Motif-based model of transcription predicts effects of sequence variants in AR enhancers and reveals distinct functions for AR-associated transcription factors

Abstract

Androgen receptor (AR)-mediated transcription plays a central role in prostate cancer development and progression, yet the contributions of individual transcription factors (TFs) to AR-dependent enhancer activity remain incompletely understood. Here we use a biophysically motivated, interpretable motif-based model to dissect these contributions from STARR-seq data in LNCaP cells. By fitting the model separately to androgen inducibility and to baseline enhancer activity, we resolve TFs into three functional classes: hormone-dependent drivers, constitutive activators, and dual-role factors that contribute to both. These patterns suggest that inducibility is associated not only with the presence of AR and co-activator motifs, but also with the relative absence of constitutive activators that may saturate enhancer output. We validate the model against an independent saturation-mutagenesis dataset spanning 40 AR enhancers, predicting mutational effects at single-base resolution (AUC = 0.76), and show that direct fitting to these data independently recovers known AR regulators. Finally, we apply the model to prostate cancer GWAS risk alleles in AR binding site regions, prioritizing four candidate variants predicted to reduce the DHT/EtOH enhancer activity ratio at these loci.