Gene Regulatory Network Inference using ensembles of Local Multiple Kernel Models
Résumé
Reconstructing gene regulatory network from high-throughput data has many potential applications, from understanding a biological organism to identifying potential drug targets. It is also a notoriously difficult problem, tackled by many scientists with various methods. In this paper, we formulate GRN inference as a sparse regression problem. We decompose the prediction of a p-genes system in p different regression problems. For each gene (target gene), we train a kernel-based regression with feature selection, predicting the expression pattern of the target gene using all the other genes (input genes). The regression will give the importance of each input gene in the prediction of the target gene. We take this importance as an indication of a putative regulatory link. Putative links are then aggregated over all genes to provide a ranking of interactions, from which we infer the GRN. Furthermore, biological data are heterogeneous. The method we propose can learn from both steady-state and time-series data, using an ensemble method that can be applied to other regression model. Finally, we compare our method, called LocKING, to state-of-the-art methods on real and realistic datasets, which are widely spread in the GRN inference community. We show that our method is competitive against individual methods. Nevertheless, best results are obtained by integrating multiple methods. We show that using LocKING among other methods significantly enhances the accuracy of the network inferred.