Relevance of different prior knowledge sources for inferring gene interaction networks

Catharina Olsen, Gianluca Bontempi, Frank Emmert-Streib, John Quackenbush, Benjamin Haibe-Kains

When inferring networks from high-throughput genomic data, one of the main challenges is the subsequent validation of these networks. In the best case scenario, the true network is partially known from previous research results published in structured databases or research articles. Traditionally, inferred networks are validated against these known interactions. Whenever the recovery rate is gauged to be high enough, subsequent high scoring but unknown inferred interactions are deemed good candidates for further experimental validation. Therefore such validation framework strongly depends on the quantity and quality of published interactions and presents serious pitfalls: (1) availability of these known interactions for the studied problem might be sparse; (2) quantitatively comparing different inference algorithms is not trivial; and (3) the use of these known interactions for validation prevents their integration in the inference procedure. The latter is particularly relevant as it has recently been showed that integration of priors during network inference significantly improves the quality of inferred networks. To overcome these problems when validating inferred networks, we recently proposed a data-driven validation framework based on single gene knock-down experiments. Using this framework, we were able to demonstrate the benefits of integrating prior knowledge and expression data. In this paper we used this framework to assess the quality of different sources of prior knowledge on their own and in combination with different genomic data sets in colorectal cancer. We observed that most prior sources lead to significant F-scores. Furthermore, their integration with genomic data leads to a significant increase in F-scores, especially for priors extracted from full text PubMed articles, known co-expression modules and genetic interactions. Lastly, we observed that the results are consistent for three different data sets: experimental knock-down data and two human tumor data sets.