A unified framework for selecting and evaluating cell-type-specific gene co-expressions in single-cell data

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 and downstream analyses 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. In this study, we systematically compared two distinct criteria for selecting correlated gene pairs from single-cell data, p-value versus correlation strength. We found that the use of p-values instead of correlation strength is more robust for both selecting meaningful gene pairs and for the fair benchmarking of co-expression estimation methods. To make this approach universally applicable, we extended and validated a simulation method that can efficiently and reliably generate empirical p-values for co-expression estimation methods that do not have corresponding or well-controlled p-values. Furthermore, we demonstrated that a fair comparison of the estimation methods requires adjusting for the varying number of gene pairs they identified and accounting for the inherent expression-level biases within ground truth biological networks. Our study provides a practical guide for researchers to select reliable correlated gene pairs for downstream study and establishes a more rigorous standard for the evaluation and comparison of gene co-expression network estimation methods.