Synthetic Time Series Resembling Human (HeLa) Cell-Cycle Gene Expression Data and Application to Gene Regulatory Network Discovery

Evaluation of gene regulatory network (GRN) discovery methods relies heavily on synthetic time series. However, synthetic data generated by traditional method deviate a lot from real data, making such evaluation questionable. Guiding by decaying sinusoids, we propose a new method that generates synthetic data resembling human (HeLa) cell-cycle gene expression data. Using the new synthetic data, a simple comparison between four GRN discovery methods reveals that Granger causality (GC) methods substantially outperform Pearson correlation coefficient (PCC), while time-shifted PCC can give comparable performance as GC methods. The new synthetic data generation would also be useful for generating other kinds of cell-cycle time series. Using data generated by our proposed method, evaluation of GRN discovery methods should be more trustworthy for real-data applications.