Generative AI-based design of hybrid transcriptional activator proteins with new DNA-binding specificity

Abstract

Transcriptional control arises from the specific recognition of promoter DNA by transcription factors (TFs), forming the basis of cellular information processing and gene regulation. In synthetic biology, TF-promoter interactions are assembled into genetic circuits to program cellular behaviors. To ensure expected circuit performance, most synthetic genetic circuits rely on combinations of well-characterized and orthogonal transcriptional regulations. This reliance minimizes crosstalk but constrains circuit complexity and information integration. Creating hybrid TFs that combine or interpolate promoter specificities could therefore expand the design space of synthetic regulatory systems. However, it remains unclear whether hybrid functions can be created by mixing amino acid sequences, and how such functional integration could be achieved in a principled manner. Here we show that a variational autoencoder (VAE) trained on LuxR-family DNA-binding domains can generate transcription factors with hybrid and partially novel promoter recognition properties. By sampling intermediate regions of the VAE-learned latent space, we designed hybrid TFs that activate both the lux and las promoters. High-throughput sort-seq assays together with individual in vivo assays revealed that a subset of functional variants exhibited dual-responsive behavior while maintaining sequence-selective DNA recognition. Together, these results provide a data-driven strategy for exploring functional intermediate sequences between closely related proteins.