A New Method to Combine Heterogeneous Data Sources for Candidate Gene Prioritization

How to effectively integrate heterogeneous data sources is becoming extremely challenging, because many useful but noisy data sources are available for the problem at hand. In this paper, for disease gene prioritization problem, we investigated multiple kernels learning (MKL) and N dimensional order statistics (NDOS) method, but found that neither could effectively extract useful information from noisy data. Especially, in MKL algorithm, ineffective data source may be given more weight, which downgrades the effectiveness of the combined kernel. We proposed an improved procedure based on NDOS. We first use cross validation to evaluate each individual data source, and only effective data sources are used in the prioritizations of candidate genes.