The reconstruction of gene regulatory networks (GRN) is a complex problem that merits attention due to the need to organize time series gene expression data into meaningful structures. Data-driven techniques allow large datasets to be analyzed without imposing assumptions on the underlying mechanisms. The authors propose inferring the GRN structure using sparse low-rank matrices, an established technique in computer vision used to repair images. An excellent analogy to the use of this technique in imaging is made throughout the paper.

Typically, time series gene expression datasets are subject to drug-induced perturbations to identify the network structure. These perturbations may introduce bias and variance in the outcome of the inference algorithm. The proposed method improves on data-driven techniques in the literature, such as Bayesian networks, as it can estimate these effects and “automatically ... repair the common graph structure of a partially perturbed GRN.” It is also robust to noise due to missing and corrupted data commonly found in biological datasets.

The authors evaluated their technique using the DREAM4 in silico network challenge dataset and found that their method can complete and repair GRN structures. Although their method avoids overfitting, it “may fail to infer the true GRN when ... dynamic responses corresponding to a certain edge do not show dominant common responses, or ... show ambiguities.” According to the authors, it will be interesting to see the tool applied to “inferring the HER2+ breast cancer signaling pathway.” This paper illustrates how cross-fertilization between disciplines such as computing and genetics can lead to significant results.