Benchmarking methods for inferring single-cell transcription factor activity using large-scale perturbation sequencing data

Abstract

Abstract Transcription factor activity (TFA) is determined not solely by the expression level of the transcription factor (TF) gene itself, but is also modulated by a series of post-transcriptional regulatory processes. Although numerous computational methods have been developed to infer TFA from single-cell transcriptomic data by constructing gene regulatory networks (GRNs), a systematic and unified evaluation of these methods using high-quality experimental data remains lacking in the field. In this study, we conducted a comprehensive evaluation of eight mainstream TFA inference methods spanning three categories—prior GRN-based, de novo GRN-based, and integrated GRN-based approaches—using large-scale, high-quality single-cell perturbation sequencing (Perturb-seq) datasets. Our results demonstrate that metaTF, which employs an integrated GRN, achieves the best performance across multiple metrics, including TF coverage, predictive accuracy for perturbed cells, and accuracy for perturbed TFs. Among de novo GRN-based methods, pySCENIC exhibits predictive accuracy second only to metaTF but with lower TF coverage; meanwhile, decoupleR, a prior GRN-based method, ranks highly across all evaluated metrics. Further investigation reveals that the enrichment of reconstructed regulons within differentially expressed genes, the selection of prior GRNs and TFA scoring algorithms, and the perturbation types of target TFs are all critical factors influencing the accuracy of TFA inference. This study provides practical recommendations for the application and development of TFA inference methods.