Abstract
BACKGROUND: Cell-type-specific gene co-expression networks are widely used to characterize gene relationships. Although many methods have been developed to infer such co-expression networks from single-cell data, the lack of consideration of false positive control in many evaluations may lead to incorrect conclusions because higher reproducibility, higher functional coherence, and a larger overlap with known biological networks may not imply better performance if the false positives are not well controlled. RESULTS: In this study, we have developed an efficient and effective simulation tool to derive empirical p-values in co-expression inference to appropriately control false positives in assessing method performance. We studied the power of the p-value-based approach in inferring cell-type-specific co-expressions from single-cell data using both simulated and real data. We also highlight the need to adjust for random overlaps between the inferred and known networks when the number of selected correlated gene pairs varies substantially across different methods. We further illustrate the expression level bias in known biological networks and the impact of such bias in method assessment. CONCLUSION: Our study indicates the importance of controlling false positives in the inference of co-expressed genes to achieve more reliable results and proposes a simulation-based p-value method to achieve this.