Ensemble based Data Fusion for Gene Function Prediction

The availability of an ever increasing amount of data sources due to recent advances in high throughput biotechnologies opens unprecedented opportunities for genome-wide gene function prediction. Several approaches to integrate heterogeneous sources of biomolecular data have been proposed in literature, but they suffer of drawbacks and limitations that we could in principle overcome by applying multiple classifier systems. In this work we evaluated the performances of three basic ensemble methods to integrate six different sources of high-dimensional biomolecular data. We also studied the performances resulting from the application of a simple greedy classifier selection scheme, and we finally repeated the entire experiment by introducing a feature filtering step. The experimental results show that data fusion realized by means of ensemble-based systems is a valuable research line for gene function prediction.