Synthetic biology needs better ways to design regulatory DNA while reducing the number of laboratory experiments. Our article in npj Systems Biology and Applications uses neural networks to find context-aware locations for inserting transcription-factor binding sites into promoters. The result connects machine-learning model design with biological constraints and experimental validation.
Designing synthetic promoters is still a laboratory-intensive process. This paper explores whether a neural model can learn enough sequence context to propose where a transcription-factor binding site should be inserted without breaking the original promoter logic.
The work grew from a long-term collaboration with Daniel Georgiev’s laboratory at the University of West Bohemia. My contribution connects machine-learning methods with a biological design problem, extending my applied-AI work beyond speech and language technologies.
Authors
Lukáš Kuhajda, Tomáš Honzík, Jan Švec, Daniel Georgiev
Abstract
Gene regulation through promoter engineering is a cornerstone of synthetic biology, enabling precise control over transcriptional networks. However, experimental approaches remain labor-intensive. While artificial neural networks (ANNs) have improved regulatory element prediction, tools for promoter–transcription factor binding site (TFBS) recombination are still lacking. We present an ANN framework for context-aware design of synthetic promoters in Saccharomyces cerevisiae. The model predicts optimal TFBS insertion sites and the extent of promoter rewriting needed for successful integration. Applying this, we screened 6,011 native yeast promoters for compatibility with the TetR TFBS, generating a ranked list of high-confidence promoter–TFBS pairs. Experimental validation showed that model-designed promoters achieved repression rates up to 98.4%, without prior experimental characterization or tuning. We further rewired the yeast transcriptional network by introducing glucose-dependent regulation of an essential gene via Mig1 TFBS insertion. These results establish a scalable, predictive method for engineering regulatory sequences and reprogramming transcriptional logic.