HSTXGB: a hyperparameter self-tuning XGBoost method integrating pre- and post-processing for gene regulatory network inference
TL;DR
A network structure refinement scheme, termed HSTXGB (a hyperparameter self-tuning XGBoost method integrating pre- and post-processing), to infer GRNs from time-course expression data by leveraging the nonlinear modeling capability of XGBoost while integrating prior knowledge and posterior statistics.
Abstract
Clarifying gene regulatory networks (GRNs) remains one of the central challenges of systems biology and is crucial for elucidating pathogenesis and curing diseases. Various machine learning techniques have been developed for gene regulatory network inference, but identifying intricate interactions is still a fundamental problem. Here, we propose a network structure refinement scheme, termed HSTXGB (a hyperparameter self-tuning XGBoost method integrating pre- and post-processing), to infer GRNs from time-course expression data by leveraging the nonlinear modeling capability of XGBoost while integrating prior knowledge (e.g., knockout data) and posterior statistics (e.g., regulation probabilities). Specifically, HSTXGB first calculates regulation relationship confidences using a self-tuning XGBoost model, which accounts for temporal dependencies in gene expression. Then, two novel strategies are designed to integrate information from prior data and to incorporate statistical information, which correspond to fluctuations in knockout experiments and to regulatory frequency and intensity, respectively. The confirmatory experiments on the benchmark datasets from the DREAM challenge as well as the E. coli datasets (8 networks in total) demonstrated that our HSTXGB scheme achieves significantly better performance compared with eight other state-of-the-art methods.