfiltered anatomy-based gene networks. The networks were constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method

The gene similarity score distributions for the zebrafish unfiltered integrated networks. The networks were constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method

The gene similarity score distributions for the mouse unfiltered integrated networks. The networks were constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method

The boxplot comparisons of the AUC distributions of ROC curves for zebrafish networks. The distributions for filtered PPI networks are compared with filtered anatomy-based gene networks and integrated networks constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The boxplot comparisons of the AUC distributions of precision-recall curves for zebrafish networks. The distributions for filtered PPI networks are compared with filtered anatomy-based gene networks and integrated networks constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The boxplot comparisons of the AUC distributions of ROC curves for mouse networks. The distributions for filtered PPI networks are compared with filtered anatomy-based gene networks and integrated networks constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The boxplot comparisons of the AUC distributions of precision-recall curves for mouse networks. The distributions for filtered PPI networks are compared with filtered anatomy-based gene networks and integrated networks constructed by a Lin method, b Resnik method, c Schlicker method, and d Wang method. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The network performance comparisons between non-randomized and randomized networks for zebrafish. The boxplot comparisons of the AUC distributions for a ROC and b precision–recall curves for the filtered non-randomized anatomy-based gene network, randomized profile anatomy-based gene network, and fully randomized anatomy-based gene network for the Wang method for the zebrafish. The boxplot comparisons of the AUC distributions for c ROC and d precision–recall curves for the filtered non-randomized integrated network, randomized profile integrated network, and fully randomized integrated network for the Wang method for the zebrafish. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The network performance comparisons for zebrafish networks when evaluated by randomly removed 30 anatomical entities. The boxplot comparisons of the AUC distributions for a ROC and b precision–recall curves for the filtered integrated network, PPI network, and anatomy-based gene network for the Wang method for zebrafish. The integrated network and the anatomy-based gene network were generated using the zebrafish anatomy profiles after randomly removing 30 anatomical entities, which had at least 10 gene annotations. The same 30 entities were used for the evaluation. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

The network performance comparisons for zebrafish networks when evaluated by Gene Ontology-Biological Process (GO-BP) entities. The boxplot comparisons of the AUC distributions of a ROC and b precision–recall curves for the filtered integrated network, PPI network, and anatomy-based gene network for the Wang method in zebrafish. The networks were evaluated using the annotation profiles containing GO-BP entities for the zebrafish genes. In the boxplots, the red line and the square represent the median and mean, respectively, and the name of the best performing network is underlined

A hypothetical representation showing how the network integration filters false positive interactions. This scenario compares candidate gene predictions between a a PPI network and an b anatomy-based gene network. The nodes A, B, and C (colored in black) in both networks represent three genes known to be associated with a certain anatomical entity denoted as entity 1. In the PPI network (a), genes D and F are predicted to be associated with entity 1 because genes D and F interact with genes A, B, and C that are known to be associated with the entity 1. In contrast, the anatomy-based gene network (b) only predicts D as a potential candidate for entity 1 because the gene F does not have any interaction with other genes annotated with entity 1. The absence of interactions of gene F in gene network (b) can be due to two reasons: (1) it is not annotated with any anatomical entities, (2) it is not annotated with entities that are similar to the anatomy entities associated with genes A, B, or C. The anatomy-based gene network (b) is built entirely on anatomy ontology information, thus it provides a different interaction structure. Hypothetically, the gene F could have formed false positive interactions in the PPI network, and the integrative use of the anatomy-based gene network may reduce the false positives by filtering them