Local causal pathway discovery for single-cell RNA sequencing count data: a benchmark study

Aim: Recent developments in single-cell RNA sequencing (scRNAseq) and analysis have revealed regulatory behaviors not previously described using bulk analysis. scRNAseq features resolution at the level of the individual cell and provides opportunities for identifying cell type-specific gene regulatory networks. The technology promises to discover biomarkers and targeted treatments with enhanced effectiveness and reduced side effects. Pathway reverse engineering and causal algorithms have been validated in bulk sequencing transcriptomic data successfully for gene regulatory network reconstruction. In the current study, we evaluated the performance of local causal discovery algorithms for de novo reconstruction of local gene regulatory networks tailored to scRNAseq count data. Method: We benchmarked the performance of the state-of-the-art local causal discovery algorithm generalized local learning with five conditional independent tests in controlled conditions (simulated count data) and realworld single-cell RNA sequencing datasets. Results: The simulation study showed that local causal discovery methods with appropriate conditional independence tests could result in excellent discovery performance (given a sufficient sample size). As expected, various conditional independence tests possess different power-sample characteristics. The discovery performance for all tested conditional independence tests on real-world data is relatively low, potentially due to imperfect standards or deviation of simulated data distribution from real-world data. Conclusion: Our findings provide insights and practical guidance for applying causal discovery methods to singlecell RNAseq data for gene regulatory network reconstruction.