DiSCGRN: Sparse Nonstationary Gene Regulatory Network Inference with Continuous-Time Propagation and Discrete Switching

Background: Reconstructing time-varying gene regulatory networks is difficult when observations are sparse, sampling intervals differ, and the timing of network changes is unknown. We aimed to jointly estimate stage-specific networks and determine whether expression time courses support a regulatory change. Methods: We developed DiSCGRN to estimate sparse stage-specific networks using an affine differential-equation model with matrix-exponential propagation over the actual sampling intervals. Stage boundaries were selected by penalized comparison of single- and two-stage models, with thresholds calibrated on independent stationary simulations. We evaluated network recovery using model-matched affine simulations, six BEELINE topologies, nonlinear Hill simulations, and original DREAM4 time series, and applied the method to mouse lung development expression data from GSE74243. Results: On model-matched affine simulations, DiSCGRN had higher mean network-recovery F1 than both full-series and split fits of four comparators, while threshold calibration reduced false segmentations in stationary test systems. In the other benchmarks, DiSCGRN achieved the highest mean F1 on four of six BEELINE topologies and the highest or tied-highest F1 in all six ten-gene Hill and DREAM4 cases. In mouse lung development, data-specific calibration supported two stages, with E18.5 selected for subsequent analysis and Meis1 and Gata2 contributing the most outgoing connections in the first and second stages, respectively. Conclusions: DiSCGRN integrates continuous-time network estimation and discrete boundary selection, allowing stage-specific regulatory differences to be examined under the conditions studied. Inferred regulatory relationships from transcription factors (TFs) to target genes in lung development provide candidates for independent experimental validation.