Computing Science. Multiple Gold Standards Address Bias in Functional Network Integration

摘要

Network integration is a widely-used method of combining large, diverse data sets. Edge weights, representing the probability that an edge actually exists, can add greatly to the value of the networks. The edge weights are usually calculated using a Gold Standard dataset. However, all Gold Standards suffer from incomplete coverage of the genome, and from bias in the type of interactions detected by different experimental techniques. Consequently the use of a single Gold Standard tends to bias the integrated network. We describe a novel Bayesian Data Fusion method for selecting and using multiple Gold Standards for scoring datasets prior to integration. We demonstrate the utility of networks scored against multiple Gold Standards for the pre-diction of Gene Ontology annotations for genes from KEGG pathways. Finally, we apply the networks to the functional prediction of genes which were uncharacterised in datasets from 2007, and evaluate the network results in the light of recent annotations.