Structure learning without context-specific ground truths: a case study in chronic low-dose radiation exposure in human cells

Methods for gene regulatory network (GRN) inference often leverage structural knowledge from curated databases to constrain the expansive genome-sized graph space or as labels for training, however, this knowledge does not necessarily pertain to the specific context being studied – in this case chronic low-dose radiation exposure. We show how this mismatch between existing knowledge and the context under investigation makes it difficult to tune and evaluate estimated context-specific GRNs. We provide a dataset of RNA-seq gene expression of human cells grown in lab exposed to low-dose ionizing radiation, and compare three algorithms (DAG-GNN, GENIE3, and GENELink) for estimating GRNs. We find that DAG-GNN, an unsupervised causal structure learning model, infers pathways that best align with existing literature. We also find that models that jointly learn from gene expression and radiation level data can directly estimate the genes most impacted by radiation, which greatly enhances downstream pathway analysis.