Uncertainty-aware transcription factor activity and perturbation inference without additional training

Abstract

Understanding how transcription factors (TFs) control gene expression is essential for deciphering cellular regulatory programs. However, estimating TF activity remains challenging. Current methods rely on either broad TF-target databases, which increase false positives, or on curated sets with limited coverage. Most methods lack uncertainty quantification. Here, we introduce TFActProfiler, an integrated resource and computational framework that learns signed TF-mRNA effects by pretraining on prior TF-mRNA interaction knowledge and large RNA-seq atlases. Given an expression profile as input, TFActProfiler infers per-sample TF activity together with a confidence score that quantifies agreement between predicted and observed mRNA levels. Across TF knockdown datasets spanning multiple human cell types, TFActProfiler consistently improves TF activity inference over approaches based on unfiltered priors or narrowly curated target sets while retaining broad TF and mRNA coverage. The learned effect map further enables training-free prediction of transcriptome responses to TF perturbations. By coupling TF activity inference with training-free perturbation modeling and confidence estimates, TFActProfiler offers a tool for dissecting regulatory programs across diverse cellular states.