BFMDDT: A Decision-Tree-Based Gene Regulatory Network Inference From Multi-Type Datasets

Gene regulatory network (GRN) reconstruction remains a great challenge in computational biology and bioinformatics. However, most existing methods focus on inferring GRNs from a single type of dataset, rather than try to integrate multi-type datasets. In this work, GRN inferences based on decision tree (DT) model are improved to handle multi-type gene expression datasets and a novel inference named BFMDDT for GRN reconstruction from multi-type datasets is proposed. In BFMDDT, a specific data-integration strategy is performed on different types of datasets to extract two types of training sets, first. Then specific DT models are applied to learn the feature importance from these training sets in two ways. On the one hand, which gene(s) would regulate a specific gene is determined by one ensemble DT model. On the other hand, which gene(s) are regulated by the specific gene is determined by one ensemble DT model of the same type. At last, by mixing the feature importance gained via the two models, a global feature importance ranking is acquired and the GRN can be constructed according to the ranking. BFMDDT shows confidence in extracting information from multi-type datasets and the extensibility by combining with other technologies. Numerous experiments are conducted which show the effectiveness of data integration, the GRN reconstruction accuracy, efficiency and extensibility of BFMDDT, respectively.