PFBNet: a priori-fused boosting method for gene regulatory network inference

摘要

In system biology, comprehending the intricate gene regulatory network (GRN) is of significant important, since it provides insights to understand the cell physiology, development and pathogenesis [1, 2]. With the advent of high-throughput experimental techniques such as RNA-Seq and DNA microarrays, inferring the GRN from such data at genomic scale is feasible. However, it is still a challenge due to the high-dimensional and noisy characteristics of the data, and the regulatory network may be obscured by the indirect connections. Another problem is that the samples of the data are often relatively few compared to the number of genes (i.e., the n≪p problem [3]). So far, various methods have been developed for inferring GRNs from expression data, including Bayesian Networks-based methods [4–9], information theory-based methods [6, 10–15], Ordinary Differential Equation (ODE) based methods [16–19], ensemble framework based methods [20–25], etc. Here we briefly review some algorithms that are related to our work. Among these approaches, the algorithms that under the ensemble framework have emerged as the strong players, such as GENIE [22], TIGRESS [21], BiXGBoost [25], etc. The key idea of the ensemble framework is to decompose the GRN inference problem into p feature selection subproblems (p is the number of genes in the data) and solve each subproblem with the corresponding regression model. As the regression model is selected, the confidences of the regulation relationships that from each candidate regulator (i.e., transcription factors (TFs)) to the corresponding target gene could be calculated as the feature weight. Finally, outputs from each subproblem are fused to reconstruct the GRN. Several algorithms (e.g., TIGRESS) chose the linear model to address the problem, however, they may not perform well if the data presents a higher-order structure. On the other hand, the algorithms that utilizing the nonlinear model can easily be computationally intractable as the number of the candidate regulators increase remarkably. Although these algorithms are successful, they inferred the GRN only used a single type of data (i.e., the gene expression), whereas other types of data (e.g., expression from the knockout) may provide non-redundant information about the directionality of regulatory relationships [23]. To this end, it is important to incorporate the prior information (e.g., the information from the knockout data) in GRN inference, which may lead the GRN to be more reliable and interpretable.